Add Batch 2aac6986-bc5b-4f18-bd6a-02503e5a0655 data

.gitattributes

@@ -556,3 +556,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/GRAD_[[:space:]]Generative[[:space:]]Retrieval-Aligned[[:space:]]Demonstration[[:space:]]Sampler[[:space:]]for[[:space:]]Efficient[[:space:]]Few-Shot[[:space:]]Reasoning/688c8cc7-a3da-439b-8f40-37ea8aadc22c_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/GRIL_[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Retrieval-Integrated[[:space:]]Learning[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/e32bf48a-ec66-46bd-950c-09b20c640a3d_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/GRPO-Guided[[:space:]]Modality[[:space:]]Selection[[:space:]]Enhanced[[:space:]]LoRA-Tuned[[:space:]]LLMs[[:space:]]for[[:space:]]Multimodal[[:space:]]Emotion[[:space:]]Recognition/f48c336a-f9fc-4f20-8d3e-6375c744850c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GRV-KBQA_[[:space:]]A[[:space:]]Three-Stage[[:space:]]Framework[[:space:]]for[[:space:]]Knowledge[[:space:]]Base[[:space:]]Question[[:space:]]Answering[[:space:]]with[[:space:]]Decoupled[[:space:]]Logical[[:space:]]Structure,[[:space:]]Semantic[[:space:]]Grounding[[:space:]]and[[:space:]]Structure-Aware[[:space:]]Validation/ca7c2f14-d7fa-4282-925f-6217c4537617_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GTA_[[:space:]]Supervised-Guided[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]for[[:space:]]Text[[:space:]]Classification[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/5317a317-b31f-4425-8151-f38a993fa23c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GUARD_[[:space:]]Glocal[[:space:]]Uncertainty-Aware[[:space:]]Robust[[:space:]]Decoding[[:space:]]for[[:space:]]Effective[[:space:]]and[[:space:]]Efficient[[:space:]]Open-Ended[[:space:]]Text[[:space:]]Generation/b5c8441d-feae-4313-aa6c-227be1980008_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GeLoRA_[[:space:]]Geometric[[:space:]]Adaptive[[:space:]]Ranks[[:space:]]For[[:space:]]Efficient[[:space:]]LoRA[[:space:]]Fine-tuning/408657ff-3659-4887-abb9-ebd6df5c83fb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GenPTQ_[[:space:]]Green[[:space:]]Post-Training[[:space:]]Quantization[[:space:]]for[[:space:]]Large-Scale[[:space:]]ASR[[:space:]]Models[[:space:]]with[[:space:]]Mixed-Precision[[:space:]]Bit[[:space:]]Allocation/99e69a29-243d-4d97-908f-6d69c507862a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GenPilot_[[:space:]]A[[:space:]]Multi-Agent[[:space:]]System[[:space:]]for[[:space:]]Test-Time[[:space:]]Prompt[[:space:]]Optimization[[:space:]]in[[:space:]]Image[[:space:]]Generation/5428225a-34af-4b71-95e2-69e7be23818e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GenPoE_[[:space:]]Generative[[:space:]]Passage-level[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]for[[:space:]]Knowledge[[:space:]]Enhancement[[:space:]]of[[:space:]]LLMs/0fe8ac59-edc9-41a5-8924-a4c69cfc95df_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Generation-Augmented[[:space:]]Retrieval_[[:space:]]Rethinking[[:space:]]the[[:space:]]Role[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Zero-Shot[[:space:]]Relation[[:space:]]Extraction/06f5d856-0d14-45d9-9bc3-7006f452ae3b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GeoChain_[[:space:]]Multimodal[[:space:]]Chain-of-Thought[[:space:]]for[[:space:]]Geographic[[:space:]]Reasoning/2884051f-2e1a-4d25-8d9b-40c2021e1eb9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GeoDANO_[[:space:]]Geometric[[:space:]]VLM[[:space:]]with[[:space:]]Domain[[:space:]]Agnostic[[:space:]]Vision[[:space:]]Encoder/fe9bd078-9e83-4103-9778-2a4a22e14fff_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GeoPQA_[[:space:]]Bridging[[:space:]]the[[:space:]]Visual[[:space:]]Perception[[:space:]]Gap[[:space:]]in[[:space:]]MLLMs[[:space:]]for[[:space:]]Geometric[[:space:]]Reasoning/a92f178e-8b3b-4323-aed1-83c6f6bc853e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Glitter_[[:space:]]A[[:space:]]Multi-Sentence,[[:space:]]Multi-Reference[[:space:]]Benchmark[[:space:]]for[[:space:]]Gender-Fair[[:space:]]German[[:space:]]Machine[[:space:]]Translation/9f6352b7-b609-41f0-bdd1-c93bbd98f33c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GmSLM[[:space:]]_[[:space:]]Generative[[:space:]]Marmoset[[:space:]]Spoken[[:space:]]Language[[:space:]]Modeling/89b253a6-162c-41b1-a7a0-7fa57530e061_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Golden[[:space:]]Touchstone_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Bilingual[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Financial[[:space:]]Large[[:space:]]Language[[:space:]]Models/aca70d61-8d88-4875-bda0-155d05a01103_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Graph-Reward-SQL_[[:space:]]Execution-Free[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]for[[:space:]]Text-to-SQL[[:space:]]via[[:space:]]Graph[[:space:]]Matching[[:space:]]and[[:space:]]Stepwise[[:space:]]Reward/671895dd-9d21-4b57-91bb-c5dca6dc74d5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GraphCheck_[[:space:]]Multipath[[:space:]]Fact-Checking[[:space:]]with[[:space:]]Entity-Relationship[[:space:]]Graphs/64081260-7407-4634-9456-c0f9eabdb364_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GreekBarBench_[[:space:]]A[[:space:]]Challenging[[:space:]]Benchmark[[:space:]]for[[:space:]]Free-Text[[:space:]]Legal[[:space:]]Reasoning[[:space:]]and[[:space:]]Citations/37c3c14c-946a-437b-b702-7f928743158a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Grounded-VideoLLM_[[:space:]]Sharpening[[:space:]]Fine-grained[[:space:]]Temporal[[:space:]]Grounding[[:space:]]in[[:space:]]Video[[:space:]]Large[[:space:]]Language[[:space:]]Models/959e284c-43d1-407f-bbdf-8175847e8aba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Guaranteed[[:space:]]Guess_[[:space:]]A[[:space:]]Language[[:space:]]Modeling[[:space:]]Approach[[:space:]]for[[:space:]]CISC-to-RISC[[:space:]]Transpilation[[:space:]]with[[:space:]]Testing[[:space:]]Guarantees/1f14fd9a-f2cf-41fc-91ac-1cc4650056ca_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Guess[[:space:]]What[[:space:]]I[[:space:]]am[[:space:]]Thinking_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Inner[[:space:]]Thought[[:space:]]Reasoning[[:space:]]of[[:space:]]Role-Playing[[:space:]]Language[[:space:]]Agents/b4888d2c-b526-4632-bd2e-4af9eb8460fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/GuiLoMo_[[:space:]]Allocating[[:space:]]Experts[[:space:]]and[[:space:]]Ranks[[:space:]]for[[:space:]]LoRA-MoE[[:space:]]via[[:space:]]Bilevel[[:space:]]Optimization[[:space:]]with[[:space:]]GuidedSelection[[:space:]]Vectors/fd67f66e-65a4-46da-915e-e341c715ac2b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Guiding[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Biomedical[[:space:]]Entity[[:space:]]Linking[[:space:]]via[[:space:]]Restrictive[[:space:]]and[[:space:]]Contrastive[[:space:]]Decoding/d3772de2-981f-4434-8679-98474ef487f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HARE_[[:space:]]an[[:space:]]entity[[:space:]]and[[:space:]]relation[[:space:]]centric[[:space:]]evaluation[[:space:]]framework[[:space:]]for[[:space:]]histopathology[[:space:]]reports/23715ed9-df41-489b-8cd6-6bd5e2990b85_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HATECAT-TR_[[:space:]]A[[:space:]]Hate[[:space:]]Speech[[:space:]]Span[[:space:]]Detection[[:space:]]and[[:space:]]Categorization[[:space:]]Dataset[[:space:]]for[[:space:]]Turkish/051fb5bb-cc6c-4440-8308-f49b85e93b1a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HAWK_[[:space:]]Highlighting[[:space:]]Entity-aware[[:space:]]Knowledge[[:space:]]for[[:space:]]Alleviating[[:space:]]Information[[:space:]]Sparsity[[:space:]]in[[:space:]]Long[[:space:]]Contexts/4882fa8c-13f2-4264-bc62-1bb9fdfea69e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HDiff_[[:space:]]Confidence-Guided[[:space:]]Denoising[[:space:]]Diffusion[[:space:]]for[[:space:]]Robust[[:space:]]Hyper-relational[[:space:]]Link[[:space:]]Prediction/5c70b82f-3a77-44dd-9426-5bc3c13149cc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HEAL_[[:space:]]A[[:space:]]Hypothesis-Based[[:space:]]Preference-Aware[[:space:]]Analysis[[:space:]]Framework/10012a38-eed8-4563-a4ef-e749efa0e600_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HEAL_[[:space:]]An[[:space:]]Empirical[[:space:]]Study[[:space:]]on[[:space:]]Hallucinations[[:space:]]in[[:space:]]Embodied[[:space:]]Agents[[:space:]]Driven[[:space:]]by[[:space:]]Large[[:space:]]Language[[:space:]]Models/dbf5a136-7a4b-4e2d-b6ae-6fb7222b500b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HEAL_[[:space:]]Hybrid[[:space:]]Enhancement[[:space:]]with[[:space:]]LLM-based[[:space:]]Agents[[:space:]]for[[:space:]]Text-attributed[[:space:]]Hypergraph[[:space:]]Self-supervised[[:space:]]Representation[[:space:]]Learning/2c117dc6-e36c-4020-8110-cfb1361fdcdc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HGAdapter_[[:space:]]Hypergraph-based[[:space:]]Adapters[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Code[[:space:]]Summarization[[:space:]]and[[:space:]]Clone[[:space:]]Detection/a4ee6435-1964-4755-a6a7-e9aab6fef982_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HIRAG_[[:space:]]Hierarchical-Thought[[:space:]]Instruction-Tuning[[:space:]]Retrieval-Augmented[[:space:]]Generation/d7e47a5d-e2e9-4f2d-b406-41169151ab13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HMCL_[[:space:]]Task-Optimal[[:space:]]Text[[:space:]]Representation[[:space:]]Adaptation[[:space:]]through[[:space:]]Hierarchical[[:space:]]Contrastive[[:space:]]Learning/7f317cfb-c8c9-4af2-a002-5e46341f387d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hallucination[[:space:]]Detection[[:space:]]in[[:space:]]Structured[[:space:]]Query[[:space:]]Generation[[:space:]]via[[:space:]]LLM[[:space:]]Self-Debating/77f974d0-8c35-42ea-bc09-53ea7662f4ad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HapticCap_[[:space:]]A[[:space:]]Multimodal[[:space:]]Dataset[[:space:]]and[[:space:]]Task[[:space:]]for[[:space:]]Understanding[[:space:]]User[[:space:]]Experience[[:space:]]of[[:space:]]Vibration[[:space:]]Haptic[[:space:]]Signals/d909e81e-0f1b-4478-9180-cf370ec80b40_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hard[[:space:]]Negatives,[[:space:]]Hard[[:space:]]Lessons_[[:space:]]Revisiting[[:space:]]Training[[:space:]]Data[[:space:]]Quality[[:space:]]for[[:space:]]Robust[[:space:]]Information[[:space:]]Retrieval[[:space:]]with[[:space:]]LLMs/bef985b4-0881-42ea-af0e-2f0eaec341a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hardware-Aware[[:space:]]Parallel[[:space:]]Prompt[[:space:]]Decoding[[:space:]]for[[:space:]]Memory-Efficient[[:space:]]Acceleration[[:space:]]of[[:space:]]LLM[[:space:]]Inference/7ffabb72-a8f7-4b9d-be93-c54961be9fa4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Harry[[:space:]]Potter[[:space:]]is[[:space:]]Still[[:space:]]Here![[:space:]]Probing[[:space:]]Knowledge[[:space:]]Leakage[[:space:]]in[[:space:]]Targeted[[:space:]]Unlearned[[:space:]]Large[[:space:]]Language[[:space:]]Models/13926e6c-9138-4ea8-a661-a18eb81e86f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HebID_[[:space:]]Detecting[[:space:]]Social[[:space:]]Identities[[:space:]]in[[:space:]]Hebrew-language[[:space:]]Political[[:space:]]Text/b48a313a-c149-45e1-9666-fbd5bd85c515_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HetGCoT_[[:space:]]Heterogeneous[[:space:]]Graph-Enhanced[[:space:]]Chain-of-Thought[[:space:]]LLM[[:space:]]Reasoning[[:space:]]for[[:space:]]Academic[[:space:]]Question[[:space:]]Answering/be416281-4f50-4e89-b250-243cd735d92d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HiMATE_[[:space:]]A[[:space:]]Hierarchical[[:space:]]Multi-Agent[[:space:]]Framework[[:space:]]for[[:space:]]Machine[[:space:]]Translation[[:space:]]Evaluation/3cc804bc-4397-43d8-95e6-5cf2a6d14444_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hidden[[:space:]]Ghost[[:space:]]Hand_[[:space:]]Unveiling[[:space:]]Backdoor[[:space:]]Vulnerabilities[[:space:]]in[[:space:]]MLLM-Powered[[:space:]]Mobile[[:space:]]GUI[[:space:]]Agents/c19e195e-69a2-4d31-978f-0540957ee9d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HierPrompt_[[:space:]]Zero-Shot[[:space:]]Hierarchical[[:space:]]Text[[:space:]]Classification[[:space:]]with[[:space:]]LLM-Enhanced[[:space:]]Prototypes/50569b6b-9b00-4931-80b9-074f9dab3e57_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hierarchical[[:space:]]Reward[[:space:]]Modeling[[:space:]]for[[:space:]]Fault[[:space:]]Localization[[:space:]]in[[:space:]]Large[[:space:]]Code[[:space:]]Repositories/487610d2-985d-4452-9b47-bf8cb39a4136_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HighMATH_[[:space:]]Evaluating[[:space:]]Math[[:space:]]Reasoning[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Breadth[[:space:]]and[[:space:]]Depth/dc21a5e3-70c7-431b-a34b-6ba1a4695311_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/HomoGraphAdapter_[[:space:]]A[[:space:]]Homogeneous[[:space:]]Graph[[:space:]]Neural[[:space:]]Network[[:space:]]as[[:space:]]an[[:space:]]Effective[[:space:]]Adapter[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/6a87aa66-fe6e-4b1d-92fd-a405570c05ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Hopscotch_[[:space:]]Discovering[[:space:]]and[[:space:]]Skipping[[:space:]]Redundancies[[:space:]]in[[:space:]]Language[[:space:]]Models/56baf399-91d2-4e2e-b36e-0001a64468bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Can[[:space:]]Input[[:space:]]Reformulation[[:space:]]Improve[[:space:]]Tool[[:space:]]Usage[[:space:]]Accuracy[[:space:]]in[[:space:]]a[[:space:]]Complex[[:space:]]Dynamic[[:space:]]Environment_[[:space:]]A[[:space:]]Study[[:space:]]on[[:space:]]tau-bench/e40f12aa-0207-4710-a7d3-58adffa93442_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Diversely[[:space:]]Can[[:space:]]Language[[:space:]]Models[[:space:]]Solve[[:space:]]Problems_[[:space:]]Exploring[[:space:]]the[[:space:]]Algorithmic[[:space:]]Diversity[[:space:]]of[[:space:]]Model-Generated[[:space:]]Code/303354e8-4e93-4cff-9153-120a81b04847_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Do[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Perform[[:space:]]on[[:space:]]PDE[[:space:]]Discovery_[[:space:]]A[[:space:]]Coarse-to-fine[[:space:]]Perspective/cb115a55-16e5-457a-bf7d-e5e4dc1b6283_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Does[[:space:]]Cognitive[[:space:]]Bias[[:space:]]Affect[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]on[[:space:]]the[[:space:]]Anchoring[[:space:]]Effect[[:space:]]in[[:space:]]Price[[:space:]]Negotiation[[:space:]]Simulations/46a3350c-ebe4-4f44-89ce-8ac3f0f946f3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Does[[:space:]]Knowledge[[:space:]]Selection[[:space:]]Help[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]Generation_/18e814b9-d7ff-42ef-bd94-0149bb9a98c1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Good[[:space:]]are[[:space:]]LLM-based[[:space:]]Rerankers_[[:space:]]An[[:space:]]Empirical[[:space:]]Analysis[[:space:]]of[[:space:]]State-of-the-Art[[:space:]]Reranking[[:space:]]Models/9339f7c4-1b56-4a1c-a566-b8a4e04eb8ee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Jailbreak[[:space:]]Defenses[[:space:]]Work[[:space:]]and[[:space:]]Ensemble_[[:space:]]A[[:space:]]Mechanistic[[:space:]]Investigation/6a1c2600-3e8a-4b82-bfe5-e244ff5e1eef_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Much[[:space:]]Do[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Know[[:space:]]about[[:space:]]Human[[:space:]]Motion_[[:space:]]A[[:space:]]Case[[:space:]]Study[[:space:]]in[[:space:]]3D[[:space:]]Avatar[[:space:]]Control/06f1af86-c517-45db-b189-5749f3cacbea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Real[[:space:]]Are[[:space:]]Synthetic[[:space:]]Therapy[[:space:]]Conversations_[[:space:]]Evaluating[[:space:]]Fidelity[[:space:]]in[[:space:]]Prolonged[[:space:]]Exposure[[:space:]]Dialogues/05962685-6e0c-4d8c-9da8-7ec25b6fd7d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Reliable[[:space:]]is[[:space:]]Multilingual[[:space:]]LLM-as-a-Judge_/88c10a96-7922-4e8a-b12e-e62281d202c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Sampling[[:space:]]Affects[[:space:]]the[[:space:]]Detectability[[:space:]]of[[:space:]]Machine-written[[:space:]]texts_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Study/bd9cd3e3-d944-48b4-bc47-1da3be59c49f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]Well[[:space:]]Can[[:space:]]Reasoning[[:space:]]Models[[:space:]]Identify[[:space:]]and[[:space:]]Recover[[:space:]]from[[:space:]]Unhelpful[[:space:]]Thoughts_/e4866f17-2d21-419c-9294-edbfab480c52_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]a[[:space:]]Bilingual[[:space:]]LM[[:space:]]Becomes[[:space:]]Bilingual_[[:space:]]Tracing[[:space:]]Internal[[:space:]]Representations[[:space:]]with[[:space:]]Sparse[[:space:]]Autoencoders/bf58ac33-799c-4b5e-950e-a64d2bc19411_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]do[[:space:]]Language[[:space:]]Models[[:space:]]Generate[[:space:]]Slang_[[:space:]]A[[:space:]]Systematic[[:space:]]Comparison[[:space:]]between[[:space:]]Human[[:space:]]and[[:space:]]Machine-Generated[[:space:]]Slang[[:space:]]Usages/a7a593a3-26a3-4052-8dac-796f069cdd53_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/How[[:space:]]to[[:space:]]Generalize[[:space:]]the[[:space:]]Detection[[:space:]]of[[:space:]]AI-Generated[[:space:]]Text_[[:space:]]Confounding[[:space:]]Neurons/8bd1a96c-eb6a-4b12-97da-ac2a2778459a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Human-Inspired[[:space:]]Obfuscation[[:space:]]for[[:space:]]Model[[:space:]]Unlearning_[[:space:]]Local[[:space:]]and[[:space:]]Global[[:space:]]Strategies[[:space:]]with[[:space:]]Hyperbolic[[:space:]]Representations/827580b9-daee-4f17-9f73-b43754214845_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Humanity’s[[:space:]]Last[[:space:]]Code[[:space:]]Exam_[[:space:]]Can[[:space:]]Advanced[[:space:]]LLMs[[:space:]]Conquer[[:space:]]Human’s[[:space:]]Hardest[[:space:]]Code[[:space:]]Competition_/8a4bafab-99a5-4429-b80f-2894c7e31f4c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Humor[[:space:]]in[[:space:]]Pixels_[[:space:]]Benchmarking[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models[[:space:]]Understanding[[:space:]]of[[:space:]]Online[[:space:]]Comics/30c39bf0-e05b-47cd-8b04-4866f611adbd_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/GRV-KBQA_ A Three-Stage Framework for Knowledge Base Question Answering with Decoupled Logical Structure, Semantic Grounding and Structure-Aware Validation/ca7c2f14-d7fa-4282-925f-6217c4537617_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82f44d34a3ca3ee3b4347671f0e39f44c6224e6610b9c7e1824647bf5f059ab4
+size 1147495

2025/GRV-KBQA_ A Three-Stage Framework for Knowledge Base Question Answering with Decoupled Logical Structure, Semantic Grounding and Structure-Aware Validation/full.md

