Add Batch 3475e95d-44e4-448d-85c9-824d4d3369b0 data

.gitattributes

@@ -565,3 +565,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/From[[:space:]]Conversation[[:space:]]to[[:space:]]Automation_[[:space:]]Leveraging[[:space:]]LLMs[[:space:]]for[[:space:]]Problem-Solving[[:space:]]Therapy[[:space:]]Analysis/4c518f1b-21b4-4a70-b051-3b45c3f23186_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/From[[:space:]]Evasion[[:space:]]to[[:space:]]Concealment_[[:space:]]Stealthy[[:space:]]Knowledge[[:space:]]Unlearning[[:space:]]for[[:space:]]LLMs/339bf05c-b046-48ee-993a-9fb0a03bab03_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/From[[:space:]]Heart[[:space:]]to[[:space:]]Words_[[:space:]]Generating[[:space:]]Empathetic[[:space:]]Responses[[:space:]]via[[:space:]]Integrated[[:space:]]Figurative[[:space:]]Language[[:space:]]and[[:space:]]Semantic[[:space:]]Context[[:space:]]Signals/5b0d4480-5078-4bcb-a57a-a23f9d188d9c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/From[[:space:]]Imitation[[:space:]]to[[:space:]]Introspection_[[:space:]]Probing[[:space:]]Self-Consciousness[[:space:]]in[[:space:]]Language[[:space:]]Models/2f324e05-4784-4971-a334-1f576bd545f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/From[[:space:]]Misleading[[:space:]]Queries[[:space:]]to[[:space:]]Accurate[[:space:]]Answers_[[:space:]]A[[:space:]]Three-Stage[[:space:]]Fine-Tuning[[:space:]]Method[[:space:]]for[[:space:]]LLMs/5dd6caac-d818-49db-bbbd-1a5eed236022_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/From[[:space:]]Observation[[:space:]]to[[:space:]]Understanding_[[:space:]]Front-Door[[:space:]]Adjustments[[:space:]]with[[:space:]]Uncertainty[[:space:]]Calibration[[:space:]]for[[:space:]]Enhancing[[:space:]]Egocentric[[:space:]]Reasoning[[:space:]]in[[:space:]]LVLMs/a56c3926-e95f-40d6-b848-790453a6e4b3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/From[[:space:]]Perception[[:space:]]to[[:space:]]Reasoning_[[:space:]]Enhancing[[:space:]]Vision-Language[[:space:]]Models[[:space:]]for[[:space:]]Mobile[[:space:]]UI[[:space:]]Understanding/5314ac89-52e2-43c4-b7c2-57f8c0282998_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/From[[:space:]]Phrases[[:space:]]to[[:space:]]Subgraphs_[[:space:]]Fine-Grained[[:space:]]Semantic[[:space:]]Parsing[[:space:]]for[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Question[[:space:]]Answering/f27726f1-38f6-4ebc-b227-36fba1f0a97b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/From[[:space:]]Specific-MLLMs[[:space:]]to[[:space:]]Omni-MLLMs_[[:space:]]A[[:space:]]Survey[[:space:]]on[[:space:]]MLLMs[[:space:]]Aligned[[:space:]]with[[:space:]]Multi-modalities/ac867663-37ff-4910-9a55-16ebe121303b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Full-Step-DPO_[[:space:]]Self-Supervised[[:space:]]Preference[[:space:]]Optimization[[:space:]]with[[:space:]]Step-wise[[:space:]]Rewards[[:space:]]for[[:space:]]Mathematical[[:space:]]Reasoning/e3c1942a-6072-41ae-9020-bff00ee3eef1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Fuzzy[[:space:]]Speculative[[:space:]]Decoding[[:space:]]for[[:space:]]a[[:space:]]Tunable[[:space:]]Accuracy-Runtime[[:space:]]Tradeoff/b38a18f9-3293-4583-8870-5f7468ebd1d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/G2S_[[:space:]]A[[:space:]]General-to-Specific[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Temporal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Forecasting[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/e8fefd89-b727-4b85-babb-df829aebdf89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GA-S3_[[:space:]]Comprehensive[[:space:]]Social[[:space:]]Network[[:space:]]Simulation[[:space:]]with[[:space:]]Group[[:space:]]Agents/30f6dd60-def5-4b2a-b93a-f30d01fc7354_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GEMS_[[:space:]]Generation-Based[[:space:]]Event[[:space:]]Argument[[:space:]]Extraction[[:space:]]via[[:space:]]Multi-perspective[[:space:]]Prompts[[:space:]]and[[:space:]]Ontology[[:space:]]Steering/05fcfb90-a8f2-4bcb-963f-c0500cd5bd2f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GIMMICK_[[:space:]]Globally[[:space:]]Inclusive[[:space:]]Multimodal[[:space:]]Multitask[[:space:]]Cultural[[:space:]]Knowledge[[:space:]]Benchmarking/395b7673-d78e-432a-a3e3-6e5fc6869bb5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GLTW_[[:space:]]Joint[[:space:]]Improved[[:space:]]Graph[[:space:]]Transformer[[:space:]]and[[:space:]]LLM[[:space:]]via[[:space:]]Three-Word[[:space:]]Language[[:space:]]for[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Completion/26f5a036-742f-4f92-8586-7e6447a37c4f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GLiM_[[:space:]]Integrating[[:space:]]Graph[[:space:]]Transformer[[:space:]]and[[:space:]]LLM[[:space:]]for[[:space:]]Document-Level[[:space:]]Biomedical[[:space:]]Relation[[:space:]]Extraction[[:space:]]with[[:space:]]Incomplete[[:space:]]Labeling/227aacee-6194-44d3-aa25-f79ce4a82c0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GNN-RAG_[[:space:]]Graph[[:space:]]Neural[[:space:]]Retrieval[[:space:]]for[[:space:]]Efficient[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Reasoning[[:space:]]on[[:space:]]Knowledge[[:space:]]Graphs/a11299d3-8957-43a6-bf01-224a6c44868b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GOLFer_[[:space:]]Smaller[[:space:]]LMs-Generated[[:space:]]Documents[[:space:]]Hallucination[[:space:]]Filter[[:space:]]&[[:space:]]Combiner[[:space:]]for[[:space:]]Query[[:space:]]Expansion[[:space:]]in[[:space:]]Information[[:space:]]Retrieval/bcce65d0-92a0-4c8a-9026-16893c6c1e2d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GOODLIAR_[[:space:]]A[[:space:]]Reinforcement[[:space:]]Learning-Based[[:space:]]Deceptive[[:space:]]Agent[[:space:]]for[[:space:]]Disrupting[[:space:]]LLM[[:space:]]Beliefs[[:space:]]on[[:space:]]Foundational[[:space:]]Principles/fba9c026-e6ff-4ea8-9d6d-22e76355400c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GRAF_[[:space:]]Graph[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]by[[:space:]]Facts[[:space:]]for[[:space:]]Romanian[[:space:]]Legal[[:space:]]Multi-Choice[[:space:]]Question[[:space:]]Answering/c3fcf348-45a9-4417-a5c6-8918603635e8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GRAMMAR-LLM_[[:space:]]Grammar-Constrained[[:space:]]Natural[[:space:]]Language[[:space:]]Generation/a0bf77f8-5651-4938-a735-569b2f832bb1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GRI-QA_[[:space:]]a[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]for[[:space:]]Table[[:space:]]Question[[:space:]]Answering[[:space:]]over[[:space:]]Environmental[[:space:]]Data/1da81c00-2314-405f-9146-3e4049fd615f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GRNFormer_[[:space:]]A[[:space:]]Biologically-Guided[[:space:]]Framework[[:space:]]for[[:space:]]Integrating[[:space:]]Gene[[:space:]]Regulatory[[:space:]]Networks[[:space:]]into[[:space:]]RNA[[:space:]]Foundation[[:space:]]Models/5d7cc1f1-4e80-4d9e-bc5f-49d9dae70532_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GSQ-Tuning_[[:space:]]Group-Shared[[:space:]]Exponents[[:space:]]Integer[[:space:]]in[[:space:]]Fully[[:space:]]Quantized[[:space:]]Training[[:space:]]for[[:space:]]LLMs[[:space:]]On-Device[[:space:]]Fine-tuning/3ff101eb-1599-4790-b659-49feceb4a914_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GUI[[:space:]]Agents_[[:space:]]A[[:space:]]Survey/b04fb9e4-73cc-4a9b-957c-06fb3a612f92_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GUIDEX_[[:space:]]Guided[[:space:]]Synthetic[[:space:]]Data[[:space:]]Generation[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Information[[:space:]]Extraction/e68993f0-566a-418d-a2fe-9649a4da42b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GUM-SAGE_[[:space:]]A[[:space:]]Novel[[:space:]]Dataset[[:space:]]and[[:space:]]Approach[[:space:]]for[[:space:]]Graded[[:space:]]Entity[[:space:]]Salience[[:space:]]Prediction/92d2e5a0-cdf1-426c-ace8-4ed4b35e3209_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GaRAGe_[[:space:]]A[[:space:]]Benchmark[[:space:]]with[[:space:]]Grounding[[:space:]]Annotations[[:space:]]for[[:space:]]RAG[[:space:]]Evaluation/d44ede99-a7f5-4b3d-9ebf-d5010e148dd2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GeAR_[[:space:]]Generation[[:space:]]Augmented[[:space:]]Retrieval/5d20f6c6-be8d-48da-a97a-11e445e81208_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GeAR_[[:space:]]Graph-enhanced[[:space:]]Agent[[:space:]]for[[:space:]]Retrieval-augmented[[:space:]]Generation/364d33d3-5dba-4b45-88e3-33830ed86b64_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GeNRe_[[:space:]]A[[:space:]]French[[:space:]]Gender-Neutral[[:space:]]Rewriting[[:space:]]System[[:space:]]Using[[:space:]]Collective[[:space:]]Nouns/bad03938-dff9-491e-8586-c9088cd5b6af_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GenTool_[[:space:]]Enhancing[[:space:]]Tool[[:space:]]Generalization[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Zero-to-One[[:space:]]and[[:space:]]Weak-to-Strong[[:space:]]Simulation/9117a6cb-1065-4c97-8ac0-8e3319f85045_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generalizable[[:space:]]Cross-Lingual[[:space:]]Cognitive[[:space:]]Distortion[[:space:]]Detection[[:space:]]with[[:space:]]Standardized[[:space:]]Annotations[[:space:]]and[[:space:]]Multi-Task[[:space:]]Learning/91d94c16-b49d-4242-9d20-9f6879562f88_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generate,[[:space:]]Discriminate,[[:space:]]Evolve_[[:space:]]Enhancing[[:space:]]Context[[:space:]]Faithfulness[[:space:]]via[[:space:]]Fine-Grained[[:space:]]Sentence-Level[[:space:]]Self-Evolution/cc17ab91-a269-4f51-843d-ea77274f418d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generating[[:space:]]Domain-Specific[[:space:]]Knowledge[[:space:]]Graphs[[:space:]]from[[:space:]]Large[[:space:]]Language[[:space:]]Models/1feadd32-b928-459e-b1fa-31c3bc67c0c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generating[[:space:]]Pedagogically[[:space:]]Meaningful[[:space:]]Visuals[[:space:]]for[[:space:]]Math[[:space:]]Word[[:space:]]Problems_[[:space:]]A[[:space:]]New[[:space:]]Benchmark[[:space:]]and[[:space:]]Analysis[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Models/b2d43f3c-65a1-44e9-b5bd-531e50c30307_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generating[[:space:]]Questions,[[:space:]]Answers,[[:space:]]and[[:space:]]Distractors[[:space:]]for[[:space:]]Videos_[[:space:]]Exploring[[:space:]]Semantic[[:space:]]Uncertainty[[:space:]]of[[:space:]]Object[[:space:]]Motions/01f73e84-907b-406b-8246-1e2395e03d6b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generative[[:space:]]Error[[:space:]]Correction[[:space:]]for[[:space:]]Emotion-aware[[:space:]]Speech-to-text[[:space:]]Translation/d91aa81b-4728-4301-93a3-61edfe9da98e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generative[[:space:]]Frame[[:space:]]Sampler[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/9d901473-4b1b-4596-90e8-2aa14988f25a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generative[[:space:]]Music[[:space:]]Models’[[:space:]]Alignment[[:space:]]with[[:space:]]Professional[[:space:]]and[[:space:]]Amateur[[:space:]]Users’[[:space:]]Expectations/65fd7a19-c843-46dc-87a4-3ec87860bf6e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GlyphPattern_[[:space:]]An[[:space:]]Abstract[[:space:]]Pattern[[:space:]]Recognition[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/923970c8-9095-448a-9a28-bfd5a6d5d564_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Going[[:space:]]Beyond[[:space:]]Your[[:space:]]Expectations[[:space:]]in[[:space:]]Latency[[:space:]]Metrics[[:space:]]for[[:space:]]Simultaneous[[:space:]]Speech[[:space:]]Translation/dc82a985-2317-4a0e-a172-b2041f92d844_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Grammar-Based[[:space:]]Code[[:space:]]Representation_[[:space:]]Is[[:space:]]It[[:space:]]a[[:space:]]Worthy[[:space:]]Pursuit[[:space:]]for[[:space:]]LLMs_/5dd6c6e4-9106-4309-883d-7a58924e621d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Graph-Assisted[[:space:]]Culturally[[:space:]]Adaptable[[:space:]]Idiomatic[[:space:]]Translation[[:space:]]for[[:space:]]Indic[[:space:]]languages/5d32ac03-563a-408b-82ba-5f6ccd916723_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Graph-guided[[:space:]]Cross-composition[[:space:]]Feature[[:space:]]Disentanglement[[:space:]]for[[:space:]]Compositional[[:space:]]Zero-shot[[:space:]]Learning/2e4c46d1-3eb7-4c3b-a0cc-b6fd6c0efb5d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Grounding[[:space:]]Task[[:space:]]Assistance[[:space:]]with[[:space:]]Multimodal[[:space:]]Cues[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Demonstration/fd75cdaf-88e1-4b84-8546-723c178a6bd3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Group[[:space:]]then[[:space:]]Scale_[[:space:]]Dynamic[[:space:]]Mixture-of-Experts[[:space:]]Multilingual[[:space:]]Language[[:space:]]Model/16041f0b-67ae-4499-80eb-fce37141c533_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HASH-RAG_[[:space:]]Bridging[[:space:]]Deep[[:space:]]Hashing[[:space:]]with[[:space:]]Retriever[[:space:]]for[[:space:]]Efficient,[[:space:]]Fine[[:space:]]Retrieval[[:space:]]and[[:space:]]Augmented[[:space:]]Generation/ad8260a5-9985-4e4f-a59e-eabbcb4d270e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HATA_[[:space:]]Trainable[[:space:]]and[[:space:]]Hardware-Efficient[[:space:]]Hash-Aware[[:space:]]Top-k[[:space:]]Attention[[:space:]]for[[:space:]]Scalable[[:space:]]Large[[:space:]]Model[[:space:]]Inference/f51d22ce-f5c1-4725-a53f-81e8b63cc767_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HG-InsightLog_[[:space:]]Context[[:space:]]Prioritization[[:space:]]and[[:space:]]Reduction[[:space:]]for[[:space:]]Question[[:space:]]Answering[[:space:]]with[[:space:]]Non-Natural[[:space:]]Language[[:space:]]Construct[[:space:]]Log[[:space:]]Data/d5c5542b-1a33-46a2-b384-380e28c02810_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HICD_[[:space:]]Hallucination-Inducing[[:space:]]via[[:space:]]Attention[[:space:]]Dispersion[[:space:]]for[[:space:]]Contrastive[[:space:]]Decoding[[:space:]]to[[:space:]]Mitigate[[:space:]]Hallucinations[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/fbf29dc6-4883-490f-84ac-cac9e283c10c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HPSS_[[:space:]]Heuristic[[:space:]]Prompting[[:space:]]Strategy[[:space:]]Search[[:space:]]for[[:space:]]LLM[[:space:]]Evaluators/44b9bd1d-2d98-4f90-b9cb-0b5497b2cecb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HTML_[[:space:]]Hierarchical[[:space:]]Topology[[:space:]]Multi-task[[:space:]]Learning[[:space:]]for[[:space:]]Semantic[[:space:]]Parsing[[:space:]]in[[:space:]]Knowledge[[:space:]]Base[[:space:]]Question[[:space:]]Answering/b1247f39-52d5-413c-92c7-37a66ddcf062_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HammerBench_[[:space:]]Fine-Grained[[:space:]]Function-Calling[[:space:]]Evaluation[[:space:]]in[[:space:]]Real[[:space:]]Mobile[[:space:]]Assistant[[:space:]]Scenarios/7de62520-1411-4e1a-b2b9-b75487ccccbb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Harnessing[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Disaster[[:space:]]Management_[[:space:]]A[[:space:]]Survey/188aa5d0-4613-4958-8996-c0255d417513_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Harnessing[[:space:]]PDF[[:space:]]Data[[:space:]]for[[:space:]]Improving[[:space:]]Japanese[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models/31a887c6-5a58-441f-a916-fece62d0578b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Harnessing[[:space:]]Whisper[[:space:]]for[[:space:]]Prosodic[[:space:]]Stress[[:space:]]Analysis/32577b24-5202-4c5b-9592-285b1c2520f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HatePRISM_[[:space:]]Policies,[[:space:]]Platforms,[[:space:]]and[[:space:]]Research[[:space:]]Integration.[[:space:]]Advancing[[:space:]]NLP[[:space:]]for[[:space:]]Hate[[:space:]]Speech[[:space:]]Proactive[[:space:]]Mitigation/03f9e079-d750-4b21-bf4e-e0a8539ec1f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hatevolution_[[:space:]]What[[:space:]]Static[[:space:]]Benchmarks[[:space:]]Don’t[[:space:]]Tell[[:space:]]Us/1b46a843-5880-44de-873d-7e177e1d1742_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HellaSwag-Pro_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Bilingual[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]the[[:space:]]Robustness[[:space:]]of[[:space:]]LLMs[[:space:]]in[[:space:]]Commonsense[[:space:]]Reasoning/af9b0dec-5c6f-4661-9b0c-cd6165e3e881_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Heuristic-based[[:space:]]Search[[:space:]]Algorithm[[:space:]]in[[:space:]]Automatic[[:space:]]Instruction-focused[[:space:]]Prompt[[:space:]]Optimization_[[:space:]]A[[:space:]]Survey/67b725ba-bb4c-4ca5-a093-e5009b2b1517_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HiCOT_[[:space:]]Improving[[:space:]]Neural[[:space:]]Topic[[:space:]]Models[[:space:]]via[[:space:]]Optimal[[:space:]]Transport[[:space:]]and[[:space:]]Contrastive[[:space:]]Learning/1079a63a-5332-4951-8042-f09090703208_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hierarchical[[:space:]]Retrieval[[:space:]]with[[:space:]]Evidence[[:space:]]Curation[[:space:]]for[[:space:]]Open-Domain[[:space:]]Financial[[:space:]]Question[[:space:]]Answering[[:space:]]on[[:space:]]Standardized[[:space:]]Documents/2b87a31a-cf99-4976-a777-7782ab2aeeab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hierarchical[[:space:]]Safety[[:space:]]Realignment_[[:space:]]Lightweight[[:space:]]Restoration[[:space:]]of[[:space:]]Safety[[:space:]]in[[:space:]]Pruned[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/35c743d5-a0bb-463f-9964-81f45ca9d989_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HopRAG_[[:space:]]Multi-Hop[[:space:]]Reasoning[[:space:]]for[[:space:]]Logic-Aware[[:space:]]Retrieval-Augmented[[:space:]]Generation/423ae31f-6da0-410c-b336-caa77e607f7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Do[[:space:]]LLMs[[:space:]]Acquire[[:space:]]New[[:space:]]Knowledge_[[:space:]]A[[:space:]]Knowledge[[:space:]]Circuits[[:space:]]Perspective[[:space:]]on[[:space:]]Continual[[:space:]]Pre-Training/f854d361-1b31-40e7-b37c-463370577a06_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/From Imitation to Introspection_ Probing Self-Consciousness in Language Models/2f324e05-4784-4971-a334-1f576bd545f0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d07e62f08d37bd70b9971232149f9b250fee4845296da9a37d2db1f086b5410
+size 4211105

2025/From Imitation to Introspection_ Probing Self-Consciousness in Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:864b25e6df9118bcde8d3f596eb021b269f59468d76d1f55483268e22540416c
+size 1956855

2025/From Misleading Queries to Accurate Answers_ A Three-Stage Fine-Tuning Method for LLMs/5dd6caac-d818-49db-bbbd-1a5eed236022_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d85a1600f54f0bfa4d82ad8c899fe65aacfa580c60ef4f197bd4aac24decf27
+size 3237401

2025/From Misleading Queries to Accurate Answers_ A Three-Stage Fine-Tuning Method for LLMs/full.md

@@ -0,0 +1,500 @@
+# From Misleading Queries to Accurate Answers: A Three-Stage Fine-Tuning Method for LLMs
+
+Guocong Li $^{1}$ , Weize Liu $^{1}$ , Yihang Wu $^{1}$ , Ping Wang $^{2}$ , Shuaihan Huang $^{1}$ , Hongxia Xu $^{1,3*}$ , Jian Wu $^{1*}$
+
+<sup>1</sup>Zhejiang University <sup>2</sup>Renmin University of China
+
+$^{3}$ AI Research Center, WeDoctor Cloud
+
+guocong_li@163.com weizeliu1115@gmail.com
+
+{einstein,wujian2000}@zju.edu.cn
+
+# Abstract
+
+Large language models (LLMs) exhibit excellent performance in natural language processing (NLP), but remain highly sensitive to the quality of input queries, especially when these queries contain misleading or inaccurate information. Existing methods focus on correcting the output, but they often overlook the potential of improving the ability of LLMs to detect and correct misleading content in the input itself. In this paper, we propose a novel three-stage fine-tuning method that enhances the ability of LLMs to detect and correct misleading information in the input, further improving response accuracy and reducing hallucinations. Specifically, the three stages include (1) training LLMs to identify misleading information, (2) training LLMs to correct the misleading information using built-in or external knowledge, and (3) training LLMs to generate accurate answers based on the corrected queries. To evaluate our method, we conducted experiments on three datasets for the hallucination detection task and the question answering (QA) task, as well as two datasets containing misleading information that we constructed. The experimental results demonstrate that our method significantly improves the accuracy and factuality of LLM responses, while also enhancing the ability to detect hallucinations and reducing the generation of hallucinations in the output, particularly when the query contains misleading information.
+
+# 1 Introduction
+
+LLMs have achieved significant success in the field of NLP, particularly excelling in NLG, where they are capable of generating coherent and reasonable responses based on user queries (Li et al., 2024b; Xuanfan and Piji, 2023). Although LLMs have
+
+![](images/58b2a402c1a63f47b122633785a6e742000f1de193ea3f102ce915ce95bc2fe7.jpg)  
+Figure 1: When LLMs directly answer a query containing misleading information (Left side), they produce incorrect responses. However, when LLMs identify and correct the misleading information in the query before generating a response, they correctly answer it (Right side). The misleading information is marked in red, while the corresponding corrections are marked in blue.
+
+made significant progress in generation ability, they are highly sensitive to the quality of the input, and the quality of the generated responses is highly susceptible (Pan et al., 2023; Wu et al., 2024a). Users often provide LLMs with queries that are imprecise or may even contain various forms of misleading information (Nie et al., 2020). Misleading information refers to inaccurate content within the input that can mislead the model during its reasoning and generation processes. When such information is present in the input, the likelihood of the model generating incorrect responses or hallucinations increases significantly (Yu et al., 2024; Song et al., 2024). This challenge is particularly prominent in knowledge-intensive domains such as medicine, law, and education, where inputs often contain inaccurate information, and the accuracy of responses is critically important (Barnard et al., 2023). Therefore, mitigating the impact of misleading information in the input on the performance of LLMs is an urgent and important issue that needs to be addressed.
+
+Current training methods for LLMs primarily
+
+focus on correcting the output after generation. An important research direction is the use of the retrieval-augmented generation (RAG) framework, which leverages external knowledge bases to verify and improve the output (Lewis et al., 2020; Gao et al., 2023). However, RAG typically requires retrieval from large external knowledge bases, a process that can be time-consuming and may not meet the low-latency requirements, especially in real-time application scenarios (Li et al., 2024a; Jin et al., 2024). In addition, the self-correction mechanism has also garnered attention, where LLMs iteratively generate and self-evaluate to critique and correct their own outputs (Madaan et al., 2024). This method has the potential to improve the accuracy and quality of the results, but it also has the drawback that the effectiveness of the correction depends on the model's accuracy in evaluating the results and may require external feedback for assistance (Kamoi et al., 2024). These existing methods focus on correcting the content generated by LLMs, without considering the correction of misleading information in the input. Although the presence of misleading information in the input may lead to incorrect responses, our proposed method is capable of identifying and detecting errors in the query, subsequently generating the correct answer, as shown in Figure 1. However, no existing work has considered improving the performance of LLMs by enhancing their ability to identify and correct misleading questions.
+
+Therefore, we propose a novel three-stage finetuning method that enhances LLMs' ability to detect and correct misleading information in the input. This method enhances the accuracy and robustness of LLMs in processing inputs containing misleading information, while mitigating the negative impact of misleading information on model outputs.
+
+The fine-tuning method that we propose consists of the following three training stages:
+
+1) Query detection training: Training LLMs to determine whether the input query contains misleading information based on built-in knowledge or external knowledge;  
+2) Query correction training: Training LLMs to combine built-in knowledge or external knowledge to correct the misleading information in the query and generate the correct query;  
+3) Query answering training: Training LLMs to generate accurate and reliable answers based on the corrected query.  
+To evaluate the performance of LLMs in answer
+
+ing queries containing misleading information, we construct two misleading datasets, HaluEval-  $QA_{mis}$  and  $CQA_{mis}$ , based on the  $QA$  subset of the HaluEval dataset and the  $CQA$  dataset, respectively. Subsequently, we conduct experiments on the misleading datasets, as well as on the three original datasets for the hallucination detection task and the question answering task. The experimental results show that when input contains misleading information, the performance of the trained model using our method is much higher than that of the original model, demonstrating that our method effectively reduces the impact of misleading information in the input on model performance. Furthermore, the model trained using our method also detects that a portion of questions in commonly used standard datasets contain misleading information. Removing these misleading questions can significantly improve the performance of the original model, whereas the model trained with our method consistently maintains stable performance, demonstrating its robustness to misleading information in the input. Regardless of whether the input contains misleading information, our method significantly improves the accuracy and factuality of LLM responses, as well as the ability to detect hallucinations in the responses, while effectively mitigating the hallucinations of the model.
+
+# 2 A Fine-Tuning Method to Mitigate the Impact of Misleading Information
+
+The framework of the proposed method is illustrated in Figure 2. It involves training LLMs to sequentially perform the following tasks: detecting misleading information in the input query, correcting the misleading query, and generating accurate answers based on the corrected query, thereby improving the model's accuracy and robustness when handling inputs containing misleading information.
+
+# 2.1 Construction of Misleading Query Datasets
+
+We used the Qwen2.5-72B-Instruct model to generate a dataset containing misleading information based on the original queries. Specifically, we constructed the misleading query dataset HaluEval  $QA_{mis}$  by augmenting the queries from the QA subset of the HaluEval dataset with misleading content, leveraging the knowledge from the original queries in the HaluEval dataset. Similarly, based on the CQA dataset, we used the Qwen2.5-
+
+![](images/4c1d8851cfc46d68b8502c81c086b05776f26c370d80dc8667a8c25e23fe7518.jpg)  
+Figure 2: The framework diagram of our method. In the first stage, we train LLMs to determine whether the query contains misleading information. In the second stage, for queries containing misleading information, we train LLMs to correct the misleading information in the query to obtain a correct query. In the third stage, we train LLMs to generate the correct answer based on the corrected query or the original query without misleading information. Both the Correct Query and Correct Answer can be directly derived from the content of the original dataset.
+
+72B-Instruct model to modify each question in the original dataset to contain misleading information, thereby constructing the misleading query dataset  $CQA_{mis}$ . For each dataset, we designed corresponding prompts to ensure that the newly generated misleading query datasets introduced misleading information while preserving the naturalness and complexity of the original queries. The prompt templates and data examples used are provided in Appendix B. At the same time, to ensure that the generated misleading query datasets are of higher quality, we adopted a batch generation strategy, generating three misleading queries per batch, and filtered the generated data based on the following two criteria:
+
+Sentence Similarity The similarity between the generated misleading query and the original query is measured using edit distance, which is defined as the minimum number of operations (including insertion, deletion, and substitution) required to transform one string into another (Niu et al., 2018).
+
+Let the original query be  $q_{\mathrm{ori}}$  and the generated misleading query be  $q_{\mathrm{mis}}$ . The sentence similarity  $S_{\mathrm{sim}}$  is defined as:
+
+$$
+S _ {\text {s i m}} = 1 - \frac {\operatorname {E d i t D i s t a n c e} \left(q _ {\text {o r i}} , q _ {\text {m i s}}\right)}{\max \left(\operatorname {L e n g t h} \left(q _ {\text {o r i}}\right) , \operatorname {L e n g t h} \left(q _ {\text {m i s}}\right)\right)}, \tag {1}
+$$
+
+where  $\text{EditDistance}(q_{\text{ori}}, q_{\text{mis}})$  represents the edit distance between  $q_{\text{ori}}$  and  $q_{\text{mis}}$ , and  $\text{Length}$  denotes the length of the sentence. We retain misleading queries with surface similarity  $S_{\text{sim}}$  higher than the threshold  $r_{\text{sim}}$  to ensure that the generated queries are close to the original queries.
+
+Answer Error Rate The answer error rate measures whether the model generates incorrect answers when responding to generated, misleading queries. We input the query into the model  $N$  times, record the number of incorrect answers  $N_{\mathrm{error}}$ , and define the error rate as:
+
+$$
+E _ {\text {e r r o r}} = \frac {N _ {\text {e r r o r}}}{N}, \tag {2}
+$$
+
+where  $E_{\text{error}}$  denotes the error rate. We only retain the misleading queries with an answer error rate  $E_{\text{error}}$  above the threshold  $r_{\text{error}}$  to ensure that these queries have a high level of misleading content.
+
+Data Filtering Strategy Specifically, for each query, we randomly generate three misleading queries that satisfy  $S_{\mathrm{sim}} > 0.8$  and  $E_{\mathrm{error}} > 0.5$ , and then retain the one with the highest value of  $S_{\mathrm{sim}} + E_{\mathrm{error}}$ . An example of randomly generated misleading queries is presented in Appendix C. This ensures that the generated misleading queries are both highly similar to the original queries and contain sufficiently misleading information, thereby enabling more effective training and evaluation of LLMs' performance when faced with highly deceptive and subtle misleading queries.
+
+# 2.2 Query Detection Training
+
+Since misleading information in the query can significantly affect the quality of LLMs' responses, the goal of this stage is to train the model to identify whether misleading information exists in the given query.
+
+Input A query. (If the dataset contains background knowledge, the background knowledge is also input to the model, such as HaluEval dataset.)
+
+Output Objective YES (indicating that the query contains misleading information) or NO (indicating that the query contains no misleading information).
+
+Training Objective The training objective is to minimize the cross-entropy loss function:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\text {c h e c k}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left(y _ {\text {t r u e}, i} \log \left(y _ {\text {p r e d}, i}\right) \right. \tag {3} \\ + \left. \left(1 - y _ {\text {t r u e}, i}\right) \log \left(1 - y _ {\text {p r e d}, i}\right)\right) \\ \end{array}
+$$
+
+where  $N$  is the number of samples in each batch,  $y_{\mathrm{true},i}$  is the true label of sample  $i$ , with a value of 1 (indicating "YES") or 0 (indicating "NO"), and  $y_{\mathrm{pred},i}$  is the model's predicted probability that the  $i$ -th sample is labeled as "YES".
+
+# 2.3 Query Correction Training
+
+Queries labeled as containing misleading information ("YES") in the query detection stage will proceed to the query correction stage. The goal of this stage is to correct the queries by leveraging the knowledge provided by the dataset or the model's
+
+internal knowledge, transforming them into versions that do not contain misleading information, thereby improving the query quality.
+
+Input A misleading query. (If the dataset contains background knowledge, the background knowledge is also input to the model, such as HaluEval dataset.)
+
+# Output Objective Corrected Query.
+
+Training Objective The loss function in the query correction stage uses semantic similarity rather than edit distance, as misleading information primarily affects semantic similarity rather than edit distance (Wu et al., 2022). This approach is used to train the model to generate high-quality corrected queries that are semantically consistent with the original correct queries. In this study, we use Sentence-BERT to obtain semantic vector representations (Reimers, 2019). Furthermore, we ensure that corrected queries are semantically and logically consistent with correct queries by minimizing the distance between corrected and correct queries in the semantic space. Mathematically, the loss function is formulated as:
+
+$$
+\mathcal {L} _ {\text {c o r r e c t i o n}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left(1 - \frac {\mathbf {u} _ {i} \cdot \mathbf {v} _ {i}}{\| \mathbf {u} _ {i} \| \| \mathbf {v} _ {i} \|}\right), \tag {4}
+$$
+
+where  $N$  is the number of training samples in each batch, and  $\mathbf{u}_i$  and  $\mathbf{v}_i$  are the semantic vector representations of the  $i$ -th correct query and corrected query, respectively.
+
+# 2.4 Query Answering Training
+
+The goal of this stage is to generate accurate and reliable answers based on the corrected query or the original query.
+
+Input The corrected query or the original query.
+
+Output Objective Correct Answers.
+
+Training Objective The model-generated answers should be aligned with the ground truth answers. The training objective is to minimize the cross-entropy loss function. Given the input sequence of the  $i$ -th data sample as  $X^{(i)} = x_{1}^{(i)},x_{2}^{(i)},\ldots ,x_{T_{i}}^{(i)}$ , the loss function is defined as:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\text {g e n e r a t i o n}} = \\ - \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\frac {1}{T _ {i}} \sum_ {t = 1} ^ {T _ {i}} \log P \left(y _ {t} ^ {(i)} \mid x _ {1} ^ {(i)}, x _ {2} ^ {(i)}, \dots , x _ {t - 1} ^ {(i)}\right)\right), \tag {5} \\ \end{array}
+$$
+
+where  $N$  is the number of training samples in each batch,  $P(y_{t}^{(i)}|x_{1}^{(i)},x_{2}^{(i)},\ldots ,x_{t - 1}^{(i)})$  denotes the probability that the model predicts the ground truth label  $y_{t}^{(i)}$  at position  $t$  in the  $i$ -th data sample and  $T_{i}$  represents the sequence length of the  $i$ -th data sample.
+
+# 2.5 Unified Framework and Training
+
+To ensure consistency across the query detection, correction, and answer generation processes, we carefully train the model for the next stage based on the model trained in the previous stage, ultimately forming a seamless pipeline. Through this three-stage training strategy, the model effectively learns to detect and correct misleading information in queries and generate accurate responses. This approach significantly mitigates the impact and potential harm of misleading information on the model's performance.
+
+# 3 Experiments
+
+# 3.1 Setup
+
+We conducted comprehensive experiments on the hallucination detection task (HaluEval), question answering tasks (CQA and TruthfulQA), and the misleading query datasets we constructed (HaluEval-QAmis and CQAmis). In all experiments, we report the mean results and standard deviations over three random runs. The experimental configuration and dataset usage details are in Appendix A.
+
+Original Datasets We conduct the construction of the misleading datasets and the evaluation of the model on three datasets:
+
+- The HaluEval dataset (Li et al., 2023) includes the QA, Summarization, and Dialogue subsets, which are used to test the model's ability to identify hallucinations and misleading information.  
+- The CQA dataset (Talmor et al., 2019) is used to evaluate the model's accuracy in answering common sense questions.  
+- The TruthfulQA dataset (Lin et al., 2022) is used to assess the truthfulness and hallucination in the generated answers.
+
+Construction of Training Datasets Our training dataset consists of 4,000 samples from the  $QA$  subset of HaluEval and the corresponding 4,000 samples from the HaluEval- $QA_{mis}$  dataset, which contains misleading information. This dataset is
+
+designed to provide diverse training samples that include both genuine and misleading queries. The specific construction process is in Appendix B.
+
+# 3.2 Main Results
+
+We conducted experiments on models from the Qwen (Yang et al., 2024) and Llama (Dubey et al., 2024) families. The experimental results for the hallucination detection task are listed in Table 1, and the experimental results for the question answering task are listed in Table 2.
+
+The Impact of Misleading Information in the Input on the Response For all the original models (those not trained with our framework), the accuracy on the HaluEval-  $QA_{mis}$  and  $CQA_{mis}$  datasets is significantly lower than that on the QA and CQA datasets. This demonstrates that the presence of misleading information in the input can significantly affect the accuracy and performance of the model's responses. After SFT, the model shows some improvement in performance on both the original dataset and the misleading query datasets. However, the accuracy on the misleading query datasets HaluEval-  $QA_{mis}$  and  $CQA_{mis}$  remains significantly lower than that on the original QA and CQA datasets. However, after training with our method, the model's performance on the misleading query datasets shows a significant reduction in the gap compared to the QA and CQA datasets. This indicates that our proposed method effectively mitigates the performance degradation caused by misleading information in the input.
+
+Performance on the Hallucination Detection Task We tested the model's performance on the hallucination detection task using the QA, Summarization, and Dialogue subsets of the HaluEval dataset, as well as the HaluEval-QAmis dataset. The experimental results show that the model trained by our method significantly outperforms the original model as well as the model after SFT on all datasets. This demonstrates that our method enhances the model's ability to detect misleading information.
+
+Performance on the Question Answering Task We evaluated the model on the commonsense question answering task using the CQA dataset and the  $CQAmis$  dataset. We found that the model trained with our method showed a significant improvement in accuracy on both datasets, demonstrating that our method can enhance the model's accuracy in question answering tasks. We tested the model's
+
+<table><tr><td rowspan="2">Instruction Model</td><td rowspan="2">Method</td><td colspan="4">HaluEval</td></tr><tr><td>Sum</td><td>Dia</td><td>QA</td><td>\( HaluEval-QA_{mis} \)</td></tr><tr><td rowspan="3">Qwen2-7B</td><td>Original</td><td>51.5 ± 0.8</td><td>70.3 ± 0.4</td><td>54.5 ± 1.3</td><td>51.1 ± 0.6</td></tr><tr><td>SFT</td><td>54.2 ± 1.8</td><td>71.9 ± 0.5</td><td>62.3 ± 1.1</td><td>58.6 ± 1.3</td></tr><tr><td>Ours</td><td>68.1 ± 1.7</td><td>72.1 ± 1.3</td><td>68.9 ± 1.6</td><td>68.0 ± 1.0</td></tr><tr><td rowspan="3">Qwen2.5-14B</td><td>Original</td><td>73.2 ± 0.6</td><td>70.5 ± 0.9</td><td>62.7 ± 1.7</td><td>54.3 ± 0.8</td></tr><tr><td>SFT</td><td>73.8 ± 0.5</td><td>71.3 ± 0.6</td><td>68.2 ± 1.3</td><td>60.7 ± 1.0</td></tr><tr><td>Ours</td><td>76.3 ± 0.8</td><td>72.2 ± 0.4</td><td>73.2 ± 1.5</td><td>71.3 ± 1.6</td></tr><tr><td rowspan="3">Llama3.2-3B</td><td>Original</td><td>27.2 ± 2.3</td><td>49.9 ± 0.5</td><td>49.5 ± 1.2</td><td>46.2 ± 0.7</td></tr><tr><td>SFT</td><td>35.0 ± 1.6</td><td>53.2 ± 1.3</td><td>51.9 ± 1.0</td><td>48.1 ± 0.4</td></tr><tr><td>Ours</td><td>57.4 ± 2.3</td><td>60.8 ± 2.1</td><td>53.9 ± 0.8</td><td>52.5 ± 0.9</td></tr><tr><td rowspan="3">Llama3-8B</td><td>Original</td><td>50.0 ± 1.8</td><td>57.3 ± 1.4</td><td>51.8 ± 0.9</td><td>49.6 ± 0.6</td></tr><tr><td>SFT</td><td>53.8 ± 0.8</td><td>59.1 ± 1.3</td><td>57.3 ± 1.3</td><td>53.2 ± 1.0</td></tr><tr><td>Ours</td><td>62.7 ± 1.5</td><td>68.6 ± 0.6</td><td>67.7 ± 1.8</td><td>66.9 ± 1.4</td></tr></table>
+
+Table 1: Performance (mean ± standard deviation, %) comparison on HaluEval dataset. In this case, Original refers to the model without any additional processing, while supervised fine-tuning (SFT) refers to the model that undergoes direct fine-tuning.  
+
+<table><tr><td rowspan="2">Instruction Model</td><td rowspan="2">Method</td><td>CQA</td><td>CQAmis</td><td colspan="3">TruthfulQA</td></tr><tr><td>ACC</td><td>ACC</td><td>BLEURT</td><td>BLEU</td><td>ROUGE</td></tr><tr><td rowspan="3">Qwen2-7B</td><td>Original</td><td>70.4 ± 0.8</td><td>68.8 ± 0.8</td><td>0.636 ± 0.013</td><td>0.543 ± 0.006</td><td>0.570 ± 0.004</td></tr><tr><td>SFT</td><td>72.1 ± 0.7</td><td>69.9 ± 1.1</td><td>0.633 ± 0.011</td><td>0.562 ± 0.002</td><td>0.577 ± 0.007</td></tr><tr><td>Ours</td><td>78.8 ± 1.6</td><td>74.9 ± 0.5</td><td>0.678 ± 0.006</td><td>0.551 ± 0.003</td><td>0.589 ± 0.006</td></tr><tr><td rowspan="3">Qwen2.5-14B</td><td>Original</td><td>80.9 ± 0.8</td><td>71.6 ± 0.9</td><td>0.709 ± 0.019</td><td>0.582 ± 0.004</td><td>0.609 ± 0.009</td></tr><tr><td>SFT</td><td>81.4 ± 0.5</td><td>73.3 ± 1.6</td><td>0.721 ± 0.005</td><td>0.577 ± 0.008</td><td>0.613 ± 0.007</td></tr><tr><td>Ours</td><td>83.1 ± 0.7</td><td>78.6 ± 1.3</td><td>0.748 ± 0.008</td><td>0.615 ± 0.011</td><td>0.642 ± 0.013</td></tr><tr><td rowspan="3">Llama3.2-3B</td><td>Original</td><td>65.1 ± 0.9</td><td>62.1 ± 0.6</td><td>0.613 ± 0.006</td><td>0.517 ± 0.005</td><td>0.527 ± 0.005</td></tr><tr><td>SFT</td><td>66.2 ± 1.0</td><td>63.7 ± 0.7</td><td>0.617 ± 0.005</td><td>0.523 ± 0.004</td><td>0.521 ± 0.008</td></tr><tr><td>Ours</td><td>68.5 ± 0.4</td><td>67.5 ± 0.5</td><td>0.632 ± 0.008</td><td>0.534 ± 0.007</td><td>0.556 ± 0.007</td></tr><tr><td rowspan="3">Llama3-8B</td><td>Original</td><td>70.8 ± 1.1</td><td>64.9 ± 0.5</td><td>0.625 ± 0.011</td><td>0.489 ± 0.012</td><td>0.553 ± 0.004</td></tr><tr><td>SFT</td><td>69.3 ± 0.5</td><td>64.3 ± 0.4</td><td>0.607 ± 0.013</td><td>0.503 ± 0.005</td><td>0.562 ± 0.008</td></tr><tr><td>Ours</td><td>76.4 ± 1.3</td><td>70.0 ± 0.8</td><td>0.668 ± 0.009</td><td>0.528 ± 0.007</td><td>0.586 ± 0.006</td></tr></table>
+
+Table 2: Performance (mean ± standard deviation, %) comparison on CQA, CQAmis and TruthfulQA dataset.
+
+ability to generate truthful responses and avoid producing false information using the TruthfulQA dataset. The experimental results show that the model trained with our method achieves improvements in BLEURT(Sellam et al., 2020), BLEU (Papineni et al., 2002), and ROUGE (Lin, 2004) metrics, demonstrating that our method can effectively reduce hallucinations in responses and improve the factual accuracy of the model's answers.
+
+Scalability Analysis We selected the 7B and 14B models from the Qwen series, as well as the 3B and 8B models from the Llama series, for experimentation. The experimental results show that, regardless of the model family or size, the performance of the models trained with our method is significantly improved. This indicates that our method is applicable to models of different scales and demonstrates good scalability. Furthermore,
+
+on the QA, HaluEval-QAmis, and Dialogue datasets, the Qwen-7B model trained with our framework even outperforms the Qwen-14B model that was not trained with our method, further validating the effectiveness and scalability of our approach.
+
+Cross-task Generalisability Analysis During the model training process, our training data only consisted of 4,000 samples from the QA dataset and 4,000 samples from the HaluEval-QAmis dataset. However, the model still shows significant performance improvements on other datasets. In comparison, we performed SFT on the model using the same training data. However, the model fine-tuned with standard methods showed significant performance improvement only on the QA and HaluEval-QAmis datasets, with little to no improvement on other datasets and even a decline in performance. This demonstrates that our method has strong cross
+
+task generalization capabilities.
+
+# 3.3 Instruction Following Capability Study
+
+We evaluated the model's instruction-following capability on the three subtasks of the hallucination detection dataset HaluEval and the HaluEval $QA_{mis}$  dataset. In these evaluations, if the model's response does not include either "YES" or "NO", or contains both, it is marked as failed. Table 3 shows the proportion of failed responses for each model in this evaluation task. The models trained with our method exhibit a significant reduction in the failed proportion across all datasets, indicating that our method significantly improves the model's instruction-following capability.
+
+<table><tr><td rowspan="2">Instruction Model</td><td rowspan="2">Method</td><td colspan="4">HaluEval</td></tr><tr><td>Sum</td><td>Dia</td><td>QA</td><td>\( HaluEval-QA_{mis} \)</td></tr><tr><td rowspan="3">Llama3.2-3B</td><td>Original</td><td>13.67</td><td>1.96</td><td>3.54</td><td>5.67</td></tr><tr><td>SFT</td><td>8.73</td><td>2.17</td><td>2.36</td><td>2.86</td></tr><tr><td>Ours</td><td>3.55</td><td>1.24</td><td>1.09</td><td>2.09</td></tr><tr><td rowspan="3">Llama3-8B</td><td>Original</td><td>2.61</td><td>0.63</td><td>1.03</td><td>1.75</td></tr><tr><td>SFT</td><td>2.17</td><td>0.58</td><td>0.95</td><td>1.42</td></tr><tr><td>Ours</td><td>0.36</td><td>0.43</td><td>0.69</td><td>0.78</td></tr></table>
+
+Table 3: Performance comparison of failure rate (%) on HaluEval dataset. A smaller failure rate (%) indicates a stronger ability of LLMs to follow instructions.
+
+# 3.4 Ablation Study
+
+We used the 3B and 8B versions of the Llama models to test the changes in model performance after each stage of training. The results are shown in Table 4.
+
+The Role of the Query Detection Stage As can be seen from the results in Table 4, the models after query detection training show significant performance improvement on all datasets. Moreover, the larger the model size, the more significant the performance improvement. This verifies that the query detection stage effectively enhances the model's ability to match query details with existing knowledge, thereby enabling it to identify potential misleading information more accurately. Therefore, after fine-tuning with the query detection stage, the model's ability to identify misleading information is significantly enhanced, providing a solid foundation for subsequent hallucination correction and answer generation.
+
+The Role of the Query Correction Stage As can be seen from the results in Table 4, the models,
+
+after further query correction training, continue to exhibit performance improvements across all datasets. In Appendix D, we present a case study that illustrates that after this step of training, the model has learned to correct queries containing misleading information.
+
+The Role of the Answer Generation Stage As can be seen from the results in Table 4, the models after further query answering training show significant improvement both on the hallucination detection task and the question answering task. The main role of the answer generation training is to help the model adapt to the corrected queries, enabling it to better understand the modified query and generate more accurate answers that align with the user's true intention. Therefore, this stage not only greatly improves the model's performance but also further consolidates the improvements achieved through the previous two fine-tuning stages, making the model more accurate when handling complex tasks.
+
+# 3.5 Case Study
+
+We provide case studies as examples in Table 18 and Appendix D, which visually demonstrate that the model trained with our method can correctly identify and correct the misleading information in the question, compared to the original model, thereby generating the correct answer. This significantly improves the accuracy of the responses and reduces hallucinations.
+
+In the example from the HaluEval-QA_mis dataset (Table 18), when faced with a query containing misleading information, the original model responds directly based on the misleading question, resulting in an incorrect answer "FA Premier League". Although this competition is indeed a British football event, it does not align with the competition that the question actually intends to ask about, which is the one attended by Malcolm Smith. This indicates that misleading information can mislead the model into providing an incorrect response. In contrast, the model trained with our method keenly detects the misleading information in the question, corrects the misleading information, and then generates a response based on the corrected question, yielding the correct result.
+
+# 3.6 Do the questions in standard datasets contain misleading information?
+
+In order to investigate whether some potentially misleading information also exists in commonly
+
+<table><tr><td rowspan="2">Instruction Model</td><td rowspan="2">Method</td><td colspan="4">HaluEval</td><td colspan="2">CQA</td></tr><tr><td>Sum</td><td>Dia</td><td>QA</td><td>HaluEval-QAmis</td><td>CQA</td><td>CQAmis</td></tr><tr><td rowspan="4">Llama-3.2-3B</td><td>Original</td><td>27.2 ± 2.3</td><td>49.9 ± 0.5</td><td>49.5 ± 1.2</td><td>46.2 ± 0.7</td><td>65.1 ± 0.9</td><td>62.1 ± 0.6</td></tr><tr><td>Stage 1</td><td>52.8 ± 1.0</td><td>56.4 ± 1.5</td><td>50.1 ± 1.0</td><td>48.4 ± 0.6</td><td>67.6 ± 0.6</td><td>66.1 ± 0.9</td></tr><tr><td>Stage 2</td><td>53.2 ± 0.7</td><td>57.2 ± 1.3</td><td>50.5 ± 0.7</td><td>48.7 ± 0.4</td><td>68.1 ± 0.8</td><td>66.5 ± 0.4</td></tr><tr><td>Stage 3</td><td>57.4 ± 2.3</td><td>60.8 ± 2.1</td><td>53.9 ± 0.8</td><td>52.5 ± 0.9</td><td>68.5 ± 0.4</td><td>67.5 ± 0.5</td></tr><tr><td rowspan="4">Llama3-8B</td><td>Original</td><td>50.0 ± 1.8</td><td>57.3 ± 1.4</td><td>51.8 ± 0.9</td><td>49.6 ± 0.6</td><td>70.8 ± 1.1</td><td>64.9 ± 0.5</td></tr><tr><td>Stage 1</td><td>60.1 ± 0.6</td><td>65.3 ± 1.2</td><td>63.1 ± 1.2</td><td>62.1 ± 1.2</td><td>73.5 ± 1.2</td><td>67.4 ± 1.1</td></tr><tr><td>Stage 2</td><td>60.5 ± 0.8</td><td>65.7 ± 0.8</td><td>63.6 ± 0.9</td><td>63.0 ± 0.7</td><td>74.8 ± 1.9</td><td>68.9 ± 0.7</td></tr><tr><td>Stage 3</td><td>62.7 ± 1.5</td><td>68.6 ± 0.6</td><td>67.7 ± 1.8</td><td>66.9 ± 1.4</td><td>76.4 ± 1.3</td><td>70.0 ± 0.8</td></tr></table>
+
+Table 4: Ablation study on HaluEval and CQA dataset. Original refers to the original model, Stage 1 refers to the model after query detection training, Stage 2 refers to the model after further query correction training, and Stage 3 refers to the model after further query answering training again.  
+
+<table><tr><td></td><td>HaluEval-QAmis Example Ground Truth Label: Super Bowl XLVIII</td></tr><tr><td>Query</td><td>&quot;knowledge&quot;: &quot; Smith was named the Most Valuable Player of Super Bowl XLVIII after they defeated the Denver Broncos. Super Bowl XLVIII was an American football game between the American Football Conference (AFC) champion Denver Broncos and National Football Conference (NFC) champion Seattle Seahawks to decide the National Football League (NFL) champion for the 2013 season.&quot; 
+&quot;question&quot;: &quot;In which England football game was Malcolm Smith named Most Valuable player?&quot;</td></tr><tr><td>Model Output (Original model)</td><td>FA Premier League</td></tr><tr><td>Model Output (Our Method)</td><td>Malcolm Smith named Most Valuable Player is American football game, not England football game, this American football game is Super Bowl XLVIII.</td></tr></table>
+
+used standard datasets (original HaluEval, CQA, and TruthfulQA), we employed the Qwen2-7B model, which had been trained in the first stage, to examine the questions from these standard datasets. Our findings revealed that some questions in the standard datasets contain misleading information. The proportions of questions containing misleading information in each dataset are summarized in Table 6.
+
+Subsequently, we removed the questions identified by the model as containing misleading information from the original datasets to form new datasets (HaluEval-Sum<sub>sub</sub>, HaluEval-Dia<sub>sub</sub>, HaluEval-QA<sub>sub</sub>, CQA<sub>sub</sub>). We then conducted evaluations on these new datasets. The experimental results are shown in Table 7.
+
+We observed that the accuracy of the original Qwen2-7B model's responses on the datasets from which questions containing misleading information had been removed showed a significant improvement compared to the original datasets. This indicates that the misleading information present in the original datasets can interfere with the model's
+
+ability to generate accurate responses. However, the Qwen2-7B model trained with our method did not show any notable performance changes, demonstrating its robustness against questions that contain misleading information and its reduced susceptibility to such interference.
+
+Table 5: Case study on the HaluEval-  $QA_{mis}$  dataset. The original model is misled by the misleading information "England football" in the query, resulting in an incorrect answer (highlighted in red). However, the model trained with our method correctly identifies that it should be "American football" and generates the correct answer (highlighted in blue).  
+
+<table><tr><td>Dataset</td><td>Percentage</td></tr><tr><td>HaluEval-Sum</td><td>3.91%</td></tr><tr><td>HaluEval-Dia</td><td>14.62%</td></tr><tr><td>HaluEval-QA</td><td>13.53%</td></tr><tr><td>CQA</td><td>10.40%</td></tr><tr><td>TruthfulQA</td><td>6.30%</td></tr></table>
+
+Table 6: The proportion of questions containing misleading information in the three standard datasets.
+
+# 4 Related Work
+
+The Impact of Misleading Information on LLMs Studies have shown that in question answering systems, when the input contains false or ambiguous
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">HaluEval-Sum</td><td colspan="2">HaluEval-Dia</td><td colspan="2">HaluEval-QA</td><td colspan="2">CQA</td></tr><tr><td>Original</td><td>Sub</td><td>Original</td><td>Sub</td><td>Original</td><td>Sub</td><td>Original</td><td>Sub</td></tr><tr><td>Qwen2-7B</td><td>51.5±0.8</td><td>53.6±0.7</td><td>70.3±0.4</td><td>71.1±0.5</td><td>54.5±1.3</td><td>57.2±1.1</td><td>70.4±0.8</td><td>73.1±0.9</td></tr><tr><td>Ours</td><td>68.1±1.7</td><td>68.5±1.4</td><td>72.1±1.3</td><td>72.3±0.8</td><td>68.9±1.6</td><td>69.7±0.7</td><td>78.8±1.6</td><td>79.1±1.1</td></tr></table>
+
+Table 7: Performance (mean ± standard deviation, %) comparison before and after addressing misleading information in the dataset. "Original" denotes the original dataset, and "Sub" refers to the dataset subset with misleading information questions removed.
+
+information, LLMs often generate inaccurate responses based on these erroneous inputs (Pan et al., 2023; Zhang et al., 2024b; Kuhn et al., 2022; Chen and Shu, 2024). To address this issue, a common approach is to detect potential errors or inconsistencies based on the syntactic and semantic analysis of the input. For example, Li and Wang (2024) proposed a detection-correction integrated framework based on a general language model. This approach integrates the detection and correction processes into a single model, enabling the automatic detection and correction of syntactic errors in the input text. To ensure the accuracy of the input information, some studies (Zhang et al., 2024a; Butala et al., 2024; Xiong et al., 2024) have attempted to validate the correctness of the input through multi-turn dialogues, ensuring that the content generated by the model aligns with reality. Although existing methods have improved the model's ability to handle misleading information to some extent, the complexity and diversity of misleading information still pose significant challenges in this area.
+
+Model Output Correction In recent years, methods aimed at correcting the content generated by LLMs to improve generation accuracy have been widely studied. Wu et al. (2024b) proposed an iterative verification-correction framework (ProCo), which identifies and corrects potential errors through subsequent validation of the generated content. Petroni et al. (2021) leveraged external knowledge bases to correct the content generated by the model. Liu et al. (2024a) explored the intrinsic self-correction capabilities of LLMs from both theoretical and empirical perspectives, proposing that zero temperature and fair prompting are key factors for successful self-correction. Liu et al. (2024b) proposed distilling the LLMs' self-evaluation capability and more comprehensive thinking into smaller models, effectively improving the performance and answer accuracy of LLMs in resource-constrained environments. In addition, knowledge-enhanced generation has also been used
+
+to reduce the likelihood of language models generating misleading content (Jiang et al., 2024).
+
+# 5 Conclusion
+
+In this paper, we propose a novel three-stage fine-tuning method to mitigate the impact of misleading information in the input queries on LLMs. Our method enhances the ability of LLMs to detect, correct, and respond accurately to queries containing misleading content, significantly improving response accuracy and reducing hallucinations. Extensive experiments across various datasets demonstrate the effectiveness of this approach, making LLMs more robust and trustworthy.
+
+# 6 Limitations
+
+Although our method has achieved significant improvement in mitigating the impact of misleading information in the input on LLMs, there are still some limitations that can be addressed in future work:
+
+1. Dependency on high-quality knowledge: Our method relies on the model's internal knowledge or requires reliable external knowledge during the query detection and Query Correction Stages. If both the model's internal knowledge and the external knowledge base are incomplete or inaccurate, the model may still generate incorrect corrections or responses.  
+2. Scalability to larger models: Although our method performs well across different model sizes (e.g., Qwen-7B vs. Qwen-14B, Llama 3B vs. 8B models), future work can explore experiments on even larger models.  
+3. Expansion of the training dataset: Future work could consider expanding the training dataset to more effectively improve model performance and facilitate transfer to additional tasks or domains.
+
+# 7 Ethical Statement
+
+The Use of AI Assistants We employed ChatGPT to assist us in polishing our paper and coding.
+
+# References
+
+Francois Barnard, Marlize Van Sittert, and Sirisha Rambhatla. 2023. Self-diagnosis and large language models: A new front for medical misinformation. arXiv preprint arXiv:2307.04910.  
+Yash Butala, Siddhant Garg, Pratyay Banerjee, and Anita Misra. 2024. Promise: A proactive multi-turn dialogue dataset for information-seeking intent resolution. In Findings of the Association for Computational Linguistics: EACL 2024, pages 1774-1789.  
+Canyu Chen and Kai Shu. 2024. Can LLM-generated misinformation be detected? In The Twelfth International Conference on Learning Representations.  
+Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997.  
+Bohan Jiang, Zhen Tan, Ayushi Nirmal, and Huan Liu. 2024. Disinformation detection: An evolving challenge in the age of llms. In Proceedings of the 2024 SIAM International Conference on Data Mining (SDM), pages 427-435. SIAM.  
+Chao Jin, Zili Zhang, Xuanlin Jiang, Fangyue Liu, Xin Liu, Xuanzhe Liu, and Xin Jin. 2024. Ragcache: Efficient knowledge caching for retrieval-augmented generation. arXiv preprint arXiv:2404.12457.  
+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. arXiv preprint arXiv:2406.01297.  
+Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2022. Clam: Selective clarification for ambiguous questions with generative language models. arXiv preprint arXiv:2212.07769.  
+Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474.  
+Dongyang Li, Junbing Yan, Taolin Zhang, Chengyu Wang, Xiaofeng He, Longtao Huang, Hui Xue, and
+
+Jun Huang. 2024a. On the role of long-tail knowledge in retrieval augmented large language models. In ACL (Short Papers).  
+Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023. Halueval: A large-scale hallucination evaluation benchmark for large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6449-6464.  
+Wei Li and Houfeng Wang. 2024. Detection-correction structure via general language model for grammatical error correction. arXiv preprint arXiv:2405.17804.  
+Zhen Li, Xiaohan Xu, Tao Shen, Can Xu, Jia-Chen Gu, Yuxuan Lai, Chongyang Tao, and Shuai Ma. 2024b. Leveraging large language models for nlg evaluation: Advances and challenges. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16028-16045.  
+Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.  
+Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252.  
+Dancheng Liu, Amir Nassereldine, Ziming Yang, Chenhui Xu, Yuting Hu, Jiajie Li, Utkarsh Kumar, Changjae Lee, and Jinjun Xiong. 2024a. Large language models have intrinsic self-correction ability. CoRR.  
+Weize Liu, Guocong Li, Kai Zhang, Bang Du, Qiyuan Chen, Xuming Hu, Hongxia Xu, Jintai Chen, and Jian Wu. 2024b. Mind's mirror: Distilling self-evaluation capability and comprehensive thinking from large language models. 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 6748-6763.  
+Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2024. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36.  
+Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial nli: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901.  
+Yilin Niu, Chao Qiao, Hang Li, and Minlie Huang. 2018. Word embedding based edit distance. arXiv preprint arXiv:1810.10752.
+
+Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Yang Wang. 2023. On the risk of misinformation pollution with large language models. In The 2023 Conference on Empirical Methods in Natural Language Processing.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.  
+F Petroni, A Piktus, A Fan, PSH Lewis, M Yazdani, ND Cao, J Thorne, Y Jernite, V Karpukhin, J Maillard, et al. 2021. Kilt: a benchmark for knowledge intensive language tasks. In NAACL-HLT, pages 2523-2544. Association for Computational Linguistics.  
+N Reimers. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084.  
+Thibault Sellam, Dipanjan Das, and Ankur P Parikh. 2020. Bleurt: Learning robust metrics for text generation. arXiv preprint arXiv:2004.04696.  
+Jongyoon Song, Sangwon Yu, and Sungroh Yoon. 2024. Large language models are skeptics: False negative problem of input-conflicting hallucination. arXiv preprint arXiv:2406.13929.  
+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, Volume 1 (Long and Short Papers), pages 4149-4158.  
+Shan Wu, Chunlei Xin, Bo Chen, Xianpei Han, and Le Sun. 2022. Semantic-aware contrastive learning for more accurate semantic parsing. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4040-4052.  
+Siye Wu, Jian Xie, Jiangjie Chen, Tinghui Zhu, Kai Zhang, and Yanghua Xiao. 2024a. How easily do irrelevant inputs skew the responses of large language models? arXiv preprint arXiv:2404.03302.  
+Zhenyu Wu, Qingkai Zeng, Zhihan Zhang, Zhaoxuan Tan, Chao Shen, and Meng Jiang. 2024b. Large language models can self-correct with key condition verification. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 12846-12867.  
+Guanming Xiong, Junwei Bao, and Wen Zhao. 2024. Interactive-kbqa: Multi-turn interactions for knowledge base question answering with large language models. arXiv preprint arXiv:2402.15131.  
+Ni Xuanfan and Li Piji. 2023. A systematic evaluation of large language models for natural language generation tasks. In Proceedings of the 22nd Chinese
+
+National Conference on Computational Linguistics (Volume 2: Frontier Forum), pages 40-56.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.  
+Tian Yu, Shaolei Zhang, and Yang Feng. 2024. Truth-aware context selection: Mitigating the hallucinations of large language models being misled by untruthful contexts. arXiv preprint arXiv:2403.07556.  
+Michael J. Q. Zhang, W. Bradley Knox, and Eunsol Choi. 2024a. Modeling future conversation turns to teach llms to ask clarifying questions. CoRR, abs/2410.13788.  
+Tong Zhang, Peixin Qin, Yang Deng, Chen Huang, Wenqiang Lei, Junhong Liu, Dingnan Jin, Hongru Liang, and Tat-Seng Chua. 2024b. Clamber: A benchmark of identifying and clarifying ambiguous information needs in large language models. arXiv e-prints, pages arXiv-2405.
+
+# A Experimental details
+
+We employ LoRA for efficient fine-tuning. The detailed setting of hyperparameters is shown in Table 8. Table 9 presents the information about the datasets used for training and testing. Regarding SFT, we use the  $QA$  and HaluEval- $QA_{mis}$  datasets for implementation.
+
+<table><tr><td>Configuration</td><td>Value</td></tr><tr><td>Number of epochs</td><td>4</td></tr><tr><td>Devices</td><td>4 Nvidia H20 GPU (80G)</td></tr><tr><td>Total Batch size</td><td>128</td></tr><tr><td>Learning rate</td><td>5 × 10-4</td></tr><tr><td>Warmup Ratio</td><td>0.1</td></tr></table>
+
+Table 8: Finetuning hyperparameters for experiments.  
+
+<table><tr><td>Datasets</td><td>Train number</td><td>Test number</td></tr><tr><td>HaluEval-QA</td><td>4000</td><td>6000</td></tr><tr><td>HaluEval-QAmis</td><td>4000</td><td>6000</td></tr><tr><td>HaluEval-Sum</td><td>0</td><td>10000</td></tr><tr><td>HaluEval-Dia</td><td>0</td><td>10000</td></tr><tr><td>CQA</td><td>0</td><td>1221</td></tr><tr><td>CQAmis</td><td>0</td><td>1221</td></tr><tr><td>TruthfulQA</td><td>0</td><td>818</td></tr></table>
+
+Table 9: Dataset Split.
+
+# B Dataset details
+
+The specific process of constructing our training dataset is as follows:
+
+- **Query Detection Stage:** We used a mixed dataset comprising  $QA$  and HaluEval- $QA_{mis}$  samples for training. Additionally, we introduced a field with - misleading as a label to indicate whether a query contains misleading information. For queries in the  $QA$  dataset, the with-misleading field is set to "NO", indicating that the query does not contain misleading information. For queries in the HaluEval- $QA_{mis}$  dataset, the with-misleading field is set to "YES", indicating that the query contains misleading information.  
+- Query Correction Stage: Since the task in this stage requires correcting queries that contain misleading information, we selected 4,000 samples from the HaluEval-QAmis dataset for training. At this point, our training objective is to correct the misleading queries
+
+and transform them into the corresponding queries from the original  $QA$  dataset.
+
+- Answer Generation Stage: In this stage, we also use a mixed dataset of  $QA$  and HaluEval $QA_{mis}$  samples to train the model to generate correct answers, where the queries containing misleading information have already been corrected in the query correction stage.
+
+This dataset construction strategy ensures that the model can focus on the corresponding data characteristics at each task stage, while also enhancing the model's ability to handle diverse queries. Next, we present the prompts that we used in the experiments as well as the HaluEval-QAmis and CQAmis datasets. Table 10 and Table 12 show the templates used to generate the HaluEval-QAmis and CQAmis datasets, while Table 11 and Table 13 display some data from the HaluEval-QAmis and CQAmis datasets. Table 14, Table 15, and Table 16 respectively show the templates for query detection, misleading query correction, and answer generation.
+
+# C Select misleading question
+
+This section demonstrates how to select high-quality misleading questions using sentence similarity and answer error rate, with the detailed process shown in Table 17.
+
+# D Case study
+
+Similar to the analysis in Section 3.5, in the example of the  $CQA_{mis}$  dataset (Table 19), the query mentions that after the mechanic made some adjustments to the engine, the car started to experience slight vibrations and emit some unusual sounds. These distracting pieces of information can easily lead the model to mistakenly interpret the sounds as coming from the horn, resulting in an incorrect answer "honk the horn". However, the model trained with our method effectively filters out the interference from the misleading information and provides a response stating that the 'unusual sounds' and 'slight vibrations' are simply distractions that might suggest the car is malfunctioning, but the fact that it's operational and no serious issues are mentioned indicates the car will start running as expected after repairs.
+
+We present a case study of the query correction stage in Table 20, the model trained using our method is able to quickly detect misleading information in the question. By leveraging the
+
+# Prompt for Generating HaluEval-QA<sub>mis</sub> Dataset
+
+You are a language expert. You should create a question that includes subtle mistakes derived from the content in the knowledge.
+
+The modified question should retain the general meaning of the original but introduce small errors or misdirecting details such as incorrect entity names, relationships, times, or terminology. These changes should be subtle and should not drastically alter the original question's structure or core intent.
+
+Do not introduce completely new concepts, terms, or laws that may cause the mistake question to deviate too far from the original. The mistake question should still be recognizable as a version of the original question, with enough misleading elements to potentially cause an incorrect answer if not identified and corrected.
+
+The goal is to create a subtle version of the original question that misleads the model, but does not make it so different than the answer is hard to recover.
+
+Do not provide any content, only the misleading question.
+
+1
+
+knowledge#: {knowledge}
+
+question#: {question}
+
+1
+
+Table 10: The prompt for generating the HaluEval-  $QA_{mis}$  dataset.
+
+knowledge from "the last man executed in Australia", it correctly modifies the misleading information in the query, "the last person executed in New Zealand", and generates the correct question.
+
+# The example of HaluEval-QA<sub>mis</sub> dataset
+
+# Example1:
+
+"knowledge": "The 2003 SEC Championship Game was won by the LSU Tigers 34-13 over the Georgia Bulldogs. The game was played in the Georgia Dome in Atlanta, Georgia. Located in Atlanta, Georgia, between downtown to the east and Vine City to the west, it was owned and operated by the State of Georgia as part of the Georgia World Congress Center Authority." "question": "The 2003 SEC Championship Game was played at a stadium owned and operated by who?"
+
+"misleading question": "The 2003 SEC Championship Game was played at a stadium in the United States owned and operated by who?"
+
+"right answer": "the State of Georgia as part of the Georgia World Congress Center Authority"
+
+"Hallucinated answer": "The film Pandorum was released in 2010." "The 2003 SEC Championship Game was played at the Mercedes-Benz Stadium in Atlanta."
+
+# Example2:
+
+"knowledge": " He was also a successful author, writing biographies of politician Al Smith and athletes Lou Gehrig and John McGraw, as well as histories of the New York Yankees, New York Giants and Brooklyn Dodgers. The team's name derived from the reputed skill of Brooklyn residents at evading the city's trolley streetcar network."
+
+"question": "What team's name derived from the reputed skill of Brooklyn residents at evading the city's trolley streetcar network, who had a history wrote of it by an American sportswriter and biographer who overed sports in New York for the New York Sunfrom 1915 to 1943?"
+
+"misleading question": "What team's name derived from the reputed skill of Manhattan residents at evading the city's trolley streetcar network?"
+
+"right answer": "Brooklyn Dodgers"
+
+"Hallucinated answer": "The New York Yankees were named after the reputed skill of Brooklyn residents at evading the city's trolley streetcar network."
+
+# Example3:
+
+"knowledge": "Sooni Taraporevala (born 1957) is an Indian screenwriter, photographer and filmmaker who is best known as the screenwriter of Mississippi Masala, The Namesake and Oscar-nominated Salaam Bombay(1988), all directed by Mira Nair. Mississippi Masala is a 1991 romantic drama film directed by Mira Nair, based upon a screenplay by Sooni Taraporevala, starring Denzel Washington, Sarita Choudhury, and Roshan Seth."
+
+"question": "Which Oscar-nominated film was written by the screenwriter of a 1990 romantic drama starring Sarita Choudhury?"
+
+"right answer": "Salaam Bombay"
+
+"hallucinated answer": "The Namesake"
+
+Table 11: Randomly sampled HaluEval-  $QA_{mis}$  dataset examples. Blue represents correct information, while red represents misleading information.
+
+# Prompt for generating  $CQA_{mis}$  dataset
+
+You are a language expert. Your task is to take a given question and modify it by introducing distracting or irrelevant details to make the problem more complex and harder to answer. These distractions should not directly lead to any of the answer choices, especially the correct one.
+
+# Steps:
+
+1. Modify the Question: Add irrelevant or distracting details that make the question more challenging. These distractions can be excessive descriptions, background information, or other elements that do not directly relate to the core of the question.  
+2. Provide a Detailed Explanation: After modifying the question, provide a clear explanation that shows why the correct answer is still valid. The explanation should demonstrate that the added distractions don't affect the reasoning behind the correct answer and why it is still the best choice.
+
+Important Notes: The distractions should increase the complexity of the question, but should not directly or indirectly guide the reader to the correct answer. The explanation should clearly outline why the correct answer is the best choice and how the added distractions do not change the fundamental reasoning.
+
+# Example:
+
+1
+
+"question": "A thoroughfare meandered through fields and woods, where was it passing though?",
+
+"choices": ["move about", "city", "country", "town", "new york city"]".
+
+1111
+
+"mistake question": "This road winds through various landscapes, sometimes open farmlands, other times dense forests, and occasionally small villages. Along the way, there are signs of modern construction like small factories and even a few billboards. The surroundings shift frequently, with the occasional sight of tall buildings in the distance. Where is this road most likely passing through?"
+
+"explanation": "The road is passing through country because, despite the modern construction and occasional tall buildings in the distance, the key elements—farmlands, forests, and small villages are characteristic of rural, countryside areas. The mention of "small factories" and "billboards" are distractions that suggest a more urban environment, but they do not outweigh the overall rural characteristics of the area described."
+
+1
+
+question#: {question}
+
+choices#: {choices}
+
+1111
+
+Table 12: The prompt for generating the  $CQA_{mis}$  dataset. In addition to the original fields, we also generated an explain field to provide a detailed description of how the question, after adding misleading information, is step by step deduced to the correct answer.
+
+# The example of  $CQA_{mis}$  dataset
+
+# Example1:
+
+"question": "Billy was reading the newspaper as he commuted to work, but once he got to his destination he balled it up and put it somewhere. Where did it put it?"
+
+"choices": ["trash", "floor", "subway", "ground", "lawn"]
+
+"answer": "trash"
+
+"misleading question": Billy was engrossed in reading the morning newspaper during his daily commute. As he arrived at his workplace, he noticed the crowded train and decided to quickly stuff the newspaper into his pocket to avoid carrying it around. However, as the day progressed, he realized he had forgotten about it and didn't know where he had placed it. Where could it be now?
+
+"explanation": Despite the added details of Billy stuffing the newspaper into his pocket and forgetting about it later, the core question remains about where Billy put the newspaper once he reached his destination. The hallucinations do not provide any information about the newspaper's final location; thus, the correct answer, 'trash', remains valid. Billy might have thrown away the newspaper if he didn't need it anymore or if he found it cluttering his belongings."
+
+# Example2:
+
+"question": "He had to wear a tuxedo while playing the keyboard instrument, so did the other hundred members of the what?"
+
+"choices": ["music store", "band", "medium", "orchestra", "piano store"]
+
+"answer": "orchestra"
+
+"misleading question": "He had to wear a tuxedo while playing the keyboard instrument, and the other hundred members also wore tuxedos. There was a piano beside him. Which group did they belong to?"
+
+"explanation": "Although there was a piano beside him, the fact that they had to wear tuxedos still leads to the correct answer being "orchestra."
+
+Table 13: Randomly sampled  $CQA_{mis}$  dataset examples. Red represents misleading information.
+
+# Prompt for question detection
+
+You are an excellent mistake detector.
+
+I will provide you with a #knowledge# and a #question#. Your task is to carefully compare the #knowledge# and the #question#, and based on the content of the #knowledge#, determine whether the #question# contains any mistake.
+
+If it contains a mistake, output 'YES', otherwise output 'NO'. Please make sure there are no extra outputs.
+
+The following is an example:
+
+knowledge#: "The area covered by South East Queensland varies, depending on the definition of the region, though it tends to include Queensland's three largest cities: the capital city Brisbane; the Gold Coast; and the Sunshine Coast. The Gold Coast is a coastal area in the Australian state of Queensland, approximately 66 km south-southeast of the state capital Brisbane and immediately north of the border with New South Wales."
+
+question#: "South East Queensland (SEQ) is a bio-geographical, political, and administrative region of the state of Queensland in Australia, the area covered by South East Queensland varies, depending on the definition of the region, though it tends to include Queensland's three largest cities, including which coastal area in the Australian state of Victoria, approximately  $66\mathrm{km}$  south-southeast of the state capital Melbourne and immediately north of the border with New South Wales?"
+
+your choice#: "YES"
+
+You need to do your best to identify whether there is any content in the #question# that contradicts the #knowledge#, and if there is, output "YES", otherwise output "NO".
+
+Table 14: The prompt for question detection.  
+
+<table><tr><td>Prompt for misleading question correction</td></tr><tr><td>You are an excellent mistake corrector. I will provide you with a #knowledge# and a #question#. Your task is to carefully compare the #knowledge# and the #question#, and based on the content of the #knowledge# and real-world information, correct any mistakes in the #question# and output the corrected content. The following is an example:</td></tr><tr><td>#knowledge#: &quot;The area covered by South East Queensland varies, depending on the definition of the region, though it tends to include Queensland&#x27;s three largest cities: the capital city Brisbane; the Gold Coast; and the Sunshine Coast. The Gold Coast is a coastal area in the Australian state of Queensland, approximately 66 km south-southeast of the state capital Brisbane and immediately north of the border with New South Wales.&quot;</td></tr><tr><td>#question#: &quot;South East Queensland (SEQ) is a bio-geographical, political, and administrative region of the state of Queensland in Australia, the area covered by South East Queensland varies, depending on the definition of the region, though it tends to include Queensland&#x27;s three largest cities, including which coastal area in the Australian state of Victoria, approximately 66 km south-southeast of the state capital Melbourne and immediately north of the border with New South Wales?&quot;</td></tr><tr><td>#your answer#: &quot;South East Queensland (SEQ) is a bio-geographical, political, and administrative region of the state of Queensland in Australia, the area covered by South East Queensland varies, depending on the definition of the region, though it tends to include Queensland&#x27;s three largest cities, including which coastal area in the Australian state of Queensland, approximately 66 km south-southeast of the state capital Brisbane and immediately north of the border with New South Wales?&quot;</td></tr><tr><td>You need to carefully identify if there is any content in the #question# that contradicts the #knowledge# or real-world knowledge. If there is, please modify the #question# to the correct form and output it. Otherwise, directly output the original #question#.</td></tr></table>
+
+Table 15: The prompt for misleading question correction.  
+
+<table><tr><td>Prompt for generate answer</td></tr><tr><td>You are an expert with broad knowledge. I will provide #knowledge# and #question#. Your task is to answer the #question# based on the #knowledge# and real-world information, and output the answer to the #question#. The following is an example:</td></tr><tr><td>#knowledge#: &quot;He was one of the pioneers of gambling in Las Vegas, where he was a partner with flamboyant mobster Bugsy Siegel at the Flamingo Hotel. Benjamin BugsySiegel (February 28, 1906 – June 20, 1947) was a Jewish American mobster. Siegel was not only influential within the Jewish mob but, like his friend and fellow gangster Meyer Lansky, he also held significant influence within the Italian-American Mafia and the largely Italian-Jewish National Crime Syndicate&quot;.&quot; #question#: &quot;What hotel did the influential National Crime Syndicate member found with David Berman?&quot;. #your answer#: &quot;Flamingo Hotel&quot;</td></tr><tr><td>You need to ensure that the answer you provide contains no content that contradicts the #knowledge# or real-world knowledge.&quot;</td></tr></table>
+
+Table 16: The prompt for generating an answer.
+
+<table><tr><td></td><td>HaluEval-QAmis Example Ground Truth Label:1965</td></tr><tr><td>Knowledge</td><td>DuPont v. Kolon Industries is an intellectual property lawsuit centering on the allegation that Kolon Industries (of ), a South Korea-based company, stole trade secrets concerning the production and marketing of Kevlar from DuPont, an American chemical company. Kevlar is a high strength synthetic fiber used in applications as diverse as bicycle tires and body armor. Developed by Stephanie Kwolek at DuPont in 1965, this high-strength material was first commercially used in the early 1970s as a replacement for steel in racing tires.</td></tr><tr><td>Original Query</td><td>What year was the high strength synthetic fiber, which was the subject of a lawsuit between DuPont and Kolon Industries, developed?</td></tr><tr><td>Misleading Query 1</td><td>What year was the low strength synthetic fiber, which was the subject of a lawsuit between DuPont and Kolon Industries, developed?Similarity Score:0.96respond1: 1867 (✗)respond2: 1965 (✓)respond3: 1433 (✗)Error Rate:0.67</td></tr><tr><td>Misleading Query 2</td><td>What year was the high strength synthetic fiber, which was the subject of a lawsuit between DuPont and Japan Industries, developed?Similarity Score:0.96respond1: 1965 (✓)respond2: 1965 (✓)respond3: 1965 (✓)Error Rate:0</td></tr><tr><td>Misleading Query 3</td><td>What year was the high strength synthetic fiber, which was the subject of a lawsuit between China and England Industries, developed?Similarity Score:0.92respond1: 1965 (✓)respond2: 1965 (✓)respond3: 1563 (✗)Error Rate:0.33</td></tr></table>
+
+Table 17: An example of randomly generated misleading queries and the calculation of  $S_{\mathrm{sim}}$  and  $E_{\mathrm{error}}$ , where the misleading information is highlighted in red. Example 1 satisfies our filtering condition of  $S_{\mathrm{sim}} > 0.8$  and  $E_{\mathrm{error}} > 0.5$ . Correct answers are marked with a correct mark  $\checkmark$ , while incorrect answers are marked with a wrong mark  $\times$ . Among them, misleading query 1 satisfies our filtering criteria of  $S_{\mathrm{sim}} > 0.8$  and  $E_{\mathrm{error}} > 0.5$ .  
+
+<table><tr><td></td><td>HaluEval-QA_mis Example Ground Truth Label: Super Bowl XLVIII</td></tr><tr><td>Query</td><td>&quot;knowledge&quot;: &quot; Smith was named the Most Valuable Player of Super Bowl XLVIII after they defeated the Denver Broncos. Super Bowl XLVIII was an American football game between the American Football Conference (AFC) champion Denver Broncos and National Football Conference (NFC) champion Seattle Seahawks to decide the National Football League (NFL) champion for the 2013 season.&quot; 
+&quot;question&quot;: &quot;In which England football game was Malcolm Smith named Most Valuable player?&quot;</td></tr><tr><td>Model Output (Original model)</td><td>FA Premier League</td></tr><tr><td>Model Output (Our Method)</td><td>Malcolm Smith named Most Valuable Player is American football game, not England football game, this American football game is Super Bowl XLVIII.</td></tr></table>
+
+Table 18: A case study on the HaluEval-  $QA_{mis}$  dataset. The original model is misled by the misleading information "England football" in the question, resulting in an incorrect answer (highlighted in red). However, the model trained with our method correctly identifies that it should be "American football" and generates the correct answer (highlighted in blue).
+
+<table><tr><td></td><td>CQAmis Example Ground Truth Label: start running</td></tr><tr><td>Query</td><td>&quot;question&quot;: &quot;After the mechanic performed several adjustments to the engine, including checking the fuel system and replacing some critical parts, the car started to behave differently. It began to vibrate slightly and made some unusual sounds. Although it seems operational, it&#x27;s not the same as it was before. What is it most likely to do next?&quot; 
+&quot;choices&quot;: [&quot;going too fast&quot;,&quot;last several years&quot;,&quot;honk the horn&quot;,&quot;go fast&quot;,&quot;start running&quot;]</td></tr><tr><td>Model Output (Original model)</td><td>honk the horn</td></tr><tr><td>Model Output (Our Method)</td><td>Even though the car is vibrating and making unusual sounds, these are typical signs of a car being started after maintenance, especially if adjustments were made to the engine. The &quot;unusual sounds&quot; and &quot;slight vibrations&quot; are simply distractions that might suggest the car is malfunctioning, but the fact that it&#x27;s operational and no serious issues are mentioned indicates the car will start running as expected after repairs.</td></tr></table>
+
+Table 19: A case study on the  $CQA_{mis}$  dataset. The original model is misled by the misleading information "slight vibrations" and "unusual sounds" in the question, resulting in an incorrect answer (highlighted in red). However, the model trained with our method correctly identifies "it's operational and no serious issues are mentioned" and generates the correct answer "the car will start running" (highlighted in blue).  
+
+<table><tr><td>Category</td><td>Content</td></tr><tr><td>Knowledge Input</td><td>His most well-known work is the award-winning stage play ‘Remember Ronald Ryan’, a dramatization of the life and subsequent death of Ronald Ryan, the last man executed in Australia. Ronald Joseph Ryan (21 February 1925 – 3 February 1967) was the last person to be legally executed in Australia.</td></tr><tr><td>Original Question</td><td>Barry Dickins wrote a film about the last person executed in New Zealand, named what?</td></tr><tr><td>Output</td><td>The knowledge mentions the last person legally executed in Australia, not New Zealand, so the question should be modified to “Barry Dickins wrote a film about the last person executed in Australia, named what?”</td></tr></table>
+
+Table 20: An example of correcting misleading inputs based on knowledge. For the input data, we used the same prompt as in the query correction training stage (Table 15). The information marked in blue represents correct information, while the information marked in red represents misleading information.