@@ -0,0 +1,421 @@
+# GRV-KBQA: A Three-Stage Framework for Knowledge Base Question Answering with Decoupled Logical Structure, Semantic Grounding and Structure-Aware Validation
+
+Yuhang Tian $^{1*}$ , Pan Yang $^{1*}$ , Dandan Song $^{1\dagger}$ , Zhijing Wu $^{1}$ , Hao Wang $^{1}$
+
+$^{1}$ School of Computer Science and Technology, Beijing Institute of Technology, China {tianyuhang, sdd}  $@$  bit.edu.cn, zaixiatongxin@gmail.com
+
+# Abstract
+
+Knowledge Base Question Answering (KBQA) is a fundamental task that enables natural language interaction with structured knowledge bases (KBs). Given a natural language question, KBQA aims to retrieve the answers from the KB. However, existing approaches, including retrieval-based, semantic parsing-based methods and large-language model-based methods often suffer from generating non-executable queries and inefficiencies in query execution. To address these challenges, we propose GRV-KBQA, a three-stage framework that decouples logical structure generation from semantic grounding and incorporates structure-aware validation to enhance accuracy. Unlike previous methods, GRV-KBQA explicitly enforces KB constraints to improve alignment between generated logical forms and KB structures. Experimental results on WebQSP and CWQ show that GRV-KBQA significantly improves performance over existing approaches. The ablation study conducted confirms the effectiveness of the decoupled logical form generation and validation mechanism of our framework.
+
+# 1 Introduction
+
+Knowledge bases (KBs) such as Freebase (Bollacker et al., 2008), Wikidata (Vrandecic and Krötzsch, 2014), and DBpedia (Auer et al., 2007), which contain vast amounts of structured data in the form of triples, are widely used due to their structured nature and the accuracy of the information they represent. Knowledge Base Questioning Answering (KBQA) is a popular application of KB, aiming to retrieve accurate answers from the KB for a given question.
+
+Previous methods for Knowledge Base Question Answering (KBQA) can be broadly categorized into two main classes: Information Retrieval-based (IR-based) (Miller et al., 2016; Sun et al.,
+
+![](images/725ba80dab090491268f07229632dbb4c32c63866a8ddb198c7e3197379b6d74.jpg)
+
+![](images/301929b3e6ab2f9f54115da4c5d30711bf89f625dcc7bc430f061aefa08da053.jpg)
+
+![](images/7f20eefcf89d1871af3556b2a98ca056ec0ab47521eddef688b5dfdeb2c24402.jpg)  
+Figure 1: Comparison of GRV-KBQA with previous methods.
+
+2019; He et al., 2021a; Zhang et al., 2022) methods and Semantic Parsing-based (SP-based) methods. IR-based methods primarily focus on retrieving relevant subgraphs from the KB and performing reasoning over these subgraphs to derive the answer. Besides, SP-based methods (Das et al., 2021; Lan and Jiang, 2020; Ye et al., 2022; Shu et al., 2022) aim to transform natural language questions into executable logical forms, such as SPARQL queries, which are then executed directly on the KB to obtain the answer. Currently, as the large language models (LLMs) have exceptional generative capabilities and powerful learning abilities,
+
+several research efforts have proposed sequence-to-sequence (seq-to-seq) approaches based on LLMs, which directly generate logical forms, align them to the KB, and then execute them on the KB to retrieve answers. KB-Coder (Nie et al., 2024) and ARG-KBQA (Tian et al., 2024) have leveraged LLMs for few-shot in-context learning, where a small number of examples guides the model to generate logical forms. To fully leverage the capabilities of LLMs while avoiding the use of black-box models, ChatKBQA (Luo et al., 2024) fine-tunes open-source models to enable them to generate logical forms based on the given questions.
+
+However, existing methods still face two key issues: 1. Challenges in Generating Correct Logical Forms: They suffer from generating non-executable logical forms in many cases. For example, the basic logical structure of generated logical form is wrong, as shown in Figure 1. This limitation primarily arises from the one-step generation paradigm, which requires the model to simultaneously learn both the structural composition of the query and the semantic grounding of its components. This joint learning process increases the model's cognitive load, making it more prone to errors in both aspects, which can lead to errors in logical forms, including incorrect logical form structures (e.g., incorrect nesting of operations or missing/extra filtering conditions) and semantic mismatches (e.g., generating non-existent entities or relations). These errors often make the generated logical forms non-executable, preventing the framework from retrieving the correct answers. 2. Inefficiency in Aligning with the Knowledge Base: Existing methods, when executing the generated logical form, require aligning the generated entities and relations to KB. However, previous methods generate all possible combinations of candidates without considering available pruning optimizations, as shown in Figure 1, which would result in an enormous search space, causing significant time to be wasted on invalid queries and greatly slowing down the query process.
+
+To address these issues, we propose GRV-KBQA: A Three-Stage Framework for Knowledge Base Question Answering with Decoupled Logical Structure Generation, Semantic Grounding, and Structure-Aware Validation. For Challenges in Generating Correct Logical Forms, we introduce Decoupled Logical Structure Generation, which simplifies the generation of nested logical structures by employing a two-phase compositional
+
+approach. This method leverages a subquery chain template with placeholders for entities, relations, and intermediate results, which are then specified based on the question semantics and instantiated with corresponding KB elements to produce the complete logical form.
+
+For Inefficiency in Aligning with the Knowledge Base, we propose Structure-Aware Validation, which employs structure-aware validation techniques to verify the consistency and correctness of the relations and entities in the logical form, ensuring that the structure of the logic form complies with the constraints of the KB.
+
+We conduct experiments on two widely used datasets, WebQSP (Yih et al., 2016a) and CWQ (Talmor and Berant, 2018a), and the results demonstrate that GRV-KBQA outperforms other baselines on these datasets, highlighting the effectiveness of Decoupled Logical Structure Generation and Structure-Aware Validation.
+
+In summary, our contributions are as follows:
+
+- We propose Decoupled Logical Structure Generation, a two-phase compositional approach that reduces the complexity of generating nested logical structures, leading to more accurate and executable logical forms.  
+- We present Structure-Aware Validation, which verifies the consistency and correctness of generated logical forms, ensuring they comply with the structural constraints of the KB.  
+- We validate the effectiveness of GRV-KBQA, through experiments on the WebQSP and CWQ datasets, demonstrating improved performance over existing methods.
+
+# 2 Related Work
+
+Traditional Knowledge Base Question Answering (KBQA) methods can generally be divided into two main categories: information retrieval (IR)-based methods and semantic parsing (SP)-based methods. In the era of LLMs, several methods have emerged that leverage LLMs to tackle the KBQA task.
+
+IR-based methods (He et al., 2021b; Shi et al., 2021; Dong et al., 2023; Zhang et al., 2022; Jiang et al., 2023b) focus on extracting a subgraph from the KB that is specific to the given question, and then apply ranking algorithms to identify the top entities or employ text decoders to directly generate the answers.
+
+![](images/bb7d19d230a020b258b9c5c9c76e11971d6d8024074de611198f5779a346d991.jpg)  
+Figure 2: Overview of GRV-KBQA
+
+SP-based methods transform natural language questions to logical forms and execute on the KB to generate answers. Yih et al. (2015) proposes a novel semantic parsing framework, where logical form generation is framed as a staged search problem, leveraging the KB to prune the search space. TIARA (Shu et al., 2022), DECAF (Yu et al., 2023), and RnG-KBQA (Ye et al., 2022) all employ sequence-to-sequence models to fully construct S-expressions and offer numerous enhancements to the semantic parsing process. FCKBQA (Zhang et al., 2023) extracts pertinent fine-grained knowledge components from the KB and reconfigures them into intermediate-level knowledge pairs, which are subsequently used to generate the final logical forms.
+
+LLM-based methods generate logical forms either by prompting or fine-tuning LLMs. StructGPT (Jiang et al., 2023a) proposes a framework that enhances the reasoning ability of LLMs by using specialized interfaces to collect relevant evidence. Li et al. (2024) enhances LLMs for
+
+KGQA by introducing question decomposition and atomic retrieval, using a decomposition tree to guide atomic-level KG subgraph retrieval for answering complex questions.. A series of training-free frameworks for few-shot in-context learning have been proposed by Li et al. (2023), Nie et al. (2024) and Tian et al. (2024) which employ LLMs to generate logical forms and ground them on the KB. Additionally, some methods (Sun et al., 2024; Huang et al., 2024) utilize strategies of stepwise query construction and search answers from the KB. In the realm of sequence-to-sequence approaches, ChatKBQA (Luo et al., 2024) proposes a generate-then-retrieve framework that fine-tunes open-source LLMs to generate logical forms.
+
+In this paper, we present the GRV-KBQA framework, which separates the processes of logical form generation and semantic grounding while integrating structure-aware validation. Unlike previous approaches, our framework leverages the inherent structure of the KB to enhance the alignment of logical forms.
+
+It is worth noting that KB-Binder also incorporates a KB-aware optimization module for execution. However, while KB-Binder (Li et al., 2023) filters candidates by retrieving a fixed two-hop subgraph, our approach considers the exact number of hops and the directionality specified by the query structure.
+
+# 3 Preliminaries
+
+# 3.1 Definitions
+
+Knowledge Base (KB) A Knowledge Base (KB) is a RDF graph that stores various types of knowledge about the real world. It can be represented as a collection of triples, where each triple encodes a fact. Formally, a KB can be represented as  $\mathcal{K} = \{(s_i,r_i,t_i)|i = 1,2,\ldots \}$  where  $s_i$  is an entity,  $t_i$  is either an entity or a literal and  $r_i$  is the relation between  $s_i$  and  $t_i$ . Each entity in the KB is associated with a unique entity ID, which allows precise identification of the corresponding entity node within the KB. Additionally, each entity is labeled with a natural language description, making it easier for humans to read and analyze. For example, the label of entity ID "m.06by7" is "Rock Music". Similarly, each relation is hierarchically labeled to accurately represent the complex relationships between entities, e.g., "music.artistGenre".
+
+Logical Form A logical form is a highly structured intermediate representation derived from knowledge base queries. It hierarchically illustrates the process of retrieving answers from the KB. Compared to SPARQL queries, logical forms are more readable and easier to understand. In this work, the logical forms we employ primarily involve operators such as JOIN, AND, TC, ARGMAX, ARGMIN, and others. The JOIN operator is used for projection and selection. For instance, (JOIN (R <rel>) s) retrieves all tail entities that have the relation "<rel>" with the given head entity  $s$ . (JOIN <rel> t) retrieves all head entities that have the relation "<rel>" with the given tail entity  $t$ . The AND operator computes the intersection of two entity sets. ARGMAX, ARGMIN, TC, LT, and GT are conditional selection operators that filter results from a given entity set based on specified conditions.
+
+# 3.2 Problem Statement
+
+Knowledge Base Question Answering (KBQA) aims to retrieve answers from a structured KB given a natural language question. Given a natural language question  $q$ , the goal of KBQA is to gener
+
+ate a structured query  $Q$  that retrieves the correct answer set  $\mathcal{A}$  from  $\kappa$ :  $Q = \text{Convert}(q)$ , where  $\text{Convert}(\cdot)$  is the convert function. In this task, the structured query  $Q$  is often formulated in SPARQL or logical form. If  $Q$  is formulated as a logical form, it must be converted into a SPARQL query using a predefined transformation function. Finally, the SPARQL query is executed on the knowledge base  $\kappa$  to obtain the final answer set  $\mathcal{A}$ .
+
+# 4 Methodology
+
+# 4.1 Overview
+
+As shown in Figure 2, GRV-KBQA is a three-stage KBQA framework which leverages the generation capability of LLMs and consists of three parts: Decoupled Logical Structure Generation, Semantic Grounding, and Structure-Aware Validation.
+
+Given a natural language question, GRV-KBQA first generates a subquery chain template. Based on the template, GRV-KBQA constructs a query structure graph which is a directed graph that defines the structural constraints that candidate entity IDs and relations must satisfy. Then, GRV-KBQA determines the semantic information for each placeholder based on the generated template and the given natural language question. It retrieves candidate entity IDs and relations from the knowledge base. GRV-KBQA leverages the generated relations to filter out entity IDs that do not maintain structural consistency with the query structure. It searches for the first valid candidate combination that generates an executable logical form  $f$ . The execution result of the generated logical form  $f$  would be the final answer set  $\mathcal{A}$ .
+
+# 4.2 Decoupled Logical Structure Generation
+
+One of the key challenges in KBQA is ensuring that logical forms are both structurally well-formed and semantically aligned with the KB. To address this, we introduce a decoupled approach that separates logical structure generation from semantic grounding. Specifically, we employ an LLM to perform two tasks. The first task is generating a subquery chain template based on the input question  $q$ . The second task aims to determine the textual content for each placeholder in the template based on the given question. Through these two tasks, we decouple the generation of the final logical form in terms of both structure and content. We fine-tune the LLM on these two tasks to enhance the performance.
+
+Input: Knowledge Base  $\mathcal{K}$ , Subquery Chain Template  $p$ , Threshold  $t_e, t_r$ , Top-k Num  $n_e, n_r$
+
+Algorithm 1 Structure-Aware Validation  
+Output: The Final Answer Set  $\mathcal{A}$  
+Task 1: Subquery Chain Template Generation  
+$\mathcal{T}\gets$  GetQueryStructure(p)   
+ $\mathcal{P}_e\gets$  GetEntityPlaceholders(p)   
+ $\mathcal{P}_r\gets$  GetRelationPlaceholders(p)   
+ $\mathcal{L}_e\gets$  GetCandidateListsWithThreshold  $(\mathcal{P}_e,n_e,t_e)$ $\mathcal{L}_r\gets$  GetCandidateLists  $(\mathcal{P}_r)$    
+for  $p_e,L_{e_i}\in \mathrm{zip}(\mathcal{P}_e,\mathcal{L}_e)$  do  $L_{e_i}'\gets \emptyset$  for  $e\in L_{e_i}$  do  $\mathcal{P}_a\gets$  GetAssociatedRelationPlaceholders  $(p_e)$  if CheckStructuralConsistency(e,  $\mathcal{P}_a,\mathcal{T},\mathcal{K})$  then  $L_{e_i}'$  .Append(e) end if end for  $L_{e}'$  .Append  $(L_{e_i}')$    
+end for   
+ $\mathcal{R}\gets$  FindNonEmptyResults  $(\mathcal{L}_e',\mathcal{L}_r,\mathcal{T},\mathcal{K})$    
+if  $\mathcal{R}\neq \emptyset$  then return R   
+end if   
+for  $c\in$  GetCombinations  $(\mathcal{L}_e)$  do  $\mathcal{L}_r'\gets \emptyset$  for  $e,p_e\in \mathrm{zip}(c,\mathcal{P}_e)$  do  $\mathcal{P}_a\gets$  GetAssociatedRelationPlaceholders  $(p_e)$  for  $p_r\in \mathcal{P}_q$  do  $r\gets$  GetContent  $(p_r)$ $n_{hop},$  direction  $\leftarrow$  GetConnection  $(p_e,p_r,T)$ $\mathcal{R}_e\gets$  GetRelations(e,  $n_{hop},$  direction, K)  $\mathcal{R}_e'\gets$  TopkSimilarRelations(r,  $\mathcal{R}_e',t_r,n_r,\mathcal{K})$ $L_{r_i}'\gets \mathcal{L}_r'$  .GetCandidates  $(p_r)$  if  $L_{r_i}' = \emptyset$  then  $\mathcal{L}_r'$  .SetCandidates  $(p_r,\mathcal{R}_e')$  else  $\mathcal{L}_r'$  .SetCandidates  $(p_r,\mathcal{R}_e'\cap L_{r_i}'')$  end if end for end for   
+ $\mathcal{R}\gets$  FindNonEmptyResults  $(\mathcal{L}_e',\mathcal{L}_r',\mathcal{T},\mathcal{K})$  if  $\mathcal{R}\neq \emptyset$  then return R   
+end if   
+end for   
+return 0
+
+The subquery chain template generation task aims to construct the structural template  $p$  of the final query based on the input question. The template  $p$  is organized as a sequence of subqueries and does not contain any specific semantic information. Instead, all entities and relations are replaced with placeholders such as [ENT_i] and [REL_j].
+
+We design this task to guide the model in learning query structural patterns that are independent of specific entities and relations. Meanwhile, the fine-grained subquery chain helps determine the overall graph structure of the query, which in turn benefits the subsequent validation process. Furthermore, the subquery chain intuitively illustrates the
+
+steps taken by the LLM to generate the final logical form, which can improve interpretability of the generation process of the final logical form.
+
+Task 2:phanholder Specification This task requires the model to determine the specific textual content for each placeholder in the template generated from Task 1, based on the natural language query, e.g. "[ENT_1] = 'Ken Barlow'," "[ENT_2] = 'Coronation Street'". The textual content is then used to retrieve entity IDs or relations from the knowledge base.
+
+We adopt parameter-efficient fine-tuning (PEFT) methods, which includes various methods such as P-Tuning v2 (Liu et al., 2022), LoRA (Hu et al., 2021), and Freeze (Geva et al., 2021), to fine-tune the backbone LLM in our framework. The input and output formats of both tasks can be found in Appendix A. For both tasks, our optimization objective is next token prediction (NSP). The final loss is defined as:
+
+$$
+\mathcal {L} = \mathcal {L} _ {1} + \mathcal {L} _ {2}, \tag {1}
+$$
+
+where  $\mathcal{L}_1$  represents the loss for the first task, and  $\mathcal{L}_2$  represents the loss for the second task.
+
+# 4.3 Retrieval and Structure-Aware Validation
+
+The LLM fine-tuned on Task 1 and Task 2 can exhibit strong semantic parsing capabilities and logical form generation ability.
+
+**Placeholder Candidate List Construction** As shown in Algorithm 1, from the output of the finetuned LLM, we first extract the query structure  $\mathcal{T}$  and the semantic information for each placeholder in the generated template. We utilize FACC1 entity linking method to retrieve entity IDs that match the textual content of entity placeholders.
+
+Entity Filtering We discard entity IDs with occurrence probabilities below the threshold  $t_e$  and obtain a refined candidate entity ID list  $L_{e_i}$  for the entity placeholder  $e_i$ . For relation placeholders, we first assume that the model-generated relations are sufficiently accurate. Therefore, the candidate list  $L_{r_i}$  for the relation placeholder  $r_i$  contains only a single element, which corresponds to the generated textual content of the placeholder. For each candidate entity ID in the candidate list of entity placeholder  $e_i$ , we validate whether its connectivity in the knowledge base maintains structural consistency with the query structure  $\mathcal{T}$ . Only entity IDs that satisfy the structure constrain are retained in
+
+the candidate list. Next, we rank the candidate entity ID list based on their occurrence probabilities in the knowledge base and select the top  $n_e$  entities as the final candidate entity set.
+
+Logical Form Generation and Execution We enumerate all possible assignments of candidate entities and relations from their respective placeholder lists and try to construct valid logical forms. Each logical form is then converted into a SPARQL query using a fixed convert function and executed on the knowledge base. The first non-empty result is returned as the final answer set  $\mathcal{A}$ .
+
+If none of the possible combinations yield a non-empty result set, we attempt to change the candidate lists for the relation placeholders. For each entity ID combination  $c_{e}$ , we take each entity ID in  $c_{e}$  as the central entity node and identify the relation placeholders associated with it and try to change the candidate lists of these placeholders. For each relation placeholder associated with the central entity, we retrieve relations that satisfy the directionality and distance constraints imposed by the query structure  $\mathcal{T}$ , from the knowledge base  $\kappa$ . We employ unsupervised retrieval methods like SimCSE (Gao et al., 2021), Conrtiver (Izacard et al., 2021) and BM25 (Robertson and Zaragoza, 2009), to compute the similarity between the retrieved relations and the model-generated relation. The retrieved relations are ranked based on their similarity to the model-generated relation. Relations with similarity scores below a predefined threshold  $t_{r}$  are discarded. From the remaining candidates, we select the top  $n_{r}$  relations as the updated candidate list for the relation placeholders. Then we try to find the first non-empty results set as our predicted answer with the updated candidate lists.
+
+# 5 Experiments
+
+In this section, we introduce the basic setup of the experiments, including the datasets, baselines, and implementation details. We also provide an analysis of the experimental results.
+
+# 5.1 Setup
+
+Datasets We evaluate the effectiveness of our approach on two widely used KBQA datasets: WebQSP (Yih et al., 2016a) and ComplexWebQuestions (CWQ) (Talmor and Berant, 2018a). Both datasets are constructed based on the Freebase (Talmor and Berant, 2018b) knowledge base and con
+
+sist of a collection of knowledge-driven natural language questions. Each question is paired with a SPARQL query that retrieves the corresponding answer set from Freebase. Table 1 presents the statistics of WebQSP and CWQ.
+
+<table><tr><td>Dataset</td><td># Train</td><td># Validation</td><td># Test</td></tr><tr><td>WebQSP</td><td>3,098</td><td>-</td><td>1,639</td></tr><tr><td>CWQ</td><td>27,639</td><td>3,519</td><td>3,531</td></tr></table>
+
+Table 1: Statistics (Number of QA Pairs) of the Datasets
+
+Baselines We compare our method with several previous approaches. The baseline methods can be categorized into three main groups: 1. Information Retrieval-based (IR-based) Methods, which are approaches that primarily rely on retrieval techniques to fetch relevant knowledge from the knowledge base; 2. Semantic Parsing-based (SP-based) Methods: which are approaches that map natural language questions to structured queries, such as SPARQL, for precise answer retrieval; 3. Large Language Model-based (LLM-based) Methods:, which are methods that leverage LLMs for semantic parsing, reasoning or knowledge retrieval. More details about the baselines can be found in Appendix B.
+
+Metrics We adopt the same evaluation metrics as previous works (Shu et al., 2022; Yu et al., 2023; Luo et al., 2024) to evaluate the overlap between the retrieved answer set and the ground truth answer set. Specifically, we use Accuracy (Acc),  $F_{1}$  score (F1), and Hits@1 as our evaluation metrics.
+
+Implementation Details While our method is generally applicable to relations of arbitrary hop distances and directions, in this paper, we restrict our consideration to relations within two hops. For WebQSP and CWQ, we use LLama2-7B and LLama2-13B (Touvron et al., 2023), respectively, as the backbone LLMs in our framework. We conduct our experiments on two NVIDIA A6000 GPUs with 48GB VRAM each. To reduce VRAM consumption during training, we apply LoRA (Hu et al., 2021). Further implementation details can be found in Appendix C.
+
+# 5.2 Results and Analysis
+
+Table 2 presents our experimental results, averaged over 3 runs. As can be observed, GRV-KBQA outperforms the existing baseline methods
+
+on most evaluation metrics and demonstrates a significant improvement over information retrieval-based and semantic parsing-based methods across all the evaluation metrics. Compared to large
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">WebQSP</td><td colspan="3">CWQ</td></tr><tr><td>F1</td><td>Hits@1</td><td>Acc</td><td>F1</td><td>Hits@1</td><td>Acc</td></tr><tr><td colspan="7">IR-based Methods</td></tr><tr><td>KV-Mem</td><td>34.5</td><td>46.7</td><td>-</td><td>15.7</td><td>21.1</td><td>-</td></tr><tr><td>Pull-Net</td><td>-</td><td>68.7</td><td>-</td><td>47.2</td><td>-</td><td></td></tr><tr><td>Embed-KGQA</td><td>-</td><td>66.6</td><td>-</td><td>-</td><td>44.7</td><td>-</td></tr><tr><td>NSM+h</td><td>67.4</td><td>74.3</td><td>-</td><td>44.0</td><td>48.8</td><td>-</td></tr><tr><td>TransferNet</td><td>-</td><td>71.4</td><td>-</td><td>-</td><td>48.6</td><td>-</td></tr><tr><td>Subgraph Retrieval</td><td>64.1</td><td>69.5</td><td>-</td><td>47.1</td><td>50.2</td><td>-</td></tr><tr><td colspan="7">SP-based Methods</td></tr><tr><td>STAGG</td><td>71.7</td><td>-</td><td>63.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>UHop</td><td>68.5</td><td>-</td><td>-</td><td>29.8</td><td>-</td><td>-</td></tr><tr><td>Topic Units</td><td>67.9</td><td>68.2</td><td>-</td><td>36.5</td><td>39.3</td><td>-</td></tr><tr><td>QGG</td><td>74.0</td><td>73.0</td><td>-</td><td>40.4</td><td>44.1</td><td>-</td></tr><tr><td>UniKGQA</td><td>72.2</td><td>77.2</td><td>-</td><td>49.4</td><td>51.2</td><td>-</td></tr><tr><td>CBR-KBQA</td><td>72.8</td><td>-</td><td>69.9</td><td>70.0</td><td>70.4</td><td>67.1</td></tr><tr><td>RnG-KBQA</td><td>75.6</td><td>-</td><td>71.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Program Transfer</td><td>76.5</td><td>74.6</td><td>-</td><td>58.7</td><td>58.1</td><td>-</td></tr><tr><td>TIARA</td><td>78.9</td><td>75.2</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GMT-KBQA</td><td>76.6</td><td>-</td><td>73.1</td><td>77.0</td><td>-</td><td>72.2</td></tr><tr><td>UnifiedSKG</td><td>73.9</td><td>-</td><td>-</td><td>68.8</td><td>-</td><td>-</td></tr><tr><td>DecAF</td><td>78.8</td><td>82.1</td><td>-</td><td>-</td><td>70.4</td><td>-</td></tr><tr><td>FC-KBQA</td><td>76.9</td><td>-</td><td>-</td><td>56.4</td><td>-</td><td>-</td></tr><tr><td colspan="7">LLM-based Methods</td></tr><tr><td>StructGPT</td><td>72.6</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Pangu</td><td>79.6</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>ToG</td><td>-</td><td>82.6</td><td>-</td><td>-</td><td>69.5</td><td>-</td></tr><tr><td>ChatKBQA</td><td>79.8</td><td>83.2</td><td>73.8</td><td>77.8</td><td>82.7</td><td>73.3</td></tr><tr><td>GRV-KBQA(ours)</td><td>80.5</td><td>83.2</td><td>75.1</td><td>78.8</td><td>82.5</td><td>74.9</td></tr></table>
+
+language model-based approaches, GRV-KBQA also achieves an improvement of approximately  $1\%$  in  $F_{1}$  and  $1.5\%$  in Accuracy. Furthermore, it is noteworthy that the improvement in Accuracy is particularly pronounced which indicates that our method is more effective at precisely identifying the correct answer.
+
+# 6 Ablation Study
+
+In order to assess the contribution of each component in GRV-KBQA, we conduct an ablation study by systematically removing or modifying key steps and observing the impact on performance. Specifically, we design a series of experiments to evaluate the effectiveness of each component. Specifically, we conduct our ablation experiments on the WebQSP dataset.
+
+# 6.1 Effectiveness of Task 1
+
+The Subquery Chain Template Generation Task (Task 1) structures the KBQA pipeline by providing an intermediate logical form before entity and relation specification. To assess its impact, we bypass this step and allow the LLM to directly generate a complete logical form template without subqueries, as in  $w / o$ $T1$  in Table 3.
+
+Table 2: Comparison of Our Method with the Baseline Methods. Most of the results are directly taken from the corresponding original papers. The best result for each evaluation metric is highlighted in bold.  
+
+<table><tr><td>Settings</td><td>F1</td><td>Hits@1</td><td>Acc</td></tr><tr><td>GRV-KBQA</td><td>80.5</td><td>83.2</td><td>75.1</td></tr><tr><td>w/o T1</td><td>80.2</td><td>83.0</td><td>74.7</td></tr><tr><td>w/o T2</td><td>78.4</td><td>82.6</td><td>70.5</td></tr><tr><td>w/o Validation</td><td>79.5</td><td>82.4</td><td>73.7</td></tr></table>
+
+Table 3: Ablation Study Results on the WebQSP Dataset
+
+Removing Task 1 leads to a drop in accuracy  $(75.1\% \rightarrow 74.7\%)$  and  $F_{1}$  score  $(0.805 \rightarrow 0.802)$ . The relatively minor impact suggests that the LLM can still infer the logical structure from the input question to some extent. However, Task 1 provides an explicit and modular breakdown of the reasoning process, which benefits interpretability and allows for easier analysis of error patterns in our framework. By decomposing complex queries into a structured form, Task 1 facilitates future refinements and debugging, making the overall KBQA pipeline more maintainable.
+
+# 6.2 Effectiveness of Task 2
+
+The Placeholder Specification task (Task 2) enables the LLM to determine the correct entities and relations for the placeholders in the structured template generated by Task 1. Instead of following this approach,  $w / o$  T2 modifies Task 1 to generate a subquery chain that already contains specific entity and relation information, which means that the LLM must directly infer KB elements during subquery chain template generation.
+
+As can be observed in Table 3, removing Task 2 results in a significant drop in accuracy (75.1% → 70.5%) and  $F_{1}$  score (0.805 → 0.784). Without explicit placeholder resolution, the LLM must simultaneously generate the subquery chain and extract the relevant semantic information from the question, rather than resolving entity and relation placeholders in a separate step. This may lead to higher entity and relation misalignment rates. The substantial accuracy drop (4.6%) demonstrates that decoupling structure generation from semantic
+
+grounding significantly improves retrieval correctness, thus leading to more precise answers.
+
+# 6.3 Effectiveness of Structure-Aware Validation
+
+The Structure-Aware Validation mechanism ensures that retrieved entities and relations comply with structural constraints derived from the generated template. In w/o Validation, this step is omitted, meaning all possible combinations are executed without structural checks.
+
+As shown in Table 3, removing Structure-Aware Validation leads to a drop in accuracy from  $75.1\%$  to  $73.7\%$  and a  $1\%$  decrease in  $F_{1}$  score. This suggests that precision declines due to an increased number of incorrect logical forms. Without structural validation, the framework is more likely to attempt to execute structurally inconsistent logical forms, which in turn results in more execution failures during query processing. This highlights the crucial role of validation in filtering structurally incompatible candidates, thus enhancing the reliability of the final logical forms.
+
+To further illustrate the impact of Structure-Aware Validation, Table 4 presents a comparative analysis of two key factors: the average number of failed execution attempts and the exact match rate (EMR) between the executed logical forms and the ground-truth logical forms. The results show that removing validation significantly increases execution failures due to structurally inconsistent queries.
+
+<table><tr><td>Settings</td><td>#Failed Execution (Average)</td><td>EMR</td></tr><tr><td>GRV-KBQA</td><td>3390119.26</td><td>66.99%</td></tr><tr><td>w/o Validation</td><td>1802539361.46</td><td>62.66%</td></tr></table>
+
+To verify the effectiveness of the Structure-Aware Validation step, we analyze the filtering rate of the candidate list for each placeholder during question processing and report the average across all questions. The filtering rates on the WebQSP and CWQ datasets are  $90.0\%$  and  $83.8\%$ , respectively. Additionally, the logical form match rate drops, which indicates that a structure-aware validation step can ensure greater alignment with the expected form. These findings confirm that enforcing structural constraints not only prevents invalid logical forms from being executed but also improves the overall accuracy of query generation.
+
+Table 4: Impact of Structure-Aware Validation on Execution Failures and Logical Form Accuracy.  
+
+<table><tr><td>Settings</td><td>F1</td><td>Hits@1</td><td>Acc</td></tr><tr><td>GRV-KBQA</td><td>80.5</td><td>83.2</td><td>75.1</td></tr><tr><td>ChatGPT(1-shot)</td><td>13.8</td><td>16.5</td><td>11.4</td></tr></table>
+
+# 6.4 Do We Have to Fine-Tune an LLM?
+
+LLMs have demonstrated impressive zero-shot and few-shot generation abilities, which also raises the question of whether fine-tuning is truly necessary for our framework.
+
+To investigate this, we compare GRVKBQA with a variant that replaces the fine-tuned LLM with an off-the-shelf LLM API (gpt-4o-2024-11-20). Instead of fine-tuning the backbone LLM on Task 1 and Task 2, this variant performs one-shot generation, where ChatGPT is prompted to directly produce the subquery chain template and the semantic information for each placeholder. By following the same validation step as our framework, this setup allows us to assess whether a non-fine-tuned LLM, guided only by one-shot prompting, can achieve comparable performance.
+
+As the results shown in Table 5, the performance gap between GRV-KBQA (fine-tuned LLM) and ChatGPT(1-shot) is substantial, which indicates the necessity of fine-tuning for our framework. The drastic performance drop suggests that one-shot generation is highly unreliable for our framework. Without explicit fine-tuning, ChatGPT struggles to generate well-formed subquery chains, which leads to a high failure rate in execution. These results demonstrate that fine-tuning plays a crucial role in our framework. While LLMs exhibit impressive generation abilities, direct prompting alone is insufficient for generating accurate, executable logical forms.
+
+# 6.5 Exact Match and Efficiency Comparison with Baseline
+
+To further demonstrate the advantages of our approach over ChatKBQA, we additionally compare the two methods in terms of exact match rate.
+
+Table 5: Comparison Between GRV-KBQA and ChatGPT API  
+
+<table><tr><td>Method</td><td>WebQSP</td><td>CWQ</td></tr><tr><td>ChatKBQA</td><td>59.40%</td><td>46.6%</td></tr><tr><td>GRV-KBQA</td><td>66.99%</td><td>60.0%</td></tr></table>
+
+Table 6: Comparison with ChatKBQA on EMR
+
+As shown in Table 6, our GRV-KBQA framework significantly outperforms ChatKBQA in terms of exact match rate (EMR) on both WebQSP and CWQ. Specifically, GRV-KBQA achieves an improvement of  $7.59\%$  points on WebQSP and  $13.4\%$  points on CWQ.
+
+These results further demonstrate that GRVKBQA can effectively increase the likelihood of generating executable logical forms that match the gold-standard answers. By separating query structure modeling from semantic grounding and verifying structural consistency before execution, GRVKBQA reduces errors in both generation and alignment, leading to more reliable query execution.
+
+We further compare the efficiency of GRVKBQA and ChatKBQA on the WebQSP dataset in terms of average processing time per question. Our method achieves an average speed of 0.9 minutes per question, compared to approximately 1.06 minutes for ChatKBQA. This further highlights the efficiency advantage of our approach.
+
+# 7 Conclusion
+
+In this paper, we propose GRV-KBQA, a three-stage framework for knowledge base question answering (KBQA) that decouples logical structure generation from semantic grounding while incorporating structure-aware validation. Existing KBQA approaches, struggle with errors in logical forms, incorrect entity and relation alignment, or inefficiencies in candidate selection, leading to execution failures. Our framework addresses these challenges by explicitly modeling the logical form through subquery chain generation and placeholder specification, and ensuring logical form validity through structural constraints.
+
+Experimental results on WebQSP and CWQ show that GRV-KBQA significantly improves performance over existing approaches, which demonstrates its effectiveness in generating more accurate and executable logical forms. The ablation study conducted further validates that decoupling logical structure generation from semantic grounding improves entity alignment, while structure-aware validation enhances execution reliability.
+
+# Limitations
+
+While GRV-KBQA improves KBQA performance by decoupling logical structure generation from semantic grounding and enforcing structure-aware validation, it still has several limitations.
+
+First, our approach relies on fine-tuning an LLM for subquery chain generation and placeholder specification. While fine-tuning enhances accuracy, it requires substantial computational resources and may limit generalization to unseen queries.
+
+Second, the structure-aware validation stage introduces computational overhead, as verifying the structural consistency of candidate entities requires multiple KB queries. This can be inefficient on large-scale KBs, potentially limiting real-world deployment.
+
+Finally, as illustrated in Appendix D, our method propagates errors across multiple stages, meaning that mistakes in subquery chain generation or placeholder resolution can negatively affect subsequent validation and ultimately lead to query failures.
+
+# Ethics Statement
+
+In this work, we use publicly available datasets and do not collect any personally identifiable information. All datasets and models are utilized in full compliance with their intended purposes and respective licenses. The primary goal of this work is to address the two key issues observed in recent large language model-based KBQA approaches. We condemn any potential misuse.
+
+# Acknowledgments
+
+This work was supported by the National Key Research and Development Program of China (Grant No. 2022YFC3302100) and the National Natural Science Foundation of China (Grant No. 62476025).
+
+# References
+
+Sören Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, and Zachary G. Ives. 2007. Dbpedia: A nucleus for a web of open data. semantic web.  
+Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIGMOD International Conference on Management of Data, SIGMOD '08, page 1247-1250, New York, NY, USA. Association for Computing Machinery.  
+Shulin Cao, Jiaxin Shi, Zijun Yao, Xin Lv, Jifan Yu, Lei Hou, Juanzi Li, Zhiyuan Liu, and Jinghui Xiao. 2022. Program transfer for answering complex questions over knowledge bases. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8128-8140, Dublin, Ireland. Association for Computational Linguistics.
+
+Zi-Yuan Chen, Chih-Hung Chang, Yi-Pei Chen, Jijnasa Nayak, and Lun-Wei Ku. 2019. UHop: An unrestricted-hop relation extraction framework for knowledge-based question answering. In Proceedings of the 2019 Conference of the North American Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 345-356, Minneapolis, Minnesota. Association for Computational Linguistics.  
+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, pages 9594-9611, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Guanting Dong, Rumei Li, Sirui Wang, Yupeng Zhang, Yunsen Xian, and Weiran Xu. 2023. Bridging the kb-text gap: Leveraging structured knowledge-aware pre-training for kbqa. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM '23, page 3854-3859, New York, NY, USA. Association for Computing Machinery.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6894-6910, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484-5495, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yu Gu, Xiang Deng, and Yu Su. 2023. Don't generate, discriminate: A proposal for grounding language models to real-world environments. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4928-4949, Toronto, Canada. Association for Computational Linguistics.  
+Gaole He, Yunshi Lan, Jing Jiang, Wayne Xin Zhao, and Ji-Rong Wen. 2021a. Improving multi-hop knowledge base question answering by learning intermediate supervision signals. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining, WSDM '21, page 553-561, New York, NY, USA. Association for Computing Machinery.  
+Gaole He, Yunshi Lan, Jing Jiang, Wayne Xin Zhao, and Ji-Rong Wen. 2021b. Improving multi-hop knowledge base question answering by learning intermediate supervision signals. In Proceedings of the 14th
+
+ACM international conference on web search and data mining, pages 553-561.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. Preprint, arXiv:2106.09685.  
+Xixin Hu, Xuan Wu, Yiheng Shu, and Yuzhong Qu. 2022. Logical form generation via multi-task learning for complex question answering over knowledge bases. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1687-1696, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Xiang Huang, Sitao Cheng, Shanshan Huang, Jiayu Shen, Yong Xu, Chaoyun Zhang, and Yuzhong Qu. 2024. Queryagent: A reliable and efficient reasoning framework with environmental feedback based self-correction. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 5014-5035. Association for Computational Linguistics.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Towards unsupervised dense information retrieval with contrastive learning. CoRR, abs/2112.09118.  
+Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Wayne Xin Zhao, and Ji-Rong Wen. 2023a. Struct-gpt: A general framework for large language model to reason over structured data. Preprint, arXiv:2305.09645.  
+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.  
+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, pages 969-974, Online. Association for Computational Linguistics.  
+Yunshi Lan, Shuohang Wang, and Jing Jiang. 2019. Knowledge base question answering with topic units. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 5046-5052. International Joint Conferences on Artificial Intelligence Organization.  
+Tianle Li, Xueguang Ma, Alex Zhuang, Yu Gu, Yu Su, and Wenhu Chen. 2023. Few-shot in-context learning for knowledge base question answering. Preprint, arXiv:2305.01750.  
+Yading Li, Dandan Song, Changzhi Zhou, Yuhang Tian, Hao Wang, Ziyi Yang, and Shuhao Zhang. 2024. A framework of knowledge graph-enhanced large
+
+language model based on question decomposition and atomic retrieval. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11472-11485, Miami, Florida, USA. Association for Computational Linguistics.  
+Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, Dublin, Ireland. Association for Computational Linguistics.  
+Haoran Luo, Haihong E, Zichen Tang, Shiyao Peng, Yikai Guo, Wentai Zhang, Chenghao Ma, Guanting Dong, Meina Song, Wei Lin, Yifan Zhu, and Anh Tuan Luu. 2024. ChatKBQA: A generate-then-retrieve framework for knowledge base question answering with fine-tuned large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pages 2039–2056, Bangkok, Thailand. Association for Computational Linguistics.  
+Alexander Miller, Adam Fisch, Jesse Dodge, Amir Hossein Karimi, Antoine Bordes, and Jason Weston. 2016. Key-value memory networks for directly reading documents. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1400-1409, Austin, Texas. Association for Computational Linguistics.  
+Zhijie Nie, Richong Zhang, Zhongyuan Wang, and Xudong Liu. 2024. Code-style in-context learning for knowledge-based question answering. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pages 18833-18841. AAAI Press.  
+Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: Bm25 and beyond. Found. Trends Inf. Retr., 3(4):333-389.  
+Apoory Saxena, Aditay Tripathi, and Partha Talukdar. 2020. Improving multi-hop question answering over knowledge graphs using knowledge base embeddings. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4498-4507, Online. 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, pages 4149-4158, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
+
+Yiheng Shu, Zhiwei Yu, Yuhan Li, Borje Karlsson, Tingting Ma, Yuzhong Qu, and Chin-Yew Lin. 2022. TIARA: Multi-grained retrieval for robust question answering over large knowledge base. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8108-8121, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Haitian Sun, Tania Bedrax-Weiss, and William Cohen. 2019. PullNet: Open domain question answering with iterative retrieval on knowledge bases and text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2380-2390, Hong Kong, China. Association for Computational Linguistics.  
+Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel 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.  
+Alon Talmor and Jonathan Berant. 2018a. 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.  
+Alon Talmor and Jonathan Berant. 2018b. 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.  
+Yuhang Tian, Dandan Song, Zhijing Wu, Changzhi Zhou, Hao Wang, Jun Yang, Jing Xu, Ruanmin Cao, and HaoYu Wang. 2024. Augmenting reasoning capabilities of LLMs with graph structures in knowledge base question answering. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11967-11977, Miami, Florida, USA. Association for Computational Linguistics.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura,
+
+Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models. Preprint, arXiv:2307.09288.
+
+Denny Vrandecic and Markus Krötzsch. 2014. Wiki-data: a free collaborative knowledgebase. Commun. ACM, 57(10):78-85.
+
+Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I. Wang, Victor Zhong, Bailin Wang, Chengzu Li, Connor Boyle, Ansong Ni, Ziyu Yao, Dragomir Radev, Caiming Xiong, Lingpeng Kong, Rui Zhang, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2022. UnifiedSKG: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 602-631, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+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), pages 6032-6043, Dublin, Ireland. 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, Chris Meek, Ming-Wei Chang, and Jina Suh. 2016a. The value of semantic parse labeling for knowledge base question answering. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 201–206, Berlin, Germany. Association for Computational Linguistics.
+
+Wen-tau Yih, Matthew Richardson, Chris Meek, Ming-Wei Chang, and Jina Suh. 2016b. The value of semantic parse labeling for knowledge base question
+
+answering. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 201-206, Berlin, Germany. Association for Computational 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.
+
+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), pages 5773-5784, Dublin, Ireland. Association for Computational Linguistics.
+
+Lingxi Zhang, Jing Zhang, Yanling Wang, Shulin Cao, Xinmei Huang, Cuiping Li, Hong Chen, and Juanzi Li. 2023. FC-KBQA: A fine-to-coarse composition framework for knowledge base question answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1002-1017, Toronto, Canada. Association for Computational Linguistics.
+
+# A Input and Output Format
+
+Table 8 and Table 9 present the input and output formats used for fine-tuning on the two tasks.
+
+# B Baselines
+
+In this section, we introduce the baselines used in our experiments.
+
+# B.1 IR-Based Methods
+
+KV-Mem (Miller et al., 2016) introduces a key-value memory network that facilitates fact encoding and reasoning for knowledge-based question answering. It enhances the model's ability to integrate relevant information and infer correct answers.
+
+PullNet (Sun et al., 2019) integrates retrieval and reasoning in a unified framework for multi-hop KBQA, iteratively constructing a question-specific subgraph from both text corpus and knowledge bases. By leveraging graph convolutional networks (GCNs) for node expansion and answer extraction, it enables effective reasoning over large, incomplete KBs.
+
+EmbedKGQA (Saxena et al., 2020) leverages knowledge graph embeddings to address multi-hop
+
+<table><tr><td></td><td colspan="2">WebQSP</td><td>CWQ</td></tr><tr><td></td><td colspan="2">Hyper-parameters for LLM Fine-Tuning</td><td></td></tr><tr><td>Base Model</td><td>LLama2-7B, LLama3-8B, QWen2.5-7B</td><td>LLama2-13B, LLama3-11B, QWen2.5-14B</td><td></td></tr><tr><td>Batch Size</td><td colspan="2">16</td><td>32</td></tr><tr><td>Learning Rate</td><td colspan="2">1e-4, 2.5e-4, 5e-4</td><td>2.5e-4, 5e-4, 1e-3</td></tr><tr><td>Epoch</td><td colspan="2">80, 100, 120</td><td>8, 10, 12</td></tr><tr><td>PEFT Method</td><td colspan="2">LoRA, P-Tuningv2, Freeze</td><td>LoRA, P-Tuningv2, Freeze</td></tr><tr><td>LoRA Rank</td><td colspan="2">8</td><td>8</td></tr><tr><td>LoRA Alpha</td><td colspan="2">8</td><td>8</td></tr><tr><td>LoRA Dropout</td><td colspan="2">0</td><td>0</td></tr><tr><td></td><td colspan="2">Hyper-parameters for Inference</td><td></td></tr><tr><td>Beam Size</td><td colspan="2">15</td><td>8</td></tr><tr><td>Inference Batch Size</td><td colspan="2">1</td><td>1</td></tr><tr><td></td><td colspan="2">Hyper-parameters for Answer Retrieval</td><td></td></tr><tr><td>ne</td><td colspan="2">50</td><td>50</td></tr><tr><td>nr</td><td colspan="2">5</td><td>5</td></tr><tr><td>te</td><td colspan="2">1e-6</td><td>1e-5</td></tr><tr><td>tr</td><td colspan="2">0</td><td>0</td></tr><tr><td>Relation Retrieval</td><td colspan="2">SimCSE</td><td>SimCSE</td></tr></table>
+
+Table 7: Implementation Details of GRV-KBQA  
+
+<table><tr><td>Input</td><td>Question: Who is keyshia cole dad ?</td></tr><tr><td rowspan="4">Output</td><td>1. result_1 = JOIN [REL_1] [ENT_1]</td></tr><tr><td>2. relation_1 = R [REL_2]</td></tr><tr><td>3. result_2 = JOIN relation_1 [ENT_2]</td></tr><tr><td>4. [ANS] = AND result_1 result_2</td></tr></table>
+
+Table 8: An Example of the Input and Output for Task 1  
+
+<table><tr><td rowspan="5">Input</td><td>Question: Who is keyshia cole dad ?</td></tr><tr><td>1. result_1 = JOIN [REL_1] [ENT_1]</td></tr><tr><td>2. relation_1 = R [REL_2]</td></tr><tr><td>3. result_2 = JOIN relation_1 [ENT_2]</td></tr><tr><td>4. [ANS] = AND result_1 result_2</td></tr><tr><td rowspan="4">Output</td><td>[ENT_1] = &#x27;Male&#x27;</td></tr><tr><td>[REL_1] = &#x27;people.person.person&#x27;</td></tr><tr><td>[REL_2] = &#x27;people.person.parents&#x27;</td></tr><tr><td>[ENT_2] = &#x27;Keyshia Cole&#x27;</td></tr></table>
+
+Table 9: An Example of the Input and Output for Task 2
+
+reasoning in KBQA, effectively handling knowledge graph sparsity without relying on external textual information. It also removes the constraint of selecting answers from a fixed local neighborhood, improving answer retrieval in incomplete KGs.
+
+$\mathbf{NSM}_{+h}$  (He et al., 2021a) employs a teacher-student framework for multi-hop KBQA, where a teacher network provides intermediate reasoning supervision to stabilize learning. By leveraging bidirectional reasoning, it enhances the reliability of intermediate entity distributions, effectively mitigating spurious reasoning issues.
+
+TransferNet (Shi et al., 2021) introduces a unified framework for multi-hop KBQA that supports
+
+both structured (knowledge graph) and unstructured (text corpus) relations. By dynamically attending to different parts of the question and transferring entity scores through activated relations, it achieves interpretable multi-step reasoning with improved optimization stability.
+
+Subgraph Retrieval (Zhang et al., 2022) introduces a trainable subgraph retriever (SR) that is decoupled from the reasoning process, mitigating reasoning bias caused by partial subgraphs. This approach enhances any subgraph-based KBQA model in a plug-and-play manner and improves retrieval effectiveness through weakly supervised pre-training and end-to-end fine-tuning.
+
+# B.2 SP-Based Methods
+
+STAGG (Yih et al., 2016b) applies supervised semantic parsing to KBQA, demonstrating that learning from structured representations yields higher accuracy than relying solely on answer annotations.
+
+UHop (Chen et al., 2019) introduces an unrestricted-hop framework for KBQA that eliminates fixed hop constraints by employing a transition-based search mechanism instead of exhaustive relation-chain search. This approach enables dynamic halting, improves efficiency, and significantly enhances performance on long relation paths.
+
+Topic Units (Lan et al., 2019) introduces topic unit linking, expanding beyond traditional entity linking to improve KBQA answer retrieval. It employs a generation-and-scoring approach to refine topic unit selection and uses reinforcement learn
+
+ing for joint optimization of topic unit linking and answer ranking.
+
+CBR-KBQA (Das et al., 2021) introduces a neuro-symbolic case-based reasoning framework for KBQA, combining a nonparametric memory to store question-logical form pairs with a retrieval-based parametric model that generates logical forms for new queries. It enables few-shot adaptability, allowing the model to handle unseen KB entities and relations without further training.
+
+QGG (Lan and Jiang, 2020) enhances query graph generation by incorporating both constraints and multi-hop reasoning in a unified framework. By integrating constraints early in the process, it effectively prunes the search space, improving efficiency in complex KBQA tasks.
+
+RnG-KBQA (Ye et al., 2022) introduces a Rank-and-Generate framework that combines contrastive ranking and generation to improve KBQA generalization. It first ranks candidate logical forms retrieved from the knowledge graph, then refines them with a tailored generation model, effectively addressing the coverage issue while maintaining strong generalization capability.
+
+Program Transfer (Cao et al., 2022) enhances program induction for KBQA by leveraging annotated programs from resource-rich KBs to supervise learning on low-resource KBs. It introduces a two-stage parsing framework, combining sketch parsing for high-level program structure generation and argument parsing for retrieving KB arguments, with ontology-guided pruning to improve search efficiency.
+
+TIARA (Shu et al., 2022) enhances PLM-based KBQA by incorporating multi-grained retrieval to provide relevant KB context and constrained decoding to improve logical form generation. This approach improves both coverage and generalization, effectively addressing semantic understanding and execution correctness challenges.
+
+GMT-KBQA (Hu et al., 2022) enhances generation-based KBQA by integrating multi-task learning to jointly optimize entity disambiguation, relation classification, and logical form generation. It first retrieves candidate entities and relations via dense retrieval, then leverages auxiliary tasks to improve generalization and reduce errors from noisy auxiliary information.
+
+UnifiedSKG (Xie et al., 2022) unifies structured knowledge grounding (SKG) tasks into a text-to-text format, enabling systematic research across diverse tasks, including KBQA. It benchmarks T5
+
+with multi-task prefix-tuning, demonstrating SOTA performance on most SKG tasks and facilitating investigations into zero-shot and few-shot learning
+
+UniKGQA (Jiang et al., 2023b) unifies retrieval and reasoning for multi-hop KGQA by integrating them at both the model architecture and parameter learning levels. It employs a semantic matching module for question-relation alignment and a matching information propagation module to guide reasoning along KG edges, ensuring a more cohesive retrieval-reasoning process.
+
+DecAF (Yu et al., 2023) combines logical form generation and direct answer prediction to enhance KBQA accuracy while mitigating execution failures. Unlike most prior methods, it relies on freetext retrieval instead of entity linking tools, improving adaptability across datasets.
+
+FC-KBQA (Zhang et al., 2023) introduces a Fine-to-Coarse Composition framework that enhances both generalization and executability in KBQA. By extracting fine-grained knowledge components from KBs and reformulating them into middle-grained knowledge pairs, it enables more robust logical form generation.
+
+# B.3 LLM-Based Methods
+
+StructGPT (Jiang et al., 2023a) improves LLMs' capability in structured data reasoning by employing an Iterative Reading-then-Reasoning (IRR) approach. It integrates specialized interfaces for structured data interaction, adopts a retrieval-augmented reasoning mechanism, and utilizes an iterative process to enhance structured data comprehension and complex question answering.
+
+Pangu (Gu et al., 2023) introduces a grounded language understanding framework that integrates a symbolic agent with a neural language model. This approach enables incremental construction of valid plans while leveraging the language model to assess the plausibility of generated plans, improving structured reasoning and decision-making.
+
+ToG (Think-on-Graph) (Sun et al., 2024) integrates large language models (LLMs) with knowledge graphs (KGs), treating the LLM as an agent that interactively explores entities and relations to enhance reasoning. It employs an iterative beam search mechanism to discover optimal reasoning paths, enabling deep reasoning, knowledge traceability, and training-free adaptability across various LLMs and KGs.
+
+ChatKBQA (Luo et al., 2024) introduces a generate-then-retrieve KBQA framework, first gen
+
+<table><tr><td>Question</td><td>Who will play Mr. Gray in the film?</td></tr><tr><td>Template Generation</td><td>1. relation_1 = R [REL_1]2. result_1 = JOIN relation_1 [ENT_1]3. relation_2 = R [REL_2]4. [ANS] = JOIN relation_2 result_1</td></tr><tr><td>Placeholder Specification</td><td>[REL_1] = &#x27;film.film_character. portrayed_in_films&#x27;[ENT_1] = &#x27;Mr. Gray&#x27;[REL_2] = &#x27;film-performance actor&#x27;</td></tr><tr><td>Golden Logical Form</td><td>(JOIN (R film-performanceActor) (JOIN (R film.film Character. portrayed_in_films) m.0nfjvrm))</td></tr><tr><td>Executed Logical Form</td><td>(JOIN (R film-performanceActor) (JOIN (R film.film Character. portrayed_in.films) m.0v4dfpr))</td></tr></table>
+
+Table 10: A Representative Case for Error Analysis
+
+erating logical forms using fine-tuned LLMs, then refining them via unsupervised entity and relation retrieval. This approach enhances retrieval efficiency, mitigates error propagation, and simplifies KBQA pipelines, achieving state-of-the-art performance on WebQSP and CWQ.
+
+# C Implementation Details
+
+Table 7 presents the implementation details of our experiments.
+
+# D Error Analysis
+
+Although our method achieves superior performance compared to existing baselines, we observe that it still struggles on certain types of questions. In this section, we present a representative example to analyze the reasons behind the failure and discuss potential directions for future improvements.
+
+Table 10 presents this representative case. In this example, the subquery chain generated by our method is structurally correct. Although the structural reasoning is logically sound, the query ultimately returns an incorrect answer due to semantic grounding failure. We find that the issue lies in the generation of the entity placeholder [ENT_1]. The model generates it as "Mr. Gray", whereas the correct entity in the knowledge base is "Christian Grey" (entity ID: m.ønfjvrm).
+
+This issue may stem from discrepancies between the parametric knowledge encoded in the large language model and the factual knowledge stored in the knowledge base. Our method relies on the model's parametric knowledge to implicitly complete contextual information from the input question and to guide the construction of logical forms. Given the vast pretraining corpora of large language models, they are often able to correctly infer background context and relevant entities, enabling accurate semantic interpretation in most cases.
+
+However, in some instances, the model's internal knowledge may interfere with the semantic grounding process, which leads to inconsistencies between the generated logical form and the actual structure of the knowledge base. Although our proposed structure-aware validation mechanism can filter out some invalid entity-relation combinations from a structural perspective, it cannot fully address cases where the input query is inherently ambiguous and the placeholder generation process lacks direct alignment with the KB. In such cases, the textual content of placeholders may remain semantically vague, and the validation mechanism is limited in its ability to correct these types of errors.