2025/From Misleading Queries to Accurate Answers_ A Three-Stage Fine-Tuning Method for LLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:636bdfa3ab93aadff6b6f04d8a91490b8e52aada44ee5db9a1c5a8b2b36e4b64
+size 1338335

2025/From Observation to Understanding_ Front-Door Adjustments with Uncertainty Calibration for Enhancing Egocentric Reasoning in LVLMs/a56c3926-e95f-40d6-b848-790453a6e4b3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:caaf59a3b47c73ed08af022fee33f2f664c41ba25918f0899af5a8295dc87e78
+size 4338858

2025/From Observation to Understanding_ Front-Door Adjustments with Uncertainty Calibration for Enhancing Egocentric Reasoning in LVLMs/full.md

@@ -0,0 +1,434 @@
+# From Observation to Understanding: Front-Door Adjustments with Uncertainty Calibration for Enhancing Egocentric Reasoning in LVLMs
+
+Shenshen Li $^{1}$ , Wenxin Meng $^{1}$ , Lei Wang $^{3}$ , Hao Yang $^{4}$ , Chong Peng $^{4}$ , Peng Yan $^{4}$ , Fumin Shen $^{1}$ , Jingkuan Song $^{1,2}$ , Heng Tao Shen $^{2}$ , Xing Xu $^{1,2*}$
+
+<sup>1</sup>University of Electronic Science and Technology of China School of Computer Science and Technology, Tongji University, <sup>3</sup>Salesforce AI Research, <sup>4</sup>Meituan
+
+# Abstract
+
+Recent progress in large vision-language models (LVLMs) has shown substantial potential across a broad spectrum of third-person tasks. However, adapting these LVLMs to egocentric scenarios remains challenging due to their third-person training bias. Existing methods that adapt LVLMs for first-person tasks often overlook critical agent-environment interactions, limiting their ability to perform egocentric reasoning. To address these challenges, we propose a novel zero-shot paradigm termed Front-Door Adjustments with Uncertainty Calibration (FRUIT) to enhance the egocentric reasoning abilities of LVLMs by simulating human causal reasoning. Specifically, the FRUIT operates in two stages: observation and understanding. Unlike conventional prompting techniques, we formalize egocentric reasoning using a structural causal model. Then, we ground interaction regions and expand them into hierarchical visual cues, augmented with corresponding captions, to form the initial observations. To reduce noise in these observations, we employ uncertainty calibration to filter out unreliable information. These refined observations as mediators are then incorporated into the prompt template, guiding the model to understand semantics from a first-person perspective. Extensive experiments conducted on the EgoThink benchmark demonstrate that our FRUIT method consistently enhances the performance of existing LVLMs on six distinct tasks. Our code is available at https://github.com/Mrshenshen/FRUIT.
+
+# 1 Introduction
+
+Recently, large vision-language models (LVLMs) (Bai et al., 2023; Zhu et al., 2024) have made significant progress in third-person scene understanding tasks (Li et al., 2023b, 2024a). Meanwhile, the growing research interest in egocentric tasks has driven the expansion of LVLMs into domains such
+
+![](images/33b2c0f3b01ed013f955f7fadd8a67e835ed3f0677cbfd7ac7f510989493a57e.jpg)  
+Figure 1: Illustrative examples of existing problems: the recent LVLM MiniGPT4 (Zhu et al., 2024) fails to accurately focus on human-object interactions and recognize the object held in the right hand as a phone rather than a suitcase. Such results exemplify the limitations of existing LVLMs in egocentric reasoning.
+
+as intelligent agents (Liu et al., 2024b) and wearable devices (Zhang et al., 2024b).
+
+However, the shift towards egocentric tasks, which requires models to interpret the dynamic interactions between agents and their environment from the subjective viewpoint of the agent, exposes limitations inherent in current LVLMs. These limitations stem from a critical mismatch: LVLMs typically trained on third-person data, lack the mechanisms necessary to prioritize agent-environment interactions that are crucial for egocentric reasoning. Such reasoning is essential for applications like intelligent agents and wearable devices.
+
+Existing approaches generally adopt two primary strategies: 1) fine-tuning models on egocentric datasets (Kukleva et al., 2024), typically limited to specific egocentric tasks; 2) applying prompt engineering techniques (Lin et al., 2024). However, both paradigms share a fundamental limitation: they fail to adequately explore the underlying causal relationships from a first-person perspective, i.e., how actions, events, and contextual factors interact to produce specific outcomes. This drawback leads to suboptimal egocentric reasoning ability (Cheng et al., 2024) across a broad spectrum of first-person tasks. (Wang et al., 2023; Li et al., 2022). As shown in Figure 1, the recent LVLM MiniGPT4 (Zhu et al., 2024) fails to accurately rec
+
+![](images/cb444e137efd5e3142e6bd58b06d2ab2d275ff43ef3573fd57016ccdaec3178e.jpg)  
+Figure 2: Pipeline representation through Causal Modeling: (a) The Structural Causal Model including prompt input  $X$ , Output  $O$  and latent confounder  $L$ ; (b) The proposed Egocentric Front-Door Adjustment that helps model better reason from a first-person perspective.
+
+![](images/73170ac8c4af8be5c6adc8c128b4fbed02b25fe18fcd4802314d548713fc462d.jpg)
+
+ognize object interactions and understand spatial relationships from a first-person perspective, as it does not prioritize active objects in the agent's view, missing key contextual cues for effective reasoning.
+
+To address this problem, inspired by the two-stage reasoning of human cognition (Clark, 2013), i.e., first gathering interaction-centric observations and then deriving contextual understanding, we propose a novel paradigm termed Front-Door Adjustments with Uncertainty Calibration (FRUIT) for enhancing the egocentric reasoning abilities of LVLMs. The FRUIT consists of two key stages: 1) Mediated Observation: Introducing hierarchical egocentric interaction cues as causal mediators, such as human-object affordance visual grounding, effectively circumventing the confounding biases from third-person training. 2) Contextual Understanding: The model refines its understanding by focusing on key observations, making more accurate and context-driven reasoning decisions.
+
+Specifically, as illustrated in Figure 2(a), instead of conventional prompting methods (Lin et al., 2024), we first capture the causal relationships between input prompts and outputs through a Structural Causal Model (SCM). Building upon this, as shown in Figure 2(b), we propose an Egocentric Front-Door Adjustment to introduce hierarchical multi-modal egocentric cues as mediated observations within the LVLM reasoning pipeline, using the grounding model (Liu et al., 2023c). Moreover, to reduce the influence of noise in cues, we develop an Uncertainty Calibration Mechanism to select reliable egocentric cues. Following the front-door criterion (Shanmugam, 2001), such cues are treated as mediator variables  $C$  between prompt  $X$  and output  $O$ , with latent confounder edges  $L$  omitted for simplicity. This framework enables the model to focus its reasoning on interaction-centric observations while maintaining a comprehensive under
+
+standing of the global semantic context. We evaluate our FRUIT method on the EgoThink benchmark dataset (Cheng et al., 2024). Extensive experimental results prove that our method can consistently improve the egocentric reasoning ability of existing LVLMs on six various egocentric tasks.
+
+Our contributions can be summarized as follows: 1) We propose a novel paradigm named Front-Door Adjustments with Uncertainty Calibration (FRUIT) to improve the egocentric reasoning ability of LVLMs by simulating the two-stage reasoning process in human cognition: observation and understanding. 2) We propose an Egocentric Front-Door Adjustment (EFDA) scheme to introduce mediated observation, which aims to guide the model in focusing on the human-object interaction, thus reasoning from a first-person perspective. 3) We design an Uncertainty Calibration Mechanism (UCM) to effectively filter out unreliable cues by modeling their inherent uncertainty.
+
+# 2 Related Work
+
+Large Vision-Language Models. Initially confined to natural language processing, LLMs have recently extended their capabilities to multi-modal tasks, particularly through the development of Large Vision-Language Models (LVLMs) (Chen et al., 2024; Su et al., 2023). These models typically undergo a two-stage training process: pretraining for feature alignment and instruction-based fine-tuning, enabling them to achieve strong performance across tasks like visual question answering (Wang et al., 2024), object detection (Li et al., 2024b), and image segmentation (Li et al., 2024c). While most LVLM evaluations are designed based on third-person data, emerging benchmarks prioritize egocentric assessments (Cheng et al., 2024). In this work, we aim to enhance egocentric processing and reasoning abilities of LVLMs.
+
+Causal Mechanism. Causal mechanisms (Rohekar et al., 2023; Zhang et al., 2024a) in LVLMs have recently garnered significant attention due to their potential to address complex reasoning tasks. To refine the reasoning process within LVLMs, several methods have utilized existing causal frameworks, such as Invariant Risk Minimization (IRM) (Arjovsky et al., 2019) and Structural Causal Model (SCM) (Schölkopf et al., 2012), which aim to explore the causal relationships. For example, SCM has provided a structured foundation to dissect causal relationships (Rohekar et al., 2023), allow
+
+![](images/81a8f9071c45b12914bbc4c8d1bff0f2f976be89a927e38f96cad486b94402e1.jpg)  
+Figure 3: The overall framework of the proposed FRUIT method. It includes two components: Egocentric Front-Door Adjustment (EFDA) that contains Interventional Cue Generation (ICG) and Output Influence via Cue Mediation (OICM), and Uncertainty Calibration Mechanism (UCM). The "Caption" denotes the deletion of caption.
+
+ing LVLMs to isolate causal factors that contribute directly to the reasoning process. In this paper, we model the reasoning pipeline of LVLMs in first-person scenarios using an SCM.
+
+Uncertainty-based Learning. Uncertainty-based learning addresses the challenge of quantifying prediction confidence, with uncertainty typically classified into epistemic and aleatoric types (Kendall and Gal, 2017). Aleatoric uncertainty arises from inherent data noise, while epistemic uncertainty reflects a lack of knowledge about the model structure. Methods like Bayesian networks (Gal and Ghahramani, 2016) and Subjective Logic (Jsang, 2018) are used to quantify epistemic uncertainty, often employing Dempster-Shafer theory (Yager and Liu, 2008). To effectively filter out the unreliable egocentric cues, we design an Uncertainty Calibration Mechanism to model their uncertainty.
+
+# 3 Proposed Method
+
+Preliminary Overview. Our primary goal is to enhance the egocentric reasoning of LVLMs to generate precise responses in first-person contexts. Given an input prompt  $X$  including an image  $I$  and text  $T$ , we first employ a grounding model to derive hierarchical visual egocentric cues  $\mathbf{H}_r$  and textual cues  $\mathbf{H}_c$ , which can be considered as mediated observations. After filtering noise in the observation, we incorporate observations into the prompt template following the front-door criterion to obtain a set of answer candidates  $A$ . Finally, we compute the CLIP score between each candidate in  $A$  and the input image, selecting the candidate
+
+with the highest score as the final output  $O$ .
+
+# 3.1 Background of Structural Causal Model
+
+Structural Causal Model. An SCM (Shanmugam, 2001) aims to capture causal mechanisms between variables. Following prior works (Rohekar et al., 2023), an SCM can be formalized as a Directed Acyclic Graph (DAG), denoted as  $G = (V, E)$ , where  $V$  represents the set of variables and  $E$  denotes the directed edges that signify direct causal relationships between these. In this context, the variables in DAG are topologically ordered, ensuring each variable precedes its causal descendants, based on fundamental assumptions outlined below: Definition 1 (Causal Markov). A variable in a causal graph that satisfies the Causal Markov Condition is independent of all other variables, excluding its direct effects, given its direct causes.
+
+Definition 2 (Faithfulness). A distribution is faithful to a graph if and only if all independence relations that hold within the distribution are also captured by the corresponding graphical structure.
+
+Pipeline Modeling through SCM. We use an SCM to capture the causal relationships and dependencies between input prompts and outputs in LVLMs for egocentric tasks. Specifically, as illustrated in Figure 2(b), the SCM is represented as a directed acyclic graph  $G = (V, E)$ . We define the components of our SCM as follows: 1)  $X$ : The input prompt, comprising visual and textual elements from the egocentric environment. 2)  $O$ : The LVLM output in response to the input prompt  $X$ , such as scene descriptions or object identifica
+
+tion. 3)  $L$ : The latent confounder, representing latent variables that simultaneously affect both input  $X$  and output  $O$ , potentially introducing biases or noise in the learned relationships. Moreover, in Figure 2(b), the notation  $X \rightarrow O$  denotes a direct causal relationship, illustrating that the varying multi-modal input prompts  $X$  in egocentric context lead to predictable adjustments in outputs  $O$ . The arrows  $L \rightarrow X$  and  $L \rightarrow O$  denote confounding influences from unobserved variables.
+
+# 3.2 Egocentric Front-Door Adjustment
+
+Due to the inaccessibility of the confounding variable  $L$ , the back-door adjustment is not feasible. Therefore, we follow the front-door criterion to integrate hierarchical multimodal egocentric cues as mediated variables, simulating the human cognition observation stage. It facilitates LVLMs focusing on human-centric interactions rather than contextual intentions on all semantics.
+
+Specifically, as illustrated in Figure 2(b), the proposed EFDA incorporates egocentric cues as a mediator  $C$  between the prompt  $X$  and output  $O$  which serves as a bridge to account for the indirect influence of the prompt on the answer. Moreover, following the front-door criterion (Shanmugam, 2001), we ignore the confounder of  $L$  with other variables. To quantify the causal effect between  $X$  and  $O$ , we employ the causal intervention through the do-operation, which is formulated as follows:
+
+$$
+P (O | d o (X)) = P (O | d o (C)) P (C | d o (X)), \tag {1}
+$$
+
+where the causal effect of  $P(O|do(X))$  between  $X$  and  $O$  can be split into:  $P(C|do(X))$  represents the probability of generating specific egocentric cues  $C$  from input  $X$ , and  $P(O|do(C))$  denotes the effect of those cues on output  $O$ .
+
+Interventional Cue Generation  $P(C|do(X))$ . To improve the egocentric reasoning ability of LVLMs, given the input  $X$  including textual prompts  $T$  and images  $I$ , we introduce the observation by constructing a hierarchical set of multi-modal egocentric cues  $C$ . Specifically, for visual cues, we utilize the grounding model (Liu et al., 2023c) to generate bounding boxes,  $b_{1}$  and  $b_{2}$ , corresponding to each hand. When no hand is detected, the middle-lower area is designated as the cropped region. Based on the boxes coordinates, we extract a cropped region,  $H_r^0$ , containing the hands and objects involved in the interaction, which can be formulated as:
+
+$$
+H _ {r} ^ {0} = f (I, b _ {1}, b _ {2}), \tag {2}
+$$
+
+where  $f$  represents the cropping function applied to image  $I$  based on the coordinates of  $b_{1}$  and  $b_{2}$ .
+
+However, focusing overly on this cropped region may exclude global-level semantic information. Therefore, we construct a hierarchical set of cropped regions  $\mathbf{H}_r = \{H_r^i\}_{i=0}^3$  by expanding each region by  $20\%$  around its center. Here  $H_r^0$  represents the minimal region focusing on the interaction, while higher values of  $i$  indicate increasingly broader regions. Based on the region set  $\mathbf{H}_r$ , LVLMs are prompted to generate a corresponding hierarchical set of captions  $\mathbf{H}_c$  as follows:
+
+$$
+\mathbf {H} _ {c} = \left\{H _ {c} ^ {i} \right\} _ {i = 0} ^ {3} = \operatorname {L V L M} \left(\mathbf {H} _ {\mathrm {r}}\right), \tag {3}
+$$
+
+The causal effect between input  $X$  and multimodal egocentric cues  $C$  can be formulated as:
+
+$$
+P (C | d o (X)) = \frac {\sum_ {i = 1} ^ {| C |} \left(S \left(H _ {r} ^ {i} , H _ {c} ^ {i}\right) + S \left(H _ {c} ^ {i} , H _ {r} ^ {i}\right)\right)}{| | C | |}, \tag {4}
+$$
+
+where  $c_{i}$  denotes the  $i$ -th pair of multi-modal cues  $H_{r}^{i}$  and  $H_{c}^{i}$ , with  $||C||$  representing the total number of multi-modal cue pairs. The function  $S(\cdot)$  is used to compute the cosine similarity.
+
+Output Influence via Cue Mediation  $P(O|do(C))$ . Based on the generated multi-modal egocentric cues  $C$ , i.e., mediated observations, we guide the LVLMs to perform egocentric scene understanding. Specifically, given the egocentric cues  $C$ , we construct the final input prompt  $P_r$  after mediation, which can be expressed as follows:
+
+$$
+P _ {r} = [ X, C ], \tag {5}
+$$
+
+The prompt details are available in supplementary materials. Then we query the LVLMs  $N$  times, obtaining N answers  $A = \{A^k\}_{k = 0}^N$  using the final prompt  $P_{r}$ . We estimate the probability of outputs by calculating CLIP score (Deng et al., 2024) between each answer and input image as follows:
+
+$$
+P (O | d o (C)) = \operatorname {C L I P} (O, I). \tag {6}
+$$
+
+# 3.3 Reasoning Uncertainty Modeling
+
+While the front-door adjustment helps the model focus on relevant egocentric cues, there remains the challenge of noise in these cues, i.e., not all egocentric cues generated from the proposed EFDA are reliable. To address this, we introduce an uncertainty-based mechanism to selectively filter out unreliable cues. In specific, we employ the Subjective Logic (SL) (Jsang, 2018) principle to quantify the uncertainty associated with the hierarchical textual
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">Object</td><td rowspan="2">Activity</td><td colspan="2">Localization</td><td rowspan="2">Fore.</td><td colspan="3">Reasoning</td></tr><tr><td>Exist</td><td>Attr</td><td>Afford</td><td>Loc</td><td>Spatial</td><td>Count</td><td>Compar</td><td>Situated</td></tr><tr><td>LLaVA-1.5-7B (2023)</td><td>33.0</td><td>47.0</td><td>54.0</td><td>35.5</td><td>35.0</td><td>49.0</td><td>27.0</td><td>20.0</td><td>47.0</td><td>37.0</td></tr><tr><td>+ FRUIT</td><td>61.0</td><td>75.0</td><td>61.0</td><td>58.0</td><td>79.0</td><td>61.0</td><td>39.5</td><td>31.0</td><td>47.0</td><td>52.0</td></tr><tr><td>LLaVA-1.5-7B-LoRA (2023)</td><td>63.0</td><td>60.0</td><td>63.0</td><td>55.0</td><td>79.0</td><td>49.0</td><td>26.0</td><td>50.0</td><td>63.0</td><td>61.0</td></tr><tr><td>+ FRUIT</td><td>85.0</td><td>71.0</td><td>66.0</td><td>77.0</td><td>85.0</td><td>63.0</td><td>58.0</td><td>53.0</td><td>70.0</td><td>70.5</td></tr><tr><td>InstructBLIP-7B (2023)</td><td>50.0</td><td>33.0</td><td>45.0</td><td>47.5</td><td>77.0</td><td>38.0</td><td>40.5</td><td>18.0</td><td>43.0</td><td>67.0</td></tr><tr><td>+ FRUIT</td><td>61.0</td><td>57.0</td><td>49.0</td><td>50.0</td><td>90.0</td><td>57.0</td><td>45.5</td><td>31.0</td><td>56.0</td><td>70.0</td></tr><tr><td>Otter-I-7B (2023)</td><td>48.0</td><td>56.0</td><td>39.0</td><td>44.0</td><td>60.0</td><td>44.0</td><td>38.0</td><td>39.0</td><td>48.0</td><td>42.0</td></tr><tr><td>+ FRUIT</td><td>57.0</td><td>62.0</td><td>45.0</td><td>44.5</td><td>63.0</td><td>49.0</td><td>45.5</td><td>42.0</td><td>50.0</td><td>44.0</td></tr><tr><td>PandaGPT-7B (2023)</td><td>40.0</td><td>56.0</td><td>41.0</td><td>37.0</td><td>61.0</td><td>52.0</td><td>43.0</td><td>19.0</td><td>52.0</td><td>44.0</td></tr><tr><td>+ FRUIT</td><td>49.0</td><td>57.0</td><td>46.0</td><td>48.0</td><td>63.0</td><td>57.0</td><td>48.0</td><td>30.0</td><td>54.0</td><td>48.0</td></tr><tr><td>MiniGPT4-7B (2024)</td><td>50.0</td><td>56.0</td><td>37.0</td><td>39.0</td><td>55.0</td><td>49.0</td><td>41.5</td><td>14.0</td><td>48.0</td><td>31.0</td></tr><tr><td>+ FRUIT</td><td>55.0</td><td>57.0</td><td>40.0</td><td>50.0</td><td>56.0</td><td>53.0</td><td>43.0</td><td>18.0</td><td>48.0</td><td>46.0</td></tr><tr><td>ShareGPT4V-7B (2024)</td><td>67.0</td><td>75.0</td><td>53.0</td><td>55.5</td><td>77.0</td><td>62.0</td><td>47.0</td><td>30.0</td><td>38.0</td><td>66.0</td></tr><tr><td>+ FRUIT</td><td>75.0</td><td>76.0</td><td>58.0</td><td>60.5</td><td>81.0</td><td>63.0</td><td>51.0</td><td>39.0</td><td>44.0</td><td>67.0</td></tr><tr><td>mPLUG-Owl2-7B (2024)</td><td>58.0</td><td>61.0</td><td>44.0</td><td>41.0</td><td>84.0</td><td>45.0</td><td>36.0</td><td>44.0</td><td>54.0</td><td>45.0</td></tr><tr><td>+ FRUIT</td><td>67.0</td><td>74.0</td><td>52.0</td><td>57.0</td><td>86.0</td><td>48.0</td><td>56.0</td><td>50.0</td><td>62.0</td><td>56.0</td></tr><tr><td>Janus-Pro-7B (2025)</td><td>71.0</td><td>76.0</td><td>54.0</td><td>66.5</td><td>88.0</td><td>71.0</td><td>55.0</td><td>56.0</td><td>68.0</td><td>56.5</td></tr><tr><td>+ FRUIT</td><td>81.0</td><td>79.0</td><td>70.0</td><td>78.5</td><td>95.0</td><td>75.0</td><td>58.0</td><td>64.0</td><td>76.0</td><td>63.0</td></tr></table>
+
+Table 1: Comparisons with existing 7B-sized LVLMs on the EgoThink benchmark (Cheng et al., 2024), including Object, Activity, Localization, Reasoning, and Forecasting. Results presented under the shaded area indicate they have been pre-trained on first-person data. Bolded values represent the best performance.
+
+egocentric cues, which can detect unreliable captions within the set  $\mathbf{H}_c$ . Initially, we calculate the similarity  $S(\mathbf{H}_c,\mathbf{H}_r)$  between the generated captions  $\mathbf{H}_c$  and corresponding regions  $\mathbf{H}_r$  to derive the caption evidence  $\mathbf{e} = \{e_i\}_{i = 0}^3$  as follows:
+
+$$
+\mathbf {e} = f _ {e} \left(S \left(\mathbf {H} _ {c}, \mathbf {H} _ {r}\right)\right) = e x p ^ {F \left(S \left(\mathbf {H} _ {c}, \mathbf {H} _ {r}\right)\right)}, \tag {7}
+$$
+
+where  $F$  denotes the ReLU activation function.
+
+Based on the caption evidence  $e_i$ , we compute  $\alpha_{i}$  and characterize the uncertainty  $\mathbf{U}$  as follows:
+
+$$
+\alpha_ {i} = e _ {i} + 1, \quad \mathbf {U} = \frac {N}{\mathbf {S} _ {d}}, \tag {8}
+$$
+
+where  $\mathbf{S}_d = \sum_{i=1}^{N} \alpha_i$  serves as the intensity parameter for the Dirichlet distribution. The distribution can be formulated as:
+
+$$
+D (\mathbf {p} | \boldsymbol {\alpha}) = \left\{ \begin{array}{l l} \frac {1}{B (\boldsymbol {\alpha})} \prod_ {j = 1} ^ {N} p _ {j} ^ {\alpha_ {j} - 1} & \mathbf {p} \in \mathcal {S} _ {N}, \\ 0 & \text {e l s e ,} \end{array} \right. \tag {9}
+$$
+
+where  $B(\alpha)$  denotes the  $N$ -dimensional beta function, and  $\mathcal{S}_N$  represents the  $N$ -dimensional unit simplex. The uncertainty  $u_i \in \mathbf{U}$  acts as a quantitative measure of the reliability of each textual cue  $c_i$ . To filter out the unreliable textual cues, we obtain the refined multi-modal cues  $C$  by setting a threshold  $\lambda$ , removing the noisy cues when  $u_i > \lambda$ .
+
+Final Outputs. Based on the above uncertainty-based refinement, the final answer is obtained by:
+
+$$
+\max  _ {C, X} P (O | d o (X)) = P (O | d o (C)) P (C | d o (X)), \tag {10}
+$$
+
+We choose the answer with the largest probability as the final output of LVLMs.
+
+# 4 Experiments
+
+# 4.1 Experimental Setup
+
+Datasets. The EgoThink dataset (Cheng et al., 2024) consists of 700 first-person images from the Ego4D (Grauman et al., 2022) video resource, each paired with a detailed question-answer set. It is organized into six core abilities, including Activity, Localization, Reasoning, Forecasting, and Planning, which are employed to evaluate the capacity of LVLMs to understand first-person images.
+
+Evaluation Metrics. Following (Cheng et al., 2024), we use GPT-4 (Achiam et al., 2023) as an automatic evaluator to assess generated answers. By comparing the model's output with the reference answer, GPT-4 determines accuracy based on answer similarity. A scoring system is applied: 0 points for incorrect answers, 0.5 for partial correctness, and 1 for correct answers.
+
+Implementation Details. In this study, we employ the PyTorch framework to implement our model. We conducted experiments utilizing 14 commonly used large vision-language models (LVLMs). To ensure a fair comparison and account for potential parameter influences, the models were divided into two groups: 7B and 13B. All models were evaluated under a zero-shot setting using the EgoThink benchmark. Given the multi-modal inputs, we first utilize the GroundingDINO (Liu et al., 2023c) identify and crop regions proximal to the hand in the images. Base on these, we generate the corresponding captions and then select appropriate image-caption pairs as hierarchical multi-modal egocentric cues.
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">Object</td><td rowspan="2">Activity</td><td colspan="2">Localization</td><td rowspan="2">Fore.</td><td colspan="3">Reasoning</td></tr><tr><td>Exist</td><td>Attr</td><td>Afford</td><td>Loc</td><td>Spatial</td><td>Count</td><td>Compar</td><td>Situated</td></tr><tr><td>LLaVA-1.5-13B (2023)</td><td>54.0</td><td>62.0</td><td>52.0</td><td>46.0</td><td>53.0</td><td>46.0</td><td>44.0</td><td>26.0</td><td>44.0</td><td>29.0</td></tr><tr><td>+ FRUIT</td><td>75.0</td><td>70.0</td><td>54.0</td><td>61.0</td><td>77.0</td><td>69.0</td><td>48.0</td><td>38.0</td><td>57.0</td><td>62.0</td></tr><tr><td>MiniGPT4-13B (2024)</td><td>47.0</td><td>48.0</td><td>31.0</td><td>28.0</td><td>60.0</td><td>40.0</td><td>38.5</td><td>33.0</td><td>33.0</td><td>46.0</td></tr><tr><td>+ FRUIT</td><td>60.0</td><td>50.0</td><td>39.0</td><td>47.5</td><td>64.0</td><td>43.0</td><td>40.0</td><td>35.0</td><td>53.0</td><td>48.0</td></tr><tr><td>LLaVA-NeXT-13B (2024)</td><td>65.0</td><td>72.0</td><td>36.0</td><td>34.0</td><td>60.0</td><td>57.0</td><td>28.0</td><td>43.0</td><td>46.0</td><td>18.5</td></tr><tr><td>+ FRUIT</td><td>86.0</td><td>79.0</td><td>69.0</td><td>69.5</td><td>78.0</td><td>69.0</td><td>43.0</td><td>44.0</td><td>65.0</td><td>40.0</td></tr></table>
+
+Table 2: Comparisons with existing 13B-sized LVLMs on the EgoThink (Cheng et al., 2024).
+
+These cues are introduced to make front-door adjustments for the LVLM reasoning process.
+
+# 4.2 Overall Comparison Results
+
+LVLM baselines. We evaluate the efficacy of our FRUIT method on 11 widely used LVLMs. Overall, we divide them into two groups: (1) 7B-sized Models: LLaVA-1.5-7B (Liu et al., 2023b), InstructBLIP-7B (Dai et al., 2023), MiniGPT4-7B (Zhu et al., 2024), Otter-I-7B (Li et al., 2023a), PandaGPT-7B (Su et al., 2023), ShareGPT4V-7B (Chen et al., 2024), mPLUG-Owl2-7B (Ye et al., 2024), Janus-Pro-7B (Chen et al., 2025). (2) 13B-sized Models: LLaVA-1.5-13B (Liu et al., 2023b), MiniGPT4-13B (Zhu et al., 2024), LLaVA-NeXT-13B (Liu et al., 2024a). More results are provided in supplementary materials.
+
+Results on the EgoThink benchmark. According to the comparison on EgoThink reported in Table 1 and Table 2, we can find that: (1) Our FRUIT method demonstrates marked improvements in accuracy across various partitions of the EgoThink dataset. These results suggest that our method enhances the model's ability to accurately extract crucial information in first-person context. (2) Moreover, even when compared to the latest LVLM Janus-Pro-7B (Chen et al., 2025) proposed by DeepSeek or LLaVA-1.5 pretrained on egocentric data, our FRUIT method shows a substantial improvement across various metrics. We hypothesize that this performance enhancement is attributed to the proposed egocentric front-door adjustments, which incorporate mediated observations to help guide LVLMs in more effectively interpreting semantics from a first-person perspective.
+
+Results on the Planning Task. As illustrated in Figure 4, we evaluated our FRUIT method on both the planning assistant and planning navigation tasks. The results reveal several findings: (1) Our method achieved superior results in two data categories across the six models evaluated, indicating its efficacy in extracting information from a first-person perspective. (2) Compared to the re
+
+![](images/1057ac086c240786ed6eeba690457dc1ba87b960aa8308726e5809d20b58325a.jpg)  
+Figure 4: Performance comparison of 7B-sized LVLMs on Planning Assistant and Planning Navigation. Six models are evaluated: LLaVA-1.5 (LLaVA), mPLUG-Owl (Owl), InstructBLIP (BLIP), ShareGPT4V (Share), Otter-Image (Otter) and MiniGPT4 (Mini).
+
+![](images/babfad75e037d9887e2f26d19387043a7e0b12a7456239d06775ad1064148116.jpg)
+
+sults obtained w/o textual cues, our FRUIT method achieves superior performance across both evaluation metrics. These improvements underscore the critical role of incorporating hierarchical, multimodal egocentric cues in capturing the key information from a first-person perspective.
+
+# 4.3 Further Analysis
+
+Ablation Study. As presented in Table 3, we list the following conclusions: (1) The comparison with No.0 and No.3 reveals that our proposed EFDA significantly enhances the performance of LVLMs on a wide range of egocentric tasks. Such results demonstrate the efficacy of integrating hierarchical multi-modal egocentric cues as mediators. It can help LVLMs capture both global-level semantics and human-object interactions, thereby facilitating reasoning from a first-person perspective. (2) From the comparison of No.1 and No.3, No.2 and No.3, we speculate that the interaction between the ICG and OICM is crucial for understanding knowledge within an egocentric context. One reasonable reason is that the ICG provides a structured representation of class hierarchies and relational mappings, while the OICM enhances this framework by adding contextual interactions that are unique to the egocentric perspective. More
+
+<table><tr><td rowspan="3">No.</td><td colspan="3">Components</td><td colspan="9">Metrics</td></tr><tr><td rowspan="2">UCM</td><td colspan="2">EFDA</td><td colspan="3">Object</td><td rowspan="2">Activity</td><td colspan="2">Localization</td><td colspan="3">Reasoning</td></tr><tr><td>ICG</td><td>OICM</td><td>Exist</td><td>Attr</td><td>Afford</td><td>Loc</td><td>Spatial</td><td>Count</td><td>Compar</td><td>Situated</td></tr><tr><td>0</td><td>-</td><td>-</td><td>-</td><td>33.0</td><td>47.0</td><td>54.0</td><td>35.5</td><td>35.0</td><td>49.0</td><td>20.0</td><td>41.0</td><td>37.0</td></tr><tr><td>1</td><td>-</td><td>✓</td><td>-</td><td>58.0</td><td>62.0</td><td>57.5</td><td>45.5</td><td>63.0</td><td>54.5</td><td>29.0</td><td>42.0</td><td>45.5</td></tr><tr><td>2</td><td>-</td><td>-</td><td>✓</td><td>56.0</td><td>61.0</td><td>56.5</td><td>42.0</td><td>61.0</td><td>58.0</td><td>28.0</td><td>47.0</td><td>42.0</td></tr><tr><td>3</td><td>-</td><td>✓</td><td>✓</td><td>69.0</td><td>70.0</td><td>58.0</td><td>57.0</td><td>76.0</td><td>58.0</td><td>37.5</td><td>45.5</td><td>50.0</td></tr><tr><td>4</td><td>✓</td><td>✓</td><td>-</td><td>63.0</td><td>65.0</td><td>58.0</td><td>55.0</td><td>68.0</td><td>55.5</td><td>32.0</td><td>43.0</td><td>46.5</td></tr><tr><td>5</td><td>✓</td><td>✓</td><td>✓</td><td>73.0</td><td>75.0</td><td>61.0</td><td>58.0</td><td>79.0</td><td>61.0</td><td>41.0</td><td>47.0</td><td>52.0</td></tr></table>
+
+Table 3: Ablation analysis of key components on the EgoThink benchmark dataset. The baseline is LLaVA-7B here.
+
+![](images/0458c89d481671b8ba345c9966799e9075c53684d550d6c7061ef123c5ea91f7.jpg)  
+Figure 5: Performance comparison of our FRUIT method against In-Context Learning (ICL) and Chain-of-Thought (COT) prompting strategies across distinct subsets: Affordance (Aff.), Activity (Act.), Spatial (Spa.), Counting (Co.), and Forecasting (Fore.).
+
+![](images/b4c4a5001ecfdd5c4611f5328a6bd1ad48ce28f1fa92fe98a813f82640671326.jpg)
+
+details are available in supplementary materials.
+
+Analysis on Different Prompting Techniques. We compare the performance of our proposed FRUIT with conventional prompting methods, including In-Context Learning (ICL) and Chain-of-Thought (CoT). Note that all methods under consideration incorporate grounded interactions to ensure a fair comparison. By observing Figure 5, we can find that: The performance comparisons reveal a substantial improvement using the FRUIT method compared to traditional prompting techniques. These results indicate that these methods fail to capture causal reasoning relationships between interacting objects and user's intentions. In contrast, FRUIT enables LVLMs to process and reason from an egocentric perspective, reinforcing our rationale that simulating human causal reasoning can enhance egocentric reasoning.
+
+Probability Estimation Selection in  $P(O|do(C))$ . We investigate the influence of different probability estimations in  $P(O|do(C))$  on Reasoning and Object tasks. Observations drawn from Figure 6 are as follows: (1) The majority voting strategy demonstrates a more accurate output in reasoning tasks involving high uncertainty, likely due to the aggregation of individual judgments, effectively mitigating the impact of outlier predictions. In contrast, the CLIP score exhibits a higher sensitivity to contextual variations, leading to more nuanced predictions, especially in tasks that require intricate
+
+![](images/178e4241ce1aca80060ee8e336b805a0da4a57010ef2bfc5cef6b6d1c9bf0110.jpg)  
+Figure 6: Effect of  $P(O|do(C))$  estimation strategies. Performance compared across w/o  $P(O|do(C))$ , majority voting, and CLIP score (Ours).
+
+![](images/206cc7783e7b0b954a9843df3cc42cfd0a15e4d0aa8f76ea974a8f25bf59100f.jpg)
+
+![](images/cd9fa9a1a0a99f3832f6127ee1faae5dacf75d2d226fc83123de5d0009f6ee72.jpg)  
+Figure 7: (a) Visualization of the uncertainty distribution of textual cues, illustrating the density of uncertainty levels. (b) Comparison on the Localization subset across different noise selection methods: without UCM, using cosine similarity (Sim), and with UCM (Ours).
+
+![](images/c6230e8ed54f549e350ea0f567f82b1cd9f3f8f7cc67e4b66d2f130fcf3cecce.jpg)
+
+understanding of semantic relationships between objects and actions. (2) When using the first predicted answer without applying  $P(O|do(C))$ , the performance of w/o  $P(O|do(C))$  consistently suffers across all tasks. This indicates that relying solely on direct predictions, without incorporating egocentric reasoning through causal intervention, undermines the accuracy of final decision-making. Visualization and Effect of Uncertain Cues. To further substantiate the presence of uncertain cues and validate the importance of our proposed Uncertainty Calibration Mechanism (UCM), we visualize the distribution of uncertainty for textual cues and compare the impact of various noise selection strategies, employing the MiniGPT4-7b model. As depicted in Figure 7, we can find that: (1) In Figure 7(a), the distribution of normalized uncertainty differs between regular tokens and hallucinations.
+
+<table><tr><td>Models</td><td>P@1</td><td>P@3</td><td>P@5</td></tr><tr><td>LLaVA-1.5-7b (Liu et al., 2023a)</td><td>31.0</td><td>46.0</td><td>52.0</td></tr><tr><td>+FRUIT</td><td>59.0</td><td>79.0</td><td>80.0</td></tr><tr><td>ShareGPT4V-7b (Chen et al., 2024)</td><td>28.0</td><td>46.0</td><td>55.0</td></tr><tr><td>+FRUIT</td><td>60.0</td><td>77.0</td><td>85.0</td></tr></table>
+
+![](images/1d37a991348a5198c624b5f55f98c16e5e11d26e8c62442ca4cca331b7c78ab0.jpg)  
+Figure 8: Performance comparison of the proposed FRUIT method against VCD (Leng et al., 2024) and OPERA (Huang et al., 2024).
+
+![](images/cfb1c2a7a3c0fcfaa8ef948aa92626c102fbfb988ef069b5f5f980dc9b3e1ce2.jpg)
+
+Hallucinations have a higher density at the upper end of the uncertainty scale, while regular tokens cluster towards lower levels. This difference validates the effectiveness of our proposed UCM in detecting unreliable textual cues. (2) As illustrated in Figure 7(b), our FRUIT method consistently outperforms the similarity-based strategy w/ Sim. The improvement reflects the efficacy of our approach in reducing the impact of noisy cues, which in turn enhances egocentric reasoning ability.
+
+Evaluation of Object Interaction Focus. To evaluate the LVLMs' ability to focus on correct objects in first-person scenarios, we define the  $\mathrm{P@K}$ , a metric that quantifies the probability that the correct answer appears in the top-K candidate predictions. For example, given 10 image-question pairs, if there are 4 correct answers within the top-5 candidate predictions from LVLMs, then  $\mathrm{P@5} = 4 / 10 = 0.4$ . From results in Table 6, we can list the following conclusions: (1) LVLMs with FRUIT demonstrate a higher likelihood of identifying relevant objects, due to the cues from the observation stage, which guide the model to focus on human-object interaction regions. (2) While multimodal cues from local interaction regions can generate descriptions centered on human-object interactions, they may lack essential global semantic context.
+
+Qualitative Analysis. As shown in Figure 9, we present a qualitative comparison of the performance between the LVLM baselines (LLaVA1.5 (Liu et al., 2023a) and PandaGPT (Su et al., 2023)) and these models enhanced with our FRUIT method. For example, when involving multiple
+
+Table 4: Evaluation of object interaction focus for LLaVA-1.5 and ShareGPT4V using our proposed P@K, with and without the FRUIT. Higher P@K means better top-K relevant object identification accuracy.  
+
+<table><tr><td>Image</td><td>Original Question &amp; Groundtruth</td></tr><tr><td></td><td>Question: What am I doing now?
+Groundtruth: Watering</td></tr><tr><td></td><td>Model Predictions</td></tr><tr><td></td><td>PandaGPT: You are holding a tennis packet and hit a tennis ball.
++FRUIT (Ours): I am standing outside, watering the lawn with a hose in my right hand.</td></tr><tr><td>Image</td><td>Original Question &amp; Groundtruth</td></tr><tr><td></td><td>Question: What&#x27;s the use of the object in my left hand? 
+Groundtruth: Painting</td></tr><tr><td></td><td>Model Predictions</td></tr><tr><td></td><td>LLaVA-1.5: In the image, a person is holding a small figurine in their left hand. The figurine appears to be a toy or a model.
++FRUIT (Ours): The object in your left hand is a paintbrush, which is being used to paint the miniature figure.</td></tr></table>
+
+Figure 9: Comparative analysis between LVLM baselines and our proposed FRUIT method in the Activity and Object subsets. Incorrect descriptions are highlighted in red. Correct predictions are bolded in blue.
+
+interacting objects, LLaVA-1.5 with our proposed FRUIT can consistently maintain accurate object localization in dynamic environments. However, the LLaVA-1.5 baseline often erroneously identifies the relationships between the object and hands in egocentric context. This comparison shows the importance of our FRUIT method in focusing on human-object interaction, thus guiding LVLMs to reason from a first-person perspective.
+
+Comparison with Hallucination Mitigation Methods. In this study, we evaluate the performance of our FRUIT method by comparing it with recent hallucination mitigation techniques, VCD (Leng et al., 2024) and OPERA (Huang et al., 20234). As illustrated in Figure 8, our FRUIT method consistently outperforms both VCD and OPERA across all evaluation metrics. This improvement indicates that while VCD and OPERA effectively address hallucination issues in traditional third-person data, they do not enhance the reasoning capabilities of large multimodal models in first-person contexts. In contrast, our FRUIT approach directs Large Vision-Language Models (LVLMs) to perform first-person reasoning in a manner that aligns with human cognitive processes.
+
+# 5 Conclusion
+
+In this work, we introduced the Front-Door Adjustments with Uncertainty Calibration (FRUIT) framework to enhance the egocentric reasoning capabilities of large vision-language models (LVLMs). By incorporating Egocentric Front-Door Adjustments, FRUIT guide the model focus on interaction objects, while the Uncertainty Calibration Mechanism filters out unreliable information. In future work, we will explore additional strategies to further enhance LVLM performance in first-person contexts.
+
+# 6 Limitations
+
+The current evaluation relies exclusively on image-text pairs from EgoThink, which cannot fully capture the temporal dynamics of real-world egocentric perception. Video understanding requires modeling continuous viewpoint changes and action-state transitions—capabilities our hierarchical cues currently lack due to fixed spatial grounding. Future work should extend the causal mediation framework to incorporate temporal attention mechanisms for long-horizon reasoning.
+
+# Acknowledgment
+
+This work was supported in part by Meituan, and National Natural Science Foundation of China under Grants, China (No.62476201 and 62222203).
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. 5, 11  
+Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. 2019. Invariant risk minimization. arXiv preprint arXiv:1907.02893. 2  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-VL: A frontier large vision-language model with versatile abilities. arXiv preprint. 1  
+Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. 2024. Sharegpt4v: Improving large multi-modal models with better captions. European Conference on Computer Vision. 2, 6, 8  
+Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. 2025. Janus-pro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811.6  
+Sijie Cheng, Zhicheng Guo, Jingwen Wu, Kechen Fang, Peng Li, Huaping Liu, and Yang Liu. 2024. Ego-think: Evaluating first-person perspective thinking capability of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition., pages 14291-14302. 1, 2, 5, 6, 11, 12, 13, 18  
+Andy Clark. 2013. Whatever next? predictive brains, situated agents, and the future of cognitive science. Behavioral and brain sciences, 36(3):181-204. 2
+
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. 2023. Instructlip: Towards general-purpose vision-language models with instruction tuning. CoRR, abs/2305.06500. 6, 13  
+Ailin Deng, Zhirui Chen, and Bryan Hooi. 2024. Seeing is believing: Mitigating hallucination in large vision-language models via clip-guided decoding. CoRR, abs/2402.15300. 4  
+Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR. 3  
+Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. 2022. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012. 5  
+Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. 20234. OPERA: alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. CVPR. 8  
+Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. 2024. Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13418-13427. 8  
+Audun Jsang. 2018. Subjective Logic: A formalism for reasoning under uncertainty. Springer Publishing Company, Incorporated. 3, 4  
+Alex Kendall and Yarin Gal. 2017. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30. 3  
+Anna Kukleva, Fadime Sener, Edoardo Remelli, Bugra Tekin, Eric Sauser, Bernt Schiele, and Shugao Ma. 2024. X-MIC: cross-modal instance conditioning for egocentric action generalization. In CVPR, pages 26354-26363. 1  
+Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. 2024. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. CVPR. 8  
+Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. 2023a. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726. 6, 12, 13
+
+Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. 2022. Bridge-prompt: Towards ordinal action understanding in instructional videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19848-19857. 1  
+Shenshen Li, Chen He, Xing Xu, Fumin Shen, Yang Yang, and Heng Tao Shen. 2024a. Adaptive uncertainty-based learning for text-based person retrieval. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, pages 3172-3180. 1  
+Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Xin Liu, and Heng Tao Shen. 2024b. Multi-grained attention network with mutual exclusion for composed query-based image retrieval. IEEE TCSVT, 34(4):2959-2972. 2  
+Shenshen Li, Xing Xu, Xun Jiang, Fumin Shen, Zhe Sun, and Andrzej Cichocki. 2024c. Cross-modal attention preservation with self-contrastive learning for composed query-based image retrieval. ACM Trans. Multim. Comput. Commun. Appl., 20(6):165:1-165:22. 2  
+Shenshen Li, Xing Xu, Yang Yang, Fumin Shen, Yijun Mo, Yujie Li, and Heng Tao Shen. 2023b. DCEL: deep cross-modal evidential learning for text-based person retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 6292-6300. ACM. 1  
+Bingqian Lin, Yunshuang Nie, Ziming Wei, Jiaqi Chen, Shikui Ma, Jianhua Han, Hang Xu, Xiaojun Chang, and Xiaodan Liang. 2024. Navcot: Boosting llm-based vision-and-language navigation via learning disentangled reasoning. CoRR, abs/2403.07376. 1, 2  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023a. Improved baselines with visual instruction tuning. CoRR, abs/2310.03744. 8  
+Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024a. Llavanext: Improved reasoning,OCR, and world knowledge. 6,11,12  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023b. Visual instruction tuning. 6, 12  
+Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. 2024b. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. CoRR, abs/2406.04339. 1  
+Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. 2023c. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499. 2, 4, 5
+
+Raanan Y. Rohekar, Yaniv Gurwicz, and Shami Nisimov. 2023. Causal interpretation of self-attention in pretrained transformers. In NeurIPS. 2, 3  
+Bernhard Scholkopf, Dominik Janzing, Jonas Peters, Eleni Sgouritsa, Kun Zhang, and Joris Mooij. 2012. On causal and anticausal learning. arXiv preprint arXiv:1206.6471. 2  
+Ram Shanmugam. 2001. Causality: Models, reasoning, and inference. Neurocomputing, 41(1-4):189-190. 2, 3, 4  
+Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. 2023. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355. 2, 6, 8, 11  
+Lei Wang, Jiabang He, Shenshen Li, Ning Liu, and Ee-Peng Lim. 2024. Mitigating fine-grained hallucination by fine-tuning large vision-language models with caption rewrites. In MultiMedia Modeling, volume 14557, pages 32-45. 2  
+Qitong Wang, Long Zhao, Liangzhe Yuan, Ting Liu, and Xi Peng. 2023. Learning from semantic alignment between unpaired multiviews for egocentric video recognition. In IEEE/CVF International Conference on Computer Vision, pages 3284-3294. 1  
+Ronald R Yager and Liping Liu. 2008. Classic works of the Dempster-Shafer theory of belief functions, volume 219. Springer. 3  
+Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. 2023. mplug-owl: Modularization empowers large language models with multimodality. Preprint, arXiv:2304.14178. 11, 13  
+Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. 2024. MPLug-owl2: Revolutionizing multimodal large language model with modality collaboration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13040-13051. 6, 12  
+Congzhi Zhang, Linhai Zhang, and Deyu Zhou. 2024a. Causal walk: Debiasing multi-hop fact verification with front-door adjustment. In AAAI, pages 19533-19541. 2  
+Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. 2024b. Groundhog grounding large language models to holistic segmentation. In CVPR, pages 14227-14238. 1  
+Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. 2024. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR. 1, 6, 11, 12, 13
+
+# A Appendix
+
+# A.1 Overview
+
+In this supplementary material, we provide more analyses of our proposed FRUIT method, which are difficult to describe in the main paper due to space limitations. The content of the additional material is shown below:
+
+- We augmented the experimental results from two state-of-the-art models using the EgoThink (Cheng et al., 2024).  
+- We reported experimental outcomes for two additional excellent models, evaluated using our proposed P@K metric.  
+- We further explore the effect of varying hyperparameter  $\lambda$ .  
+- We conducted additional experiments to analyze the optimal number of visual egocentric cues. Specifically, we evaluated the effectiveness of the FRUIT method on images that do not contain hands.  
+- We included further examples of experimental results based on the EgoThink (Cheng et al., 2024).  
+- We provided a comprehensive list of all prompts utilized in the experiment.
+
+# A.2 Results on the EgoThink benchmark
+
+To further validate the efficacy of our FRUIT method in enhancing the interpretation of egocentric images, we conducted experiments on the EgoThink (Cheng et al., 2024) dataset using three state-of-the-art models, mPLUG-Owl (Ye et al., 2023), LLaVA-NeXT (Liu et al., 2024a) and PandaGPT (Su et al., 2023). The results of these experiments are summarized in Table 5.
+
+Our evaluation covered 10 categories within EgoThink (Cheng et al., 2024), excluding the planning category. Notably, our method consistently improved model performance across all 10 categories. These findings suggest that our approach significantly enhances the model's ability to interpret first-person perspective images, including various aspects such as object identification, spatial context, and behavioral understanding.
+
+![](images/9e725d7fd4ee2c1b6fe8e3710ee0b456885801138478908d897d50eaab43a954.jpg)  
+Figure 10: Comparative analysis of Localization (Loc.) and Forecasting (Fore.) performance metrics across varying values of the parameter  $\lambda$  within the range [0.1, 0.6].
+
+# A.3 Results on the  $\mathbf{P}@\mathbf{K}$  benchmark
+
+Our P@K metric employs 100 labeled images, each representing an object with the highest degree of interaction with the character. By requiring the model to identify the top K objects with the highest human interaction and utilizing GPT-4 (Achiam et al., 2023) to verify the presence of labeled object among the K selected objects, we evaluate the model's capability to interpret egocentric images from a first-person perspective. We implemented this metric using the two latest state-of-the-art models, LLaVA-NeXT(Liu et al., 2024a) and MiniGPT4 (Zhu et al., 2024). The experimental results in Table 6 demonstrate that our method significantly enhances the performance of the model, particularly in identifying key items. This finding substantiates the fact that our approach effectively improves the model's comprehension of first-person images.
+
+# A.4 Effect of Varying Hyper-parameter  $\lambda$
+
+In this section, we further explore the influence of the hyper-parameter  $\lambda$ , which is designed to filter out unreliable cues when the uncertainty exceeds a threshold, with  $\lambda = 0.4$  serving as a critical value. Figure 10 presents the performance trends for Localization (Loc.) and Forecasting (Fore.) metrics as  $\lambda$  varies from 0.1 to 0.6. Localization performance peaks around  $\lambda = 0.4$ , demonstrating stability and robustness within the range  $\lambda \in [0.2, 0.5]$ . This suggests that the model effectively handles moderate variations in  $\lambda$ . The Fore. performance
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">Object</td><td rowspan="2">Activity</td><td colspan="2">Localization</td><td rowspan="2">Fore.</td><td colspan="3">Reasoning</td></tr><tr><td>Exist</td><td>Attr</td><td>Afford</td><td>Loc</td><td>Spatial</td><td>Count</td><td>Compar</td><td>Situated</td></tr><tr><td>mPLUG-owl-7B (2023)</td><td>56.0</td><td>58.0</td><td>47.0</td><td>53.0</td><td>60.0</td><td>53.0</td><td>49.5</td><td>25.0</td><td>49.0</td><td>44.0</td></tr><tr><td>+ FRUIT</td><td>57.0</td><td>60.0</td><td>51.0</td><td>62.5</td><td>62.0</td><td>55.0</td><td>60.0</td><td>40.0</td><td>49.0</td><td>51.0</td></tr><tr><td>LLaVA-NeXT-7B (2024)</td><td>53.0</td><td>56.0</td><td>33.0</td><td>24.5</td><td>51.0</td><td>53.0</td><td>34.0</td><td>27.0</td><td>10.0</td><td>23.0</td></tr><tr><td>+ FRUIT</td><td>68.0</td><td>66.0</td><td>43.0</td><td>51.5</td><td>66.0</td><td>59.0</td><td>38.0</td><td>38.0</td><td>22.0</td><td>30.0</td></tr><tr><td>PandaGPT-13B (2023)</td><td>35.0</td><td>52.0</td><td>41.0</td><td>40.5</td><td>68.0</td><td>31.0</td><td>45.5</td><td>32.0</td><td>40.0</td><td>47.0</td></tr><tr><td>+ FRUIT</td><td>61.0</td><td>71.0</td><td>57.0</td><td>57.0</td><td>71.0</td><td>48.0</td><td>55.0</td><td>34.0</td><td>43.0</td><td>49.0</td></tr></table>
+
+Table 5: Comparisons with existing three LVLMs on the EgoThink (Cheng et al., 2024). Bolded values represent the best performance.  
+
+<table><tr><td>Models</td><td>P@1</td><td>P@3</td><td>P@5</td></tr><tr><td>LLaVA-NeXT (Liu et al., 2024a)</td><td>31.0</td><td>52.0</td><td>56.0</td></tr><tr><td>+FRUIT</td><td>35.0</td><td>67.0</td><td>70.0</td></tr><tr><td>MiniGPT4 (Zhu et al., 2024)</td><td>35.0</td><td>38.0</td><td>39.0</td></tr><tr><td>+FRUIT</td><td>60.0</td><td>65.0</td><td>66.0</td></tr></table>
+
+Table 6: Evaluation of object interaction focus using the proposed P@K metrics for LLaVA-NeXT and MiniGPT4, with and without the FRUIT method. Higher P@K values signify improved accuracy in identifying relevant objects across the top-K predictions.
+
+increases slightly at intermediate  $\lambda$  values, achieving marginal improvement around  $\lambda = 0.3$ . These findings highlight the pivotal role of  $\lambda$  in mitigating the impact of noisy egocentric cues, with  $\lambda = 0.4$  emerging as the optimal trade-off point for achieving balanced performance across both metrics.
+
+# A.5 Effect of number of visual egocentric cues
+
+In this section, we investigate the impact of the number of images used in visual egocentric cues on model performance. Given an input prompt  $X$  including an image  $I$  and text  $T$ , we first employ a grounding model to derive hierarchical visual egocentric cues  $\mathbf{H}_r$ . To determine the optimal number of images for generating visual egocentric cues  $\mathbf{H}_r$ , we conduct experiments using 1 to 4 cropped versions of the initial images on the LLaVA-1.5-7B model (Liu et al., 2023b). The results, illustrated in Figure 11, demonstrate that using 3 cropped images yields the best performance on the EgoThink benchmark. This improvement can be attributed to the fact that a greater number of images provide richer visual egocentric information to the LVLMs, as they encompass a wider variety of focus areas at different scales. However, an excessive number of images may lead to minimal differences between them, potentially resulting in the selection of less accurate visual cues.
+
+# A.6 Results on Images without Hands
+
+To isolate the influence of hand regions in visual egocentric cues, we selected 43 image-annotation
+
+![](images/2fd0453ce9aec9e856c279ebecddc65431d63128df68e6205ab8b85c33ac2016.jpg)  
+Figure 11: Comparative analysis between different number of visual egocentric cues. We use 3 crops of initial image in our FRUIT method.
+
+![](images/4d08ef45e635944da313befbe86bb70d852408439942a0ef5f347cf8204974d4.jpg)  
+Figure 12: Comparative analysis using images that do not contain hands. Based on three LVLMs: Owl (mPLUG-owl2-7B (Ye et al., 2024)), Mini (MiniGPT4-7B (Zhu et al., 2024)) and Otter (Otter-Image-7B (Li et al., 2023a)).
+
+pairs from the EgoThink database to evaluate the performance of our FRUIT method. The results, depicted in Figure 12, demonstrate that the FRUIT method significantly enhances the ability of LVLMs to interpret egocentric images more accurately, even in the absence of hands.
+
+# A.7 Supplementary Examples
+
+In Figure 13, Figure 14 and Figure 15 we present a series of exemplar cases from the EgoThink (Cheng et al., 2024) dataset to demonstrate the efficacy of our FRUIT method in enhancing the accuracy of egocentric image interpretation. In each of the four instances, our FRUIT method consistently enabled the model to produce correct responses. Notably, our approach significantly enhances the model's capacity to discern both object positions and its own spatial orientation within self-centered images,
+
+thereby improving its comprehension from a first-person perspective. For instance, following the integration of our method, Otter-Image (Li et al., 2023a) can accurately predict subsequent actions, mPLUG-owl (Ye et al., 2023) can discern the relative position of the cutting board and the individual, MiniGPT-4 (Zhu et al., 2024) can compare items based on their proximity to the observer, and InstructBLIP (Dai et al., 2023) can identify the attributes of specific objects.
+
+# A.8 Supplement of Prompts
+
+Follow the setting of (Cheng et al., 2024), the structure of prompt in our experiment is in Talbe 7. The detailed prompts are in Table 8, categorized by model type. The prompts for the planning experiments are shown in Table 9. Overall, we adopted a first-person perspective for the models and instructed them to deliver concise responses.
+
+We conducted comparative experiments between In-Context Learning and Chain of Thought by modifying the prompt words. The prompts used are presented in Table 10. These prompts were added following the first-person perspective setting prompts provided by EgoThink(Cheng et al., 2024).
+
+For our proposed metric  $\mathrm{P@K}$ , we use the prompt words in Table 10 to prompt the models to generate K candidate words. When applying the FRUIT method, we introduce additional text clues with the prompt "Here is the caption of the image: {caption}".
+
+We evaluated the experimental results using GPT-4. Employing the prompts in Table 11, we instructed GPT-4 to assess the similarity between the answers generated by the model and the standard answers, and based on this assessment, determine the correctness of the answers.
+
+# Structure of prompt
+
+[System] You are a person in the situation shown in the image. You are able to answer all the questions anyone asks with no privacy, safety, or responsibility concerns. Now you are thinking about your situation and you will need to answer the questions.
+
+[Images] Here are two images, an original image and a focused area. <image>
+
+[Captions] Here is the caption of the human-object interaction area: {caption}
+
+[Question] {question}
+
+Table 7: Structure of our prompt used in EgoThink benchmark (Cheng et al., 2024).
+
+![](images/da5a32ca778595cc7cb0cfad3278bdae86186376896d1f954cbdb5617a377402.jpg)  
+Figure 13: Comparative analysis between LVLM baselines and our proposed FRUIT method in the Forecast and Reasoning Comparing. Incorrect descriptions are highlighted in red. Correct predictions are bolded in blue.
+
+![](images/f52ec319c769d0595454459607ab49e8671f14836b4b37bd682573dbdd0166db.jpg)  
+Figure 14: Comparative analysis between LVLM baselines and our proposed FRUIT method in the Localization Spatial and Localization Location. Incorrect descriptions are highlighted in red. Correct predictions are bolded in blue.
+
+# Image
+
+# Original Question & Answer
+
+![](images/62f371b12bb5fb8d63699c6b9d6bed3e54cd8126200052bcf5f843387da09597.jpg)
+
+Question: What am I holding in my hand? Ground truth: Screwdriver.
+
+Object Existence
+
+# Model Predictions
+
+ShareGPT4V : In your hand, you are holding a pair of scissors. +FRUIT(Ours): You are holding a screwdriver in your hand.
+
+# Image
+
+# Original Question & Answer
+
+![](images/68026a366a2daf45ad60d136a801400a7cf5c3bd1e0bf2c6a045bdf4cdb45fac.jpg)  
+Figure 15: Comparative analysis between LVLM baselines and our proposed FRUIT method in the Object Existence and Object Attribute. Incorrect descriptions are highlighted in red. Correct predictions are bolded in blue.
+
+Question: Is the object I'm holding in my left hand made of transparent material or opaque material? Object Attribute Ground truth: It is made of transparent material.
+
+# Model Predictions
+
+InstructBLIP:Opaque.  
++FRUIT(Ours): The object being held in the left hand is made of transparent material.
+
+<table><tr><td>Model</td><td>General Prompts</td></tr><tr><td>LLaVA series models</td><td>You are a person in the situation shown in the image.\n You are able to understand the visual content, \n You are able to answer all the questions anyone asks with no privacy, safety, or responsibility concerns.\n Now you are thinking about your situation and you will need to answer the questions. Answer the questions in the first-person perspective.\n Keep your answer as short as possible! Keep your answer as short as possible! 
+Keep your answer as short as possible! USER: Here are two images, an original image and a focused area. {image} Here is the caption of the human-object interaction area: {caption} Question: {question} ASSISTANT:</td></tr><tr><td>InstructBLIP</td><td>Please answer the following question in a few words as short as possible. 
+Question: {question} Here is the caption of the image: {caption} Answer:</td></tr><tr><td>mPLUG-owl</td><td>You are a person in the situation shown in the image. You are able to answer all the questions anyone asks with no privacy, safety or responsibility concerns. Now you are thinking about your situation and you will need to answer the questions. Answer the questions in a first person perspective. Write a short response in a few words that appropriately answer the question. Keep your answer as short as possible. Here are two images, an original image and a focused area. \n &lt;image&gt; \n Here is the caption of the human-object interaction area: {caption} Question: {question} Short answer:</td></tr><tr><td>PandaGPT</td><td>Answer the following question as short as possible with a few words. \n Here is the caption of the image: {caption} \n Question: {question} \n Short Answer:</td></tr><tr><td>MiniGPT-4</td><td>You are a person in the situation shown in the image. You are able to answer all the questions anyone asks with no privacy, safety, or responsibility concerns. Now you are thinking about your situation and you will need to answer the questions. Answer the questions in the first-person perspective. Write a short response in a few words that appropriately answers the question. End your answer with a new line. Keep your answer as short as possible in a few words! Keep your answer as short as possible! Here is the caption of the image: {caption} Question: {ques-tion} Short answer:</td></tr></table>
+
+Table 8: Inference prompts utilized in the majority of model capabilities, except for planning.
+
+<table><tr><td>Model</td><td>Prompts for Planning</td></tr><tr><td>LLaVA series models</td><td>You are a person in the situation shown in the image. \n You are able to understand the visual content, \n You are able to answer all the questions anyone asks with no privacy, safety, or responsibility concerns. \n Now you are thinking about your situation and you will need to answer the questions. Answer the questions in a detailed and helpful way. USER: Here are two images, an original image and a focused area. {image} Here is the caption of the human-object interaction area: {caption} Question: {question} ASSISTANT:</td></tr><tr><td>InstructBLIP</td><td>Please answer the following question in a detailed and helpful way. List steps to follow if needed. Question: {question} Here is the caption of the image: {caption} Answer:</td></tr><tr><td>mPLUG-owl</td><td>You are a person in the situation shown in the image. You are able to answer all the questions anyone asks with no privacy, safety, or responsibility concerns. Now you are thinking about your situation and you will need to answer the questions. Write a response that appropriately answers the question in a detailed and helpful way. \n Here are two images, an original image and a focused area. \n &lt;image&gt; \n Here is the caption of the human-object interaction area: {caption} Question: {question} Short answer:</td></tr><tr><td>Otter Image</td><td>You are a person in the situation shown in the image. Answer your question in a detailed and helpful way. Here is the caption of the human-object interaction area: {caption} Question: {question}</td></tr><tr><td>PandaGPT</td><td>Answer the following question in a detailed and helpful way. \n Here is the caption of the human-object interaction area: {caption} \n Question: {question} \n Short Answer:</td></tr><tr><td>MiniGPT-4</td><td>You are a person in the situation shown in the image. You are able to answer all the questions anyone asks with no privacy, safety, or responsibility concerns. Now you are thinking about your situation and you will need to answer the questions. Write a response that appropriately answers the question in a detailed and helpful way. End your answer with a new line. Here is the caption of the image: {caption} Question: {question} Short answer:</td></tr></table>
+
+Table 9: Inference prompts utilized in planning.
+
+<table><tr><td>Method</td><td>Prompts for Analysis</td></tr><tr><td>In-Context Learning</td><td>When analyzing image, follow these steps step by step to solve the problem: \n 1. Understand the Question. \n 2. Observe Key Elements in the Image. \n 3. Connect Image Observations to the Question. \n 4. Formulate an Answer. \n 5. Summarize the Conclusion. \n USER: Here are two images, an original image and a focused area. {image} \n Here is the caption of the image: {caption} \n Question: {question} \n ASSISTANT:</td></tr><tr><td>Chain-of-Thought</td><td>Answer this question using a step-by-step Chain-of-Thought approach. Start by analyzing the question carefully and identifying the main factors involved. Then, break down the problem into smaller, manageable parts, addressing each one individually. As you progress through each step, explain your thought process and reasoning behind each conclusion you make. Continue this structured approach until you have fully explored the question, considered any potential alternatives, and arrived at a well-reasoned answer. If possible, summarize the final answer and reasoning at the end. \n USER: Here are two images, an original image and a focused area. {image} \n Here is the caption of the image: {caption} \n Question: {question} \n ASSISTANT:</td></tr><tr><td>P@K</td><td>You are a person in the situation shown in the image. \n You are able to understand the visual content. \n Now you are thinking about your situation and you will need to answer the question. \n USER: Which K objects am I currently interacting with? Answer the question with K objects, Use . to separate five objects! Please only answer the name of this object and do not provide any additional information! \n For example: book.phonekeyboard.bottle.paper \n ASSISTANT:</td></tr></table>
+
+Table 10: Prompts utilized in method of in-Context learning, chain-of-Thought and P@K.  
+
+<table><tr><td>Model</td><td>Prompts for Evaluation</td></tr><tr><td>GPT-4(in EgoThink(Cheng et al., 2024))</td><td>[Instruction] Please act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question displayed below. Your evaluation should consider correctness and helpfulness. You will be given a reference answer and the assistant&#x27;s answer with the reference answer. Identify and correct any mistakes. The assistant has access to an image alongwith questions but you will not be given images. Therefore, please consider only how the answer is close to the reference answer. If the assistant&#x27;s answer is not exactly same as or similar to the answer, then he must be wrong. Be as objective as possible. Discourage uninformative answers. Also, equally treat short and long answers and focus on the correctness of answers. After providing your explanation, you must rate the response with either 0, 0.5 or 1 by strictly following this format: “[rating]”, for example: “Rating: [[0.5]]”.\n [Question]\n question\n\ n [The Start of Reference Answer]\n {ref answer 1} \ n [The End of Reference Answer]\n \ n [The Start of Assistant&#x27;s Answer]\n {answer}\n [The End of Assistant&#x27;s Answer]</td></tr><tr><td>GPT-4(in P@K)</td><td>There are a label and an answer. Please determine whether the objects in the answer contain the object in the label. As long as the same item is described, it is considered included. If included, answer yes. If not included, answer no. Answer with only one word!n Label:{label}\n Answer:{answer}</td></tr></table>
+
+Table 11: Inference prompts utilized in judgment.