2025/GRV-KBQA_ A Three-Stage Framework for Knowledge Base Question Answering with Decoupled Logical Structure, Semantic Grounding and Structure-Aware Validation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f99a2ee248b1e1d7ed2e86a6d0e3177373170c64ddd4d24c935c559246e7573
+size 576621

2025/GTA_ Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models/5317a317-b31f-4425-8151-f38a993fa23c_content_list.json

@@ -0,0 +1,1570 @@
+[
+    {
+        "type": "text",
+        "text": "GTA: Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models",
+        "text_level": 1,
+        "bbox": [
+            122,
+            89,
+            875,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Min Zeng, Jingfei Sun, Xueyou Luo, Caiquan Liu, Shiqi Zhang, Li Xie, Xiaoxin Chen",
+        "bbox": [
+            132,
+            156,
+            868,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "vivo AI Lab",
+        "bbox": [
+            448,
+            175,
+            549,
+            189
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "zengmin.ai@vivo.com",
+        "bbox": [
+            428,
+            193,
+            568,
+            206
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In natural language processing tasks, pure reinforcement learning (RL) fine-tuning methods often suffer from inefficient exploration and slow convergence; while supervised fine-tuning (SFT) methods, although efficient in training, have limited performance ceiling and less solid theoretical foundation compared to RL. To address efficiency-capability trade-off, we propose the Guess-Think-Answer (GTA) framework that combines the efficiency of SFT with the capability gains of RL in a unified training paradigm. GTA works by having the model first produce a provisional guess (optimized via cross-entropy loss), then reflect on this guess before generating the final answer, with RL rewards shaping both the final output and the format of the entire GTA structure. This hybrid approach achieves both faster convergence than pure RL and higher performance ceiling than pure SFT. To mitigate gradient conflicts between the two training signals, we employ loss masking and gradient constraints. Empirical results on four text classification benchmarks demonstrate that GTA substantially accelerates convergence while outperforming both standalone SFT and RL baselines.",
+        "bbox": [
+            144,
+            288,
+            460,
+            657
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            670,
+            258,
+            686
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Text classification, as a foundational task in natural language processing (NLP), has been widely employed for sentiment analysis (Pang et al., 2002), intent recognition (Chen et al., 2016), and news categorization (Johnson and Zhang, 2015). Early NLP solutions primarily relied on rule-based systems and statistical models—including hidden Markov models (HMMs) and support vector machines (SVMs)—to learn patterns from annotated corpora (Joachims, 1998). The emergence of deep learning and Transformer (Vaswani et al., 2017) architectures dramatically enhanced classification performance, with Bidirectional Encoder Representations from Transformers (BERT)'s (Devlin,",
+        "bbox": [
+            112,
+            696,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2018) bidirectional pre-training paradigm capturing rich contextual representations and achieving breakthroughs across multiple tasks.",
+        "bbox": [
+            507,
+            261,
+            884,
+            309
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs)—such as GPT (Achiam et al., 2023), Llama (Grattafori et al., 2024), Qwen (Yang et al., 2024), and DeepSeek (Guo et al., 2025)—have demonstrated remarkable capabilities across numerous NLP tasks, including text classification (Kostina et al., 2025). While SFT has been the predominant approach to adapt these models for specific tasks, it faces inherent limitations: SFT methods directly learn to produce correct answers without explicit reasoning, leading to limited generalization capabilities and performance ceilings. Chain-of-thought (CoT) prompting (Wei et al., 2022)—a technique that guides models to generate intermediate reasoning steps before producing final answers—has shown significant improvements across various reasoning tasks (Kojima et al., 2022; Xia et al., 2025). However, applying CoT within the SFT paradigm requires extensive human annotation of reasoning chains, resulting in substantial costs and susceptibility to annotator biases and quality inconsistencies (Tan et al., 2024; Byun et al., 2024).",
+        "bbox": [
+            507,
+            309,
+            884,
+            663
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "RL offers a promising alternative that can theoretically overcome these limitations by combining the benefits of CoT reasoning with optimization-based learning (Xu et al., 2025; Wang et al., 2024). From reinforcement learning from human feedback (RLHF) (Ouyang et al., 2022) to advanced frameworks like group relative policy optimization (GRPO) (Shao et al., 2024), RL techniques can explore and optimize intermediate reasoning processes without requiring manually annotated reasoning chains. This approach holds particular promise for enhancing model performance beyond what SFT can achieve. However, the application of RL to text classification tasks remains challenging due to fundamental limitations: unlike supervised learning's direct approach, RL methods",
+        "bbox": [
+            507,
+            664,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1050",
+        "bbox": [
+            482,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 1050-1060",
+        "bbox": [
+            216,
+            945,
+            779,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            280,
+            958,
+            715,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "must discover optimal reasoning strategies through self-guided exploration—a process often hampered by inefficient exploration, slow convergence, and potential training instability (Chen et al., 2025a). These efficiency challenges have hindered pure RL-based methods from consistently outperforming SFT approaches despite their stronger theoretical foundation.",
+        "bbox": [
+            112,
+            84,
+            487,
+            211
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges, this work introduces a novel Guess-Think-Answer (GTA) framework that seamlessly integrates the advantages of SFT and RL within a unified single-stage training process. In our approach, the model first generates an intuitive guessed answer, then engages in a \"think\" step—reasoning explicitly over the guessed answer and the input question—before producing a refined final answer. Our main contributions can be summarized as follows:",
+        "bbox": [
+            115,
+            212,
+            489,
+            372
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We propose a novel GTA framework that structures the reasoning process into three distinct phases: an initial intuitive guess, an explicit reasoning step that reflects on this preliminary prediction, and a refined final answer that incorporates this reasoning.",
+            "2. We develop a unified training approach that seamlessly integrates SFT and RL within a single-stage process. Our method applies cross-entropy loss to the guessed answer while optimizing the reasoning process and final answer through RL-based rewards. To ensure effective cooperation between these learning paradigms, we introduce a specialized loss masking strategy and gradient cosine adjustment technique that mitigates potential gradient conflicts.",
+            "3. Our framework eliminates the need for manual annotation of reasoning chains by enabling the model to spontaneously learn effective reasoning patterns through reinforcement. Experimental results demonstrate that GTA significantly outperforms both standard SFT baselines and state-of-the-art RL methods across multiple text classification benchmarks."
+        ],
+        "bbox": [
+            127,
+            382,
+            489,
+            804
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            815,
+            270,
+            831
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "CoT prompting has emerged as a powerful technique to enhance the reasoning capabilities of LLMs by guiding them to generate intermediate reasoning steps before producing final answers (Wei et al., 2022; Kojima et al., 2022; Xia et al.,",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2025). While CoT significantly improves performance across various reasoning tasks, its application in text classification often requires extensive human annotation of reasoning chains, leading to substantial costs and quality inconsistencies (Tan et al., 2024; Byun et al., 2024). To address these limitations, researchers have explored RL approaches that can optimize model behavior without requiring manual annotation of intermediate steps (Xu et al., 2025; Wang et al., 2024).",
+        "bbox": [
+            507,
+            84,
+            884,
+            243
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional RL methods in NLP, however, often suffer from inefficient exploration and slow convergence, making it challenging for pure RL-based methods to consistently outperform SFT approaches despite their stronger theoretical foundation (Chen et al., 2025a). Recent advancements like the GRPO algorithm (Shao et al., 2024) have improved RL efficiency by estimating baselines from group scores, thereby reducing computational costs. Additionally, when combining multiple learning objectives—such as SFT and RL—gradient conflicts can arise, leading to suboptimal convergence. Techniques such as gradient masking and cosine similarity adjustments have been developed to mitigate these conflicts by aligning gradients toward compatible directions (Yu et al., 2020; Chen et al., 2018), which inspires our approach to harmonizing supervised and RL signals within our proposed GTA framework.",
+        "bbox": [
+            507,
+            244,
+            884,
+            549
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            507,
+            563,
+            658,
+            579
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present the proposed GTA framework in detail. We begin with an overview of the overall design, followed by the description of each component and training objective, highlighting how the method effectively integrates reinforcement learning with reasoning-oriented supervision.",
+        "bbox": [
+            507,
+            588,
+            884,
+            684
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Prompt Design",
+        "text_level": 1,
+        "bbox": [
+            507,
+            695,
+            672,
+            711
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To accelerate convergence during RL training, we propose the GTA, which introduces a novel Guess stage to the conventional reasoning process. As illustrated in Figure 1, our prompt design guides the model to sequentially produce three components: Guess, Think, and Answer. The right side of the figure presents an example of a model-generated response adhering to this structure.",
+        "bbox": [
+            507,
+            715,
+            882,
+            844
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Guess: In this initial stage, the model generates a preliminary answer based on intuition or prior knowledge. This guess serves as a reference point for subsequent reasoning and",
+        "bbox": [
+            531,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1051",
+        "bbox": [
+            482,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/0af04e2d463858b2d422c3459d0549d8579ad93ffa2b917012639d55554018ec.jpg",
+        "image_caption": [
+            "Figure 1: Overview of the Guess-Think-Answer framework."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            84,
+            868,
+            292
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the final answer. During training, the Guess component is supervised using cross-entropy loss.",
+        "bbox": [
+            147,
+            346,
+            485,
+            394
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Think: Building upon the initial guess and the input question, the model produces a sequence of reasoning steps. This process aids in task comprehension and enhances the model's generalization capabilities.",
+            "- Answer: The final answer is generated by integrating insights from both the Guess and Think stages. This answer may align with or differ from the initial guess. The quality of the Answer, along with the overall output structure, is evaluated using a reward signal, which guides the RL component of the training."
+        ],
+        "bbox": [
+            136,
+            409,
+            489,
+            617
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "By incorporating supervised signals in the Guess stage, our framework accelerates RL convergence and fosters the generation of interpretable reasoning processes.",
+        "bbox": [
+            112,
+            632,
+            489,
+            697
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Training Objective",
+        "text_level": 1,
+        "bbox": [
+            114,
+            709,
+            309,
+            726
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We propose a unified training framework that combines SFT and RL within a single optimization process. As shown in Figure 2, our approach exploits the GTA output format to assign distinct training objectives to the model's various outputs.",
+        "bbox": [
+            112,
+            732,
+            489,
+            813
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "SFT Loss. For the Guess segment, which represents the model's initial prediction based on intuition or prior knowledge, we employ a standard cross-entropy loss. To ensure that the loss computation focuses solely on this segment, we apply a masking strategy that assigns a special token (e.g.,",
+        "bbox": [
+            112,
+            824,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "-100) to tokens outside the Guess span. Formally, the SFT loss is defined as:",
+        "bbox": [
+            505,
+            346,
+            882,
+            376
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S F T}} = - \\sum_ {t \\in \\mathcal {G}} \\log P _ {\\theta} \\left(y _ {t} \\mid y _ {<   t}, x\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            574,
+            385,
+            882,
+            419
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{G}$  denotes the set of token positions corresponding to the Guess segment,  $y_{t}$  is the target token at position  $t$ ,  $y_{< t}$  represents the sequence of preceding tokens, and  $x$  is the input text.",
+        "bbox": [
+            507,
+            426,
+            882,
+            492
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "RL objective function. In optimizing the model's final output, we introduce the GRPO (Shao et al., 2024) algorithm with minor modifications. GRPO improves sample efficiency by generating multiple candidate outputs for the same input prompt and computing relative advantages without a separate value function, thereby reducing training resource consumption. In LLMs, reward signals can be categorized into model-based and rule-based rewards; text classification tasks are particularly amenable to rule-based rewards, which are assigned by directly comparing the model's final prediction to the ground-truth label. The reward definitions are as follows:",
+        "bbox": [
+            507,
+            499,
+            882,
+            724
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {\\text {f o r m a t}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f f o r m a t c o r r e c t ,} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            728,
+            882,
+            772
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {\\text {a c c u r a c y}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f c l a s s i f i c a t i o n c o r r e c t ,} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            524,
+            780,
+            882,
+            821
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {\\text {t o t a l}} = R _ {\\text {f o r m a t}} + R _ {\\text {a c c u r a c y}}. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            551,
+            831,
+            880,
+            847
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $R_{\\mathrm{format}}$  denotes the format reward, which is granted whenever the model's output adheres to the prescribed GTA format.  $R_{\\mathrm{accuracy}}$  denotes the accuracy reward, which is awarded when the model's",
+        "bbox": [
+            507,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1052",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/e611e900219ef873b64390fad53f04721dba741950aa5ac3b2b3e703860cc765.jpg",
+        "image_caption": [
+            "Figure 2: Illustration of the loss masking strategy applied during training."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            87,
+            863,
+            219
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "final prediction matches the true label.  $R_{\\mathrm{total}}$  represents the overall reward signal. The overall RL training objective is defined as follows:",
+        "bbox": [
+            112,
+            273,
+            489,
+            321
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {J} (\\theta) = \\mathbb {E} \\big [ q \\sim P (Q), \\{o _ {i} \\} _ {i = 1} ^ {G} \\sim \\pi_ {\\theta_ {\\mathrm {o l d}}} (\\cdot | q) \\big ] \\left[ \\frac {1}{G} \\right] \\\\ \\sum_ {i = 1} ^ {G} \\frac {1}{| o _ {i} |} \\sum_ {t \\not \\in \\mathcal {G}} \\min \\left(r _ {i} (\\theta), \\mathrm {c l i p} (r _ {i} (\\theta), \\right. \\\\ \\left. \\left. 1 - \\epsilon , 1 + \\epsilon\\right) \\hat {A} _ {i, t} \\right] - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ \\pi_ {\\theta} \\| \\pi_ {\\text {r e f}} \\right] \\tag {5} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            119,
+            395,
+            485,
+            545
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $G$  denotes the group size, representing the number of output sequences sampled in parallel for the same prompt. The  $i$ -th output sequence is denoted as  $o_i$ , where a loss mask is applied to the text within the Guess segment to selectively include tokens in the loss computation.  $\\hat{A}_{i,t} = (R_{\\mathrm{total},i} - \\mu) / \\sigma$  denotes the advantage function, obtained by subtracting the group's mean reward  $\\mu$  from each individual reward and then dividing by the group's reward standard deviation  $\\sigma$ . The ratio  $r_i(\\theta) = \\frac{\\pi_\\theta(o_i|q)}{\\pi_{\\theta_{\\mathrm{old}}}(o_i|q)}$  corresponds to the probability ratio between the new and old policies. The hyperparameter  $\\epsilon$  defines the clipping range for gradient updates, and  $\\beta$  adjusts the weight of the Kullback-Leibler (KL) divergence term. Since our training involves backpropagating two distinct loss components, we jointly constrain the magnitude of model updates using both the clipping mechanism and the KL divergence term. While  $\\mathbb{D}_{\\mathrm{KL}}$  represents the KL divergence term and can be further",
+        "bbox": [
+            112,
+            595,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "expressed as:",
+        "bbox": [
+            509,
+            274,
+            611,
+            288
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbb {D} _ {\\mathrm {K L}} \\left[ \\pi_ {\\theta} \\| \\pi_ {\\text {r e f}} \\right] = \\frac {\\pi_ {\\text {r e f}} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)}{\\pi_ {\\theta} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)} - (6) \\\\ \\log \\frac {\\pi_ {\\text {r e f}} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)}{\\pi_ {\\theta} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)} - 1. (7) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            300,
+            882,
+            393
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "we adopt the same KL divergence term computation as in the original GRPO, but instead of using a static base model as the reference, we periodically update the reference model with the current model during training. This strategy prevents the model's updates from being constrained too closely to the base model, which could otherwise hinder performance improvements. Policy optimization algorithms maximize the objective function  $\\mathcal{J}(\\theta)$  via gradient ascent, which is equivalent to finding the minimum of  $-\\mathcal{J}(\\theta)$  by gradient descent; accordingly, the RL loss function can be expressed as follows:",
+        "bbox": [
+            507,
+            407,
+            884,
+            613
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R L}} = - \\mathcal {J} (\\theta) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            638,
+            615,
+            882,
+            633
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Total loss function. The total loss function is defined as the sum of two distinct loss components, and is computed as follows:",
+        "bbox": [
+            507,
+            642,
+            882,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {T o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {S F T}} + \\lambda_ {2} \\mathcal {L} _ {\\text {R L}}, \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            594,
+            705,
+            882,
+            720
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_{1}$  and  $\\lambda_{2}$  are two hyperparameters that balance the weights of the SFT and RL loss terms.",
+        "bbox": [
+            507,
+            734,
+            882,
+            766
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Loss Mask. During training, the masking strategy ensures that each loss component only affects its intended segment. When computing the SFT loss, tokens outside the Guess section are masked, enabling the language model to learn the correct labels via cross-entropy loss solely on the Guess portion. Conversely, during RL, tokens within the Guess section are masked in the loss calculation to prevent adverse learning signals from the guessed",
+        "bbox": [
+            507,
+            776,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1053",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "labels. This approach effectively isolates the two loss computations, reducing gradient conflicts during backpropagation.",
+        "bbox": [
+            112,
+            84,
+            489,
+            134
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Gradient Conflict Detection and Resolution",
+        "text_level": 1,
+        "bbox": [
+            112,
+            143,
+            413,
+            175
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Despite the loss masking mechanism described above and the assignment of distinct weights to each objective to reduce gradient conflicts, multitask learning cannot fully avoid such interference. To further mitigate gradient conflicts, we analyze the gradients of the two losses during backpropagation and integrate theoretical insights from PCGrad (Yu et al., 2020) into the training process. As illustrated in Figure 3, we use cosine similarity to detect gradient conflicts: a positive cosine similarity between gradients from the two losses indicates no conflict during backpropagation, whereas a negative cosine similarity denotes the occurrence of a gradient conflict. The calculation formula of cosine similarity under the gradient can be expressed as follows:",
+        "bbox": [
+            112,
+            181,
+            489,
+            439
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\cos (\\theta) = \\frac {\\nabla \\mathcal {L} _ {\\mathrm {S F T}} \\cdot \\nabla \\mathcal {L} _ {\\mathrm {R L}}}{\\| \\nabla \\mathcal {L} _ {\\mathrm {S F T}} \\| \\cdot \\| \\nabla \\mathcal {L} _ {\\mathrm {R L}} \\|}. \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            181,
+            447,
+            487,
+            483
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta$  is the angle between the two gradient vectors, and  $\\nabla$  denotes the gradient calculated by backpropagation under the corresponding loss. The final loss update rule is:",
+        "bbox": [
+            112,
+            493,
+            490,
+            558
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {F i n a l}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\text {T o t a l}}, & \\text {i f} \\nabla \\mathcal {L} _ {\\mathrm {S F T}} \\cdot \\nabla \\mathcal {L} _ {\\mathrm {R L}} > 0, \\\\ \\mathcal {L} _ {\\mathrm {R L}}, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            126,
+            569,
+            487,
+            612
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During parameter updates, when gradient conflicts arise, we mitigate such conflicts by retaining only the loss component associated with the RL objective.",
+        "bbox": [
+            112,
+            623,
+            489,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            700,
+            321,
+            718
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This section outlines the experimental setup used to evaluate our approach. We describe the datasets, baseline models, and implementation details.",
+        "bbox": [
+            112,
+            727,
+            489,
+            776
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            787,
+            228,
+            802
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Considering the resource constraints of our experimental process, we selected four datasets reflecting distinct classification scenarios based on recent studies (Chen et al., 2024; Menon and Srivastava, 2024; Chen et al., 2025b): SST-5 (Socher et al., 2013), Amazon (FitzGerald et al., 2022), Emotion (Saravia et al., 2018), and BBC News (Greene",
+        "bbox": [
+            112,
+            808,
+            490,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3d311ecf656058ad60cb12cfd55346742343806b837ccf976a45b9b2ad5acca5.jpg",
+        "image_caption": [
+            "Figure 3: Illustration of gradient conflict analysis via gradient cosine similarity"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            524,
+            84,
+            870,
+            187
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and Cunningham, 2006), each containing multiple categories. The detailed descriptions of these datasets are presented in Table 1; they cover four domains—movie reviews, intent recognition, English tweets, and News—with class counts ranging from five to eighteen.",
+        "bbox": [
+            507,
+            256,
+            884,
+            354
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Models",
+        "text_level": 1,
+        "bbox": [
+            507,
+            368,
+            613,
+            382
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Below we provide a concise summary of LLMs used in our experiments:",
+        "bbox": [
+            507,
+            391,
+            880,
+            422
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Qwen2.5 (3B) $^{1}$  is Alibaba's open-source 3 billion parameters instruction-tuned LLM supporting long-context understanding (up to 128K tokens) and generation (up to 8K tokens), making it adept at handling extended dialogues and complex prompts. It features robust multilingual comprehension across 29 languages, ensuring broad applicability in diverse language settings. It excels in structured output generation (e.g., JSON) and instruction following.",
+        "bbox": [
+            507,
+            423,
+            882,
+            583
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Qwen3  $(4\\mathbf{B})^{2}$  is Alibaba's latest open-source large language model with 4 billion parameters. Trained on a substantially larger corpus of 36 trillion tokens across 119 languages and dialects, it delivers robust performance across diverse reasoning and understanding tasks. It supports hybrid reasoning modes, seamlessly switching between CoT thinking and direct-response generation.",
+        "bbox": [
+            507,
+            585,
+            882,
+            714
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Llama3.2  $(3\\mathbf{B})^{3}$  is a 3 billion parameter open-source model released by Meta. It is designed as a lightweight variant of the Llama 3 family, trained on a diverse multilingual corpus and optimized for efficiency on resource-constrained environments. Despite its relatively small size, Llama3.2 demonstrates competitive performance on a wide range of reasoning and classification tasks.",
+        "bbox": [
+            507,
+            715,
+            884,
+            845
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "1https://huggingface.co/Qwen/Qwen2.",
+            "5-3B-Instruct",
+            "$^{2}$ https://huggingface.co/Qwen/Qwen3-4B",
+            "3https://huggingface.co/meta-llama/Llama-3. 2-3B"
+        ],
+        "bbox": [
+            507,
+            857,
+            855,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1054",
+        "bbox": [
+            482,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e43c3924a3fd63441080cdcdbca30b10da2081b608893c36192e53d871e21219.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Description</td><td>Training</td><td>Testing</td><td>Classes</td></tr><tr><td>SST-5</td><td>The Stanford Sentiment Treebank five-class benchmark is a standard corpus for fine-grained sentiment classification. It consists of sentences drawn from movie reviews, each annotated at the sentence level with one of five sentiment labels—very negative, negative, neutral, positive, and very positive.</td><td>8,544</td><td>2,210</td><td>5</td></tr><tr><td>Amazon</td><td>This is the English (en-US) subset of the Massive Sceno-nario Classification task from the Massive Text Embed-ding Benchmark (MTEB), aimed at intent prediction in voice assistant interactions. The dataset covers 18 scenario classes (such as alarm, audio, iot, calendar, play, news, and weather).</td><td>11,514</td><td>2,974</td><td>18</td></tr><tr><td>Emotion</td><td>The emotion dataset is a carefully curated subset of English tweets annotated with six basic emotions—sadness, joy, love, anger, fear, and surprise—providing a stan-dardized benchmark for evaluating emotion recognition models. Each sample consists of a tweet paired with its corresponding label.</td><td>16,000</td><td>2,000</td><td>6</td></tr><tr><td>BBC News</td><td>Dataset on BBC News Topic Classification published on the BBC News website corresponding during 2004-2005. Each article is labeled under one of 5 categories: business, entertainment, politics, sport or tech.</td><td>1,225</td><td>1,000</td><td>5</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            877,
+            470
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Detailed description of datasets utilized in the experimental process. Each dataset differs in terms of the number of classes, training samples, and test samples.",
+        "bbox": [
+            112,
+            478,
+            882,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Hyperparameters",
+        "text_level": 1,
+        "bbox": [
+            112,
+            533,
+            302,
+            549
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "All experiments were carried out on a multi-node cluster, each node hosting four NVIDIA L40s GPUs (48 GB each) and coordinated via DeepSpeed with ZeRO Stage 2. For SFT, we employed the open-source ModelScope Swift framework, while RL baselines used the TRL library's GRPO implementation and our proposed GTA built atop GRPO. Inputs were truncated to a maximum of 4,096 tokens, and models were trained in bfloat16 precision with a per-device batch size of 4. In the RL phase, each prompt generated 16 candidate answers, and we applied importance sampling with a reuse factor of 4, and included a KL penalty weighted by  $\\beta = 0.01$  to stabilize policy updates. In our GTA, losses are assigned equal weights of 1. Each dataset is trained for 3-4 epochs.",
+        "bbox": [
+            112,
+            556,
+            489,
+            815
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            828,
+            329,
+            846
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we report the main experimental results and provide in-depth analyses. We first compare GTA with baseline methods on multiple benchmarks, then investigate convergence behav",
+        "bbox": [
+            112,
+            857,
+            490,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ior, reasoning robustness, and case studies to gain further insights into its effectiveness.",
+        "bbox": [
+            507,
+            533,
+            882,
+            565
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            577,
+            658,
+            592
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The experimental results presented in Table 2 evaluate the fine-tuning performance of three methods—SFT, GRPO, and GTA—across two model scales: Qwen2.5 (3B), Qwen3 (4B), and Llama3.2 (3B). The evaluation spans four classification benchmarks: SST-5, Amazon, Emotion, and BBC News. Performance metrics include accuracy and weighted  $\\mathrm{F}_1$  scores. Accuracy reflects the proportion of correct predictions over the total number of instances. Weighted  $\\mathrm{F}_1$  score computes the harmonic mean of precision and recall, weighted by the number of instances in each class. This metric handles class imbalance by assigning higher weights to classes with more instances.",
+        "bbox": [
+            505,
+            598,
+            884,
+            822
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Across all datasets and different models, GTA consistently outperforms both SFT and GRPO in terms of accuracy and weighted  $\\mathrm{F_1}$  Score. Notably, on the Emotion dataset, GTA achieves an  $\\mathrm{F_1}$  score of 92.47% with Qwen2.5, 92.94% with Qwen3, and 93.36% with Llama3.2 surpassing the other",
+        "bbox": [
+            507,
+            824,
+            885,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1055",
+        "bbox": [
+            482,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4f17a518f36c0b1b922ef60f8e852ae0312c986d712918414987fd602186c9be.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"2\">Base</td><td colspan=\"2\">SFT</td><td colspan=\"2\">GRPO</td><td colspan=\"2\">GTA</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td colspan=\"9\">Qwen2.5 (3B)</td></tr><tr><td>SST-5</td><td>11.76</td><td>13.34</td><td>60.72</td><td>59.59</td><td>58.60</td><td>57.05</td><td>61.58</td><td>61.52</td></tr><tr><td>Amazon</td><td>54.84</td><td>55.48</td><td>91.96</td><td>91.92</td><td>90.82</td><td>91.03</td><td>92.47</td><td>92.46</td></tr><tr><td>Emotion</td><td>58.75</td><td>58.63</td><td>91.35</td><td>91.41</td><td>82.50</td><td>81.54</td><td>92.45</td><td>92.47</td></tr><tr><td>BBC News</td><td>81.50</td><td>82.88</td><td>97.70</td><td>97.70</td><td>95.40</td><td>95.47</td><td>98.50</td><td>98.50</td></tr><tr><td colspan=\"9\">Qwen3 (4B)</td></tr><tr><td>SST-5</td><td>45.88</td><td>39.80</td><td>61.67</td><td>60.87</td><td>59.28</td><td>58.70</td><td>61.95</td><td>60.94</td></tr><tr><td>Amazon</td><td>68.96</td><td>70.28</td><td>92.57</td><td>92.58</td><td>90.55</td><td>90.32</td><td>92.87</td><td>92.92</td></tr><tr><td>Emotion</td><td>51.15</td><td>54.00</td><td>92.20</td><td>92.09</td><td>84.55</td><td>84.23</td><td>92.95</td><td>92.94</td></tr><tr><td>BBC News</td><td>80.40</td><td>81.79</td><td>97.70</td><td>97.70</td><td>94.90</td><td>95.01</td><td>97.90</td><td>97.91</td></tr><tr><td colspan=\"9\">Llama3.2 (3B)</td></tr><tr><td>SST-5</td><td>38.42</td><td>36.11</td><td>59.91</td><td>50.65</td><td>56.33</td><td>53.40</td><td>61.18</td><td>60.13</td></tr><tr><td>Amazon</td><td>19.60</td><td>18.79</td><td>91.69</td><td>91.52</td><td>84.13</td><td>83.01</td><td>92.84</td><td>92.84</td></tr><tr><td>Emotion</td><td>41.65</td><td>42.57</td><td>93.00</td><td>92.92</td><td>74.40</td><td>74.42</td><td>93.30</td><td>93.36</td></tr><tr><td>BBC News</td><td>34.30</td><td>25.15</td><td>97.40</td><td>97.41</td><td>91.30</td><td>91.49</td><td>97.50</td><td>97.60</td></tr></table>",
+        "bbox": [
+            236,
+            80,
+            759,
+            370
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Fine-tuning performance (%) on four benchmarks using SFT, GRPO, and GTA across two model sizes (Base refers to the model without fine-tuning).",
+        "bbox": [
+            112,
+            381,
+            882,
+            411
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4347cc5238af92f5fcabfb574d3b7b881789e346c36ae337448fc674d4690225.jpg",
+        "image_caption": [
+            "Figure 4: Accuracy comparison of Qwen3 (4B) think and no think across multiple datasets"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            439,
+            480,
+            593
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "methods by a significant margin. Similarly, on the Amazon dataset, GTA attains the highest accuracy and  $\\mathrm{F_1}$  scores, indicating its robustness across different domains. These results underscore the effectiveness of the proposed GTA method in enhancing model performance across diverse datasets and model scales. The consistent improvements in both accuracy and  $\\mathrm{F_1}$  scores highlight GTA's robustness.",
+        "bbox": [
+            112,
+            670,
+            489,
+            815
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Exploring Performance Boundaries in Zero-Shot",
+        "text_level": 1,
+        "bbox": [
+            112,
+            832,
+            460,
+            865
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We first investigated whether reasoning-enhanced prompting could improve classification performance without model fine-tuning, as this would",
+        "bbox": [
+            112,
+            873,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "establish important baselines and validate a core premise of our GTA framework: that explicit reasoning steps can elevate model capabilities. This exploration addresses a fundamental question in LLM deployment—whether performance limitations stem from model capabilities themselves or from suboptimal reasoning processes that can be enhanced through structured prompting. Using the base Qwen3 (4B) model with its native \"think\" mode toggle, we compared standard direct answering against explicit reasoning-then-answering across four benchmarks: SST-5, Amazon, Emotion, and BBC News. Figure 4 demonstrates consistent performance improvements across all datasets when reasoning steps are incorporated. The accuracy on SST-5 increases from  $45.88\\%$  to  $50.59\\%$ , Amazon review classification improves from  $68.96\\%$  to  $71.25\\%$ , Emotion classification improves from  $51.15\\%$  to  $58.00\\%$ , and BBC News classification shows the most significant gain from  $80.40\\%$  to  $90.00\\%$ . These results reveal that models operating in a deliberate reasoning mode possess substantially higher performance ceilings than those constrained to direct response generation.",
+        "bbox": [
+            507,
+            436,
+            884,
+            822
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Convergence Analysis of RL",
+        "text_level": 1,
+        "bbox": [
+            507,
+            834,
+            779,
+            850
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "RL-based fine-tuning typically suffers from slow convergence due to lengthy exploration cycles. To assess how our GTA method addresses this limitation, we compared its convergence speed against",
+        "bbox": [
+            507,
+            856,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1056",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/00084a15327a4f757549455cbb74f8e0e2b4e8dfa8b1ea4b13a3c73f044ec91d.jpg",
+        "image_caption": [
+            "(a) SST-5 (3B)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            93,
+            349,
+            200
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/20865649cf4ab21cce4226ce94b8d6e641d984126b5af2cf090ccd94007885a9.jpg",
+        "image_caption": [
+            "(b) Amazon (3B)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            381,
+            93,
+            600,
+            200
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/d520288942a3a8e0c5650be3620164b60508161ea6efdf55f397888780317f7d.jpg",
+        "image_caption": [
+            "(c) Emotion (3B)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            633,
+            95,
+            848,
+            200
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d4e2a653a1321c1531abe6450451d3d0206904fe829e9d544e249a6d0b5d997.jpg",
+        "image_caption": [
+            "(d) SST-5 (4B)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            230,
+            349,
+            335
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7d84c8c35ddc00486019aa74d30a2896fd07854aad06c002fae23f8de78d892f.jpg",
+        "image_caption": [
+            "(e) Amazon (4B)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            381,
+            230,
+            600,
+            335
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9178812e1ffd66c85f52dc5d7b7dcd5a45aa93df02b16c0ef5d075d9d908f856.jpg",
+        "image_caption": [
+            "(f) Emotion (4B)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            633,
+            230,
+            848,
+            336
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/e9222ca87e74d13de4d99478e2ef2074046ee27147ac393a696a9a47d1311bee.jpg",
+        "image_caption": [
+            "(a) SST-5 (3B) Accuracy Reward",
+            "Figure 6: Accuracy reward and test set accuracy trends of GRPO and GTA on SST-5"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            139,
+            423,
+            462,
+            583
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/09236163d5dece0a48e7018dc5c439c21a1017b38dc81a42cb148c389fb7d476.jpg",
+        "image_caption": [
+            "Figure 5: Variations in accuracy reward values on the SST-5, Amazon, and Emotion datasets during GTA and GRPO fine-tuning",
+            "(b) Accuracy  $(\\%)$  on the SST-5 (3B) Test Set"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            414,
+            858,
+            575
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the GRPO baseline across multiple datasets and model scales. Since BBC News is a relatively easy dataset with fewer categories and less training data, its convergence trends are less informative for analyzing RL training dynamics. We therefore focus our convergence analysis on SST-5, Amazon, and Emotion. Figure 5 illustrates the accuracy reward trajectories for both 3B and 4B models on SST-5, Emotion, and Amazon datasets, tracking three curves: GTA's answer segment (blue), GTA's guess segment (green), and GRPO (orange). On SST-5 and Emotion datasets, GTA demonstrates a substantial convergence advantage, with both guess and answer rewards consistently outperforming GRPO. While GRPO achieves faster initial convergence on the Amazon dataset, GTA maintains superior",
+        "bbox": [
+            112,
+            651,
+            489,
+            908
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "overall performance throughout training. These results confirm that the supervised guess segment effectively guides the answer segment, dramatically accelerating convergence while preserving the optimization benefits of RL—offering significant advantages for applying RL to language models in classification tasks.",
+        "bbox": [
+            507,
+            651,
+            884,
+            762
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To investigate whether GRPO might eventually match GTA's performance given sufficient training time, we extended our experiments on the SST-5 dataset to 10,000 steps (approximately 10 epochs). Figure 6 presents this extended analysis using the 3B-parameter model, with Figure 6(a) showing training reward progression and Figure 6(b) depicting test accuracy evolution. Even in this extended training scenario, GRPO's convergence remains",
+        "bbox": [
+            507,
+            776,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1057",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/15f5b0cf44b98337a9003fcfc0eda19248de74b47a34dc273bc962b7d3025e2d.jpg",
+        "image_caption": [
+            "Figure 7: Illustration of GTA reasoning: from incorrect guess to correct answer"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            137,
+            83,
+            858,
+            294
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/52a495c38afac294e9a6840b603e4dadb39cef17783d170690bb33c754d59395.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 3: Performance  $(\\%)$  of Qwen2.5 (3B) on GTA trained by RL on SST-5. GTA (RL) means that all stages are trained using RL, and GTA  $(\\mathrm{SFT} + \\mathrm{RL})$  indicates that the loss of the guess is calculated through SFT."
+        ],
+        "table_body": "<table><tr><td></td><td>GRPO</td><td>GTA (RL)</td><td>GTA (SFT+RL)</td></tr><tr><td>Acc</td><td>58.60</td><td>57.42</td><td>61.58</td></tr><tr><td>F1</td><td>57.05</td><td>56.52</td><td>61.52</td></tr></table>",
+        "bbox": [
+            136,
+            344,
+            463,
+            398
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "substantially slower than GTA's. Despite showing gradual improvements in both reward and accuracy, GRPO fails to match GTA's performance even after 10,000 steps. These findings demonstrate that GTA not only accelerates training but also achieves a higher performance ceiling, offering significant practical advantages for deploying RL in language model fine-tuning for classification tasks.",
+        "bbox": [
+            112,
+            488,
+            487,
+            617
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.4 Ablation on Guess",
+        "text_level": 1,
+        "bbox": [
+            112,
+            627,
+            305,
+            642
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To validate the contribution of the supervised Guess stage, we replace the Guess component with RL updates. The results are shown in Table 3, We can observe that replacing supervised Guess with RL training does not improve the final accuracy. In fact, it yields a drop. This indicates the effectiveness of using supervised loss in the Guess stage.",
+        "bbox": [
+            112,
+            649,
+            487,
+            762
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.5 Reasoning Process Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            771,
+            376,
+            788
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We find that although making a guess under the guidance of supervision can accelerate convergence, the model does not blindly commit to the guessed answer as the final prediction. As show in Figure 7, when the model produces an incorrect guess, subsequent reasoning steps end to allow it to correct previous mistakes. In Case A, the model first predicts an incorrect label but gradually revises",
+        "bbox": [
+            112,
+            791,
+            487,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "its reasoning and ultimately derives the correct answer. In Case B, the model not only outputs the correct final label but also explicitly indicates that the candidate label set does not include \"sport\", thereby mitigating hallucination issues commonly observed in SFT. These examples, together with the overall accuracy improvements, highlight that GTA exhibits stronger robustness than purely SFT methods that only optimize the Guess segment.",
+        "bbox": [
+            507,
+            349,
+            884,
+            494
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            507,
+            794,
+            523
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we present GTA, a novel training framework that addresses the efficiency-capability trade-off between SFT and RL by introducing a Guess stage to traditional CoT outputs. Under this framework, model outputs are organized into Guess, Think, and Answer segments, where the Guess is optimized via SFT while the overall format and final output are optimized through RL. To mitigate gradient conflicts, we apply loss masking and cosine-similarity constraints. Experimental results on four text-classification benchmarks show that GTA consistently outperforms pure SFT and GRPO baselines in accuracy and  $\\mathrm{F_1}$  scores while achieving significantly faster convergence than pure RL approaches. Through analysis of training dynamics, we demonstrate that GTA successfully combines the efficiency of supervised learning with the performance gains of RL, substantially accelerating convergence and addressing the exploration inefficiency inherent to RL-based fine-tuning. Theoretically, our approach is not limited to text classification tasks, and we plan to explore extending GTA to broader NLP tasks in future work.",
+        "bbox": [
+            505,
+            535,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1058",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "While publicly available text classification datasets often contain noisy data, we did not perform preprocessing or sample high-quality subsets in this study. Additionally, due to resource constraints, we validated our proposed method only on 3B and 4B parameter models, without extending the evaluation to larger-scale models. Given that RL heavily relies on the underlying model's capacity, more powerful and generalizable models may yield greater benefits.",
+        "bbox": [
+            112,
+            110,
+            490,
+            269
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            299,
+            213,
+            313
+        ],
+        "page_idx": 9
+    },
+    {
+        "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, and Shyamal et al. Anadkat. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.",
+            "Ju-Seung Byun, Jiyun Chun, Jihyung Kil, and Andrew Perrault. 2024. Ares: Alternating reinforcement learning and supervised fine-tuning for enhanced multi-modal chain-of-thought reasoning through diverse ai feedback. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4410-4430.",
+            "Junfan Chen, Richong Zhang, Junchi Chen, and Chunming Hu. 2024. Open-set semi-supervised text classification via adversarial disagreement maximization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2170-2180.",
+            "Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, and Fan et al. Yang. 2025a. Research: Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470.",
+            "Si-An Chen, Hsuan-Tien Lin, and Chih-Jen Lin. 2025b. Preserving zero-shot capability in supervised finetuning for multi-label text classification. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 5699-5712.",
+            "Yun-Nung Chen, Dilek Hakkani-Tür, and Xiaodong He. 2016. Zero-shot learning of intent embeddings for expansion by convolutional deep structured semantic models. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6045-6049. IEEE.",
+            "Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. 2018. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR."
+        ],
+        "bbox": [
+            115,
+            323,
+            489,
+            921
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.",
+            "Jack FitzGerald, Christopher Hench, Charith Peris, Scott Mackie, Kay Rottmann, Ana Sanchez, Aaron Nash, Liam Urbach, Vishesh Kakarala, and Richa et al. Singh. 2022. Massive: A 1m-example multilingual natural language understanding dataset with 51 typologically-diverse languages. arXiv preprint arXiv:2204.08582.",
+            "Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, and Alex et al. Vaughan. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv-2407.",
+            "Derek Greene and Pádraig Cunningham. 2006. Practical solutions to the problem of diagonal dominance in kernel document clustering. In Proceedings of the 23rd international conference on Machine learning, pages 377-384.",
+            "Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, and Xiao et al. Bi. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.",
+            "Thorsten Joachims. 1998. Text categorization with support vector machines: Learning with many relevant features. In European conference on machine learning, pages 137-142. Springer.",
+            "Rie Johnson and Tong Zhang. 2015. Effective use of word order for text categorization with convolutional neural networks. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL HLT 2015, pages 103-112. Association for Computational Linguistics (ACL).",
+            "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.",
+            "Arina Kostina, Marios D Dikaiakos, Dimosthenis Stefanidis, and George Pallis. 2025. Large language models for text classification: Case study and comprehensive review. arXiv preprint arXiv:2501.08457.",
+            "Rakesh R Menon and Shashank Srivastava. 2024. Discern: Decoding systematic errors in natural language for text classifiers. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 19565-19583.",
+            "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, and Alex et al. Ray. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            921
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1059",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? sentiment classification using machine learning techniques. In Proceedings of the ACL-02 conference on Empirical methods in natural language processing-Volume 10, pages 79-86.",
+            "Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. 2018. Carer: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 3687-3697.",
+            "Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, and Y et al. Wu. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.",
+            "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642.",
+            "Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansoresh Karami, Jundong Li, Lu Cheng, and Huan Liu. 2024. Large language models for data annotation and synthesis: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 930-957.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.",
+            "Shuhe Wang, Shengyu Zhang, Jie Zhang, Runyi Hu, Xiaoya Li, Tianwei Zhang, Jiwei Li, Fei Wu, Guoyin Wang, and Eduard Hovy. 2024. Reinforcement learning enhanced llms: A survey. arXiv preprint arXiv:2412.10400.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, and Denny et al. Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.",
+            "Yu Xia, Rui Wang, Xu Liu, Mingyan Li, Tong Yu, Xiang Chen, Julian McAuley, and Shuai Li. 2025. Beyond chain-of-thought: A survey of chain-of-x paradigms for llms. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10795-10809.",
+            "Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, and Fanjin et al. Meng. 2025. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint arXiv:2501.09686."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, and Haoran et al. Wei. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.",
+            "Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836."
+        ],
+        "bbox": [
+            512,
+            85,
+            880,
+            212
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1060",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2025/GTA_ Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models/5317a317-b31f-4425-8151-f38a993fa23c_model.json