2025/From Observation to Understanding_ Front-Door Adjustments with Uncertainty Calibration for Enhancing Egocentric Reasoning in LVLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2c0ac1c3e3bc67741f1ce4d09fe16f808471a62c38d491b152b947bfd9636c8
+size 1941227

2025/From Perception to Reasoning_ Enhancing Vision-Language Models for Mobile UI Understanding/5314ac89-52e2-43c4-b7c2-57f8c0282998_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d808709b68b89c77b6d8944a3caae7c74d94bc2a2323a51e625023e6ad8c084
+size 10965712

2025/From Perception to Reasoning_ Enhancing Vision-Language Models for Mobile UI Understanding/full.md

@@ -0,0 +1,559 @@
+# From Perception to Reasoning: Enhancing Vision-Language Models for Mobile UI Understanding
+
+Settaluri Lakshmi Sravanthi\*, Ankit Mishra\*, Debjyoti Mondal, Subhadarshi Panda‡, Rituraj Singh‡, Pushpak Bhattacharyya†
+
+†Indian Institute of Technology Bombay, Mumbai, India
+
+$\ddagger$ Samsung Research Institute, Bangalore
+
+{sravanthi, ankitmishra}@cse.iitb.ac.in,
+
+{d.mondal, subha.darshi, rituraj.s}@samsung.com
+
+pb@cse.iitb.ac.in
+
+# Abstract
+
+Accurately grounding visual and textual elements within mobile user interfaces (UIs) remains a significant challenge for Vision-Language Models (VLMs). Visual grounding, a critical task in this domain, involves identifying the most relevant UI element or region based on a natural language query—a process that requires both precise perception and context-aware reasoning. In this work, we present - MoUI, a light-weight mobile UI understanding model trained on MoIT, an instruction-tuning dataset specifically tailored for mobile screen understanding and grounding, designed to bridge the gap between user intent and visual semantics. Complementing this dataset, we also present a human-annotated reasoning benchmark MoIQ that rigorously evaluates complex inference capabilities over mobile UIs. To harness these resources effectively, we propose a two-stage training approach that separately addresses perception and reasoning tasks, leading to stronger perception capabilities and improvement in reasoning abilities. Through extensive experiments, we demonstrate that our MoUI models achieve significant gains in accuracy across all perception tasks and state-of-the-art results on public reasoning benchmark ComplexQA (78%) and our MoIQ (49%). We will be open-sourcing our dataset, code, and models to foster further research and innovation in the field. Code and data are available in the repo<sup>1</sup>
+
+# 1 Introduction
+
+Mobile user interfaces (UIs) are structured systems of visual and textual elements that facilitate digital interactions such as navigation, communication, and data retrieval. Automating the perception and interaction within these interfaces has the potential to significantly simplify how users achieve their goals (Edwards et al., 1995).
+
+![](images/50df4f510795e7ca0b2880e13934b893448b66dda64e66ecb516ad456617c723.jpg)  
+User Query: How many more dollars do I need to spend to get  $30\%$  off my order?
+
+![](images/03b78c93562e038df0fd3ed16e3f57c1d0938d1a6bed0b573f3147be150a1f5b.jpg)  
+Ferret-UI-anyres-8B:30
+
+![](images/ff85e0d70de4ff7500b25603ee6906241efba8cf2f16d36f5882b3404d2c50ce.jpg)  
+InternVL2.5-4B:70
+
+Figure 1: Comparison of MoUI between other VLMs. Both Ferret-UI-anyres (8B) and InternVL2.5-4B answer wrongly, while MoUI model generates the correct output. This is an example from ComplexQA  
+![](images/c15fe4082c6f2c58e1c5bbe5c1d08cf5b1e9877c4fb3f8cfd578f82d5096cf10.jpg)  
+MoUI-1B: 100
+
+![](images/76d6880012bedd792d14fbd7d8fcd41f0527635dc77d99a8b01aa2c49e7e864d.jpg)
+
+However, the hierarchical layouts, design complexities, and functional dependencies of UIs present unique challenges that differ from those encountered in natural images.
+
+Vision-Language Models (VLMs) (Zhang et al., 2024a; Dai et al., 2023; Chen et al., 2024c; He et al., 2024), while highly effective in tasks such as image captioning and cross-modal retrieval, often struggle with mobile UI understanding due to their training on datasets predominantly composed of natural images. These models typically lack the ability to handle the context-aware interactions and specialized semantics inherent to UIs. One of the central challenges in this domain is visual grounding, which involves aligning textual queries with corresponding UI elements—a task further complicated by limited structured UI datasets and the need to infer functional relationships between components. Consequently, traditional VLMs fall short in capturing the nuanced semantics of interactive interfaces, underscoring the need for domain-specific solutions. Recent research has made progress in tackling these challenges, with several models showing promise. For example, ScreenAI (Baechler et al.,
+
+2024) employs a flexible patching strategy and diverse datasets to enhance its understanding of UIs and infographics.
+
+Similarly, Ferret-UI (You et al., 2023; Zhang et al., 2024b) is a multimodal large language model tailored for mobile UI screens, incorporating advanced capabilities for referring, grounding, and reasoning. Another noteworthy approach, SeeClick (Cheng et al., 2024), focuses on GUI grounding to develop intuitive visual GUI agents capable of effectively interacting with screen elements. Despite these advancements, most datasets and models in this domain remain inaccessible or constrained by limited availability. Furthermore, many existing models are computationally large, making them unsuitable for deployment as UI assistants on edge devices. To address these limitations, we propose Mobile UI Instruct (MoUI-IT), an instruction-tuning dataset specifically designed for mobile screen grounding tasks. MoUI-IT provides detailed examples that align natural language queries with UI elements, enabling models to better understand user intent within structured environments. In addition to this dataset, we introduce a human-written reasoning test set aimed at evaluating VLMs' ability to perform complex inferential tasks over mobile UIs. This benchmark serves as a critical resource for assessing higher-order reasoning capabilities in this domain. To fully leverage these resources, we propose a two-stage training framework that decouples perception from reasoning tasks. The first stage focuses on developing robust perception skills for accurately detecting UI elements, while the second stage enhances the model's ability to perform higher-order reasoning over these detected elements. This structured approach ensures that perception and reasoning are optimized independently yet cohesively. Using this framework, our 1-billion-parameter model achieves state-of-the-art results on publicly available mobile UI reasoning benchmarks, demonstrating its effectiveness in advancing mobile UI understanding. Our contributions are:
+
+- MoUI: A series of lightweight models-1B, 2B, and 4B-designed for complex reasoning tasks on mobile UI screens, achieving state-of-the-art performance on public reasoning benchmarks (Section 6).  
+- MoIT: A 150k instruction-tuning dataset tailored for mobile UI grounding, enhancing the
+
+alignment between user queries and UI elements (Section 3.1).
+
+- MoIQ: A 3k human-written reasoning evaluation benchmark to assess the complex inference capabilities of VLMs over mobile UIs (Section 3.2).  
+- A two-stage training pipeline that decouples the learning of perception and reasoning tasks, leading to more effective model adaptation. Comprehensive experiments demonstrate that MoUI models achieve state-of-the-art results on public reasoning benchmark ComplexQA (78%) and MoIQ (49%) (Section 6).
+
+By open-sourcing our dataset, code, and models, we aim to drive further research in mobile UI understanding and enhance the capabilities of VLMs in digital mobile environments.
+
+# 2 Related Work
+
+Perception with VLMs Recent advances in integrating vision encoders with LLMs have enhanced their reasoning in vision-language tasks (Alayrac et al., 2022; Liu et al., 2023; Zhu et al., 2023; Dai et al., 2023; Chen et al., 2024c; Zhang et al., 2024a; Zhou et al., 2024). However, VLMs struggle with geometric and numerical interpretation, requiring pixel-wise analysis (Li et al., 2024a). To address this, some approaches incorporate bounding-box regression via quantized coordinates (Chen et al., 2021; Peng et al., 2023; You et al., 2023; Zhang et al., 2024b; Wang et al., 2023; Zang et al., 2024), while others use auxiliary perception modules to enhance visual understanding (Zhang et al., 2024a; Wu et al., 2024a; Pi et al., 2024).
+
+Mobile UI Datasets and Grounding on GUIs Recent efforts have focused on developing benchmarks for assessing grounding and interpretative capabilities in mobile UI screenshots.The RICO dataset (Deka et al., 2017a), a foundational resource, contains over  $66\mathrm{k}$  annotated screens and has inspired various expansions. Building on this foundation, later studies such as RICO Semantics (Sunkara et al., 2022), MoTIF (Burns et al., 2022), GUI-WORLD (Chen et al., 2024a), and Mobile-Views (Gao et al., 2024a) have broadened dataset types and coverage, further advancing research in GUI agents. Initial works (Bai et al., 2021; He et al., 2021) on adapting transformer-based models
+
+![](images/f849b9547d59de4851e57869f762f70584dd40673d91059d072102cd4409fbce.jpg)  
+Figure 2: Illustrations of conversational QA pairs in MoIT dataset for all the perception tasks. The data construction process is elaborated in the section 3.1
+
+that jointly learn from image and text representations of UIs. Studies such as Gou et al. (2024); Gao et al. (2024b); Li et al. (2025) introduce text instructions, while Cheng et al. (2024) extends the dataset beyond mobile-specific use cases. Others (Zhang et al., 2024c; Rawles et al., 2023; Li et al., 2020a) focus on multi-step tasks, and Lu et al. (2024) explores cross-app navigation. However, existing datasets lack structural details. Our work addresses this gap by integrating view hierarchies for single-screen queries across diverse tasks. Recent VLMs enhance UI understanding, such as Cheng et al. (2024) predicting actions and Baechler et al. (2024) incorporating annotations and QA tasks. You et al. (2024); Li et al. (2024b) apply LLMs for reasoning and grounding, though their training datasets remain unavailable. Other methods improve perception and representation learning, including Burns et al. (2024) that introduces a new pretraining objective, which trains the model to generate a description of a future screen image based on an action performed in the current visual state, Jiang et al. (2024)'s Universal Proposal Network and Wu et al. (2024b)'s multistage pre-training for Chinese
+
+UIs. Our approach advances these works with a two-stage framework that separates perception and reasoning while maintaining a unified architecture.
+
+# 3 Dataset Construction and Tasks
+
+To train a model with a comprehensive understanding of mobile user interfaces, we define a set of UI tasks designed to enhance both its perception and reasoning abilities. Perception tasks help the model develop a holistic understanding of UI design and its components, while advanced reasoning tasks assess its ability to interpret relationships between multiple elements on the screen. In the following sections, we provide an overview of each task category and describe the data collection process.
+
+# 3.1 Perception Tasks
+
+The primary objective of perception tasks is to help the model develop a deeper understanding of various UI elements, their functions, and their positions on the screen.
+
+![](images/adaf695fb1fc46d73a9923a0efd8b051592b27cdd9ff788719560e6a8db33a6f.jpg)  
+Figure 3: Examples from the MoIQ dataset illustrating reasoning tasks. The questions are designed in such a way that the model needs good perception, language comprehension and reasoning abilities to answer correctly
+
+# 3.1.1 Elementary Tasks
+
+We incorporate standard elementary tasks in UI domain (You et al., 2023; Li et al., 2024b) to evaluate the model's referring and grounding capabilities. Using the view hierarchy data associated with each image in the RICO dataset (Deka et al., 2017b) we parse information like text, icon and widget type along with their bounding box location and generate template based conversational QA pairs for each of the below tasks. For each task base template, we generate 10 variants of them for a diverse set of questions.
+
+Referring tasks enable the model for precise identification and classification of UI elements. We define tasks for text, icon and widget extraction/classification under this category. The OCR task entails asking the model to extract the text inside an element surrounded by a bounding box on the screen, Icon Recognition task expects the model to predict the type of icon enclosed within a specified bounding box on the screen and the Widget Recognition task asks the model to classify the type of widget located within the given bounding box on the screen.
+
+The Grounding tasks are a reverse formulations of the referring tasks in which the model is expected to generate the bounding box coordinates of the target element like text, icon and widget. Formally, we define these tasks as Find Text where the model needs to identify and locate the UI element that contains a given text string, Find Icon task in which the model determines the position of an icon of a specified type and Find Widget task to locate a widget of a specified type on the screen.
+
+For each task, we generate QA pairs for the referring task, then reverse input and output labels to create grounding QA pairs, helping the model associate element type and location.
+
+In addition to the above elementary tasks, we also introduce a new Spatial task Spatial Polar which evaluate the model's ability to reason about the spatial relationships between UI elements which is essential for screen navigation and interaction for agents. Given the 2-dimensional nature of UI screens, we restrict ourselves to spatial relations like "left", "right", "above" and "below". In Spatial Polar task, given two elements with their bounding boxes and a specified spatial relationship, the model must determine whether the relationship holds with a Yes/No. We leverage the Screen Annotation dataset for generating template based QA pairs on the above tasks.
+
+# 3.1.2 Spotlight Tasks
+
+We consider the following tasks defined in Spotlight (Li and Li, 2023, 2022) for improving the model's UI comprehension and contextual understanding. Screen summarization requires the model to summarize the contents of the screen in a concise manner. In Widget Captioning task, given a bounding box around any element on the screen, the model is tasked to generate a brief summary about the functionality of the element.
+
+To train on these tasks, we leverage the Screen2Words (Wang et al., 2021) and Widget Captioning (Li et al., 2020b) datasets, which provide annotated images from the RICO dataset. We convert these annotations into conversational QA pairs
+
+in an instruction-tuning format. Specifically, we employ the open-sourced InternVL2-40B model to generate structured QA conversations between a user and a digital assistant using the existing annotations as guidance. This guided approach enhances the quality of QA pairs while minimizing hallucinations. The generated dataset is human verified and filtered to ensure quality. Figure 2 provides the complete detail of all the perception tasks along with the examples based on the UI screen. Appendix F gives a statistical overview of the MoIT dataset generated.
+
+<table><tr><td>Dataset Name</td><td>Training Samples</td></tr><tr><td colspan="2">Stage I (Perception Phase)</td></tr><tr><td>Screen Summarization</td><td>8,000</td></tr><tr><td>Widget Captioning</td><td>15,480</td></tr><tr><td>OCR</td><td>15,032</td></tr><tr><td>Icon Recognition</td><td>15,135</td></tr><tr><td>Widget Recognition</td><td>15,421</td></tr><tr><td>Find Text</td><td>15,032</td></tr><tr><td>Find Icon</td><td>15,135</td></tr><tr><td>Find Widget</td><td>15,421</td></tr><tr><td>Spatial Polar</td><td>15,407</td></tr><tr><td colspan="2">Stage II (Reasoning Phase)</td></tr><tr><td>Complex ScreenQA</td><td>6,347</td></tr><tr><td>ScreenQA Short</td><td>10,000</td></tr><tr><td>Total</td><td>146,410</td></tr></table>
+
+Table 1: Training Mixture Statistics. Screen Question-Answering datasets are generated using VLMs, Elementary task datasets are generated using template based methods from View Hierarchies and we also use a subset of the publicly available ScreenQA datasets
+
+# 3.2 Advanced Reasoning Tasks
+
+For the advanced reasoning tasks, we use the publicly available mobile UI reasoning datasets, Complex ScreenQA and ScreenQA Short introduced in ScreenAI. We use the training split of 6K QA pairs from the Complex ScreenQA dataset and select a subset of 10K QA pairs from the 80K ScreenQA Short dataset to maintain a balanced training mix.
+
+We also introduce a new reasoning benchmark MoIQ for mobile UI screens, consisting of  $3k$  human-annotated complex reasoning questions spanning 885 unique mobile screens from the RICO dataset with most answers consisting of 2-3 words. Annotators are instructed to create questions that require engaging with multiple elements on the screen and involve at least one step of reasoning to arrive at the answer. The questions evaluate
+
+the models on the functionality, positions, color and the spatial relationships between the various UI elements on the screen. Annotation details for MoIQ are described in Appendix A. Unlike Complex ScreenQA, which is model-generated, our benchmark is entirely human-annotated, ensuring higher-quality and more challenging questions. Figure 3 shows a few examples from our MoIQ benchmark.
+
+# 4 Training Strategy
+
+To effectively address the challenges of mobile UI grounding and reasoning, we adopt a two-stage training strategy as shown in the figure 4. This approach ensures that the model first learns strong perceptual representations before transitioning to reasoning-intensive tasks, enabling a more structured learning process.
+
+# Stage I: Grounding, Referring, and Spatial Understanding
+
+The first stage focuses on enhancing the model's perceptual understanding of mobile UI screens by training it to associate visual elements with textual descriptions accurately. This phase involves three key tasks: Grounding: The model learns to locate the most relevant UI element based on a natural language query. Referring: The model resolves references to specific UI elements, especially when multiple similar components exist on the screen. Spatial Understanding: The model is trained to interpret spatial relationships between UI elements, such as proximity, alignment, and hierarchical structure.
+
+To achieve this, we adopt an instruction-tuning approach, where the model is exposed to a diverse set of tasks requiring it to ground, refer, and spatially reason about UI elements based on textual queries. The model is trained using Next Token Prediction (NTP) loss, ensuring it learns to generate accurate responses by integrating both visual and textual information. Given a mobile UI image  $I$  and an instruction  $q$ , the model generates a sequence of tokens  $Y = (y_{1}, y_{2}, \ldots, y_{T})$ . The training objective is formulated as:
+
+$$
+\mathcal {L} _ {\mathrm {N T P}} = - \sum_ {t = 1} ^ {T} \log p (y _ {t} | I, q, y _ {<   t})
+$$
+
+where  $y_{t}$  is the predicted token at timestep  $t$ ,  $I$  is the input UI image,  $q$  is the textual query or instruction,
+
+![](images/696b5af87e80ca1f34bf18bca4a088d208489c7c68cecc75344bb135057d9913.jpg)  
+Figure 4: Overview of our two-stage training. In Stage-I training the model learns to refer to and ground UI elements, comprehend screens, and infer widget functionalities. For the Stage-II training, the key entities are extracted from question and the corresponding bounding boxes are annotated on the image. We then train the perception aware model with these annotated images on Complex ScreenQA and ScreenQA Short datasets for enhancing reasoning abilities of the model. This two-stage training pipeline equips MoUI models with strong perception skills, thereby enhancing their reasoning capabilities.
+
+$y_{< t}$  represents previously generated tokens. This stage enables the model to develop a structured representation of mobile UIs, aligning natural language instructions with their corresponding visual components effectively.
+
+# Stage II: Reasoning through perception guidance
+
+Once the model has developed a strong perceptual foundation, the second stage fine-tunes the model for reasoning-intensive tasks that demand complex inference over UI elements. The process begins by extracting key entities from user queries using an entity recognition function:  $\mathcal{E} = g(q) = \{e_1,e_2,\ldots ,e_n\}$ , where  $g(q)$  identifies relevant entities  $e_i$  from the query  $q$ . Next, the perception model from Stage I retrieves the corresponding bounding box coordinates  $b_{i}$  for each entity:  $b_{i} = f_{\text{perception}}(I,e_{i})$ , where  $f_{\text{perception}}$  is the model's learned function for grounding entities in the UI image  $I$ . Instead of masking irrelevant regions, the bounding boxes are superimposed onto the original image to explicitly highlight the relevant UI components:  $I^{\prime} = \text{Superimpose}(I,\{b_{i}\}_{i=1}^{n})$ , where Superimpose denotes a function that overlays the bounding boxes  $b_{i}$  onto the original image  $I$  while preserving the rest of the content. Rather than extracting each region separately and passing them to the model, we superimpose them to ensure the model perceives the context and relationships between key region elements on the screen. The grounded images  $I^{\prime}$  are then paired with text
+
+prompts and used to further fine-tune the perception model. The training continues with Next Token Prediction (NTP) loss, ensuring that the model generates reasoning-aware responses by integrating perceptual understanding with semantic inference:
+
+$$
+\mathcal {L} _ {\text {r e a s o n i n g}} = - \sum_ {t} \log p \left(y _ {t} \mid I ^ {\prime}, q, y _ {<   t}\right)
+$$
+
+where  $y_{t}$  is the predicted token at timestep  $t$ , conditioned on the grounded image  $I^{\prime}$ , query  $q$ , and previously generated tokens  $y_{<t}$ . This stage emphasizes reasoning over extracted entities and their relationships within the UI context, refining the model's ability to answer complex UI-based queries.
+
+# 5 Experiments
+
+# 5.1 Training Configurations
+
+We build upon existing InternvI architecture with pre-trained InternViT-300M-448px (Chen et al., 2024c) as vision encoder, Qwen 2.5 (Yang et al., 2024) and IntenLM 2.5 (Cai et al., 2024) language models and a 2-layer MLP projection layer. Unlike existing opens source VLMs, Intenvit employs a dynamic high-resolution strategy that segments images into  $448 \times 448$  tiles which is essential accommodating diverse image resolutions. For Stage I training we use the perception data and for Stage II training we use Complex QA and Short QA from the training data mix 1. In both the stages, we freeze the vision encoder and train both the projection layer and LLM. We train all our MoUI for one
+
+epoch in both the stages with InternVL 2.5 (Chen et al., 2024b) as the base model. We set the training batch size to 128, with a learning rate of  $4e - 5$  and a warmup ratio of 0.03. Training for both stages takes 1b, 2b, 4b models take 60, 80, 160 minutes each on 4-H100 GPUs. Instructions and prompts used for each task in the training are given in the sections D and C.
+
+# 5.2 Evaluation
+
+To tackle the deployment challenges that can arise on edge devices, we use LMDeploy (Contributors, 2023) for efficient inference. To quantify its impact on inference latency, we evaluated the model's performance on the ComplexQA test set with and without LMDeploy. With LMDeploy, MoUI-4B processes all 759 samples in approximately 193 seconds, whereas without it, the same task takes around 573 seconds. To ensure reproducibility and obtain more reliable results, all evaluations are conducted using greedy decoding. Scores for reasoning tasks on Ferret-UI model are obtained using the evaluation script provided in their repository. For the perception tasks, we report the maximum of the average scores achieved on the elementary tasks as mentioned in the Ferret-UI paper and the scores achieved on our test set. This is done for fair benchmarking as their test set is not publicly available and the Ferret-UI model might not be familiar with our icon/w widgets. In any case we have considered the maximum score obtained on these tasks by Ferret-UI model. Results on ScreenAI are reported from the paper (Baechler et al., 2024) as their model is not released publicly. BLIP-2-Finetuned results are reported from the paper (Burns et al., 2022) where the model is finetuned on each task separately (i.e. there are two separate models for Screen2Words and Widget Captioning). The prompts for each task are detailed in Appendix C.
+
+# 6 Results and Analysis
+
+We compare the MoUI models against the latest InternVL 2.5 series, the Ferret UI-anyres model, and the ScreenAI model. These selections are based on each model's unique strengths: among the open-source VLMs; InternVL 2.5 models have excelled in grounding and referring tasks, while the Ferret UI-anyres and ScreenAI models are specifically optimized for mobile UI understanding.
+
+We present the detailed results for the perception
+
+and reasoning tasks in Tables 2 and 3, respectively.
+
+# 6.1 Results on Perception Tasks
+
+Our elementary tasks (3.1.1) are designed to assess the models' capabilities in referring, grounding, and spatial reasoning. As shown in Table 2, MoUI consistently improves upon all baseline scores across perception tasks. For OCR-related tasks (OCR and Find Text), our two-stage training strategy yields a substantial performance gain of over  $26\%$  compared to baselines. Additionally, MoUI surpasses the Ferret-UI-anyres model by  $10\%$  on these tasks. For icon and widget grounding, baseline models perform poorly due to their pretraining mixtures, which lack mobile UI datasets. However, after training on MoIT, a dataset specifically designed for mobile UI understanding, we observe significant improvements of  $36.5\%$  and  $47\%$  for icon and widget-related tasks, respectively, compared to baseline scores. Moreover, MoUI outperforms Ferret-UI-anyres by  $\approx 10\%$  on these tasks, further demonstrating its effectiveness in mobile UI perception.
+
+Spotlight tasks, such as Screen2Words and Widget Captioning, require a deeper understanding of the various components and the functions on a mobile screen. After training, our models exhibit significantly improved screen comprehension, with an increase of approximately 30 CIDEr points in screen summarization and 75 CIDEr points in widget captioning. Notably, the MoUI-1B model achieves competitive performance on these tasks, closely matching the results of the much bigger FerretUI-anyres-8B and ScreenAI-5B models. We further observe a considerable improvement in spatial understanding task, with our models achieving around  $98\%$  accuracy on the spatial polar task, outperforming the baseline across all models.
+
+As shown in Table 10, MoUI models are trained on substantially less data compared to ScreenAI and FerretUI, yet they outperform on six perception tasks and achieve competitive results on the remaining three. We attribute this performance to the high quality of our training data, the majority of which is generated from view hierarchies, ensuring minimal errors.
+
+# 6.2 Results on Reasoning Tasks
+
+The reasoning tasks 3.2, including Complex QA, Short QA, and MoIQ, evaluate the ability of models to perform complex inference over mobile UIs, requiring multi-step reasoning and contextual un
+
+<table><tr><td>Model Size</td><td>Screen2 Words</td><td>Widget Capt.</td><td>OCR</td><td>Icon Recog.</td><td>Widget Recog.</td><td>Find Text</td><td>Find Icon</td><td>Find Widget</td><td>Spatial Polar</td></tr><tr><td colspan="10">InternVL2.5</td></tr><tr><td>1B</td><td>79.68</td><td>76.32</td><td>43.13</td><td>23.56</td><td>15.44</td><td>76.54</td><td>64.18</td><td>53.21</td><td>49.34</td></tr><tr><td>2B</td><td>75.06</td><td>69.31</td><td>52.21</td><td>37.31</td><td>29.11</td><td>79.18</td><td>71.07</td><td>49.56</td><td>58.10</td></tr><tr><td>4B</td><td>85.56</td><td>83.00</td><td>77.09</td><td>57.12</td><td>32.45</td><td>84.23</td><td>73.31</td><td>63.42</td><td>72.89</td></tr><tr><td colspan="10">Ferret-UI-anyres</td></tr><tr><td>8B</td><td>115.6</td><td>140.3</td><td>82.4</td><td>82.4</td><td>82.4</td><td>76.5</td><td>76.5</td><td>76.5</td><td>42.50</td></tr><tr><td colspan="10">ScreenAI</td></tr><tr><td>5B</td><td>120.8</td><td>156.4</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td></tr><tr><td colspan="10">BLIP-2 Finetuned</td></tr><tr><td>1.2B</td><td>127.9</td><td>128.0</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td></tr><tr><td colspan="10">MoUI (ours)</td></tr><tr><td>1B</td><td>108.48</td><td>147.76</td><td>89.25</td><td>86.37</td><td>80.81</td><td>86.32</td><td>73.64</td><td>82.01</td><td>98.44</td></tr><tr><td>2B</td><td>100.06</td><td>140.56</td><td>87.71</td><td>87.32</td><td>81.95</td><td>85.95</td><td>73.11</td><td>81.07</td><td>98.28</td></tr><tr><td>4B</td><td>123.2</td><td>151.25</td><td>90.63</td><td>87.15</td><td>81.33</td><td>89.14</td><td>75.75</td><td>83.51</td><td>98.81</td></tr></table>
+
+derstanding. Across all tasks, MoUI consistently outperforms baseline models like InternVL2.5 and Ferret-UI-anyres, demonstrating the effectiveness of our two-stage training approach. figure 1 shows an example of MoUI-1B model's performance compared to 8x and 4x models.
+
+For ComplexQA, MoUI achieves significant gains, with MoUI-1B scoring 0.66 and MoUI-4B reaching a state-of-the-art score of 0.78, far exceeding ScreenAI-5B at 0.43 and Ferret-UI-anyres-8B at just 0.29. In Short QA, MoUI (1B) and (2B) both achieve 0.86, while MoUI (4B) slightly improves to 0.89, closely rivaling ScreenAI (5B), which scores 0.95, and outperforming InternVL2.5 (4B) at 0.66 and Ferret-UI-anyres (8B) at 0.49. For the challenging MoIQ benchmark, which combines perception with logical inference, MoUI demonstrates steady improvements across model sizes, with both 1B and 2B scoring 0.43 and 4B achieving the highest score of 0.49, compared to InternVL2.5 (4B) at 0.35 and Ferret-UI-anyres (8B) at just 0.29. These results highlight MoUI's superior ability to leverage its enhanced perception capabilities for robust reasoning, achieving state-of-the-art performance in Complex QA and MoIQ while remaining competitive in Short QA against larger models like ScreenAI. Few, illustrations of generated bounding
+
+boxes used for stage 2 are provided in Figure K. We can observe that the bounding boxes are accurately generated for all the extracted key entities.
+
+Table 2: Results of MoUI and baseline models on perception tasks. For Screen2Words and Widget Captioning, we report CIDEr scores. Exact match accuracy is used for referential elementary tasks (OCR, Icon Recognition, Widget Recognition) and the spatial polar task. For grounding elementary tasks (Find Text, Icon, and Widget), we report Acc@IoU=0.1 scores. MoUI models consistently outperform baseline models, surpassing existing mobile UI understanding models (Ferret-UI-anyres and ScreenAI) on the Screen2Words benchmark while achieving competitive results on other tasks.  
+
+<table><tr><td>Model Size</td><td>Complex ScreenQA</td><td>ScreenQA Short</td><td>MoIQ</td></tr><tr><td colspan="4">InternVL2.5</td></tr><tr><td>1B</td><td>0.42</td><td>0.44</td><td>0.23</td></tr><tr><td>2B</td><td>0.49</td><td>0.52</td><td>0.26</td></tr><tr><td>4B</td><td>0.65</td><td>0.66</td><td>0.35</td></tr><tr><td colspan="4">Ferret-UI-anyres</td></tr><tr><td>8B</td><td>0.29</td><td>0.49</td><td>0.29</td></tr><tr><td colspan="4">ScreenAI</td></tr><tr><td>5B</td><td>0.43</td><td>0.95</td><td>-</td></tr><tr><td colspan="4">MoUI (1-stage training)</td></tr><tr><td>4B</td><td>0.74</td><td>0.83</td><td>0.42</td></tr><tr><td colspan="4">MoUI (2-stage training)</td></tr><tr><td>1B</td><td>0.66</td><td>0.86</td><td>0.43</td></tr><tr><td>2B</td><td>0.67</td><td>0.86</td><td>0.43</td></tr><tr><td>4B</td><td>0.78</td><td>0.89</td><td>0.49</td></tr></table>
+
+Table 3: Results of MoUI and baseline models on reasoning tasks. We report the SQuAD F1 scores for all three reasoning tasks. MoUI outperforms ScreenAI and Ferret-UI-anyres (8B) on the complexQA benchmark and achieves competitive scores on the Short ScreenQA benchmark.
+
+# 6.3 Qualitative Error Analysis
+
+For OCR-related tasks, the model occasionally confuses text adjacent to the given bounding box with the intended text. In icon-related tasks, errors arise when distinguishing visually similar icon types, such as menu vs. list or expand_more vs. arrow_downward. Additionally, in some cases, the model incorrectly includes the red bounding box enclosing the icon as part of the icon itself, leading to misclassification. In widget recognition, the model struggles with fine-grained distinctions, often confusing text with text button or image with icon. Furthermore, since many widgets contain illustrations, they are sometimes misinterpreted as images. For reasoning tasks, the model faces challenges when multiple entities are involved, often failing to infer relationships between them and struggles to construct the logical reasoning necessary to answer the question correctly. We have added a few error cases in the Appendix G
+
+# 6.4 Ablation Studies
+
+# Comparison with one stage training
+
+In this setting, instead of training the model in two different stages for perception and reasoning tasks as shown in figure 4, we train the model only once on the entire data. This experiment helps us to establish the importance of perception guidance during reasoning training stage. We observe that in table 3 there is an increase in the accuracy of  $\approx 4\%$  on all the three reasoning tasks. We conclude that integrating perception guidance enhances the reasoning capabilities of MoUI. As shown in Appendix, Section 8 we find that Stage II training not only improves the model's reasoning abilities but also enhances certain perception capabilities acquired during Stage I, resulting in overall better performance.
+
+# Comparison with other vision encoders
+
+To enhance the perception and reasoning abilities of VLMS in use-cases like Mobile UI understanding, along with high quality instruction tuning data; a strong vision encoder that can accommodate images of any aspect ratio is important. To empirically verify this, we experimented with TinyLLaVA-3.1B model with SigLIP (Zhai et al., 2023) as vision encoder and Phi-2 (Li et al., 2023) as language decoder. As shown in Appendix E we can observe that even though there is an increase in the accuracy after finetuning on MoUI data the models are performing below par on perception tasks which
+
+results in lower performance in reasoning tasks as well.
+
+# 7 Conclusion and Future Work
+
+In this work, we introduce a series of lightweight MoUI models, along with the MoIT instruction-tuning dataset and the MoIQ reasoning benchmark. We also propose a two-stage training pipeline that first enhances the model's perception capabilities, and then leverages these improvements to strengthen its reasoning abilities. Our results demonstrate that the careful selection of a suitable vision encoder, high-quality training data, and the proposed training pipeline collectively contribute to achieving state-of-the-art or competitive performance on referring, grounding, and reasoning tasks—all while using smaller models. Despite these advances, the MoIQ benchmark remains challenging, with the current state-of-the-art model reaching only  $49\%$  accuracy. This highlights the need for more robust reasoning models in the UI domain. Looking ahead, we aim to close this gap by developing models capable of multi-command grounding, spatial reasoning, and understanding implicit user intent in complex UI contexts. This will involve extending our current instruction tuning paradigm to support multi-turn interactions and compositional tasks, while incorporating richer contextual cues from both visual and semantic modalities. By pushing the boundaries of reasoning in multimodal UI understanding, we hope to enable more intelligent and adaptable user interfaces. To foster continued research in this area, we open-source our datasets, models, and code.
+
+# 8 Limitations
+
+Current models are designed to process instructions in a single language. Extending them to support multiple languages would significantly enhance their applicability, enabling deployment to a broader, global audience. This multilingual capability would not only improve accessibility for users in non-English-speaking regions but also facilitate cross-lingual knowledge transfer, especially benefiting low-resource languages. Our dataset generation pipeline currently relies on view hierarchies or similar metadata about UI screens to construct meaningful instruction-response pairs. This structural information is crucial for grounding instructions to specific interface elements and understanding the semantic layout of the screen. At present,
+
+we use spaCy for entity extraction; however, more advanced models are available that offer improved accuracy, albeit at the cost of higher computational requirements.
+
+# 9 Acknowledgments
+
+Authors from Indian Institute of Technology, Bombay would like to thank Neysa for providing seamless GPU support.
+
+# References
+
+Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736.  
+Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Carbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. 2024. Screenai: A vision-language model for ui and infographics understanding. Preprint, arXiv:2402.04615.  
+Chongyang Bai, Xiaoxue Zang, Ying Xu, Srinivas Sunkara, Abhinav Rastogi, Jindong Chen, et al. 2021. Uibert: Learning generic multimodal representations for ui understanding. arXiv preprint arXiv:2107.13731.  
+Andrea Burns, Deniz Arsan, Sanjna Agrawal, Ranjitha Kumar, Kate Saenko, and Bryan A Plummer. 2022. A dataset for interactive vision-language navigation with unknown command feasibility. In European Conference on Computer Vision, pages 312-328. Springer.  
+Andrea Burns, Kate Saenko, and Bryan A Plummer. 2024. Tell me what's next: Textual foresight for generic ui representations. arXiv preprint arXiv:2406.07822.  
+Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yin ing Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong
+
+Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao, and Dahua Lin. 2024. Internl m2 technical report. Preprint, arXiv:2403.17297.  
+Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. 2024a. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv e-prints, pages arXiv-2406.  
+Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. 2021. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852.  
+Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. 2024b. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271.  
+Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. 2024c. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198.  
+Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. 2024. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935.  
+LMDeploy Contributors. 2023. Lmdeploy: A toolkit for compressing, deploying, and serving llm. https://github.com/InternLM/lmdeploy.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven CH Hoi. 2023. Instructclip: Towards general-purpose vision-language models with instruction tuning. arxiv abs/2305.06500 (2023).  
+Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. 2017a. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th Annual Symposium on User Interface Software and Technology, UIST '17.  
+Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. 2017b. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual ACM symposium on user interface software and technology, pages 845-854.
+
+W Keith Edwards, Elizabeth D Mynatt, and Kathryn Stockton. 1995. Access to graphical interfaces for blind users. *Interactions*, 2(1):54-67.  
+Longxi Gao, Li Zhang, Shihe Wang, Shangguang Wang, Yuanchun Li, and Mengwei Xu. 2024a. Mobileviews: A large-scale mobile gui dataset. arXiv preprint arXiv:2409.14337.  
+Longxi Gao, Li Zhang, Shihe Wang, Shangguang Wang, Yuanchun Li, and Mengwei Xu. 2024b. Mobile-views: A large-scale mobile gui dataset. Preprint, arXiv:2409.14337.  
+Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. 2024. Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243.  
+Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. 2024. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530.  
+Zecheng He, Srinivas Sunkara, Xiaoxue Zang, Ying Xu, Lijuan Liu, Nevan Wichers, Gabriel Schubiner, Ruby Lee, and Jindong Chen. 2021. Actionbert: Leveraging user actions for semantic understanding of user interfaces. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 5931-5938.  
+Qing Jiang, Yuqin Yang, Yuda Xiong, Yihao Chen, Zhaoyang Zeng, Tianhe Ren, Lei Zhang, et al. 2024. Chatrex: Taming multimodal llm for joint perception and understanding. arXiv preprint arXiv:2411.18363.  
+Gang Li and Yang Li. 2022. Spotlight: Mobile ui understanding using vision-language models with a focus. arXiv preprint arXiv:2209.14927.  
+Gang Li and Yang Li. 2023. Spotlight: Mobile ui understanding using vision-language models with a focus. In The Eleventh International Conference on Learning Representations.  
+Guanzhen Li, Yuxi Xie, and Min-Yen Kan. 2024a. MVP-bench: Can large vision-language models conduct multi-level visual perception like humans? In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13505-13527, Miami, Florida, USA. Association for Computational Linguistics.  
+Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. 2025. Screenshot-pro: Gui grounding for professional high-resolution computer use. Preprint.  
+Yang Li, Jiacong He, Xin Zhou, Yuan Zhang, and Jason Baldridge. 2020a. Mapping natural language instructions to mobile ui action sequences. arXiv preprint arXiv:2005.03776.
+
+Yang Li, Gang Li, Luheng He, Jingjie Zheng, Hong Li, and Zhiwei Guan. 2020b. Widget captioning: Generating natural language description for mobile user interface elements. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5495-5510.  
+Yuanzhi Li, Sebastien Bubeck, Ronen Eldan, Allie Del Giorno, Suriya Gunasekar, and Yin Tat Lee. 2023. Textbooks are all you need II: phi-1.5 technical report. CoRR, abs/2309.05463.  
+Zhangheng Li, Keen You, Haotian Zhang, Di Feng, Harsh Agrawal, Xiujun Li, Mohana Prasad Sathya Moorthy, Jeff Nichols, Yinfei Yang, and Zhe Gan. 2024b. Ferret-ui 2: Mastering universal user interface understanding across platforms. arXiv preprint arXiv:2410.18967.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. In Advances in Neural Information Processing Systems, volume 36, pages 34892-34916. Curran Associates, Inc.  
+Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo. 2024. Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451.  
+Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. 2023. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824.  
+Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. 2024. Perceptiongpt: Effectively fusing visual perception into lvm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27124-27133.  
+Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. 2023. Androidinthewild: A large-scale dataset for android device control. In Advances in Neural Information Processing Systems, volume 36, pages 59708-59728. Curran Associates, Inc.  
+Srinivas Sunkara, Maria Wang, Lijuan Liu, Gilles Baechler, Yu-Chung Hsiao, Abhanshu Sharma, James Stout, et al. 2022. Towards better semantic understanding of mobile interfaces. arXiv preprint arXiv:2210.02663.  
+Bryan Wang, Gang Li, Xin Zhou, Zhourong Chen, Tovi Grossman, and Yang Li. 2021. Screen2words: Automatic mobile ui summarization with multimodal learning. In The 34th Annual ACM Symposium on User Interface Software and Technology, pages 498-510.  
+Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. 2023. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079.
+
+Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, et al. 2024a. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. arXiv preprint arXiv:2406.08394.
+
+Qinzhuo Wu, Weikai Xu, Wei Liu, Tao Tan, Jianfeng Liu, Ang Li, Jian Luan, Bin Wang, and Shuo Shang. 2024b. Mobilevm: A vision-language model for better intra-and inter-ui understanding. arXiv preprint arXiv:2409.14818.
+
+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. arXiv preprint arXiv:2412.15115.
+
+Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. 2023. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704.
+
+Keen You, Haotian Zhang, Eldon Schoop, Floris Weers, Amanda Swearngin, Jeffrey Nichols, Yinfei Yang, and Zhe Gan. 2024. Ferret-ui: Grounded mobile ui understanding with multimodal llms. In European Conference on Computer Vision, pages 240-255. Springer.
+
+Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. 2024. Contextual object detection with multimodal large language models. Preprint, arXiv:2305.18279.
+
+Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sigmoid loss for language image pre-training. Preprint, arXiv:2303.15343.
+
+Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Leizhang, Chunyuan Li, et al. 2024a. Llavage: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer.
+
+Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. 2024b. Ferret-v2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973.
+
+Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. 2024c. Android in the zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713.
+
+Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. 2024. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289.
+
+Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592.
+
+# A Annotation Details
+
+For MoUI we employed five annotators for writing QAs on MobileUI screens. They are experts in writing reasoning based datasets and have annotated several such datasets before. Three annotators are tasked for writing QA and two annotators are tasked for verification in a double-blinded manner. Each annotator is tasked to write 2-3 questions per screen. Each question will be verified by two other annotators. A question will be considered only if both the verifiers mark it as valid. A total of  $4.5k$  questions were written out of which  $3k$  are retained. The annotators were compensated as per the laws of the geographical location.
+
+# B Evaluation Metrics
+
+We use the below evaluation metrics for evaluating our models on various tasks.
+
+<table><tr><td>Task</td><td>Metric Used</td></tr><tr><td colspan="2">Spotlight Tasks</td></tr><tr><td>Screen Summarization</td><td>CIDEr</td></tr><tr><td>Widget Captioning</td><td>CIDEr</td></tr><tr><td colspan="2">Elementary Tasks</td></tr><tr><td>OCR</td><td>Exact Match</td></tr><tr><td>Icon Recognition</td><td>Exact Match</td></tr><tr><td>Widget Recognition</td><td>Exact Match</td></tr><tr><td>Find Text</td><td>Acc@IoU=0.1</td></tr><tr><td>Find Icon</td><td>Acc@IoU=0.1</td></tr><tr><td>Find Widget</td><td>Acc@IoU=0.1</td></tr><tr><td>Spatial Polar</td><td>Exact Match</td></tr><tr><td colspan="2">Reasoning Tasks</td></tr><tr><td>Complex ScreenQA</td><td>SQuAD F1</td></tr><tr><td>ScreenQA Short</td><td>SQuAD F1</td></tr><tr><td>MoIQ</td><td>SQuAD F1</td></tr></table>
+
+Table 4: Evaluation Metrics Used
+
+# C Prompts for Data Generation and Zero-Shot Inference
+
+# C.1 Spotlight Tasks Data Generation
+
+# Prompt 1: Screen Summarization Dataset Generation Prompt
+
+Design a question-answer pair where a user asks the assistant to provide a summary of the screen. Keep in mind that the interaction should be between a mobile user and a visual AI assistant. You are given a screenshot of the mobile screen where the user is and a short summary describing the contents of the screen. Use the provided summary as a starting point to describe the screen for the user and keep it short and concise.
+
+Format the output in the following form:
+
+USER:"<question for providing summary>"
+
+AI ASSISTANT: "<generated summary output>"
+
+Provided Summary: "Discussion forum app for anime and K-dramas"
+
+# C.2 Ferret-UI-anyres Prompts
+
+# Prompt 1: Complex ScreenQA Base Prompt
+
+Answer only the question asked in text and if the query is counting or arithmetic based only output the numerical value as the answer. Here is the question you shall answer:  $\{user\_ query\}$
+
+# Prompt 2: ScreenQA Short Base Prompt
+
+If the question is about counting, only answer in a single number. Here is the question you shall answer: {user_query}
+
+# Prompt 2: Widget Captioning Dataset Generation Prompt
+
+Design a question-answer pair where a user asks the assistant to provide a phrase that best describes the functionality of the interactive element [bbox]. Keep in mind that the interaction should be between a mobile user and a visual AI assistant. You are given a screenshot of the mobile screen where the user is with a_bbox in red colour around the widget and its location in the form of [x1, y1, x2, y2] normalised to the size of the screen. You will also receive a sample caption that describes the widget enclosed in the_bbox at the provided location. Use the provided caption as a starting point to describe the functionality of the widget for the user and keep it short and concise.
+
+Format the output in the following form:
+
+USER: "<question for describing the functionality of the widget at [bbox]>"
+
+AI ASSISTANT: "<generated caption for the functionality of the widget>"
+
+Provided inputs:
+
+bbox location: [0, 320, 840, 189]
+
+caption: "advertisement"
+
+# Prompt 3: Spatial Polar Task Base Prompt
+
+Please make sure that the answer is only Yes/No. Here is the question you shall answer: {user_query}
+
+# C.3 Zero-Shot Inference Prompts
+
+# Prompt 1: OCR Task Base Prompt
+
+You are an AI visual assistant designed to help users with questions regarding content on a UI screen.
+
+You will be provided with an image of the UI screen with a red bounding box surrounding a region.
+
+Please extract the text inside the bounding box region from the given image.
+
+Only output the text and nothing else.
+
+Here is the question you shall answer:  $\{user_query\}$
+
+# Prompt 2: Screen Summarization Task Base Prompt
+
+You are an AI visual assistant designed to help users with questions regarding content on a UI screen.
+
+You will be provided with an image of the UI screen. Your task is to very briefly summarize the mobile screen given. Only output the screen summary and nothing else.
+
+Here is the question you shall answer:  $\{user_query\}$
+
+# Prompt 3: Widget Captioning Base Prompt
+
+You are a digital assistant that needs to answer queries from the user about a mobile UI screen. You will be given a query from user and the screen as inputs. Carefully focus on the screen and the red bounding box given. Keep your answers short and concise.
+
+Please only use the following format: AN-SWER: <answer> User: {user_query}
+
+# Prompt 4: ScreenQA Short Base Prompt
+
+Answer user queries about the mobile screen. The answer should be as short as possible. User: {user_query}
+
+# Prompt 6: ScreenQA Short Base Prompt
+
+You are a digital assistant that needs to answer queries from the user about a mobile UI screen. You will be given a query from user and the screen as inputs. Carefully focus on the screen and the red bounding box given. Your task is to predict the icon type of the icon given in the red bounding box.
+
+The icon type can only be of the types: expand-less, redo, follow, bluetooth, add, notifications, avatar, edit, arrow_backward, call, lock, font, microphone, menu, globe, thumbs_down, skipprevious, folder, playlist, filter, settings, close, emoji, delete, build, wallpaper, thumbs_up, explore, swap, refresh, star, search, volume, sliders, time, photo, zoom_out, weather, date_range, send, videocam, more, info, national_flag, bookmark, gift, power, reply, launch, email, dialpad, copy, group, filter_list, home, repeat, warning, minus, cart, compare, music, arrow_downward, arrow_forward, visibility, expand_more, book, shop, help, save, dashboard, share, favorite, facebook, pause, list, location_crosshair, location, check, description, skip_next, label, undo, fullscreen file_download, play, attach_file, navigation, network_wifi, flash, twitter, av_rewind, history, chat, arrow_upward layers, switcher, av_forward, flight.
+
+Please only answer from the given icon types.
+
+Keep your answers short and concise.
+
+Please only use the following format:
+
+ANSWER: <answer>
+
+User: {user_query}
+
+# Prompt 5: Grounding Tasks Base Prompt
+
+Please provide the bounding box coordinates of the region this sentence describes: <ref>user_query</ref>
+
+# Prompt 7: Widget Recognition Base Prompt
+
+You are a digital assistant that needs to answer queries from the user about a mobile UI screen.
+
+You will be given a query from user and the screen as inputs.
+
+Carefully focus on the screen and the red bounding box given.
+
+Your task is to predict the widget type of the widget given in the red bounding box.
+
+The widget type can only be of the types: Image, Web View, Text, Icon, List Item, Painter, Video, On/Off Switch, Input, Checkbox, Slider, Background Image, Pager Indicator, Advertisement, Text Button, Card, Radio Button, Map View.
+
+Please only answer from the given widget types and print nothing else other than the widget type. Keep your answers short and concise.
+
+Please only use the following format:
+
+ANSWER: <answer>
+
+User: {user_query}
+
+# D Instruction Templates
+
+<table><tr><td>Task</td><td>Instruction Template</td></tr><tr><td>Screen Summarization</td><td>&lt;image&gt;{Question}</td></tr><tr><td>Widget Captioning</td><td>&lt;image&gt;{Question}</td></tr><tr><td>OCR</td><td>&lt;image&gt;Given the screenshot image, extract the text in bounding box: {Question}</td></tr><tr><td>Icon Recognition</td><td>&lt;image&gt;Given the screenshot image, output the icon type for element in bounding box: {Question}</td></tr><tr><td>Widget Recognition</td><td>&lt;image&gt;Given the screenshot image, output the widget type for element in bounding box: {Question}</td></tr><tr><td>Find Text</td><td>&lt;image&gt;Please give the bounding box coordinates for the question: {Question}</td></tr><tr><td>Find Icon</td><td>&lt;image&gt;Please give the bounding box coordinates for the question: {Question}</td></tr><tr><td>Find Widget</td><td>&lt;image&gt;Please give the bounding box coordinates for the question: {Question}</td></tr><tr><td>Spatial Polar</td><td>&lt;image&gt;Given the image and two bounding boxes, answer Yes/No for the question: {Question}</td></tr><tr><td>Complex ScreenQA</td><td>Given the image, answer the following question with no more than five words. {Question}</td></tr><tr><td>ScreenQA Short</td><td>Based on the image, respond to this question with a short answer: {Question}</td></tr></table>
+
+# E TinyLLava-Gemma Results
+
+Table 5: Instruction Templates Used for MoIT dataset  
+
+<table><tr><td>Setting</td><td>Screen2 Words</td><td>Widget Capt.</td><td>OCR</td><td>Icon Recog.</td><td>Widget Recog.</td><td>Complex Short QA</td><td>QA</td><td>Spatial Polar</td></tr><tr><td>Zero-Shot</td><td>34.23</td><td>25.66</td><td>4.20</td><td>0</td><td>8.29</td><td>34.76</td><td>5.26</td><td>54.8</td></tr><tr><td>Finetuned TinyLLava</td><td>78.9</td><td>117.32</td><td>21.6</td><td>36.03</td><td>41.58</td><td>41.23</td><td>4.64</td><td>92.68</td></tr></table>
+
+# F MoIT Dataset Statistics
+
+Table 6: TinyLLava-Gemma results for perception tasks before and after training on our MoIT dataset  
+
+<table><tr><td>UI Task</td><td>Method</td><td>Count of QA pairs</td></tr><tr><td>Screen Summarization</td><td>InternVL2</td><td>10000</td></tr><tr><td>Widget Captioning</td><td>InternVL2</td><td>19475</td></tr><tr><td>Find Icon</td><td>Template</td><td>16796</td></tr><tr><td>Widget Classification</td><td>Template</td><td>19299</td></tr><tr><td>Icon Recognition</td><td>Template</td><td>16796</td></tr><tr><td>Find Widget</td><td>Template</td><td>19299</td></tr><tr><td>OCR</td><td>Template</td><td>18871</td></tr><tr><td>Spatial Polar</td><td>Template</td><td>19261</td></tr></table>
+
+Table 7: MoIT dataset statistics
+
+# G Error Cases
+
+# OCR-Related Error
+
+![](images/a9eb7dc26b78668e9b2765ac0a9358ac99d99dee143437ef80ec47fb2135074c.jpg)
+
+# User Query
+
+What does the text inside the red bounding box on the given screen say?
+
+# Model Output
+
+Underworld: Blood Wars
+
+# Error Analysis
+
+Here, the ground truth answer is "Dec 1, 2016." However, the model misinterprets the adjacent text as the intended text within the bounding box and outputs "Underworld: Blood Wars" instead. We hypothesize that this occurs because the adjacent text is significantly larger, leading the model to prioritize it over the smaller text inside the bounding box.
+
+# Icon Recognition Error
+
+![](images/30f95f4965b05eca985dde37cb5d34ac6c6d300fd82e2a24eef60a1ea9b778aa.jpg)
+
+# User Query
+
+Can you identify the icon type of the element enclosed by the red bounding box in the provided screen?
+
+# Model Output
+
+play
+
+# Error Analysis
+
+Here, the correct answer is "arrow_forward". Since the bounding box is tightly wrapped around the icon, the model mistakenly interprets it as part of the icon itself, leading it to incorrectly classify the icon as play.
+
+# Reasoning-Based Error
+
+![](images/c6410f66ed0ff1bddbceae2ff1a3e72a6ff4973f8eef24401520096a5637662a.jpg)
+
+![](images/feca993332818dc472b752b23c29ca7d4cf3e6ae568d205f9450693993c7cfdd.jpg)
+
+![](images/a2005bcfa73a7c330e02ff4ceba53bccdd6b98ab83d7e8cda509fc541a8bef1a.jpg)
+
+![](images/d917d4d730ff4e5c265ebebe579dc4aa09a42dd1e8b3d8b72befa1b8adb604e3.jpg)
+
+![](images/041fe101f4686993f85366d43109e76a7c5e95d58b6de0d513995f177ae2e99b.jpg)
+
+![](images/571b7a14c4d2e3c39f22d3135e798b66839c0da311915719bd10d82e3b6eebfc.jpg)
+
+![](images/f33aa7d14ece28bf03d7bb86bd6d3dac10ca6eb77ee7bbc227a20a34a3b9b579.jpg)
+
+![](images/a2f82603353b976fe28fcafb3133fc2afe6d6e3ff73a1f03b247ff498abaefc9.jpg)
+
+![](images/aa202a8c4475f8b0cfc2955313b56a8562b92a7fb1bc896d381f56c9d8ed2fc0.jpg)
+
+![](images/c95606c93cb573fd48dc6c4c091bd090799f11d0511b49224e2ab79e11058c8e.jpg)
+
+# User Query
+
+How many more actors are in the movie's cast than actresses?
+
+# Model Output
+
+2
+
+# Error Analysis
+
+Here, the ground truth answer is 1. However, the model misidentifies the actress count, likely due to incomplete cast images on the screen, resulting in an incorrect output of 2.
+
+# Reasoning-Based Error
+
+![](images/e67e6d5b85dae520676e54819b5c865bc89d40ab70335bbe0b68843d8b147b0e.jpg)
+
+# User Query
+
+Which option is selected: Login or Sign Up?
+
+# Model Output
+
+Login
+
+# Error Analysis
+
+Since the "Sign Up" button is highlighted, the correct answer is Sign Up. But, the model misidentifies the active button as Login which is not highlighted, likely due to varying UI styles across applications.
+
+# H Stage I vs Stage II Comparison for Perception Tasks
+
+<table><tr><td></td><td>Screen2 Words</td><td>Widget Capt.</td><td>OCR</td><td>Icon Recog.</td><td>Widget Recog.</td><td>Find Text</td><td>Find Icon</td><td>Find Widget</td><td>Spatial Polar</td></tr><tr><td>After Stage I</td><td>123.15</td><td>133.41</td><td>90.53</td><td>83.33</td><td>79.67</td><td>89.25</td><td>75.96</td><td>79.16</td><td>98.85</td></tr><tr><td>After Stage II</td><td>123.2</td><td>151.25</td><td>90.63</td><td>87.15</td><td>81.33</td><td>89.14</td><td>75.75</td><td>83.51</td><td>98.81</td></tr></table>
+
+Table 8: Stage I vs Stage II performance comparison on the various perception tasks.
+
+# I Performance on the ScreenSpot Benchmark
+
+<table><tr><td>LVLMs</td><td>Model Size</td><td>Training Data Size</td><td>Mobile (Text)</td><td>Mobile (Icon/Widget)</td></tr><tr><td>MoUI</td><td>1B</td><td>150K</td><td>53.43%</td><td>33.56%</td></tr><tr><td>MoUI</td><td>4B</td><td>150K</td><td>71.85%</td><td>47.65%</td></tr><tr><td>SeeClick</td><td>9.6B</td><td>1M</td><td>78.0%</td><td>52.0%</td></tr></table>
+
+Table 9: Comparison of LVLMs across model size, training data, and mobile performance
+
+# J Comparison with other Mobile UI datasets
+
+<table><tr><td rowspan="2">Data</td><td rowspan="2">Size</td><td colspan="3">Task Coverage</td><td colspan="2">Dataset Generation Method</td></tr><tr><td>Referring</td><td>Grounding</td><td>Spatial</td><td>Perception</td><td>Reasoning</td></tr><tr><td>ScreenAI</td><td>&gt;400M</td><td>✓</td><td>✗</td><td>✗</td><td>Model generated</td><td>Model generated</td></tr><tr><td>FerretUI</td><td>250K</td><td>✓</td><td>✓</td><td>✗</td><td>Model generated</td><td>Model generated</td></tr><tr><td>MoUI (ours)</td><td>150K</td><td>✓</td><td>✓</td><td>✓</td><td>Template based generation</td><td>Human written</td></tr></table>
+
+Table 10: Comparison of Mobile UI datasets
+
+# K Annotated Bounding Boxes after Stage I Training
+
+![](images/a0766745242970b8a76f055083a6e5096f8a282c4736fc91b63490bb03cb242a.jpg)  
+Figure 5: Examples of bounding box annotations of the extracted key entities from the user query by the perception aware model after the Stage I training.
+
+![](images/493c6bac6157755968519b3fc63b5e61a929f7c35779eb1ea045983f9abf3497.jpg)
+
+![](images/995d09e0db1117daa0687ee0ca2247f42b03e93b2b123dc19d0365178a335166.jpg)
+
+Question: How many text inputs are there after the text 'We need your name and email'?  
+Ground Truth: 2  
+Key Entities Extracted: "How many text inputs", "the text", "your name and email"
+
+Question: How long is the current automatic refresh interval?  
+Ground Truth: 60 sec  
+Key Entities Extracted: "the current automatic refresh interval"
+
+Question: How many years does the age range span?  
+Ground Truth: 5  
+Key Entities Extracted: "How many years", "the age range span"

2025/From Perception to Reasoning_ Enhancing Vision-Language Models for Mobile UI Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d974fc56be444ea0d82ec97c1f650f6811929677174b7e0a4c9ee3ae47d3d6a
+size 1038327

2025/From Phrases to Subgraphs_ Fine-Grained Semantic Parsing for Knowledge Graph Question Answering/f27726f1-38f6-4ebc-b227-36fba1f0a97b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1711e5cc401f0cb4e61e537b9edb99cf5157c2855e0b77e3f4f4f9e9761cb81d
+size 946412

2025/From Phrases to Subgraphs_ Fine-Grained Semantic Parsing for Knowledge Graph Question Answering/full.md

@@ -0,0 +1,427 @@
+# From Phrases to Subgraphs: Fine-Grained Semantic Parsing for Knowledge Graph Question Answering
+
+Yurun Song*, Xiangqing Shen*, Rui Xia†
+
+School of Computer Science and Engineering,
+
+Nanjing University of Science and Technology, China
+
+{yrsong, xqshen, rxia}@njust.edu.cn
+
+# Abstract
+
+The recent emergence of large language models (LLMs) has brought new opportunities to knowledge graph question answering (KGQA), but also introduces challenges such as semantic misalignment and reasoning noise. Semantic parsing (SP), previously a mainstream approach for KGQA, enables precise graph pattern matching by mapping natural language queries to executable logical forms. However, it faces limitations in scalability and generalization, especially when dealing with complex, multi-hop reasoning tasks. In this work, we propose a Fine-Grained Semantic Parsing (FGSP) framework for KGQA. Our framework constructs a fine-grained mapping library via phrase-level segmentation of historical question-logical form pairs, and performs online retrieval and fusion of relevant subgraph fragments to answer complex queries. This fine-grained, compositional approach ensures tighter semantic alignment between questions and knowledge graph structures, enhancing both interpretability and adaptability to diverse query types. Experimental results on two KGQA benchmarks demonstrate the effectiveness of FGSP, with a notable  $18.5\%$  relative F1 performance improvement over the SOTA on the complex multi-hop CWQ dataset. Our code is available at https://github.com/NUSTM/From-Phrases-to-Subgraphs.
+
+# 1 Introduction
+
+Knowledge Graph Question Answering (KGQA) aims to leverage structured information stored in knowledge graphs such as Freebase (Bollacker et al., 2008) and Wikidata (Vrandecic and Krötzsch, 2014) to answer natural language questions. It has been widely applied in domains including Web search (Jang et al., 2017), medical consultation (Wu et al., 2024), and legal analysis (Cui et al.,
+
+![](images/9a3afcd3aefeaa779106105dac13a136e393b55cb11feb5a681fdc1c5a160fc8.jpg)
+
+![](images/7d2d0a6b9452b937491e91593f800f1970be8eb13157afc765a689ef515b04fd.jpg)
+
+![](images/a9dd2314e69fd4ba1ee7ae6e3bf1cbda385fc8a75ee9dcd255b96d80395c5e98.jpg)  
+Figure 1: Comparison of three types of KGQA approaches.
+
+2024). Knowledge graphs (KGs), which represent information via triple-based structures of entities and relations, provide a structured and semantically rich knowledge source. This inherent structure enables precise reasoning and facilitates multi-hop inference.
+
+In recent years, the emergence of large language models (LLMs) has significantly influenced the research paradigm of KGQA. Existing approaches that combine LLMs with KGs can be broadly categorized into two types: agent exploration frameworks (Sun et al., 2024; Zhu et al., 2024; Jiang et al., 2024) and path generation frameworks (Luo et al., 2024a; Sui et al., 2024). The former generates reasoning paths through continuous interaction be
+
+tween the LLMs and KGs to solve problems, while the latter has the LLMs generate reasoning paths first, which are then mapped to KGs instances to retrieve relevant information. Both approaches face the dual challenges of semantic gaps and paths: a) the unstructured nature of LLMs training data contrasts with the structured text of KGs, leading to hallucination; b) the LLMs' sequential reasoning paradigm generates redundant paths when handling multi-hop problems, causing noise accumulation in reasoning subgraphs. As illustrated in Figure 1, the model may generate irrelevant paths such as "Germany  $\rightarrow$  Scotland  $\rightarrow$  Pound sterling", thereby introducing spurious connections that hinder accurate reasoning. The root cause of these issues lies in the difference between natural language and graph-structured data, which motivates us to explore solutions that are more suited to the characteristics of graph structures.
+
+Before the rise of LLM-based methods, semantic parsing (SP) was widely regarded as a mainstream approach for KGQA, known for its precision and efficiency (Berant and Liang, 2014; Dong and Lapata, 2016; Shaw et al., 2019). Semantic parsing transforms natural language questions into executable logical forms (LFs), which are then used to query knowledge graphs. This approach offers several distinct advantages over LLM-based methods. First, LFs align naturally with the entity-relationship structure of KGs, ensuring strict semantic alignment. This mitigates the hallucination problem caused by the unstructured training data of LLMs. Second, the syntax of LFs supports direct graph pattern matching, enabling precise and efficient reasoning. This avoids the redundant and noisy paths often produced by LLMs' sequential reasoning in multi-hop tasks. However, mapping between natural language and graph structures via LFs still poses challenges in scalability and generalization. As illustrated in Figure 1, when the model encounters an unseen semantic pattern—such as a question involving "currency"—it may incorrectly map it to a related but incorrect relation like "economy," leading to an erroneous LF and ultimately failing to retrieve the correct answer.
+
+To address this issue, we propose a Fine-Grained Semantic Parsing (FGSP) framework for KGQA. FGSP enables the decomposition of complex questions into atomic semantic units, allowing for more precise alignment between natural language and KG structures. The framework operates in two stages. In the offline stage, a fine-grained map
+
+ping library is constructed by extracting and curating historical question-LF pairs, where questions are decomposed into sub-questions representing reasoning steps, and the LF is parsed according to the syntax structure. In the online reasoning stage, hierarchical reasoning is performed, where the question decomposition module breaks down complex queries into sub-questions; the phrase retrieval module searches for relevant phrases and instantiates LF fragments; the subgraph fusion module implements novel sequential fusion and combination fusion rules to construct a complete informational subgraph. Through a fine-grained semantic alignment mechanism, the framework achieves hierarchical analysis and compositional resolution of complex questions, enhancing the flexibility of problem-solving and ultimately improving the accuracy of complex semantic parsing by precisely matching atomic semantic units. As shown in Figure 1, fragments such as "countries bordering Germany" and "currencies used in countries" can be extracted from historical questions. Merging their corresponding subgraphs yields a complete structure for answering the current question.
+
+Experiments on the WebQSP (Yih et al., 2016) and CWQ (Talmor and Berant, 2018) benchmark datasets demonstrate that our approach achieves strong performance, with a notable  $18.5\%$  relative F1 score improvement on the complex multi-hop CWQ dataset. This demonstrates the architectural advantages of our framework in handling multi-hop reasoning tasks and effectively reducing the retrieval of irrelevant information. Ablation studies further validate that the framework maintains robust performance across various base models and retrieval mechanisms.
+
+# 2 Preliminary
+
+In this section, we first introduce the definition of KG, KGQA and SP.
+
+Knowledge Graph (KG) is a directed graph  $\mathcal{G} = (\mathcal{E},\mathcal{R},\mathcal{T})$  , where  $\mathcal{E}$  is the set of entities,  $\mathcal{R}$  is the set of relations, and  $\mathcal{T}\subseteq \mathcal{E}\times \mathcal{R}\times \mathcal{E}$  is the set of triples. Each triple  $(h,r,t)\in \mathcal{T}$  encodes a factual assertion, where  $h,t\in \mathcal{E}$  are the heads and tail entities, and  $r\in \mathcal{R}$  is the relation linking them.
+
+Knowledge Graph Question Answering (KGQA) is a multi-hop reasoning task based on a knowledge graph (KG). Its goal is to translate a natural language question  $q \in \mathcal{Q}$  to an answer  $a \in \mathcal{A}$ , where  $\mathcal{A} \subseteq \mathcal{E}$ . Given a question  $q$ , the model
+
+![](images/eedad5fb5f7c1acc242f01289fab827700882d25aade9ba67008bac1606a5c43.jpg)
+
+![](images/2a68dc4c9ffd751943e6394ef42c709a3aa65616d7de6237575e36cf0581336a.jpg)  
+Figure 2: The overall framework of FGSP. Our framework is primarily divided into two stages. The offline stage (top-right) involves constructing a phrase-level question-SPARQL mapping library. In the online phase, the question is decomposed into sub-questions. Then the relevant phrases are retrieved and instantiated into sub-graphs, and finally the complete information sub-graph is constructed through sub-graph fusion.
+
+extracts a subgraph from the KG  $\mathcal{G}$  that contains the path-relevant evidence to infer the answer  $a$ .
+
+Semantic Parsing (SP) transforms natural language queries  $q \in \mathcal{Q}$  into logical forms  $z \in \mathcal{Z}$ , where  $\mathcal{Z}$  denotes a space of formal representations such as lambda-DCS, SPARQL, or s-expressions. In the context of KGQA, the logical form  $z$  is a structured query executable on a knowledge graph  $\mathcal{G}$  to retrieve the answer  $a$ .
+
+# 3 Methodology
+
+This section outlines the framework, which adopts SPARQL as the specific form of LF, and comprises three components: (1) the phrase-level question-SPARQL mapping library construction module, which analyzes questions and SPARQL to build a mapping library; (2) the phrase-level retrieval module, which retrieves relevant phrase pairs from the library based on sub-questions generated by LLMs; (3) the subgraph fusion module, which integrates retrieved SPARQL phrases into global subgraphs to derive accurate answers. Figure 2 presents an overview of the framework.
+
+# 3.1 Phrase-level Question-SPARQL Mapping Library Construction
+
+In the offline phase, we will construct a mapping library, with the phrases contained within it constituting the basis for phrase-level retrieval in the online phase. This stage consists of two key steps:
+
+(1) rule-based preprocessing of SPARQL queries, and (2) generation of context-aware phrase-level segmentation using LLMs.
+
+The direct implementation of LLM-driven phrase-level segmentation introduces two key challenges: (1) over-aggregation, where a single phrase contains multiple reasoning steps, and (2) over-fragmentation, which leads to phrase disintegration and compromises semantic integrity. To address this, we design a rule-based SPARQL parser for preprocessing. Initially, we identify the Basic Graph Patterns (BGPs) representing the core semantics of SPARQL as primary candidates for reasoning paths. Next, BGP components with endpoints that are non-entity nodes are merged to form long phrases with complete semantics. These phrases maintain structural consistency, enabling flexible assembly for semantic extension during multi-hop reasoning, while also ensuring that each phrase serves as the minimal reasoning unit, preserving the robustness of atomic reasoning.
+
+Building on preprocessed SPARQL fragments, we construct prompts that include the preprocessed phrases, the original SPARQL query, and the question. These prompts guide the LLM to establish semantic alignment through bidirectional phrase alignment. This method differs fundamentally from traditional approaches that convert triples into natural language. For instance, given the triple (Zazaki, iso_639_3_code, zza), traditional methods might
+
+generate descriptions such as "Zazaki's ISO 639-3 code is zza" whereas our framework generates a sentence more suited to the question, such as "The abbreviation for the Zazaki language is zza". It is important to note that, while the preprocessing step extracts the primary reasoning paths, the LLM still performs semantic mapping for other SPARQL syntactic structures to ensure the full utilization of SPARQL. In this approach, the LLM does not need to fully comprehend the entire KG; instead, it precisely maps the reasoning process of problem-solving to the closed phrase set, alleviating semantic mapping biases caused by incomplete KG understanding.
+
+# Phrase-level Segmentation Prompt
+
+Given a question, SPARQL, mid and its entity name, and the main reasoning path, split the question and SPARQL into pairs of clauses based on the given reasoning path. The parts that do not belong to the reasoning path also need to be split.
+
+# 3.2 Question Decomposition and Phrase-level Retrieval
+
+For classic multi-hop reasoning tasks such as KGQA, inspired by previous studies (Khot et al., 2023; Wang et al., 2023), we use concise prompt templates to utilize LLMs to decompose a multi-hop question into sub-questions. As shown in Figure 2 a, the question "What currencies do countries bordering Germany use?" is decomposed into two sub-questions: "What countries border Germany?" and "What are the currencies used in a country?". During this process, the LLM is not exposed to KG data.
+
+For each sub-question, we retrieve relevant NL phrases from the phrase-level question-SPARQL mapping library by computing embedding cosine similarity, selecting the top-k candidates, and converting them into corresponding SPARQL phrases. As shown in Figure 2 b, "Countries bordering Germany" and "Currencies used in the country" are retrieved as related clauses.
+
+# Question Decomposition Prompt
+
+Decompose the given question into sub-question. There should be no nesting between sub-questions.
+
+# 3.3 Sub-Knowledge Graph Instantiation and Fusion
+
+To establish a connection between the retrieved SPARQL phrases and the final question answer, this module focuses on converting the retrieved
+
+![](images/c0885a5b1079b1b32d15852202c71c5112534b35c87e09808c622ac0dfd73cb4.jpg)
+
+![](images/6396002a84e13902adbde8f771078176ec25c5794988261d095e6aa0307c8b3f.jpg)  
+Figure 3: Two fusion paradigms, sequential (top) and combinatorial (bottom)
+
+phrases into a coherent subgraph that accurately represents the original question.
+
+Before subgraph fusion, the first step is to instantiate the retrieved SPARQL phrases into question-relevant subgraphs. SPARQL fragments capture local subgraph information within a KG. For example, the SPARQL phrase " $?country\_location\_country\_currency\_used\_Euro"$  corresponds to a subgraph containing all countries that use Euro as their currency. The instantiation of the subgraph is achieved by performing entity replacement on the entities within the retrieved SPARQL phrases.
+
+A rule-based approach is then employed to fuse the subgraphs, transforming them into an integrated subgraph that represents the query's semantic content. This process builds on two classic reasoning paradigms in KGQA: (a) sequential reasoning  $(A\to B\to C)$  and (b) compositional reasoning  $(A + B\rightarrow C)$ . Specifically, sequential reasoning extends the semantics of a preceding subgraph with the information from subsequent subgraphs, forming a reasoning path that eventually leads to the answer. In contrast, compositional reasoning synthesizes multi-subgraph constraints to generate a subgraph that satisfies multiple constraints simultaneously, as shown in Figure 3. The rule-based scheme designed based on these two reasoning modes systematically integrates local subgraphs into a unified subgraph that reflects the semantics of the original query, from which the final answer is derived, enabling trustworthy inference on the knowledge graph.
+
+<table><tr><td rowspan="2">Type</td><td rowspan="2">Methods</td><td colspan="2">WebQSP</td><td colspan="2">CWQ</td></tr><tr><td>F1</td><td>Hit</td><td>F1</td><td>Hit</td></tr><tr><td rowspan="4">LLM-only</td><td>Llama3.1-8b (Meta, 2024)</td><td>34.8</td><td>55.5</td><td>22.4</td><td>28.1</td></tr><tr><td>Qwen2-7B (Qwen et al., 2024)</td><td>35.5</td><td>50.8</td><td>21.6</td><td>25.3</td></tr><tr><td>ChatGPT (OpenAI, 2022)</td><td>59.3</td><td>67.4</td><td>43.2</td><td>47.5</td></tr><tr><td>GPT-4 (OpenAI, 2023)</td><td>62.3</td><td>73.2</td><td>49.9</td><td>55.6</td></tr><tr><td rowspan="4">Inference-based</td><td>StructGPT (Jiang et al., 2023a)</td><td>63.7</td><td>72.6</td><td>49.6</td><td>54.3</td></tr><tr><td>Readi w/GPT-4 (Cheng et al., 2024)</td><td>-</td><td>78.7</td><td>-</td><td>67.0</td></tr><tr><td>ToG w/GPT-4 (Sun et al., 2024)</td><td>-</td><td>82.6</td><td>-</td><td>67.6</td></tr><tr><td>KG-CoT (Zhao et al., 2024)</td><td>-</td><td>84.9</td><td>-</td><td>62.3</td></tr><tr><td rowspan="6">Training-based</td><td>NSM (He et al., 2021)</td><td>62.8</td><td>68.7</td><td>42.4</td><td>47.6</td></tr><tr><td>TransferNet (Shi et al., 2021)</td><td>-</td><td>71.4</td><td>-</td><td>48.6</td></tr><tr><td>SR+NSM w E2E (Zhang et al., 2022)</td><td>64.1</td><td>69.5</td><td>46.3</td><td>49.3</td></tr><tr><td>UniKGQA (Jiang et al., 2023b)</td><td>70.2</td><td>75.1</td><td>48.0</td><td>50.7</td></tr><tr><td>DECAF (Yu et al., 2023)</td><td>78.8</td><td>82.1</td><td>-</td><td>70.4</td></tr><tr><td>RoG (Luo et al., 2024a)</td><td>70.8</td><td>85.7</td><td>56.2</td><td>62.6</td></tr><tr><td></td><td>Ours</td><td>73.3</td><td>88.4</td><td>66.6</td><td>91.6</td></tr></table>
+
+Table 1: QA performance (F1 and Hit) of FGSP on WebQSP and CWQ datasets. Bold fonts indicate the best performance. The results of lama3.1-8b and Qwen2-7B are from Luo et al. (2024b). The results of ChatGPT and GPT-4 are from Jiang et al. (2024). others are from the origin paper
+
+# 4 Experiment
+
+# 4.1 Datasets
+
+We use two classic benchmarks, WebQuestionSP (WebQSP) (Yih et al., 2016) and Complex WebQuestions (CWQ) (Talmor and Berant, 2018), to evaluate the effectiveness of our proposed framework. Freebase (Bollacker et al., 2008) is the underlying knowledge graph of these two datasets, containing about 88 million entities, 20,000 relation types, and 126 million RDF triples. The details of datasets are provided in the appendix A
+
+# 4.2 Baselines
+
+We compared our framework with 14 baseline models, categorized into three groups: 1) LLM-only, where the model relies solely on the LLM to answer questions without the use of KGs; 2) inference-based, where the language model and KGs interact to enable reasoning; and 3) training-based, where the language model is fine-tuned using a training dataset. The details of these approaches are provided in the appendix B.
+
+# 4.3 Evaluation Metrics
+
+Consistent with previous studies (Tan et al., 2023), we use Hit and F1 as our evaluation metrics. Hit verifies whether at least one correct answer appears in the generated predictions, while F1 measures the overall coverage of all correct answers by balancing precision and recall.
+
+# 4.4 Implementations
+
+For constructing the phrase-level question-SPARQL mapping library and question decomposition, we used GPT4o (OpenAI, 2024) as the base model, and the temperature was set to 0.4. BGE-m3 (Chen et al., 2024) was used as the embedding model in phrase retrieval. For the retrieved similar phrases, we used Top-5 as the subsequent subgraph fusion. For the baseline results in the table, we directly used the results reported in the corresponding paper for comparison. The LLM-only method used zero-shot prompts.
+
+# 4.5 Main results
+
+Table 1 presents the comparative experimental results of the proposed framework on two standard KGQA datasets. Compared to existing approaches that integrate LLMs with KGs, our framework demonstrates significant advantages in overall performance, which validates the method of deep integration of natural language and graph structures through semantic parsing. This approach effectively improves the efficiency and accuracy of information extraction from knowledge graphs. Specifically, in comparison to the LLM-only baseline models, our method achieves an F1 score improvement of  $20.8\%$  on the CWQ dataset, respectively, demonstrating that the structured knowledge retrieval mechanism can effectively address the
+
+![](images/9e0e9f5d489d878468d040a4e8a27f49b53827ba1671a2e649d88e69328867c2.jpg)  
+Figure 4: Comparison of fine-grained retrieval and question retrieval on WebQSP (left) and CWQ (right). Performance is shown for varying retrieval Top-k phrases.
+
+knowledge gaps of LLMs. Notably, on the more complex CWQ dataset, our method further surpasses the current state-of-the-art approach, RoG, by  $18.5\%$  in F1 score. This improvement is primarily attributed to the technical advantages of the standardized SPARQL query language, unlike traditional methods that rely on autonomous reasoning path generation by the model, the structured query mechanism ensures higher accuracy through a standardized information extraction process. It is particularly worth noting that this substantial improvement empirically demonstrates the flexibility and scalability advantages of our framework, which are realized through fine-grained semantic decomposition. Additionally, our approach adopts a decoupled architectural design, enabling the LLM to operate independently of the KG during question reasoning and sub-question decomposition.
+
+# 5 Analysis
+
+In this section, we further analyze our framework by addressing the following six questions:
+
+- Q1: How does the top-k selection in the retrieval stage affect the performance?  
+- Q2: How does fine-grained retrieval mapping affect the performance relative to the complete question retrieval?  
+Q3: How does the scale of phrases affect the performance?  
+- Q4: How do different models for phrase-level segmentation affect the performance?  
+- Q5: How do different models for question decomposition affect the performance?  
+Q6: How do different retrieval methods affect the performance?
+
+![](images/efc1faf37a2a5f50972c19cc44cc365d71aaf6a10c2ed477d4bc06fd72993df3.jpg)  
+Figure 5: The F1 scores (Left) and Hit (Right) of FGSP on the test set of CWQ with a various amount of phrases.
+
+# 5.1 Impact of Top-k and Fine-grained (Q1, Q2)
+
+This section validates the necessity and effectiveness of the fine-grained framework through comparative experiments, focusing on the performance differences between the proposed method and question retrieval-based approaches under various Top-k configurations. The baseline method based on question retrieval uses all questions from the training set as retrieval sources. Its implementation consists of two key steps: (1) replacing the entities in SPARQL with mentioned entities; (2) executing the knowledge graph query. As shown in Figure 4, the experimental results demonstrate that our method outperforms the baseline across all Top-k values, confirming the effectiveness of the fine-grained framework. Notably, on multi-hop complex questions in the CWQ dataset, our method shows a more significant performance gain, with an average F1 improvement of  $49.1\%$ . This is attributed to the modular nature of fine-grained phrases compared to the original queries. Through the flexible phrase composition, the system can flexibly retrieve the knowledge fragments required for multi-hop reasoning. Further analysis reveals that fine-grained retrieval is more sensitive to the Top-k parameter. Specifically, the growth rate of F1 in the range of  $k = 1$  to  $k = 10$  in the CWQ dataset is  $55.9\%$ , while the baseline method experiences a relatively limited performance improvement, with only a  $25\%$  increase in the same range. This phenomenon suggests that the fine-grained phrases can better leverage an expanded search space.
+
+# 5.2 Impact of Phrase Count (Q3)
+
+This section systematically explores the impact mechanism of phrase size on the model's robustness. To simulate the data availability differences in real-world scenarios, The experimental setup
+
+<table><tr><td rowspan="2">Models</td><td rowspan="2">Parameter</td><td colspan="2">CWQ</td></tr><tr><td>F1</td><td>Hit</td></tr><tr><td>GPT4o</td><td>-</td><td>66.6</td><td>91.6</td></tr><tr><td>gpt4o-mini</td><td>-</td><td>66.3</td><td>91.9</td></tr><tr><td></td><td>14B</td><td>66.4</td><td>91.8</td></tr><tr><td></td><td>7B</td><td>66.3</td><td>91.0</td></tr><tr><td>Qwen2.5</td><td>3B</td><td>63.9</td><td>86.9</td></tr><tr><td></td><td>1.5B</td><td>54.1</td><td>77.7</td></tr><tr><td>LLAMA 3.1</td><td>8B</td><td>63.7</td><td>89.5</td></tr><tr><td>LLAMA 3.2</td><td>3B</td><td>62.0</td><td>87.8</td></tr><tr><td></td><td>1B</td><td>52.0</td><td>77.7</td></tr></table>
+
+constructs a progressive data size scenario ranging from sparse phrases (5%) to complete phrases (100%). Figure 5 presents a performance comparison curve across different phrase sizes, revealing two important patterns: first, model performance shows a significant positive correlation with the number of phrases, with accuracy increasing linearly in the  $5\% -30\%$  range; second, once the number of phrases exceeds  $30\%$ , the improvement rate slows down, and as the phrase count approaches  $100\%$ , marginal gains are observed, indicating a performance saturation threshold. It is worth noting that under extreme data-limited conditions (e.g.,  $5\%$  phrases), the framework still exhibits excellent robustness: key metrics, F1 and Hit, remain in the range of  $58.2\% -66.6\%$  and  $85.1\% -91.6\%$ . This phenomenon validates the dual advantages of the framework design: scalability via performance improvements driven by incremental data, and strong adaptability under low-resource environments. The experimental results empirically demonstrate that the framework can maintain stable performance in resource-constrained scenarios, which is of significant practical value for real-world applications.
+
+# 5.3 Impact of Phrase-level Segmentation (Q4)
+
+We systematically evaluated the impact of various open-source and closed-source LLMs on the framework's performance in the phrase-level segmentation. In the experimental design, SPARQL statements preprocessed by rules were consistently input into different LLMs for fine-grained phrase decomposition. As shown in Table 2, GPT-4o demonstrated the best performance in terms of F1 score, while GPT-4o-mini outperformed others in the Hit metric. Notably, open-source models also exhibited performance comparable to that of closed-source models: Qwen-14B's decomposition results
+
+Table 2: Performance of different LLMs on phrase-level segmentation.  
+
+<table><tr><td rowspan="2">Models</td><td rowspan="2">Parameter</td><td colspan="2">CWQ</td></tr><tr><td>F1</td><td>Hit</td></tr><tr><td>GPT4o</td><td>-</td><td>66.6</td><td>91.6</td></tr><tr><td>GPT4o-mini</td><td>-</td><td>66.2</td><td>91.7</td></tr><tr><td rowspan="5">Qwen2.5</td><td>14B</td><td>65.2</td><td>89.8</td></tr><tr><td>7B</td><td>64.6</td><td>89.0</td></tr><tr><td>3B</td><td>63.4</td><td>88.4</td></tr><tr><td>1.5B</td><td>60.7</td><td>87.2</td></tr><tr><td>0.5B</td><td>58.1</td><td>84.3</td></tr><tr><td>LLAMA 3.1</td><td>8B</td><td>64.7</td><td>90.1</td></tr><tr><td rowspan="2">LLAMA 3.2</td><td>3B</td><td>62.0</td><td>87.8</td></tr><tr><td>1B</td><td>52.0</td><td>77.7</td></tr><tr><td>no deco</td><td>-</td><td>57.5</td><td>80.9</td></tr></table>
+
+Table 3: Performance of different models in the question decomposition. The no-deco means don't decompose the question.
+
+were closest to GPT-4o, and LLaMA-8B reached  $95.6\%$  of GPT-4o's performance, which validates the compatibility of this framework with both closed-source (GPT-4o) and open-source (Qwen, LLaMA) models. Further analysis revealed a sharp performance decline when the model's parameter size dropped below 1.5B. Diagnosing the outputs of smaller models, we found that, even with the aid of rule preprocessing, low-parameter models still struggled to accurately capture fine-grained semantic boundaries, frequently resulting in errors of over-aggregation and over-fragmentation. This issue is the key factor limiting their performance.
+
+# 5.4 Impact of Question Decomposition (Q5)
+
+We investigated the impact of different question decomposition models on the phrase-level retrieval stage. Four approaches were compared: GPT-4o, Qwen, LLaMA, and a baseline without decomposition strategy. All experiments were conducted under identical conditions. The same semantic embedding model and Top-5 retrieval parameters were maintained throughout the comparison. As shown in Table 3, when using LLMs for question decomposition, our framework exhibited F1 scores consistently in the 52.0-66.6 range. Remarkably, even 3B-parameter models achieved  $93 - 95\%$  relative performance compared to GPT-4o, which provides strong evidence for the robustness of this framework in the question decomposition stage. Ablation experiments showed that when the question decomposition module was removed, Hit and F1 scores decreased by  $9.2\%$  and  $9.1\%$ , respectively, compared to GPT-4o's decomposition approach, demonstrating the critical role of the ques
+
+<table><tr><td rowspan="2">Top-K</td><td rowspan="2">Methods</td><td colspan="2">WebQSP</td><td colspan="2">CWQ</td></tr><tr><td>F1</td><td>Hit</td><td>F1</td><td>Hit</td></tr><tr><td rowspan="4">Top-10</td><td>BM25</td><td>24.6</td><td>35.9</td><td>25.3</td><td>47.2</td></tr><tr><td>Sentence-BERT</td><td>73.9</td><td>90.2</td><td>70.3</td><td>94.8</td></tr><tr><td>conan</td><td>66.6</td><td>83.0</td><td>67.4</td><td>92.7</td></tr><tr><td>BGE-m3</td><td>78.1</td><td>91.9</td><td>71.7</td><td>95.3</td></tr><tr><td rowspan="4">Top-5</td><td>BM25</td><td>16.5</td><td>24.4</td><td>17.0</td><td>33.2</td></tr><tr><td>Sentence-BERT</td><td>63.9</td><td>82.4</td><td>64.9</td><td>90.7</td></tr><tr><td>conan</td><td>57.9</td><td>74.6</td><td>62.0</td><td>88.4</td></tr><tr><td>BGE-m3</td><td>73.3</td><td>88.4</td><td>66.6</td><td>91.6</td></tr></table>
+
+Table 4: Performance of different retrieval methods.
+
+tion decomposition module within this framework.
+
+# 5.5 Impact of Retrieval Methods (Q6)
+
+In phrase-level retrieval method comparison experiments, we systematically evaluated the performance differences between BM25 (Robertson and Walker, 1994) and other dense retrieval models such as BGE-M3 (Chen et al., 2024), SentenceBERT (Reimers and Gurevych, 2019), and Conan (Li et al., 2024). As shown in Table 4, vector-based retrieval methods exhibited slight performance variations between different embedding models: on the CWQ dataset, the best-performing BGE-M3 demonstrated a modest improvement of  $6.3\% -7.4\%$  in F1 and a  $2.8\% -3.6\%$  improvement in Hit over Conan, indicating that the choice of embedding model has a limited impact on this framework. The BM25 method significantly degraded system performance. This can be attributed to two interrelated factors. First, entities in historical questions exhibit heterogeneous characteristics, where different questions involve substantially distinct entities. Second, the phrase structure introduces additional complexity - specifically, the sparsity of word distributions in short phrases creates modeling challenges. Term-frequency-based methods particularly struggle to establish effective probability distribution models under these conditions of lexical sparsity.
+
+# 6 Related Work
+
+Currently, research on KGQA can be broadly categorized into two main approaches: one leverages LLMs for KGQA tasks, while the other is based on semantic parsing methods.
+
+# 6.1 LLMs for KGQA
+
+KGs, as structured knowledge bases, can effectively supplement the factual knowledge of LLMs and reduce hallucinations (Pan et al., 2024). Existing KG-enhanced approaches are primarily divided into two categories: inference-based and training-
+
+based methods. Inference-based models, such as StructGPT (Jiang et al., 2023a), ToG (Sun et al., 2024), and KnowAgent (Zhu et al., 2024), facilitate multi-round interactions between LLMs and KGs, treating LLMs as agents. Although this approach effectively leverages LLMs' reasoning abilities, the lack of sufficient understanding of relationships or entities within the KG by the LLM further amplifies hallucinations, leading to incorrect reasoning results. Training-based methods aim to integrate the knowledge of KG into LLMs. RoG (Luo et al., 2024a) samples the reasoning paths from the graph, injects relational information from the KG into the LLM, and generates reasoning paths to retrieve useful information from the KG for inference. GNN-RAG (Mavromatis and Karypis, 2024) uses a lightweight GNN to effectively extract information from the KG to assist LLM reasoning. Training-based methods, in order to ensure coverage of correct reasoning paths, tend to generate additional noise paths, which may affect subsequent LLM reasoning.
+
+# 6.2 Semantic Parsing
+
+In the KGQA task, semantic parsing (SP) aims to convert natural language questions into structured logical forms, such as lambda-DCS (Liang, 2013), SPARQL queries (Das et al., 2021), graph queries (Yih et al., 2015; Lan and Jiang, 2020), and s-expressions (Gu et al., 2021). Early methods employed grammar-based parsers (Mitra et al., 2022; Sun et al., 2020; Liang et al., 2017), whereas recent research has focused on leveraging pretrained language models to enhance semantic parsing (Scholak et al., 2021; Zhang et al., 2019). SPARQA (Sun et al., 2020) introduces skeleton grammar to represent the high-level structure of complex questions and incorporates BERT to improve the accuracy of semantic dependency relations. SR (Zhang et al., 2022) employs a trainable subgraph retriever that utilizes a dual-encoder model to expand reasoning paths and dynamically determine termination conditions for expansion. RNG-KBQA (Ye et al., 2022) combines a contrastive ranker with a generative model, leveraging high-confidence candidates to optimize the final composition of logical forms. These methods optimize latent space representations to improve question generalization. In contrast, our approach tackles scalability challenges in semantic parsing through an explicit phrase alignment and retrieval mechanism.
+
+# 7 Conclusion
+
+In this paper, we propose a two-stage fine-grained phrase alignment framework to address the issues of semantic gaps and redundant paths that arise in the integration of LLMs with KGs. By leveraging SPARQL, we establish a connection between natural language and graph structures, facilitating the precise and efficient retrieval of relevant information from the KG. Experimental results on two standard KGQA benchmarks validate the effectiveness of our approach, demonstrating that the application of semantic parsing effectively bridges the semantic gap and reasoning discrepancies between LLMs and KGs, leading to faithful and reliable inference.
+
+# Limitations
+
+Although we have demonstrated the effectiveness and stability of our framework across various experimental settings, several limitations remain. First, our approach depends on Freebase, which may limit its applicability to knowledge bases with different structures and entity distributions, such as Wikidata and DBpedia. Future work should explore its generalization capability across different knowledge bases. Second, we employ rule-based substitution for retrieved SPARQL phrases, a process that could be improved in terms of accuracy and adaptability through generative models. Lastly, we use SPARQL as the LF for fine-grained alignment; extending LF to S-expression and lambda-DCS could further validate the framework's generality and compatibility.
+
+# Ethics Statement
+
+This work does not pose any ethical issues. All the data and models used in this paper are publicly available and are used under following licenses: Creative Commons BY 4.0 License, MIT License, Apache license 2.0, Llama 3.1 Community License Agreement, Llama 3.2 Community License Agreement.
+
+# Acknowledgments
+
+This work was supported by the Natural Science Foundation of China (No. 62476134).
+
+# References
+
+Jonathan Berant and Percy Liang. 2014. Semantic parsing via paraphrasing. In Proceedings of the 52nd Annual Meeting of the Association for Computational
+
+Linguistics, ACL 2014, June 22-27, 2014, Baltimore, MD, USA, Volume 1: Long Papers, pages 1415-1425. The Association for Computer Linguistics.  
+Kurt D. Bollacker, Colin Evans, Praveen K. Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the ACM SIGMOD International Conference on Management of Data, SIGMOD 2008, Vancouver, BC, Canada, June 10-12, 2008, pages 1247-1250. ACM.  
+Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 2318-2335. Association for Computational Linguistics.  
+Sitao Cheng, Ziyuan Zhuang, Yong Xu, Fangkai Yang, Chaoyun Zhang, Xiaoting Qin, Xiang Huang, Ling Chen, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan, and Qi Zhang. 2024. Call me when necessary: Llms can efficiently and faithfully reason over structured environments. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 4275-4295. Association for Computational Linguistics.  
+Jiaxi Cui, Munan Ning, Zongjian Li, Bohua Chen, Yang Yan, Hao Li, Bin Ling, Yonghong Tian, and Li Yuan. 2024. Chatlaw: A multi-agent collaborative legal assistant with knowledge graph enhanced mixture-of-experts large language model. Preprint, arXiv:2306.16092.  
+Rajarshi Das, Manzil Zaheer, Dung Thai, Ameya Godbole, Ethan Perez, Jay Yoon Lee, Lizhen Tan, Lazaros Polymenakos, and Andrew McCallum. 2021. Case-based reasoning for natural language queries over knowledge bases. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 9594-9611. Association for Computational Linguistics.  
+Li Dong and Mirella Lapata. 2016. Language to logical form with neural attention. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics.  
+Yu Gu, Sue Kase, Michelle Vanni, Brian M. Sadler, Percy Liang, Xifeng Yan, and Yu Su. 2021. Beyond I.I.D.: three levels of generalization for question answering on knowledge bases. In WWW '21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021, pages 3477-3488. ACM / IW3C2.
+
+Gaole He, Yunshi Lan, Jing Jiang, Wayne Xin Zhao, and Ji-Rong Wen. 2021. Improving multi-hop knowledge base question answering by learning intermediate supervision signals. In WSDM '21, The Fourteenth ACM International Conference on Web Search and Data Mining, Virtual Event, Israel, March 8-12, 2021, pages 553-561. ACM.  
+Heewon Jang, Yeongtaek Oh, Seunghee Jin, Haemin Jung, Hyesoo Kong, Dokyung Lee, Dongkyu Jeon, and Wooju Kim. 2017. Kbqa: Constructing structured query graph from keyword query for semantic search. In Proceedings of the International Conference on Electronic Commerce, pages 1-8.  
+Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Xin Zhao, and Ji-Rong Wen. 2023a. Structgpt: A general framework for large language model to reason over structured data. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 9237-9251. Association for Computational Linguistics.  
+Jinhao Jiang, Kun Zhou, Wayne Xin Zhao, Yang Song, Chen Zhu, Hengshu Zhu, and Ji-Rong Wen. 2024. Kg-agent: An efficient autonomous agent framework for complex reasoning over knowledge graph. CoRR, abs/2402.11163.  
+Jinhao Jiang, Kun Zhou, Xin Zhao, and Ji-Rong Wen. 2023b. Unikgqa: Unified retrieval and reasoning for solving multi-hop question answering over knowledge graph. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. 2023. Decomposed prompting: A modular approach for solving complex tasks. In *The Eleventh International Conference on Learning Representations*, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Yunshi Lan and Jing Jiang. 2020. Query graph generation for answering multi-hop complex questions from knowledge bases. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 969-974. Association for Computational Linguistics.  
+Shiyu Li, Yang Tang, Shizhe Chen, and Xi Chen. 2024. Conan-embedding: General text embedding with more and better negative samples. Preprint, arXiv:2408.15710.  
+Chen Liang, Jonathan Berant, Quoc V. Le, Kenneth D. Forbus, and Ni Lao. 2017. Neural symbolic machines: Learning semantic parsers on freebase with weak supervision. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers, pages 23-33. Association for Computational Linguistics.
+
+Percy Liang. 2013. Lambda dependency-based compositional semantics. CoRR, abs/1309.4408.  
+Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. 2024a. Reasoning on graphs: Faithful and interpretable large language model reasoning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.  
+Linhao Luo, Zicheng Zhao, Chen Gong, Gholamreza Haffari, and Shirui Pan. 2024b. Graph-constrained reasoning: Faithful reasoning on knowledge graphs with large language models. CoRR, abs/2410.13080.  
+Costas Mavromatis and George Karypis. 2024. GNN-RAG: graph neural retrieval for large language model reasoning. CoRR, abs/2405.20139.  
+Meta. 2024. Build the future of ai with meta llama 3.  
+Sayantan Mitra, Roshni R. Ramnani, and Shubhashis Sengupta. 2022. Constraint-based multi-hop question answering with knowledge graph. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track, NAACL 2022, Hybrid: Seattle, Washington, USA + Online, July 10-15, 2022, pages 280-288. Association for Computational Linguistics.  
+OpenAI. 2022. Introducing chatgpt.  
+OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.  
+OpenAI. 2024. Hello gpt-4o.  
+Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. 2024. Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering, 36(7):3580-3599.  
+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, 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. Preprint, arXiv:2412.15115.  
+Nils Reimers and Iryna Gurevych. 2019. 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.  
+Stephen E. Robertson and Steve Walker. 1994. Some simple effective approximations to the 2-poisson
+
+model for probabilistic weighted retrieval. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. Dublin, Ireland, 3-6 July 1994 (Special Issue of the SIGIR Forum), pages 232-241. ACM/Springer.  
+Torsten Scholak, Nathan Schucher, and Dzmitry Bah-danau. 2021. PICARD: parsing incrementally for constrained auto-regressive decoding from language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November; 2021, pages 9895-9901. Association for Computational Linguistics.  
+Peter Shaw, Philip Massey, Angelica Chen, Francesco Piccinno, and Yasemin Altun. 2019. Generating logical forms from graph representations of text and entities. 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 95-106. Association for Computational Linguistics.  
+Jiaxin Shi, Shulin Cao, Lei Hou, Juanzi Li, and Hanwang Zhang. 2021. Transfernet: An effective and transparent framework for multi-hop question answering over relation graph. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 4149-4158. Association for Computational Linguistics.  
+Yuan Sui, Yufei He, Nian Liu, Xiaoxin He, Kun Wang, and Bryan Hooi. 2024. Fidelis: Faithful reasoning in large language model for knowledge graph question answering. CoRR, abs/2405.13873.  
+Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel M. Ni, Heung-Yeung Shum, and Jian Guo. 2024. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. Open-Review.net.  
+Yawei Sun, Lingling Zhang, Gong Cheng, and Yuzhong Qu. 2020. SPARQA: skeleton-based semantic parsing for complex questions over knowledge bases. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 8952-8959. AAAI Press.  
+Alon Talmor and Jonathan Berant. 2018. The web as a knowledge-base for answering complex questions. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,
+
+Volume 1 (Long Papers), pages 641-651, New Orleans, Louisiana. Association for Computational Linguistics.  
+Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023. Can chatgpt replace traditional KBQA models? an in-depth analysis of the question answering performance of the GPT LLM family. In The Semantic Web - ISWC 2023 - 22nd International Semantic Web Conference, Athens, Greece, November 6-10, 2023, Proceedings, Part I, volume 14265 of Lecture Notes in Computer Science, pages 348-367. Springer.  
+Denny Vrandecic and Markus Krötzsch. 2014. Wiki-data: a free collaborative knowledgebase. Commun. ACM, 57(10):78-85.  
+Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. 2023. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. 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 2609-2634. Association for Computational Linguistics.  
+Junde Wu, Jiayuan Zhu, and Yunli Qi. 2024. Medical graph RAG: towards safe medical large language model via graph retrieval-augmented generation. CoRR, abs/2408.04187.  
+Xi Ye, Semih Yavuz, Kazuma Hashimoto, Yingbo Zhou, and Caiming Xiong. 2022. RNG-KBQA: generation augmented iterative ranking for knowledge base question answering. 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 6032-6043. Association for Computational Linguistics.  
+Wen-tau Yih, Ming-Wei Chang, Xiaodong He, and Jianfeng Gao. 2015. Semantic parsing via staged query graph generation: Question answering with knowledge base. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 1: Long Papers, pages 1321-1331. The Association for Computer Linguistics.  
+Wen-tau Yih, Matthew Richardson, Christopher Meek, Ming-Wei Chang, and Jina Suh. 2016. The value of semantic parse labeling for knowledge base question answering. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 2: Short Papers. The Association for Computer Linguistics.  
+Donghan Yu, Sheng Zhang, Patrick Ng, Henghui Zhu, Alexander Hanbo Li, Jun Wang, Yiqun Hu,
+
+William Yang Wang, Zhiguo Wang, and Bing Xiang. 2023. Decaf: Joint decoding of answers and logical forms for question answering over knowledge bases. In *The Eleventh International Conference on Learning Representations*, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Haoyu Zhang, Jingjing Cai, Jianjun Xu, and Ji Wang. 2019. Complex question decomposition for semantic parsing. 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 4477-4486. Association for Computational Linguistics.  
+Jing Zhang, Xiaokang Zhang, Jifan Yu, Jian Tang, Jie Tang, Cuiping Li, and Hong Chen. 2022. Subgraph retrieval enhanced model for multi-hop knowledge base question answering. 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 5773-5784. Association for Computational Linguistics.  
+Ruilin Zhao, Feng Zhao, Long Wang, Xianzhi Wang, and Guandong Xu. 2024. Kg-cot: Chain-of-thought prompting of large language models over knowledge graphs for knowledge-aware question answering. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI 2024, Jeju, South Korea, August 3-9, 2024, pages 6642-6650. ijcai.org.  
+Yuqi Zhu, Shuofei Qiao, Yixin Ou, Shumin Deng, Ningyu Zhang, Shiwei Lyu, Yue Shen, Lei Liang, Jinjie Gu, and Huajun Chen. 2024. Knowa-gent: Knowledge-augmented planning for llm-based agents. CoRR, abs/2403.03101.
+
+<table><tr><td>Datasets</td><td>#Train</td><td>#Test</td><td>Max #hop</td></tr><tr><td>WebQSP</td><td>2,826</td><td>1,628</td><td>2</td></tr><tr><td>CWQ</td><td>27,639</td><td>3,531</td><td>4</td></tr></table>
+
+Table 5: Statistics of datasets.  
+
+<table><tr><td>Dataset</td><td>#Ans = 1</td><td>2 ≥ #Ans ≤ 4</td><td>5 ≥ #Ans ≤ 9</td><td>#Ans ≥ 10</td></tr><tr><td>WebQSP</td><td>51.2%</td><td>27.4%</td><td>8.3%</td><td>12.1%</td></tr><tr><td>CWQ</td><td>70.6%</td><td>19.4%</td><td>6%</td><td>4%</td></tr></table>
+
+Table 6: Statistics of the number of answers for questions in WebQSP and CWQ.  
+
+<table><tr><td>Dataset</td><td>1 hop</td><td>2 hop</td><td>≥3 hop</td></tr><tr><td>WebQSP</td><td>65.49 %</td><td>34.51%</td><td>0.00%</td></tr><tr><td>CWQ</td><td>40.91 %</td><td>38.34%</td><td>20.75%</td></tr></table>
+
+Table 7: Statistics of the question hops in WebQSP and CWQ.
+
+# A Datasets
+
+We utilize two standard datasets, WebQSP and CWQ. To ensure fairness, we adopt the same training and test splits as in previous studies. The detailed statistics of the datasets are presented in Table 5. Both WebQSP and CWQ are based on Freebase. In our approach, we employ the complete Freebase as the knowledge graph for information retrieval.
+
+# B Baselines
+
+We compared FGSP against 14 baselines, which can be categorized into three groups: (1) LLM-only methods, (2) inference-based methods, and (3) training-based methods. The specific details of these approaches are as follows.
+
+# B.1 LLM-only method
+
+The results of lama3.1-8b (Meta, 2024) and Qwen2-7B (Qwen et al., 2024) are from Luo et al. (2024b). The results of ChatGPT (OpenAI, 2022) and GPT-4 (OpenAI, 2023) are from Jiang et al. (2024).
+
+# B.2 Inference-based methods
+
+- StructGPT (Jiang et al., 2023a) enhances large language models' reasoning ability on structured data by integrating the iterative reading
+
+reasoning (IRR) method with structured data interfaces.
+
+- Readi (Cheng et al., 2024) enables efficient and faithful reasoning in structured environments by allowing large language models (LLMs) to generate and edit reasoning paths.  
+- ToG (Sun et al., 2024) tightly integrates large language models with knowledge graphs, leveraging graphs for deep reasoning and employing iterative beam search execution to improve reasoning capability and interpretability.  
+- KG-CoT (Zhao et al., 2024) enhances the knowledge reasoning ability of LLMs by incorporating small-scale stepwise graph reasoning models and leveraging knowledge graphs (KGs), thereby improving LLMs' performance on knowledge-intensive question-answering tasks without requiring fine-tuning.
+
+# B.3 Training-based methods
+
+- NSM (He et al., 2021) proposes a multi-hop knowledge base question-answering method based on a teacher-student framework, improving reasoning ability by learning intermediate supervision signals and generating more reliable intermediate entity distributions through bidirectional reasoning mechanisms.  
+- GraftNet (Shi et al., 2021) is an efficient and transparent framework for multi-hop question answering that infers answers by propagating entity scores across entities, supporting the processing of both labeled and textual entity relations within a unified framework.  
+- SR+NSM (Zhang et al., 2022) decouples the retrieval and reasoning processes to enhance the performance of embedding-based KBQA models.  
+- UniKGQA (Jiang et al., 2023b) addresses multi-hop knowledge graph question answering by integrating semantic matching and information propagation modules while unifying retrieval and reasoning in model architecture and parameter learning.  
+- DECAF (Yu et al., 2023) improves the accuracy of knowledge base question-answering tasks by jointly generating logical forms and
+
+<table><tr><td>Type</td><td>Split results</td><td>Remark</td></tr><tr><td>over-aggregation</td><td>[&quot;The character in Cars played by the actor who played Porco Rosso.&quot;, 
+?&quot;c ns:filmActor.dubbing_performance?k .&quot;] 
+?k ns:film.dubbing_performance.character ns:m.0nfpg4s . 
+?c ns:filmActor.film ?y 
+?y ns:film-performance.character ?x&quot;]</td><td>No separation of questions and SPARQL at all</td></tr><tr><td rowspan="3">over-fragmentation</td><td>[&quot;What character&quot;, &quot;character ?x&quot;]</td><td>The split result is not a triple</td></tr><tr><td>[&quot;Porco Rosso&quot;, &quot;m.0nfpg4s&quot;]</td><td>Split to only a single entity</td></tr><tr><td>[&quot;in Cars&quot;, &quot;?y ns:film-performance.film ns:m.03q0r1 .&quot;&quot;]</td><td>The split result cannot support complete semantics</td></tr></table>
+
+Table 8: Two kinds of mistakes in the phrase-level segmentation: a)over-encapsulation b)excessive fragmentation  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Parameter</td><td rowspan="2">Number</td><td colspan="2">Average length</td></tr><tr><td>NL clauses</td><td>Question</td></tr><tr><td>GPT4o</td><td>-</td><td>43066</td><td>6.12</td><td></td></tr><tr><td>gpt4o-mini</td><td>-</td><td>42373</td><td>6.01</td><td></td></tr><tr><td rowspan="5">Qwen2.5</td><td>14B</td><td>40310</td><td>6.30</td><td rowspan="5">13.21</td></tr><tr><td>7B</td><td>41330</td><td>5.96</td></tr><tr><td>3B</td><td>41395</td><td>6.24</td></tr><tr><td>1.5B</td><td>46943</td><td>6.72</td></tr><tr><td>0.5B</td><td>23393</td><td>7.13</td></tr><tr><td>LLAMA 3.1</td><td>8B</td><td>44135</td><td>6.38</td><td></td></tr><tr><td rowspan="2">LLAMA 3.2</td><td>3B</td><td>58886</td><td>8.63</td><td></td></tr><tr><td>1B</td><td>61623</td><td>6.53</td><td></td></tr></table>
+
+Table 9: Scale details of phrase-level question-SPARQL mapping libraries built by different models
+
+direct answers, leveraging the strengths of both approaches while simplifying model adaptation across different datasets.
+
+- ROG (Luo et al., 2024a) combines large language models (LLMs) and knowledge graphs (KGs), employing a plan-retrieveresason framework to generate faithful and interpretable reasoning results.
+
+# C Error in Phrase-level Segmentation
+
+In phrase-level segmentation, we identify two issues that arise when directly relying on LLMs for phrase-level segmentation: over-encapsulation and excessive fragmentation. These two issues are presented in detail in the table 8.
+
+# D Details of the Phrase-level Question-SPARQL Mapping Libraries
+
+The details of the phrase-level question-SPARQL mapping libraries are presented in the table 9. The constructed library comprises approximately 40,000 phrase pairs, with an average natural language phrase length of around six words and an average original question length of 13.21 words.
+
+# E Prompt
+
+We present the prompts used throughout our framework.
+
+The prompt employed for the phrase pair decomposition process is shown in table 10. We adopt a few-shot approach to ensure that the phrase pairs generated by the LLM adhere to our predefined format. The prompt includes preprocessed Basic Graph Patterns (BGPs), the original SPARQL query, and the complete question. Notably, the natural language phrases derived from the question are not sub-questions but rather declarative statements.
+
+For the online stage, the prompt used for LLM-based question decomposition is presented in table 10. To minimize constraints on the LLM's reasoning process, we refrain from designing overly complex prompts for question decomposition.
+
+Prompt for Phrase-level Segmentation Given a question, a SPARQL query, mid and its entity name in the SPARQL statement, and the main reasoning path implied by SPARQL, split the question and SPARQL into pairs of clauses based on the given reasoning path. The parts that do not belong to the reasoning path also need to be split.
+
+Examples
+
+Few-shot Question
+
+<Question>
+
+SPARQL:
+
+<SPARQL>
+
+mention entity:
+
+<mention entity>
+
+inference chain:
+
+<inference chain>
+
+pairs:
+
+# Demonstration Example
+
+Question: What is the national currency of the country where Bajan is spoken?
+
+SPARQL:
+
+PREFIX ns: <http://rdfs.freerefbase.com/ns/>
+
+SELECT DISTINCT ?x
+
+WHERE{
+
+FILTER(?x != ?c)
+
+FILTER (!isLiteral(?x) OR lang(?x) = " OR langMatches(lang(?x), 'en'))
+
+?c ns:location countrylanguages_spoken ns:m.03xx69 .
+
+?c ns:location countrycurrency_used ?x .
+
+mention entity:
+
+m.03xx69: Bajan Language
+
+inference chain:
+
+?c ns:location countrylanguages_spoken ns:m.03xx69 .
+
+?c ns:location countrycurrency_used ?x .
+
+pairs:
+
+[["Countries that speak Barbadian", "?c ns:location.countrieslanguages_spoken ns:m.03xx69."],
+
+["The country's currency",?"c ns:location countrycurrency_used ?x ."]
+
+# Prompt for Question Decomposition
+
+Decompose the given problem into subproblems. There should be no nesting between subproblems.
+
+Examples:
+
+<Few-shot>
+
+Question
+
+<Question>
+
+# Demonstration Example
+
+Question: Which city of residence for Tom Hanks was the birthplace of Elon Musk?
+
+Sub-questions: ["Which city did Tom Hanks live in?", "Where was Elon Musk born?]
+
+Table 10: Detailed prompts for modules of FGSP

2025/From Phrases to Subgraphs_ Fine-Grained Semantic Parsing for Knowledge Graph Question Answering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:876435686b1c4029382f897ff23504df6bbf9c4985c3115df51a0205d88686d2
+size 589252

2025/From Specific-MLLMs to Omni-MLLMs_ A Survey on MLLMs Aligned with Multi-modalities/ac867663-37ff-4910-9a55-16ebe121303b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46b931012c36ac0c160fdac52e124700da8cc597d352498ce7032116f75da646
+size 1390896

2025/From Specific-MLLMs to Omni-MLLMs_ A Survey on MLLMs Aligned with Multi-modalities/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c23972292bff36b06c16c1a7cb451919981f7c8747d23fc2e42ad7c6d9532fad
+size 1532285

2025/Full-Step-DPO_ Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning/e3c1942a-6072-41ae-9020-bff00ee3eef1_content_list.json