@@ -0,0 +1,2048 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.123,
+                0.09,
+                0.877,
+                0.131
+            ],
+            "angle": 0,
+            "content": "GTA: Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.158,
+                0.87,
+                0.175
+            ],
+            "angle": 0,
+            "content": "Min Zeng, Jingfei Sun, Xueyou Luo, Caiquan Liu, Shiqi Zhang, Li Xie, Xiaoxin Chen"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.449,
+                0.177,
+                0.55,
+                0.19
+            ],
+            "angle": 0,
+            "content": "vivo AI Lab"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.429,
+                0.195,
+                0.569,
+                0.208
+            ],
+            "angle": 0,
+            "content": "zengmin.ai@vivo.com"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.261,
+                0.341,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.289,
+                0.461,
+                0.658
+            ],
+            "angle": 0,
+            "content": "In natural language processing tasks, pure reinforcement learning (RL) fine-tuning methods often suffer from inefficient exploration and slow convergence; while supervised fine-tuning (SFT) methods, although efficient in training, have limited performance ceiling and less solid theoretical foundation compared to RL. To address efficiency-capability trade-off, we propose the Guess-Think-Answer (GTA) framework that combines the efficiency of SFT with the capability gains of RL in a unified training paradigm. GTA works by having the model first produce a provisional guess (optimized via cross-entropy loss), then reflect on this guess before generating the final answer, with RL rewards shaping both the final output and the format of the entire GTA structure. This hybrid approach achieves both faster convergence than pure RL and higher performance ceiling than pure SFT. To mitigate gradient conflicts between the two training signals, we employ loss masking and gradient constraints. Empirical results on four text classification benchmarks demonstrate that GTA substantially accelerates convergence while outperforming both standalone SFT and RL baselines."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.671,
+                0.26,
+                0.687
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.697,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Text classification, as a foundational task in natural language processing (NLP), has been widely employed for sentiment analysis (Pang et al., 2002), intent recognition (Chen et al., 2016), and news categorization (Johnson and Zhang, 2015). Early NLP solutions primarily relied on rule-based systems and statistical models—including hidden Markov models (HMMs) and support vector machines (SVMs)—to learn patterns from annotated corpora (Joachims, 1998). The emergence of deep learning and Transformer (Vaswani et al., 2017) architectures dramatically enhanced classification performance, with Bidirectional Encoder Representations from Transformers (BERT)'s (Devlin,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.262,
+                0.885,
+                0.31
+            ],
+            "angle": 0,
+            "content": "2018) bidirectional pre-training paradigm capturing rich contextual representations and achieving breakthroughs across multiple tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.31,
+                0.885,
+                0.664
+            ],
+            "angle": 0,
+            "content": "Large language models (LLMs)—such as GPT (Achiam et al., 2023), Llama (Grattafori et al., 2024), Qwen (Yang et al., 2024), and DeepSeek (Guo et al., 2025)—have demonstrated remarkable capabilities across numerous NLP tasks, including text classification (Kostina et al., 2025). While SFT has been the predominant approach to adapt these models for specific tasks, it faces inherent limitations: SFT methods directly learn to produce correct answers without explicit reasoning, leading to limited generalization capabilities and performance ceilings. Chain-of-thought (CoT) prompting (Wei et al., 2022)—a technique that guides models to generate intermediate reasoning steps before producing final answers—has shown significant improvements across various reasoning tasks (Kojima et al., 2022; Xia et al., 2025). However, applying CoT within the SFT paradigm requires extensive human annotation of reasoning chains, resulting in substantial costs and susceptibility to annotator biases and quality inconsistencies (Tan et al., 2024; Byun et al., 2024)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.665,
+                0.886,
+                0.922
+            ],
+            "angle": 0,
+            "content": "RL offers a promising alternative that can theoretically overcome these limitations by combining the benefits of CoT reasoning with optimization-based learning (Xu et al., 2025; Wang et al., 2024). From reinforcement learning from human feedback (RLHF) (Ouyang et al., 2022) to advanced frameworks like group relative policy optimization (GRPO) (Shao et al., 2024), RL techniques can explore and optimize intermediate reasoning processes without requiring manually annotated reasoning chains. This approach holds particular promise for enhancing model performance beyond what SFT can achieve. However, the application of RL to text classification tasks remains challenging due to fundamental limitations: unlike supervised learning's direct approach, RL methods"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1050"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.781,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 1050-1060"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.282,
+                0.959,
+                0.716,
+                0.972
+            ],
+            "angle": 0,
+            "content": "November 4-9, 2025 ©2025 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.212
+            ],
+            "angle": 0,
+            "content": "must discover optimal reasoning strategies through self-guided exploration—a process often hampered by inefficient exploration, slow convergence, and potential training instability (Chen et al., 2025a). These efficiency challenges have hindered pure RL-based methods from consistently outperforming SFT approaches despite their stronger theoretical foundation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.214,
+                0.49,
+                0.373
+            ],
+            "angle": 0,
+            "content": "To address these challenges, this work introduces a novel Guess-Think-Answer (GTA) framework that seamlessly integrates the advantages of SFT and RL within a unified single-stage training process. In our approach, the model first generates an intuitive guessed answer, then engages in a \"think\" step—reasoning explicitly over the guessed answer and the input question—before producing a refined final answer. Our main contributions can be summarized as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.383,
+                0.49,
+                0.48
+            ],
+            "angle": 0,
+            "content": "1. We propose a novel GTA framework that structures the reasoning process into three distinct phases: an initial intuitive guess, an explicit reasoning step that reflects on this preliminary prediction, and a refined final answer that incorporates this reasoning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.49,
+                0.49,
+                0.666
+            ],
+            "angle": 0,
+            "content": "2. We develop a unified training approach that seamlessly integrates SFT and RL within a single-stage process. Our method applies cross-entropy loss to the guessed answer while optimizing the reasoning process and final answer through RL-based rewards. To ensure effective cooperation between these learning paradigms, we introduce a specialized loss masking strategy and gradient cosine adjustment technique that mitigates potential gradient conflicts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.676,
+                0.49,
+                0.805
+            ],
+            "angle": 0,
+            "content": "3. Our framework eliminates the need for manual annotation of reasoning chains by enabling the model to spontaneously learn effective reasoning patterns through reinforcement. Experimental results demonstrate that GTA significantly outperforms both standard SFT baselines and state-of-the-art RL methods across multiple text classification benchmarks."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.129,
+                0.383,
+                0.49,
+                0.805
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.816,
+                0.271,
+                0.832
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.842,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "CoT prompting has emerged as a powerful technique to enhance the reasoning capabilities of LLMs by guiding them to generate intermediate reasoning steps before producing final answers (Wei et al., 2022; Kojima et al., 2022; Xia et al.,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.244
+            ],
+            "angle": 0,
+            "content": "2025). While CoT significantly improves performance across various reasoning tasks, its application in text classification often requires extensive human annotation of reasoning chains, leading to substantial costs and quality inconsistencies (Tan et al., 2024; Byun et al., 2024). To address these limitations, researchers have explored RL approaches that can optimize model behavior without requiring manual annotation of intermediate steps (Xu et al., 2025; Wang et al., 2024)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.246,
+                0.885,
+                0.55
+            ],
+            "angle": 0,
+            "content": "Traditional RL methods in NLP, however, often suffer from inefficient exploration and slow convergence, making it challenging for pure RL-based methods to consistently outperform SFT approaches despite their stronger theoretical foundation (Chen et al., 2025a). Recent advancements like the GRPO algorithm (Shao et al., 2024) have improved RL efficiency by estimating baselines from group scores, thereby reducing computational costs. Additionally, when combining multiple learning objectives—such as SFT and RL—gradient conflicts can arise, leading to suboptimal convergence. Techniques such as gradient masking and cosine similarity adjustments have been developed to mitigate these conflicts by aligning gradients toward compatible directions (Yu et al., 2020; Chen et al., 2018), which inspires our approach to harmonizing supervised and RL signals within our proposed GTA framework."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.564,
+                0.659,
+                0.58
+            ],
+            "angle": 0,
+            "content": "3 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.589,
+                0.885,
+                0.685
+            ],
+            "angle": 0,
+            "content": "In this section, we present the proposed GTA framework in detail. We begin with an overview of the overall design, followed by the description of each component and training objective, highlighting how the method effectively integrates reinforcement learning with reasoning-oriented supervision."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.696,
+                0.673,
+                0.712
+            ],
+            "angle": 0,
+            "content": "3.1 Prompt Design"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.717,
+                0.884,
+                0.845
+            ],
+            "angle": 0,
+            "content": "To accelerate convergence during RL training, we propose the GTA, which introduces a novel Guess stage to the conventional reasoning process. As illustrated in Figure 1, our prompt design guides the model to sequentially produce three components: Guess, Think, and Answer. The right side of the figure presents an example of a model-generated response adhering to this structure."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.858,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "- Guess: In this initial stage, the model generates a preliminary answer based on intuition or prior knowledge. This guess serves as a reference point for subsequent reasoning and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1051"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.133,
+                0.085,
+                0.87,
+                0.293
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.289,
+                0.307,
+                0.705,
+                0.322
+            ],
+            "angle": 0,
+            "content": "Figure 1: Overview of the Guess-Think-Answer framework."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.149,
+                0.347,
+                0.487,
+                0.395
+            ],
+            "angle": 0,
+            "content": "the final answer. During training, the Guess component is supervised using cross-entropy loss."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.41,
+                0.489,
+                0.492
+            ],
+            "angle": 0,
+            "content": "- Think: Building upon the initial guess and the input question, the model produces a sequence of reasoning steps. This process aids in task comprehension and enhances the model's generalization capabilities."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.504,
+                0.49,
+                0.618
+            ],
+            "angle": 0,
+            "content": "- Answer: The final answer is generated by integrating insights from both the Guess and Think stages. This answer may align with or differ from the initial guess. The quality of the Answer, along with the overall output structure, is evaluated using a reward signal, which guides the RL component of the training."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.41,
+                0.49,
+                0.618
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.633,
+                0.49,
+                0.698
+            ],
+            "angle": 0,
+            "content": "By incorporating supervised signals in the Guess stage, our framework accelerates RL convergence and fosters the generation of interpretable reasoning processes."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.711,
+                0.31,
+                0.727
+            ],
+            "angle": 0,
+            "content": "3.2 Training Objective"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.733,
+                0.49,
+                0.814
+            ],
+            "angle": 0,
+            "content": "We propose a unified training framework that combines SFT and RL within a single optimization process. As shown in Figure 2, our approach exploits the GTA output format to assign distinct training objectives to the model's various outputs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.825,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "SFT Loss. For the Guess segment, which represents the model's initial prediction based on intuition or prior knowledge, we employ a standard cross-entropy loss. To ensure that the loss computation focuses solely on this segment, we apply a masking strategy that assigns a special token (e.g.,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.506,
+                0.347,
+                0.883,
+                0.378
+            ],
+            "angle": 0,
+            "content": "-100) to tokens outside the Guess span. Formally, the SFT loss is defined as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.576,
+                0.386,
+                0.883,
+                0.42
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {S F T}} = - \\sum_ {t \\in \\mathcal {G}} \\log P _ {\\theta} \\left(y _ {t} \\mid y _ {<   t}, x\\right) \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.428,
+                0.884,
+                0.493
+            ],
+            "angle": 0,
+            "content": "where \\(\\mathcal{G}\\) denotes the set of token positions corresponding to the Guess segment, \\(y_{t}\\) is the target token at position \\(t\\), \\(y_{< t}\\) represents the sequence of preceding tokens, and \\(x\\) is the input text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.5,
+                0.884,
+                0.725
+            ],
+            "angle": 0,
+            "content": "RL objective function. In optimizing the model's final output, we introduce the GRPO (Shao et al., 2024) algorithm with minor modifications. GRPO improves sample efficiency by generating multiple candidate outputs for the same input prompt and computing relative advantages without a separate value function, thereby reducing training resource consumption. In LLMs, reward signals can be categorized into model-based and rule-based rewards; text classification tasks are particularly amenable to rule-based rewards, which are assigned by directly comparing the model's final prediction to the ground-truth label. The reward definitions are as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.538,
+                0.73,
+                0.883,
+                0.773
+            ],
+            "angle": 0,
+            "content": "\\[\nR _ {\\text {f o r m a t}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f f o r m a t c o r r e c t ,} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.526,
+                0.781,
+                0.883,
+                0.822
+            ],
+            "angle": 0,
+            "content": "\\[\nR _ {\\text {a c c u r a c y}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f c l a s s i f i c a t i o n c o r r e c t ,} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {3}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.552,
+                0.832,
+                0.882,
+                0.848
+            ],
+            "angle": 0,
+            "content": "\\[\nR _ {\\text {t o t a l}} = R _ {\\text {f o r m a t}} + R _ {\\text {a c c u r a c y}}. \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.858,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "where \\( R_{\\mathrm{format}} \\) denotes the format reward, which is granted whenever the model's output adheres to the prescribed GTA format. \\( R_{\\mathrm{accuracy}} \\) denotes the accuracy reward, which is awarded when the model's"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1052"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.134,
+                0.089,
+                0.864,
+                0.221
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.248,
+                0.234,
+                0.747,
+                0.25
+            ],
+            "angle": 0,
+            "content": "Figure 2: Illustration of the loss masking strategy applied during training."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.274,
+                0.49,
+                0.322
+            ],
+            "angle": 0,
+            "content": "final prediction matches the true label. \\( R_{\\mathrm{total}} \\) represents the overall reward signal. The overall RL training objective is defined as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.12,
+                0.397,
+                0.487,
+                0.546
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathcal {J} (\\theta) = \\mathbb {E} \\big [ q \\sim P (Q), \\{o _ {i} \\} _ {i = 1} ^ {G} \\sim \\pi_ {\\theta_ {\\mathrm {o l d}}} (\\cdot | q) \\big ] \\left[ \\frac {1}{G} \\right] \\\\ \\sum_ {i = 1} ^ {G} \\frac {1}{| o _ {i} |} \\sum_ {t \\not \\in \\mathcal {G}} \\min \\left(r _ {i} (\\theta), \\mathrm {c l i p} (r _ {i} (\\theta), \\right. \\\\ \\left. \\left. 1 - \\epsilon , 1 + \\epsilon\\right) \\hat {A} _ {i, t} \\right] - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ \\pi_ {\\theta} \\| \\pi_ {\\text {r e f}} \\right] \\tag {5} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.596,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "where \\(G\\) denotes the group size, representing the number of output sequences sampled in parallel for the same prompt. The \\(i\\)-th output sequence is denoted as \\(o_i\\), where a loss mask is applied to the text within the Guess segment to selectively include tokens in the loss computation. \\(\\hat{A}_{i,t} = (R_{\\mathrm{total},i} - \\mu) / \\sigma\\) denotes the advantage function, obtained by subtracting the group's mean reward \\(\\mu\\) from each individual reward and then dividing by the group's reward standard deviation \\(\\sigma\\). The ratio \\(r_i(\\theta) = \\frac{\\pi_\\theta(o_i|q)}{\\pi_{\\theta_{\\mathrm{old}}}(o_i|q)}\\) corresponds to the probability ratio between the new and old policies. The hyperparameter \\(\\epsilon\\) defines the clipping range for gradient updates, and \\(\\beta\\) adjusts the weight of the Kullback-Leibler (KL) divergence term. Since our training involves backpropagating two distinct loss components, we jointly constrain the magnitude of model updates using both the clipping mechanism and the KL divergence term. While \\(\\mathbb{D}_{\\mathrm{KL}}\\) represents the KL divergence term and can be further"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.275,
+                0.613,
+                0.29
+            ],
+            "angle": 0,
+            "content": "expressed as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.515,
+                0.301,
+                0.884,
+                0.394
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathbb {D} _ {\\mathrm {K L}} \\left[ \\pi_ {\\theta} \\| \\pi_ {\\text {r e f}} \\right] = \\frac {\\pi_ {\\text {r e f}} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)}{\\pi_ {\\theta} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)} - (6) \\\\ \\log \\frac {\\pi_ {\\text {r e f}} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)}{\\pi_ {\\theta} \\left(o _ {i , t} \\mid q , o _ {i , <   t}\\right)} - 1. (7) \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.408,
+                0.885,
+                0.614
+            ],
+            "angle": 0,
+            "content": "we adopt the same KL divergence term computation as in the original GRPO, but instead of using a static base model as the reference, we periodically update the reference model with the current model during training. This strategy prevents the model's updates from being constrained too closely to the base model, which could otherwise hinder performance improvements. Policy optimization algorithms maximize the objective function \\(\\mathcal{J}(\\theta)\\) via gradient ascent, which is equivalent to finding the minimum of \\(-\\mathcal{J}(\\theta)\\) by gradient descent; accordingly, the RL loss function can be expressed as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.64,
+                0.617,
+                0.883,
+                0.634
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {R L}} = - \\mathcal {J} (\\theta) \\tag {8}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.643,
+                0.884,
+                0.691
+            ],
+            "angle": 0,
+            "content": "Total loss function. The total loss function is defined as the sum of two distinct loss components, and is computed as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.595,
+                0.706,
+                0.883,
+                0.721
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {T o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {S F T}} + \\lambda_ {2} \\mathcal {L} _ {\\text {R L}}, \\tag {9}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.735,
+                0.884,
+                0.767
+            ],
+            "angle": 0,
+            "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are two hyperparameters that balance the weights of the SFT and RL loss terms."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.777,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Loss Mask. During training, the masking strategy ensures that each loss component only affects its intended segment. When computing the SFT loss, tokens outside the Guess section are masked, enabling the language model to learn the correct labels via cross-entropy loss solely on the Guess portion. Conversely, during RL, tokens within the Guess section are masked in the loss calculation to prevent adverse learning signals from the guessed"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1053"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.135
+            ],
+            "angle": 0,
+            "content": "labels. This approach effectively isolates the two loss computations, reducing gradient conflicts during backpropagation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.145,
+                0.414,
+                0.176
+            ],
+            "angle": 0,
+            "content": "3.3 Gradient Conflict Detection and Resolution"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.183,
+                0.49,
+                0.44
+            ],
+            "angle": 0,
+            "content": "Despite the loss masking mechanism described above and the assignment of distinct weights to each objective to reduce gradient conflicts, multitask learning cannot fully avoid such interference. To further mitigate gradient conflicts, we analyze the gradients of the two losses during backpropagation and integrate theoretical insights from PCGrad (Yu et al., 2020) into the training process. As illustrated in Figure 3, we use cosine similarity to detect gradient conflicts: a positive cosine similarity between gradients from the two losses indicates no conflict during backpropagation, whereas a negative cosine similarity denotes the occurrence of a gradient conflict. The calculation formula of cosine similarity under the gradient can be expressed as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.182,
+                0.448,
+                0.488,
+                0.484
+            ],
+            "angle": 0,
+            "content": "\\[\n\\cos (\\theta) = \\frac {\\nabla \\mathcal {L} _ {\\mathrm {S F T}} \\cdot \\nabla \\mathcal {L} _ {\\mathrm {R L}}}{\\| \\nabla \\mathcal {L} _ {\\mathrm {S F T}} \\| \\cdot \\| \\nabla \\mathcal {L} _ {\\mathrm {R L}} \\|}. \\tag {10}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.494,
+                0.491,
+                0.559
+            ],
+            "angle": 0,
+            "content": "where \\(\\theta\\) is the angle between the two gradient vectors, and \\(\\nabla\\) denotes the gradient calculated by backpropagation under the corresponding loss. The final loss update rule is:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.127,
+                0.57,
+                0.488,
+                0.613
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\text {F i n a l}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\text {T o t a l}}, & \\text {i f} \\nabla \\mathcal {L} _ {\\mathrm {S F T}} \\cdot \\nabla \\mathcal {L} _ {\\mathrm {R L}} > 0, \\\\ \\mathcal {L} _ {\\mathrm {R L}}, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {11}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.624,
+                0.49,
+                0.687
+            ],
+            "angle": 0,
+            "content": "During parameter updates, when gradient conflicts arise, we mitigate such conflicts by retaining only the loss component associated with the RL objective."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.701,
+                0.322,
+                0.719
+            ],
+            "angle": 0,
+            "content": "4 Experimental Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.728,
+                0.49,
+                0.777
+            ],
+            "angle": 0,
+            "content": "This section outlines the experimental setup used to evaluate our approach. We describe the datasets, baseline models, and implementation details."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.788,
+                0.229,
+                0.803
+            ],
+            "angle": 0,
+            "content": "4.1 Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.809,
+                0.491,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Considering the resource constraints of our experimental process, we selected four datasets reflecting distinct classification scenarios based on recent studies (Chen et al., 2024; Menon and Srivastava, 2024; Chen et al., 2025b): SST-5 (Socher et al., 2013), Amazon (FitzGerald et al., 2022), Emotion (Saravia et al., 2018), and BBC News (Greene"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.526,
+                0.085,
+                0.872,
+                0.188
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.202,
+                0.883,
+                0.231
+            ],
+            "angle": 0,
+            "content": "Figure 3: Illustration of gradient conflict analysis via gradient cosine similarity"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.258,
+                0.885,
+                0.355
+            ],
+            "angle": 0,
+            "content": "and Cunningham, 2006), each containing multiple categories. The detailed descriptions of these datasets are presented in Table 1; they cover four domains—movie reviews, intent recognition, English tweets, and News—with class counts ranging from five to eighteen."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.369,
+                0.614,
+                0.383
+            ],
+            "angle": 0,
+            "content": "4.2 Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.392,
+                0.882,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Below we provide a concise summary of LLMs used in our experiments:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.424,
+                0.884,
+                0.585
+            ],
+            "angle": 0,
+            "content": "Qwen2.5 (3B)\\(^{1}\\) is Alibaba's open-source 3 billion parameters instruction-tuned LLM supporting long-context understanding (up to 128K tokens) and generation (up to 8K tokens), making it adept at handling extended dialogues and complex prompts. It features robust multilingual comprehension across 29 languages, ensuring broad applicability in diverse language settings. It excels in structured output generation (e.g., JSON) and instruction following."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.586,
+                0.884,
+                0.715
+            ],
+            "angle": 0,
+            "content": "Qwen3 \\((4\\mathbf{B})^{2}\\) is Alibaba's latest open-source large language model with 4 billion parameters. Trained on a substantially larger corpus of 36 trillion tokens across 119 languages and dialects, it delivers robust performance across diverse reasoning and understanding tasks. It supports hybrid reasoning modes, seamlessly switching between CoT thinking and direct-response generation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.716,
+                0.885,
+                0.846
+            ],
+            "angle": 0,
+            "content": "Llama3.2 \\((3\\mathbf{B})^{3}\\) is a 3 billion parameter open-source model released by Meta. It is designed as a lightweight variant of the Llama 3 family, trained on a diverse multilingual corpus and optimized for efficiency on resource-constrained environments. Despite its relatively small size, Llama3.2 demonstrates competitive performance on a wide range of reasoning and classification tasks."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.509,
+                0.858,
+                0.797,
+                0.87
+            ],
+            "angle": 0,
+            "content": "1https://huggingface.co/Qwen/Qwen2."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.871,
+                0.614,
+                0.882
+            ],
+            "angle": 0,
+            "content": "5-3B-Instruct"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.534,
+                0.884,
+                0.815,
+                0.896
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)https://huggingface.co/Qwen/Qwen3-4B"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.897,
+                0.857,
+                0.919
+            ],
+            "angle": 0,
+            "content": "3https://huggingface.co/meta-llama/Llama-3. 2-3B"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.509,
+                0.858,
+                0.857,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1054"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.12,
+                0.082,
+                0.878,
+                0.472
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Dataset</td><td>Description</td><td>Training</td><td>Testing</td><td>Classes</td></tr><tr><td>SST-5</td><td>The Stanford Sentiment Treebank five-class benchmark is a standard corpus for fine-grained sentiment classification. It consists of sentences drawn from movie reviews, each annotated at the sentence level with one of five sentiment labels—very negative, negative, neutral, positive, and very positive.</td><td>8,544</td><td>2,210</td><td>5</td></tr><tr><td>Amazon</td><td>This is the English (en-US) subset of the Massive Sceno-nario Classification task from the Massive Text Embed-ding Benchmark (MTEB), aimed at intent prediction in voice assistant interactions. The dataset covers 18 scenario classes (such as alarm, audio, iot, calendar, play, news, and weather).</td><td>11,514</td><td>2,974</td><td>18</td></tr><tr><td>Emotion</td><td>The emotion dataset is a carefully curated subset of English tweets annotated with six basic emotions—sadness, joy, love, anger, fear, and surprise—providing a stan-dardized benchmark for evaluating emotion recognition models. Each sample consists of a tweet paired with its corresponding label.</td><td>16,000</td><td>2,000</td><td>6</td></tr><tr><td>BBC News</td><td>Dataset on BBC News Topic Classification published on the BBC News website corresponding during 2004-2005. Each article is labeled under one of 5 categories: business, entertainment, politics, sport or tech.</td><td>1,225</td><td>1,000</td><td>5</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.479,
+                0.883,
+                0.51
+            ],
+            "angle": 0,
+            "content": "Table 1: Detailed description of datasets utilized in the experimental process. Each dataset differs in terms of the number of classes, training samples, and test samples."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.534,
+                0.303,
+                0.55
+            ],
+            "angle": 0,
+            "content": "4.3 Hyperparameters"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.557,
+                0.49,
+                0.816
+            ],
+            "angle": 0,
+            "content": "All experiments were carried out on a multi-node cluster, each node hosting four NVIDIA L40s GPUs (48 GB each) and coordinated via DeepSpeed with ZeRO Stage 2. For SFT, we employed the open-source ModelScope Swift framework, while RL baselines used the TRL library's GRPO implementation and our proposed GTA built atop GRPO. Inputs were truncated to a maximum of 4,096 tokens, and models were trained in bfloat16 precision with a per-device batch size of 4. In the RL phase, each prompt generated 16 candidate answers, and we applied importance sampling with a reuse factor of 4, and included a KL penalty weighted by \\(\\beta = 0.01\\) to stabilize policy updates. In our GTA, losses are assigned equal weights of 1. Each dataset is trained for 3-4 epochs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.829,
+                0.33,
+                0.847
+            ],
+            "angle": 0,
+            "content": "5 Results and Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.858,
+                0.491,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In this section, we report the main experimental results and provide in-depth analyses. We first compare GTA with baseline methods on multiple benchmarks, then investigate convergence behav"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.534,
+                0.883,
+                0.566
+            ],
+            "angle": 0,
+            "content": "ior, reasoning robustness, and case studies to gain further insights into its effectiveness."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.578,
+                0.66,
+                0.593
+            ],
+            "angle": 0,
+            "content": "5.1 Performance"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.599,
+                0.885,
+                0.824
+            ],
+            "angle": 0,
+            "content": "The experimental results presented in Table 2 evaluate the fine-tuning performance of three methods—SFT, GRPO, and GTA—across two model scales: Qwen2.5 (3B), Qwen3 (4B), and Llama3.2 (3B). The evaluation spans four classification benchmarks: SST-5, Amazon, Emotion, and BBC News. Performance metrics include accuracy and weighted \\( \\mathrm{F}_1 \\) scores. Accuracy reflects the proportion of correct predictions over the total number of instances. Weighted \\( \\mathrm{F}_1 \\) score computes the harmonic mean of precision and recall, weighted by the number of instances in each class. This metric handles class imbalance by assigning higher weights to classes with more instances."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.825,
+                0.887,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Across all datasets and different models, GTA consistently outperforms both SFT and GRPO in terms of accuracy and weighted \\( \\mathrm{F_1} \\) Score. Notably, on the Emotion dataset, GTA achieves an \\( \\mathrm{F_1} \\) score of 92.47% with Qwen2.5, 92.94% with Qwen3, and 93.36% with Llama3.2 surpassing the other"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1055"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.237,
+                0.081,
+                0.761,
+                0.372
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"2\">Base</td><td colspan=\"2\">SFT</td><td colspan=\"2\">GRPO</td><td colspan=\"2\">GTA</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td colspan=\"9\">Qwen2.5 (3B)</td></tr><tr><td>SST-5</td><td>11.76</td><td>13.34</td><td>60.72</td><td>59.59</td><td>58.60</td><td>57.05</td><td>61.58</td><td>61.52</td></tr><tr><td>Amazon</td><td>54.84</td><td>55.48</td><td>91.96</td><td>91.92</td><td>90.82</td><td>91.03</td><td>92.47</td><td>92.46</td></tr><tr><td>Emotion</td><td>58.75</td><td>58.63</td><td>91.35</td><td>91.41</td><td>82.50</td><td>81.54</td><td>92.45</td><td>92.47</td></tr><tr><td>BBC News</td><td>81.50</td><td>82.88</td><td>97.70</td><td>97.70</td><td>95.40</td><td>95.47</td><td>98.50</td><td>98.50</td></tr><tr><td colspan=\"9\">Qwen3 (4B)</td></tr><tr><td>SST-5</td><td>45.88</td><td>39.80</td><td>61.67</td><td>60.87</td><td>59.28</td><td>58.70</td><td>61.95</td><td>60.94</td></tr><tr><td>Amazon</td><td>68.96</td><td>70.28</td><td>92.57</td><td>92.58</td><td>90.55</td><td>90.32</td><td>92.87</td><td>92.92</td></tr><tr><td>Emotion</td><td>51.15</td><td>54.00</td><td>92.20</td><td>92.09</td><td>84.55</td><td>84.23</td><td>92.95</td><td>92.94</td></tr><tr><td>BBC News</td><td>80.40</td><td>81.79</td><td>97.70</td><td>97.70</td><td>94.90</td><td>95.01</td><td>97.90</td><td>97.91</td></tr><tr><td colspan=\"9\">Llama3.2 (3B)</td></tr><tr><td>SST-5</td><td>38.42</td><td>36.11</td><td>59.91</td><td>50.65</td><td>56.33</td><td>53.40</td><td>61.18</td><td>60.13</td></tr><tr><td>Amazon</td><td>19.60</td><td>18.79</td><td>91.69</td><td>91.52</td><td>84.13</td><td>83.01</td><td>92.84</td><td>92.84</td></tr><tr><td>Emotion</td><td>41.65</td><td>42.57</td><td>93.00</td><td>92.92</td><td>74.40</td><td>74.42</td><td>93.30</td><td>93.36</td></tr><tr><td>BBC News</td><td>34.30</td><td>25.15</td><td>97.40</td><td>97.41</td><td>91.30</td><td>91.49</td><td>97.50</td><td>97.60</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.382,
+                0.884,
+                0.412
+            ],
+            "angle": 0,
+            "content": "Table 2: Fine-tuning performance (%) on four benchmarks using SFT, GRPO, and GTA across two model sizes (Base refers to the model without fine-tuning)."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.12,
+                0.441,
+                0.482,
+                0.594
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.613,
+                0.489,
+                0.645
+            ],
+            "angle": 0,
+            "content": "Figure 4: Accuracy comparison of Qwen3 (4B) think and no think across multiple datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.671,
+                0.49,
+                0.816
+            ],
+            "angle": 0,
+            "content": "methods by a significant margin. Similarly, on the Amazon dataset, GTA attains the highest accuracy and \\( \\mathrm{F_1} \\) scores, indicating its robustness across different domains. These results underscore the effectiveness of the proposed GTA method in enhancing model performance across diverse datasets and model scales. The consistent improvements in both accuracy and \\( \\mathrm{F_1} \\) scores highlight GTA's robustness."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.833,
+                0.462,
+                0.866
+            ],
+            "angle": 0,
+            "content": "5.2 Exploring Performance Boundaries in Zero-Shot"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.874,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We first investigated whether reasoning-enhanced prompting could improve classification performance without model fine-tuning, as this would"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.437,
+                0.885,
+                0.823
+            ],
+            "angle": 0,
+            "content": "establish important baselines and validate a core premise of our GTA framework: that explicit reasoning steps can elevate model capabilities. This exploration addresses a fundamental question in LLM deployment—whether performance limitations stem from model capabilities themselves or from suboptimal reasoning processes that can be enhanced through structured prompting. Using the base Qwen3 (4B) model with its native \"think\" mode toggle, we compared standard direct answering against explicit reasoning-then-answering across four benchmarks: SST-5, Amazon, Emotion, and BBC News. Figure 4 demonstrates consistent performance improvements across all datasets when reasoning steps are incorporated. The accuracy on SST-5 increases from \\(45.88\\%\\) to \\(50.59\\%\\), Amazon review classification improves from \\(68.96\\%\\) to \\(71.25\\%\\), Emotion classification improves from \\(51.15\\%\\) to \\(58.00\\%\\), and BBC News classification shows the most significant gain from \\(80.40\\%\\) to \\(90.00\\%\\). These results reveal that models operating in a deliberate reasoning mode possess substantially higher performance ceilings than those constrained to direct response generation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.835,
+                0.78,
+                0.851
+            ],
+            "angle": 0,
+            "content": "5.3 Convergence Analysis of RL"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.857,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "RL-based fine-tuning typically suffers from slow convergence due to lengthy exploration cycles. To assess how our GTA method addresses this limitation, we compared its convergence speed against"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1056"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.133,
+                0.095,
+                0.35,
+                0.201
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.202,
+                0.205,
+                0.295,
+                0.219
+            ],
+            "angle": 0,
+            "content": "(a) SST-5 (3B)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.382,
+                0.095,
+                0.6,
+                0.201
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.445,
+                0.206,
+                0.553,
+                0.219
+            ],
+            "angle": 0,
+            "content": "(b) Amazon (3B)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.635,
+                0.096,
+                0.85,
+                0.201
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.696,
+                0.206,
+                0.803,
+                0.218
+            ],
+            "angle": 0,
+            "content": "(c) Emotion (3B)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.133,
+                0.231,
+                0.35,
+                0.336
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.202,
+                0.342,
+                0.295,
+                0.354
+            ],
+            "angle": 0,
+            "content": "(d) SST-5 (4B)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.383,
+                0.231,
+                0.6,
+                0.336
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.446,
+                0.342,
+                0.551,
+                0.354
+            ],
+            "angle": 0,
+            "content": "(e) Amazon (4B)"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.635,
+                0.231,
+                0.85,
+                0.337
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.697,
+                0.342,
+                0.802,
+                0.354
+            ],
+            "angle": 0,
+            "content": "(f) Emotion (4B)"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.365,
+                0.882,
+                0.395
+            ],
+            "angle": 0,
+            "content": "Figure 5: Variations in accuracy reward values on the SST-5, Amazon, and Emotion datasets during GTA and GRPO fine-tuning"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.141,
+                0.424,
+                0.463,
+                0.584
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.211,
+                0.588,
+                0.415,
+                0.601
+            ],
+            "angle": 0,
+            "content": "(a) SST-5 (3B) Accuracy Reward"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.51,
+                0.415,
+                0.86,
+                0.576
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.547,
+                0.588,
+                0.822,
+                0.601
+            ],
+            "angle": 0,
+            "content": "(b) Accuracy \\((\\%)\\) on the SST-5 (3B) Test Set"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.212,
+                0.612,
+                0.785,
+                0.626
+            ],
+            "angle": 0,
+            "content": "Figure 6: Accuracy reward and test set accuracy trends of GRPO and GTA on SST-5"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.652,
+                0.49,
+                0.909
+            ],
+            "angle": 0,
+            "content": "the GRPO baseline across multiple datasets and model scales. Since BBC News is a relatively easy dataset with fewer categories and less training data, its convergence trends are less informative for analyzing RL training dynamics. We therefore focus our convergence analysis on SST-5, Amazon, and Emotion. Figure 5 illustrates the accuracy reward trajectories for both 3B and 4B models on SST-5, Emotion, and Amazon datasets, tracking three curves: GTA's answer segment (blue), GTA's guess segment (green), and GRPO (orange). On SST-5 and Emotion datasets, GTA demonstrates a substantial convergence advantage, with both guess and answer rewards consistently outperforming GRPO. While GRPO achieves faster initial convergence on the Amazon dataset, GTA maintains superior"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.652,
+                0.885,
+                0.763
+            ],
+            "angle": 0,
+            "content": "overall performance throughout training. These results confirm that the supervised guess segment effectively guides the answer segment, dramatically accelerating convergence while preserving the optimization benefits of RL—offering significant advantages for applying RL to language models in classification tasks."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.777,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "To investigate whether GRPO might eventually match GTA's performance given sufficient training time, we extended our experiments on the SST-5 dataset to 10,000 steps (approximately 10 epochs). Figure 6 presents this extended analysis using the 3B-parameter model, with Figure 6(a) showing training reward progression and Figure 6(b) depicting test accuracy evolution. Even in this extended training scenario, GRPO's convergence remains"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1057"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.139,
+                0.084,
+                0.86,
+                0.296
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.229,
+                0.309,
+                0.768,
+                0.325
+            ],
+            "angle": 0,
+            "content": "Figure 7: Illustration of GTA reasoning: from incorrect guess to correct answer"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.137,
+                0.346,
+                0.465,
+                0.399
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>GRPO</td><td>GTA (RL)</td><td>GTA (SFT+RL)</td></tr><tr><td>Acc</td><td>58.60</td><td>57.42</td><td>61.58</td></tr><tr><td>F1</td><td>57.05</td><td>56.52</td><td>61.52</td></tr></table>"
+        },
+        {
+            "type": "table_footnote",
+            "bbox": [
+                0.113,
+                0.408,
+                0.489,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Table 3: Performance \\((\\%)\\) of Qwen2.5 (3B) on GTA trained by RL on SST-5. GTA (RL) means that all stages are trained using RL, and GTA \\((\\mathrm{SFT} + \\mathrm{RL})\\) indicates that the loss of the guess is calculated through SFT."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.489,
+                0.489,
+                0.618
+            ],
+            "angle": 0,
+            "content": "substantially slower than GTA's. Despite showing gradual improvements in both reward and accuracy, GRPO fails to match GTA's performance even after 10,000 steps. These findings demonstrate that GTA not only accelerates training but also achieves a higher performance ceiling, offering significant practical advantages for deploying RL in language model fine-tuning for classification tasks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.629,
+                0.307,
+                0.643
+            ],
+            "angle": 0,
+            "content": "5.4 Ablation on Guess"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.65,
+                0.489,
+                0.763
+            ],
+            "angle": 0,
+            "content": "To validate the contribution of the supervised Guess stage, we replace the Guess component with RL updates. The results are shown in Table 3, We can observe that replacing supervised Guess with RL training does not improve the final accuracy. In fact, it yields a drop. This indicates the effectiveness of using supervised loss in the Guess stage."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.772,
+                0.378,
+                0.789
+            ],
+            "angle": 0,
+            "content": "5.5 Reasoning Process Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.793,
+                0.489,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We find that although making a guess under the guidance of supervision can accelerate convergence, the model does not blindly commit to the guessed answer as the final prediction. As show in Figure 7, when the model produces an incorrect guess, subsequent reasoning steps end to allow it to correct previous mistakes. In Case A, the model first predicts an incorrect label but gradually revises"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.35,
+                0.885,
+                0.495
+            ],
+            "angle": 0,
+            "content": "its reasoning and ultimately derives the correct answer. In Case B, the model not only outputs the correct final label but also explicitly indicates that the candidate label set does not include \"sport\", thereby mitigating hallucination issues commonly observed in SFT. These examples, together with the overall accuracy improvements, highlight that GTA exhibits stronger robustness than purely SFT methods that only optimize the Guess segment."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.508,
+                0.796,
+                0.524
+            ],
+            "angle": 0,
+            "content": "6 Conclusion and Future Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.536,
+                0.885,
+                0.921
+            ],
+            "angle": 0,
+            "content": "In this work, we present GTA, a novel training framework that addresses the efficiency-capability trade-off between SFT and RL by introducing a Guess stage to traditional CoT outputs. Under this framework, model outputs are organized into Guess, Think, and Answer segments, where the Guess is optimized via SFT while the overall format and final output are optimized through RL. To mitigate gradient conflicts, we apply loss masking and cosine-similarity constraints. Experimental results on four text-classification benchmarks show that GTA consistently outperforms pure SFT and GRPO baselines in accuracy and \\(\\mathrm{F_1}\\) scores while achieving significantly faster convergence than pure RL approaches. Through analysis of training dynamics, we demonstrate that GTA successfully combines the efficiency of supervised learning with the performance gains of RL, substantially accelerating convergence and addressing the exploration inefficiency inherent to RL-based fine-tuning. Theoretically, our approach is not limited to text classification tasks, and we plan to explore extending GTA to broader NLP tasks in future work."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1058"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.221,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.111,
+                0.491,
+                0.271
+            ],
+            "angle": 0,
+            "content": "While publicly available text classification datasets often contain noisy data, we did not perform preprocessing or sample high-quality subsets in this study. Additionally, due to resource constraints, we validated our proposed method only on 3B and 4B parameter models, without extending the evaluation to larger-scale models. Given that RL heavily relies on the underlying model's capacity, more powerful and generalizable models may yield greater benefits."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.3,
+                0.214,
+                0.315
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.324,
+                0.49,
+                0.391
+            ],
+            "angle": 0,
+            "content": "Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, and Shyamal et al. Anadkat. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.402,
+                0.49,
+                0.495
+            ],
+            "angle": 0,
+            "content": "Ju-Seung Byun, Jiyun Chun, Jihyung Kil, and Andrew Perrault. 2024. Ares: Alternating reinforcement learning and supervised fine-tuning for enhanced multi-modal chain-of-thought reasoning through diverse ai feedback. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4410-4430."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.506,
+                0.49,
+                0.586
+            ],
+            "angle": 0,
+            "content": "Junfan Chen, Richong Zhang, Junchi Chen, and Chunming Hu. 2024. Open-set semi-supervised text classification via adversarial disagreement maximization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2170-2180."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.596,
+                0.49,
+                0.675
+            ],
+            "angle": 0,
+            "content": "Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, and Fan et al. Yang. 2025a. Research: Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.686,
+                0.49,
+                0.754
+            ],
+            "angle": 0,
+            "content": "Si-An Chen, Hsuan-Tien Lin, and Chih-Jen Lin. 2025b. Preserving zero-shot capability in supervised finetuning for multi-label text classification. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 5699-5712."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.764,
+                0.49,
+                0.844
+            ],
+            "angle": 0,
+            "content": "Yun-Nung Chen, Dilek Hakkani-Tür, and Xiaodong He. 2016. Zero-shot learning of intent embeddings for expansion by convolutional deep structured semantic models. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6045-6049. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.855,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. 2018. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.324,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.127
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.134,
+                0.885,
+                0.227
+            ],
+            "angle": 0,
+            "content": "Jack FitzGerald, Christopher Hench, Charith Peris, Scott Mackie, Kay Rottmann, Ana Sanchez, Aaron Nash, Liam Urbach, Vishesh Kakarala, and Richa et al. Singh. 2022. Massive: A 1m-example multilingual natural language understanding dataset with 51 typologically-diverse languages. arXiv preprint arXiv:2204.08582."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.235,
+                0.885,
+                0.302
+            ],
+            "angle": 0,
+            "content": "Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, and Alex et al. Vaughan. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv-2407."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.309,
+                0.885,
+                0.376
+            ],
+            "angle": 0,
+            "content": "Derek Greene and Pádraig Cunningham. 2006. Practical solutions to the problem of diagonal dominance in kernel document clustering. In Proceedings of the 23rd international conference on Machine learning, pages 377-384."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.383,
+                0.885,
+                0.462
+            ],
+            "angle": 0,
+            "content": "Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, and Xiao et al. Bi. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.47,
+                0.885,
+                0.525
+            ],
+            "angle": 0,
+            "content": "Thorsten Joachims. 1998. Text categorization with support vector machines: Learning with many relevant features. In European conference on machine learning, pages 137-142. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.532,
+                0.885,
+                0.624
+            ],
+            "angle": 0,
+            "content": "Rie Johnson and Tong Zhang. 2015. Effective use of word order for text categorization with convolutional neural networks. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL HLT 2015, pages 103-112. Association for Computational Linguistics (ACL)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.632,
+                0.885,
+                0.698
+            ],
+            "angle": 0,
+            "content": "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.706,
+                0.885,
+                0.761
+            ],
+            "angle": 0,
+            "content": "Arina Kostina, Marios D Dikaiakos, Dimosthenis Stefanidis, and George Pallis. 2025. Large language models for text classification: Case study and comprehensive review. arXiv preprint arXiv:2501.08457."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.768,
+                0.885,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Rakesh R Menon and Shashank Srivastava. 2024. Discern: Decoding systematic errors in natural language for text classifiers. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 19565-19583."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.842,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, and Alex et al. Ray. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1059"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? sentiment classification using machine learning techniques. In Proceedings of the ACL-02 conference on Empirical methods in natural language processing-Volume 10, pages 79-86."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.16,
+                0.487,
+                0.238
+            ],
+            "angle": 0,
+            "content": "Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. 2018. Carer: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 3687-3697."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.247,
+                0.487,
+                0.313
+            ],
+            "angle": 0,
+            "content": "Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, and Y et al. Wu. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.321,
+                0.487,
+                0.412
+            ],
+            "angle": 0,
+            "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.421,
+                0.487,
+                0.513
+            ],
+            "angle": 0,
+            "content": "Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansoresh Karami, Jundong Li, Lu Cheng, and Huan Liu. 2024. Large language models for data annotation and synthesis: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 930-957."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.521,
+                0.486,
+                0.586
+            ],
+            "angle": 0,
+            "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.595,
+                0.487,
+                0.66
+            ],
+            "angle": 0,
+            "content": "Shuhe Wang, Shengyu Zhang, Jie Zhang, Runyi Hu, Xiaoya Li, Tianwei Zhang, Jiwei Li, Fei Wu, Guoyin Wang, and Eduard Hovy. 2024. Reinforcement learning enhanced llms: A survey. arXiv preprint arXiv:2412.10400."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.669,
+                0.487,
+                0.746
+            ],
+            "angle": 0,
+            "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, and Denny et al. Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.755,
+                0.487,
+                0.834
+            ],
+            "angle": 0,
+            "content": "Yu Xia, Rui Wang, Xu Liu, Mingyan Li, Tong Yu, Xiang Chen, Julian McAuley, and Shuai Li. 2025. Beyond chain-of-thought: A survey of chain-of-x paradigms for llms. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10795-10809."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.842,
+                0.487,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, and Fanjin et al. Meng. 2025. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint arXiv:2501.09686."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.513,
+                0.086,
+                0.882,
+                0.139
+            ],
+            "angle": 0,
+            "content": "An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, and Haoran et al. Wei. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.513,
+                0.149,
+                0.882,
+                0.214
+            ],
+            "angle": 0,
+            "content": "Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.513,
+                0.086,
+                0.882,
+                0.214
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.483,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "1060"
+        }
+    ]
+]