@@ -0,0 +1,1788 @@
+[
+    {
+        "type": "text",
+        "text": "Full-Step-DPO: Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning",
+        "text_level": 1,
+        "bbox": [
+            124,
+            84,
+            872,
+            124
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Huimin Xu $^{1}$ , Xin Mao $^{1}$ , Feng-Lin Li $^{2}$ , Xiaobao Wu $^{1*}$  Wang Chen $^{2}$ , Wei Zhang $^{3}$ , Anh Tuan Luu $^{1*}$",
+        "bbox": [
+            242,
+            135,
+            757,
+            168
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Nanyang Technological University, Singapore  $^{2}$ Shopee Pte. Ltd, Singapore,  $^{3}$ SEA Group, Singapore {huimin.xu, xin.mao, xiaobao.wu, anhtuan.luu}@ntu.edu.sg {fenglin.li, chen.wang}@shopee.com, terry.zhang@sea.com",
+        "bbox": [
+            216,
+            171,
+            784,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Direct Preference Optimization (DPO) often struggles with long-chain mathematical reasoning. Existing approaches, such as Step-DPO, typically improve this by focusing on the first erroneous step in the reasoning chain. However, they overlook all other steps and rely heavily on humans or GPT-4 to identify erroneous steps. To address these issues, we propose Full-Step-DPO, a novel DPO framework tailored for mathematical reasoning. Instead of optimizing only the first erroneous step, it leverages step-wise rewards from the entire reasoning chain. This is achieved by training a self-supervised process reward model, which automatically scores each step, providing rewards while avoiding reliance on external signals. Furthermore, we introduce a novel stepwise DPO loss, which dynamically updates gradients based on these step-wise rewards. This endows stronger reasoning capabilities to language models. Extensive evaluations on both in-domain and out-of-domain mathematical reasoning benchmarks across various base language models, demonstrate that Full-Step-DPO achieves superior performance compared to state-of-the-art baselines<sup>1</sup>.",
+        "bbox": [
+            141,
+            282,
+            460,
+            652
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            659,
+            258,
+            674
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models (LLMs) have attracted massive interest due to their remarkable capabilities across various tasks (Kaddour et al., 2023; Song et al., 2023; Wang et al., 2023a; Zheng et al., 2024; Wang et al., 2023b; Pan et al., 2024; Wu et al., 2024b,c,d,a). However, they commonly encounter difficulties when tackling complex and symbolic multi-step reasoning, particularly in mathematical problem reasoning (Lightman et al., 2023; Huang et al., 2023; Wu, 2025). To improve the mathematical reasoning ability, some studies use Direct Preference Optimization (DPO) (Rafailov et al., 2024)",
+        "bbox": [
+            112,
+            684,
+            489,
+            876
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/fd81edc99a96354f12a23b158cd2d1f9163d807962eb26ed65d1cf508162317f.jpg",
+        "image_caption": [
+            "DPO"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            536,
+            297,
+            863,
+            338
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1a7eddef3220f336d8fe0af692f35ea07a98f1c0a020f3f7959f6fb50ff6a469.jpg",
+        "image_caption": [
+            "Step-DPO"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            529,
+            369,
+            863,
+            429
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/54a3d6658f15b77fce39e517a17da27e88133fe06185115c86afd845f199bab1.jpg",
+        "image_caption": [
+            "Full-Step-DPO",
+            "Figure 1: Comparison between DPO, Step-DPO, and our Full-Step-DPO. DPO operates on solution-wise preference data. Step-DPO advances to step-wise data but optimizes only a single step. Full-Step-DPO optimizes all steps with a novel step-wise DPO loss, effectively enhancing the model's reasoning capability."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            532,
+            467,
+            867,
+            531
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "with solution-wise preference data but find its benefit limited (Pal et al., 2024; Xu et al., 2024; Jiao et al., 2024). Recent works attribute this limitation to DPO's inability to perform process supervision and instead build preference data based on reasoning steps rather than entire solutions (Lai et al., 2024; Chen et al., 2024; Xie et al., 2024a; Lu et al., 2024). For example, Step-DPO (Lai et al., 2024) focuses on optimizing only the first erroneous step in the reasoning chain, demonstrating notable improvements.",
+        "bbox": [
+            507,
+            663,
+            884,
+            838
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, despite their improvements, these existing methods face the following limitations: (1) Some focus solely on the first erroneous step and ignore all other useful steps in the reasoning chain (Lai et al., 2024), as shown in Figure 1.",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding Authors.",
+        "bbox": [
+            137,
+            882,
+            290,
+            895
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1Our code, data, and models are available at https:// github.com/Anna7355/Full-Step-DPO.",
+        "bbox": [
+            115,
+            895,
+            487,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "24343",
+        "bbox": [
+            473,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2025, pages 24343-24356 July 27 - August 1, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "As a result, they fail to fully optimize the reasoning chains, leading to suboptimal performance. (2) Their loss function still follows the early vanilla DPO in a solution-wise approach (Lu et al., 2024; Xie et al., 2024a). Consequently, it cannot directly leverage rewards in a step-wise fashion for learning. (3) They heavily rely on costly and resource-intensive annotations from GPT-4 or humans to detect erroneous steps (Lai et al., 2024; Lightman et al., 2023), significantly limiting their practicality.",
+        "bbox": [
+            110,
+            84,
+            489,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the above limitations, we propose Full-Step-DPO, a novel DPO framework for mathematical reasoning. As illustrated in Figure 1, unlike vanilla DPO, which operates solution-wise, or Step-DPO, which focuses solely on the first erroneous step, Full-Step-DPO utilizes each step in the entire reasoning chain and optimizes them using step-wise rewards. We first train a Process Reward Model (PRM) (Lightman et al., 2023; Wang et al., 2023c) in a self-supervised way, utilizing data generated by the model itself. This approach enables the PRM to automatically score each step in the reasoning chain, eliminating the reliance on external annotations such as GPT-4 or humans. Then, we propose a novel Step-wise DPO Loss, which employs dynamic gradient updates to optimize each step based on its corresponding reward. This approach shifts the optimization focus from solution-wise to step-wise, enabling the policy model to achieve superior reasoning capabilities.",
+        "bbox": [
+            115,
+            246,
+            489,
+            565
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments on both in-domain and out-of-domain mathematical reasoning datasets with four widely used backbone LLMs. Experimental results demonstrate that our Full-Step DPO consistently outperforms the DPO and Step-DPO baselines, validating its effectiveness in enhancing reasoning performance. Our contributions can be summarized as follows:",
+        "bbox": [
+            112,
+            567,
+            489,
+            695
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose the Full-Step-DPO framework with a novel step-wise DPO loss that dynamically adjusts each step's gradient based on its reward, enabling step-wise optimization rather than solution-wise and enhancing reasoning ability.",
+            "- We train a self-supervised PRM to provide stepwise rewards for preference learning and explore a more efficient approach for automatically constructing PRM training data.",
+            "- Extensive experiments on widely used mathematical benchmarks and base language models showcase the remarkable effectiveness of our method."
+        ],
+        "bbox": [
+            114,
+            705,
+            487,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            665,
+            98
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Mathematical Reasoning Mathematical reasoning task is one of the most challenging tasks for LLMs. Various approaches have been explored to improve or elicit the mathematical reasoning ability of LLMs. A number of approaches have either continually pre-trained the base model on a vast of mathematical datasets (Azerbayev et al., 2023; Shao et al., 2024) or used supervised fine-tuning with substantial synthetic datasets distilled from cutting-edge models (Luo et al., 2023; Yu et al., 2023b; Mitra et al., 2024; Xu et al., 2024). Another line of work focuses on enhancing test-time computation by generating multiple solutions, developing separate reward models at either the outcome or process level to rerank these solutions (Cobbe et al., 2021a; Lightman et al., 2023), or employing decoding strategies guided by the reward model (Yu et al., 2023a; Xie et al., 2024b; Wang et al., 2023c; Wu et al., 2024e). In addition, Reinforcement Learning's potential in general domains, demonstrated by Achiam et al. (2023) and Touvron et al. (2023), some studies have explored its use in mathematical reasoning (Wang et al., 2023c; Mitra et al., 2024; Pal et al., 2024).",
+        "bbox": [
+            507,
+            121,
+            884,
+            507
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Preference Learning Recently, preference learning (Ethayarajh et al., 2024; Mao et al., 2024b,a) has attracted significant attention due to its ability to align with human preferences and distinguish between positive and negative examples. While these methods, like DPO (Rafailov et al., 2024), have proven effective in general domains, it offers only marginal benefits for mathematical reasoning (Pal et al., 2024). Some works (Chen et al., 2024; Lai et al., 2024) suggest that DPO's focus on coarse solution-wise preferences makes it less effective at correcting errors in multi-step reasoning, hindering reasoning improvement. Therefore, Step-DPO (Lai et al., 2024) was proposed, which first identifies the first erroneous step, and then optimizes only this erroneous step along with the corresponding correct one. Although this approach enhances mathematical reasoning capabilities, it totally overlooks the other steps in long-chain reasoning, which also provide valuable information and should not be completely disregarded. Building on this consideration, we propose Full-Step-DPO, which fully accounts for each step by dynamically optimizing all steps in the reasoning process.",
+        "bbox": [
+            507,
+            535,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "24344",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Step-wise Supervision Recent findings by Lightman et al. (2023) suggest that step-wise supervision outperforms outcome-wise, due to the provision of more detailed feedback. However, training a PRM requires either costly manual annotation (Lightman et al., 2023) or significant computational resources (Khalifa et al., 2023; Wang et al., 2023c), which hinders the advancement and practical application of PRM. Therefore, in this paper, we aim to build a PRM for mathematical reasoning without relying on human annotation and with reduced computational resources. Additionally, we explore the effectiveness of the PRM in decoding and preference learning scenarios.",
+        "bbox": [
+            112,
+            84,
+            489,
+            309
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Full-Step DPO",
+        "text_level": 1,
+        "bbox": [
+            112,
+            319,
+            278,
+            337
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we elaborate the proposed Full-Step DPO framework. We begin by reviewing the background of previous DPO and Step-DPO. Then we introduce the novel Step-wise DPO Loss which optimizes with step-wise rewards, and the Process Reward Model which automatically generate these step-wise rewards. Finally we outline the complete training pipeline of our Full-Step-DPO.",
+        "bbox": [
+            112,
+            344,
+            489,
+            475
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            112,
+            483,
+            257,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "DPO. Direct Preference Optimization (DPO) (Rafailov et al., 2024) is one of the most popular preference optimization methods. Instead of learning an explicit reward model, DPO directly uses pair-wise preference data to optimize the policy model with an equivalent optimization objective. Specifically, given an input prompt  $x$ , and a preference data pair  $(y^w, y^l)$ , DPO aims to maximize the probability of the entire preferred solution  $y^w$  and minimize that of the dispreferred solution  $y^l$ . The optimization objective of DPO is:",
+        "bbox": [
+            112,
+            504,
+            489,
+            682
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {(x, y ^ {w}, y ^ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\right. \\right. \\\\ \\left. \\beta \\log \\frac {\\pi_ {\\theta} (y ^ {w} \\mid x)}{\\pi_ {\\mathrm {r e f}} (y ^ {w} \\mid x)} - \\beta \\log \\frac {\\pi_ {\\theta} (y ^ {l} \\mid x)}{\\pi_ {\\mathrm {r e f}} (y ^ {l} \\mid x)}\\right) \\Bigg ] \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            115,
+            688,
+            482,
+            753
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\pi_{\\theta}(\\cdot |x)$  is the policy model to be optimized,  $\\pi_{\\mathrm{ref}}(\\cdot |x)$  is the reference model,  $(x,y^{w},y^{l})$  are preference pairs,  $\\sigma$  is the sigmoid function,  $\\beta$  is a parameter controlling the deviation from the reference model.",
+        "bbox": [
+            112,
+            759,
+            489,
+            839
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Step-DPO. Although DPO performs well on chat benchmarks, it is less effective for long-chain reasoning tasks like mathematical problems. StepDPO (Lai et al., 2024) attributes this to DPO's",
+        "bbox": [
+            112,
+            856,
+            489,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "inability to consider the sequential nature of mathematical reasoning, as rejecting an entire dispreferred solution may inadvertently penalize correct preceding steps, introducing significant noise. To address this, Step-DPO optimizes only the first incorrect step. As shown in Figure 1, given a math problem and a series of initial correct reasoning steps  $\\{s_1,\\dots,s_{k - 1}\\}$ , Step-DPO aims to maximize the probability of the correct next step  $s_k^w$  and minimize the probability of the incorrect one  $s_k^l$ . Note that  $s_k^w$  and  $s_k^l$  refer to single steps, not all subsequent steps. The loss function used is still the vanilla DPO loss.",
+        "bbox": [
+            507,
+            84,
+            884,
+            292
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Step-wise DPO Loss",
+        "text_level": 1,
+        "bbox": [
+            507,
+            303,
+            714,
+            318
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We now introduce the novel Step-wise DPO loss, which performs step-wise optimization using stepwise rewards. Although the motivation behind Step-DPO is reasonable, focusing solely on optimizing the first erroneous step and neglecting the valuable information provided by other steps may not be optimal. Additionally, we contend that it is not truly a step-wise DPO, as it still relies on the standard solution-wise DPO loss and resembles more of a data construction method.",
+        "bbox": [
+            507,
+            323,
+            884,
+            482
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To address this, we modify the vanilla DPO loss to the step-wise DPO loss, dynamically weighting the gradients of each step based on its reward, thereby enabling true step-wise optimization. Let's start with the gradient of the loss function  $\\mathcal{L}_{\\mathrm{DPO}}$ . The gradient with respect to the parameters  $\\theta$  can be written as:",
+        "bbox": [
+            507,
+            483,
+            884,
+            595
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} = - \\beta \\mathbb {E} _ {(x, y ^ {w}, y ^ {l}) \\sim \\mathcal {D}} [ \\sigma (\\hat {r} _ {\\theta} (x, y ^ {l}) - \\hat {r} _ {\\theta} (x, y ^ {w})) \\\\ \\left[ \\nabla_ {\\theta} \\log \\pi_ {\\theta} (y ^ {w} \\mid x) - \\nabla_ {\\theta} \\log \\pi_ {\\theta} (y ^ {l} \\mid x) \\right] \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            510,
+            598,
+            878,
+            646
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\hat{r}_{\\theta}(x,y) = \\beta \\log \\frac{\\pi_{\\theta}(y|x)}{\\pi_{\\mathrm{ref}}(y|x)}$ . Intuitively, the gradient indiscriminately increases the likelihood of whole  $y^{w}$  and decreases the likelihood of whole  $y^{l}$ . To achieve dynamically weighting, we break  $\\nabla_{\\theta} \\log \\pi_{\\theta}(y \\mid x)$  into a step-wise form and weight the gradient as follows:",
+        "bbox": [
+            507,
+            653,
+            882,
+            751
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} = - \\beta \\mathbb {E} _ {(x, y ^ {w}, y ^ {l}) \\sim \\mathcal {D}} \\left[ \\sigma \\left(\\hat {r} _ {\\theta} (x, y ^ {l}) - \\hat {r} _ {\\theta} (x, y ^ {w})\\right) \\right. \\\\ \\left[ \\sum_ {i = 1} ^ {K ^ {w}} \\alpha_ {i} ^ {w} \\nabla_ {\\theta} \\log \\pi_ {\\theta} \\left(s _ {i} ^ {w} \\mid x, s _ {: <   i} ^ {w}\\right) \\right. \\\\ \\left. \\left. - \\sum_ {i = 1} ^ {K ^ {l}} \\alpha_ {i} ^ {l} \\nabla_ {\\theta} \\log \\pi_ {\\theta} \\left(s _ {i} ^ {l} \\mid x, s _ {: <   i} ^ {l}\\right) \\right] \\right] \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            755,
+            860,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $s_i$  represents the  $i$ -th reasoning step of the solution  $y$ ,  $s_{: < i}$  denotes all reasoning steps preceding  $s_i$ ,  $K$  is the total number of steps, and  $\\alpha_i$  is",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "24345",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the weight coefficient of  $s_i$ , calculated based on the reward of  $s_i$  as shown below:",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\alpha_ {i} = \\left\\{ \\begin{array}{l l} \\frac {e ^ {\\gamma r _ {s i}}}{\\sum_ {j} e ^ {\\gamma r _ {s j}}}, & s _ {i} \\in y ^ {w} \\\\ \\frac {e ^ {- \\gamma r _ {s i}}}{\\sum_ {j} e ^ {- \\gamma r _ {s j}}}, & s _ {i} \\in y ^ {l} \\end{array} \\right.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            188,
+            128,
+            411,
+            183
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $r_{s_i}$  is the reward of the step  $s_i$ , which will be introduced in the next subsection, and  $\\gamma$  is the temperature of the Softmax operation. It is important to note that the calculation of  $\\alpha_{i}$  differs between the preferred solution  $y^{w}$  and the dispreferred solution  $y^{l}$ . For preferred solutions, a higher reward indicates a greater likelihood of correct reasoning in that step, so the model should perform gradient ascent with greater intensity. Conversely, for dispreferred solutions, a lower reward suggests a higher chance of incorrect reasoning, and the model should apply gradient descent with greater intensity accordingly. This approach allows us to leverage all steps and adaptively adjust the weight of each step based on its probability of correctness, achieving true step-wise optimization.",
+        "bbox": [
+            112,
+            193,
+            489,
+            450
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Compared to Step-DPO methods that focus solely on a single step, our method optimizes all steps simultaneously, enabling better global optimization. Noted that as  $\\gamma \\rightarrow 0$ , all steps will have equal weights, making Full-Step-DPO equivalent to vanilla DPO.",
+        "bbox": [
+            112,
+            451,
+            489,
+            546
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Process Reward Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            561,
+            349,
+            575
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To obtain step-wise rewards, we train a Process Reward Model (PRM). The biggest challenge in training a PRM is constructing a process supervision dataset. Previous studies (Uesato et al., 2022; Lightman et al., 2023) utilize human annotators to obtain step-wise labels, which requires advanced annotator skills and is quite costly. Later, MathShepherd (Wang et al., 2023c) proposes using Monte Carlo estimation (Coulom, 2006) to automatically gather step-wise supervision, but it remained computationally expensive. In this section, we first examine the principles of Monte Carlo estimation, then present our simplified solution that significantly improves the efficiency of data construction.",
+        "bbox": [
+            112,
+            583,
+            489,
+            822
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Monte Carlo estimation. This approach assumes that the gold label  $y_{s_i}$  of a step  $s_i$  can be defined as the probability to deduce the correct answer  $a^*$ , and it includes both sampling and simulation phase. Specifically, given a math problem, it first",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "randomly samples  $M$  solutions, with each solution consisting of  $K$  reasoning steps  $\\{s_1,s_2,\\ldots ,s_K\\}$  and  $a$  represents the decoded answer from the last step  $s_K$ . Then, to estimate the quality of reasoning step  $s_i$  in a given solution, it simulates  $N$  subsequent reasoning processes from this step:  $\\{(s_{i + 1,j},\\dots ,s_{K,j})\\}_{j = 1}^{N}$ . The golden label for  $s_i$  is calculated as follows:",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ny _ {s _ {i}} = \\frac {\\sum_ {j = 1} ^ {N} \\mathbb {I} (a _ {j} = a ^ {*})}{N}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            601,
+            221,
+            789,
+            258
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $a_{j}$  is the decoded answer for the  $j$ -th simulated solution, and  $\\mathbb{I}$  is the indicator function that returns 1 if  $a_{j} = a^{*}$  and 0 otherwise. This two-stage approach is highly time-consuming, as it requires  $N$  simulations for each of  $K$  steps across all  $M$  solutions, resulting in a time complexity of  $O(MNK)$ .",
+        "bbox": [
+            507,
+            267,
+            884,
+            381
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our efficient approach. It is important to note that there is a trade-off between the sampling number  $M$  and the simulation number  $N$  when computational resources are limited. A larger  $M$  can provide more data for training the PRM, while a larger  $N$  can result in higher accuracy of the labels  $y_{i}$ . In this paper, we found that the trained PRM performs reasonably well even with  $N = 1$  when  $M$  is large, such as 32. This is likely because a larger  $M$  introduces more diversity into the training data, making the PRM more tolerant to slight reductions in data precision caused by the limited simulation number. This setting simplifies the PRM data construction by requiring only the sampling of  $M$  solutions without the need for simulation, significantly reducing computational resources and lowering the time complexity to  $O(M)$ . As a result, the gold label for step  $s_{i}$  can be simplified as follows:",
+        "bbox": [
+            507,
+            395,
+            884,
+            697
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ny _ {s _ {i}} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} a = a ^ {*} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            611,
+            699,
+            779,
+            740
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "then the PRM could be trained as shown below:",
+        "bbox": [
+            507,
+            745,
+            865,
+            760
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P R M}} = - \\sum_ {i = 1} ^ {K} \\left[ y _ {s _ {i}} \\log r _ {s _ {i}} + \\left(1 - y _ {s _ {i}}\\right) \\log \\left(1 - r _ {s _ {i}}\\right) \\right]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            507,
+            770,
+            890,
+            814
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $y_{s_i}$  is the golden label for  $s_i$ ,  $r_{s_i}$  is the sigmoid score assigned by the PRM. With the above PRM, we can automatically score each step in the reasoning chain, providing reward signals for the step-wise DPO loss and enabling step-wise optimization.",
+        "bbox": [
+            507,
+            825,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "24346",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/94f7b2ed08d392c2a986402851dfd3ff4a3527f706c0293679c0d21805beb02b.jpg",
+        "image_caption": [
+            "Step 1. PRM Training"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            110,
+            468,
+            170
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/a33da6cd2ff7d6cb191a43e4add57e3868a5fcf6518547b4bbe3470607c649a9.jpg",
+        "image_caption": [
+            "Step 2. Preference Building"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            137,
+            210,
+            473,
+            262
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/710bf4ed45dc2699222692f9fb06f316cf4093d99f12d37573a778d018d37a19.jpg",
+        "image_caption": [
+            "Step 3. Preference Learning",
+            "step-wise preferences"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            147,
+            303,
+            215,
+            321
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/f0c004655858efdac27a6ac493b1e17d6ca200aaeba9edee555a68455998f421.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            240,
+            307,
+            374,
+            332
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/15b3c65c5c05446c819f463c6d585c478edbc455c9f1dae83cedbd801acfff8b.jpg",
+        "image_caption": [
+            "Figure 2: The overall framework of Full-Step-DPO consists of three steps: (1) Training the PRM using the model itself and generated solutions. (2) Using the PRM to score and filter solutions to form preference data with step-wise rewards. (3) Training the policy model with the proposed step-wise DPO loss."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            386,
+            299,
+            455,
+            346
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Training Pipeline",
+        "text_level": 1,
+        "bbox": [
+            112,
+            481,
+            295,
+            495
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Following previous methods (Wang et al., 2023c; Shao et al., 2024), we adopt a standard training pipeline illustrated in Figure 2: (1) We begin by training a PRM with self-generated data, where higher reward values indicate a stronger likelihood of correct reasoning, while lower values suggest potential errors. (2) The trained PRM is then used to construct preference pairs with step-wise rewards. Specifically, we generate  $M$  solutions for each math problem, score each step of these solutions with the PRM to produce a reward sequence, and calculate the average reward across all steps as the overall reward for each solution. We select the top  $T$  correct solutions with the highest rewards and the bottom  $T$  incorrect solutions with the lowest rewards to form  $T^2$  step-wise preference pairs. (3) Finally, we update the policy model using the proposed step-wise DPO loss and the step-wise preference pairs, as described in Section 3.2.",
+        "bbox": [
+            112,
+            502,
+            487,
+            807
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During the inference, a well-trained PRM can guide the decoding process and enhance the model's performance. Therefore, in addition to the standard greedy decoding, we explore three alternative decoding methods: (1) Self-Consistency (SC) (Wang et al., 2022): given a problem in the test set, we sample  $K$  candidate solutions from the policy",
+        "bbox": [
+            112,
+            809,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "model. Instead of relying on the first decoded solution, we select the final answer based on majority voting over the answers provided by all sampled solutions. SC is a simple yet highly effective verification strategy. (2) Best-of-N (BoN): we similarly sample  $K$  candidate solutions, score them using the reward model, and select the highest-scoring solution as the final answer. Following previous work (Lightman et al., 2023; Wang et al., 2023c), we use the minimum score across all steps as the final score assigned to a solution by the PRM. (3) Stepwise Beam Search (SBS) (Yu et al., 2023a): the PRM provides feedback at each step, offering more fine-grained guidance. Specifically, for each step, we first sample  $b_{1}$  candidate subsequent steps, then score them using the PRM. The top  $b_{2}$  steps are retained, and decoding continues until  $b_{2}$  final solutions are reached. The detailed algorithm is provided in Appendix A.",
+        "bbox": [
+            507,
+            84,
+            884,
+            390
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            404,
+            653,
+            420
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            430,
+            712,
+            445
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Backbones. To comprehensively validate the effectiveness of our proposed method, we adopt four popular open-source LLMs as the backbone models: MetaMath-Mistral-7B (Yu et al., 2023b), Llama-3-8B (Touvron et al., 2023), DeepSeekMath-Base-7B (Shao et al., 2024) and Qwen2-7B (Bai et al., 2023). To improve these backbones' reasoning ability, Step-DPO (Lai et al., 2024) finetunes DeepSeekMath-Base-7B and Qwen2-7B on two open-source synthetic math datasets, MetaMath (Yu et al., 2023b) and MMIQC (Liu and Yao, 2024), resulting in DeepSeekMath-Base-SFT $^{2}$  and Qwen2-7B-SFT $^{3}$ , which greatly outperform their previous versions. Following Step-DPO, we further finetune Llama3-8B to produce Llama3-8B-SFT. MetaMath-Mistral-7B has already been finetuned on MetaMath, so no additional finetuning was performed.",
+        "bbox": [
+            507,
+            451,
+            882,
+            741
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baselines. For closed-source baselines, we compare our approach with OpenAI's GPT-3.5 and GPT-4 (Achiam et al., 2023). We also benchmarked our method against recent high-performing mathematical LLMs, including WizardMath (Luo et al., 2023), MetaMath (Yu et al., 2023b), InternLM-Math-7B (Ying et al., 2024), Qwen2-7B",
+        "bbox": [
+            507,
+            758,
+            882,
+            871
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://huggingface.co/xinlai/DeepSeekMath-Base-SFT",
+        "bbox": [
+            507,
+            881,
+            766,
+            906
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "3https://huggingface.co/xinlai/Qwen2-7B-SFT",
+        "bbox": [
+            532,
+            906,
+            858,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "24347",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Instruct(Bai et al., 2023), DeepSeekMath-Instruct (Shao et al., 2024), InternLM-Math-20B (Ying et al., 2024), and Llama-3-70B-Instruct (Touvron et al., 2023).",
+        "bbox": [
+            112,
+            84,
+            487,
+            146
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we compare it against DPO (Rafailov et al., 2024) and Step-DPO (Lai et al., 2024). Among these, Lai et al. (2024) publicly release DeepSeekMath-Base-SFT-Step-DPO  $^{4}$  and Qwen2-7B-SFT-Step-DPO  $^{5}$ , which we directly used for evaluation. Additionally, we trained MetaMath-Mistral-7B-Step-DPO and Llama-3-8B-SFT-Step-DPO using their publicly available code and dataset.",
+        "bbox": [
+            112,
+            149,
+            485,
+            294
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets. To ensure a fair comparison, we use the same training dataset  $^{6}$  provided by Step-DPO (Lai et al., 2024), which is synthesized from the training set of GSM8K (Cobbe et al., 2021b) and MATH (Hendrycks et al., 2021). Noted that we only use the problem prompts in this dataset and do not use the step labels marked by GPT-4.",
+        "bbox": [
+            112,
+            312,
+            485,
+            425
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For in-domain evaluation, we conduct experiments on GSM8K and MATH, which contain 1,319 and 5,000 test problems, respectively. We also evaluate on two more challenging out-of-domain (OOD) test sets OCWCourses (OCW) (Lewkowycz et al., 2022) and GaoKao2023 (GK2023) (Liao et al., 2024). OCW contains of 272 undergraduate-level STEM problems requiring multi-step reasoning for most questions, while GK2023 includes 385 mathematics problems from the 2023 Chinese higher education entrance exam, translated into English. Accuracy serves as the evaluation metric.",
+        "bbox": [
+            112,
+            426,
+            489,
+            634
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. During PRM training, we first randomly sample  $M = 32$  solutions for each math problem using Qwen2-7B-SFT and then label them as described in Section 3.3, resulting in the PRM training set. Then, we add a classification-head to Qwen2-7B-SFT and train it on the PRM training set for one epoch. The batch size is 256, and the learning rate is 5e-7.",
+        "bbox": [
+            112,
+            653,
+            487,
+            781
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To build preference learning datasets, we first sample  $M = 32$  solutions for each math problem. The trained PRM then scores each solution, and",
+        "bbox": [
+            112,
+            784,
+            487,
+            832
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "we select  $T = 4$  solutions with the highest average rewards and  $T = 4$  with the lowest average rewards to randomly form 16 preference pairs.",
+        "bbox": [
+            507,
+            84,
+            880,
+            131
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "During preference learning, the batch size is 64, the learning rate is  $5\\mathrm{e} - 7$ ,  $\\beta$  is 0.05, and the reward temperature  $\\gamma$  is 0.5. We use the AdamW (Loshchilov and Hutter, 2017) optimizer with a linear decay learning rate scheduler and only train one epoch. The warm-up ratio is 0.05.",
+        "bbox": [
+            507,
+            134,
+            882,
+            230
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "During the decoding phase, we conduct experiments with two settings for SC and BoN, using  $K = 5$  and  $K = 15$ . For Step-wise Beam Search, to ensure fair comparison, we test two configurations:  $b_{1} = 5$ ,  $b_{2} = 1$  (corresponding to  $K = 5$ ) and  $b_{1} = 5$ ,  $b_{2} = 3$  (corresponding to  $K = 15$ ). The sampling temperature is set to 0.8.",
+        "bbox": [
+            507,
+            231,
+            884,
+            344
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "All the experiments are conducted on a server equipped with 8 NVIDIA A100-80GB GPUs and 512GB of system RAM. The implementation frameworks are PyTorch (Paszke et al., 2017), DeepSpeed (Rasley et al., 2020), and Huggingface (Wolf et al., 2019).",
+        "bbox": [
+            507,
+            346,
+            882,
+            441
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Main Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            456,
+            660,
+            470
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1 provides a comprehensive comparison of various models on both MATH and GSM8K, including open-source and closed-source LLMs. We find that: (1) Consistent with previous studies (Pal et al., 2024), DPO exhibits notable instability. Its performance shows slight degradation on MetaMath-Mistral-7B and Llama-3-8B-SFT backbones, while the accuracy drops sharply to around  $20\\%$  on Qwen2-7B-SFT. It achieves a slight performance improvement only when applied to the DeepSeekMath-Base-SFT. (2) Step-DPO achieves only minimal improvements across all backbones, with gains generally around  $1\\%$  and, in some settings, even slight performance drops. We evaluate the publicly released Step-DPO model using its official script, and the results may differ slightly from those reported in the Step-DPO paper. Similar issues have also been observed by other researchers [7]. (3) Our Full-Step-DPO consistently outperforms Step-DPO across all backbones. Specifically, when applied to MetaMath-Mistral-7B and Llama-3-8B-SFT, our model achieves improvements of approximately  $2.3\\%$  to  $3.7\\%$ , while applied to the stronger backbones, DeepSeekMath-Base-SFT and Qwen2-7B-SFT, our method still delivers gains exceeding",
+        "bbox": [
+            507,
+            479,
+            882,
+            882
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{7}$ https://github.com/dvlab-research/Step-DPO/issues/2",
+        "bbox": [
+            509,
+            894,
+            867,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "4https://huggingface.co/xinlai/ DeepSeekMath-Base-SFT-Step-DPO 5https://huggingface.co/xinlai/ Qwen2-7B-SFT-Step-DPO 6https://huggingface.co/datasets/xinlai/ Math-Step-DPO-10K",
+        "bbox": [
+            112,
+            844,
+            440,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "24348",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f5597d86125272774b28509c9b27e1fde6c32d8bee1c731b3a317d6d22a6f8c3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>MATH (%)</td><td>GSM8K (%)</td></tr><tr><td>GPT-3.5</td><td>34.1</td><td>80.8</td></tr><tr><td>GPT-4</td><td>53.6</td><td>93.6</td></tr><tr><td>WizardMath</td><td>10.7</td><td>54.9</td></tr><tr><td>MetaMath</td><td>19.8</td><td>66.5</td></tr><tr><td>InternLM-Math-7B</td><td>34.6</td><td>78.1</td></tr><tr><td>Qwen2-7B-Instruct</td><td>49.6</td><td>82.3</td></tr><tr><td>DeepSeekMath-Instruct</td><td>46.8</td><td>82.9</td></tr><tr><td>InternLM-Math-20B</td><td>37.7</td><td>82.6</td></tr><tr><td>Llama-3-70B-Instruct</td><td>50.4</td><td>93.0</td></tr><tr><td>MetaMath-Mistral-7B</td><td>28.2</td><td>77.7</td></tr><tr><td>+ DPO</td><td>24.8 -3.4</td><td>70.7 -7.0</td></tr><tr><td>+ Step-DPO</td><td>28.9 +0.7</td><td>79.6 +1.9</td></tr><tr><td>+ Full-Step-DPO</td><td>30.5 +2.3</td><td>81.4 +3.7</td></tr><tr><td>Llama-3-8B-SFT</td><td>32.6</td><td>78.5</td></tr><tr><td>+ DPO</td><td>23.4 -9.2</td><td>62.3 -16.2</td></tr><tr><td>+ Step-DPO</td><td>31.8 -0.8</td><td>80.1 +1.6</td></tr><tr><td>+ Full-Step-DPO</td><td>35.0 +2.4</td><td>82.0 +3.5</td></tr><tr><td>DeepSeekMath-Base-SFT</td><td>51.7</td><td>86.4</td></tr><tr><td>+ DPO</td><td>51.7 -0</td><td>87.3 +0.9</td></tr><tr><td>+ Step-DPO</td><td>52.9 +1.2</td><td>86.6 +0.2</td></tr><tr><td>+ Full-Step-DPO</td><td>53.2 +1.5</td><td>87.9 +1.5</td></tr><tr><td>Qwen2-7B-SFT</td><td>53.9</td><td>88.3</td></tr><tr><td>+ DPO</td><td>20.0 -23.9</td><td>27.3 -61.0</td></tr><tr><td>+ Step-DPO</td><td>54.9 +1.0</td><td>88.4 +0.1</td></tr><tr><td>+ Full-Step-DPO</td><td>55.4 +1.5</td><td>89.3 +1.0</td></tr></table>",
+        "bbox": [
+            119,
+            82,
+            485,
+            463
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "$1\\%$ . These results clearly demonstrate the effectiveness of our proposed approach, which considers all steps in the reasoning process rather than focusing on solution-wise preferences or only a single step. A case study can be found in Appendix B.3.",
+        "bbox": [
+            112,
+            527,
+            489,
+            609
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Results on OOD Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            624,
+            359,
+            639
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To further demonstrate the superiority of Full-Step-DPO, we evaluate the models on OOD datasets GK2023 and OCW, as shown in Table 2. On these competition-level math problems, DPO and Step-DPO often exhibit performance degradation under various settings, while our Full-Step-DPO consistently achieves performance improvements. The only exception occurs on the OCW dataset with MetaMath-Mistral-7B, where Full-Step-DPO shows a slight  $0.8\\%$  drop in accuracy. However, this drop is notably smaller than  $3.0\\%$  with DPO and the  $3.7\\%$  with Step-DPO. These results demonstrate the superior stability and resilience of Full-Step-DPO, particularly in handling challenging mathematical reasoning tasks. More experimental results on additional datasets can be found in Appendix B.1.",
+        "bbox": [
+            112,
+            646,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/8333c4d896def8c7ccdb8336487f65682a7d4f196dc693534eef402751ba805a.jpg",
+        "table_caption": [
+            "Table 1: Performance comparison of various models on MATH and GSM8K with greedy decoding."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>GK2023 (%)</td><td>OCW (%)</td></tr><tr><td>MetaMath-Mistral-7B</td><td>15.8</td><td>10.7</td></tr><tr><td>+ DPO</td><td>15.8 -0</td><td>7.7 -3.0</td></tr><tr><td>+ Step-DPO</td><td>15.1 -0.7</td><td>7.0 -3.7</td></tr><tr><td>+ Full-Step-DPO</td><td>20.5 +4.7</td><td>9.9 -0.8</td></tr><tr><td>Llama-3-8B-SFT</td><td>20.5</td><td>12.5</td></tr><tr><td>+ DPO</td><td>11.7 -8.8</td><td>9.9 -2.6</td></tr><tr><td>+ Step-DPO</td><td>19.7 -0.8</td><td>13.6 +1.1</td></tr><tr><td>+ Full-Step-DPO</td><td>22.1 +1.6</td><td>15.1 +2.6</td></tr><tr><td>DeepSeekMath-Base-SFT</td><td>30.4</td><td>19.1</td></tr><tr><td>+ DPO</td><td>31.2 +0.8</td><td>18.4 -0.7</td></tr><tr><td>+ Step-DPO</td><td>31.2 +0.8</td><td>18.0 -1.1</td></tr><tr><td>+ Full-Step-DPO</td><td>31.7 +1.3</td><td>20.2 +1.1</td></tr><tr><td>Qwen2-7B-SFT</td><td>33.0</td><td>15.8</td></tr><tr><td>+ DPO</td><td>8.8 -24.2</td><td>8.1 -7.7</td></tr><tr><td>+ Step-DPO</td><td>32.5 -0.5</td><td>15.8 -0</td></tr><tr><td>+ Full-Step-DPO</td><td>33.5 +0.5</td><td>18.4 +2.6</td></tr></table>",
+        "bbox": [
+            517,
+            82,
+            884,
+            338
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Performance comparison on out-of-domain math problems.",
+        "bbox": [
+            507,
+            348,
+            882,
+            378
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Results on Various Verification Strategies",
+        "text_level": 1,
+        "bbox": [
+            507,
+            403,
+            878,
+            419
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 presents the performance of different verification strategies on GSM8K under two settings:  $K = 5$  and  $K = 15$ . We find that: (1) SC serves as a simple yet powerful validation method that significantly improves performance across all models. Even for the high-performing Qwen2-7B-SFT, which achieves an accuracy of  $89.3\\%$  with greedy decoding, SC further improves the accuracy to  $93\\%$  when the sampling size  $K = 15$ . This result is already comparable to GPT-4's accuracy of  $93.6\\%$ . (2) Compared to SC, BoN often achieves further improvements on MetaMath-Mistral-7B and Llama-3-8B-SFT. However, on the highly capable DeepSeekMath-Base-SFT and Qwen2-7B-SFT, BoN underperforms SC, indicating that the benefits of the reward model diminish for very strong baseline models. (3) SBS performs worse than both SC and BoN across most settings, yet consistently surpasses Greedy decoding, aligning with findings from previous studies (Yu et al., 2023a; Khalifa et al., 2023). This may be because, during the early stages of inference, the reward model struggles to effectively distinguish the correctness of steps.",
+        "bbox": [
+            507,
+            423,
+            884,
+            793
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5 Analysis of PRMs",
+        "text_level": 1,
+        "bbox": [
+            507,
+            804,
+            695,
+            819
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As discussed in Section 3.3, the quality of the PRM may be influenced by the sampling number  $M$  and simulation number  $N$ . To assess this, we conducted a controlled experiment with fixed  $M = 32$  and varying  $N$ . The trained PRM is then used to guide the decoding of MetaMath-Mistral-7B-Full-Step",
+        "bbox": [
+            507,
+            824,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "24349",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/dfc7c378f85279b5e5608382a0092fda43f8faeae674c1fe48ce57045ab05165.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            82,
+            480,
+            190
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/36cb754288f4f480d9b2ef03e68c13a4604700801cb677fe1042d0c3f6823244.jpg",
+        "image_caption": [
+            "Figure 3: Performance comparison of various verifications on GSM8K, with all models trained using our Full-Step-DPO."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            196,
+            480,
+            324
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "DPO using BoN strategy with  $K = 15$ .",
+        "bbox": [
+            112,
+            403,
+            406,
+            419
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 4, the purple bars indicate the A100-Hours cost for constructing PRM training data, while blue and orange lines show GSM8K and MATH accuracy. When  $N = 1$ , which represents our proposed method, the model achieves competitive performance with  $85.3\\%$  accuracy on GSM8K and  $34.2\\%$  on MATH, while only requiring 8.5h with a time complexity of  $O(M)$ . As  $N$  increases, we observe that the performance fluctuation remains relatively minimal (within approximately  $1\\%$  range), while the sampling cost grows substantially with a complexity of  $O(MNK)$ . This empirical evidence suggests that our approach with  $N = 1$  achieves a good balance between sampling efficiency and model performance. Additionally, we provide a comprehensive comparison between our PRM and other publicly available PRMs in Appendix B.2.",
+        "bbox": [
+            112,
+            420,
+            489,
+            709
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6 Sensitivity of Hyperparameters",
+        "text_level": 1,
+        "bbox": [
+            112,
+            722,
+            405,
+            738
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In step-wise DPO loss, the reward temperature  $\\gamma$  reflects the level of trust in the PRM. As  $\\gamma$  increases, the PRM model has a greater impact on the gradients. When  $\\gamma \\rightarrow 0$ , it indicates complete distrust in the PRM model, assigning equal weight coefficient to all steps, degrading in vanilla DPO. Conversely, when  $\\gamma \\rightarrow \\infty$ , the loss function optimizes only the single step with the maximum or minimum reward in the solution, similar to Step-DPO. Figure 5 presents the accuracy of MetaMath-Mistral-7B-Full-Step-DPO",
+        "bbox": [
+            112,
+            744,
+            487,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1ea9128f17b92b6c7ff46027421834ad91b4cbd4513550e934253b1dcd20d856.jpg",
+        "image_caption": [
+            "Figure 4: Accuracy of MetaMath-Mistral-7B-Full-Step-DPO using BoN decoding with  $K = 15$ . The PRM uses a fixed sampling number  $M = 32$ , while the simulation number  $N$  varies. Purple bars indicate the A100-Hours cost for constructing PRM training data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            80,
+            884,
+            223
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/2e7936fa99ff31f2c6a7b412a9d798f690318a4458a7b0459caa0c4884dd34cf.jpg",
+        "image_caption": [
+            "Figure 5: Accuracy of MetaMath-Mistral-7B-Full-Step-DPO with different reward temperature  $\\gamma$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            319,
+            873,
+            460
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "with different  $\\gamma$  values. The experimental results indicate that introducing a PRM to weight the gradients indeed effectively enhances optimization efficiency and improves performance. Additionally, this experiment demonstrates that there is a sweet spot for the reward temperature  $\\gamma$ ; excessively high or low  $\\gamma$  will reduce accuracy.",
+        "bbox": [
+            507,
+            525,
+            882,
+            639
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            668,
+            640,
+            684
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose Full-Step-DPO, a novel framework for mathematical reasoning that optimizes each step in the entire reasoning chain using step-wise rewards. To achieve this, we train a self-supervised Process Reward Model to automatically score reasoning steps, eliminating reliance on external annotations. We also propose a novel Step-Wise DPO Loss that dynamically updates gradients based on the rewards for individual steps, enabling step-wise optimization and enhancing the reasoning ability of policy models. Experimental results on various benchmarks validate the effectiveness of Full-Step-DPO, paving the way for its application to other reasoning-intensive tasks.",
+        "bbox": [
+            505,
+            696,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "24350",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While we have conducted comprehensive experiments to demonstrate the effectiveness of Full-Step-DPO, several limitations remain. First, recent advancements suggest that generative reward models outperform the discriminative reward model used in this work. Exploring how generative reward models can further enhance mathematical reasoning capabilities would be a valuable direction for future research. Second, during preference data construction, the current strategy of selecting samples based on average reward is relatively simple. Investigating more advanced sample selection strategies may lead to further improvements. Finally, the step-wise DPO loss proposed in this paper is highly adaptable to other reasoning tasks, such as code generation. Conducting experiments on a broader range of tasks would provide additional evidence of the advantages of our method.",
+        "bbox": [
+            115,
+            109,
+            489,
+            398
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            411,
+            278,
+            426
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-TC2022-005). We also wish to extend their heartfelt gratitude to the Sea AI Lab for their generous support in providing the necessary equipment and computational resources critical for the successful completion of this research.",
+        "bbox": [
+            112,
+            436,
+            489,
+            565
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            592,
+            213,
+            606
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.",
+            "Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q Jiang, Jia Deng, Stella Biderman, and Sean Welleck. 2023. Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631.",
+            "Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.",
+            "Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024. Step-level value preference optimization for mathematical reasoning. arXiv preprint arXiv:2406.10858.",
+            "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias"
+        ],
+        "bbox": [
+            115,
+            615,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021a. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.",
+            "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021b. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.",
+            "Rémi Coulom. 2006. Efficient selectivity and backup operators in monte-carlo tree search. In International conference on computers and games, pages 72-83. Springer.",
+            "Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306.",
+            "Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874.",
+            "Mohammad Javad Hosseini, Hannaneh Hajishirzi, Oren Etzioni, and Nate Kushman. 2014. Learning to solve arithmetic word problems with verb categorization. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 523-533.",
+            "Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2023. Large language models cannot self-correct reasoning yet. arXiv preprint arXiv:2310.01798.",
+            "Fangkai Jiao, Chengwei Qin, Zhengyuan Liu, Nancy F Chen, and Shafiq Joty. 2024. Learning planning-based reasoning by trajectories collection and process reward synthesizing. arXiv preprint arXiv:2402.00658.",
+            "Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169.",
+            "Muhammad Khalifa, Lajanugen Logeswaran, Moontae Lee, Honglak Lee, and Lu Wang. 2023. Grace: Discriminator-guided chain-of-thought reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15299-15328.",
+            "Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xi-angru Peng, and Jiaya Jia. 2024. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. arXiv preprint arXiv:2406.18629.",
+            "Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. 2022. Solving quantitative reasoning problems with language models. Advances"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "24351",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "in Neural Information Processing Systems, 35:3843-3857.",
+            "Minpeng Liao, Wei Luo, Chengxi Li, Jing Wu, and Kai Fan. 2024. Mario: Math reasoning with code interpreter output-a reproducible pipeline. arXiv preprint arXiv:2401.08190.",
+            "Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let's verify step by step. arXiv preprint arXiv:2305.20050.",
+            "Haoxiong Liu and Andrew Chi-Chih Yao. 2024. Aug-mentation math word problems via iterative question composing. arXiv preprint arXiv:2401.09003.",
+            "Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101.",
+            "Zimu Lu, Aojun Zhou, Ke Wang, Houxing Ren, Weikang Shi, Junting Pan, Mingjie Zhan, and Hongsheng Li. 2024. Step-controlled dpo: Leveraging stepwise error for enhanced mathematical reasoning. arXiv preprint arXiv:2407.00782.",
+            "Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583.",
+            "Xin Mao, Feng-Lin Li, Huimin Xu, Wei Zhang, Wang Chen, and Anh Tuan Luu. 2024a. As simple as fine-tuning: Llm alignment via bidirectional negative feedback loss. arXiv preprint arXiv:2410.04834.",
+            "Xin Mao, Feng-Lin Li, Huimin Xu, Wei Zhang, and Anh Tuan Luu. 2024b. Don't forget your reward values: Language model alignment via value-based calibration. arXiv preprint arXiv:2402.16030.",
+            "Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772.",
+            "Arindam Mitra, Hamed Khanpour, Corby Rosset, and Ahmed Awadallah. 2024. Orca-math: Unlocking the potential of slms in grade school math. arXiv preprint arXiv:2402.14830.",
+            "Arka Pal, Deep Karkhanis, Samuel Dooley, Manley Roberts, Siddartha Naidu, and Colin White. 2024. Smaug: Fixing failure modes of preference optimisation with dpo-positive. arXiv preprint arXiv:2402.13228.",
+            "Fengjun Pan, Xiaobao Wu, Zongrui Li, and Anh Tuan Luu. 2024. Are LLMs good zero-shot fallacy classifiers? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14338-14364, Miami, Florida, USA. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch. NIPS 2017 Workshop Autodiff Submission.",
+            "Arkil Patel, Satwik Bhattachamishra, and Navin Goyal. 2021. Are nlp models really able to solve simple math word problems? arXiv preprint arXiv:2103.07191.",
+            "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.",
+            "Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 3505-3506.",
+            "Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, YK Li, Y Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.",
+            "Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Schärli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210-31227. PMLR.",
+            "Yifan Song, Weimin Xiong, Dawei Zhu, Cheng Li, Ke Wang, Ye Tian, and Sujian Li. 2023. Rest-gpt: Connecting large language models with real-world applications via restful apis. arXiv preprint arXiv:2306.06624.",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.",
+            "Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275.",
+            "Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023a. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291.",
+            "Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. 2023b. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "24352",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Peiyi Wang, Lei Li, Zhihong Shao, RX Xu, Damai Dai, Yifei Li, Deli Chen, Y Wu, and Zhifang Sui. 2023c. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. CoRR, abs/2312.08935.",
+            "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.",
+            "Xiaobao Wu. 2025. Sailing ai by the stars: A survey of learning from rewards in post-training and test-time scaling of large language models. arXiv preprint arXiv:2505.02686.",
+            "Xiaobao Wu, Thong Nguyen, and Anh Tuan Luu. 2024a. A survey on neural topic models: Methods, applications, and challenges. Artificial Intelligence Review.",
+            "Xiaobao Wu, Thong Thanh Nguyen, Delvin Ce Zhang, William Yang Wang, and Anh Tuan Luu. 2024b. FASTopic: Pretrained transformer is a fast, adaptive, stable, and transferable topic model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.",
+            "Xiaobao Wu, Liangming Pan, William Yang Wang, and Anh Tuan Luu. 2024c. AKEW: Assessing knowledge editing in the wild. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15118-15133, Miami, Florida, USA. Association for Computational Linguistics.",
+            "Xiaobao Wu, Liangming Pan, Yuxi Xie, Ruiwen Zhou, Shuai Zhao, Yubo Ma, Mingzhe Du, Rui Mao, Anh Tuan Luu, and William Yang Wang. 2024d. Antileak-bench: Preventing data contamination by automatically constructing benchmarks with updated real-world knowledge. arXiv preprint arXiv:2412.13670.",
+            "Zhenyu Wu, Qingkai Zeng, Zhihan Zhang, Zhaoxuan Tan, Chao Shen, and Meng Jiang. 2024e. Enhancing mathematical reasoning in llms by stepwise correction. arXiv preprint arXiv:2410.12934.",
+            "Yuxi Xie, Anirudh Goyal, Wenyue Zheng, Min-Yen Kan, Timothy P Lillicrap, Kenji Kawaguchi, and Michael Shieh. 2024a. Monte carlo tree search boosts reasoning via iterative preference learning. arXiv preprint arXiv:2405.00451.",
+            "Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2024b. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems, 36."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yifan Xu, Xiao Liu, Xinghan Liu, Zhenyu Hou, Yueyan Li, Xiaohan Zhang, Zihan Wang, Aohan Zeng, Zhengxiao Du, Wenyi Zhao, et al. 2024. Chatglmmath: Improving math problem-solving in large language models with a self-critique pipeline. arXiv preprint arXiv:2404.02893.",
+            "Huaiyuan Ying, Shuo Zhang, Linyang Li, Zhejian Zhou, Yunfan Shao, Zhaoye Fei, Yichuan Ma, Jiawei Hong, Kuikun Liu, Ziyi Wang, et al. 2024. Internlm-math: Open math large language models toward verifiable reasoning. arXiv preprint arXiv:2402.06332.",
+            "Fei Yu, Anningzhe Gao, and Benyou Wang. 2023a. Outcome-supervised verifiers for planning in mathematical reasoning. arXiv preprint arXiv:2311.09724.",
+            "Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 2023b. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284.",
+            "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."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            467
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "24353",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Step-wise Beam Search",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            354,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Unlike conventional beam search, which relies on token-level probabilities, our method integrates the reward model with an associated reranking criterion. This enables for step-wise beam search (SBS) (Yu et al., 2023a; Chen et al., 2024), effectively selecting the preferred solution path in mathematical reasoning, while incurring a lower computational cost compared to Monte Carlo Tree Search. Specifically, for each step  $t$ , suppose the sampling size is  $b_{1}$ , the policy model  $\\pi_{\\theta}$  produces a set of candidate steps  $\\mathbb{S}^{(1:t+1)} = \\{S_i^{(1:t+1)}\\}_{i=1}^{b_1}$ , where  $S_i^{(1:t+1)} = [s_i^1, \\dots, s_i^{t+1}]$  is the  $i$ -th partial solution up to step  $t+1$ . Given the PRM  $\\pi_r$  that can score each step, we select the top-scoring steps with beam size  $b_{2}$ . The algorithm is detailed in Algorithm 1. By focusing on the quality of each reasoning step rather than just the final solution, our method enhances the overall reasoning capabilities of the model.",
+        "bbox": [
+            112,
+            108,
+            884,
+            255
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1 Step-wise Beam Search"
+        ],
+        "code_body": "1: Input: Math problem  $q$ , Sampling size  $b_{1}$ , Beam size  $b_{2}$ , Maximum step  $C$   \n2: Output: Best solution for  $q$   \n3: Models: Policy model  $\\pi_{\\theta}$  and PRM  $\\pi_r$   \n4: function STEPLEVELBEAMSEARCH( $q, b_1, b_2, C$ )  \n5: Initialize step sequences  $\\mathbb{S} \\gets \\{\\}$   \n6: Use  $\\pi_{\\theta}$  to sample initial steps  $\\{s_1^1, \\ldots, s_{b_1}^1\\}$   \n7: Use  $\\pi_r$  to score all initial steps  $\\{r_1^1, \\ldots, r_{b_1}^1\\}$   \n8: Select top- $b_1$  steps and add to  $\\mathbb{S}$   \n9: Set current step counter  $t \\gets 1$   \n10: while  $t < C$  do  \n11: if All sequences in  $\\mathbb{S}$  are complete then  \n12: Break  \n13: end if  \n14:  $\\mathbb{S}_{\\mathrm{new}} \\gets \\{\\}$   \n15:  $\\mathbb{R} \\gets \\{\\}$ ;  \n16: for each solution  $\\mathcal{S}^{(1:t)}$  in  $\\mathbb{S}$  do  \n17: for  $i = 1$  to  $b_1$  do  \n18:  $\\mathcal{S}_i^{(1:t + 1)} = \\pi_{\\theta}(\\mathcal{S}^{(1:t)}; q)$   \n19:  $r_i^{(1:t + 1)} = \\pi_r(\\mathcal{S}_i^{(1:t + 1)}; q)$   \n20:  $\\mathbb{S}_{\\mathrm{new}} \\gets \\mathbb{S}_{\\mathrm{new}} + \\{\\mathcal{S}_i^{(1:t + 1)}\\}$   \n21:  $\\mathbb{R} \\gets \\mathbb{R} + \\{r_i^{(1:t + 1)}\\}$   \n22: end for  \n23: end for  \n24:  $\\mathbb{S}_{\\mathrm{new}} \\gets$  top- $b_2$  rewarded solutions in ( $\\mathbb{S}_{\\mathrm{new}}, \\mathbb{R}$ )  \n25:  $\\mathbb{S} \\gets \\mathbb{S}_{\\mathrm{new}}$   \n26:  $t \\gets t + 1$ ;  \n27: end while  \n28: return solution with highest final reward in  $\\mathbb{S}$   \n29: end function",
+        "bbox": [
+            115,
+            285,
+            882,
+            765
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Additional Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            788,
+            359,
+            804
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.1 More OOD Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            813,
+            324,
+            828
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our method on five additional OOD mathematical reasoning datasets as our testbed. As shown in Table 3, our Full-Step-DPO consistently improves performance across all datasets, demonstrating the effectiveness and generalization ability of our approach on OOD mathematical reasoning tasks.",
+        "bbox": [
+            112,
+            834,
+            882,
+            882
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "- SVAMP (Patel et al., 2021) includes 1000 math questions of up to fourth grade difficulty. These questions can be solved by expressions requiring no more than two operators.",
+        "bbox": [
+            134,
+            889,
+            882,
+            921
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "24354",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- AddSub (Hosseini et al., 2014) contains 395 math questions that involve addition and subtraction operations.",
+            "- ASDiv (Miao et al., 2021) contains 2215 English math questions of different problem types. Each question provides the corresponding equation and answer.",
+            "- GSM-IC2 and GSM-ICM (Shi et al., 2023) are mathematical reasoning datasets containing irrelevant conditions within the problem descriptions each consisting of 1000 problems. Problems in GSM-IC2 require two steps to solve, while problems in GSM-ICM require more than two steps to solve."
+        ],
+        "bbox": [
+            134,
+            84,
+            884,
+            228
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/f77d75c0be6bd0ead2534ff3e9db62586313a4f8720907511746e40408de3c8d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>SVAMP (%)</td><td>AddSub (%)</td><td>ASDiv (%)</td><td>GSM-IC2 (%)</td><td>GSM-ICM (%)</td><td>Average (%)</td></tr><tr><td>MetaMath-Mistral-7B</td><td>79.1</td><td>86.6</td><td>81.2</td><td>77.9</td><td>76.5</td><td>80.3</td></tr><tr><td>+ DPO</td><td>72.7</td><td>53.4</td><td>73.8</td><td>60.5</td><td>65.2</td><td>65.1</td></tr><tr><td>+ Step-DPO</td><td>80.3</td><td>86.9</td><td>82.7</td><td>76.4</td><td>76.3</td><td>80.5</td></tr><tr><td>+ Full-Step-DPO</td><td>81.7</td><td>88.6</td><td>83.9</td><td>75.9</td><td>76.6</td><td>81.3</td></tr><tr><td>Llama-3-8B-SFT</td><td>82.8</td><td>88.4</td><td>85.0</td><td>80.1</td><td>79.5</td><td>83.2</td></tr><tr><td>+ DPO</td><td>72.2</td><td>60.0</td><td>74.2</td><td>64.1</td><td>66.3</td><td>67.4</td></tr><tr><td>+ Step-DPO</td><td>81.7</td><td>88.5</td><td>85.3</td><td>80.7</td><td>81.2</td><td>83.5</td></tr><tr><td>+ Full-Step-DPO</td><td>82.9</td><td>88.8</td><td>86.4</td><td>82.1</td><td>81.6</td><td>84.4</td></tr><tr><td>DeepSeekMath-Base-SFT</td><td>84.2</td><td>87.6</td><td>91.0</td><td>85.4</td><td>85.2</td><td>86.7</td></tr><tr><td>+ DPO</td><td>85.1</td><td>86.6</td><td>90.6</td><td>85.2</td><td>85.0</td><td>86.5</td></tr><tr><td>+ Step-DPO</td><td>85.3</td><td>85.3</td><td>90.7</td><td>85.9</td><td>86.2</td><td>86.7</td></tr><tr><td>+ Full-Step-DPO</td><td>85.9</td><td>86.6</td><td>91.2</td><td>87.8</td><td>85.3</td><td>87.4</td></tr><tr><td>Qwen2-7B-SFT</td><td>88.7</td><td>92.7</td><td>91.6</td><td>93.7</td><td>91.6</td><td>91.7</td></tr><tr><td>+ DPO</td><td>23.5</td><td>25.6</td><td>27.0</td><td>30.1</td><td>29.5</td><td>27.1</td></tr><tr><td>+ Step-DPO</td><td>88.1</td><td>92.2</td><td>91.8</td><td>93.8</td><td>91.9</td><td>91.6</td></tr><tr><td>+ Full-Step-DPO</td><td>89.5</td><td>93.1</td><td>92.4</td><td>93.6</td><td>92.7</td><td>92.3</td></tr></table>",
+        "bbox": [
+            117,
+            253,
+            878,
+            501
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B.2 Performance Comparison of Different PRMs",
+        "text_level": 1,
+        "bbox": [
+            112,
+            571,
+            515,
+            588
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We compare our PRM with Math-Shepherd-PRM-7B using MetaMath-Mistral-7B as the base model under the Best-of-8 decoding strategy. During decoding, we sampled eight responses with temperature  $= 0.7$ . For both PRMs, each response was scored by the minimum reward among its reasoning steps, and the response with the highest score was selected as the final output.",
+        "bbox": [
+            110,
+            594,
+            882,
+            658
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 4, our PRM achieves  $84.3\\%$  accuracy on GSM8K and  $33.2\\%$  on MATH, outperforming Math-Shepherd-PRM-7B (82.4% on GSM8K, 32.6% on MATH). While these results suggest the effectiveness of our approach, we acknowledge that the comparison may not be entirely comprehensive due to differences in backbone architecture and training data between the PRMs. Nevertheless, these findings, combined with our analysis in Section 4.5, demonstrate that our PRM can achieve competitive performance while significantly reducing the computational cost of PRM construction.",
+        "bbox": [
+            110,
+            659,
+            884,
+            756
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/1e50fa0b00344d39c01db8f13a004e234284bab9b4b0894ca7cf5e1146ea2512.jpg",
+        "table_caption": [
+            "Table 3: Performance comparison on five additional OOD math problems. The best and second-best results are highlighted in bold and underlined, respectively."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>MATH</td><td>GSM8K</td></tr><tr><td>MetaMath-Mistral-7B</td><td>28.2</td><td>77.7</td></tr><tr><td>+ Self-Consistency</td><td>32.4</td><td>81.3</td></tr><tr><td>+ Math-Shepherd-PRM-7B</td><td>32.6</td><td>82.4</td></tr><tr><td>+ Our-PRM-7B</td><td>33.2</td><td>84.3</td></tr></table>",
+        "bbox": [
+            310,
+            769,
+            692,
+            863
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Performance comparison of different scoring strategies using Best-of-8 decoding on MATH and GSM8K benchmarks.",
+        "bbox": [
+            112,
+            872,
+            882,
+            900
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "24355",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B.3 Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            252,
+            99
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "We conducted a detailed case study using a sample problem from the AddSub dataset, as shown in Table 5. We compared outputs from DeepSeekMath-Base-SFT model and its Step-DPO and Full-Step-DPO variants. While both the base SFT model and Step-DPO model made incorrect reasoning by including the cost of unbought CDs, our Full-Step-DPO model generated correct reasoning steps. Our PRM scores accurately reflect the correctness of each reasoning step, with higher scores for correct steps and lower scores for incorrect ones.",
+        "bbox": [
+            110,
+            105,
+            885,
+            202
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/3994df9bd95eed550b814344034b1aa7b3433308cf2d8ead65bf2931232be403.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Question</td><td>For his car, Mike spent $118.54 on speakers and $106.33 on new tires. Mike wanted 3 CD&#x27;s for $4.58 but decided not to. In total, how much did Mike spend on car parts?</td></tr><tr><td>Gold Answer</td><td>224.87</td></tr><tr><td rowspan=\"3\">SFT Response</td><td>Step 1: Mike spent 118.54onspeaker sand106.33 on new tires. So the total cost of these two items is $118.54 + $106.33 = $224.87. Score: 0.9348</td></tr><tr><td>Step 2: Mike wanted to buy 3 CD&#x27;s for $4.58 each, but he decided not to. So the total cost of the CD&#x27;s he didn&#x27;t buy is 3 * $4.58 = $13.74. Score: 0.3646</td></tr><tr><td>Step 3: To find out how much Mike spent in total on car parts, we need to add the total cost of the speakers and tires to the total cost of the CD&#x27;s he didn&#x27;t buy. So the total cost is $224.87 + $13.74 = $238.61. The answer is: 238.61 Score: 0.0427</td></tr><tr><td rowspan=\"3\">Step-DPO Response</td><td>Step 1: First, Mike spent $118.54 on speakers and $106.33 on new tires. So the total cost of these two items is $118.54 + $106.33 = $224.87. Score: 0.9406</td></tr><tr><td>Step 2: Mike also wanted 3 CD&#x27;s, but he decided not to buy them. Each CD costs $4.58, so the total cost of the CD&#x27;s he didn&#x27;t buy is 3 * $4.58 = $13.74. Score: 0.3036</td></tr><tr><td>Step 3: To find out how much Mike spent in total on car parts, we need to add the total cost of the speakers and tires to the total cost of the CD&#x27;s he didn&#x27;t buy. So, the total cost is $224.87 + $13.74 = $238.61. The answer is: 238.61 Score: 0.0445</td></tr><tr><td rowspan=\"3\">Full-Step-DPO Response</td><td>Step 1: Mike spent $118.54 on speakers and $106.33 on new tires. So the total amount he spent on these two items is $118.54 + $106.33 = $224.87. Score: 0.9313</td></tr><tr><td>Step 2: Mike wanted to buy 3 CD&#x27;s for $4.58 each, but he decided not to. So the total amount he would have spent on the CD&#x27;s is 3 * $4.58 = $13.74. Score: 0.7408</td></tr><tr><td>Step 3: To find out how much Mike spent in total on car parts, we need to add the total amount he spent on speakers and tires to the amount he would have spent on CD&#x27;s. So the total amount is $224.87 + $0 = $224.87. The answer is: 224.87 Score: 0.7363</td></tr></table>",
+        "bbox": [
+            119,
+            209,
+            885,
+            536
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Comparison of reasoning steps and their corresponding PRM scores across different models for AddSub dataset. The base model is DeepSeekMath-Base-SFT.",
+        "bbox": [
+            112,
+            546,
+            882,
+            576
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "24356",
+        "bbox": [
+            475,
+            928,
+            526,
+            941
+        ],
+        "page_idx": 13
+    }
+]

2025/Full-Step-DPO_ Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning/e3c1942a-6072-41ae-9020-bff00ee3eef1_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae739001c370ebd124830df77becd9e64282e82f374cd07e2828060d54f6bed3
+size 859915

2025/Full-Step-DPO_ Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning/full.md

@@ -0,0 +1,363 @@
+# Full-Step-DPO: Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning
+
+Huimin Xu $^{1}$ , Xin Mao $^{1}$ , Feng-Lin Li $^{2}$ , Xiaobao Wu $^{1*}$  Wang Chen $^{2}$ , Wei Zhang $^{3}$ , Anh Tuan Luu $^{1*}$
+
+$^{1}$ Nanyang Technological University, Singapore  $^{2}$ Shopee Pte. Ltd, Singapore,  $^{3}$ SEA Group, Singapore {huimin.xu, xin.mao, xiaobao.wu, anhtuan.luu}@ntu.edu.sg {fenglin.li, chen.wang}@shopee.com, terry.zhang@sea.com
+
+# Abstract
+
+Direct Preference Optimization (DPO) often struggles with long-chain mathematical reasoning. Existing approaches, such as Step-DPO, typically improve this by focusing on the first erroneous step in the reasoning chain. However, they overlook all other steps and rely heavily on humans or GPT-4 to identify erroneous steps. To address these issues, we propose Full-Step-DPO, a novel DPO framework tailored for mathematical reasoning. Instead of optimizing only the first erroneous step, it leverages step-wise rewards from the entire reasoning chain. This is achieved by training a self-supervised process reward model, which automatically scores each step, providing rewards while avoiding reliance on external signals. Furthermore, we introduce a novel stepwise DPO loss, which dynamically updates gradients based on these step-wise rewards. This endows stronger reasoning capabilities to language models. Extensive evaluations on both in-domain and out-of-domain mathematical reasoning benchmarks across various base language models, demonstrate that Full-Step-DPO achieves superior performance compared to state-of-the-art baselines<sup>1</sup>.
+
+# 1 Introduction
+
+Large Language Models (LLMs) have attracted massive interest due to their remarkable capabilities across various tasks (Kaddour et al., 2023; Song et al., 2023; Wang et al., 2023a; Zheng et al., 2024; Wang et al., 2023b; Pan et al., 2024; Wu et al., 2024b,c,d,a). However, they commonly encounter difficulties when tackling complex and symbolic multi-step reasoning, particularly in mathematical problem reasoning (Lightman et al., 2023; Huang et al., 2023; Wu, 2025). To improve the mathematical reasoning ability, some studies use Direct Preference Optimization (DPO) (Rafailov et al., 2024)
+
+![](images/fd81edc99a96354f12a23b158cd2d1f9163d807962eb26ed65d1cf508162317f.jpg)  
+DPO
+
+![](images/1a7eddef3220f336d8fe0af692f35ea07a98f1c0a020f3f7959f6fb50ff6a469.jpg)  
+Step-DPO
+
+![](images/54a3d6658f15b77fce39e517a17da27e88133fe06185115c86afd845f199bab1.jpg)  
+Full-Step-DPO  
+Figure 1: Comparison between DPO, Step-DPO, and our Full-Step-DPO. DPO operates on solution-wise preference data. Step-DPO advances to step-wise data but optimizes only a single step. Full-Step-DPO optimizes all steps with a novel step-wise DPO loss, effectively enhancing the model's reasoning capability.
+
+with solution-wise preference data but find its benefit limited (Pal et al., 2024; Xu et al., 2024; Jiao et al., 2024). Recent works attribute this limitation to DPO's inability to perform process supervision and instead build preference data based on reasoning steps rather than entire solutions (Lai et al., 2024; Chen et al., 2024; Xie et al., 2024a; Lu et al., 2024). For example, Step-DPO (Lai et al., 2024) focuses on optimizing only the first erroneous step in the reasoning chain, demonstrating notable improvements.
+
+However, despite their improvements, these existing methods face the following limitations: (1) Some focus solely on the first erroneous step and ignore all other useful steps in the reasoning chain (Lai et al., 2024), as shown in Figure 1.
+
+As a result, they fail to fully optimize the reasoning chains, leading to suboptimal performance. (2) Their loss function still follows the early vanilla DPO in a solution-wise approach (Lu et al., 2024; Xie et al., 2024a). Consequently, it cannot directly leverage rewards in a step-wise fashion for learning. (3) They heavily rely on costly and resource-intensive annotations from GPT-4 or humans to detect erroneous steps (Lai et al., 2024; Lightman et al., 2023), significantly limiting their practicality.
+
+To address the above limitations, we propose Full-Step-DPO, a novel DPO framework for mathematical reasoning. As illustrated in Figure 1, unlike vanilla DPO, which operates solution-wise, or Step-DPO, which focuses solely on the first erroneous step, Full-Step-DPO utilizes each step in the entire reasoning chain and optimizes them using step-wise rewards. We first train a Process Reward Model (PRM) (Lightman et al., 2023; Wang et al., 2023c) in a self-supervised way, utilizing data generated by the model itself. This approach enables the PRM to automatically score each step in the reasoning chain, eliminating the reliance on external annotations such as GPT-4 or humans. Then, we propose a novel Step-wise DPO Loss, which employs dynamic gradient updates to optimize each step based on its corresponding reward. This approach shifts the optimization focus from solution-wise to step-wise, enabling the policy model to achieve superior reasoning capabilities.
+
+We conduct experiments on both in-domain and out-of-domain mathematical reasoning datasets with four widely used backbone LLMs. Experimental results demonstrate that our Full-Step DPO consistently outperforms the DPO and Step-DPO baselines, validating its effectiveness in enhancing reasoning performance. Our contributions can be summarized as follows:
+
+- We propose the Full-Step-DPO framework with a novel step-wise DPO loss that dynamically adjusts each step's gradient based on its reward, enabling step-wise optimization rather than solution-wise and enhancing reasoning ability.  
+- We train a self-supervised PRM to provide stepwise rewards for preference learning and explore a more efficient approach for automatically constructing PRM training data.  
+- Extensive experiments on widely used mathematical benchmarks and base language models showcase the remarkable effectiveness of our method.
+
+# 2 Related Work
+
+Mathematical Reasoning Mathematical reasoning task is one of the most challenging tasks for LLMs. Various approaches have been explored to improve or elicit the mathematical reasoning ability of LLMs. A number of approaches have either continually pre-trained the base model on a vast of mathematical datasets (Azerbayev et al., 2023; Shao et al., 2024) or used supervised fine-tuning with substantial synthetic datasets distilled from cutting-edge models (Luo et al., 2023; Yu et al., 2023b; Mitra et al., 2024; Xu et al., 2024). Another line of work focuses on enhancing test-time computation by generating multiple solutions, developing separate reward models at either the outcome or process level to rerank these solutions (Cobbe et al., 2021a; Lightman et al., 2023), or employing decoding strategies guided by the reward model (Yu et al., 2023a; Xie et al., 2024b; Wang et al., 2023c; Wu et al., 2024e). In addition, Reinforcement Learning's potential in general domains, demonstrated by Achiam et al. (2023) and Touvron et al. (2023), some studies have explored its use in mathematical reasoning (Wang et al., 2023c; Mitra et al., 2024; Pal et al., 2024).
+
+Preference Learning Recently, preference learning (Ethayarajh et al., 2024; Mao et al., 2024b,a) has attracted significant attention due to its ability to align with human preferences and distinguish between positive and negative examples. While these methods, like DPO (Rafailov et al., 2024), have proven effective in general domains, it offers only marginal benefits for mathematical reasoning (Pal et al., 2024). Some works (Chen et al., 2024; Lai et al., 2024) suggest that DPO's focus on coarse solution-wise preferences makes it less effective at correcting errors in multi-step reasoning, hindering reasoning improvement. Therefore, Step-DPO (Lai et al., 2024) was proposed, which first identifies the first erroneous step, and then optimizes only this erroneous step along with the corresponding correct one. Although this approach enhances mathematical reasoning capabilities, it totally overlooks the other steps in long-chain reasoning, which also provide valuable information and should not be completely disregarded. Building on this consideration, we propose Full-Step-DPO, which fully accounts for each step by dynamically optimizing all steps in the reasoning process.
+
+Step-wise Supervision Recent findings by Lightman et al. (2023) suggest that step-wise supervision outperforms outcome-wise, due to the provision of more detailed feedback. However, training a PRM requires either costly manual annotation (Lightman et al., 2023) or significant computational resources (Khalifa et al., 2023; Wang et al., 2023c), which hinders the advancement and practical application of PRM. Therefore, in this paper, we aim to build a PRM for mathematical reasoning without relying on human annotation and with reduced computational resources. Additionally, we explore the effectiveness of the PRM in decoding and preference learning scenarios.
+
+# 3 Full-Step DPO
+
+In this section, we elaborate the proposed Full-Step DPO framework. We begin by reviewing the background of previous DPO and Step-DPO. Then we introduce the novel Step-wise DPO Loss which optimizes with step-wise rewards, and the Process Reward Model which automatically generate these step-wise rewards. Finally we outline the complete training pipeline of our Full-Step-DPO.
+
+# 3.1 Preliminary
+
+DPO. Direct Preference Optimization (DPO) (Rafailov et al., 2024) is one of the most popular preference optimization methods. Instead of learning an explicit reward model, DPO directly uses pair-wise preference data to optimize the policy model with an equivalent optimization objective. Specifically, given an input prompt  $x$ , and a preference data pair  $(y^w, y^l)$ , DPO aims to maximize the probability of the entire preferred solution  $y^w$  and minimize that of the dispreferred solution  $y^l$ . The optimization objective of DPO is:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {D P O}} (\theta) = - \mathbb {E} _ {(x, y ^ {w}, y ^ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(\right. \right. \\ \left. \beta \log \frac {\pi_ {\theta} (y ^ {w} \mid x)}{\pi_ {\mathrm {r e f}} (y ^ {w} \mid x)} - \beta \log \frac {\pi_ {\theta} (y ^ {l} \mid x)}{\pi_ {\mathrm {r e f}} (y ^ {l} \mid x)}\right) \Bigg ] \\ \end{array}
+$$
+
+where  $\pi_{\theta}(\cdot |x)$  is the policy model to be optimized,  $\pi_{\mathrm{ref}}(\cdot |x)$  is the reference model,  $(x,y^{w},y^{l})$  are preference pairs,  $\sigma$  is the sigmoid function,  $\beta$  is a parameter controlling the deviation from the reference model.
+
+Step-DPO. Although DPO performs well on chat benchmarks, it is less effective for long-chain reasoning tasks like mathematical problems. StepDPO (Lai et al., 2024) attributes this to DPO's
+
+inability to consider the sequential nature of mathematical reasoning, as rejecting an entire dispreferred solution may inadvertently penalize correct preceding steps, introducing significant noise. To address this, Step-DPO optimizes only the first incorrect step. As shown in Figure 1, given a math problem and a series of initial correct reasoning steps  $\{s_1,\dots,s_{k - 1}\}$ , Step-DPO aims to maximize the probability of the correct next step  $s_k^w$  and minimize the probability of the incorrect one  $s_k^l$ . Note that  $s_k^w$  and  $s_k^l$  refer to single steps, not all subsequent steps. The loss function used is still the vanilla DPO loss.
+
+# 3.2 Step-wise DPO Loss
+
+We now introduce the novel Step-wise DPO loss, which performs step-wise optimization using stepwise rewards. Although the motivation behind Step-DPO is reasonable, focusing solely on optimizing the first erroneous step and neglecting the valuable information provided by other steps may not be optimal. Additionally, we contend that it is not truly a step-wise DPO, as it still relies on the standard solution-wise DPO loss and resembles more of a data construction method.
+
+To address this, we modify the vanilla DPO loss to the step-wise DPO loss, dynamically weighting the gradients of each step based on its reward, thereby enabling true step-wise optimization. Let's start with the gradient of the loss function  $\mathcal{L}_{\mathrm{DPO}}$ . The gradient with respect to the parameters  $\theta$  can be written as:
+
+$$
+\begin{array}{l} \nabla_ {\theta} \mathcal {L} = - \beta \mathbb {E} _ {(x, y ^ {w}, y ^ {l}) \sim \mathcal {D}} [ \sigma (\hat {r} _ {\theta} (x, y ^ {l}) - \hat {r} _ {\theta} (x, y ^ {w})) \\ \left[ \nabla_ {\theta} \log \pi_ {\theta} (y ^ {w} \mid x) - \nabla_ {\theta} \log \pi_ {\theta} (y ^ {l} \mid x) \right] \\ \end{array}
+$$
+
+where  $\hat{r}_{\theta}(x,y) = \beta \log \frac{\pi_{\theta}(y|x)}{\pi_{\mathrm{ref}}(y|x)}$ . Intuitively, the gradient indiscriminately increases the likelihood of whole  $y^{w}$  and decreases the likelihood of whole  $y^{l}$ . To achieve dynamically weighting, we break  $\nabla_{\theta} \log \pi_{\theta}(y \mid x)$  into a step-wise form and weight the gradient as follows:
+
+$$
+\begin{array}{l} \nabla_ {\theta} \mathcal {L} = - \beta \mathbb {E} _ {(x, y ^ {w}, y ^ {l}) \sim \mathcal {D}} \left[ \sigma \left(\hat {r} _ {\theta} (x, y ^ {l}) - \hat {r} _ {\theta} (x, y ^ {w})\right) \right. \\ \left[ \sum_ {i = 1} ^ {K ^ {w}} \alpha_ {i} ^ {w} \nabla_ {\theta} \log \pi_ {\theta} \left(s _ {i} ^ {w} \mid x, s _ {: <   i} ^ {w}\right) \right. \\ \left. \left. - \sum_ {i = 1} ^ {K ^ {l}} \alpha_ {i} ^ {l} \nabla_ {\theta} \log \pi_ {\theta} \left(s _ {i} ^ {l} \mid x, s _ {: <   i} ^ {l}\right) \right] \right] \\ \end{array}
+$$
+
+where  $s_i$  represents the  $i$ -th reasoning step of the solution  $y$ ,  $s_{: < i}$  denotes all reasoning steps preceding  $s_i$ ,  $K$  is the total number of steps, and  $\alpha_i$  is
+
+the weight coefficient of  $s_i$ , calculated based on the reward of  $s_i$  as shown below:
+
+$$
+\alpha_ {i} = \left\{ \begin{array}{l l} \frac {e ^ {\gamma r _ {s i}}}{\sum_ {j} e ^ {\gamma r _ {s j}}}, & s _ {i} \in y ^ {w} \\ \frac {e ^ {- \gamma r _ {s i}}}{\sum_ {j} e ^ {- \gamma r _ {s j}}}, & s _ {i} \in y ^ {l} \end{array} \right.
+$$
+
+where  $r_{s_i}$  is the reward of the step  $s_i$ , which will be introduced in the next subsection, and  $\gamma$  is the temperature of the Softmax operation. It is important to note that the calculation of  $\alpha_{i}$  differs between the preferred solution  $y^{w}$  and the dispreferred solution  $y^{l}$ . For preferred solutions, a higher reward indicates a greater likelihood of correct reasoning in that step, so the model should perform gradient ascent with greater intensity. Conversely, for dispreferred solutions, a lower reward suggests a higher chance of incorrect reasoning, and the model should apply gradient descent with greater intensity accordingly. This approach allows us to leverage all steps and adaptively adjust the weight of each step based on its probability of correctness, achieving true step-wise optimization.
+
+Compared to Step-DPO methods that focus solely on a single step, our method optimizes all steps simultaneously, enabling better global optimization. Noted that as  $\gamma \rightarrow 0$ , all steps will have equal weights, making Full-Step-DPO equivalent to vanilla DPO.
+
+# 3.3 Process Reward Models
+
+To obtain step-wise rewards, we train a Process Reward Model (PRM). The biggest challenge in training a PRM is constructing a process supervision dataset. Previous studies (Uesato et al., 2022; Lightman et al., 2023) utilize human annotators to obtain step-wise labels, which requires advanced annotator skills and is quite costly. Later, MathShepherd (Wang et al., 2023c) proposes using Monte Carlo estimation (Coulom, 2006) to automatically gather step-wise supervision, but it remained computationally expensive. In this section, we first examine the principles of Monte Carlo estimation, then present our simplified solution that significantly improves the efficiency of data construction.
+
+Monte Carlo estimation. This approach assumes that the gold label  $y_{s_i}$  of a step  $s_i$  can be defined as the probability to deduce the correct answer  $a^*$ , and it includes both sampling and simulation phase. Specifically, given a math problem, it first
+
+randomly samples  $M$  solutions, with each solution consisting of  $K$  reasoning steps  $\{s_1,s_2,\ldots ,s_K\}$  and  $a$  represents the decoded answer from the last step  $s_K$ . Then, to estimate the quality of reasoning step  $s_i$  in a given solution, it simulates  $N$  subsequent reasoning processes from this step:  $\{(s_{i + 1,j},\dots ,s_{K,j})\}_{j = 1}^{N}$ . The golden label for  $s_i$  is calculated as follows:
+
+$$
+y _ {s _ {i}} = \frac {\sum_ {j = 1} ^ {N} \mathbb {I} (a _ {j} = a ^ {*})}{N}
+$$
+
+where  $a_{j}$  is the decoded answer for the  $j$ -th simulated solution, and  $\mathbb{I}$  is the indicator function that returns 1 if  $a_{j} = a^{*}$  and 0 otherwise. This two-stage approach is highly time-consuming, as it requires  $N$  simulations for each of  $K$  steps across all  $M$  solutions, resulting in a time complexity of  $O(MNK)$ .
+
+Our efficient approach. It is important to note that there is a trade-off between the sampling number  $M$  and the simulation number  $N$  when computational resources are limited. A larger  $M$  can provide more data for training the PRM, while a larger  $N$  can result in higher accuracy of the labels  $y_{i}$ . In this paper, we found that the trained PRM performs reasonably well even with  $N = 1$  when  $M$  is large, such as 32. This is likely because a larger  $M$  introduces more diversity into the training data, making the PRM more tolerant to slight reductions in data precision caused by the limited simulation number. This setting simplifies the PRM data construction by requiring only the sampling of  $M$  solutions without the need for simulation, significantly reducing computational resources and lowering the time complexity to  $O(M)$ . As a result, the gold label for step  $s_{i}$  can be simplified as follows:
+
+$$
+y _ {s _ {i}} = \left\{ \begin{array}{l l} 1 & \text {i f} a = a ^ {*} \\ 0 & \text {o t h e r w i s e} \end{array} \right.
+$$
+
+then the PRM could be trained as shown below:
+
+$$
+\mathcal {L} _ {\mathrm {P R M}} = - \sum_ {i = 1} ^ {K} \left[ y _ {s _ {i}} \log r _ {s _ {i}} + \left(1 - y _ {s _ {i}}\right) \log \left(1 - r _ {s _ {i}}\right) \right]
+$$
+
+where  $y_{s_i}$  is the golden label for  $s_i$ ,  $r_{s_i}$  is the sigmoid score assigned by the PRM. With the above PRM, we can automatically score each step in the reasoning chain, providing reward signals for the step-wise DPO loss and enabling step-wise optimization.
+
+![](images/94f7b2ed08d392c2a986402851dfd3ff4a3527f706c0293679c0d21805beb02b.jpg)  
+Step 1. PRM Training
+
+![](images/a33da6cd2ff7d6cb191a43e4add57e3868a5fcf6518547b4bbe3470607c649a9.jpg)  
+Step 2. Preference Building
+
+![](images/710bf4ed45dc2699222692f9fb06f316cf4093d99f12d37573a778d018d37a19.jpg)  
+Step 3. Preference Learning  
+step-wise preferences
+
+![](images/f0c004655858efdac27a6ac493b1e17d6ca200aaeba9edee555a68455998f421.jpg)
+
+![](images/15b3c65c5c05446c819f463c6d585c478edbc455c9f1dae83cedbd801acfff8b.jpg)  
+Figure 2: The overall framework of Full-Step-DPO consists of three steps: (1) Training the PRM using the model itself and generated solutions. (2) Using the PRM to score and filter solutions to form preference data with step-wise rewards. (3) Training the policy model with the proposed step-wise DPO loss.
+
+# 3.4 Training Pipeline
+
+Following previous methods (Wang et al., 2023c; Shao et al., 2024), we adopt a standard training pipeline illustrated in Figure 2: (1) We begin by training a PRM with self-generated data, where higher reward values indicate a stronger likelihood of correct reasoning, while lower values suggest potential errors. (2) The trained PRM is then used to construct preference pairs with step-wise rewards. Specifically, we generate  $M$  solutions for each math problem, score each step of these solutions with the PRM to produce a reward sequence, and calculate the average reward across all steps as the overall reward for each solution. We select the top  $T$  correct solutions with the highest rewards and the bottom  $T$  incorrect solutions with the lowest rewards to form  $T^2$  step-wise preference pairs. (3) Finally, we update the policy model using the proposed step-wise DPO loss and the step-wise preference pairs, as described in Section 3.2.
+
+During the inference, a well-trained PRM can guide the decoding process and enhance the model's performance. Therefore, in addition to the standard greedy decoding, we explore three alternative decoding methods: (1) Self-Consistency (SC) (Wang et al., 2022): given a problem in the test set, we sample  $K$  candidate solutions from the policy
+
+model. Instead of relying on the first decoded solution, we select the final answer based on majority voting over the answers provided by all sampled solutions. SC is a simple yet highly effective verification strategy. (2) Best-of-N (BoN): we similarly sample  $K$  candidate solutions, score them using the reward model, and select the highest-scoring solution as the final answer. Following previous work (Lightman et al., 2023; Wang et al., 2023c), we use the minimum score across all steps as the final score assigned to a solution by the PRM. (3) Stepwise Beam Search (SBS) (Yu et al., 2023a): the PRM provides feedback at each step, offering more fine-grained guidance. Specifically, for each step, we first sample  $b_{1}$  candidate subsequent steps, then score them using the PRM. The top  $b_{2}$  steps are retained, and decoding continues until  $b_{2}$  final solutions are reached. The detailed algorithm is provided in Appendix A.
+
+# 4 Experiments
+
+# 4.1 Experimental Setup
+
+Backbones. To comprehensively validate the effectiveness of our proposed method, we adopt four popular open-source LLMs as the backbone models: MetaMath-Mistral-7B (Yu et al., 2023b), Llama-3-8B (Touvron et al., 2023), DeepSeekMath-Base-7B (Shao et al., 2024) and Qwen2-7B (Bai et al., 2023). To improve these backbones' reasoning ability, Step-DPO (Lai et al., 2024) finetunes DeepSeekMath-Base-7B and Qwen2-7B on two open-source synthetic math datasets, MetaMath (Yu et al., 2023b) and MMIQC (Liu and Yao, 2024), resulting in DeepSeekMath-Base-SFT $^{2}$  and Qwen2-7B-SFT $^{3}$ , which greatly outperform their previous versions. Following Step-DPO, we further finetune Llama3-8B to produce Llama3-8B-SFT. MetaMath-Mistral-7B has already been finetuned on MetaMath, so no additional finetuning was performed.
+
+Baselines. For closed-source baselines, we compare our approach with OpenAI's GPT-3.5 and GPT-4 (Achiam et al., 2023). We also benchmarked our method against recent high-performing mathematical LLMs, including WizardMath (Luo et al., 2023), MetaMath (Yu et al., 2023b), InternLM-Math-7B (Ying et al., 2024), Qwen2-7B
+
+Instruct(Bai et al., 2023), DeepSeekMath-Instruct (Shao et al., 2024), InternLM-Math-20B (Ying et al., 2024), and Llama-3-70B-Instruct (Touvron et al., 2023).
+
+Additionally, we compare it against DPO (Rafailov et al., 2024) and Step-DPO (Lai et al., 2024). Among these, Lai et al. (2024) publicly release DeepSeekMath-Base-SFT-Step-DPO  $^{4}$  and Qwen2-7B-SFT-Step-DPO  $^{5}$ , which we directly used for evaluation. Additionally, we trained MetaMath-Mistral-7B-Step-DPO and Llama-3-8B-SFT-Step-DPO using their publicly available code and dataset.
+
+Datasets. To ensure a fair comparison, we use the same training dataset  $^{6}$  provided by Step-DPO (Lai et al., 2024), which is synthesized from the training set of GSM8K (Cobbe et al., 2021b) and MATH (Hendrycks et al., 2021). Noted that we only use the problem prompts in this dataset and do not use the step labels marked by GPT-4.
+
+For in-domain evaluation, we conduct experiments on GSM8K and MATH, which contain 1,319 and 5,000 test problems, respectively. We also evaluate on two more challenging out-of-domain (OOD) test sets OCWCourses (OCW) (Lewkowycz et al., 2022) and GaoKao2023 (GK2023) (Liao et al., 2024). OCW contains of 272 undergraduate-level STEM problems requiring multi-step reasoning for most questions, while GK2023 includes 385 mathematics problems from the 2023 Chinese higher education entrance exam, translated into English. Accuracy serves as the evaluation metric.
+
+Implementation Details. During PRM training, we first randomly sample  $M = 32$  solutions for each math problem using Qwen2-7B-SFT and then label them as described in Section 3.3, resulting in the PRM training set. Then, we add a classification-head to Qwen2-7B-SFT and train it on the PRM training set for one epoch. The batch size is 256, and the learning rate is 5e-7.
+
+To build preference learning datasets, we first sample  $M = 32$  solutions for each math problem. The trained PRM then scores each solution, and
+
+we select  $T = 4$  solutions with the highest average rewards and  $T = 4$  with the lowest average rewards to randomly form 16 preference pairs.
+
+During preference learning, the batch size is 64, the learning rate is  $5\mathrm{e} - 7$ ,  $\beta$  is 0.05, and the reward temperature  $\gamma$  is 0.5. We use the AdamW (Loshchilov and Hutter, 2017) optimizer with a linear decay learning rate scheduler and only train one epoch. The warm-up ratio is 0.05.
+
+During the decoding phase, we conduct experiments with two settings for SC and BoN, using  $K = 5$  and  $K = 15$ . For Step-wise Beam Search, to ensure fair comparison, we test two configurations:  $b_{1} = 5$ ,  $b_{2} = 1$  (corresponding to  $K = 5$ ) and  $b_{1} = 5$ ,  $b_{2} = 3$  (corresponding to  $K = 15$ ). The sampling temperature is set to 0.8.
+
+All the experiments are conducted on a server equipped with 8 NVIDIA A100-80GB GPUs and 512GB of system RAM. The implementation frameworks are PyTorch (Paszke et al., 2017), DeepSpeed (Rasley et al., 2020), and Huggingface (Wolf et al., 2019).
+
+# 4.2 Main Results
+
+Table 1 provides a comprehensive comparison of various models on both MATH and GSM8K, including open-source and closed-source LLMs. We find that: (1) Consistent with previous studies (Pal et al., 2024), DPO exhibits notable instability. Its performance shows slight degradation on MetaMath-Mistral-7B and Llama-3-8B-SFT backbones, while the accuracy drops sharply to around  $20\%$  on Qwen2-7B-SFT. It achieves a slight performance improvement only when applied to the DeepSeekMath-Base-SFT. (2) Step-DPO achieves only minimal improvements across all backbones, with gains generally around  $1\%$  and, in some settings, even slight performance drops. We evaluate the publicly released Step-DPO model using its official script, and the results may differ slightly from those reported in the Step-DPO paper. Similar issues have also been observed by other researchers [7]. (3) Our Full-Step-DPO consistently outperforms Step-DPO across all backbones. Specifically, when applied to MetaMath-Mistral-7B and Llama-3-8B-SFT, our model achieves improvements of approximately  $2.3\%$  to  $3.7\%$ , while applied to the stronger backbones, DeepSeekMath-Base-SFT and Qwen2-7B-SFT, our method still delivers gains exceeding
+
+<table><tr><td>Model</td><td>MATH (%)</td><td>GSM8K (%)</td></tr><tr><td>GPT-3.5</td><td>34.1</td><td>80.8</td></tr><tr><td>GPT-4</td><td>53.6</td><td>93.6</td></tr><tr><td>WizardMath</td><td>10.7</td><td>54.9</td></tr><tr><td>MetaMath</td><td>19.8</td><td>66.5</td></tr><tr><td>InternLM-Math-7B</td><td>34.6</td><td>78.1</td></tr><tr><td>Qwen2-7B-Instruct</td><td>49.6</td><td>82.3</td></tr><tr><td>DeepSeekMath-Instruct</td><td>46.8</td><td>82.9</td></tr><tr><td>InternLM-Math-20B</td><td>37.7</td><td>82.6</td></tr><tr><td>Llama-3-70B-Instruct</td><td>50.4</td><td>93.0</td></tr><tr><td>MetaMath-Mistral-7B</td><td>28.2</td><td>77.7</td></tr><tr><td>+ DPO</td><td>24.8 -3.4</td><td>70.7 -7.0</td></tr><tr><td>+ Step-DPO</td><td>28.9 +0.7</td><td>79.6 +1.9</td></tr><tr><td>+ Full-Step-DPO</td><td>30.5 +2.3</td><td>81.4 +3.7</td></tr><tr><td>Llama-3-8B-SFT</td><td>32.6</td><td>78.5</td></tr><tr><td>+ DPO</td><td>23.4 -9.2</td><td>62.3 -16.2</td></tr><tr><td>+ Step-DPO</td><td>31.8 -0.8</td><td>80.1 +1.6</td></tr><tr><td>+ Full-Step-DPO</td><td>35.0 +2.4</td><td>82.0 +3.5</td></tr><tr><td>DeepSeekMath-Base-SFT</td><td>51.7</td><td>86.4</td></tr><tr><td>+ DPO</td><td>51.7 -0</td><td>87.3 +0.9</td></tr><tr><td>+ Step-DPO</td><td>52.9 +1.2</td><td>86.6 +0.2</td></tr><tr><td>+ Full-Step-DPO</td><td>53.2 +1.5</td><td>87.9 +1.5</td></tr><tr><td>Qwen2-7B-SFT</td><td>53.9</td><td>88.3</td></tr><tr><td>+ DPO</td><td>20.0 -23.9</td><td>27.3 -61.0</td></tr><tr><td>+ Step-DPO</td><td>54.9 +1.0</td><td>88.4 +0.1</td></tr><tr><td>+ Full-Step-DPO</td><td>55.4 +1.5</td><td>89.3 +1.0</td></tr></table>
+
+$1\%$ . These results clearly demonstrate the effectiveness of our proposed approach, which considers all steps in the reasoning process rather than focusing on solution-wise preferences or only a single step. A case study can be found in Appendix B.3.
+
+# 4.3 Results on OOD Datasets
+
+To further demonstrate the superiority of Full-Step-DPO, we evaluate the models on OOD datasets GK2023 and OCW, as shown in Table 2. On these competition-level math problems, DPO and Step-DPO often exhibit performance degradation under various settings, while our Full-Step-DPO consistently achieves performance improvements. The only exception occurs on the OCW dataset with MetaMath-Mistral-7B, where Full-Step-DPO shows a slight  $0.8\%$  drop in accuracy. However, this drop is notably smaller than  $3.0\%$  with DPO and the  $3.7\%$  with Step-DPO. These results demonstrate the superior stability and resilience of Full-Step-DPO, particularly in handling challenging mathematical reasoning tasks. More experimental results on additional datasets can be found in Appendix B.1.
+
+Table 1: Performance comparison of various models on MATH and GSM8K with greedy decoding.  
+
+<table><tr><td>Model</td><td>GK2023 (%)</td><td>OCW (%)</td></tr><tr><td>MetaMath-Mistral-7B</td><td>15.8</td><td>10.7</td></tr><tr><td>+ DPO</td><td>15.8 -0</td><td>7.7 -3.0</td></tr><tr><td>+ Step-DPO</td><td>15.1 -0.7</td><td>7.0 -3.7</td></tr><tr><td>+ Full-Step-DPO</td><td>20.5 +4.7</td><td>9.9 -0.8</td></tr><tr><td>Llama-3-8B-SFT</td><td>20.5</td><td>12.5</td></tr><tr><td>+ DPO</td><td>11.7 -8.8</td><td>9.9 -2.6</td></tr><tr><td>+ Step-DPO</td><td>19.7 -0.8</td><td>13.6 +1.1</td></tr><tr><td>+ Full-Step-DPO</td><td>22.1 +1.6</td><td>15.1 +2.6</td></tr><tr><td>DeepSeekMath-Base-SFT</td><td>30.4</td><td>19.1</td></tr><tr><td>+ DPO</td><td>31.2 +0.8</td><td>18.4 -0.7</td></tr><tr><td>+ Step-DPO</td><td>31.2 +0.8</td><td>18.0 -1.1</td></tr><tr><td>+ Full-Step-DPO</td><td>31.7 +1.3</td><td>20.2 +1.1</td></tr><tr><td>Qwen2-7B-SFT</td><td>33.0</td><td>15.8</td></tr><tr><td>+ DPO</td><td>8.8 -24.2</td><td>8.1 -7.7</td></tr><tr><td>+ Step-DPO</td><td>32.5 -0.5</td><td>15.8 -0</td></tr><tr><td>+ Full-Step-DPO</td><td>33.5 +0.5</td><td>18.4 +2.6</td></tr></table>
+
+Table 2: Performance comparison on out-of-domain math problems.
+
+# 4.4 Results on Various Verification Strategies
+
+Figure 3 presents the performance of different verification strategies on GSM8K under two settings:  $K = 5$  and  $K = 15$ . We find that: (1) SC serves as a simple yet powerful validation method that significantly improves performance across all models. Even for the high-performing Qwen2-7B-SFT, which achieves an accuracy of  $89.3\%$  with greedy decoding, SC further improves the accuracy to  $93\%$  when the sampling size  $K = 15$ . This result is already comparable to GPT-4's accuracy of  $93.6\%$ . (2) Compared to SC, BoN often achieves further improvements on MetaMath-Mistral-7B and Llama-3-8B-SFT. However, on the highly capable DeepSeekMath-Base-SFT and Qwen2-7B-SFT, BoN underperforms SC, indicating that the benefits of the reward model diminish for very strong baseline models. (3) SBS performs worse than both SC and BoN across most settings, yet consistently surpasses Greedy decoding, aligning with findings from previous studies (Yu et al., 2023a; Khalifa et al., 2023). This may be because, during the early stages of inference, the reward model struggles to effectively distinguish the correctness of steps.
+
+# 4.5 Analysis of PRMs
+
+As discussed in Section 3.3, the quality of the PRM may be influenced by the sampling number  $M$  and simulation number  $N$ . To assess this, we conducted a controlled experiment with fixed  $M = 32$  and varying  $N$ . The trained PRM is then used to guide the decoding of MetaMath-Mistral-7B-Full-Step
+
+![](images/dfc7c378f85279b5e5608382a0092fda43f8faeae674c1fe48ce57045ab05165.jpg)
+
+![](images/36cb754288f4f480d9b2ef03e68c13a4604700801cb677fe1042d0c3f6823244.jpg)  
+Figure 3: Performance comparison of various verifications on GSM8K, with all models trained using our Full-Step-DPO.
+
+DPO using BoN strategy with  $K = 15$ .
+
+As shown in Figure 4, the purple bars indicate the A100-Hours cost for constructing PRM training data, while blue and orange lines show GSM8K and MATH accuracy. When  $N = 1$ , which represents our proposed method, the model achieves competitive performance with  $85.3\%$  accuracy on GSM8K and  $34.2\%$  on MATH, while only requiring 8.5h with a time complexity of  $O(M)$ . As  $N$  increases, we observe that the performance fluctuation remains relatively minimal (within approximately  $1\%$  range), while the sampling cost grows substantially with a complexity of  $O(MNK)$ . This empirical evidence suggests that our approach with  $N = 1$  achieves a good balance between sampling efficiency and model performance. Additionally, we provide a comprehensive comparison between our PRM and other publicly available PRMs in Appendix B.2.
+
+# 4.6 Sensitivity of Hyperparameters
+
+In step-wise DPO loss, the reward temperature  $\gamma$  reflects the level of trust in the PRM. As  $\gamma$  increases, the PRM model has a greater impact on the gradients. When  $\gamma \rightarrow 0$ , it indicates complete distrust in the PRM model, assigning equal weight coefficient to all steps, degrading in vanilla DPO. Conversely, when  $\gamma \rightarrow \infty$ , the loss function optimizes only the single step with the maximum or minimum reward in the solution, similar to Step-DPO. Figure 5 presents the accuracy of MetaMath-Mistral-7B-Full-Step-DPO
+
+![](images/1ea9128f17b92b6c7ff46027421834ad91b4cbd4513550e934253b1dcd20d856.jpg)  
+Figure 4: Accuracy of MetaMath-Mistral-7B-Full-Step-DPO using BoN decoding with  $K = 15$ . The PRM uses a fixed sampling number  $M = 32$ , while the simulation number  $N$  varies. Purple bars indicate the A100-Hours cost for constructing PRM training data.
+
+![](images/2e7936fa99ff31f2c6a7b412a9d798f690318a4458a7b0459caa0c4884dd34cf.jpg)  
+Figure 5: Accuracy of MetaMath-Mistral-7B-Full-Step-DPO with different reward temperature  $\gamma$ .
+
+with different  $\gamma$  values. The experimental results indicate that introducing a PRM to weight the gradients indeed effectively enhances optimization efficiency and improves performance. Additionally, this experiment demonstrates that there is a sweet spot for the reward temperature  $\gamma$ ; excessively high or low  $\gamma$  will reduce accuracy.
+
+# 5 Conclusion
+
+In this work, we propose Full-Step-DPO, a novel framework for mathematical reasoning that optimizes each step in the entire reasoning chain using step-wise rewards. To achieve this, we train a self-supervised Process Reward Model to automatically score reasoning steps, eliminating reliance on external annotations. We also propose a novel Step-Wise DPO Loss that dynamically updates gradients based on the rewards for individual steps, enabling step-wise optimization and enhancing the reasoning ability of policy models. Experimental results on various benchmarks validate the effectiveness of Full-Step-DPO, paving the way for its application to other reasoning-intensive tasks.
+
+# Limitations
+
+While we have conducted comprehensive experiments to demonstrate the effectiveness of Full-Step-DPO, several limitations remain. First, recent advancements suggest that generative reward models outperform the discriminative reward model used in this work. Exploring how generative reward models can further enhance mathematical reasoning capabilities would be a valuable direction for future research. Second, during preference data construction, the current strategy of selecting samples based on average reward is relatively simple. Investigating more advanced sample selection strategies may lead to further improvements. Finally, the step-wise DPO loss proposed in this paper is highly adaptable to other reasoning tasks, such as code generation. Conducting experiments on a broader range of tasks would provide additional evidence of the advantages of our method.
+
+# Acknowledgement
+
+This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-TC2022-005). We also wish to extend their heartfelt gratitude to the Sea AI Lab for their generous support in providing the necessary equipment and computational resources critical for the successful completion of this research.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q Jiang, Jia Deng, Stella Biderman, and Sean Welleck. 2023. Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631.  
+Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.  
+Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024. Step-level value preference optimization for mathematical reasoning. arXiv preprint arXiv:2406.10858.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias
+
+Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021a. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021b. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.  
+Rémi Coulom. 2006. Efficient selectivity and backup operators in monte-carlo tree search. In International conference on computers and games, pages 72-83. Springer.  
+Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306.  
+Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874.  
+Mohammad Javad Hosseini, Hannaneh Hajishirzi, Oren Etzioni, and Nate Kushman. 2014. Learning to solve arithmetic word problems with verb categorization. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 523-533.  
+Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2023. Large language models cannot self-correct reasoning yet. arXiv preprint arXiv:2310.01798.  
+Fangkai Jiao, Chengwei Qin, Zhengyuan Liu, Nancy F Chen, and Shafiq Joty. 2024. Learning planning-based reasoning by trajectories collection and process reward synthesizing. arXiv preprint arXiv:2402.00658.  
+Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169.  
+Muhammad Khalifa, Lajanugen Logeswaran, Moontae Lee, Honglak Lee, and Lu Wang. 2023. Grace: Discriminator-guided chain-of-thought reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15299-15328.  
+Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xi-angru Peng, and Jiaya Jia. 2024. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. arXiv preprint arXiv:2406.18629.  
+Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. 2022. Solving quantitative reasoning problems with language models. Advances
+
+in Neural Information Processing Systems, 35:3843-3857.  
+Minpeng Liao, Wei Luo, Chengxi Li, Jing Wu, and Kai Fan. 2024. Mario: Math reasoning with code interpreter output-a reproducible pipeline. arXiv preprint arXiv:2401.08190.  
+Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let's verify step by step. arXiv preprint arXiv:2305.20050.  
+Haoxiong Liu and Andrew Chi-Chih Yao. 2024. Aug-mentation math word problems via iterative question composing. arXiv preprint arXiv:2401.09003.  
+Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101.  
+Zimu Lu, Aojun Zhou, Ke Wang, Houxing Ren, Weikang Shi, Junting Pan, Mingjie Zhan, and Hongsheng Li. 2024. Step-controlled dpo: Leveraging stepwise error for enhanced mathematical reasoning. arXiv preprint arXiv:2407.00782.  
+Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583.  
+Xin Mao, Feng-Lin Li, Huimin Xu, Wei Zhang, Wang Chen, and Anh Tuan Luu. 2024a. As simple as fine-tuning: Llm alignment via bidirectional negative feedback loss. arXiv preprint arXiv:2410.04834.  
+Xin Mao, Feng-Lin Li, Huimin Xu, Wei Zhang, and Anh Tuan Luu. 2024b. Don't forget your reward values: Language model alignment via value-based calibration. arXiv preprint arXiv:2402.16030.  
+Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772.  
+Arindam Mitra, Hamed Khanpour, Corby Rosset, and Ahmed Awadallah. 2024. Orca-math: Unlocking the potential of slms in grade school math. arXiv preprint arXiv:2402.14830.  
+Arka Pal, Deep Karkhanis, Samuel Dooley, Manley Roberts, Siddartha Naidu, and Colin White. 2024. Smaug: Fixing failure modes of preference optimisation with dpo-positive. arXiv preprint arXiv:2402.13228.  
+Fengjun Pan, Xiaobao Wu, Zongrui Li, and Anh Tuan Luu. 2024. Are LLMs good zero-shot fallacy classifiers? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14338-14364, Miami, Florida, USA. Association for Computational Linguistics.
+
+Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch. NIPS 2017 Workshop Autodiff Submission.  
+Arkil Patel, Satwik Bhattachamishra, and Navin Goyal. 2021. Are nlp models really able to solve simple math word problems? arXiv preprint arXiv:2103.07191.  
+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.  
+Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 3505-3506.  
+Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, YK Li, Y Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.  
+Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Schärli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210-31227. PMLR.  
+Yifan Song, Weimin Xiong, Dawei Zhu, Cheng Li, Ke Wang, Ye Tian, and Sujian Li. 2023. Rest-gpt: Connecting large language models with real-world applications via restful apis. arXiv preprint arXiv:2306.06624.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275.  
+Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023a. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291.  
+Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. 2023b. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926.
+
+Peiyi Wang, Lei Li, Zhihong Shao, RX Xu, Damai Dai, Yifei Li, Deli Chen, Y Wu, and Zhifang Sui. 2023c. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. CoRR, abs/2312.08935.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.  
+Xiaobao Wu. 2025. Sailing ai by the stars: A survey of learning from rewards in post-training and test-time scaling of large language models. arXiv preprint arXiv:2505.02686.  
+Xiaobao Wu, Thong Nguyen, and Anh Tuan Luu. 2024a. A survey on neural topic models: Methods, applications, and challenges. Artificial Intelligence Review.  
+Xiaobao Wu, Thong Thanh Nguyen, Delvin Ce Zhang, William Yang Wang, and Anh Tuan Luu. 2024b. FASTopic: Pretrained transformer is a fast, adaptive, stable, and transferable topic model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.  
+Xiaobao Wu, Liangming Pan, William Yang Wang, and Anh Tuan Luu. 2024c. AKEW: Assessing knowledge editing in the wild. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15118-15133, Miami, Florida, USA. Association for Computational Linguistics.  
+Xiaobao Wu, Liangming Pan, Yuxi Xie, Ruiwen Zhou, Shuai Zhao, Yubo Ma, Mingzhe Du, Rui Mao, Anh Tuan Luu, and William Yang Wang. 2024d. Antileak-bench: Preventing data contamination by automatically constructing benchmarks with updated real-world knowledge. arXiv preprint arXiv:2412.13670.  
+Zhenyu Wu, Qingkai Zeng, Zhihan Zhang, Zhaoxuan Tan, Chao Shen, and Meng Jiang. 2024e. Enhancing mathematical reasoning in llms by stepwise correction. arXiv preprint arXiv:2410.12934.  
+Yuxi Xie, Anirudh Goyal, Wenyue Zheng, Min-Yen Kan, Timothy P Lillicrap, Kenji Kawaguchi, and Michael Shieh. 2024a. Monte carlo tree search boosts reasoning via iterative preference learning. arXiv preprint arXiv:2405.00451.  
+Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2024b. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems, 36.
+
+Yifan Xu, Xiao Liu, Xinghan Liu, Zhenyu Hou, Yueyan Li, Xiaohan Zhang, Zihan Wang, Aohan Zeng, Zhengxiao Du, Wenyi Zhao, et al. 2024. Chatglmmath: Improving math problem-solving in large language models with a self-critique pipeline. arXiv preprint arXiv:2404.02893.  
+Huaiyuan Ying, Shuo Zhang, Linyang Li, Zhejian Zhou, Yunfan Shao, Zhaoye Fei, Yichuan Ma, Jiawei Hong, Kuikun Liu, Ziyi Wang, et al. 2024. Internlm-math: Open math large language models toward verifiable reasoning. arXiv preprint arXiv:2402.06332.  
+Fei Yu, Anningzhe Gao, and Benyou Wang. 2023a. Outcome-supervised verifiers for planning in mathematical reasoning. arXiv preprint arXiv:2311.09724.  
+Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 2023b. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284.  
+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.
+
+# A Step-wise Beam Search
+
+Unlike conventional beam search, which relies on token-level probabilities, our method integrates the reward model with an associated reranking criterion. This enables for step-wise beam search (SBS) (Yu et al., 2023a; Chen et al., 2024), effectively selecting the preferred solution path in mathematical reasoning, while incurring a lower computational cost compared to Monte Carlo Tree Search. Specifically, for each step  $t$ , suppose the sampling size is  $b_{1}$ , the policy model  $\pi_{\theta}$  produces a set of candidate steps  $\mathbb{S}^{(1:t+1)} = \{S_i^{(1:t+1)}\}_{i=1}^{b_1}$ , where  $S_i^{(1:t+1)} = [s_i^1, \dots, s_i^{t+1}]$  is the  $i$ -th partial solution up to step  $t+1$ . Given the PRM  $\pi_r$  that can score each step, we select the top-scoring steps with beam size  $b_{2}$ . The algorithm is detailed in Algorithm 1. By focusing on the quality of each reasoning step rather than just the final solution, our method enhances the overall reasoning capabilities of the model.
+
+Algorithm 1 Step-wise Beam Search  
+1: Input: Math problem  $q$ , Sampling size  $b_{1}$ , Beam size  $b_{2}$ , Maximum step  $C$   
+2: Output: Best solution for  $q$   
+3: Models: Policy model  $\pi_{\theta}$  and PRM  $\pi_r$   
+4: function STEPLEVELBEAMSEARCH( $q, b_1, b_2, C$ )  
+5: Initialize step sequences  $\mathbb{S} \gets \{\}$   
+6: Use  $\pi_{\theta}$  to sample initial steps  $\{s_1^1, \ldots, s_{b_1}^1\}$   
+7: Use  $\pi_r$  to score all initial steps  $\{r_1^1, \ldots, r_{b_1}^1\}$   
+8: Select top- $b_1$  steps and add to  $\mathbb{S}$   
+9: Set current step counter  $t \gets 1$   
+10: while  $t < C$  do  
+11: if All sequences in  $\mathbb{S}$  are complete then  
+12: Break  
+13: end if  
+14:  $\mathbb{S}_{\mathrm{new}} \gets \{\}$   
+15:  $\mathbb{R} \gets \{\}$ ;  
+16: for each solution  $\mathcal{S}^{(1:t)}$  in  $\mathbb{S}$  do  
+17: for  $i = 1$  to  $b_1$  do  
+18:  $\mathcal{S}_i^{(1:t + 1)} = \pi_{\theta}(\mathcal{S}^{(1:t)}; q)$   
+19:  $r_i^{(1:t + 1)} = \pi_r(\mathcal{S}_i^{(1:t + 1)}; q)$   
+20:  $\mathbb{S}_{\mathrm{new}} \gets \mathbb{S}_{\mathrm{new}} + \{\mathcal{S}_i^{(1:t + 1)}\}$   
+21:  $\mathbb{R} \gets \mathbb{R} + \{r_i^{(1:t + 1)}\}$   
+22: end for  
+23: end for  
+24:  $\mathbb{S}_{\mathrm{new}} \gets$  top- $b_2$  rewarded solutions in ( $\mathbb{S}_{\mathrm{new}}, \mathbb{R}$ )  
+25:  $\mathbb{S} \gets \mathbb{S}_{\mathrm{new}}$   
+26:  $t \gets t + 1$ ;  
+27: end while  
+28: return solution with highest final reward in  $\mathbb{S}$   
+29: end function
+
+# B Additional Experiments
+
+# B.1 More OOD Datasets
+
+We evaluate our method on five additional OOD mathematical reasoning datasets as our testbed. As shown in Table 3, our Full-Step-DPO consistently improves performance across all datasets, demonstrating the effectiveness and generalization ability of our approach on OOD mathematical reasoning tasks.
+
+- SVAMP (Patel et al., 2021) includes 1000 math questions of up to fourth grade difficulty. These questions can be solved by expressions requiring no more than two operators.
+
+- AddSub (Hosseini et al., 2014) contains 395 math questions that involve addition and subtraction operations.  
+- ASDiv (Miao et al., 2021) contains 2215 English math questions of different problem types. Each question provides the corresponding equation and answer.  
+- GSM-IC2 and GSM-ICM (Shi et al., 2023) are mathematical reasoning datasets containing irrelevant conditions within the problem descriptions each consisting of 1000 problems. Problems in GSM-IC2 require two steps to solve, while problems in GSM-ICM require more than two steps to solve.
+
+<table><tr><td>Model</td><td>SVAMP (%)</td><td>AddSub (%)</td><td>ASDiv (%)</td><td>GSM-IC2 (%)</td><td>GSM-ICM (%)</td><td>Average (%)</td></tr><tr><td>MetaMath-Mistral-7B</td><td>79.1</td><td>86.6</td><td>81.2</td><td>77.9</td><td>76.5</td><td>80.3</td></tr><tr><td>+ DPO</td><td>72.7</td><td>53.4</td><td>73.8</td><td>60.5</td><td>65.2</td><td>65.1</td></tr><tr><td>+ Step-DPO</td><td>80.3</td><td>86.9</td><td>82.7</td><td>76.4</td><td>76.3</td><td>80.5</td></tr><tr><td>+ Full-Step-DPO</td><td>81.7</td><td>88.6</td><td>83.9</td><td>75.9</td><td>76.6</td><td>81.3</td></tr><tr><td>Llama-3-8B-SFT</td><td>82.8</td><td>88.4</td><td>85.0</td><td>80.1</td><td>79.5</td><td>83.2</td></tr><tr><td>+ DPO</td><td>72.2</td><td>60.0</td><td>74.2</td><td>64.1</td><td>66.3</td><td>67.4</td></tr><tr><td>+ Step-DPO</td><td>81.7</td><td>88.5</td><td>85.3</td><td>80.7</td><td>81.2</td><td>83.5</td></tr><tr><td>+ Full-Step-DPO</td><td>82.9</td><td>88.8</td><td>86.4</td><td>82.1</td><td>81.6</td><td>84.4</td></tr><tr><td>DeepSeekMath-Base-SFT</td><td>84.2</td><td>87.6</td><td>91.0</td><td>85.4</td><td>85.2</td><td>86.7</td></tr><tr><td>+ DPO</td><td>85.1</td><td>86.6</td><td>90.6</td><td>85.2</td><td>85.0</td><td>86.5</td></tr><tr><td>+ Step-DPO</td><td>85.3</td><td>85.3</td><td>90.7</td><td>85.9</td><td>86.2</td><td>86.7</td></tr><tr><td>+ Full-Step-DPO</td><td>85.9</td><td>86.6</td><td>91.2</td><td>87.8</td><td>85.3</td><td>87.4</td></tr><tr><td>Qwen2-7B-SFT</td><td>88.7</td><td>92.7</td><td>91.6</td><td>93.7</td><td>91.6</td><td>91.7</td></tr><tr><td>+ DPO</td><td>23.5</td><td>25.6</td><td>27.0</td><td>30.1</td><td>29.5</td><td>27.1</td></tr><tr><td>+ Step-DPO</td><td>88.1</td><td>92.2</td><td>91.8</td><td>93.8</td><td>91.9</td><td>91.6</td></tr><tr><td>+ Full-Step-DPO</td><td>89.5</td><td>93.1</td><td>92.4</td><td>93.6</td><td>92.7</td><td>92.3</td></tr></table>
+
+# B.2 Performance Comparison of Different PRMs
+
+We compare our PRM with Math-Shepherd-PRM-7B using MetaMath-Mistral-7B as the base model under the Best-of-8 decoding strategy. During decoding, we sampled eight responses with temperature  $= 0.7$ . For both PRMs, each response was scored by the minimum reward among its reasoning steps, and the response with the highest score was selected as the final output.
+
+As shown in Table 4, our PRM achieves  $84.3\%$  accuracy on GSM8K and  $33.2\%$  on MATH, outperforming Math-Shepherd-PRM-7B (82.4% on GSM8K, 32.6% on MATH). While these results suggest the effectiveness of our approach, we acknowledge that the comparison may not be entirely comprehensive due to differences in backbone architecture and training data between the PRMs. Nevertheless, these findings, combined with our analysis in Section 4.5, demonstrate that our PRM can achieve competitive performance while significantly reducing the computational cost of PRM construction.
+
+Table 3: Performance comparison on five additional OOD math problems. The best and second-best results are highlighted in bold and underlined, respectively.  
+
+<table><tr><td>Model</td><td>MATH</td><td>GSM8K</td></tr><tr><td>MetaMath-Mistral-7B</td><td>28.2</td><td>77.7</td></tr><tr><td>+ Self-Consistency</td><td>32.4</td><td>81.3</td></tr><tr><td>+ Math-Shepherd-PRM-7B</td><td>32.6</td><td>82.4</td></tr><tr><td>+ Our-PRM-7B</td><td>33.2</td><td>84.3</td></tr></table>
+
+Table 4: Performance comparison of different scoring strategies using Best-of-8 decoding on MATH and GSM8K benchmarks.
+
+# B.3 Case Study
+
+We conducted a detailed case study using a sample problem from the AddSub dataset, as shown in Table 5. We compared outputs from DeepSeekMath-Base-SFT model and its Step-DPO and Full-Step-DPO variants. While both the base SFT model and Step-DPO model made incorrect reasoning by including the cost of unbought CDs, our Full-Step-DPO model generated correct reasoning steps. Our PRM scores accurately reflect the correctness of each reasoning step, with higher scores for correct steps and lower scores for incorrect ones.
+
+<table><tr><td>Question</td><td>For his car, Mike spent $118.54 on speakers and $106.33 on new tires. Mike wanted 3 CD&#x27;s for $4.58 but decided not to. In total, how much did Mike spend on car parts?</td></tr><tr><td>Gold Answer</td><td>224.87</td></tr><tr><td rowspan="3">SFT Response</td><td>Step 1: Mike spent 118.54onspeaker sand106.33 on new tires. So the total cost of these two items is $118.54 + $106.33 = $224.87. Score: 0.9348</td></tr><tr><td>Step 2: Mike wanted to buy 3 CD&#x27;s for $4.58 each, but he decided not to. So the total cost of the CD&#x27;s he didn&#x27;t buy is 3 * $4.58 = $13.74. Score: 0.3646</td></tr><tr><td>Step 3: To find out how much Mike spent in total on car parts, we need to add the total cost of the speakers and tires to the total cost of the CD&#x27;s he didn&#x27;t buy. So the total cost is $224.87 + $13.74 = $238.61. The answer is: 238.61 Score: 0.0427</td></tr><tr><td rowspan="3">Step-DPO Response</td><td>Step 1: First, Mike spent $118.54 on speakers and $106.33 on new tires. So the total cost of these two items is $118.54 + $106.33 = $224.87. Score: 0.9406</td></tr><tr><td>Step 2: Mike also wanted 3 CD&#x27;s, but he decided not to buy them. Each CD costs $4.58, so the total cost of the CD&#x27;s he didn&#x27;t buy is 3 * $4.58 = $13.74. Score: 0.3036</td></tr><tr><td>Step 3: To find out how much Mike spent in total on car parts, we need to add the total cost of the speakers and tires to the total cost of the CD&#x27;s he didn&#x27;t buy. So, the total cost is $224.87 + $13.74 = $238.61. The answer is: 238.61 Score: 0.0445</td></tr><tr><td rowspan="3">Full-Step-DPO Response</td><td>Step 1: Mike spent $118.54 on speakers and $106.33 on new tires. So the total amount he spent on these two items is $118.54 + $106.33 = $224.87. Score: 0.9313</td></tr><tr><td>Step 2: Mike wanted to buy 3 CD&#x27;s for $4.58 each, but he decided not to. So the total amount he would have spent on the CD&#x27;s is 3 * $4.58 = $13.74. Score: 0.7408</td></tr><tr><td>Step 3: To find out how much Mike spent in total on car parts, we need to add the total amount he spent on speakers and tires to the amount he would have spent on CD&#x27;s. So the total amount is $224.87 + $0 = $224.87. The answer is: 224.87 Score: 0.7363</td></tr></table>
+
+Table 5: Comparison of reasoning steps and their corresponding PRM scores across different models for AddSub dataset. The base model is DeepSeekMath-Base-SFT.

2025/Full-Step-DPO_ Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fcd8ba0468af492d3c3bf6d39b6ba44cfa44fdd71f104cd79bdf18a1dee18e2
+size 638316

2025/Fuzzy Speculative Decoding for a Tunable Accuracy-Runtime Tradeoff/b38a18f9-3293-4583-8870-5f7468ebd1d3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dffc175d58a2aa393d146ece99bc21ccc45e8f9803145af9289fdbfde6dba789
+size 588811