2025/GTA_ Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models/5317a317-b31f-4425-8151-f38a993fa23c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa99a9ef08db55ae152543dd52e7e798467807f106db4ad16f4f7f7177b6f77f
+size 737335

2025/GTA_ Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models/full.md

@@ -0,0 +1,300 @@
+# GTA: Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models
+
+Min Zeng, Jingfei Sun, Xueyou Luo, Caiquan Liu, Shiqi Zhang, Li Xie, Xiaoxin Chen
+
+vivo AI Lab
+
+zengmin.ai@vivo.com
+
+# Abstract
+
+In natural language processing tasks, pure reinforcement learning (RL) fine-tuning methods often suffer from inefficient exploration and slow convergence; while supervised fine-tuning (SFT) methods, although efficient in training, have limited performance ceiling and less solid theoretical foundation compared to RL. To address efficiency-capability trade-off, we propose the Guess-Think-Answer (GTA) framework that combines the efficiency of SFT with the capability gains of RL in a unified training paradigm. GTA works by having the model first produce a provisional guess (optimized via cross-entropy loss), then reflect on this guess before generating the final answer, with RL rewards shaping both the final output and the format of the entire GTA structure. This hybrid approach achieves both faster convergence than pure RL and higher performance ceiling than pure SFT. To mitigate gradient conflicts between the two training signals, we employ loss masking and gradient constraints. Empirical results on four text classification benchmarks demonstrate that GTA substantially accelerates convergence while outperforming both standalone SFT and RL baselines.
+
+# 1 Introduction
+
+Text classification, as a foundational task in natural language processing (NLP), has been widely employed for sentiment analysis (Pang et al., 2002), intent recognition (Chen et al., 2016), and news categorization (Johnson and Zhang, 2015). Early NLP solutions primarily relied on rule-based systems and statistical models—including hidden Markov models (HMMs) and support vector machines (SVMs)—to learn patterns from annotated corpora (Joachims, 1998). The emergence of deep learning and Transformer (Vaswani et al., 2017) architectures dramatically enhanced classification performance, with Bidirectional Encoder Representations from Transformers (BERT)'s (Devlin,
+
+2018) bidirectional pre-training paradigm capturing rich contextual representations and achieving breakthroughs across multiple tasks.
+
+Large language models (LLMs)—such as GPT (Achiam et al., 2023), Llama (Grattafori et al., 2024), Qwen (Yang et al., 2024), and DeepSeek (Guo et al., 2025)—have demonstrated remarkable capabilities across numerous NLP tasks, including text classification (Kostina et al., 2025). While SFT has been the predominant approach to adapt these models for specific tasks, it faces inherent limitations: SFT methods directly learn to produce correct answers without explicit reasoning, leading to limited generalization capabilities and performance ceilings. Chain-of-thought (CoT) prompting (Wei et al., 2022)—a technique that guides models to generate intermediate reasoning steps before producing final answers—has shown significant improvements across various reasoning tasks (Kojima et al., 2022; Xia et al., 2025). However, applying CoT within the SFT paradigm requires extensive human annotation of reasoning chains, resulting in substantial costs and susceptibility to annotator biases and quality inconsistencies (Tan et al., 2024; Byun et al., 2024).
+
+RL offers a promising alternative that can theoretically overcome these limitations by combining the benefits of CoT reasoning with optimization-based learning (Xu et al., 2025; Wang et al., 2024). From reinforcement learning from human feedback (RLHF) (Ouyang et al., 2022) to advanced frameworks like group relative policy optimization (GRPO) (Shao et al., 2024), RL techniques can explore and optimize intermediate reasoning processes without requiring manually annotated reasoning chains. This approach holds particular promise for enhancing model performance beyond what SFT can achieve. However, the application of RL to text classification tasks remains challenging due to fundamental limitations: unlike supervised learning's direct approach, RL methods
+
+must discover optimal reasoning strategies through self-guided exploration—a process often hampered by inefficient exploration, slow convergence, and potential training instability (Chen et al., 2025a). These efficiency challenges have hindered pure RL-based methods from consistently outperforming SFT approaches despite their stronger theoretical foundation.
+
+To address these challenges, this work introduces a novel Guess-Think-Answer (GTA) framework that seamlessly integrates the advantages of SFT and RL within a unified single-stage training process. In our approach, the model first generates an intuitive guessed answer, then engages in a "think" step—reasoning explicitly over the guessed answer and the input question—before producing a refined final answer. Our main contributions can be summarized as follows:
+
+1. We propose a novel GTA framework that structures the reasoning process into three distinct phases: an initial intuitive guess, an explicit reasoning step that reflects on this preliminary prediction, and a refined final answer that incorporates this reasoning.  
+2. We develop a unified training approach that seamlessly integrates SFT and RL within a single-stage process. Our method applies cross-entropy loss to the guessed answer while optimizing the reasoning process and final answer through RL-based rewards. To ensure effective cooperation between these learning paradigms, we introduce a specialized loss masking strategy and gradient cosine adjustment technique that mitigates potential gradient conflicts.  
+3. Our framework eliminates the need for manual annotation of reasoning chains by enabling the model to spontaneously learn effective reasoning patterns through reinforcement. Experimental results demonstrate that GTA significantly outperforms both standard SFT baselines and state-of-the-art RL methods across multiple text classification benchmarks.
+
+# 2 Related Work
+
+CoT prompting has emerged as a powerful technique to enhance the reasoning capabilities of LLMs by guiding them to generate intermediate reasoning steps before producing final answers (Wei et al., 2022; Kojima et al., 2022; Xia et al.,
+
+2025). While CoT significantly improves performance across various reasoning tasks, its application in text classification often requires extensive human annotation of reasoning chains, leading to substantial costs and quality inconsistencies (Tan et al., 2024; Byun et al., 2024). To address these limitations, researchers have explored RL approaches that can optimize model behavior without requiring manual annotation of intermediate steps (Xu et al., 2025; Wang et al., 2024).
+
+Traditional RL methods in NLP, however, often suffer from inefficient exploration and slow convergence, making it challenging for pure RL-based methods to consistently outperform SFT approaches despite their stronger theoretical foundation (Chen et al., 2025a). Recent advancements like the GRPO algorithm (Shao et al., 2024) have improved RL efficiency by estimating baselines from group scores, thereby reducing computational costs. Additionally, when combining multiple learning objectives—such as SFT and RL—gradient conflicts can arise, leading to suboptimal convergence. Techniques such as gradient masking and cosine similarity adjustments have been developed to mitigate these conflicts by aligning gradients toward compatible directions (Yu et al., 2020; Chen et al., 2018), which inspires our approach to harmonizing supervised and RL signals within our proposed GTA framework.
+
+# 3 Methodology
+
+In this section, we present the proposed GTA framework in detail. We begin with an overview of the overall design, followed by the description of each component and training objective, highlighting how the method effectively integrates reinforcement learning with reasoning-oriented supervision.
+
+# 3.1 Prompt Design
+
+To accelerate convergence during RL training, we propose the GTA, which introduces a novel Guess stage to the conventional reasoning process. As illustrated in Figure 1, our prompt design guides the model to sequentially produce three components: Guess, Think, and Answer. The right side of the figure presents an example of a model-generated response adhering to this structure.
+
+- Guess: In this initial stage, the model generates a preliminary answer based on intuition or prior knowledge. This guess serves as a reference point for subsequent reasoning and
+
+![](images/0af04e2d463858b2d422c3459d0549d8579ad93ffa2b917012639d55554018ec.jpg)  
+Figure 1: Overview of the Guess-Think-Answer framework.
+
+the final answer. During training, the Guess component is supervised using cross-entropy loss.
+
+- Think: Building upon the initial guess and the input question, the model produces a sequence of reasoning steps. This process aids in task comprehension and enhances the model's generalization capabilities.  
+- Answer: The final answer is generated by integrating insights from both the Guess and Think stages. This answer may align with or differ from the initial guess. The quality of the Answer, along with the overall output structure, is evaluated using a reward signal, which guides the RL component of the training.
+
+By incorporating supervised signals in the Guess stage, our framework accelerates RL convergence and fosters the generation of interpretable reasoning processes.
+
+# 3.2 Training Objective
+
+We propose a unified training framework that combines SFT and RL within a single optimization process. As shown in Figure 2, our approach exploits the GTA output format to assign distinct training objectives to the model's various outputs.
+
+SFT Loss. For the Guess segment, which represents the model's initial prediction based on intuition or prior knowledge, we employ a standard cross-entropy loss. To ensure that the loss computation focuses solely on this segment, we apply a masking strategy that assigns a special token (e.g.,
+
+-100) to tokens outside the Guess span. Formally, the SFT loss is defined as:
+
+$$
+\mathcal {L} _ {\mathrm {S F T}} = - \sum_ {t \in \mathcal {G}} \log P _ {\theta} \left(y _ {t} \mid y _ {<   t}, x\right) \tag {1}
+$$
+
+where  $\mathcal{G}$  denotes the set of token positions corresponding to the Guess segment,  $y_{t}$  is the target token at position  $t$ ,  $y_{< t}$  represents the sequence of preceding tokens, and  $x$  is the input text.
+
+RL objective function. In optimizing the model's final output, we introduce the GRPO (Shao et al., 2024) algorithm with minor modifications. GRPO improves sample efficiency by generating multiple candidate outputs for the same input prompt and computing relative advantages without a separate value function, thereby reducing training resource consumption. In LLMs, reward signals can be categorized into model-based and rule-based rewards; text classification tasks are particularly amenable to rule-based rewards, which are assigned by directly comparing the model's final prediction to the ground-truth label. The reward definitions are as follows:
+
+$$
+R _ {\text {f o r m a t}} = \left\{ \begin{array}{l l} 1, & \text {i f f o r m a t c o r r e c t ,} \\ 0, & \text {o t h e r w i s e ,} \end{array} \right. \tag {2}
+$$
+
+$$
+R _ {\text {a c c u r a c y}} = \left\{ \begin{array}{l l} 1, & \text {i f c l a s s i f i c a t i o n c o r r e c t ,} \\ 0, & \text {o t h e r w i s e ,} \end{array} \right. \tag {3}
+$$
+
+$$
+R _ {\text {t o t a l}} = R _ {\text {f o r m a t}} + R _ {\text {a c c u r a c y}}. \tag {4}
+$$
+
+where  $R_{\mathrm{format}}$  denotes the format reward, which is granted whenever the model's output adheres to the prescribed GTA format.  $R_{\mathrm{accuracy}}$  denotes the accuracy reward, which is awarded when the model's
+
+![](images/e611e900219ef873b64390fad53f04721dba741950aa5ac3b2b3e703860cc765.jpg)  
+Figure 2: Illustration of the loss masking strategy applied during training.
+
+final prediction matches the true label.  $R_{\mathrm{total}}$  represents the overall reward signal. The overall RL training objective is defined as follows:
+
+$$
+\begin{array}{l} \mathcal {J} (\theta) = \mathbb {E} \big [ q \sim P (Q), \{o _ {i} \} _ {i = 1} ^ {G} \sim \pi_ {\theta_ {\mathrm {o l d}}} (\cdot | q) \big ] \left[ \frac {1}{G} \right] \\ \sum_ {i = 1} ^ {G} \frac {1}{| o _ {i} |} \sum_ {t \not \in \mathcal {G}} \min \left(r _ {i} (\theta), \mathrm {c l i p} (r _ {i} (\theta), \right. \\ \left. \left. 1 - \epsilon , 1 + \epsilon\right) \hat {A} _ {i, t} \right] - \beta \mathbb {D} _ {\mathrm {K L}} \left[ \pi_ {\theta} \| \pi_ {\text {r e f}} \right] \tag {5} \\ \end{array}
+$$
+
+where  $G$  denotes the group size, representing the number of output sequences sampled in parallel for the same prompt. The  $i$ -th output sequence is denoted as  $o_i$ , where a loss mask is applied to the text within the Guess segment to selectively include tokens in the loss computation.  $\hat{A}_{i,t} = (R_{\mathrm{total},i} - \mu) / \sigma$  denotes the advantage function, obtained by subtracting the group's mean reward  $\mu$  from each individual reward and then dividing by the group's reward standard deviation  $\sigma$ . The ratio  $r_i(\theta) = \frac{\pi_\theta(o_i|q)}{\pi_{\theta_{\mathrm{old}}}(o_i|q)}$  corresponds to the probability ratio between the new and old policies. The hyperparameter  $\epsilon$  defines the clipping range for gradient updates, and  $\beta$  adjusts the weight of the Kullback-Leibler (KL) divergence term. Since our training involves backpropagating two distinct loss components, we jointly constrain the magnitude of model updates using both the clipping mechanism and the KL divergence term. While  $\mathbb{D}_{\mathrm{KL}}$  represents the KL divergence term and can be further
+
+expressed as:
+
+$$
+\begin{array}{l} \mathbb {D} _ {\mathrm {K L}} \left[ \pi_ {\theta} \| \pi_ {\text {r e f}} \right] = \frac {\pi_ {\text {r e f}} \left(o _ {i , t} \mid q , o _ {i , <   t}\right)}{\pi_ {\theta} \left(o _ {i , t} \mid q , o _ {i , <   t}\right)} - (6) \\ \log \frac {\pi_ {\text {r e f}} \left(o _ {i , t} \mid q , o _ {i , <   t}\right)}{\pi_ {\theta} \left(o _ {i , t} \mid q , o _ {i , <   t}\right)} - 1. (7) \\ \end{array}
+$$
+
+we adopt the same KL divergence term computation as in the original GRPO, but instead of using a static base model as the reference, we periodically update the reference model with the current model during training. This strategy prevents the model's updates from being constrained too closely to the base model, which could otherwise hinder performance improvements. Policy optimization algorithms maximize the objective function  $\mathcal{J}(\theta)$  via gradient ascent, which is equivalent to finding the minimum of  $-\mathcal{J}(\theta)$  by gradient descent; accordingly, the RL loss function can be expressed as follows:
+
+$$
+\mathcal {L} _ {\mathrm {R L}} = - \mathcal {J} (\theta) \tag {8}
+$$
+
+Total loss function. The total loss function is defined as the sum of two distinct loss components, and is computed as follows:
+
+$$
+\mathcal {L} _ {\text {T o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {S F T}} + \lambda_ {2} \mathcal {L} _ {\text {R L}}, \tag {9}
+$$
+
+where  $\lambda_{1}$  and  $\lambda_{2}$  are two hyperparameters that balance the weights of the SFT and RL loss terms.
+
+Loss Mask. During training, the masking strategy ensures that each loss component only affects its intended segment. When computing the SFT loss, tokens outside the Guess section are masked, enabling the language model to learn the correct labels via cross-entropy loss solely on the Guess portion. Conversely, during RL, tokens within the Guess section are masked in the loss calculation to prevent adverse learning signals from the guessed
+
+labels. This approach effectively isolates the two loss computations, reducing gradient conflicts during backpropagation.
+
+# 3.3 Gradient Conflict Detection and Resolution
+
+Despite the loss masking mechanism described above and the assignment of distinct weights to each objective to reduce gradient conflicts, multitask learning cannot fully avoid such interference. To further mitigate gradient conflicts, we analyze the gradients of the two losses during backpropagation and integrate theoretical insights from PCGrad (Yu et al., 2020) into the training process. As illustrated in Figure 3, we use cosine similarity to detect gradient conflicts: a positive cosine similarity between gradients from the two losses indicates no conflict during backpropagation, whereas a negative cosine similarity denotes the occurrence of a gradient conflict. The calculation formula of cosine similarity under the gradient can be expressed as follows:
+
+$$
+\cos (\theta) = \frac {\nabla \mathcal {L} _ {\mathrm {S F T}} \cdot \nabla \mathcal {L} _ {\mathrm {R L}}}{\| \nabla \mathcal {L} _ {\mathrm {S F T}} \| \cdot \| \nabla \mathcal {L} _ {\mathrm {R L}} \|}. \tag {10}
+$$
+
+where  $\theta$  is the angle between the two gradient vectors, and  $\nabla$  denotes the gradient calculated by backpropagation under the corresponding loss. The final loss update rule is:
+
+$$
+\mathcal {L} _ {\text {F i n a l}} = \left\{ \begin{array}{l l} \mathcal {L} _ {\text {T o t a l}}, & \text {i f} \nabla \mathcal {L} _ {\mathrm {S F T}} \cdot \nabla \mathcal {L} _ {\mathrm {R L}} > 0, \\ \mathcal {L} _ {\mathrm {R L}}, & \text {o t h e r w i s e}, \end{array} \right. \tag {11}
+$$
+
+During parameter updates, when gradient conflicts arise, we mitigate such conflicts by retaining only the loss component associated with the RL objective.
+
+# 4 Experimental Setup
+
+This section outlines the experimental setup used to evaluate our approach. We describe the datasets, baseline models, and implementation details.
+
+# 4.1 Datasets
+
+Considering the resource constraints of our experimental process, we selected four datasets reflecting distinct classification scenarios based on recent studies (Chen et al., 2024; Menon and Srivastava, 2024; Chen et al., 2025b): SST-5 (Socher et al., 2013), Amazon (FitzGerald et al., 2022), Emotion (Saravia et al., 2018), and BBC News (Greene
+
+![](images/3d311ecf656058ad60cb12cfd55346742343806b837ccf976a45b9b2ad5acca5.jpg)  
+Figure 3: Illustration of gradient conflict analysis via gradient cosine similarity
+
+and Cunningham, 2006), each containing multiple categories. The detailed descriptions of these datasets are presented in Table 1; they cover four domains—movie reviews, intent recognition, English tweets, and News—with class counts ranging from five to eighteen.
+
+# 4.2 Models
+
+Below we provide a concise summary of LLMs used in our experiments:
+
+Qwen2.5 (3B) $^{1}$  is Alibaba's open-source 3 billion parameters instruction-tuned LLM supporting long-context understanding (up to 128K tokens) and generation (up to 8K tokens), making it adept at handling extended dialogues and complex prompts. It features robust multilingual comprehension across 29 languages, ensuring broad applicability in diverse language settings. It excels in structured output generation (e.g., JSON) and instruction following.
+
+Qwen3  $(4\mathbf{B})^{2}$  is Alibaba's latest open-source large language model with 4 billion parameters. Trained on a substantially larger corpus of 36 trillion tokens across 119 languages and dialects, it delivers robust performance across diverse reasoning and understanding tasks. It supports hybrid reasoning modes, seamlessly switching between CoT thinking and direct-response generation.
+
+Llama3.2  $(3\mathbf{B})^{3}$  is a 3 billion parameter open-source model released by Meta. It is designed as a lightweight variant of the Llama 3 family, trained on a diverse multilingual corpus and optimized for efficiency on resource-constrained environments. Despite its relatively small size, Llama3.2 demonstrates competitive performance on a wide range of reasoning and classification tasks.
+
+1https://huggingface.co/Qwen/Qwen2.  
+5-3B-Instruct  
+$^{2}$ https://huggingface.co/Qwen/Qwen3-4B  
+3https://huggingface.co/meta-llama/Llama-3. 2-3B
+
+<table><tr><td>Dataset</td><td>Description</td><td>Training</td><td>Testing</td><td>Classes</td></tr><tr><td>SST-5</td><td>The Stanford Sentiment Treebank five-class benchmark is a standard corpus for fine-grained sentiment classification. It consists of sentences drawn from movie reviews, each annotated at the sentence level with one of five sentiment labels—very negative, negative, neutral, positive, and very positive.</td><td>8,544</td><td>2,210</td><td>5</td></tr><tr><td>Amazon</td><td>This is the English (en-US) subset of the Massive Sceno-nario Classification task from the Massive Text Embed-ding Benchmark (MTEB), aimed at intent prediction in voice assistant interactions. The dataset covers 18 scenario classes (such as alarm, audio, iot, calendar, play, news, and weather).</td><td>11,514</td><td>2,974</td><td>18</td></tr><tr><td>Emotion</td><td>The emotion dataset is a carefully curated subset of English tweets annotated with six basic emotions—sadness, joy, love, anger, fear, and surprise—providing a stan-dardized benchmark for evaluating emotion recognition models. Each sample consists of a tweet paired with its corresponding label.</td><td>16,000</td><td>2,000</td><td>6</td></tr><tr><td>BBC News</td><td>Dataset on BBC News Topic Classification published on the BBC News website corresponding during 2004-2005. Each article is labeled under one of 5 categories: business, entertainment, politics, sport or tech.</td><td>1,225</td><td>1,000</td><td>5</td></tr></table>
+
+Table 1: Detailed description of datasets utilized in the experimental process. Each dataset differs in terms of the number of classes, training samples, and test samples.
+
+# 4.3 Hyperparameters
+
+All experiments were carried out on a multi-node cluster, each node hosting four NVIDIA L40s GPUs (48 GB each) and coordinated via DeepSpeed with ZeRO Stage 2. For SFT, we employed the open-source ModelScope Swift framework, while RL baselines used the TRL library's GRPO implementation and our proposed GTA built atop GRPO. Inputs were truncated to a maximum of 4,096 tokens, and models were trained in bfloat16 precision with a per-device batch size of 4. In the RL phase, each prompt generated 16 candidate answers, and we applied importance sampling with a reuse factor of 4, and included a KL penalty weighted by  $\beta = 0.01$  to stabilize policy updates. In our GTA, losses are assigned equal weights of 1. Each dataset is trained for 3-4 epochs.
+
+# 5 Results and Analysis
+
+In this section, we report the main experimental results and provide in-depth analyses. We first compare GTA with baseline methods on multiple benchmarks, then investigate convergence behav
+
+ior, reasoning robustness, and case studies to gain further insights into its effectiveness.
+
+# 5.1 Performance
+
+The experimental results presented in Table 2 evaluate the fine-tuning performance of three methods—SFT, GRPO, and GTA—across two model scales: Qwen2.5 (3B), Qwen3 (4B), and Llama3.2 (3B). The evaluation spans four classification benchmarks: SST-5, Amazon, Emotion, and BBC News. Performance metrics include accuracy and weighted  $\mathrm{F}_1$  scores. Accuracy reflects the proportion of correct predictions over the total number of instances. Weighted  $\mathrm{F}_1$  score computes the harmonic mean of precision and recall, weighted by the number of instances in each class. This metric handles class imbalance by assigning higher weights to classes with more instances.
+
+Across all datasets and different models, GTA consistently outperforms both SFT and GRPO in terms of accuracy and weighted  $\mathrm{F_1}$  Score. Notably, on the Emotion dataset, GTA achieves an  $\mathrm{F_1}$  score of 92.47% with Qwen2.5, 92.94% with Qwen3, and 93.36% with Llama3.2 surpassing the other
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="2">Base</td><td colspan="2">SFT</td><td colspan="2">GRPO</td><td colspan="2">GTA</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td colspan="9">Qwen2.5 (3B)</td></tr><tr><td>SST-5</td><td>11.76</td><td>13.34</td><td>60.72</td><td>59.59</td><td>58.60</td><td>57.05</td><td>61.58</td><td>61.52</td></tr><tr><td>Amazon</td><td>54.84</td><td>55.48</td><td>91.96</td><td>91.92</td><td>90.82</td><td>91.03</td><td>92.47</td><td>92.46</td></tr><tr><td>Emotion</td><td>58.75</td><td>58.63</td><td>91.35</td><td>91.41</td><td>82.50</td><td>81.54</td><td>92.45</td><td>92.47</td></tr><tr><td>BBC News</td><td>81.50</td><td>82.88</td><td>97.70</td><td>97.70</td><td>95.40</td><td>95.47</td><td>98.50</td><td>98.50</td></tr><tr><td colspan="9">Qwen3 (4B)</td></tr><tr><td>SST-5</td><td>45.88</td><td>39.80</td><td>61.67</td><td>60.87</td><td>59.28</td><td>58.70</td><td>61.95</td><td>60.94</td></tr><tr><td>Amazon</td><td>68.96</td><td>70.28</td><td>92.57</td><td>92.58</td><td>90.55</td><td>90.32</td><td>92.87</td><td>92.92</td></tr><tr><td>Emotion</td><td>51.15</td><td>54.00</td><td>92.20</td><td>92.09</td><td>84.55</td><td>84.23</td><td>92.95</td><td>92.94</td></tr><tr><td>BBC News</td><td>80.40</td><td>81.79</td><td>97.70</td><td>97.70</td><td>94.90</td><td>95.01</td><td>97.90</td><td>97.91</td></tr><tr><td colspan="9">Llama3.2 (3B)</td></tr><tr><td>SST-5</td><td>38.42</td><td>36.11</td><td>59.91</td><td>50.65</td><td>56.33</td><td>53.40</td><td>61.18</td><td>60.13</td></tr><tr><td>Amazon</td><td>19.60</td><td>18.79</td><td>91.69</td><td>91.52</td><td>84.13</td><td>83.01</td><td>92.84</td><td>92.84</td></tr><tr><td>Emotion</td><td>41.65</td><td>42.57</td><td>93.00</td><td>92.92</td><td>74.40</td><td>74.42</td><td>93.30</td><td>93.36</td></tr><tr><td>BBC News</td><td>34.30</td><td>25.15</td><td>97.40</td><td>97.41</td><td>91.30</td><td>91.49</td><td>97.50</td><td>97.60</td></tr></table>
+
+Table 2: Fine-tuning performance (%) on four benchmarks using SFT, GRPO, and GTA across two model sizes (Base refers to the model without fine-tuning).
+
+![](images/4347cc5238af92f5fcabfb574d3b7b881789e346c36ae337448fc674d4690225.jpg)  
+Figure 4: Accuracy comparison of Qwen3 (4B) think and no think across multiple datasets
+
+methods by a significant margin. Similarly, on the Amazon dataset, GTA attains the highest accuracy and  $\mathrm{F_1}$  scores, indicating its robustness across different domains. These results underscore the effectiveness of the proposed GTA method in enhancing model performance across diverse datasets and model scales. The consistent improvements in both accuracy and  $\mathrm{F_1}$  scores highlight GTA's robustness.
+
+# 5.2 Exploring Performance Boundaries in Zero-Shot
+
+We first investigated whether reasoning-enhanced prompting could improve classification performance without model fine-tuning, as this would
+
+establish important baselines and validate a core premise of our GTA framework: that explicit reasoning steps can elevate model capabilities. This exploration addresses a fundamental question in LLM deployment—whether performance limitations stem from model capabilities themselves or from suboptimal reasoning processes that can be enhanced through structured prompting. Using the base Qwen3 (4B) model with its native "think" mode toggle, we compared standard direct answering against explicit reasoning-then-answering across four benchmarks: SST-5, Amazon, Emotion, and BBC News. Figure 4 demonstrates consistent performance improvements across all datasets when reasoning steps are incorporated. The accuracy on SST-5 increases from  $45.88\%$  to  $50.59\%$ , Amazon review classification improves from  $68.96\%$  to  $71.25\%$ , Emotion classification improves from  $51.15\%$  to  $58.00\%$ , and BBC News classification shows the most significant gain from  $80.40\%$  to  $90.00\%$ . These results reveal that models operating in a deliberate reasoning mode possess substantially higher performance ceilings than those constrained to direct response generation.
+
+# 5.3 Convergence Analysis of RL
+
+RL-based fine-tuning typically suffers from slow convergence due to lengthy exploration cycles. To assess how our GTA method addresses this limitation, we compared its convergence speed against
+
+![](images/00084a15327a4f757549455cbb74f8e0e2b4e8dfa8b1ea4b13a3c73f044ec91d.jpg)  
+(a) SST-5 (3B)
+
+![](images/20865649cf4ab21cce4226ce94b8d6e641d984126b5af2cf090ccd94007885a9.jpg)  
+(b) Amazon (3B)
+
+![](images/d520288942a3a8e0c5650be3620164b60508161ea6efdf55f397888780317f7d.jpg)  
+(c) Emotion (3B)
+
+![](images/6d4e2a653a1321c1531abe6450451d3d0206904fe829e9d544e249a6d0b5d997.jpg)  
+(d) SST-5 (4B)
+
+![](images/7d84c8c35ddc00486019aa74d30a2896fd07854aad06c002fae23f8de78d892f.jpg)  
+(e) Amazon (4B)
+
+![](images/9178812e1ffd66c85f52dc5d7b7dcd5a45aa93df02b16c0ef5d075d9d908f856.jpg)  
+(f) Emotion (4B)
+
+![](images/e9222ca87e74d13de4d99478e2ef2074046ee27147ac393a696a9a47d1311bee.jpg)  
+(a) SST-5 (3B) Accuracy Reward  
+Figure 6: Accuracy reward and test set accuracy trends of GRPO and GTA on SST-5
+
+![](images/09236163d5dece0a48e7018dc5c439c21a1017b38dc81a42cb148c389fb7d476.jpg)  
+Figure 5: Variations in accuracy reward values on the SST-5, Amazon, and Emotion datasets during GTA and GRPO fine-tuning  
+(b) Accuracy  $(\%)$  on the SST-5 (3B) Test Set
+
+the GRPO baseline across multiple datasets and model scales. Since BBC News is a relatively easy dataset with fewer categories and less training data, its convergence trends are less informative for analyzing RL training dynamics. We therefore focus our convergence analysis on SST-5, Amazon, and Emotion. Figure 5 illustrates the accuracy reward trajectories for both 3B and 4B models on SST-5, Emotion, and Amazon datasets, tracking three curves: GTA's answer segment (blue), GTA's guess segment (green), and GRPO (orange). On SST-5 and Emotion datasets, GTA demonstrates a substantial convergence advantage, with both guess and answer rewards consistently outperforming GRPO. While GRPO achieves faster initial convergence on the Amazon dataset, GTA maintains superior
+
+overall performance throughout training. These results confirm that the supervised guess segment effectively guides the answer segment, dramatically accelerating convergence while preserving the optimization benefits of RL—offering significant advantages for applying RL to language models in classification tasks.
+
+To investigate whether GRPO might eventually match GTA's performance given sufficient training time, we extended our experiments on the SST-5 dataset to 10,000 steps (approximately 10 epochs). Figure 6 presents this extended analysis using the 3B-parameter model, with Figure 6(a) showing training reward progression and Figure 6(b) depicting test accuracy evolution. Even in this extended training scenario, GRPO's convergence remains
+
+![](images/15f5b0cf44b98337a9003fcfc0eda19248de74b47a34dc273bc962b7d3025e2d.jpg)  
+Figure 7: Illustration of GTA reasoning: from incorrect guess to correct answer
+
+<table><tr><td></td><td>GRPO</td><td>GTA (RL)</td><td>GTA (SFT+RL)</td></tr><tr><td>Acc</td><td>58.60</td><td>57.42</td><td>61.58</td></tr><tr><td>F1</td><td>57.05</td><td>56.52</td><td>61.52</td></tr></table>
+
+Table 3: Performance  $(\%)$  of Qwen2.5 (3B) on GTA trained by RL on SST-5. GTA (RL) means that all stages are trained using RL, and GTA  $(\mathrm{SFT} + \mathrm{RL})$  indicates that the loss of the guess is calculated through SFT.
+
+substantially slower than GTA's. Despite showing gradual improvements in both reward and accuracy, GRPO fails to match GTA's performance even after 10,000 steps. These findings demonstrate that GTA not only accelerates training but also achieves a higher performance ceiling, offering significant practical advantages for deploying RL in language model fine-tuning for classification tasks.
+
+# 5.4 Ablation on Guess
+
+To validate the contribution of the supervised Guess stage, we replace the Guess component with RL updates. The results are shown in Table 3, We can observe that replacing supervised Guess with RL training does not improve the final accuracy. In fact, it yields a drop. This indicates the effectiveness of using supervised loss in the Guess stage.
+
+# 5.5 Reasoning Process Analysis
+
+We find that although making a guess under the guidance of supervision can accelerate convergence, the model does not blindly commit to the guessed answer as the final prediction. As show in Figure 7, when the model produces an incorrect guess, subsequent reasoning steps end to allow it to correct previous mistakes. In Case A, the model first predicts an incorrect label but gradually revises
+
+its reasoning and ultimately derives the correct answer. In Case B, the model not only outputs the correct final label but also explicitly indicates that the candidate label set does not include "sport", thereby mitigating hallucination issues commonly observed in SFT. These examples, together with the overall accuracy improvements, highlight that GTA exhibits stronger robustness than purely SFT methods that only optimize the Guess segment.
+
+# 6 Conclusion and Future Work
+
+In this work, we present GTA, a novel training framework that addresses the efficiency-capability trade-off between SFT and RL by introducing a Guess stage to traditional CoT outputs. Under this framework, model outputs are organized into Guess, Think, and Answer segments, where the Guess is optimized via SFT while the overall format and final output are optimized through RL. To mitigate gradient conflicts, we apply loss masking and cosine-similarity constraints. Experimental results on four text-classification benchmarks show that GTA consistently outperforms pure SFT and GRPO baselines in accuracy and  $\mathrm{F_1}$  scores while achieving significantly faster convergence than pure RL approaches. Through analysis of training dynamics, we demonstrate that GTA successfully combines the efficiency of supervised learning with the performance gains of RL, substantially accelerating convergence and addressing the exploration inefficiency inherent to RL-based fine-tuning. Theoretically, our approach is not limited to text classification tasks, and we plan to explore extending GTA to broader NLP tasks in future work.
+
+# Limitations
+
+While publicly available text classification datasets often contain noisy data, we did not perform preprocessing or sample high-quality subsets in this study. Additionally, due to resource constraints, we validated our proposed method only on 3B and 4B parameter models, without extending the evaluation to larger-scale models. Given that RL heavily relies on the underlying model's capacity, more powerful and generalizable models may yield greater benefits.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, and Shyamal et al. Anadkat. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Ju-Seung Byun, Jiyun Chun, Jihyung Kil, and Andrew Perrault. 2024. Ares: Alternating reinforcement learning and supervised fine-tuning for enhanced multi-modal chain-of-thought reasoning through diverse ai feedback. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4410-4430.  
+Junfan Chen, Richong Zhang, Junchi Chen, and Chunming Hu. 2024. Open-set semi-supervised text classification via adversarial disagreement maximization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2170-2180.  
+Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, and Fan et al. Yang. 2025a. Research: Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470.  
+Si-An Chen, Hsuan-Tien Lin, and Chih-Jen Lin. 2025b. Preserving zero-shot capability in supervised finetuning for multi-label text classification. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 5699-5712.  
+Yun-Nung Chen, Dilek Hakkani-Tür, and Xiaodong He. 2016. Zero-shot learning of intent embeddings for expansion by convolutional deep structured semantic models. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6045-6049. IEEE.  
+Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. 2018. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR.
+
+Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Jack FitzGerald, Christopher Hench, Charith Peris, Scott Mackie, Kay Rottmann, Ana Sanchez, Aaron Nash, Liam Urbach, Vishesh Kakarala, and Richa et al. Singh. 2022. Massive: A 1m-example multilingual natural language understanding dataset with 51 typologically-diverse languages. arXiv preprint arXiv:2204.08582.  
+Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, and Alex et al. Vaughan. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv-2407.  
+Derek Greene and Pádraig Cunningham. 2006. Practical solutions to the problem of diagonal dominance in kernel document clustering. In Proceedings of the 23rd international conference on Machine learning, pages 377-384.  
+Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, and Xiao et al. Bi. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.  
+Thorsten Joachims. 1998. Text categorization with support vector machines: Learning with many relevant features. In European conference on machine learning, pages 137-142. Springer.  
+Rie Johnson and Tong Zhang. 2015. Effective use of word order for text categorization with convolutional neural networks. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL HLT 2015, pages 103-112. Association for Computational Linguistics (ACL).  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.  
+Arina Kostina, Marios D Dikaiakos, Dimosthenis Stefanidis, and George Pallis. 2025. Large language models for text classification: Case study and comprehensive review. arXiv preprint arXiv:2501.08457.  
+Rakesh R Menon and Shashank Srivastava. 2024. Discern: Decoding systematic errors in natural language for text classifiers. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 19565-19583.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, and Alex et al. Ray. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744.
+
+Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? sentiment classification using machine learning techniques. In Proceedings of the ACL-02 conference on Empirical methods in natural language processing-Volume 10, pages 79-86.  
+Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. 2018. Carer: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 3687-3697.  
+Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, and Y et al. Wu. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642.  
+Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansoresh Karami, Jundong Li, Lu Cheng, and Huan Liu. 2024. Large language models for data annotation and synthesis: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 930-957.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.  
+Shuhe Wang, Shengyu Zhang, Jie Zhang, Runyi Hu, Xiaoya Li, Tianwei Zhang, Jiwei Li, Fei Wu, Guoyin Wang, and Eduard Hovy. 2024. Reinforcement learning enhanced llms: A survey. arXiv preprint arXiv:2412.10400.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, and Denny et al. Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Yu Xia, Rui Wang, Xu Liu, Mingyan Li, Tong Yu, Xiang Chen, Julian McAuley, and Shuai Li. 2025. Beyond chain-of-thought: A survey of chain-of-x paradigms for llms. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10795-10809.  
+Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, and Fanjin et al. Meng. 2025. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint arXiv:2501.09686.
+
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, and Haoran et al. Wei. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.  
+Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836.