2025/Fuzzy Speculative Decoding for a Tunable Accuracy-Runtime Tradeoff/full.md

@@ -0,0 +1,481 @@
+# Fuzzy Speculative Decoding for a Tunable Accuracy-Runtime Tradeoff
+
+Maximilian Holsman
+
+Duke University maximilian.holsman@duke.edu
+
+Yukun Huang
+
+Duke University  
+yukun.huang@duke.edu
+
+Bhuwan Dhingra
+
+Duke University bdningra@cs.duke.edu
+
+# Abstract
+
+Speculative Decoding (SD) enforces strict distributional equivalence to the target model when accepting candidate tokens. While it maintains the target model's generation quality, this strict equivalence limits the speedup achievable by SD and prevents users from trading deviations from the target distribution in exchange for further inference speed gains. To address these limitations, we introduce Fuzzy Speculative Decoding (FSD) - a decoding algorithm that generalizes SD by accepting candidate tokens based on the divergences between the target and draft model distributions. By allowing for controlled divergence from the target model, FSD enables users to flexibly trade generation quality for inference speed. Across several benchmarks, our method is able to achieve significant runtime improvements of over 5 tokens per second faster than SD at only an approximate  $2\%$  absolute reduction in benchmark accuracy. In many cases, FSD is even able to match SD benchmark accuracy at over 2 tokens per second faster, demonstrating that distributional equivalence is not necessary to maintain target model performance. Furthermore, FSD can be seamlessly integrated into existing SD extensions; we demonstrate this by applying FSD to EAGLE-2, greatly enhancing this existing extension's efficiency while allowing it to leverage FSD's tunable quality-speed tradeoff.
+
+# 1 Introduction
+
+Speculative decoding (SD), introduced by Leviathan et al. (2023) and Chen et al. (2023), is a large language model (LLM) inference acceleration algorithm that leverages a smaller, faster draft model to generate sequences of candidate tokens which are then verified and accepted in parallel by a larger target model. The speculative sampling rule that SD employs to determine which candidates to accept enforces a strict equivalence
+
+of the final sampling distribution and the original target model distribution. Thus, by cutting out the expensive sequential generation from the large target model, SD can lead to inference time reductions of around 2-3X while maintaining the same generation quality as the target model.
+
+Despite this impressive speedup, SD suffers from two major flaws. Firstly, in order to maintain strict distributional equivalence to the target model, the SD candidate acceptance rule is overly strict, and in many cases may reject tokens that if accepted would have no impact on final generation quality (Lin et al., 2025), thereby unnecessarily limiting potential speed-ups. Secondly, the enforced distributional equivalence means that users cannot tune the SD acceptance rule to be more or less lenient in its candidate acceptance, preventing users from trading deviations from the target model distribution in exchange for further inference speed gains. However, the flexibility for users to tune their LLM generation along an inference speed -generation quality tradeoff would be highly beneficial in real-world applications, as the relative importance of inference speed compared to generation quality may vary across different contexts within an application.
+
+To address these limitations of SD, we introduce Fuzzy Speculative Decoding - a generalized SD algorithm that determines token acceptance based on the divergence between the target and draft model distributions, allowing users to tune the generation quality - inference time tradeoff of their model. With FSD, users have the flexibility to tune a threshold parameter  $T$  that determines how lenient candidate acceptance should be, and thus can control how much they are willing to deviate from the target model's distribution in exchange for further runtime reductions. As it doesn't enforce strict distributional equivalence, FSD can achieve significant runtime improvements over SD by accepting a higher percentage of candidate tokens.
+
+![](images/70bd63a49367a5ef7f4a569ce88627634ba1f1daa282a9e4cf67c34d044df6e1.jpg)  
+Figure 1: Visual comparison between FSD and SD. SD accepts candidate token with a probability dependent on the candidate's relative likelihood under  $M_D$  and  $M_T$ . FSD determines candidate acceptance deterministically based on whether the divergence between the  $M_D$  and  $M_T$  distributions at the candidate's position exceeds a given threshold  $T$ , allowing users to determine how many candidate tokens to accept by setting the threshold  $T$  accordingly.
+
+We conduct extensive experiments across four diverse benchmarks—spanning factoid QA, math, and coding—using three different model pairs. Our key findings are:
+
+1. FSD matches SD's accuracy while achieving over 2 tokens per second speedup by relaxing strict distributional equivalence.  
+2. FSD enables greater speedups (up to 5 tokens per second) when a slight accuracy tradeoff is acceptable (approximately  $2\%$  absolute drop).  
+3. FSD offers a superior tunability mechanism, enabling a flexible tradeoff between accuracy and inference speed that consistently achieves higher accuracy than assigning queries between the target and draft models based on a predefined proportion.
+
+We also perform a broad range of ablation studies, demonstrating that FSD's performance shares many similarities with SD, including dependence on draft and target model alignment for a given text and the ability to use both sample-based and greedy decoding strategies. We also show that FSD can be applied on top of existing SD extensions like EAGLE-2, bringing the aforementioned tunability to these otherwise inflexible methods.
+
+# 2 Previous works
+
+Several works have sought to improve speculative decoding, primarily by increasing the acceptance rate of draft-generated tokens, including but not limited to (1) Verifying more tokens with tree-structured proposals: Some methods improve efficiency by allowing the draft model to propose tokens in a tree structure, enabling the target model to verify multiple candidates in parallel using tree attention mechanisms. This expands the search space and increases the likelihood of accepting a valid token (Li et al., 2024c,a; Cai et al., 2024; Ankner et al., 2024; Miao et al., 2023; Chen et al., 2024). (2) Aligning the draft model with the target model: Methods include fine-tuning the draft model to mimic the target model's outputs (Zhou et al., 2023), granting the draft model access to additional representation information from the target model (AishwaryaP et al., 2024; Zhang et al., 2024b; Du et al., 2024), or even using a partial version of the target model as the draft model itself—such as using partial layers (Liu et al., 2024a; Elhoushi et al., 2024; Zhang et al., 2024a) or augmenting the target model with lightweight extensions to improve alignment (Monea et al., 2023; Fu et al., 2024; Santilli et al., 2023; Cai et al., 2024). (3) Adaptive candidate length selection: Instead
+
+of fixing the number of candidate tokens per step, some methods allow the draft model to determine when to stop generating (Kim et al., 2023; Huang et al., 2024), or enable the target model to verify tokens before the draft model has finished drafting (Liu et al., 2024b), leading to more flexible and efficient speculative decoding. While these methods enhance SD efficiency, they enforce strict distributional guarantees and offer limited flexibility in balancing accuracy and efficiency. In contrast, our framework demonstrates that such guarantees are unnecessary and provides tunable tradeoffs. Moreover, its flexibility allows seamless integration with existing approaches, paving the way for further research and optimization.
+
+The most similar method to ours is concurrent work Judge Decoding (JD) (Bachmann et al., 2025), an SD variant where a compact module is trained on token embeddings to 'judge' and accept candidate tokens based on correctness rather than strict alignment with the target model. This allows JD to accept more tokens than SD with minimal performance loss. However, JD has two major limitations. First, it generalizes poorly to unseen data, as token acceptance relies on a trained judgment module. Its performance drops significantly on out-of-distribution text (Bachmann et al., 2025). Second, JD requires per-model training, preventing out-of-the-box use for new model pairs. In contrast, FSD is training-free, generalizes across datasets, and can be applied to any model pair out-of-the-box, effectively addressing JD's weaknesses.
+
+# 3 Speculative Decoding
+
+We start by reviewing how SD works in order to properly introduce FSD as an extension of this method.
+
+Consider a larger target model  $M_T$  and a smaller draft model  $M_D$ . The biggest bottleneck when generating from  $M_T$  individually is that tokens are sequentially dependent, and therefore each token will require a full  $M_T$  forward pass to be generate conditional on the previously generated tokens. SD mitigates this bottleneck by first generating a sequence of candidate tokens sequentially from the faster  $M_D$ , and then uses a single  $M_T$  forward pass only to verify which of these tokens to accept. Provided that  $M_D$  is a good enough approximation of
+
+$M_T$  such that a significant portion of these candidates are accepted, the runtime saved by avoiding sequential generation from  $M_T$  outweighs the additional runtime of running  $M_D$ , resulting in an overall speedup. In order to maintain  $M_T$ 's full generation quality, SD accepts candidate tokens based on an acceptance rule that guarantees the final sequence of sampled tokens will still be distributed the same as they would under  $M_T$ .
+
+Specifically, at each SD step  $M_D$  first generates a sequence of  $L$  candidate tokens,  $k = [x_0, x_1 \dots x_L]$ , which are then passed through  $M_T$  to calculate the likelihood of each candidate token  $x_i$  under  $M_T$ . Using this likelihood, each candidate  $x_i$  is accepted with the probability:
+
+$$
+P _ {a c c e p t} (x _ {i}) = \min \left(1, \frac {P _ {M _ {T}} (x _ {i} | x _ {<   i})}{P _ {M _ {D}} (x _ {i} | x _ {<   i})}\right)
+$$
+
+making the final candidate token SD acceptance rule:
+
+$$
+F _ {a c c e p t} (x _ {i}) = \left\{ \begin{array}{l l} 1 & \text {i f} P _ {a c c e p t} (x _ {i}) > y \sim \mathcal {U} (0, 1) \\ 0, & \text {e l s e} \end{array} \right.
+$$
+
+Once SD reaches the first rejection of the candidate sequence, it resamples a token at the rejected candidate position from the adjusted distribution:
+
+$$
+M _ {r e s a m p l e} = P _ {M _ {T}} (x _ {i} | x _ {<   i}) - P _ {M _ {D}} (x _ {i} | x _ {<   i})
+$$
+
+(Note that  $P_{M_T}$  and  $P_{M_D}$  will already have been calculated to determine the acceptance probability.)
+
+By accepting tokens that are more likely under  $M_D$  than under  $M_T$  with a probability of  $\frac{P_{M_T}(x_i|x_{<i})}{P_{M_D}(x_i|x_{<i})}$  and resampling rejected tokens from an adjusted distribution, SD corrects for the bias introduced by  $M_D$ , ensuring that the final distribution remains the same as that of  $M_T$ .
+
+# 3.1 Determining SD speed-ups
+
+The inference speed-up of SD heavily depends on the percentage of candidate tokens accepted. Given a fixed candidate length  $L$ , the more similar the distributions of  $M_{D}$  and  $M_{T}$  tend to be over a given generation, the more frequently candidate tokens will be accepted, and thus the greater the inference acceleration. This makes the speed-ups achieved by SD highly dependent on the distribution of text the model is generating, which we can see in Table 1. This variation in acceptance percentages based
+
+<table><tr><td>Dataset</td><td></td><td colspan="3">Candidate length</td></tr><tr><td></td><td></td><td>5</td><td>10</td><td>15</td></tr><tr><td rowspan="2">CSQA</td><td>Tk. / sec</td><td>9.3</td><td>9.3</td><td>7.9</td></tr><tr><td>% MD Tk.</td><td>75.7</td><td>82.8</td><td>84.6</td></tr><tr><td rowspan="2">GSM8K</td><td>Tk. / sec</td><td>11.3</td><td>13.2</td><td>13.0</td></tr><tr><td>% MD Tk.</td><td>81.5</td><td>89.2</td><td>91.4</td></tr><tr><td rowspan="2">MMLU</td><td>Tk. / sec</td><td>7.2</td><td>7.2</td><td>6.3</td></tr><tr><td>% MD Tk.</td><td>78.7</td><td>85.6</td><td>87.5</td></tr><tr><td rowspan="2">HumanEval</td><td>Tk. / sec</td><td>13.7</td><td>16.0</td><td>16.3</td></tr><tr><td>% MD Tk.</td><td>81.5</td><td>88.7</td><td>91.4</td></tr></table>
+
+Table 1: Inference speeds and percent of tokens originating from  $M_D$  under SD on Llama3.1 8B + 70B. Tk. /s denotes tokens per second;  $\% M_D$  Tk. denotes percentage of total generated tokens originating from  $M_D$ .
+
+on text distributions means that each text will have an optimal candidate length  $L$  for which the SD inference speed is maximized. However, once the optimal  $L$  has been found for the given text distribution, the percentage of tokens accepted is effectively fixed, capping the inference speed of SD to a level beyond which it cannot be increased further. This is the limitation of SD that FSD addresses.
+
+# 4 Fuzzy Speculative Decoding
+
+The defining difference of FSD is that it employs a different token acceptance rule that can be tuned to be more or less lenient in its acceptance decisions based on a threshold parameter  $T$ , which can be arbitrarily set by the user. This effectively allows users to determine how much they are willing to diverge from the target distribution  $M_T$  in exchange for a higher percentage of candidates accepted, resulting in speed-ups beyond SD.
+
+While SD determines acceptance based on the likelihood of candidate  $x_{i}$  under  $P_{M_T}$  and  $P_{M_D}$ , FSD calculates the distribution-level divergence between these two distributions at each candidate position. Then, based on the tunable divergence threshold  $T$ , FSD will accept a candidate token if the models' divergence at the corresponding position is less than  $T$ . This makes the FSD acceptance rule:
+
+$$
+F _ {a c c e p t} (x _ {i}) = \left\{ \begin{array}{l l} 1 & \text {i f D i v} (P _ {M _ {T}} [ i ], P _ {M _ {D}} [ i ]) <   T \\ 0, & \text {e l s e} \end{array} \right.
+$$
+
+where  $P_{M_T}[i]$  and  $P_{M_D}[i]$  are the  $M_T$  and  $M_D$  next token distributions at candidate position  $i$  respectively.
+
+In the case of candidate token rejection, FSD will sample from  $P_{M_T}[i]$ , that is the original target model distribution at the rejected position, with whatever sampling method the user sets for the generation. The full FSD algorithm is depicted if Figure 1 as a side-by-side comparison with SD.
+
+# 4.1 Intuitive motivation
+
+Just like SD, FSD aims to accept candidate tokens at positions for which  $M_T$  and  $M_D$  are similar. Instead of relying on strict equivalence in final distribution, FSD relies on the fact that across an entire generation,  $M_T$  and  $M_D$  will produce similar tokens when the divergence between their distributions is low. This in turn means that at positions with low divergence, we can likely use tokens sampled from  $M_D$  in place of those sampled from  $M_T$  with minimal impact on the final generation.
+
+By tuning  $T$ , users can directly dictate how lenient candidate acceptance rule should be, thereby implicitly determining how much they are willing to allow the final sampling distribution to diverge from  $M_T$  in exchange for further runtime reductions. In addition, as the FSD acceptance rule becomes more relaxed, users can also increase the candidate length  $L$  past the value that was optimal for SD to realize even further reductions in inference time.
+
+As a general framework, FSD can use any divergence type that relies solely on  $P_{M_T}$  and  $P_{M_D}$ . In this work, we focused on KL divergence, JS divergence, and total variation distance. We define these divergences in Appendix C. We also perform an empirical evaluation of FSD performance under these different divergence types in Appendix E, and find that JS divergence is the best performing divergence type for FSD.
+
+# 4.2 Final divergence from  $M_T$  under FSD
+
+Unlike SD, FSD does not enforce distributional equivalence to  $M_T$ . Tokens generated via FSD are sampled from a distribution that has diverged from  $M_T$  by an amount dependent on the threshold  $T$ . Specifically, when generating a sequence of N tokens, the divergence between FSD sequence-level distribution and the  $M_T$  sequence level distribution is upper bounded by:
+
+$$
+\operatorname {Div}\left(P_{M_{T}}\left(x_{1:N}\right),P_{\mathrm{FSD}}\left(x_{1:N}\right)\right)\leq N\cdot \%_{M_{D}}\cdot T
+$$
+
+where  $P_{\mathrm{FSD}}$  is the distribution of a sequence sampled from  $M_D$  and  $M_T$  using FSD,  $N$  is the
+
+sequence length,  $\%_{M_D}$  is the percentage of final tokens originating from  $M_D$ , and  $T$  is the divergence threshold set by the user. We show the derivation of this bound in Appendix D.
+
+While this bound establishes a theoretical limit on divergence, it doesn't directly indicate how FSD impacts downstream performance. The relationship between sequence-level divergence and generation quality is non-trivial, as performance degradation depends not only on the magnitude of sequence-level divergence, but also on which token the models diverge on. Thus, an empirical evaluation is necessary to quantify how different choices of  $T$  impact model performance.
+
+# 4.3 Reducing FSD to standard SD
+
+As described, FSD accepts tokens purely based on the divergence between distributions. While we will show empirically that a sufficiently low divergence threshold can retrieve SD performance at matching or higher throughput, it may still be desirable to have FSD reduce to SD at a sufficiently low  $T$ . Therefore, we also introduce the following FSD variant - reducible FSD (rFSD) - that accepts tokens identically to SD when  $T = 0$ :
+
+$$
+F _ {\text {a c c e p t}} (x _ {i}) = \left\{ \begin{array}{l l} 1 & \text {i f} \operatorname {D i v} (P _ {M _ {T}} [ i ], P _ {M _ {D}} [ i ]) <   T \\ & \text {o r} P _ {\text {a c c e p t}} (x _ {i}) > y \sim \mathcal {U} (0, 1) \\ 0 & \text {e l s e r e s a m p l e f r o m} M _ {\text {r e s a m p l e}} \end{array} \right.
+$$
+
+When  $T = 0$ ,  $\mathrm{Div}(P_{M_T}[i], P_{M_D}[i]) < T$  will always be false (since divergences are strictly positive for all nonequivalent distributions), reducing the acceptance rule to that of traditional SD  $(P_{\mathrm{accept}}(x_i) > y \sim \mathcal{U}(0,1))$ . In cases of rejection, rFSD would resample from the adjusted distribution  $M_{\mathrm{resample}}$  much like traditional SD would. While our main results will focus on regular FSD, we have included results in appendix B showing that empirically rFSD performs the same as traditional FSD.
+
+# 5 Main experiments
+
+# 5.1 Experiment design
+
+We tested FSD at various thresholds in comparison to SD on a range of benchmarks, reporting benchmark accuracy, inference speed (tokens/second), and average length of accepted candidates sequences for three of these threshold levels (denoted FSD (Low), (Med.), and (High)). We evaluated
+
+on CommonsenseQA (Talmor et al., 2019) for factual knowledge, GSM8K (Cobbe et al., 2021) for math, MMLU (Hendrycks et al., 2021) $^{2}$  for general knowledge and reasoning, and HumanEval (Chen et al., 2021) for code generation. We performed experiments on  $3M_{D} - M_{T}$  model pairs of varying size: Llama3.1 8B + 70B (Grattafori et al., 2024), Gemma2 2B + 27B (Team et al., 2024), and Qwen2.5 7B + 32B (Qwen et al., 2025). All Gemma2 and Qwen2.5 tests were performed on 2 A6000s, while the Llama3.1 tests were performed on 2 A100s. We use a batch size of 1 for all experiments. JS divergence was chosen as the divergence type following a preliminary experiments that indicated it performed the best. An in depth explanation of the experiment design can be found in Appendix F, and the results of our preliminary divergence type comparison in Appendix E.
+
+# 5.2 Implementation
+
+To perform our experiments, we modified huggingface's transformers library (Wolf et al., 2020) to implemented FSD within the library's assisted generation functionality. This allows us to easily test FSD using the transformers library and allows for a fair comparison to SD, which is implemented in the library by default. We share our FSD implementation at https://github.com/maxholsman/fsd.
+
+# 5.3 FSD performance
+
+We present our experimental results in Table 2 and in Figure 2.
+
+FSD generally matches SD accuracy at noticeably faster inference speeds. When setting  $T$  to lower values, FSD's accuracy converges to the level of SD, often reaching this level while accepting more candidate tokens and thereby realizing greater runtime improvements. This clearly demonstrates that in many cases, the distributional equivalence enforced by SD is not necessary to maintain the full  $M_T$  performance level. Particularly notable are the Llama3.1 and Qwen2.5 GSM8K results, in which FSD is able to outperform SD at around 3 and 4 tokens per second faster, respectively.
+
+As mentioned in section 2, many other SD extensions have been able to achieve SD performance at faster generation speeds, so this finding
+
+<table><tr><td></td><td colspan="3">GSM8K</td><td colspan="3">CSQA</td><td colspan="3">MMLU</td><td colspan="3">HumanEval</td></tr><tr><td colspan="13">Llama3.1 8B + 70B</td></tr><tr><td></td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td></tr><tr><td>M_D</td><td>84.6</td><td>31.8</td><td>-</td><td>73.8</td><td>32.5</td><td>-</td><td>72.2</td><td>32.8</td><td>-</td><td>63.2</td><td>33.0</td><td>-</td></tr><tr><td>M_T</td><td>94.9</td><td>8.5</td><td>-</td><td>83.6</td><td>8.9</td><td>-</td><td>86.2</td><td>9.3</td><td>-</td><td>79.1</td><td>9.3</td><td>-</td></tr><tr><td>SD</td><td>95.1</td><td>16.8</td><td>9.7</td><td>84.1</td><td>13.5</td><td>1.96</td><td>84.8</td><td>15.8</td><td>3.37</td><td>77.4</td><td>20.5</td><td>7.6</td></tr><tr><td>FSD (Low)</td><td>95.2</td><td>19.5</td><td>11.8</td><td>84.0</td><td>14.4</td><td>3.32</td><td>84.0</td><td>17.0</td><td>3.9</td><td>78.9</td><td>22.3</td><td>8.5</td></tr><tr><td>FSD (Med.)</td><td>94.3</td><td>21.2</td><td>12.4</td><td>83.7</td><td>17.5</td><td>4.3</td><td>83.0</td><td>18.1</td><td>4.1</td><td>77.6</td><td>23.2</td><td>8.6</td></tr><tr><td>FSD (High)</td><td>93.1</td><td>22.0</td><td>13.5</td><td>82.1</td><td>19.5</td><td>8.14</td><td>82.6</td><td>18.8</td><td>4.2</td><td>77.4</td><td>23.6</td><td>8.9</td></tr><tr><td colspan="13">Gemma2 2B + 27B</td></tr><tr><td></td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td></tr><tr><td>M_D</td><td>57.5</td><td>28.5</td><td>-</td><td>64.6</td><td>31.3</td><td>-</td><td>55.2</td><td>24.3</td><td>-</td><td>40.9</td><td>17.9</td><td>-</td></tr><tr><td>M_T</td><td>90.7</td><td>8.8</td><td>-</td><td>83.0</td><td>9.1</td><td>-</td><td>75.3</td><td>9.4</td><td>-</td><td>75.6</td><td>9.6</td><td>-</td></tr><tr><td>SD</td><td>90.8</td><td>16.2</td><td>5.7</td><td>83.1</td><td>11.5</td><td>2.07</td><td>76.8</td><td>12.2</td><td>2.7</td><td>76.2</td><td>12.4</td><td>3.7</td></tr><tr><td>FSD (Low)</td><td>89.6</td><td>18.4</td><td>6.8</td><td>82.3</td><td>13.9</td><td>2.5</td><td>75.6</td><td>13.3</td><td>2.9</td><td>78.7</td><td>13.6</td><td>4.02</td></tr><tr><td>FSD (Med.)</td><td>88.5</td><td>19.4</td><td>7.1</td><td>81.6</td><td>15.7</td><td>3.2</td><td>75.4</td><td>15.5</td><td>3.5</td><td>77.8</td><td>14.1</td><td>4.2</td></tr><tr><td>FSD (High)</td><td>86.1</td><td>21.5</td><td>11.1</td><td>79.5</td><td>17.5</td><td>3.9</td><td>74.2</td><td>16.1</td><td>3.7</td><td>75.8</td><td>14.3</td><td>4.3</td></tr><tr><td colspan="13">Qwen2.5 7B + 32B</td></tr><tr><td></td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td><td>Acc</td><td>Spd</td><td>ALen</td></tr><tr><td>M_D</td><td>89.9</td><td>34.8</td><td>-</td><td>80.2</td><td>36.6</td><td>-</td><td>71.9</td><td>35.6</td><td>-</td><td>68.1</td><td>26.9</td><td>-</td></tr><tr><td>M_T</td><td>94.9</td><td>8.8</td><td>-</td><td>86.9</td><td>9.1</td><td>-</td><td>82.7</td><td>9.6</td><td>-</td><td>80.9</td><td>9.6</td><td>-</td></tr><tr><td>SD</td><td>95.1</td><td>17.4</td><td>6.6</td><td>86.8</td><td>14.0</td><td>2.7</td><td>82.2</td><td>16.0</td><td>3.2</td><td>82.1</td><td>15.2</td><td>3.7</td></tr><tr><td>FSD (Low)</td><td>94.7</td><td>21.4</td><td>8.2</td><td>86.6</td><td>16.1</td><td>3.3</td><td>82.0</td><td>18.0</td><td>3.7</td><td>81.9</td><td>17.1</td><td>4.3</td></tr><tr><td>FSD (Med.)</td><td>94.2</td><td>22.4</td><td>9.2</td><td>86.1</td><td>19.5</td><td>6.6</td><td>81.6</td><td>19.5</td><td>4.0</td><td>79.0</td><td>17.2</td><td>4.4</td></tr><tr><td>FSD (High)</td><td>94.0</td><td>22.0</td><td>9.3</td><td>85.9</td><td>20.9</td><td>6.9</td><td>81.7</td><td>20.7</td><td>4.46</td><td>78.3</td><td>17.7</td><td>4.6</td></tr></table>
+
+Table 2: Benchmark performance of FSD at varying threshold levels compared to  $M_D$ ,  $M_T$ , and SD. "Acc" refers to the QA accuracy. "Spd" refers to Inference Speed (tokens/sec.). "ALen" refers to the average accepted sequence length.
+
+isn't necessarily unique to FSD. However, these prior methods all still enforce strict distributional equivalence to  $M_T$ , making our findings notable as they demonstrate this equivalence is often not necessary. Furthermore, given this fundamental difference, our method can easily be applied to these existing SD extensions in order to further extend their respective speedups, as we will demonstrate in section 6.4.
+
+FSD achieves even greater runtime improvements over SD when slight accuracy loss is acceptable. As  $T$  increases, FSD is able to achieve runtime speedups far greater than SD while only sacrificing small reductions in benchmark accuracy. The higher the divergence from  $M_T$  we are willing to tolerate when accepting tokens, the greater the runtime improvement over SD. While benchmark accuracy does eventually degrade as  $T$  increases, we note how minimal this deterioration is. For instance, FSD with Llama3.1 8B + 70B on CSQA achieves a 6 token per second increase over the
+
+inference speed of SD in exchange for only a  $2\%$  absolute reduction in accuracy. We expect that in many applications of LLMs, such a runtime improvement would likely justify these small reductions in generation quality.
+
+FSD allows for a previously unattainable accuracy - runtime tunability. The accuracy - runtime tunability of FSD is demonstrated in Figure 2. A model with good tunability should satisfy two key requirements: (1) it should allow flexibility in adjusting the speed-accuracy tradeoff across the speed axis, and (2) it should achieve the highest possible accuracy compared to other methods at the same speed. Unlike SD, which has a fixed efficiency, FSD enables flexible adjustments along the speed axis while maintaining minimal accuracy degradation, thereby meeting the first requirement. To evaluate the second requirement, we introduce a random allocation baseline, where queries are randomly assigned between the target and draft models, allowing tunability by
+
+![](images/44c7b2ed4d398ab0979bba94364b2726205a5098e8002554ff96a2fa70df7f78.jpg)  
+Figure 2: FSD Benchmark accuracy - inference speed tradeoff compared to SD. Results were collected with Llama3.1 8B + 70B as model pair
+
+![](images/8602e82860754d77e7aa0756c236ae3d333bf3d3511d950f84da65027bda984e.jpg)
+
+![](images/5bd41cf86fcc5820d21934160ac5a1f0a8d3106bac366e171c6f701002ed565e.jpg)
+
+![](images/e9cab41c6a336648c33d9dde0944fc0f3f6056cb39085e8cf41eae6ad2bb9e36.jpg)
+
+![](images/47312d786314a6998daff9c5450f3baae074df02289d7ff7b819b6c520131b97.jpg)
+
+adjusting the proportion of queries sent to the target model. We represent this baseline with a greyline interpolating between the target and draft models. As shown in Figure 2, FSD consistently outperforms the random allocation method across all speeds, demonstrating not only its flexibility but also its superior tunability.
+
+In addition to the results above, we also share our full, more detailed results in Appendix A to help characterize benchmark accuracy as a function of  $T$  across the tested model pairs.
+
+# 6 Ablation studies
+
+# 6.1 FSD and SD variation across datasets
+
+As expected, the acceptance percentages and thereby the runtime improvements of both FSD and SD are highly dependent on the benchmark. We observe that FSD follows the same trends in acceptance percentages across datasets that SD does. That is, the benchmarks on which SD accept more candidate tokens (of course at the  $M_T$  accuracy level) are also the benchmarks on which FSD can accept more candidates when set to match this  $M_T$  accuracy level.
+
+This trend points to an underlying difference in draft and target model alignment across datasets which is affecting both methods ability to accept tokens. We illustrate this difference in  $M_D$  and  $M_T$  model alignment across datasets in Figure 3, which shows the distribution of JS divergences between the Llama models on a subset of question from each dataset. As we can see, the divergences are much more heavily skewed to be much lower on datasets for which both SD and FSD accept more tokens, such as GSM8K and HumanEval. Intuitively, this makes sense: the more similar distributions tend to be across a given text generation, the lower their
+
+![](images/4fa3b301eca46e0c16be15f40290d1613e56781ff8f25e79c91b7e684c2674e9.jpg)  
+Figure 3: Distributions of JS divergences between  $M_{D}$  and  $M_{T}$  across tested datasets. Long tail of distributions (JS div.  $\geq 0.3$ ) truncated for better visibility.
+
+divergences, and thus the more candidates FSD will accept at a given threshold. Likewise, the more similar the distributions, the more likely it is that SD accepts a candidate, since the acceptance probabilities will tend to be higher. Thus, it makes sense that both FSD and SD's runtimes follow the same trend across benchmarks.
+
+# 6.2 Selecting  $T$
+
+As previously discussed, the relationship between an FSD threshold  $T$  and both the inference speed and downstream benchmark performance is dependent on the dataset and the candidate length  $L$ . Thus, users will not preemptively know what inference speed and downstream performance correspond to each threshold  $T$  and candidate length  $L$ .
+
+However, we make two key observations that allow users to easily select  $T$  to achieve a desired performance or inference speed. First, we find that the performance level corresponding to a given threshold  $T$  loosely generalizes across datasets, giving users a good starting point when setting
+
+$T$  on an unknown distribution. Second, we find that  $T$  can be tuned to achieve a desired inference speed with a dev set as small as just 8-16 questions.
+
+Table 3 shows the performance of FSD for all three model pairs at a single selected threshold held constant across datasets for each pair. We can see that for all three pairs, FSD with this constant threshold consistently achieves approximately SD accuracy at around 1-3 tokens per second faster than SD across all datasets. Thus, similar to how certain candidate lengths are known to be good starting points for SD and can later be tuned based on the specific text distribution, we show that the similar out-of-the-box values thresholds values exist for FSD.
+
+Tuning T on small dev set. We demonstrate that  $T$  can be accurately tuned on a small development (dev) set in Table 4. For each dataset, we sampled 10 dev sets of increasing size (4 - 32 questions) from the respective training splits. Then, for a given threshold, we determine the inference speed of FSD across the dev sets, as well as on the full test set. We then calculate the average difference between the dev set inference speeds and the actual inference speed on the test set as a measure of how accurately  $T$  can be tuned to achieve a given inference speed on the dev sets alone. We can see that across both datasets, even a dev set as small as just 4-8 questions can effectively be used to tune  $T$  with only minor error to "true" inference speed, demonstrating that users can easily tune  $T$  to achieve a desired inference speed with very little overhead.
+
+# 6.3 Greedy decoding vs. sample-based decoding
+
+As described in the experiment setup, we used greedy decoding to generate candidate sequences from  $M_D$ , and used sample-based decoding to sample from the  $M_T$  in the case of candidate rejection. While greedy decoding from  $M_D$  is standard practice when using SD, both greedy and sample-based decoding are regularly used in SD to sample from the adjusted distribution in case of rejection. Thus, the question arises whether FSD is also able to accommodate for greedy decoding, in addition to sample-based, in the case of candidate rejection.
+
+To test this, we evaluated FSD performance on GSM8K and CSQA with greedy decoding and com
+
+![](images/10852c0ce7a62b9005b0cdbe55df32ef3d73d67e15e7977c50c5c862c733e428.jpg)
+
+![](images/96792c100f0ecdcf00a23d5f9945d0b1f2ae04ba0698d3bc1b3a7dba7de68af9.jpg)  
+Figure 4: FSD performance compared on GSM8K and CSQA with greedy decoding from  $M_T$  distribution in case of rejection. SD baselines also used greedy decoding. Model pair used was Gemma2 2B + 27B.
+
+![](images/f06099c3e3e2185f0f45760d74badf650a5dbd75d61a6aa398cb200d69844f8b.jpg)
+
+![](images/49dfe5023a72d0b6dd6ec2b4fdb4d1f8c9b7421f2d371d1031e582d44ec06d13.jpg)  
+Figure 5: Accuracy-routine tradeoff of FSD + EAGLE-2 on Llama 3.1 8B Instruct
+
+pared this performance to that of SD under greedy decoding, to see whether the performance trend is similar to what we observe in Table 2. As we can see in Figure 4, FSD seems to follow the same performance trend observed in the main results under greedy decoding. We can again see FSD converge to SD performance at lower thresholds, and achieve significant runtime improvements at the cost of accuracy at higher thresholds. Again we can also see that the higher model alignment on GSM8K we discussed above allows FSD to achieve more impressive results over SD on this dataset, while the performance on CSQA is slightly weaker. This all is consistent with our main results in Table 2.
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="3">Qwen2.5 7B + 32B</td><td colspan="3">Llama3.1 8B + 70B</td><td colspan="3">Gemma2 2B + 27B</td></tr><tr><td>SD Acc.</td><td>FSD Acc. @ T = 0.4</td><td>Speedup over SD (tokens / second)</td><td>SD Acc.</td><td>FSD Acc. @ T = 0.3</td><td>Speedup over SD (tokens / second)</td><td>SD Acc.</td><td>FSD Acc. @ T = 0.7</td><td>Speedup over SD (tokens / second)</td></tr><tr><td>GSM8K</td><td>95.1</td><td>94.7</td><td>3.4</td><td>95.1</td><td>94.7</td><td>3.7</td><td>90.8</td><td>89.6</td><td>2.1</td></tr><tr><td>CSQA</td><td>86.8</td><td>86.4</td><td>3.6</td><td>84.2</td><td>83.8</td><td>2.8</td><td>83.1</td><td>82.3</td><td>2.4</td></tr><tr><td>MMLU</td><td>82.3</td><td>82.1</td><td>2.8</td><td>84.8</td><td>84.1</td><td>1.2</td><td>76.8</td><td>74.8</td><td>1.9</td></tr><tr><td>HumanEval</td><td>82.1</td><td>81.9</td><td>2.9</td><td>77.4</td><td>77.6</td><td>2.7</td><td>76.2</td><td>77.8</td><td>1.7</td></tr></table>
+
+Table 3: FSD performance comparison across datasets at set thresholds  
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">T</td><td colspan="4">Avg. % error w/ dev. set size n</td><td rowspan="2">Test set spd (t/s)</td></tr><tr><td>n = 4</td><td>n = 8</td><td>n = 16</td><td>n = 32</td></tr><tr><td rowspan="4">CSQA</td><td>0.5</td><td>6.1</td><td>2.6</td><td>3.9</td><td>1.2</td><td>1.4</td></tr><tr><td>0.7</td><td>4.2</td><td>2.9</td><td>2.7</td><td>1.2</td><td>1.5</td></tr><tr><td>0.9</td><td>4.6</td><td>1.9</td><td>2.2</td><td>1.4</td><td>2.0</td></tr><tr><td>1.1</td><td>3.6</td><td>2.7</td><td>2.2</td><td>1.5</td><td>1.3</td></tr><tr><td rowspan="4">GSM8K</td><td>0.5</td><td>3.3</td><td>3.0</td><td>2.5</td><td>2.2</td><td>1.8</td></tr><tr><td>0.7</td><td>5.4</td><td>3.7</td><td>2.2</td><td>1.3</td><td>0.9</td></tr><tr><td>0.9</td><td>4.4</td><td>2.9</td><td>2.3</td><td>1.6</td><td>1.2</td></tr><tr><td>1.1</td><td>4.2</td><td>3.4</td><td>1.2</td><td>1.3</td><td>1.4</td></tr></table>
+
+Table 4: Average percentage error in inference speed between test and dev. set at increasing dev. set sizes
+
+# 6.4 FSD on existing SD extensions
+
+Given its flexibility, FSD can be applied on top of most existing SD extensions, bringing tunability and even further speed-ups to the already accelerated inference that these methods achieve. Since FSD only needs the distributions of the draft and target models to determine acceptance, virtually all SD extensions can be used with FSD. To demonstrate this, we apply FSD on top of EAGLE-2 (Li et al., 2024b), and report the accuracy-routine tradeoff on CSQA and GSM8K in Figure 5. As we can see, FSD preserves the same benefits of tunability and further acceleration by foregoing distributional equivalence, even when applied to already significantly accelerated SD extensions. We expect this to hold true for most other SD extensions.
+
+# 7 Discussion
+
+# 7.1 Potential further developments
+
+Unlike the probabilistic acceptance rule of SD, the FSD acceptance criteria is deterministic given the  $M_D$  and  $M_T$  logits. This means that FSD allows for the generation of a token-level dataset of acceptance / rejected labels, since the FSD acceptance decision relies solely on the  $M_D$  and  $M_T$  distributions at each tokens position. This unlocks the possibility of training a classifier to predict which tokens will be accepted and which will be rejected, based purely on the tokens up to the position being
+
+generated. Such a classifier can be used to dynamically set the candidate length generated by the draft model, reducing the number of rejected tokens at each SD step and thereby further increasing the inference time speed ups.
+
+The second area that we feel has potential for future development is the testing and development of a novel divergence types to identify which candidate tokens should be accepted with limited impact on generation quality. Given that FSD was already able to achieve very impressive results with simple divergence types like KL divergence and JS divergence, we expect that the divergences tailored specifically to this methods are likely to further mitigate the deterioration of generation quality as the acceptance threshold  $T$  increases and allow FSD to maintain quality at even higher generation speeds. Judge Decoding (Bachmann et al., 2025) attempts a similar approach to this by using learned token correctness to determine acceptance, however as discussed this method doesn't generalize, leaving this direction open for further research.
+
+# 8 Conclusion
+
+We have introduced FSD - a modified SD algorithm that can accept divergence allows users to tune how much divergence from  $M_T$  they are willing to accept in exchange for runtime improvement beyond SD. This flexibility to achieve significantly higher runtimes, in addition an ability to match SD generation quality at faster inference in certain scenarios, makes FSD novel alternative to SD that we expect can be valuable in many LLM applications. We have shown that FSD is able to achieve very strong empirical results on-par with SD, and is able to achieve considerably higher generation speeds the cost of only minor deteriorations in generation quality.
+
+# 9 Limitations
+
+The biggest limitation of our method is that it is not preemptively known what threshold  $T$  will result in
+
+what downstream generation performance, as this relationship is highly dependent on the distribution of the text being generated and the candidate length  $L$ . Thus, as we have discussed, a practical application of FSD will have to either perform calibration tests on a text distribution similar to the eventual generation's distribution, or will have to use a potentially suboptimal out-of-the-box value similar to those discussed in section 6.2. However, we have demonstrated that tuning  $T$  can be done with minimal computational overhead on a very small dev. set, and therefore does not pose a substantial barrier to the application of FSD. We also note that SD suffers from a similar reliance on hyperparameter tuning, as its inference speed is highly dependent on using the correct  $L$ . In fact, an incorrect selection of  $L$  can result in SD having no impact on or even decreasing the generation speed compared to  $M_T$ . Therefore, while FSD's reliance on  $T$  does represent an additional hyperparameter sensitivity, it's akin to SD's dependency on  $L$ , and thereby is not a limitation unique to our method.
+
+# 10 Acknowledgements
+
+Yukun Huang is supported by NSF award IIS-2211526.
+
+# References
+
+S AishwaryaP, Pranav Ajit Nair, Yashas Samaga, Toby Boyd, Sanjiv Kumar, Prateek Jain, and Praneeth Netrapalli. 2024. Tandem transformers for inference efficient llms. ArXiv, abs/2402.08644.  
+Zack Ankner, Rishab Parthasarathy, Aniruddha Nrusimha, Christopher Rinard, Jonathan RaganKelley, and William Brandon. 2024. Hydra: Sequentially-dependent draft heads for medusa decoding. ArXiv, abs/2402.05109.  
+Gregor Bachmann, Sotiris Anagnostidis, Albert Pumarola, Markos Georgopoulos, Artsiom Sanakoyeu, Yuming Du, Edgar Schonfeld, Ali Thabet, and Jonas Kohler. 2025. Judge decoding: Faster speculative sampling requires going beyond model alignment. Preprint, arXiv:2501.19309.  
+Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Deming Chen, and Tri Dao. 2024. Medusa: Simple llm inference acceleration framework with multiple decoding heads. Preprint, arXiv:2401.10774.  
+Charlie Chen, Sebastian Borgeaud, Geoffrey Irving, Jean-Baptiste Lespiau, Laurent Sifre, and John Jumper. 2023. Accelerating large language model decoding with speculative sampling. Preprint, arXiv:2302.01318.
+
+Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidi Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino, Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders, Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. 2021. Evaluating large language models trained on code. Preprint, arXiv:2107.03374.  
+Zhuoming Chen, Avner May, Ruslan Svirschevski, Yuhsun Huang, Max Ryabinin, Zhihao Jia, and Beidi Chen. 2024. Sequoia: Scalable, robust, and hardware-aware speculative decoding. ArXiv, abs/2402.12374.  
+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.  
+Cunxiao Du, Jing Jiang, Yuanchen Xu, Jiawei Wu, Sicheng Yu, Yongqi Li, Shenggui Li, Kai Xu, Liqiang Nie, Zhaopeng Tu, and Yang You. 2024. Glide with a cape: A low-hassle method to accelerate speculative decoding. ArXiv, abs/2402.02082.  
+Mostafa Elhoushi, Akshit Shrivastava, Diana Liskovich, Basil Hosmer, Bram Wasti, Liangzhen Lai, Anas Mahmoud, Bilge Acun, Saurabh Agarwal, Ahmed Roman, Ahmed Aly, Beidi Chen, and Carole-Jean Wu. 2024. LayerSkip: Enabling early exit inference and self-speculative decoding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12622-12642, Bangkok, Thailand. Association for Computational Linguistics.  
+Yichao Fu, Peter D. Bailis, Ion Stoica, and Hao Zhang. 2024. Break the sequential dependency of llm inference using lookahead decoding. ArXiv, abs/2402.02057.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste
+
+Roziere, 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, Danny Wyatt, 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, Francisco Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Govind Thattai, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jack Zhang, 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, Karthik Prasad, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Maria Tsimpoukelli, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nikolay Bashlykov, Nikolay Bogoychev, Niladri Chatterji Ning Zhang,Olivier Duchenne Onur Celebi Patrick Alrassy,Pengchuan Zhang,Pengwei Li,Petar Vasic,Peter Weng,Prajjwal Bhargava,Pratik Dubal. Praveen Krishnan,Punit Singh Koura,Puxin Xu Qing He,Qingxiao Dong,Ragavan Srinivasan,Raj Ganapathy,Ramon Calderer,Ricardo Silveira Cabral Robert Stojnic,Roberta Raileanu,Rohan Maheswari Rohit Girdhar,Rohit Patel,Romain Sauvestre,Ronnie Polidoro,Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang,Saghar Hosseini,Sahana Chennabasappa,Sanjay Singh Sean Bell Seohyun Sonia Kim,Sergey Edunov,Shaoliang Nie. Sharan NarangSharath Raparthy Sheng ShenShengye Wan Shruti BhosaleShun ZhangSimon Vandenhende Soumya Batra Spencer Whitman Sten Sootla Stephane Collot Suchin Gururangan Sydney Borodinsky,Tamar Herman,Tara Fowler,Tarek SheashaThomas Georgiou Thomas ScialomTobias Speckbacher,Todor MihaylovTong Xiao,Ujjwal Karn,Vedanuj GoswamiVibhor Gupta,Vignesh Ramanathan,Viktor KerkezVincent Gonguet,Virginie Do,Vish Vogeti,Vitor Albiero,Vladan PetrovicWeiwei Chu,Wenhan Xiong,Wenyin Fu.Whitney MeersXavier Martinet,Xiaodong Wang,Xiaofang Wang,Xiaoqing Ellen Tan Xide XiaXinfeng XieXuchao JiaXuewei Wang,Yaelle Goldschlag,Yashesh Gaur,Yasmine Babaei,Yi Wen
+
+Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Srivastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Amos Teo, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Dong, Annie Franco, Anuj Goyal, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, ChingHsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery, Eleonora Presani, Emily Hahn, Emily Wood, Eric-Tuan Le, Erik Brinkman, Esteban Arcaute, Evan Dunbar, Evan Smothers, Fei Sun Felix Kreuk, Feng Tian Filippos Kokkinos, Firat Ozgenel, Francesco Caggioni Frank Kanayet Frank Seide,Gabriela Medina Flores,Gabriella Schwarz Gada Badeer, Georgia Swee, Gil Halpern Grant Herman Grigory SizovGuangyiZhangGuna Lakshminarayanan,Hakan Inan Hamid Shojanazeri Han Zou Hannah WangHanwen Zha Haroun Habeeb,Harrison Rudolph,Helen Suk,Henry Aspegren Hunter Goldman Hongyuan Zhan Ibrahim Damlaj,Igor Molybog,Igor Tufanov,Ilias Leontiadis Irina-Elena Veliche Itai Gat Jake WeissmanJames Geboski James Kohli Janice Lam Japhet Asher Jean-Baptiste GayaJeff MarcusJeff TangJennifer Chan Jenny Zhen Jeremy Reizenstein Jeremy Teboul Jessica Zhong Jian Jin Jingyi Yang Joe Cummings Jon Carvill Jon Shepard Jonathan McPhie Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu,Kam Hou U,Karan Saxena Kartikay Khandwal Katayoun Zand Kathy Matosich Kaushik Veeraraghavan Kelly Michelena Keqian Li Kiran Jagadeesh Kun HuangKunal Chawla Kyle HuangLailin ChenLakshya Garg Lavender A Leandro Silva Lee Bell Lei Zhang Liangpeng Guo,Licheng Yu,Liron MoshkovichLuca Wehrstedt,Madian Khabsa Manav Avalani Manish Bhatt Martynas Mankus Matan Hasson Matthew Lennie Matthias Reso Maxim Groshev Maxim Naumov Maya Lathi Meghan Keneally Miao Liu Michael L. SeltzerMichal Valko Michelle Restrepo Mihir Patel Mik Vyatskov Mikayel Samvelyan Mike Clark Mike Macey Mike WangMiquel Jubert Hermoso Mo Metanat Mohammad Rastegari Munish Bansal Nandhini Santhanam,Natascha Parks Natasha White,Navyata BawaNayan Singhal Nick Egebo Nicolas Usunier Nikhil Mehta Nikolay Pavlovich
+
+Laptev, Ning Dong, Norman Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, British Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Ragtoham Murthy, Raghu Nayani, Rahul Mitra, Rangaprabhu Parthasarathy, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Russ Howes, Rudy Rinott, Sachin Mehta, Sachin Siby, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Summer Deng, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Koehler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihalescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaojian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait, Zachary DeVito, Zef Rosnbrick, Zhaoduo Wen, Zhenyu Yang Zhiwei Zhao and Zhiyu Ma. 2024. The llama 3 herd of models. Preprint arXiv:2407.21783.
+
+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.
+
+Kaixuan Huang, Xudong Guo, and Mengdi Wang. 2024. Specdec++: Boosting speculative decoding via adaptive candidate lengths. ArXiv, abs/2405.19715.
+
+Sehoon Kim, Karttikeya Mangalam, Suhong Moon, Jitendra Malik, Michael W. Mahoney, Amir Gholami, and Kurt Keutzer. 2023. Speculative decoding with big little decoder. In Neural Information Processing Systems.
+
+Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023. Fast inference from transformers via speculative decoding. Preprint, arXiv:2211.17192.
+
+Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2024a. EAGLE-2: Faster inference of language models with dynamic draft trees. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 7421-7432, Miami, Florida, USA. Association for Computational Linguistics.
+
+Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2024b. Eagle-2: Faster inference of language models with dynamic draft trees. Preprint, arXiv:2406.16858.
+
+Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2024c. Eagle: Speculative sampling requires rethinking feature uncertainty. ArXiv, abs/2401.15077.
+
+Zicheng Lin, Tian Liang, Jiahao Xu, Qiuzhi Lin, Xing Wang, Ruilin Luo, Chufan Shi, Siheng Li, Yujiu Yang, and Zhaopeng Tu. 2025. Critical tokens matter: Token-level contrastive estimation enhances llm's reasoning capability. Preprint, arXiv:2411.19943.
+
+Jiahao Liu, Qifan Wang, Jingang Wang, and Xunliang Cai. 2024a. Speculative decoding via early-exiting for faster LLM inference with Thompson sampling control mechanism. In *Findings of the Association for Computational Linguistics: ACL* 2024, pages 3027-3043, Bangkok, Thailand. Association for Computational Linguistics.
+
+Tianyu Liu, Yun Li, Qitan Lv, Kai Liu, Jianchen Zhu, and Winston Hu. 2024b. Parallel speculative decoding with adaptive draft length. ArXiv, abs/2408.11850.
+
+Xupeng Miao, Gabriele Oliaro, Zhihao Zhang, Xinhao Cheng, Zeyu Wang, Zhengxin Zhang, Rae Ying Yee Wong, Alan Zhu, Lijie Yang, Xiaoxiang Shi, Chunan Shi, Zhuoming Chen, Daiyaan Arfeen, Reyna Abhyankar, and Zhihao Jia. 2023. Speciner: Accelerating large language model serving with tree-based speculative inference and verification. Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3.
+
+Giovanni Monea, Armand Joulin, and Edouard Grave. 2023. Pass: Parallel speculative sampling. *ArXiv*, abs/2311.13581.
+
+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.
+
+Andrea Santilli, Silvio Severino, Emilian Postolache, Valentino Maiorca, Michele Mancusi, Riccardo Marin, and Emanuele Rodola. 2023. Accelerating transformer inference for translation via parallel decoding. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12336-12355, Toronto, Canada. Association for Computational Linguistics.
+
+Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. Commonsenseqa: A question answering challenge targeting commonsense knowledge. Preprint, arXiv:1811.00937.
+
+Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan, Sammy Jerome, Anton Tsitsulin, Nino Vieillard, Piotr Stanczyk, Sertan Girgin, Nikola Momchev, Matt Hoffman, Shantanu Thakoor, Jean-Bastien Grill, Behnam Neyshabur, Olivier Bachem, Alanna Walton, Aliaksei Severyn, Alicia Parrish, Aliya Ahmad, Allen Hutchison, Alvin Abdagic, Amanda Carl, Amy Shen, Andy Brock, Andy Coenen, Anthony Laforge, Antonia Paterson, Ben Bastian, Bilal Piot, Bo Wu, Brandon Royal, Charlie Chen, Chintu Kumar, Chris Perry, Chris Welty, Christopher A. Choquette-Choo, Danila Sinopalnikov, David Weinberger, Dimple Vijaykumar, Dominika Rogozinska, Dustin Herbison, Elisa Bandy, Emma Wang, Eric Noland, Erica Moreira, Evan Senter, Evgenii Eltyshev, Francesco Visin, Gabriel Rasskin, Gary Wei, Glenn Cameron, Gus Martins, Hadi Hashemi, Hanna Klimczak-Plucińska, Harleen Batra, Harsh Dhand, Ivan Nardini, Jacinda Mein, Jack Zhou, James Svensson, Jeff Stanway, Jetha Chan, Jin Peng Zhou, Joana Carrasqueira, Joana Iljazi, Jocelyn Becker, Joe Fernandez, Joost van Amersfoort, Josh Gordon, Josh Lipschultz, Josh Newlan, Ju yeong Ji, Kareem Mohamed, Kartikeya Badola, Kat Black, Katie Millican, Keelin McDonell, Kelvin Nguyen, Kiranbir Sodhia, Kish Greene, Lars Lowe Sjoesund, Lauren Usui, Laurent Sifre, Lena Heuermann, Leticia Lago, Lilly McNealus, Livio Baldini Soares, Logan Kilpatrick, Lucas Dixon, Luciano Martins, Michael Reid, Manvinder Singh, Mark Iverson, Martin Gorner, Mat Velloso, Mateo Wirth, Matt Davidow, Matt Miller, Matthew Rahtz, Matthew Watson, Meg Risdal, Mehran Kazemi, Michael Moynihan, Ming Zhang, Minsuk Kahng, Minwoo Park, Mofi Rahman Mohit Khatwani, Natalie Dao, Nenshad Bardoliwalla, Nesh Devanathan, Neta Dumai, Nilay Chauhan, Oscar Wahltinez, Pankil Botarda Parker Barnes Paul Barham Paul Michel Pengchong Jin Petko Georgiev Phil Culliton Pradeep Kuppala Ramona Comanescu Ramona Merhej Reena Jana Reza Ardeshir Rokni Rishabh Agarwal Ryan Mullins Samaneh Saadat Sara Mc Carthy Sarah Cogan Sarah Perrin Sebastien M. R. Arnold Sebastian Krause Shengyang Dai Shruti Garg Shruti Sheth Sue Ronstrom Susan Chan Timothy Jordan Ting Yu Tom Eccles Tom Hennigan Tomas Kocisky Tulsee Doshi Vihan Jain Vikas Yadav Vilobh Meshram Vishal Dharmadhikari Warren Barkley Wei Wei Wenming Ye Woohyun Han Woosuk Kwon Xiang Xu Zhe Shen Zhitao Gong Zichuan Wei Victor Cotruta Phoebe Kirk Anand Rao Minh Giang Ludovic Peran Tris Warkentin Eli Collins Joelle Barral Zoubin Ghahramani Raia Hadsell D. Sculley Jeanine Banks Anca Dragan
+
+Slav Petrov, Oriol Vinyals, Jeff Dean, Demis Hassabis, Koray Kavukcuoglu, Clement Farabet, Elena Buchatskaya, Sebastian Borgeaud, Noah Fiedel, Armand Joulin, Kathleen Kenealy, Robert Dadashi, and Alek Andreev. 2024. *Gemma 2: Improving open language models at a practical size.* Preprint, arXiv:2408.00118.
+
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Huggingface's transformers: State-of-the-art natural language processing. Preprint, arXiv:1910.03771.
+
+Jun Zhang, Jue Wang, Huan Li, Lidan Shou, Ke Chen, Gang Chen, and Sharad Mehrotra. 2024a. Draft & verify: Lossless large language model acceleration via self-speculative decoding. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11263-11282, Bangkok, Thailand. Association for Computational Linguistics.
+
+Lefan Zhang, Xiaodan Wang, Yanhua Huang, and Ruiwen Xu. 2024b. Learning harmonized representations for speculative sampling.
+
+Yongchao Zhou, Kaifeng Lyu, Ankit Singh Rawat, Aditya Krishna Menon, Afshin Rostamizadeh, Sanjiv Kumar, Jean-François Kagy, and Rishabh Agarwal. 2023. Distillspec: Improving speculative decoding via knowledge distillation. ArXiv, abs/2310.08461.
+
+# A Characterization of benchmark accuracy as function of  $T$
+
+We share our full results with specific values of  $T$  in Tables 5, 6, and 7
+
+# B rFSD performance
+
+We share performance of rFSD in Table 8, demonstrating that this FSD modification performs similarly to traditional FSD.
+
+# C Divergence definitions
+
+# C.1 Kullback-Leibler (KL) Divergence:
+
+$$
+\begin{array}{l} D _ {\mathrm {K L}} \left(P _ {M _ {T}} \| P _ {M _ {D}}\right) = \sum_ {t \in \mathcal {V}} P _ {M _ {T}} (t \mid x) \\ \cdot \log \left(\frac {P _ {M _ {T}} (t \mid x)}{P _ {M _ {D}} (t \mid x)}\right) \\ \end{array}
+$$
+
+where  $\mathcal{V}$  is the vocabulary,  $P_{M_T}(t\mid x)$  is the probability assigned by model  $M_T$  to token  $t$  given context  $x$ ,  $P_{M_D}(t\mid x)$  is the probability assigned by model  $M_D$  to token  $t$  given context  $x$ .
+
+<table><tr><td></td><td>Dataset</td><td>Type</td><td>T</td><td>Accuracy</td><td>Inf. Spd. (tok. / sec)</td><td>Acceptance length</td><td>Candidate length</td><td>Acceptance %</td></tr><tr><td rowspan="35">Llama 3.1 8B + 70B Instruct</td><td rowspan="9">GSM8K</td><td>MT</td><td>-</td><td>94.9</td><td>8.5</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>84.9</td><td>31.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>95.1</td><td>16.8</td><td>9.7</td><td>14.2</td><td>91.4</td></tr><tr><td>FSD</td><td>0.3</td><td>95.2</td><td>19.5</td><td>11.8</td><td>14.2</td><td>92.9</td></tr><tr><td>FSD</td><td>0.4</td><td>94.7</td><td>20.6</td><td>12.1</td><td>14.2</td><td>93.1</td></tr><tr><td>FSD</td><td>0.5</td><td>94.3</td><td>21.2</td><td>12.4</td><td>14.2</td><td>93.3</td></tr><tr><td>FSD</td><td>0.6</td><td>94.2</td><td>21.5</td><td>12.6</td><td>14.2</td><td>93.4</td></tr><tr><td>FSD</td><td>0.7</td><td>93.9</td><td>21.8</td><td>13.5</td><td>14.2</td><td>93.9</td></tr><tr><td>FSD</td><td>0.8</td><td>93.1</td><td>22.0</td><td>13.6</td><td>14.2</td><td>93.9</td></tr><tr><td rowspan="8">CSQA</td><td>MT</td><td>-</td><td>83.6</td><td>8.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>73.8</td><td>32.5</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>84.2</td><td>13.5</td><td>3.0</td><td>4.9</td><td>75.4</td></tr><tr><td>FSD</td><td>0.3</td><td>84.0</td><td>14.4</td><td>3.3</td><td>4.9</td><td>77.5</td></tr><tr><td>FSD</td><td>0.35</td><td>83.0</td><td>15.5</td><td>3.6</td><td>4.9</td><td>79.0</td></tr><tr><td>FSD</td><td>0.4</td><td>83.9</td><td>16.3</td><td>3.9</td><td>4.9</td><td>80.2</td></tr><tr><td>FSD</td><td>0.5</td><td>83.7</td><td>17.5</td><td>4.3</td><td>4.9</td><td>81.8</td></tr><tr><td>FSD</td><td>0.6</td><td>82.3</td><td>18.1</td><td>4.5</td><td>4.9</td><td>82.5</td></tr><tr><td rowspan="9">MMLU</td><td>MT</td><td>-</td><td>86.2</td><td>9.3</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>72.2</td><td>32.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>84.8</td><td>15.8</td><td>3.4</td><td>5.0</td><td>77.4</td></tr><tr><td>FSD</td><td>0.3</td><td>84.1</td><td>15.6</td><td>3.4</td><td>5.0</td><td>77.6</td></tr><tr><td>FSD</td><td>0.35</td><td>83.9</td><td>16.5</td><td>3.6</td><td>5.0</td><td>78.6</td></tr><tr><td>FSD</td><td>0.4</td><td>84.1</td><td>17.0</td><td>3.8</td><td>5.0</td><td>79.6</td></tr><tr><td>FSD</td><td>0.5</td><td>83.0</td><td>18.2</td><td>4.1</td><td>5.0</td><td>80.7</td></tr><tr><td>FSD</td><td>0.6</td><td>82.6</td><td>18.8</td><td>4.2</td><td>5.0</td><td>81.0</td></tr><tr><td>FSD</td><td>0.7</td><td>81.3</td><td>20.5</td><td>11.0</td><td>14.6</td><td>92.1</td></tr><tr><td rowspan="9">HumanEval</td><td>MT</td><td>-</td><td>79.1</td><td>9.3</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>63.2</td><td>33.0</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>77.4</td><td>20.5</td><td>7.6</td><td>9.8</td><td>88.7</td></tr><tr><td>FSD</td><td>0.2</td><td>79.6</td><td>20.7</td><td>8.2</td><td>9.8</td><td>89.4</td></tr><tr><td>FSD</td><td>0.3</td><td>78.9</td><td>22.3</td><td>8.5</td><td>9.8</td><td>89.8</td></tr><tr><td>FSD</td><td>0.4</td><td>77.6</td><td>23.2</td><td>8.8</td><td>9.8</td><td>90.1</td></tr><tr><td>FSD</td><td>0.5</td><td>77.4</td><td>23.6</td><td>8.9</td><td>9.8</td><td>90.3</td></tr><tr><td>FSD</td><td>0.6</td><td>75.0</td><td>24.8</td><td>9.6</td><td>9.8</td><td>90.9</td></tr><tr><td>FSD</td><td>0.7</td><td>73.9</td><td>24.9</td><td>9.6</td><td>9.8</td><td>90.9</td></tr></table>
+
+Table 5: Characterized performance-accuracy tradeoff of FSD with Llama 3.1 8B + 70B Instruct  
+
+<table><tr><td></td><td>Dataset</td><td>Type</td><td>T</td><td>Accuracy</td><td>Inf. Spd. (tok. / sec)</td><td>Acceptance length</td><td>Candidate length</td><td>Acceptance %</td></tr><tr><td rowspan="9">GSM8K</td><td rowspan="9">GSM8K</td><td>MT</td><td>-</td><td>90.7</td><td>8.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>57.5</td><td>28.5</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>90.8</td><td>16.3</td><td>5.7</td><td>9.1</td><td>85.8</td></tr><tr><td>FSD</td><td>0.7</td><td>89.6</td><td>18.4</td><td>6.8</td><td>9.2</td><td>87.8</td></tr><tr><td>FSD</td><td>0.8</td><td>88.8</td><td>18.9</td><td>7.0</td><td>9.2</td><td>88.1</td></tr><tr><td>FSD</td><td>0.9</td><td>88.6</td><td>19.4</td><td>7.1</td><td>9.2</td><td>88.3</td></tr><tr><td>FSD</td><td>1.0</td><td>87.8</td><td>20.3</td><td>7.7</td><td>9.2</td><td>89.0</td></tr><tr><td>FSD</td><td>1.25</td><td>85.9</td><td>21.3</td><td>8.1</td><td>9.2</td><td>89.6</td></tr><tr><td>FSD</td><td>1.5</td><td>83.1</td><td>21.3</td><td>7.9</td><td>9.2</td><td>89.4</td></tr><tr><td rowspan="11">GSQA</td><td rowspan="11">CSQA</td><td>MT</td><td>-</td><td>83.0</td><td>9.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>64.6</td><td>31.3</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>83.1</td><td>11.5</td><td>2.1</td><td>4.7</td><td>68.7</td></tr><tr><td>FSD</td><td>0.5</td><td>82.4</td><td>12.0</td><td>2.2</td><td>4.6</td><td>70.5</td></tr><tr><td>FSD</td><td>0.6</td><td>82.1</td><td>13.0</td><td>2.4</td><td>4.6</td><td>71.7</td></tr><tr><td>FSD</td><td>0.7</td><td>82.3</td><td>13.9</td><td>2.5</td><td>4.5</td><td>73.0</td></tr><tr><td>FSD</td><td>0.8</td><td>81.5</td><td>15.1</td><td>3.2</td><td>4.5</td><td>77.3</td></tr><tr><td>FSD</td><td>0.9</td><td>81.7</td><td>15.7</td><td>3.2</td><td>4.8</td><td>77.8</td></tr><tr><td>FSD</td><td>1.0</td><td>81.1</td><td>16.2</td><td>3.3</td><td>4.5</td><td>78.5</td></tr><tr><td>FSD</td><td>1.25</td><td>79.5</td><td>17.5</td><td>3.9</td><td>4.5</td><td>80.8</td></tr><tr><td>FSD</td><td>1.5</td><td>77.1</td><td>18.0</td><td>3.9</td><td>4.5</td><td>81.1</td></tr><tr><td rowspan="8">MMLU</td><td rowspan="8">MMLU</td><td>MT</td><td>-</td><td>75.3</td><td>9.4</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>55.2</td><td>24.3</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>76.8</td><td>12.2</td><td>2.7</td><td>4.8</td><td>73.6</td></tr><tr><td>FSD</td><td>0.6</td><td>75.6</td><td>13.3</td><td>2.9</td><td>4.8</td><td>74.9</td></tr><tr><td>FSD</td><td>0.7</td><td>74.8</td><td>14.1</td><td>3.0</td><td>4.8</td><td>75.7</td></tr><tr><td>FSD</td><td>0.8</td><td>75.4</td><td>15.5</td><td>3.5</td><td>4.8</td><td>78.4</td></tr><tr><td>FSD</td><td>0.9</td><td>74.2</td><td>15.8</td><td>3.6</td><td>4.8</td><td>78.7</td></tr><tr><td>FSD</td><td>1.0</td><td>74.2</td><td>16.1</td><td>3.7</td><td>4.8</td><td>79.1</td></tr><tr><td rowspan="9">HumanEval</td><td rowspan="9">HumanEval</td><td>MT</td><td>-</td><td>75.6</td><td>9.6</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>40.9</td><td>17.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>76.2</td><td>12.4</td><td>3.7</td><td>4.9</td><td>78.9</td></tr><tr><td>FSD</td><td>0.4</td><td>75.6</td><td>13.2</td><td>4.0</td><td>4.9</td><td>80.1</td></tr><tr><td>FSD</td><td>0.5</td><td>78.7</td><td>13.6</td><td>4.0</td><td>4.9</td><td>80.3</td></tr><tr><td>FSD</td><td>0.6</td><td>77.4</td><td>14.0</td><td>4.1</td><td>4.9</td><td>80.6</td></tr><tr><td>FSD</td><td>0.7</td><td>77.8</td><td>14.1</td><td>4.2</td><td>4.9</td><td>80.9</td></tr><tr><td>FSD</td><td>0.8</td><td>76.0</td><td>14.2</td><td>4.3</td><td>4.9</td><td>81.1</td></tr><tr><td>FSD</td><td>0.9</td><td>75.8</td><td>14.3</td><td>4.3</td><td>4.9</td><td>81.4</td></tr></table>
+
+Table 6: Characterized performance-accuracy tradeoff of FSD with Gemma  ${22}\mathrm{\;B} + {27}\mathrm{\;B}$  Instruct
+
+<table><tr><td></td><td>Dataset</td><td>Type</td><td>T</td><td>Accuracy</td><td>Inf. Spd. (tok. / sec)</td><td>Acceptance length</td><td>Candidate length</td><td>Acceptance %</td></tr><tr><td rowspan="30">Qwen 2.5 7B + 32B Instruct</td><td rowspan="8">GSM8K</td><td>MT</td><td>-</td><td>94.9</td><td>8.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>89.9</td><td>34.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>95.1</td><td>17.4</td><td>6.5</td><td>9.7</td><td>87.1</td></tr><tr><td>FSD</td><td>0.3</td><td>94.8</td><td>20.0</td><td>7.8</td><td>9.7</td><td>89.1</td></tr><tr><td>FSD</td><td>0.4</td><td>94.7</td><td>20.9</td><td>8.0</td><td>9.7</td><td>89.4</td></tr><tr><td>FSD</td><td>0.5</td><td>94.7</td><td>21.4</td><td>8.2</td><td>9.7</td><td>89.6</td></tr><tr><td>FSD</td><td>0.8</td><td>94.3</td><td>22.4</td><td>9.2</td><td>9.7</td><td>90.8</td></tr><tr><td>FSD</td><td>0.9</td><td>93.9</td><td>22.0</td><td>9.3</td><td>9.7</td><td>90.8</td></tr><tr><td rowspan="6">CSQA</td><td>MT</td><td>-</td><td>86.9</td><td>9.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>80.2</td><td>36.6</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>86.1</td><td>14.0</td><td>2.7</td><td>4.9</td><td>73.6</td></tr><tr><td>FSD</td><td>0.4</td><td>86.2</td><td>17.6</td><td>6.3</td><td>9.7</td><td>87.0</td></tr><tr><td>FSD</td><td>0.5</td><td>86.1</td><td>19.5</td><td>6.6</td><td>9.7</td><td>87.5</td></tr><tr><td>FSD</td><td>0.6</td><td>85.9</td><td>20.9</td><td>6.9</td><td>9.7</td><td>88.1</td></tr><tr><td rowspan="7">MMLU</td><td>MT</td><td>-</td><td>82.7</td><td>9.6</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>71.9</td><td>35.6</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>82.3</td><td>16.0</td><td>3.2</td><td>5.0</td><td>76.5</td></tr><tr><td>FSD</td><td>0.3</td><td>82.0</td><td>18.0</td><td>3.7</td><td>5.0</td><td>79.3</td></tr><tr><td>FSD</td><td>0.4</td><td>82.1</td><td>18.8</td><td>3.9</td><td>5.0</td><td>80.0</td></tr><tr><td>FSD</td><td>0.5</td><td>81.6</td><td>19.5</td><td>4.0</td><td>5.0</td><td>80.6</td></tr><tr><td>FSD</td><td>0.6</td><td>81.7</td><td>20.7</td><td>4.5</td><td>4.9</td><td>82.1</td></tr><tr><td rowspan="9">HumanEval</td><td>MT</td><td>-</td><td>80.9</td><td>9.6</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MD</td><td>-</td><td>68.1</td><td>26.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SD</td><td>-</td><td>82.1</td><td>15.2</td><td>3.7</td><td>4.9</td><td>79.0</td></tr><tr><td>FSD</td><td>0.4</td><td>81.9</td><td>17.1</td><td>4.3</td><td>4.9</td><td>81.2</td></tr><tr><td>FSD</td><td>0.5</td><td>79.5</td><td>17.2</td><td>4.4</td><td>4.9</td><td>81.6</td></tr><tr><td>FSD</td><td>0.6</td><td>79.1</td><td>17.4</td><td>4.4</td><td>4.9</td><td>81.8</td></tr><tr><td>FSD</td><td>0.7</td><td>80.5</td><td>17.7</td><td>4.5</td><td>4.9</td><td>82.0</td></tr><tr><td>FSD</td><td>0.8</td><td>79.9</td><td>17.7</td><td>4.5</td><td>4.9</td><td>82.1</td></tr><tr><td>FSD</td><td>0.9</td><td>78.3</td><td>17.7</td><td>4.6</td><td>4.9</td><td>82.3</td></tr></table>
+
+Table 7: Characterized performance-accuracy tradeoff of FSD with Qwen  ${2.57}\mathrm{\;B} + {32}\mathrm{\;B}$  Instruct  
+
+<table><tr><td>Dataset</td><td>Type</td><td>Acc.</td><td>Spd. (t/s)</td><td>ALen</td><td>CLen</td></tr><tr><td rowspan="4">CSQA</td><td>SD</td><td>82.9</td><td>7.78</td><td>2.06</td><td>5</td></tr><tr><td>FSD (Low)</td><td>81.9</td><td>9.82</td><td>2.85</td><td>5</td></tr><tr><td>FSD (Med.)</td><td>80.8</td><td>11.0</td><td>3.03</td><td>5</td></tr><tr><td>FSD (High)</td><td>79.6</td><td>11.7</td><td>3.22</td><td>5</td></tr><tr><td rowspan="4">GSM8K</td><td>SD</td><td>90.0</td><td>10.5</td><td>5.72</td><td>10</td></tr><tr><td>FSD (Low)</td><td>89.3</td><td>12.7</td><td>7.23</td><td>10</td></tr><tr><td>FSD (Med.)</td><td>88.2</td><td>13.0</td><td>7.38</td><td>10</td></tr><tr><td>FSD (High)</td><td>87.0</td><td>13.5</td><td>7.56</td><td>10</td></tr></table>
+
+Table 8: Performance-accuracy tradeoff of rFSD compared to regular SD. Acc. indicates benchmark accuracy, Spd. indicates inference speed in tokens / second, ALen indicates acceptance length, CLen indicates candidate length
+
+# C.2 Jensen-Shannon (JS) Divergence:
+
+$$
+\begin{array}{l} D _ {\mathrm {J S}} \left(P _ {M _ {T}} \| P _ {M _ {D}}\right) = \frac {1}{2} D _ {\mathrm {K L}} \left(P _ {M _ {T}} \| M\right) \\ + \frac {1}{2} D _ {\mathrm {K L}} \left(P _ {M _ {D}} \| M\right) \\ \end{array}
+$$
+
+where  $M(t \mid x)$  is the mixture distribution (average of  $P_{M_T}$  and  $P_{M_D}$ ):
+
+$$
+M (t \mid x) = \frac {P _ {M _ {T}} (t \mid x) + P _ {M _ {D}} (t \mid x)}{2}
+$$
+
+and  $D_{\mathrm{KL}}$  is the Kullback-Leibler divergence as defined above.
+
+# C.3 Total Variation (TV) Distance:
+
+$$
+\begin{array}{l} D _ {\mathrm {T V}} \left(P _ {M _ {T}}, P _ {M _ {D}}\right) = \frac {1}{2} \sum_ {t \in \mathcal {V}} \left| P _ {M _ {T}} (t \mid x) \right. \\ - P _ {M _ {D}} (t \mid x) | \\ \end{array}
+$$
+
+where:  $P_{M_T}(t\mid x)$  and  $P_{M_D}(t\mid x)$  are the probabilities from models  $M_T$  and  $M_D$  respectively, as defined above.
+
+# D Derivation of FSD sequence-level divergence bound
+
+# D.1 KL divergence bound
+
+Assume that  $M^{*}$  denotes a model using FSD to decode from target model  $M_T$  and draft model  $M_D$ . Starting with the sequence-level KL divergence decomposed autoregressively:
+
+$$
+\begin{array}{l} D _ {\mathrm {K L}} \left(P _ {M _ {T}} \| P _ {M ^ {*}}\right) \\ = \sum_ {t = 1} ^ {T} E _ {P _ {M _ {T}}} \left[ D _ {\mathrm {K L}} \left(P _ {M _ {T}} (t \mid x) \| P _ {M ^ {*}} (t \mid x)\right) \right] \\ \end{array}
+$$
+
+where the expectation is taken over the sequence of tokens  $x_{1:t-1}$ . By assumption, at each step when the  $M_D - M_T$  divergence exceeds  $\tau$ , model  $M^*$  will use  $P_{M_T}$  instead of  $P_{M_D}$ , making the divergence between  $P_{M^*}$  and  $P_{M_T}$  0. Let  $p_{\mathrm{use}}$  be the probability that  $P_{M_D}$  is used:
+
+$$
+D _ {\mathrm {K L}} \left(P _ {M _ {T}} (t \mid x) \| P _ {M ^ {*}} (t \mid x)\right) \leq p _ {\text {u s e}} \tau
+$$
+
+Summing over  $T$  steps:
+
+$$
+D _ {\mathrm {K L}} \left(P _ {M _ {T}} \| P _ {M ^ {*}}\right) \leq \sum_ {t = 1} ^ {T} p _ {\text {u s e}} \tau = T p _ {\text {u s e}} \tau
+$$
+
+# D.2 JS divergence bound
+
+The JS divergence is defined as:
+
+$$
+\begin{array}{l} D _ {\mathrm {J S}} \left(P _ {M _ {T}} \| P _ {M ^ {*}}\right) = \frac {1}{2} D _ {\mathrm {K L}} \left(P _ {M _ {T}} \| M\right) \\ + \frac {1}{2} D _ {\mathrm {K L}} \left(P _ {M ^ {*}} \| M\right) \\ \end{array}
+$$
+
+Using the KL decomposition for both terms and applying the same per-step bound  $\tau$  for when  $P_{M_D}$  is used:
+
+$$
+\begin{array}{l} D _ {\mathrm {J S}} \left(P _ {M _ {T}} \| P _ {M ^ {*}}\right) \leq \frac {1}{2} \sum_ {t = 1} ^ {T} p _ {\text {u s e}} \tau + \frac {1}{2} \sum_ {t = 1} ^ {T} p _ {\text {u s e}} \tau \\ = T p _ {\text {u s e}} \tau \\ \end{array}
+$$
+
+# D.3 TV distance bound
+
+The sequence-level TV distance decomposes similarly via subadditivity:
+
+$$
+D _ {\mathrm {T V}} (P _ {M _ {T}}, P _ {M ^ {*}}) \leq
+$$
+
+$$
+\sum_ {t = 1} ^ {T} E _ {P _ {M _ {T}} (x _ {1: t - 1})} [ D _ {\mathrm {T V}} (P _ {M _ {T}} (t \mid x), P _ {M ^ {*}} (t \mid x)) ]
+$$
+
+By assumption, if  $P_{M_D}$  is used, the per-step TV distance is bounded by  $\tau$ :
+
+$$
+D _ {\mathrm {T V}} \left(P _ {M _ {T}} (t \mid x), P _ {M ^ {*}} (t \mid x)\right) \leq p _ {\text {u s e}} \tau
+$$
+
+Summing over  $T$  steps:
+
+$$
+D _ {\mathrm {T V}} \left(P _ {M _ {T}}, P _ {M ^ {*}}\right) \leq \sum_ {t = 1} ^ {T} p _ {\text {u s e}} \tau = T p _ {\text {u s e}} \tau
+$$
+
+Final Result: For all three divergences, the upper bound is:
+
+$$
+D \left(P _ {M _ {T}} \| P _ {M ^ {*}}\right) \leq T p _ {\text {u s e}} \tau .
+$$
+
+# E Divergence comparison under FSD
+
+We referenced in the section 5.1, we performed preliminary tests on the different divergence types to see which divergence was best able to maintain SD accuracy as  $T$  increases. The results of this preliminary experiments can be seen below.
+
+![](images/34ef77fdf20af7748f72961fdb04e2c38190a793b1ea4668f2ac04def71b6983.jpg)  
+(a) FSD performance of different divergence types on GSM8K.
+
+![](images/b75eaf5bd8a702b47fa5e4385e50d3aa70dc298ad02dbe4ef33a6614c4aeaa4d.jpg)  
+(b) FSD performance of different divergence types on CSQA.  
+Figure 6: Comparison of FSD performance on GSM8K and CSQA with different divergence types.
+
+# F In-depth experiment design
+
+Below are more details on the procedure we used to collect our main results. While this procedure was generally followed throughout, some additional data points were collected and reported (e.g., reporting  $T = 0.35$  results).
+
+For each benchmark, we start by empirically determining the approximately optimal SD candidate length  $L$  by testing SD with  $L = [5,10,15,20]$  on a small subset of questions, and select the  $L$  with the fastest inference speed as the candidate length to be used in our SD baseline. We denote this SD optimal candidate length as  $L'$ . We then test FSD with threshold  $T = [0.1,0.2,\dots 0.9,1.0]$  at  $L'$  on the same small subset of question to determine the threshold  $T_{SD}$  that accepts approximately the same percentage of candidate tokens as SD. Starting from this 'equivalent'  $T_{SD}$ , we then evaluate FSD's benchmark performance at threshold increasing in increments of 0.1, until benchmark performance has degraded by approximately 20% of the performance difference between  $M_D$  and  $M_T$ . (e.g. if  $M_T$  scores 90%,  $M_D$  scores 80%,
+
+![](images/5362ad0740afd815e026bd2edce9c8c7f2a7465153f90eddf64f917818d5fdf5.jpg)  
+Figure 7: CSQA performance of regular FSD vs FSD with random token acceptance at varying percentages of  $M_D$  tokens.  $\%$  MD Tok. denotes the percentage of final generated tokens originating from  $M_D$ . Experiment was performed on Gemma2 2B + 27B model pair
+
+we test FSD at increasing  $T$  until accuracy reaches  $90 - ((90 - 80)*0.2) = 88\%$ . For each threshold, we complete three trials, using greedy decoding to generate the candidate sequences from  $M_D$  and sample-based decoding to sample from  $M_T$  in the case of candidate rejection. We use the same sampling strategy for our SD baseline, as this is the default for the huggingface assisted generation implementation we used.
+
+Importantly, as the acceptance percentage increases beyond that of SD,  $L'$  may no longer be the optimal candidate length. Thus, we increased  $L'$  to the next highest length in [5, 10, 15, 20] if we observed that FSD is accepting close to all candidates.
+
+To quantify the performance-routine tunability of our method, we report the FSD benchmark accuracy, inference speed (tokens/second), and average length of accepted candidates sequences at three increasing threshold levels (denoted FSD (Low), (Med.), and (High)). These three levels are meant to simulate scenarios in which users are willing to accept increasing drops in generation quality in exchange for increasing generation speeds.
+
+We would also like to note that a single Llama IT token was accidentally included in the Gemma2 prompt for the CSQA evaluations. We've verified that this erroneous inclusion had minimal impact on the benchmark accuracy, and have therefore retained the previous results.
+
+# G Random baseline
+
+In Table 2, we can clearly see that benchmark accuracy is highly sensitive to the percentage of can
+
+didate tokens accepted. For every benchmark, FSD accuracy is almost identical to SD accuracy when the threshold  $T$  is set such FSD accepts a similar percentage of candidate tokens. This begs the question: is benchmark performance simply a function of the candidate acceptance percentage, irrespective of which tokens are being accepted?
+
+To test this, we performed a random FSD baseline, in which FSD was set to randomly accept a certain percentage of candidate tokens. By doing this, we are able to determine whether the divergences between distributions is an effective method of determining which tokens can be accepted with minimal impact on downstream performance, or whether this performance is mostly determined by how many  $M_D$  tokens are accepted. We report these results in Figure 7. As expected, we can see that FSD with random candidate acceptance does significantly worse than regular divergence-based FSD, even when significantly fewer candidates from  $M_D$  are being accepted. Thus, it does seem that  $M_D - M_T$  divergence is an effective criteria for deciding which candidates to accept, implying that the development of better divergences will likely improve FSD performance even further.