2025/GTA_ Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:642d205af0c35e49fc85759658b08cd1ed94afa6943d37aa4c53099f131fc380
+size 702576

2025/GUARD_ Glocal Uncertainty-Aware Robust Decoding for Effective and Efficient Open-Ended Text Generation/b5c8441d-feae-4313-aa6c-227be1980008_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48ba31c8d2105a80e9cb324e91a262f7f212c3fbeffd0b498347db1be40e10dc
+size 639995

2025/GUARD_ Glocal Uncertainty-Aware Robust Decoding for Effective and Efficient Open-Ended Text Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fe061ffe496f5e605126eafe755db68706f91a3405afa808e5460b3d2406046
+size 1972249

2025/GeLoRA_ Geometric Adaptive Ranks For Efficient LoRA Fine-tuning/408657ff-3659-4887-abb9-ebd6df5c83fb_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc9800eba415606fd46fcbefcd1c919c111763d3ccb8e0d9e133be3d1dc2a2a5
+size 881944

2025/GeLoRA_ Geometric Adaptive Ranks For Efficient LoRA Fine-tuning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e81898fd4c538c0e97bcc0557f44f34d5b6e76708df43dcea70f37874ca7c140
+size 923900

2025/GenPTQ_ Green Post-Training Quantization for Large-Scale ASR Models with Mixed-Precision Bit Allocation/99e69a29-243d-4d97-908f-6d69c507862a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:287935780a24025b55505fc432a5bee5a3193664bd0b4e668aa025186443079d
+size 1375977

2025/GenPTQ_ Green Post-Training Quantization for Large-Scale ASR Models with Mixed-Precision Bit Allocation/full.md