2025/Fuzzy Speculative Decoding for a Tunable Accuracy-Runtime Tradeoff/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acc21a8aa147f6ac6707952db41575532eac6286521d400e5c9db14063772ac1
+size 918531

2025/G2S_ A General-to-Specific Learning Framework for Temporal Knowledge Graph Forecasting with Large Language Models/e8fefd89-b727-4b85-babb-df829aebdf89_content_list.json

@@ -0,0 +1,1942 @@
+[
+    {
+        "type": "text",
+        "text": "G2S: A General-to-Specific Learning Framework for Temporal Knowledge Graph Forecasting with Large Language Models",
+        "text_level": 1,
+        "bbox": [
+            115,
+            79,
+            880,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Long Bai $^{1,2,4}$ , Zixuan Li $^{1,2}$ , Xiaolong Jin $^{1,2,3}$ , Jiafeng Guo $^{1,2,3}$ , Xueqi Cheng $^{1,2,3}$ , Tat-Seng Chua $^{4}$",
+        "bbox": [
+            278,
+            124,
+            715,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Key Laboratory of Network Data Science and Technology, Institute of Computing Technology, Chinese Academy of Sciences",
+        "bbox": [
+            228,
+            159,
+            768,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ State Key Laboratory of AI Safety",
+        "bbox": [
+            352,
+            192,
+            642,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$ School of Computer Science, University of Chinese Academy of Sciences",
+        "bbox": [
+            194,
+            208,
+            801,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{4}$ National University of Singapore",
+        "bbox": [
+            359,
+            225,
+            638,
+            243
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{bailong,jinxiaolong}@ict.ac.cn",
+        "bbox": [
+            342,
+            243,
+            655,
+            258
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            277,
+            339,
+            292
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Forecasting over Temporal Knowledge Graphs (TKGs) which predicts future facts based on historical ones has received much attention. Recent studies have introduced Large Language Models (LLMs) for this task to enhance the models' generalization abilities. However, these models perform forecasting via simultaneously learning two kinds of entangled knowledge in the TKG: (1) general patterns, i.e., invariant temporal structures shared across different scenarios; and (2) scenario information, i.e., factual knowledge engaged in specific scenario, such as entities and relations. As a result, the learning processes of these two kinds of knowledge may interfere with each other, which potentially impact the generalization abilities of the models. To enhance the generalization ability of LLMs on this task, in this paper, we propose a General-to-Specific learning framework (G2S) that disentangles the learning processes of the above two kinds of knowledge. In the general learning stage, we mask the scenario information in different TKGs and convert it into anonymous temporal structures. After training on these structures, the model is able to capture the general patterns across different TKGs. In the specific learning stage, we inject the scenario information into the structures via either in-context learning or fine-tuning modes. Experimental results show that G2S effectively improves the generalization abilities of LLMs.",
+        "bbox": [
+            141,
+            305,
+            460,
+            746
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            760,
+            258,
+            776
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Temporal Knowledge Graph (TKG), which consists of facts in the form of (subject, relation, object, timestamp), is often used to represent the changes of facts over time. Forecasting future facts based on historical ones in TKG is critical for many time-sensitive scenarios, such as finance (Feng et al., 2019), healthcare (Ernst et al., 2014), and poli",
+        "bbox": [
+            112,
+            785,
+            490,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/d3edb1068d4913ba08b67dbe70c2565c152c16fdfe4d5ca4dbdac174dba9bd75.jpg",
+        "image_caption": [
+            "Figure 1: An example of learning general patterns and applying to unseen datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            261,
+            877,
+            439
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tics (Morstatter et al., 2019). These scenarios involve different entities, relations, and time granularities, which drive us to seek a model that is able to adapt to new scenarios efficiently and effectively.",
+        "bbox": [
+            507,
+            495,
+            884,
+            558
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Previous methods view TKG as a temporally ordered sequence of knowledge graphs and adopt graph neural networks (Li et al., 2021b) or recurrent neural networks (Jin et al., 2020) to model it. These methods mainly focus on fitting a model to each specific TKG and overlook the generalization abilities. Recent methods attempt to model different TKGs via a unified Large Language Model (LLM), which shows considerable improvement in generalization ability. Some methods append historical facts into the prompt and utilize the incontext learning technique to predict the answer in unseen scenarios (Lee et al., 2023; Xia et al., 2024). Other methods learn temporal knowledge via supervised fine-tuning LLMs on a certain dataset and try to transfer this knowledge to unseen scenarios (Liao et al., 2024; Ding et al., 2024).",
+        "bbox": [
+            507,
+            563,
+            882,
+            837
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The generalization abilities of the TKG forecasting models mainly come from their abilities to learn general knowledge from seen scenarios and apply it to unseen scenarios. However, different TKGs describe different scenarios that involve different en",
+        "bbox": [
+            507,
+            841,
+            884,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding author",
+        "bbox": [
+            139,
+            906,
+            289,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "20927",
+        "bbox": [
+            473,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2025, pages 20927-20938",
+        "bbox": [
+            220,
+            945,
+            774,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 27 - August 1, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            268,
+            958,
+            727,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tities, relations, and time granularities. Taking Figure 1 as an example, GDELT is about conflicts and crisis scenarios in the geopolitics domain, while WIKI contains sports and academic scenarios in the general domain. This leads us to imagine: what is the general knowledge that can be shared across the scenarios in these TKGs? We observe that, if we anonymize the specific entity and relation names into abstract IDs, some temporal structures in different scenarios reveal a common substructure. Following this observation, we think that the TKGs contain two kinds of entangled knowledge, namely, general patterns and scenario information. The general patterns are common structures that can be generalized to different scenarios, while scenario information is factual knowledge, such as specific entities and relations, which only help TKG forecasting in a certain scenario. If the models directly learn the TKG (i.e., learn these two kinds of knowledge simultaneously), their learning processes may interfere with each other. From this point of view, previous methods, even LLM-based ones, may not learn the general patterns comprehensively, which limits their generalization abilities.",
+        "bbox": [
+            115,
+            84,
+            490,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Motivated by this, in this paper, we aim to disentangle the above two kinds of knowledge in TKG, and propose a general-to-specific learning framework for TKG forecasting, named G2S. G2S consists of two stages, namely, the general learning stage and the specific learning stage. In the general learning stage, we convert multiple TKGs into anonymous temporal structures via anonymizing the specific entities, relations, and timestamps into abstract IDs. After being trained on these structures, the model is able to learn the general patterns across multiple TKGs comprehensively. In the specific learning stage, we split the unseen TKG into anonymous temporal structures and their mappings between the specific entities/relations and the abstract IDs, as shown in the bottom right of Figure 1. In addition to general patterns, the model further learns the above information in either the in-context learning mode or the fine-tuning mode. To verify the effectiveness of G2S, we evaluate the proposed framework under three settings, namely, zero-shot, low-resource, and standard learning. Experimental results show that G2S effectively improves the generalization abilities of LLMs.",
+        "bbox": [
+            115,
+            472,
+            489,
+            856
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The contributions of this paper are three-fold:",
+        "bbox": [
+            132,
+            859,
+            472,
+            873
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We identify the two kinds of entangled knowledge in TKG, namely, the general patterns and",
+        "bbox": [
+            136,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "scenario information, and disentangle them via several anonymization strategies.",
+        "bbox": [
+            544,
+            84,
+            880,
+            118
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a general-to-specific learning framework for TKG forecasting, which prevents the learning processes of general patterns and scenario information from interfering with each other. In this way, the generalization ability of the model is enhanced.",
+            "- We conduct extensive experiments under three different settings. The experimental results demonstrate the effectiveness of the proposed framework."
+        ],
+        "bbox": [
+            531,
+            129,
+            884,
+            300
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            313,
+            665,
+            330
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "There are two different task settings for TKG reasoning, namely, interpolation and extrapolation settings. TKG reasoning under interpolation setting, also known as TKG completion, aims to predict the missing entities at past timestamps (Jiang et al., 2016), while TKG reasoning under extrapolation setting, also known as TKG forecasting, aims to predict the missing entities at future timestamps (Jin et al., 2020). This paper focuses on the latter one.",
+        "bbox": [
+            507,
+            341,
+            884,
+            500
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional TKG forecasting methods can be categorized into three different types: (1) Neural-based methods utilize deep neural networks to model the historical facts. RE-Net (Jin et al., 2020) develops a recurrent neural network to model query related historical facts. RE-GCN (Li et al., 2021b) utilizes graph convolutional networks to model recent KG snapshots. TANGO (Han et al., 2021b) adopts neural ordinary differential equations to model the structure information among entities. xERTE (Han et al., 2021a) designs a subgraph expansion and pruning mechanism to find explainable supports for forecasting. (2) Reinforcement learning-based methods utilize reinforcement learning to find a path between query entity and answer entity, such as CluSTeR (Li et al., 2021a) and TITer (Sun et al., 2021). (3) Rule-based methods mine logical rules to derive the answer from historical facts, such as TLogic (Liu et al., 2022). These methods mainly focus on enhancing the performance of the models on each specific TKG and overlook the generalization ability of the models.",
+        "bbox": [
+            507,
+            502,
+            885,
+            872
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, some studies have introduced LLM for TKG forecasting, which can be categorized into in-context learning and supervised fine-tuning",
+        "bbox": [
+            507,
+            873,
+            882,
+            922
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "20928",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "methods. Lee et al. (2023) propose an in-context learning method, which provides a basic input format to fit TKG data to LLMs. Xia et al. (2024) investigate the LLM's ability to find useful historical facts and reasoning answers based on them. GenTKG (Liao et al., 2024) proposes to select historical facts via rules and trains LLMs to answer the query via supervised fine-tuning. These methods show generalization abilities on TKG forecasting to some extent. However, as analyzed in Section 1, their learning processes of general patterns and scenario information may interfere with each other, which limits their generalization abilities.",
+        "bbox": [
+            112,
+            84,
+            492,
+            294
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            306,
+            337,
+            322
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A TKG  $\\mathcal{G}$  is formalized as a sequence of KG snapshots from timestamps 1 to  $T$ , i.e.,  $\\mathcal{G} = \\{\\mathcal{G}_1, \\mathcal{G}_2, \\dots, \\mathcal{G}_T\\}$ . Each TKG snapshot  $\\mathcal{G}_t = \\{\\mathcal{E}, \\mathcal{R}, \\mathcal{F}_t\\}$  consists of the entity set  $\\mathcal{E}$ , relation set  $\\mathcal{R}$ , and the fact set  $\\mathcal{F}_t$  at timestamp  $t \\in \\{1, \\dots, T\\}$ . Each fact  $f_i \\in \\mathcal{F}_t$  is denoted as a quadruple  $f_i = (s_i, r_i, o_i, t)$ , where  $s_i, o_i \\in \\mathcal{E}$  are subject and object entities;  $r_i \\in \\mathcal{R}$  is the relation. Currently, the TKG forecasting task focuses on predicting the missing object entity in a query  $q = (s_q, r_q, ?, t_q)$ . The model  $\\mathcal{M}$  is required to score possible object entities based on the query and historical TKG snapshots:  $\\text{score}(o) = \\mathcal{M}(q, \\mathcal{G}_{< t_q})$ . Note that, when predicting the missing subject entity, the query is formulated as  $q = (?, r_q, o_q, t_q)$ .",
+        "bbox": [
+            112,
+            332,
+            489,
+            576
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            588,
+            263,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To enhance the generalization ability of LLMs on TKG forecasting, we propose G2S that disentangles the learning process of the general patterns and specific information into two stages, namely, the general learning and specific learning stages. Figure 2 shows the overall diagram of G2S. We will describe the design of these two learning stages and provide other training and inference details.",
+        "bbox": [
+            112,
+            615,
+            489,
+            744
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 General Learning Stage",
+        "text_level": 1,
+        "bbox": [
+            112,
+            756,
+            349,
+            772
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the general learning stage, there are three main processing steps, namely, query construction, anonymous temporal structure conversion, and sample construction. In what follows, we will describe these steps in detail.",
+        "bbox": [
+            112,
+            778,
+            489,
+            858
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1.1 Query Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            869,
+            332,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this step, we construct training queries from the TKGs and find historical facts related to",
+        "bbox": [
+            112,
+            889,
+            487,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "these queries. For each fact  $f_{q} = (s_{q},r_{q},o_{q},t_{q})$  in a TKG  $\\mathcal{G}$ , we construct two queries  $q_{1} = (s_{q},r_{q},?,t_{q})$  and  $q_{2} = (?,r_{q},o_{q},t_{q})$  to predict the object and subject entities, respectively. Then, we select one-hop historical facts  $\\mathcal{H}_{q}$  for each query  $q$  which is a commonly used history selection strategy in existing approaches (Jin et al., 2020; Lee et al., 2023):",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {H} _ {q _ {1}} = \\bigcup_ {t <   t _ {q}} \\{(s _ {q}, r ^ {\\prime}, o ^ {\\prime}, t) \\in \\mathcal {F} _ {t} \\} \\\\ \\mathcal {H} _ {q _ {2}} = \\bigcup_ {t <   t _ {q}} \\left\\{\\left(s ^ {\\prime}, r ^ {\\prime}, o _ {q}, t\\right) \\in \\mathcal {F} _ {t} \\right\\}, \\tag {1} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            219,
+            880,
+            290
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $s', o' \\in \\mathcal{E}$  are arbitrary entities,  $r' \\in \\mathcal{R}$  is arbitrary relation. The historical facts are sorted according to their timestamps in ascending order. Following the convention, we retain the most recent  $L$  historical facts. To better learn general patterns, we collect queries across multiple TKGs with different entities, relations, and time granularities. For all the TKGs, we adopt the same processing method.",
+        "bbox": [
+            507,
+            297,
+            884,
+            428
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1.2 Anonymous Temporal Structure Conversion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            435,
+            823,
+            464
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To eliminate the impact of scenario information in the general learning stage, we apply an anonymization strategy that converts entities, relations, and timestamps in queries and facts into abstract IDs.",
+        "bbox": [
+            507,
+            469,
+            882,
+            533
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With respect to the timestamps, previous studies (Lee et al., 2023; Liao et al., 2024) utilize the number of timestamps from the start time of each TKG as the ID. For example, in dataset ICEWS14, the start time is 2014-01-01, and each timestamp is 1 day, so the ID of the timestamp 2014-01-10 is 9. However, such IDs may create bias between training data and testing data, since testing timestamps are not in the same period as training timestamps. Therefore, we set the timestamp ID of a query as 0, and the timestamp ID of each historical fact to the number of timestamps from the query time:",
+        "bbox": [
+            507,
+            535,
+            882,
+            727
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {A} (t) = t _ {q} - t. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            638,
+            734,
+            880,
+            753
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With respect to entities and relations, we investigate three different strategies:",
+        "bbox": [
+            507,
+            760,
+            884,
+            791
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Frequency ID (FID) uses the rank of entity frequency (descending order) for each query and corresponding historical facts as ID. This strategy was proposed in previous studies (Lee et al., 2023; Liao et al., 2024).",
+        "bbox": [
+            507,
+            793,
+            882,
+            871
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Global ID (GID) uses the original ID provided in each dataset. In this strategy, the ID of each entity and each relation is unique for each TKG.",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "20929",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/f5c65d98238f155540e2df0a540a3a59a5cae55ef5701275dba61c306f596f33.jpg",
+        "image_caption": [
+            "Figure 2: The proposed G2S learning framework. Here, \"ATS\", \"SFT\" and \"ICL\" denote the anonymous temporal structures, supervised fine-tuning, and in-context learning, respectively, and each color in KG denotes a specific entity or relation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            152,
+            84,
+            843,
+            318
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Random ID (RID) uses the randomly assigned IDs for entities and relations within a query and corresponding historical facts.",
+        "bbox": [
+            112,
+            381,
+            487,
+            429
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Detailed analyses of the advantages and disadvantages of each strategy are described in Section 6.5",
+        "bbox": [
+            112,
+            431,
+            489,
+            476
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.3 Sample Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            492,
+            341,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Based on the anonymous temporal structure, we construct training samples to supervised fine-tune the LLM. Each training sample consists of an input-output pair, where an input consists of two parts, namely, history and query. In the history part, each line contains a historical fact in the form of “ $\\mathcal{A}(t) : [\\mathcal{A}(s), \\mathcal{A}(r), \\mathcal{A}(o)]$ ”, where  $\\mathcal{I}$  denotes the anonymization strategy described above. The historical facts are listed by their temporal order. In the query part, the query  $(s_q, r_q, ?, t_q)$  is presented in the form of “ $0 : [\\mathcal{A}(s_q), \\mathcal{A}(r_q), ?]$ \". In addition, the input contains an empty answer part at the end to prompt the model to generate the answer. The output of the training sample is the correct answer entity ID “ $\\mathcal{A}(o_q)$ \". For FID and RID strategies, if the answer entity does not occur in historical facts, it will not be assigned with an ID, so we set the output to be “None”. After being fine-tuned on such training samples, the model captures the general patterns across multiple TKGs.",
+        "bbox": [
+            112,
+            514,
+            489,
+            835
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Specific Learning Stage",
+        "text_level": 1,
+        "bbox": [
+            112,
+            850,
+            346,
+            866
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the specific learning stage, the processing steps are similar to those presented in Section 4.1. However, there are two key differences:",
+        "bbox": [
+            112,
+            873,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Scenario information mapping. In the specific learning stage, the model is required to combine the scenario information with the general patterns to derive more accurate answers. Therefore, when applying an anonymization strategy to convert entities/relations into IDs, we retain the mappings between entities/relations and IDs. Then, we add two parts at the beginning of the input, namely, entity and relation parts, according to the mappings. In the entity part, each line contains the mapping between an entity and its abstract ID in the form “ $\\mathcal{A}(e):e$ ”, as shown in Figure 2. The relation part is done similarly, which contains the mappings between relations and their IDs.",
+        "bbox": [
+            505,
+            381,
+            884,
+            605
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Two learning modes. The specific learning stage can run in two different modes:",
+        "bbox": [
+            507,
+            607,
+            882,
+            639
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In-context learning (ICL) mode. In this mode, the model is prohibited from utilizing the training samples. Thus, the model can only learn scenario information via in-context learning.",
+        "bbox": [
+            507,
+            640,
+            882,
+            705
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Supervised Fine-tuning (SFT) mode. In this mode, the model is allowed to utilize the training samples. Thus, the model can learn the entities and relations via updating the parameters.",
+        "bbox": [
+            507,
+            706,
+            880,
+            771
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 Training and Inference Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            785,
+            796,
+            800
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The training objective is to minimize the cross-entropy loss between generated token sequence  $O$  and ground-truth token sequence  $\\hat{O}$ , i.e.,  $\\mathcal{L} = CE(O, \\hat{O})$ .",
+        "bbox": [
+            507,
+            808,
+            884,
+            872
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use LLaMA3-8B as the backbone model and adopt Low-Rank Adaptation (LoRA) (Hu et al., 2022) method to effectively fine-tune it.",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "20930",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the inference phase, considering the efficiency, we only require the model to generate one step. Then, we adopt the generation score of the tokens provided by the LLM as the ranking score:",
+        "bbox": [
+            112,
+            84,
+            489,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s c o r e} (o) = \\Pr (o | I n p u t), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            203,
+            159,
+            487,
+            175
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $o$  is a generated token. Finally, we retain the top 10 valid predictions after filtering the duplicate predictions.",
+        "bbox": [
+            112,
+            187,
+            487,
+            235
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            246,
+            319,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            272,
+            228,
+            286
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use 5 widely-used datasets in this paper: ICEWS14 (Li et al., 2021b), ICEWS18 (Jin et al., 2020), YAGO (Mahdisoltani et al., 2013), GDELT (Jin et al., 2020), and WIKI (Leblay and Chekol, 2018). The basic statistics of the datasets are shown in Appendix B. Especially, ICEWS14, ICEWS18, and GDELT follow CAMEO<sup>1</sup> schema, which means relation types in the three datasets are almost the same. YAGO and WIKI are constructed from open knowledge graphs of the same name, and thus, they have their own schema definitions.",
+        "bbox": [
+            112,
+            292,
+            489,
+            468
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Among these datasets, the training parts of GDELT and WIKI are used in the general learning stage, while ICEWS14, ICEWS18, and YAGO are used in the specific learning stage. The proposed model is evaluated on the later three datasets. We sample 100,000 training samples from GDELT and 30,000 training samples from WIKI, approximately according to the ratio of their total training facts.",
+        "bbox": [
+            112,
+            470,
+            487,
+            598
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.2 Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            609,
+            221,
+            621
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We employ widely used H@1/3/10 to evaluate the proposed model. These metrics represent the proportion of queries that the model ranks the correct answer entity at top 1/3/10. Moreover, we adopt the time-aware filtering setting, where the valid entities except the correct answer entity are removed from the predictions. Mean Reciprocal Rank (MRR) is another metric that has often been used in previous studies. However, this metric is not applicable for LLMs, since LLMs can not produce the full rank of all entities for a query.",
+        "bbox": [
+            112,
+            629,
+            489,
+            806
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            816,
+            233,
+            831
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We compare the proposed method with several baseline methods. The baseline methods can be divided into two different categories, i.e., traditional methods and LLM-based methods. With",
+        "bbox": [
+            112,
+            837,
+            487,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "respect to traditional methods, we choose REGCN (Li et al., 2021b), xERTE (Han et al., 2021a), TANGO (Han et al., 2021b),TITer (Sun et al., 2021), and TLogic (Liu et al., 2022). With respect to LLM-based methods, we choose incontext learning (Lee et al., 2023) method and GenTKG (Liao et al., 2024). For in-context learning method, we use GPT-NeoX-20B (Black et al., 2022), Llama2-7B (Touvron et al., 2023), and Llama3-8B (AI@Meta, 2024) as backbone models, noted as GPT-NeoX-ICL, Llama2-ICL, and Llama3-ICL, respectively.",
+        "bbox": [
+            507,
+            84,
+            884,
+            277
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.4 Experiment Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            288,
+            717,
+            305
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our model under three different settings: (1) Standard setting. Under this setting, all training samples can be used to fine-tune the model in the specific learning stage. (2) Zero-shot setting. Under this setting, no training samples can be used to fine-tune the model in the specific learning stage. (3) Low-resource setting. Under this setting, the earliest  $5\\% / 20\\% / 50\\%$  training samples can be used to fine-tune the model in the specific learning stage.",
+        "bbox": [
+            507,
+            310,
+            884,
+            455
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.5 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            466,
+            741,
+            482
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For experiments on all datasets, the number of historical facts  $L$  is set to 50; context length of prompts is set to 1024; Learning rate is set to 1e-4; batch size is set to 8; number of epochs is set to 1. We tune the model on the validation set and report the best results. Unless otherwise specified, G2S utilizes both GDELT and WIKI datasets and adopts RID strategy in general learning stage. In specific learning stage, G2S adopts GID strategy on ICEWS14 and ICEWS18, and adopts RID strategy on YAGO<sup>2</sup>. We develop our experiment code based on LlamaFactory (Zheng et al., 2024) and run the experiments on 4 NVIDIA A800 GPUs. The general learning stage takes about 2 hours. Under the standard setting (fine-tuning on all training samples), specific learning on ICEWS14 takes about 2 hours; ICEWS18 takes about 20 hours; YAGO takes about 7 hours.",
+        "bbox": [
+            507,
+            487,
+            882,
+            776
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Results and Analyses",
+        "text_level": 1,
+        "bbox": [
+            507,
+            790,
+            726,
+            806
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6.1 Standard Setting",
+        "text_level": 1,
+        "bbox": [
+            507,
+            816,
+            690,
+            832
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the performance of G2S when sufficient training data are available, we conduct experiments under standard settings, where all training",
+        "bbox": [
+            507,
+            837,
+            882,
+            885
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>https://eventdata.parusanalytics.com/data.dir/cameo.html",
+        "bbox": [
+            134,
+            906,
+            485,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ GID strategy is not suitable for YAGO, see Appendix A for detailed analyses.",
+        "bbox": [
+            507,
+            894,
+            882,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "20931",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/150be768798ed5e298de575c2f677f453345a375a67ee97e0ddbb23aea2c9af9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model\\Dataset</td><td colspan=\"3\">ICEWS14</td><td colspan=\"3\">ICEWS18</td><td colspan=\"3\">YAGO</td></tr><tr><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td></tr><tr><td>RE-GCN</td><td>31.3</td><td>47.3</td><td>62.6</td><td>22.3</td><td>36.7</td><td>52.5</td><td>78.7</td><td>84.2</td><td>88.4</td></tr><tr><td>xERTE</td><td>33.0</td><td>45.4</td><td>57.0</td><td>20.9</td><td>33.5</td><td>46.2</td><td>84.2</td><td>90.2</td><td>91.2</td></tr><tr><td>TANGO</td><td>27.2</td><td>40.8</td><td>55.0</td><td>19.1</td><td>31.8</td><td>46.2</td><td>59.0</td><td>64.6</td><td>67.7</td></tr><tr><td>TIter</td><td>31.9</td><td>45.4</td><td>57.5</td><td>21.2</td><td>32.5</td><td>43.9</td><td>84.5</td><td>90.8</td><td>91.2</td></tr><tr><td>TLogic</td><td>33.2</td><td>47.6</td><td>60.2</td><td>20.4</td><td>33.6</td><td>48.0</td><td>74.0</td><td>78.9</td><td>79.1</td></tr><tr><td>llama2-ICL</td><td>25.8</td><td>43.0</td><td>51.0</td><td>13.5</td><td>27.6</td><td>32.6</td><td>67.7</td><td>79.0</td><td>81.8</td></tr><tr><td>llama3-ICL</td><td>31.9</td><td>42.4</td><td>44.1</td><td>18.1</td><td>27.2</td><td>28.8</td><td>56.9</td><td>57.9</td><td>57.9</td></tr><tr><td>GPT-NeoX-ICL</td><td>32.4</td><td>46.0</td><td>56.5</td><td>19.2</td><td>31.3</td><td>41.4</td><td>78.7</td><td>89.2</td><td>92.6</td></tr><tr><td>GenTKG</td><td>36.85</td><td>47.95</td><td>53.50</td><td>24.25</td><td>37.25</td><td>42.10</td><td>79.15*</td><td>83.00*</td><td>84.25*</td></tr><tr><td>G2S</td><td>38.33</td><td>54.12</td><td>68.58</td><td>23.04</td><td>35.32</td><td>46.62</td><td>87.88</td><td>90.89</td><td>90.93</td></tr></table>",
+        "bbox": [
+            122,
+            80,
+            873,
+            312
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Results of TKG reasoning on datasets ICEWS14, ICEWS18, and YAGO. The best results among all models are in boldface. The results of GenTKG on YAGO are marked with “*” since they use self-developed subset of the standard YAGO dataset. Thus, they are not directly comparable with other methods.",
+        "bbox": [
+            112,
+            321,
+            882,
+            367
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "data can be used to fine-tune the model in the specific learning stage. The experimental results are shown in Table 1. The methods are split into three groups, the first group contains the traditional methods, the second contains the LLM-based methods with in-context learning, and the third contains the LLM-based methods with supervised fine-tuning. From the table, we can observe that:",
+        "bbox": [
+            112,
+            380,
+            487,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) G2S outperforms the baselines on ICEWS14 and YAGO under the H@1 metric, and achieves the second best performance on ICEWS18, which shows the forecasting ability of G2S.",
+            "(2) G2S and GenTKG outperform ICL methods in general, implying that supervised fine-tuning methods may better learn datasets than ICL methods if sufficient training data are available.",
+            "(3) G2S slightly underperforms GenTKG on ICEWS18, possibly due to the multi-token ID generation issue in G2S. A detailed analysis of this result can be found in Appendix A."
+        ],
+        "bbox": [
+            112,
+            510,
+            485,
+            705
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.2 Zero-shot Setting",
+        "text_level": 1,
+        "bbox": [
+            112,
+            720,
+            295,
+            736
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the generalization ability of G2S, we conduct TKG forecasting under the zero-shot setting. Since the traditional methods can hardly run under the zero-shot settings, we only compare G2S with LLM-based methods. In addition, we adopt a simple frequency baseline proposed by Lee et al. (2023), which predicts answers according to entity frequency in historical facts.",
+        "bbox": [
+            112,
+            743,
+            487,
+            871
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare several variants of G2S to show the impact of different anonymization strategies in general learning stage: (1)  $\\mathrm{G2S}_{GL(F)}$  utilizes GDELT",
+        "bbox": [
+            112,
+            873,
+            487,
+            922
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "and adopts FID strategy; (2)  $\\mathrm{G2S}_{GL(F + Map)}$  add mappings between entity/relation and IDs in addition to  $\\mathrm{G2S}_{GL(F)}$ ; (3)  $\\mathrm{G2S}_{GL(R)}$  utilizes GDELT and adopts RID strategy; (4)  $\\mathrm{G2S}_{GL(R)} w. WIKI$  jointly utilizes GDELT and YAGO, and adopts RID strategy. In specific learning stage, the anonymization strategies are selected via the performance on validation set.",
+        "bbox": [
+            505,
+            380,
+            884,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We exclude the GenTKG from the baselines since, strictly speaking, its generalization setting is not a zero-shot setting. The differences between the two settings lie in the following aspects: (1) The training data (ICEWS14) and testing data (ICEWS18) of GenTKG (generalization) originate from the same source but correspond to different years; and (2) GenTKG adopts logic-rules mined from ICEWS18.",
+        "bbox": [
+            507,
+            512,
+            882,
+            656
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The experimental results are shown in Table 2. From the table, we can observe that:",
+        "bbox": [
+            507,
+            659,
+            882,
+            690
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "(1) Llama3-ICL (8B) underperforms GPT-NeoX-ICL (20B) even though llama3 is a more recent model. These results imply that the scale of the model has a relatively large impact on the generalization ability in this task. Therefore, it is somehow unfair to directly compare G2S (8B) to GPT-NeoX-ICL. However,  $\\mathrm{G2S}_{GL(R)}w.$  WIKI still obtains comparable performance with GPT-NeoX-ICL on ICEWS datasets (especially a higher Hit@10 on ICEWS18) and relatively higher performance on YAGO dataset. Compared to the baselines except GPT-NeoX-ICL,  $\\mathrm{G2S}_{GL(R)}w.$  WIKI obtains significantly higher performance on all three datasets, which demonstrates the strong gen",
+        "bbox": [
+            507,
+            696,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "20932",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/2e40154513039946d249d96c5319421a8e5b2b97b14ba220e27f2652ab0f6e83.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset\nModel</td><td colspan=\"3\">ICEWS14</td><td colspan=\"3\">ICEWS18</td><td colspan=\"3\">YAGO</td></tr><tr><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td></tr><tr><td>Frequency</td><td>24.3</td><td>38.7</td><td>53.2</td><td>14.1</td><td>26.5</td><td>40.9</td><td>76.6</td><td>85.9</td><td>92.1</td></tr><tr><td>llama2-ICL</td><td>25.8</td><td>43.0</td><td>51.0</td><td>13.5</td><td>27.6</td><td>32.6</td><td>67.7</td><td>79.0</td><td>81.8</td></tr><tr><td>llama3-ICL</td><td>31.9</td><td>42.4</td><td>44.1</td><td>18.1</td><td>27.2</td><td>28.8</td><td>56.9</td><td>57.9</td><td>57.9</td></tr><tr><td>GPT-NeoX-ICL</td><td>32.4</td><td>46.0</td><td>56.5</td><td>19.2</td><td>31.3</td><td>41.4</td><td>78.4</td><td>89.1</td><td>92.7</td></tr><tr><td>G2SGL(F)</td><td>31.52</td><td>44.80</td><td>55.62</td><td>19.47</td><td>31.17</td><td>43.05</td><td>81.50</td><td>87.54</td><td>90.79</td></tr><tr><td>G2SGL(F+Map)</td><td>32.13</td><td>45.56</td><td>56.59</td><td>19.43</td><td>31.87</td><td>44.52</td><td>81.00</td><td>87.05</td><td>90.76</td></tr><tr><td>G2SGL(R)</td><td>31.95</td><td>44.87</td><td>55.61</td><td>19.54</td><td>31.34</td><td>43.29</td><td>80.40</td><td>87.32</td><td>90.86</td></tr><tr><td>G2SGL(R) w.WIKI</td><td>32.02</td><td>44.99</td><td>55.84</td><td>19.71</td><td>31.58</td><td>43.45</td><td>86.07</td><td>90.82</td><td>91.02</td></tr></table>",
+        "bbox": [
+            127,
+            80,
+            870,
+            275
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Results of zero-shot TKG reasoning on datasets ICEWS14, ICEWS18, and YAGO. The best results among all models are in boldface and the second best results are underlined.",
+        "bbox": [
+            112,
+            282,
+            882,
+            312
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "eralization ability of this model.",
+        "bbox": [
+            112,
+            338,
+            354,
+            353
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(2)  $\\mathrm{G2S}_{GL(F)}$  underperforms  $\\mathrm{G2S}_{GL(F + Map)}$  on ICEWS14 and ICEWS18, while outperforms it on YAGO. We guess it is because GDELT shares the same schema with ICEWS14 and ICEWS18 (i.e., relation types are almost similar in these datasets),  $\\mathrm{G2S}_{GL(F + Map)}$  may learn useful scenario information from GDELT and apply it to ICEWS datasets. However, this scenario information may not be so helpful to YAGO. These results verify our motivation that simultaneously learning general patterns and scenario information may affect the generalization ability of the model.",
+            "(3) The performance of  $\\mathrm{G2S}_{GL(F)}$  and  $\\mathrm{G2S}_{GL(R)}$  are similar, which inspires us to use the RID strategy in the general learning stage of the overall framework, since the RID strategy is simpler in practice.",
+            "(4)  $\\mathrm{G2S}_{GL(R)}$  w. WIKI outperforms  $\\mathrm{G2S}_{GL(R)}$  on all three datasets, especially on YAGO. We guess it is because WIKI may contain more useful general patterns for YAGO than GDELT. However, the improvements on ICEWS14 and ICEWS18 verify our motivation that, although TKGs may differ in entities, relations, and time granularities, they still share some general patterns."
+        ],
+        "bbox": [
+            110,
+            355,
+            489,
+            758
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.3 Low-resource Setting",
+        "text_level": 1,
+        "bbox": [
+            112,
+            771,
+            329,
+            785
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the generalization ability of G2S under the low-resource setting, where the earliest  $5\\% / 20\\% / 50\\%$  samples in ICEWS14 are used to train the model, and all test samples in ICEWS14 are used to evaluate the model. In this experiment, we use  $\\mathrm{G2S}_{GL(R)} w.$  WIKI in Section 6.2 as the initial model for the specific learning stage. We compare three variants of G2S with different",
+        "bbox": [
+            112,
+            791,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "anonymization strategies in specific learning stage:  $\\mathrm{G2S}_{SL(F)}$ ,  $\\mathrm{G2S}_{SL(R)}$ , and  $\\mathrm{G2S}_{SL(G)}$  represent the variants of G2S that adopt FID, RID, and GID strategy in specific learning stage, respectively. Other baselines are the same as those in Liao et al. (2024).",
+        "bbox": [
+            507,
+            338,
+            884,
+            418
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The experimental results are shown in Table 3. From the table, we can observe that:",
+        "bbox": [
+            507,
+            419,
+            882,
+            450
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) All three variants of G2S outperform the baseline methods via training on  $5\\%$  training samples. Compared to GenTKG, which simultaneously learns general patterns and scenario information, the proposed two-stage learning framework obtains better generalization ability.",
+            "(2) When using  $5\\%$  training samples,  $\\mathrm{G2S}_{SL(G)}$  underperforms  $\\mathrm{G2S}_{SL(F)}$  and  $\\mathrm{G2S}_{SL(R)}$ . As the number of training samples increases,  $\\mathrm{G2S}_{SL(G)}$  gradually outperforms the other two. This observation implies that, when there are very few training samples, the model is unable to learn the static mapping between entities/relations with global IDs. However, when more training samples are provided, GID performs more effectively than the other two strategies.",
+            "(3) LLM-based methods, such as GenTKG and variants of G2S, outperform traditional methods with a large margin, implying that the LLMs generally obtain better generalization abilities than traditional models. Thus, they require fewer training samples to reach the same performance as traditional models."
+        ],
+        "bbox": [
+            507,
+            451,
+            884,
+            821
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.4 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            835,
+            673,
+            850
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To verify the effectiveness of each design in G2S, we conduct an ablation study on ICEWS14 under all three settings. For standard setting, we evaluate three variants of G2S:  $w / o$ . Ent means the",
+        "bbox": [
+            507,
+            856,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "20933",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cdd0e294ec8c4a4be4f1d65c2d192853d432c62229ca403772b97e1313e1a6ef.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">5%</td><td colspan=\"3\">20%</td><td colspan=\"3\">50%</td></tr><tr><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td></tr><tr><td>RE-GCN</td><td>13.79</td><td>22.09</td><td>30.27</td><td>19.63</td><td>29.67</td><td>39.83</td><td>24.05</td><td>36.72</td><td>48.84</td></tr><tr><td>xERTE</td><td>6.95</td><td>14.17</td><td>25.46</td><td>17.80</td><td>29.26</td><td>42.08</td><td>22.51</td><td>34.15</td><td>46.59</td></tr><tr><td>TANGO</td><td>11.29</td><td>17.18</td><td>22.97</td><td>11.25</td><td>17.38</td><td>23.38</td><td>14.37</td><td>17.51</td><td>22.77</td></tr><tr><td>TITer</td><td>21.06</td><td>34.78</td><td>49.10</td><td>26.69</td><td>39.42</td><td>51.78</td><td>30.05</td><td>42.82</td><td>54.74</td></tr><tr><td>TLogic</td><td>26.03</td><td>37.42</td><td>46.50</td><td>28.72</td><td>40.48</td><td>50.71</td><td>29.84</td><td>42.40</td><td>53.37</td></tr><tr><td>GenTKG</td><td>30.60</td><td>42.20</td><td>49.30</td><td>34.90</td><td>46.60</td><td>54.00</td><td>36.00</td><td>48.70</td><td>55.50</td></tr><tr><td>G2SSL(F)</td><td>33.65</td><td>45.88</td><td>55.81</td><td>34.12</td><td>46.63</td><td>56.63</td><td>35.10</td><td>48.08</td><td>58.07</td></tr><tr><td>G2SSL(R)</td><td>33.22</td><td>45.75</td><td>56.08</td><td>33.70</td><td>46.29</td><td>56.24</td><td>34.12</td><td>46.50</td><td>56.44</td></tr><tr><td>G2SSL(G)</td><td>31.84</td><td>44.95</td><td>56.21</td><td>33.84</td><td>47.72</td><td>61.39</td><td>36.01</td><td>50.93</td><td>65.29</td></tr></table>",
+        "bbox": [
+            161,
+            80,
+            836,
+            290
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/4de752c45417387527c19d83e78e71df7a1408a3891371683e9370b5e4619b02.jpg",
+        "table_caption": [
+            "Table 3: Results of low-resource TKG reasoning on ICEWS14. The best results among all models are in boldface."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>H@1</td><td>H@3</td></tr><tr><td colspan=\"3\">Standard Setting</td></tr><tr><td>G2S</td><td>38.33</td><td>54.12</td></tr><tr><td>w/o. Ent</td><td>35.672.66↓</td><td>50.243.88↓</td></tr><tr><td>w/o. Rel</td><td>38.000.33↓</td><td>53.300.82↓</td></tr><tr><td>w/o. Ent&amp;Rel</td><td>35.692.64↓</td><td>50.203.92↓</td></tr><tr><td colspan=\"3\">Zero-shot Setting</td></tr><tr><td>G2SGL</td><td>32.02</td><td>44.99</td></tr><tr><td>w/o. GL</td><td>13.6718.35↓</td><td>28.8916.10↓</td></tr><tr><td colspan=\"3\">Low-resource Setting</td></tr><tr><td>G2SLSL</td><td>33.65</td><td>45.88</td></tr><tr><td>w/o. GL</td><td>32.131.52↓</td><td>44.970.91↓</td></tr></table>",
+        "bbox": [
+            121,
+            336,
+            480,
+            563
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Ablation study on ICEWS14 under all three settings.",
+        "bbox": [
+            112,
+            571,
+            485,
+            602
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "mappings between entities and IDs are removed in the specific learning stage;  $w / o$ . Rel means the mappings between relations and IDs are removed in the specific learning stage; and  $w / o$ . Ent&Rel means both kinds of mappings are removed in the specific learning stage. For zero-shot and low-resource settings, we evaluate the variants that remove the general learning stage and use llama3 as the initial model in the specific learning stage, which is denoted as  $w / o$ . GL. Here, G2SGL is G2SGL(R) w. WIKI in Section 6.2; and G2SSL is G2SLS(F) in Section 6.3 that use 5% training samples.",
+        "bbox": [
+            112,
+            629,
+            489,
+            837
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The experimental results are shown in Table 4, where we can observe that:",
+        "bbox": [
+            112,
+            839,
+            487,
+            870
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(1) G2S outperforms  $w / o$ . Ent,  $w / o$ . Rel, and  $w / o$ . Ent&Rel, which shows the effectiveness of scenario information in TKG forecasting. In",
+        "bbox": [
+            112,
+            873,
+            487,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "addition, the scenario information about entities are usually more important than that about relations.",
+        "bbox": [
+            507,
+            340,
+            880,
+            370
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(2)  $\\mathrm{G2S}_{GL}$  and  $\\mathrm{G2S}_{SL}$  outperform corresponding  $w / o$ .  $GL$  models, respectively, which shows the effectiveness of the general learning stage in enhancing the generalization ability.",
+        "bbox": [
+            507,
+            370,
+            882,
+            436
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.5 Analyses on Anonymization Strategies",
+        "text_level": 1,
+        "bbox": [
+            507,
+            445,
+            855,
+            462
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "From the above experimental results, we can conclude that:",
+        "bbox": [
+            507,
+            467,
+            882,
+            495
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 2, in the general learning stage, the performance of FID  $(\\mathrm{G2S}_{GL(F)})$  and RID  $(\\mathrm{G2S}_{GL(R)})$  strategies are comparable under the zero-shot setting. Both strategies enhance the generalization ability of the model, compared to the model that is additionally trained on the scenario information  $(\\mathrm{G2S}_{GL(F + Map)})$ .",
+        "bbox": [
+            507,
+            499,
+            880,
+            612
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In the specific learning stage, the three strategies show different performances under each setting. As shown in Table 1, GID shows the best performance under the standard setting, where sufficient samples are available to learn the static mapping between an entity/relation and a unique ID. However, the different results on  $5\\%$  and  $50\\%$  training data shown in Table 3 imply that, GID requires more training data than the other two strategies. As shown in Tables 2 and 3, FID performs best under the low-resource and zero-shot settings. We guess the reason is that FID contains the frequency features in its IDs, so it outperforms RID.",
+        "bbox": [
+            507,
+            613,
+            882,
+            820
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            832,
+            648,
+            847
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This paper investigates the generalization ability of LLMs for temporal knowledge graph forecasting. Firstly, We identified two kinds of entangled knowledge in TKGs, namely, general patterns and",
+        "bbox": [
+            507,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "20934",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "scenario information. Specifically, we adopted several anonymization strategies to convert the entities and relations in TKGs into abstract IDs. Via these structures, the LLMs are able to learn the general patterns more comprehensively. Further, we proposed a general-to-specific learning framework (G2S), which disentangles the learning of these two kinds of knowledge into two learning stages: the general and specific learning stages. To verify the generalization ability of the learned G2S model, we conducted experiments under standard, zero-shot, and low-resource settings. The results demonstrate that the model is able to outperform the baseline methods under all three settings, which shows the effectiveness of the proposed method.",
+        "bbox": [
+            112,
+            84,
+            492,
+            326
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            342,
+            220,
+            357
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We find the following limitations:",
+        "bbox": [
+            112,
+            372,
+            366,
+            387
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) This paper only studies the generalization ability of G2S on 5 TKGs about geopolitics and general domains. The generalization ability of the model in other specific domains requires further exploration.",
+            "(2) Previous studies have proposed different strategies to select relevant historical facts for queries, including rule-based and reinforcement learning-based strategies. In this paper, we only investigate the generalization ability under one-hop strategy for simplicity. Investigating the generalization ability under other history selection strategies and multi-hop facts will be an important direction for future work.",
+            "(3) The GID strategy in G2S usually assigns larger IDs for entities, which sometimes requires the model to generate a complete ID through multi-step decoding. Currently, LLM-based TKG forecasting methods, especially those adopting SFT, cannot deal with this problem well."
+        ],
+        "bbox": [
+            112,
+            388,
+            489,
+            715
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            731,
+            285,
+            747
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is partially funded by the National Natural Science Foundation of China under grant 62306299, 62441229 and 62406308, the LenovocAS Joint Lab Youth Scientist Project, the project under Grants No. JCKY2022130C039, and the Strategic Priority Research Program of the CAS under Grants No. XDB0680102. Thanks to the reviewers for the constructive discussions and suggestions.",
+        "bbox": [
+            112,
+            760,
+            489,
+            904
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            83,
+            608,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "AI@Meta. 2024. Llama 3 model card.",
+        "bbox": [
+            510,
+            107,
+            771,
+            121
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Sidney Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, Usvsn Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, and Samuel Weinbach. 2022. GPT-NeoX-20B: An open-source autoregressive language model. In Proceedings of BigScience Episode #5 – Workshop on Challenges & Perspectives in Creating Large Language Models, pages 95–136, virtual+Dublin. Association for Computational Linguistics.",
+            "Bin Chen, Chunjing Xiao, and Fan Zhou. 2024. Natural evolution-based dual-level aggregation for temporal knowledge graph reasoning. In *Findings of the Association for Computational Linguistics: EMNLP* 2024, pages 9274–9284, Miami, Florida, USA. Association for Computational Linguistics.",
+            "Zifeng Ding, Heling Cai, Jingpei Wu, Yunpu Ma, Ruotong Liao, Bo Xiong, and Volker Tresp. 2024. zrLLM: Zero-shot relational learning on temporal knowledge graphs with large language models. 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 1877-1895, Mexico City, Mexico. Association for Computational Linguistics.",
+            "Patrick Ernst, Cynthia Meng, Amy Siu, and Gerhard Weikum. 2014. Knowlife: A knowledge graph for health and life sciences. In 2014 IEEE 30th International Conference on Data Engineering, pages 1254-1257.",
+            "Fuli Feng, Xiangnan He, Xiang Wang, Cheng Luo, Yiqun Liu, and Tat-Seng Chua. 2019. Temporal relational ranking for stock prediction. ACM Trans. Inf. Syst., 37(2).",
+            "Zhen Han, Peng Chen, Yunpu Ma, and Volker Tresp. 2021a. Explainable subgraph reasoning for forecasting on temporal knowledge graphs. In International Conference on Learning Representations.",
+            "Zhen Han, Zifeng Ding, Yunpu Ma, Yujia Gu, and Volker Tresp. 2021b. Learning neural ordinary equations for forecasting future links on temporal knowledge graphs. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8352-8364, Online and Punta Cana, Dominican Republic. Association for 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 International Conference on Learning Representations."
+        ],
+        "bbox": [
+            509,
+            134,
+            885,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "20935",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Tingsong Jiang, Tianyu Liu, Tao Ge, Lei Sha, Sujian Li, Baobao Chang, and Zhifang Sui. 2016. Encoding temporal information for time-aware link prediction. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2350-2354, Austin, Texas. Association for Computational Linguistics.",
+            "Woojeong Jin, Meng Qu, Xisen Jin, and Xiang Ren. 2020. Recurrent event network: Autoregressive structure inference over temporal knowledge graphs. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6669-6683, Online. Association for Computational Linguistics.",
+            "Julien Leblay and Melisachew Wudage Chekol. 2018. Deriving validity time in knowledge graph. In *Companion Proceedings of the The Web Conference* 2018, WWW '18, page 1771-1776, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee.",
+            "Dong-Ho Lee, Kian Ahrabian, Woojeong Jin, Fred Morstatter, and Jay Pujara. 2023. Temporal knowledge graph forecasting without knowledge using incontext learning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 544-557, Singapore. Association for Computational Linguistics.",
+            "Yujia Li, Shiliang Sun, and Jing Zhao. 2022. Tirgn: Time-guided recurrent graph network with local-global historical patterns for temporal knowledge graph reasoning. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 2152-2158. International Joint Conferences on Artificial Intelligence Organization. Main Track.",
+            "Zixuan Li, Xiaolong Jin, Saiping Guan, Wei Li, Jiafeng Guo, Yuanzhuo Wang, and Xueqi Cheng. 2021a. Search from history and reason for future: Two-stage reasoning on temporal knowledge graphs. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4732-4743, Online. Association for Computational Linguistics.",
+            "Zixuan Li, Xiaolong Jin, Wei Li, Saiping Guan, Jiafeng Guo, Huawei Shen, Yuzhuo Wang, and Xueqi Cheng. 2021b. Temporal knowledge graph reasoning based on evolutionary representation learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, page 408-417, New York, NY, USA. Association for Computing Machinery.",
+            "Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, and Xinwang Liu. 2023. Learn from relational correlations and periodic events for temporal knowledge graph reasoning. In Proceedings of the 46th International ACM SIGIR"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Conference on Research and Development in Information Retrieval, SIGIR '23, page 1559-1568, New York, NY, USA. Association for Computing Machinery.",
+            "Ruotong Liao, Xu Jia, Yangzhe Li, Yunpu Ma, and Volker Tresp. 2024. GenTKG: Generative forecasting on temporal knowledge graph with large language models. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 4303-4317, Mexico City, Mexico. Association for Computational Linguistics.",
+            "Kangzheng Liu, Feng Zhao, Guandong Xu, Xianzhi Wang, and Hai Jin. 2023. Retia: Relation-entity twin-interact aggregation for temporal knowledge graph extrapolation. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 1761-1774.",
+            "Yushan Liu, Yunpu Ma, Marcel Hildebrandt, Mitchell Joblin, and Volker Tresp. 2022. Tlogic: Temporal logical rules for explainable link forecasting on temporal knowledge graphs. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4):4120-4127.",
+            "Farzaneh Mahdisoltani, Joanna Biega, and Fabian M. Suchanek. 2013. YAGO3: A Knowledge Base from Multilingual Wikipedias. In CIDR, Asilomar, United States.",
+            "Fred Morstatter, Aram Galstyan, Gleb Satyukov, Daniel Benjamin, Andres Abeliuk, Mehrnoosh Mirtaheri, KSM Tozammel Hossain, Pedro Szekely, Emilio Ferrara, Akira Matsui, Mark Steyvers, Stephen Bennet, David Budescu, Mark Himmelstein, Michael Ward, Andreas Beger, Michele Catasta, Rok Sosic, Jure Leskovec, Pavel Atanasov, Regina Joseph, Rajiv Sethi, and Ali Abbas. 2019. Sage: A hybrid geopolitical event forecasting system. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 6557-6559. International Joint Conferences on Artificial Intelligence Organization.",
+            "Haohai Sun, Jialun Zhong, Yunpu Ma, Zhen Han, and Kun He. 2021. TimeTraveler: Reinforcement learning for temporal knowledge graph forecasting. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8306-8319, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.",
+            "Yuwei Xia, Ding Wang, Qiang Liu, Liang Wang, Shu Wu, and Xiaoyu Zhang. 2024. Chain-of-history reasoning for temporal knowledge graph forecasting. Preprint, arXiv:2402.14382."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "20936",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Yi Xu, Junjie Ou, Hui Xu, and Luoyi Fu. 2023. Temporal knowledge graph reasoning with historical contrastive learning. Proceedings of the AAAI Conference on Artificial Intelligence, 37(4):4765-4773.",
+        "bbox": [
+            115,
+            85,
+            489,
+            139
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Mengqi Zhang, Yuwei Xia, Qiang Liu, Shu Wu, and Liang Wang. 2023. Learning latent relations for temporal knowledge graph reasoning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12617-12631, Toronto, Canada. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            152,
+            489,
+            244
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "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.",
+        "bbox": [
+            115,
+            258,
+            489,
+            363
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Multi-Token ID Generation Issue",
+        "text_level": 1,
+        "bbox": [
+            114,
+            381,
+            440,
+            395
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "G2S is designed to generate single-token IDs, which may fail on the queries whose answers are multi-token IDs. This issue, referred to as the multi-token ID generation issue, particularly occurs when adopting the GID strategy, since the IDs assigned by this strategy are usually much larger. This issue arises in certain cases, potentially affecting the model's overall performance. As shown in Table 5, only  $15.4\\%$  of the queries in ICEWS14 have an answer with a multi-token ID, so the impact of this issue is limited. For ICEWS18, this percentage increases to  $40.2\\%$ , which impacts the overall performance of the model to some extent. For YAGO, this percentage is  $91.8\\%$ , which means GID will fail on most queries in this dataset.",
+        "bbox": [
+            112,
+            409,
+            489,
+            651
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/9b68107d46d30710be519adbb238970658accc7d662df82aa46a063105923ea7.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 5: Statistics on percentage of multi-token answers in validation set of ICEWS14, ICEWS18, and YAGO. \"# Queries\" denotes the number of total queries; and \"# MT IDs\" denotes the number of queries where the answers are multi-token IDs."
+        ],
+        "table_body": "<table><tr><td></td><td>ICEWS14</td><td>ICEWS18</td><td>YAGO</td></tr><tr><td># Queries</td><td>17,028</td><td>91,990</td><td>39,046</td></tr><tr><td># MT IDs</td><td>2,621</td><td>36,981</td><td>35,852</td></tr><tr><td>Percentage</td><td>15.4%</td><td>40.2%</td><td>91.8%</td></tr></table>",
+        "bbox": [
+            117,
+            664,
+            485,
+            747
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Dataset Statistics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            860,
+            302,
+            876
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The basic statistics of the datasets are shown in Table 7.",
+        "bbox": [
+            112,
+            889,
+            487,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/2cd055e20d4785fa71b92c605c5238570815dd5f7cd3cca984473f94cc6debd1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>##Entities ###2:Citizen (Nigeria)</td></tr><tr><td>...28:Boko Haram</td></tr><tr><td>...192:Cameroon</td></tr><tr><td>...</td></tr><tr><td>##Relations ##</td></tr><tr><td>...11:Use conventional military force</td></tr><tr><td>...25:Abduct, hijack, or take hostage</td></tr><tr><td>...</td></tr><tr><td>##History ##40:[28,25,2]</td></tr><tr><td>...19:[28,25,192]</td></tr><tr><td>...17:[28,25,2]15:[28,25,2]</td></tr><tr><td>...2:[28,25,2]1:[28,11,2]</td></tr><tr><td>##Query ##0:[28,25,?]</td></tr><tr><td>##Answer ##2</td></tr></table>",
+        "bbox": [
+            569,
+            85,
+            826,
+            527
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Figure 3: An example of the prompt together with answer in specific learning stage.",
+        "bbox": [
+            507,
+            538,
+            884,
+            568
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Additional Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            593,
+            709,
+            608
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 6 shows experimental results which compare G2S with additional baseline models under standard setting. The baseline models include TiRGN (Li et al., 2022), CENET (Xu et al., 2023), RETIA (Liu et al., 2023), L2TKG (Zhang et al., 2023), and RPC (Liang et al., 2023). As shown in the table, G2S obtains higher or competitive performance compared to the baselines.",
+        "bbox": [
+            507,
+            620,
+            884,
+            750
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "D Prompt Example and Case Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            763,
+            835,
+            780
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "An example of the prompt in specific learning stage is shown in Figure 3. The prompt in general learning stage excludes \"Entity\" and \"Relation\" parts.",
+        "bbox": [
+            507,
+            791,
+            882,
+            839
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We found a training query in GDELT (in RID) which is similar in structure but different in semantic, as shown in Figure 4. The specific scenario information for this query is shown in Figure 5:",
+        "bbox": [
+            507,
+            840,
+            882,
+            904
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Note that the \"history\" part of the above two",
+        "bbox": [
+            526,
+            904,
+            880,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "20937",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/0e1865b734343f89fcf4266fbbd05858eab0e33f15e046c8e5668ee1caee5d64.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model\\Dataset</td><td colspan=\"3\">ICEWS14</td><td colspan=\"3\">ICEWS18</td><td colspan=\"3\">ICEWS05-15</td></tr><tr><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td><td>H@1</td><td>H@3</td><td>H@10</td></tr><tr><td>RiRGN</td><td>33.73</td><td>49.85</td><td>64.46</td><td>23.10</td><td>37.90</td><td>54.20</td><td>39.07</td><td>55.75</td><td>70.11</td></tr><tr><td>CENET</td><td>30.18</td><td>41.79</td><td>55.20</td><td>23.47</td><td>34.05</td><td>47.27</td><td>38.14</td><td>49.89</td><td>62.53</td></tr><tr><td>RETIA</td><td>32.47</td><td>48.01</td><td>63.64</td><td>21.96</td><td>36.18</td><td>51.95</td><td>34.02</td><td>49.64</td><td>64.42</td></tr><tr><td>L2TKG</td><td>35.36</td><td>-</td><td>71.05</td><td>22.15</td><td>-</td><td>55.04</td><td>41.86</td><td>-</td><td>80.69</td></tr><tr><td>RPC</td><td>34.87</td><td>49.80</td><td>65.08</td><td>24.34</td><td>38.74</td><td>55.89</td><td>39.47</td><td>57.11</td><td>71.75</td></tr><tr><td>G2S</td><td>38.33</td><td>54.12</td><td>68.58</td><td>23.04</td><td>35.32</td><td>46.62</td><td>41.39</td><td>58.15</td><td>73.34</td></tr></table>",
+        "bbox": [
+            132,
+            80,
+            862,
+            240
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/760ab1bf0c295a1fdaabaa813d115cfc2e48930b5b9ef7adee9ad5c8175784f3.jpg",
+        "table_caption": [
+            "Table 6: Results of TKG reasoning on datasets ICEWS14, ICEWS18, and ICEWS05-15. The best results among all models are in boldface. The experimental results of baselines are cited from Chen et al. (2024)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Schema</td><td># Entity</td><td># Relation</td><td># Train</td><td># Valid</td><td># Test</td><td>Granularity</td></tr><tr><td>ICEWS14</td><td>CAMEO</td><td>7,128</td><td>230</td><td>74,845</td><td>8,514</td><td>7,371</td><td>1 day</td></tr><tr><td>ICEWS18</td><td>CAMEO</td><td>23,033</td><td>256</td><td>373,018</td><td>45,995</td><td>49,545</td><td>1 day</td></tr><tr><td>ICEWS05-15</td><td>CAMEO</td><td>10,488</td><td>251</td><td>368,868</td><td>46,302</td><td>46,159</td><td>1 day</td></tr><tr><td>YAGO</td><td>YAGO</td><td>10,623</td><td>10</td><td>161,540</td><td>19,523</td><td>20,026</td><td>1 year</td></tr><tr><td>GDELT</td><td>CAMEO</td><td>7,691</td><td>240</td><td>1,734,399</td><td>238,765</td><td>305,241</td><td>15 min</td></tr><tr><td>WIKI</td><td>Wikidata</td><td>12,554</td><td>24</td><td>539,286</td><td>67,538</td><td>63,110</td><td>1 year</td></tr></table>",
+        "bbox": [
+            127,
+            292,
+            868,
+            423
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Statistics of the datasets. \"# Entity/Relation\" denotes the number of entities/relations, and \"# Train/Valid/Test\" denotes the number of facts in train-valid/test set, respectively.",
+        "bbox": [
+            112,
+            432,
+            878,
+            462
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": ">>> History\n      \n...\n...\n294:[0,1,6]\n...\n151:[0,1,1]\n...\n78:[0,1,1]\n...\n32:[0,1,1]\n...\n7:[0,5,1]\n>>> Query\n      \n0:[0,1,?] \n>>> Answer\n      \n1",
+        "guess_lang": "txt",
+        "bbox": [
+            173,
+            489,
+            426,
+            734
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "queries are represented in anonymous temporal structures, thus the model can mine the patterns more easily. Though these two samples use different anonymization strategies and are different in time granularity, they follow the same pattern. This pattern may contain facts related to the query ([28,25, 192], [28,25,2], [0,1,6] and [0,1,1]) and the correct answer depends on the frequency of tail",
+        "bbox": [
+            112,
+            791,
+            487,
+            921
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/bbf084394fe2af32c32db3357131f037924b4ee0dc61d98cf9b8f5ff211ced6d.jpg",
+        "image_caption": [
+            "Figure 4: A training sample in general learning stage.",
+            "Figure 5: The corresponding entity and relation ID mapping."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            569,
+            491,
+            823,
+            678
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "entity. This case shows the possibility of learning general patterns across datasets.",
+        "bbox": [
+            507,
+            750,
+            880,
+            783
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "20938",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    }
+]