@@ -0,0 +1,360 @@
+# GenPTQ: Green Post-Training Quantization for Large-Scale ASR Models with Mixed-Precision Bit Allocation *
+
+Beom Jin Kang $^{1}$ , Hyun Kim $^{\dagger,1}$
+
+$^{1}$  Department of Electrical and Information Engineering, Research Center for EIT Seoul National University of Science and Technology, Seoul, Korea {beomjin, hyunkim}@seoultech.ac.kr
+
+# Abstract
+
+Large-scale models have achieved state-of-the-art performance in automatic speech recognition (ASR), but their high memory and computation demands pose significant challenges for deployment. To address these challenges, weight-only quantization is widely adopted in large-scale models, where weights dominate memory usage, as it enables efficient compression with minimal accuracy degradation compared to activation quantization. Accordingly, most prior quantization studies for ASR models have focused on weights and employed quantization-aware training (QAT) to restore accuracy. However, QAT incurs substantial additional training costs, posing clear limitations for practical application to large-scale models. Moreover, despite the varying quantization sensitivity across layers, mixed-precision quantization (MPQ) remains underexplored in ASR. In this paper, we propose GenPTQ, a mixed-precision post-training quantization method that optimizes the trade-off among accuracy, model size, and optimization cost by leveraging gradient-based sensitivity measurement and transforming the search space into a continuous domain for efficient numerical optimization. Applied to Whisper and Conformer models across multiple speech datasets, GenPTQ achieves up to  $89.1\%$  model size reduction (2.5-bit average precision) with only a  $0.8\%$  increase in WER, and completes optimization in just 15 seconds. These results demonstrate its effectiveness for low-resource ASR deployment.
+
+# 1 Introduction
+
+Recently, large-scale end-to-end models (Radford et al., 2023; Gulati et al., 2020; Peng et al., 2022) based on the transformer architecture (Vaswani et al., 2017) have achieved state-of-the-art (SOTA) recognition accuracy in automatic speech recognition (ASR), driving their widespread adoption.
+
+tion across various speech processing applications (Zheng et al., 2024; Bhagtani et al., 2024). However, as model size increases, memory usage and computational cost grow substantially, leading to excessive power consumption and inference latency in real-world deployment scenarios (Lee et al., 2024; Wagner et al., 2024). These issues extend beyond efficiency concerns, presenting a critical challenge in terms of environmental impact and degradation of user experience (Aquino-Britez et al., 2025; Li et al., 2024a). Accordingly, various model compression methods have been actively explored to reduce the size and computational cost of ASR models, while preserving recognition accuracy (Kim and Kim, 2022; Kim et al., 2024; Choi and Kim, 2025). Quantization reduces the bitprecision of weights or activations—originally represented in 32-bit floating-point format—thereby effectively reducing the overall model size (Kim et al., 2022; Gholami et al., 2022). In particular, for large-scale models such as Whisper (Radford et al., 2023), memory bandwidth often emerges as the primary bottleneck rather than computational cost. Consequently, reducing weight bit-precision is crucial for improving practical efficiency (Kim et al., 2023). Moreover, as weights consume over  $3.4 \times$  more memory than activations, applying lower-bit quantization to weights is more efficient, especially given that activations tend to be more sensitive to accuracy degradation. Accordingly, most ASR quantization studies have primarily focused on weight-only quantization, reflecting this characteristic (Rybakov et al., 2023).
+
+Quantization methods are generally categorized into post-training quantization (PTQ) and quantization-aware training (QAT) (Kim and Kim, 2025; Choi et al., 2024). PTQ applies quantization directly to pre-trained weights without requiring retraining, making it attractive for its simplicity and efficiency. However, the rounding involved in converting floating-point values to lower-precision
+
+patterns can introduce quantization noise, leading to potential performance degradation. To address this, QAT jointly optimizes model weights and quantization parameters (e.g., scaling factors) during training, which allows the model to better adapt to quantization and typically results in higher accuracy (Rybakov et al., 2023; Ding et al., 2022). As a result, most prior ASR quantization studies have focused on QAT. Nevertheless, QAT imposes substantial training overhead, often requiring tens of GPU-hours. For example, applying QAT to the Conformer-large model demands over 84 hours of training (Gulati et al., 2020), which presents a major barrier for practical deployment, especially in user-adaptive or low-resource settings. Furthermore, as model sizes continue to grow, the training cost of QAT is expected to scale accordingly (Guan et al., 2024; Hasan, 2024). These limitations highlight the need for PTQ methods that offer a better trade-off between model size and accuracy, without incurring the high cost of full retraining.
+
+Meanwhile, mixed-precision quantization (MPQ) mitigates accuracy degradation by assigning optimal bit-precision to each layer based on its quantization sensitivity (Ranjan and Savakis, 2025; Dong et al., 2019; Xu et al., 2025). While MPQ has shown strong potential, its application is limited by the prohibitively large search space involved in identifying layer-wise precision configurations. This results in significant computational and optimization overhead, posing a major barrier to practical adoption, especially in large-scale ASR models. Consequently, these challenges have slowed the progress of MPQ research in the ASR domain, despite its theoretical advantages.
+
+To address the limitations of prior ASR quantization methods, we propose GenPTQ, a novel mixed-precision PTQ approach that achieves an improved trade-off among accuracy, model size, and optimization efficiency. GenPTQ efficiently allocates per-layer bit precision with minimal optimization overhead, achieving robust quantization performance across various ASR models and datasets—all without requiring additional training. The contributions of this work are summarized as follows:
+
+- Gradient-Based Layer-Wise Sensitivity Measurement: We quantitatively assess the impact of quantization perturbation on model accuracy by leveraging the gradient information of each layer. This metric guides the allocation of layer-wise optimal bit precision,
+
+thereby enabling effective optimization of the trade-off between accuracy and model size.
+
+- Low-cost Bit-Precision Allocation Method: We propose a method to minimize the optimization cost associated with bit-precision search. Specifically, we introduce a continuous bit-precision representation (CBPR) based on layer-wise sensitivity values to parameterize bit-precision configurations. By exploiting the differentiability of CBPR, we design a loss function that enforces a target average precision constraint. Additionally, we introduce a sensitivity regularization (SR) to prevent excessive bit-precision reduction in highly sensitive layers. This approach enables the effective derivation of optimal bit-precision configurations with minimal optimization overhead.
+
+- Evaluation of Generalization and Scalability: We applied GenPTQ to both the Whisper and Conformer models to evaluate its performance. Across various speech datasets, the method maintained an average weight bit precision of 2.5 bits, reducing model size by  $89.1\%$ , with only an average increase of  $0.8\%$  in word error rate (WER) and  $0.3\%$  in character error rate (CER).
+
+# 2 Related Works
+
+# 2.1 Quantization
+
+Overview of Quantization: Quantization reduces the model size by converting weights and activations, originally represented in 32-bit floating-point format, into low-bit precision (e.g., INT2, INT4). In general, min-max uniform quantization is employed, which can be formulated as follows:
+
+$$
+x _ {q} = \operatorname {c l a m p} \left(\left\lfloor \frac {x}{s _ {f}} \right\rfloor , - 2 ^ {b - 1}, 2 ^ {b - 1} - 1\right) \tag {1}
+$$
+
+$$
+w h e r e s _ {f} = \frac {\alpha - \beta}{2 ^ {b} - 1}
+$$
+
+Here,  $\alpha$  and  $\beta$  denote the maximum and minimum values of a 32-bit floating-point input  $x$ , respectively. The operator  $\lfloor \cdot \rfloor$  represents the rounding function, and  $b$  indicates the bit-precision. Through this process,  $x$  is mapped to a low-bit integer  $x_{q}$  within the range  $[-2^{b - 1}, 2^{b - 1} - 1]$ . Dequantization is then applied to restore the value to the original floating-point domain, as follows:
+
+$$
+\tilde {x} = s _ {f} \cdot x _ {q} \tag {2}
+$$
+
+Quantization inevitably introduces perturbation between the original value  $x$  and its quantized representation  $\tilde{x}$  due to the rounding operation, which can potentially lead to accuracy degradation. Notably, as bit-precision decreases, model size can be significantly reduced; however, the representational range of the quantized value  $x_{q}$  becomes increasingly limited, leading to greater accuracy degradation. Accordingly, we focus on developing a methodology that effectively improves the tradeoff between model size and accuracy.
+
+Previous Quantization Methods for ASR Models: To date, most quantization methods for ASR models have been developed within the QAT paradigm. For example, (Ding et al., 2022) applied naive 4-bit QAT to the Conformer, a prominent architecture in ASR, achieving minimal degradation in recognition accuracy. Similarly, (Rybakov et al., 2023) proposed a sub-channel asymmetric, weight-only 2-bit QAT method to further improve recognition accuracy. Furthermore, (Kim et al., 2022) proposed a data-free quantization method that synthesizes mel-spectrograms using batch normalization statistics. Although QAT-based quantization methods have demonstrated high recognition accuracy, their application to large-scale ASR models such as Whisper remains limited due to the substantial training cost, posing challenges for practical deployment. Motivated by the limitations, we propose a PTQ method that minimizes accuracy degradation without requiring additional training, even for large-scale ASR models.
+
+Previous PTQ Methods for Other Applications: AdaRound (Nagel et al., 2020) introduced an adaptive rounding method that combines Hessian-based layer-wise task loss approximation with soft relaxation. The method was validated on various convolutional neural networks (CNNs), including ResNet (He et al., 2016) and MobileNetV2 (Sandler et al., 2018), using the ImageNet (Deng et al., 2009) dataset. In addition, BRECQ (Li et al., 2021) proposed a PTQ method based on Hessian-guided block-wise weight reconstruction, targeting various CNNs. Although AdaRound and BRECQ were primarily developed for computer vision tasks, both methods leverage Hessian-based optimization frameworks to precisely control quantization perturbations at the layer or block level. This design plays a critical role in determining the overall performance of PTQ. Nevertheless, these methods typically perform layer- or block-wise optimization and involve complex Hessian-based computations,
+
+which lead to substantial optimization costs. Consequently, these methods face practical limitations when directly applied to large-scale ASR models such as Whisper. To emphasize the suitability of the proposed method for ASR tasks, we conduct a comparative evaluation against AdaRound and BRECQ by applying them to ASR models and assessing both efficiency and accuracy.
+
+# 2.2 Mixed Precision Quantization
+
+The primary objective of MPQ is to assign the optimal bit-precision to each layer based on its sensitivity, in order to minimize accuracy degradation under a given target bit-precision constraint. Specifically, higher bit-precisions (e.g., INT6, INT8) are allocated to sensitivity-critical layers to preserve accuracy, while lower bit-precisions (e.g., INT2, INT4) are assigned to less sensitive layers to aggressively reduce model size. This enables effective optimization of the trade-off between accuracy and model size. For example, HTQ (Li et al., 2024b) evaluates the importance of each layer based on connection sensitivity and assigns optimal bit-precision through Pareto frontier exploration. AMPA (Ding et al., 2024) performs sensitivity analysis by integrating gradient magnitude and quantization loss, applying a thresholding mechanism for dynamically bit-precision assignment. However, these methods are specifically tailored for CNNs and face limitations when applied to the structurally more complex transformer-based ASR models. Meanwhile, (Xu et al., 2025) proposed a Gumbel-softmax (Maddison et al., 2016)-based mixed-precision QAT method for wav2vec2. 0(Baevski et al., 2020) and HuBERT-large (Hsu et al., 2021). However, these methods commonly involve complex bit-precision search spaces, resulting in high optimization costs during bit allocation. Such issues become increasingly pronounced for large-scale models, posing a significant barrier to practical deployment in real-world systems. To address this, we propose a mixed-precision PTQ method that efficiently derives an optimal bit-precision configuration with extremely low optimization cost.
+
+# 3 Proposed Methods
+
+# 3.1 Gradient-based Sensitivity Metric
+
+In general, layer-wise sensitivity can be estimated using quantization perturbation based on mean square error (MSE). Specifically, for a layer  $L_{t}$  where  $t\in \{1,2,3,\dots,N\}$ , with a weight matrix
+
+$w\in \mathbb{R}^{m\times n}$ , the MSE-based layer-wise quantization perturbation  $\Phi_t$  is defined as follows:
+
+$$
+\Phi_ {t} = \frac {1}{m n} \sum_ {i = 1} ^ {m} \sum_ {j = 1} ^ {n} \left(\tilde {w} _ {i, j} - w _ {i, j}\right) ^ {2} \tag {3}
+$$
+
+That is,  $\Phi_t$  quantitatively measures the perturbation introduced by quantization in layer  $L_t$ , which guides the assignment of an appropriate bit-precision. However, even with the same level of quantization perturbation, the impact of each layer on the overall accuracy can vary significantly. That is, even when the quantization perturbation is small, certain layers may still have a considerable impact on overall accuracy. This highlights the need for a fine-grained sensitivity metric that can accurately reflect such effects. Therefore, we evaluate the relative importance of each layer to model accuracy by leveraging the mean gradient of the weights with respect to the task loss function  $\mathcal{L}_{\mathrm{task}}$ , as follows:
+
+$$
+g _ {t} (w) = \frac {1}{K} \sum_ {i = 1} ^ {K} \left| \frac {\partial \mathcal {L} _ {t a s k}}{\partial w _ {i}} \right| \tag {4}
+$$
+
+Here,  $g_{t}(w)$  denotes the average absolute gradient of the weights in layer  $L_{t}$ , which consists of  $K$  weight elements. The operator  $\partial$  represents the partial derivative of the task loss function  $\mathcal{L}_{\mathrm{task}}$  with respect to each weight element  $w_{i}$ . This metric quantitatively reflects the relative influence of the weights in layer  $L_{t}$  on the task loss  $\mathcal{L}_{\mathrm{task}}$  (Chauhan et al., 2023). Based on this gradient  $g_{t}(w)$ , we define a more precise sensitivity metric as follows:
+
+$$
+S _ {t} = \frac {1}{K} \sum_ {i = 1} ^ {K} \left(\left| \frac {\partial \mathcal {L} _ {t a s k}}{\partial w _ {i}} \right| (\tilde {w} _ {i} - w _ {i}) ^ {2}\right) \tag {5}
+$$
+
+The proposed sensitivity metric is formulated as the product of the quantization perturbation and  $g_{t}(w)$ , enabling it to capture not only the magnitude of quantization error but also the influence of each layer's weights on model accuracy. Specifically, even when the quantization perturbation is small, a high bit-precision can be assigned if the corresponding weights have a significant impact on the task performance. This makes the metric a crucial indicator for guiding the search for bit-precision configurations that achieve optimal accuracy.
+
+# 3.2 Continuous Bit-Precision Representation
+
+The sensitivity  $S_{t}$ , derived from Eq. (5), is used to select the optimal bit-precision for each
+
+layer from a predefined candidate set  $b \in \{\mathrm{INT2, INT3, \ldots, INT8}\}$ . However, since  $b$  is a discrete variable, it is non-differentiable, and the optimal bit-precision is typically determined using search-based algorithms (Kloberdanz and Le, 2023; Tai et al., 2024). To reduce the computational cost associated with such search procedures, we propose CBPR, which transforms layer-wise bit-precision into continuous and differentiable variables, enabling efficient numerical optimization via a loss function. To this end, we normalize  $S_{t}$  and define CBPR within the range bounded by the minimum and maximum bit-precisions,  $b_{\mathrm{min}}$  and  $b_{\mathrm{max}}$ , as follows:
+
+$$
+P _ {t} = b _ {\text {m i n}} + \left(1 - \frac {S _ {t} - S _ {\text {m i n}}}{S _ {\text {m a x}} - S _ {\text {m i n}}}\right) \left(b _ {\text {m a x}} - b _ {\text {m i n}}\right) \tag {6}
+$$
+
+Here,  $S_{\mathrm{min}}$  and  $S_{\mathrm{max}}$  denote the minimum and maximum sensitivity values across all layers, respectively. Accordingly, the initial value of  $P_{t}$  is determined based on the sensitivity distribution, such that layers with higher sensitivity are mapped to higher bit precisions. The computed  $P_{t}$  values are then used to assign bit-precisions to each layer, and the average bit-precision of the model,  $b_{m}$ , is defined as follows:
+
+$$
+b _ {m} = \frac {1}{N} \sum_ {t = 1} ^ {N} \left\lfloor P _ {t} \right\rceil \tag {7}
+$$
+
+The compression ratio of a model is commonly evaluated using the average bit-precision  $b_{m}$ . However, the initial value of  $b_{m}$  may not satisfy the target bit-precision constraint  $b_{tg}$ . To address this, we exploit the differentiability of the continuous variable  $P_{t}$  to design an estimable loss function. This enables the derivation of an optimal bit-precision configuration  $\hat{P}_{t}$  that minimizes performance degradation while adhering to the target bit-precision constraint.
+
+# 3.3 Mixed-Precision Bit Allocation with Sensitivity Regularization
+
+To ensure that the average bit-precision  $b_{m}$  satisfies the target constraint  $b_{tg}$ , we formulate the following loss function  $\mathcal{L}_p$ :
+
+$$
+\mathcal {L} _ {p} = \left| \left(\frac {1}{N} \sum_ {t = 1} ^ {N} \lfloor P _ {t} \rceil\right) - b _ {t g} \right| \tag {8}
+$$
+
+By optimizing  $P_{t}$  with respect to  $\mathcal{L}_p$ , the average bit-precision of the model can be adjusted to converge toward the target constraint  $b_{tg}$ . However, for
+
+![](images/09f058452bc00993ae48bdf244866796b373cdd162ee09bdc7faf9c2c74a00b9.jpg)  
+Figure 1: Overview of the Whisper Model Architecture and the Proposed GenPTQ Framework.
+
+![](images/f6f7e4a52a167bddfb0390d7bdb97c6a8e8db878cf6790e7a99a6d87a466aef2.jpg)
+
+![](images/76a01ba90d0b585788e349d57b0487e0ba3f6eac6dfc5b276064514fdebcb81e.jpg)
+
+layers with high sensitivity, the initial value of  $P_{t}$  tends to be large, which may result in excessive reduction during the gradient descent. Consequently, low bit-precision could be assigned to highly sensitive layers, leading to accuracy degradation. To mitigate this, we introduce an SR term as follows:
+
+$$
+\mathcal {L} _ {S R} = \sum_ {t = 1} ^ {N} S _ {t} \left(b _ {\max } - P _ {t}\right) \tag {9}
+$$
+
+The final loss function, including the SR term, is defined as follows:
+
+$$
+\mathcal {L} _ {m p q} = \lambda_ {1} \mathcal {L} _ {p} + \lambda_ {2} \mathcal {L} _ {S R} \tag {10}
+$$
+
+The loss function  $\mathcal{L}_{mpq}$  mitigates the assignment of excessively low bit-precisions to highly sensitive layers during the optimization of  $P_{t}$ , thereby guiding the search toward an optimal layer-wise bit-precision configuration  $\hat{P}_t$  that balances the trade-off between accuracy and model size.
+
+# 3.4 The Overall Proposed Framework
+
+Figure 1 presents a visual overview of the overall GenPTQ framework, while Algorithm 1 outlines its complete operational procedure. We begin by computing the gradient of the task loss and calculating the layer-wise sensitivity  $S_{t}$  using Eq. (5) (Algorithm1, Line 7). Subsequently, to minimize optimization cost, CBPR is derived from the sensitivity values (Algorithm 1, Line 11), followed by numerical optimization guided by a loss function to obtain the optimal layer-wise bit-precision configuration that satisfies the target bit constraint (Algorithm 1, Lines 13-18). In this process, the final loss function  $\mathcal{L}_{\mathrm{mpq}}$ , incorporating the SR term, suppresses the assignment of low-precision bits to highly sensitive layers and facilitates the derivation of the optimal bit allocation  $\hat{P} t$  that adheres
+
+# Algorithm 1 GenPTQ Framework
+
+1: Input: Pre-trained weights  $w$ , Target bit-precision constraints  $b_{tg}$ , Initial bit-precision  $b$ , Scaling factor  $s_f$ , Optimization iteration  $iter$ , Hyper parameter  $\lambda_1, \lambda_2$ , Total number of layers  $N$ , Learning rate  $\eta$  
+2: for  $L_{t}$  in Layers do  
+3: /* Weight quantization & dequantization*/  
+4:  $w_{q} = clamp(\left\lfloor \frac{w}{s_{f}} \right\rfloor, -2^{b}, 2^{b} - 1)$  
+5:  $\tilde{w} = s_f\cdot w_q$  
+6:  $/^{*}$  Compute gradient-based sensitivity using Eq. (5)\*/  
+7:  $S_{t} = \frac{1}{K}\sum_{i = 1}^{K}\left(\left|\frac{\partial\mathcal{L}_{task}}{\partial w_{i}}\right|(\tilde{w}_{i} - w_{i})^{2}\right)$  
+8: end for  
+9: for  $t$  in  $N$  do  
+10:  $/*$  Calculate CBPR using Eq. (6)*/  
+11:  $P_{t} = b_{min} + (1 - \frac{S_{t} - S_{max}}{S_{max} - S_{min}})(b_{max} - b_{min})$  
+12: end for  
+13: for  $i$  in iter do  
+14: /\* Calculate MPQ loss using Eq.(10)\*/  
+15:  $\mathcal{L}_{mpq} = \lambda_1\mathcal{L}_p + \lambda_2\mathcal{L}_{SR}$  
+16: /\* Update  $P_{t}$  based on backpropagation\*/  
+17:  $P_{t}\gets P_{t} - \eta \cdot \frac{\partial L_{mpq}}{\partial P_{t}}$  
+18: end for  
+19: # Get optimal bit-precision  $\tilde{P}_t$  
+20:  $\hat{P}_t = P_t$  
+21: for  $L_{t}$  in Layers do  
+22: /\* Assign optimal bit precision per Layer\*/  
+23:  $b_{opt} = \left\lfloor \hat{P}_t\right]$  
+24: /* Scaling Factor Update/  
+25:  $s_f = \frac{\alpha - \bar{p}}{2^{b_{opt} - 1}}$  
+26:  $/^{*}$  Weight quantization & dequantization \*/  
+27:  $w_{q} = clamp(\left|\frac{w}{s_{f}}\right|, - 2^{b_{opt}},2^{b_{opt}} - 1)$  
+28:  $\tilde{w} = s_f\cdot w_q$  
+29: end for  
+30: Return  $\tilde{w}$
+
+to the target constraint  $b_{tg}$ . A key advantage of the proposed method is that it requires only a small calibration dataset for gradient computation and achieves efficient bit-precision allocation without relying on combinatorial search, resulting in minimal optimization overhead. This makes it a practical solution for MPQ, effectively reducing optimization cost while achieving a balanced trade-off between accuracy and model size.
+
+# 4 Experimental Results
+
+# 4.1 Experimental Environment
+
+We evaluate the performance of GenPTQ on the Whisper-medium and Conformer-large models. The pre-trained weights for Whisper-medium were loaded from Hugging Face (Wolf et al., 2020), while those for Conformer-large were generated by training the model using the ESPnet toolkit (Watanabe et al., 2018). All experiments were conducted on a single H100 GPU. During optimization, the learning rate was fixed at 0.1 with 150 iterations. To simplify the setup, the hyperparameters  $\lambda_{1}$  and  $\lambda_{2}$  in Eq. (10) were both set to 1. Following the configuration in (Rybakov et al., 2023), quantization was implemented at the block-wise granularity.
+
+# 4.2 Performance Evaluation on Various English Speech Datasets
+
+We apply the proposed method to the Whisper-medium model and evaluate its recognition error (i.e., WER and CER) and generalization performance across multiple English speech datasets, including Librispeech test-clean (Panayotov et al., 2015), CommonVoice-test (Ardila et al., 2019), VoxPopuli-test (Wang et al., 2021), Tedmium-test (Rousseau et al., 2012), GigaSpeech-test (Chen et al., 2021), and SPGISpeech-test (O'Neill et al., 2021).
+
+Table 1 compares the performance of the 32-bit floating-point baseline, a standard min-max PTQ method (MinMax), and the proposed GenPTQ (Ours). Here, the bit-precision of Ours refers to the average bit-precision across the entire model. As shown in Table 1, our method achieves an average bit-precision of 2.5-bit while maintaining WER and CER increases within  $1\%$  across all datasets except CommonVoice. On CommonVoice, relatively higher increases of  $2.4\%$  in WER and  $1.1\%$  in CER are observed. This can be attributed to the higher level of noise in CommonVoice, which makes the model more susceptible to quantization-induced perturbations compared to cleaner datasets such as Librispeech test-clean. Nevertheless, the proposed method maintains high recognition accuracy on most datasets, demonstrating robust generalization capability across diverse domains. Notably, it achieves a significantly reduced model size of 323.2MB while maintaining WER and CER comparable to MinMax-based INT8 and INT4 PTQ. Furthermore, it consistently outperforms MinMax-based INT2 PTQ across all datasets in terms of
+
+recognition accuracy. These results indicate that the proposed method offers a highly effective tradeoff between model size and accuracy, making it well-suited for practical deployment in memory-constrained ASR systems.
+
+Additionally, to demonstrate the portability of the proposed method, we evaluate its performance not only on the Whisper model but also on the Wav2Vec2 (Baevski et al., 2020) model. Specifically, we load the pre-trained Wav2Vec2 weights from Hugging Face, apply GenPTQ, and evaluate the performance on the LibriSpeech test-clean dataset. Table 2 presents the performance of GenPTQ on the Wav2Vec2-L model. Consistent with the results observed on the Whisper model, GenPTQ achieves substantially higher speech recognition accuracy at an effective precision of 2.5 bits, compared to the conventional MinMax 2-bit baseline. These findings demonstrate that the proposed method is not only effective for Whisper but also generalizable to diverse speech recognition models.
+
+# 4.3 Performance Evaluation on Conformer
+
+We apply GenPTQ to the Conformer-large model and compare it with SOTA QAT methods on the Librispeech test-clean dataset in terms of training cost, accuracy, and model size trade-offs. Since baseline WER values vary across methods, accuracy is compared based on the relative change in WER. Accordingly, in Table 3, Baseline<sup>1</sup> denotes the baseline performance used in prior works, while Baseline<sup>2</sup> represents the baseline performance of our method. Additionally, the training time for QAT methods is measured over 70 epochs, following the official training configuration of the baseline Conformer-large model.
+
+As shown in Table 3, our method incurs only a  $0.3\%$  increase in WER when applying 2-bit precision to the Conformer model, representing a marginal increase compared to the SOTA QAT-based method (Rybakov et al., 2023) at the same bit precision. However, it is noteworthy that QAT requires over 84 hours of training, while our method achieves comparable performance with just 7.8 seconds of optimization—focused solely on CBPR tuning rather than weight updates—demonstrating the minimal overhead of our approach. These results demonstrate that our method achieves a favorable trade-off between optimization cost and accuracy. Moreover, compared to prior methods (Kim et al., 2022; Ding et al., 2022), our method in
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Prec.</td><td rowspan="2">Size(MB)</td><td colspan="2">Librispeech</td><td colspan="2">CommonVoice</td><td colspan="2">Voxpopuli</td><td colspan="2">TedLium</td><td colspan="2">GigaSpeech</td><td colspan="2">SPGISpeech</td></tr><tr><td>WER</td><td>CER</td><td>WER</td><td>CER</td><td>WER</td><td>CER</td><td>WER</td><td>CER</td><td>WER</td><td>CER</td><td>WER</td><td>CER</td></tr><tr><td rowspan="2">Baseline</td><td>FP32</td><td>2955.9</td><td>4.1</td><td>1.6</td><td>11.4</td><td>5.5</td><td>10.2</td><td>6.0</td><td>11.9</td><td>4.5</td><td>12.6</td><td>7.5</td><td>3.7</td><td>1.8</td></tr><tr><td>INT8</td><td>864.5</td><td>4.2</td><td>1.6</td><td>11.5</td><td>5.5</td><td>10.2</td><td>6.1</td><td>12.0</td><td>4.8</td><td>12.8</td><td>7.6</td><td>3.7</td><td>1.8</td></tr><tr><td rowspan="2">MinMax</td><td>INT4</td><td>489.2</td><td>4.2</td><td>1.7</td><td>11.7</td><td>5.6</td><td>10.3</td><td>6.1</td><td>12.2</td><td>5.0</td><td>12.7</td><td>7.6</td><td>3.8</td><td>1.9</td></tr><tr><td>INT2</td><td>301.5</td><td>15.5</td><td>9.8</td><td>28.7</td><td>16.2</td><td>28.8</td><td>19.8</td><td>27.8</td><td>18.0</td><td>43.4</td><td>37.0</td><td>19.2</td><td>13.2</td></tr><tr><td>Ours</td><td>INT2.5</td><td>323.2</td><td>4.3</td><td>1.7</td><td>13.8</td><td>6.6</td><td>11.0</td><td>6.2</td><td>12.3</td><td>4.7</td><td>12.8</td><td>7.6</td><td>4.4</td><td>2.1</td></tr></table>
+
+Table 1: Performance comparison of quantization methods across various datasets (WER and CER in %).  
+
+<table><tr><td>Model</td><td>Method</td><td>Prec.</td><td>WER(↓)</td><td>CER(↓)</td></tr><tr><td rowspan="5">Wav2Vec2-L (Baevski et al., 2020)</td><td>Baseline</td><td>FP32</td><td>11.9</td><td>2.9</td></tr><tr><td rowspan="3">MinMax</td><td>INT8</td><td>12.0</td><td>2.9</td></tr><tr><td>INT4</td><td>12.5</td><td>2.7</td></tr><tr><td>INT2</td><td>21.9</td><td>5.1</td></tr><tr><td>Ours</td><td>INT2.5</td><td>12.6</td><td>2.7</td></tr></table>
+
+Table 2: Performance of quantization methods on Wav2Vec2-Large using the LibriSpeech dataset  
+
+<table><tr><td>Method</td><td>Prec.</td><td>Training</td><td>Training Time</td><td>Size (MB)</td><td>WER(↓)</td><td>Diff.</td></tr><tr><td>Baseline1</td><td>FP32</td><td>-</td><td>-</td><td>474.5</td><td>2.0</td><td>-</td></tr><tr><td>[1]</td><td>INT8</td><td></td><td></td><td>123.7</td><td>3.1</td><td>+1.1</td></tr><tr><td>[2]</td><td>INT8</td><td rowspan="3">✓</td><td rowspan="3">84h 11m</td><td>138.1</td><td>2.0</td><td>-</td></tr><tr><td>[3]</td><td>INT4</td><td>81.9</td><td>2.0</td><td>-</td></tr><tr><td>[4]</td><td>INT2</td><td>55.3</td><td>2.0</td><td>-</td></tr><tr><td>Baseline2</td><td>FP32</td><td>-</td><td>-</td><td>475.2</td><td>2.2</td><td>-</td></tr><tr><td rowspan="2">Ours</td><td>INT2</td><td rowspan="2">-</td><td rowspan="2">7.8s</td><td>58.6</td><td>2.5</td><td>+0.3</td></tr><tr><td>INT2.5</td><td>62.6</td><td>2.2</td><td>-</td></tr></table>
+
+curs only a slight increase in WER under lower bit-precision settings, despite requiring no additional training. Notably, at the 2.5-bit setting, our method achieves optimal accuracy with no increase in WER, matching the performance of SOTA QAT methods. These results highlight that our method offers a superior trade-off among optimization cost, accuracy, and model size compared to existing QAT-based approaches.
+
+Additionally, we compare the performance of different quantization methods for the Conformer model across multiple datasets. In particular, we load pre-trained weights trained on the VoxPopuli, Multilingual Librispeech (MLS) (Pratap et al., 2020) datasets from Hugging Face, and apply GenPTQ to evaluate the performance on the corresponding test sets. As shown in Table 4, the MinMax method exhibits a substantial increase in WER at INT2 precision, whereas our approach effectively mitigates this degradation at a comparable precision. Furthermore, on the MLS-German and MLS-Spanish datasets, our method achieves
+
+Table 3: Performance comparison between the proposed method and SOTA QAT methods on Conformer-large using the Librispeech dataset  
+
+<table><tr><td>Dataset</td><td>Method</td><td>Prec.</td><td>WER(↓)</td><td>CER(↓)</td></tr><tr><td rowspan="5">Voxpopuli</td><td rowspan="2">Baseline</td><td>FP32</td><td>6.9</td><td>3.8</td></tr><tr><td>INT8</td><td>6.9</td><td>3.9</td></tr><tr><td rowspan="2">MinMax</td><td>INT4</td><td>7.3</td><td>4.1</td></tr><tr><td>INT2</td><td>15.3</td><td>6.8</td></tr><tr><td>Ours</td><td>INT2.5</td><td>7.8</td><td>4.3</td></tr><tr><td rowspan="5">MLS-German</td><td rowspan="2">Baseline</td><td>FP32</td><td>4.9</td><td>1.6</td></tr><tr><td>INT8</td><td>4.9</td><td>1.6</td></tr><tr><td rowspan="2">MinMax</td><td>INT4</td><td>5.0</td><td>1.7</td></tr><tr><td>INT2</td><td>36.5</td><td>12.6</td></tr><tr><td>Ours</td><td>INT2.5</td><td>5.7</td><td>1.8</td></tr><tr><td rowspan="5">MLS-Spanish</td><td rowspan="2">Baseline</td><td>FP32</td><td>4.4</td><td>1.8</td></tr><tr><td>INT8</td><td>4.5</td><td>1.8</td></tr><tr><td rowspan="2">MinMax</td><td>INT4</td><td>4.8</td><td>1.9</td></tr><tr><td>INT2</td><td>54.2</td><td>39.3</td></tr><tr><td>Ours</td><td>INT2.5</td><td>5.4</td><td>2.1</td></tr></table>
+
+Table 4: Performance of quantization methods on Conformer-Large across various datasets  
+
+<table><tr><td>Method</td><td>Prec.</td><td>WER(↓)</td><td>CER(↓)</td><td>Optimization Time</td></tr><tr><td>Baseline</td><td>FP32</td><td>4.1</td><td>1.6</td><td>-</td></tr><tr><td>AdaRound</td><td rowspan="2">INT2</td><td>5.8</td><td>2.4</td><td>8h 10m</td></tr><tr><td>BRECQ</td><td>5.5</td><td>2.5</td><td>2h 41m</td></tr><tr><td>Ours</td><td>INT2.5</td><td>4.3</td><td>1.7</td><td>15.16s</td></tr></table>
+
+Table 5: Performance comparison between the proposed method and SOTA PTQ methods on Whisper-medium using the Librispeech dataset
+
+markedly higher recognition accuracy than the MinMax 2-bit method, demonstrating its superior generalization capability.
+
+# 4.4 Accuracy-optimization cost trade-off Evaluation
+
+We compare SOTA PTQ methods (e.g., AdaRound (Nagel et al., 2020), BRECQ (Li et al., 2021)) with the proposed GenPTQ on Whisper-medium model, focusing on the trade-off between optimization cost and accuracy. Table 5 presents the optimization time, WER, and CER of each PTQ method applied to the Whisper-medium model on the Librispeech test-clean dataset. We directly applied AdaRound and BRECQ to Whisper-medium using their official
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Prec.</td><td rowspan="2">Size(MB)</td><td colspan="2">MLS-German</td><td colspan="2">MLS-Portuguese</td><td colspan="2">MLS-Spanish</td><td colspan="2">MLS-Polish</td></tr><tr><td>WER</td><td>CER</td><td>WER</td><td>CER</td><td>WER</td><td>CER</td><td>WER</td><td>CER</td></tr><tr><td>Baseline</td><td>FP32</td><td>2955.9</td><td>6.8</td><td>2.4</td><td>8.7</td><td>3.1</td><td>5.1</td><td>1.6</td><td>6.5</td><td>1.4</td></tr><tr><td rowspan="3">MinMax</td><td>INT8</td><td>864.5</td><td>6.7</td><td>2.3</td><td>8.9</td><td>3.2</td><td>5.1</td><td>1.6</td><td>6.6</td><td>1.4</td></tr><tr><td>INT4</td><td>489.2</td><td>6.9</td><td>2.4</td><td>9.2</td><td>3.6</td><td>5.2</td><td>1.7</td><td>6.8</td><td>1.6</td></tr><tr><td>INT2</td><td>301.5</td><td>51.7</td><td>39.8</td><td>64.5</td><td>38.4</td><td>30.4</td><td>16.4</td><td>69.7</td><td>54.5</td></tr><tr><td>Ours</td><td>INT2.5</td><td>323.2</td><td>7.9</td><td>3.1</td><td>10.3</td><td>3.6</td><td>6.3</td><td>2.0</td><td>7.9</td><td>1.9</td></tr></table>
+
+implementations and evaluated performance under consistent experimental conditions. In terms of optimization cost, AdaRound and BRECQ involve computationally expensive layer-wise and block-wise procedures, requiring approximately 8 hours and 2 hours, respectively. In contrast, our method completes optimization in only 15.16 seconds, demonstrating a substantial reduction in optimization cost. From the accuracy perspective, GenPTQ also achieves the best performance, achieving a WER of 4.3 under the 2.5-bit precision setting. Meanwhile, directly comparing the proposed method under the 2.5-bit precision setting with existing SOTA PTQ methods under the 2-bit precision setting may appear somewhat unfair. However, as shown in Table 1, the proposed method results in a model size that is only about 22.3MB larger than 2-bit quantization, yet achieves much better recognition performance (i.e.,  $\mathrm{WER} = 4.3$  vs.  $\mathrm{WER} = 5.8$ ) with significantly lower optimization cost (i.e., 15.16s vs. 8h10m). Considering these factors together, the proposed method offers a highly practical advantage in terms of the trade-off between optimization cost and accuracy, and can be regarded as an efficient PTQ solution for large-scale models.
+
+# 4.5 Ablation Study
+
+Performance Evaluation on Multilingual Datasets: Unlike conventional ASR models, Whisper achieves high multilingual recognition accuracy without additional fine-tuning. Accordingly, to assess generalization, we evaluate the proposed method on the MLS dataset, which includes German, Spanish, Portuguese, and Polish. As shown in Table 6, the proposed method achieves significantly lower WER and CER compared to MinMax-based INT2 PTQ. These results demonstrate that the proposed method maintains strong performance even in multilingual scenarios, highlighting its potential for language generalization on Whisper.
+
+Table 6: Performance comparison of quantization methods across various multilingual datasets  
+
+<table><tr><td>Dataset</td><td>SR</td><td>WER(↓)</td><td>WER Diff.</td><td>CER(↓)</td><td>CER Diff.</td></tr><tr><td rowspan="2">Librispeech</td><td>×</td><td>4.9</td><td>+0.8</td><td>2.2</td><td>+0.6</td></tr><tr><td>✓</td><td>4.3</td><td>+0.2</td><td>1.7</td><td>+0.1</td></tr><tr><td rowspan="2">Voxpopuli</td><td>×</td><td>11.5</td><td>+1.3</td><td>6.7</td><td>+0.7</td></tr><tr><td>✓</td><td>11.0</td><td>+0.8</td><td>6.2</td><td>+0.2</td></tr><tr><td rowspan="2">TedLium</td><td>×</td><td>13.0</td><td>+1.1</td><td>5.2</td><td>+0.7</td></tr><tr><td>✓</td><td>12.3</td><td>+0.4</td><td>4.7</td><td>+0.2</td></tr></table>
+
+Table 7: Analysis of the impact of SR on WER and CER for Whisper-medium across Librispeech, Voxpopuli, and TedLium Datasets
+
+Analysis of the Effectiveness of Sensitivity Regularization: We further analyze the performance contribution of the proposed SR. Specifically, we assess its impact by comparing results with and without SR in Eq. (10), using the Whisper-medium model across three datasets: LibriSpeech, VoxPopuli, and Tedlium. As shown in Table 7, the application of SR leads to reductions in WER and CER by approximately  $0.2 \sim 0.8$  and  $0.1 \sim 0.2$ , respectively, across all datasets. These results indicate that SR effectively mitigates performance degradation by preventing excessive bit-precision reduction in highly sensitive layers, even under a fixed target bit-precision constraint.
+
+A Comprehensive Analysis of Sensitivity Regularization: To more comprehensively analyze the effectiveness of the proposed SR, we conduct a visual analysis of the relationship between layerwise sensitivity and the assigned bit-precision. Figure 2(a) and (b) illustrate the bit-precision allocation without and with SR, respectively. As shown in Figure 2, the application of SR leads to a clear difference in bit-precision assignment, particularly for highly sensitive layers. In Figure 2(a), where SR is not applied, we observe that layers with high sensitivity (e.g., the first and third layers) are assigned excessively low bit-precisions (i.e., 2-bit, 3-bit) to satisfy the global bit constraint  $b_{tg}$ . This suggests that although higher bit-precisions may be initially assigned during the CBPR computation, they are aggressively reduced during optimization, compromising the quantization expressiveness of sensitive
+
+![](images/117fd66309f3a15be5208561c2a0385995436146475f04e0c92c1b9ad7c4f158.jpg)  
+(a) Layer-wise Sensitivity and Precision Analysis (w/o SR)
+
+![](images/21c8dd38f779979f223253b52bb3d2b0fbdda70ffcfb583c4e67569d76d20c15.jpg)  
+(b) Layer-wise Sensitivity and Precision Analysis (w SR)  
+Figure 2: Visualization of bit allocation patterns for Layers 0-70 of the Whisper-medium model under SR-applied and SR-free conditions. The x-axis denotes the layer index, the y-axis indicates the assigned bitprecision, and the secondary axis represents sensitivity.
+
+layers and potentially degrading recognition performance. In contrast, Figure 2(b) shows that the application of SR stabilizes bit allocation, preserving relatively high bit-precision (i.e., 8-bit) for layers with high sensitivity. This highlights the role of SR in preserving bit-precision where it is most needed and in stabilizing bit-precision allocation during CBPR-based optimization.
+
+Analysis of performance variation with respect to the size of the calibration dataset: We further provide a detailed analysis of GenPTQ performance on the Whisper model with varying sizes of calibration data. As shown in Table 8, our method requires only a very small number of calibration samples (i.e., 1 or 2 samples), which highlights the advantage of GenPTQ in incurring extremely low optimization cost. This is because, unlike conventional approaches that utilize calibration datasets to update specific quantization parameters through optimization, the GenPTQ method employs them solely for gradient computation to measure sensitivity. As a result, even a single sample is sufficient for effective gradient estimation, rendering GenPTQ a highly practical method with very low dependence on calibration data.
+
+# 5 Conclusion
+
+We presented GenPTQ, a mixed-precision posttraining quantization framework that effectively balances recognition accuracy, model size, and op
+
+<table><tr><td>Calibration Samples</td><td>Dataset</td><td>WER(↓)</td><td>CER(↓)</td></tr><tr><td>1</td><td></td><td>4.3</td><td>1.7</td></tr><tr><td>2</td><td>Librispeech</td><td>4.3</td><td>1.7</td></tr><tr><td>4</td><td></td><td>OOM</td><td>OOM</td></tr><tr><td>1</td><td></td><td>11.0</td><td>6.2</td></tr><tr><td>2</td><td>Voxpopuli</td><td>11.0</td><td>6.2</td></tr><tr><td>4</td><td></td><td>OOM</td><td>OOM</td></tr></table>
+
+Table 8: Performance analysis of GenPTQ on the Whisper model with respect to the number of calibration samples
+
+tization cost for large-scale ASR models. By leveraging a gradient-based sensitivity metric and a CBPR, GenPTQ enables efficient layer-wise bit allocation without requiring any retraining. Experimental results on both the Whisper and Conformer models demonstrate that GenPTQ achieves an average of 2.5-bit weight quantization while incurring less than one percent degradation in WER. Moreover, it completes the entire optimization process in just 15 seconds, significantly outperforming existing PTQ and QAT methods in terms of both accuracy and efficiency. These findings highlight the practicality and scalability of GenPTQ as a quantization solution for real-world ASR deployment.
+
+# Limitations
+
+The proposed method demonstrates strong performance across various speech recognition datasets, significantly reducing optimization cost while maintaining high recognition accuracy. However, it follows a gradient-based sensitivity metric to guide bit-precision allocation, which relies on the assumption that the metric accurately reflects the actual impact on recognition accuracy. This assumption may not always hold in models with complex inter-layer dependencies or strong long-range interactions, potentially leading to suboptimal bit allocation.
+
+# Acknowledgement
+
+This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2025-02304537, Development of Intelligent Home On-Device AI Matter Hub System Technology) and K-CHIPS (Korea Collaborative & High-tech Initiative for Prospective Semiconductor Research) (RS-2025-02305531) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea).
+
+# References
+
+Sergio Aquino-Bríguez, Pablo García-Sánchez, Andrés Ortiz, and Diego Aquino-Bríguez. 2025. Towards an energy consumption index for deep learning models: A comparative analysis of architectures, gpus, and measurement tools. Sensors, 25(3):846.  
+Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. 2019. Common voice: A massively-multilingual speech corpus. arXiv preprint arXiv:1912.06670.  
+Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460.  
+Kratika Bhagtani, Amit Kumar Singh Yadav, Paolo Bestagini, and Edward J Delp. 2024. Sslct: A convolutional transformer for synthetic speech localization. In 2024 IEEE 7th International Conference on Multimedia Information Processing and Retrieval (MIPR), pages 134-140. IEEE.  
+Arun Chauhan, Utsav Tiwari, and 1 others. 2023. Post training mixed precision quantization of neural networks using first-order information. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1343-1352.  
+Guoguo Chen, Shuzhou Chai, Guanbo Wang, Jiayu Du, Wei-Qiang Zhang, Chao Weng, Dan Su, Daniel Povey, Jan Trmal, Junbo Zhang, and 1 others. 2021. Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio. arXiv preprint arXiv:2106.06909.  
+Dahun Choi and Hyun Kim. 2025. Gradq-vit: Robust and efficient gradient quantization for vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 16019-16027.  
+Dahun Choi, Juntae Park, and Hyun Kim. 2024. Hlq: Hardware-friendly logarithmic quantization aware training for power-efficient low-precision cnn models. IEEE Access.  
+Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE.  
+Li Ding, Wen Fei, Yuyang Huang, Shuangrui Ding, Wenrui Dai, Chenglin Li, Junni Zou, and Hongkai Xiong. 2024. Ampa: Adaptive mixed precision allocation for low-bit integer training. In *Forty-first International Conference on Machine Learning*.  
+Shaojin Ding, Phoenix Meadowlark, Yanzhang He, Lukasz Lew, Shivani Agrawal, and Oleg Rybakov.
+
+2022. 4-bit conformer with native quantization aware training for speech recognition. arXiv preprint arXiv:2203.15952.  
+Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. 2019. Hawq: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 293-302.  
+Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. 2022. A survey of quantization methods for efficient neural network inference. In *Low-power computer vision*, pages 291-326. Chapman and Hall/CRC.  
+Ziyi Guan, Hantao Huang, Yupeng Su, Hong Huang, Ngai Wong, and Hao Yu. 2024. Aptq: Attention-aware post-training mixed-precision quantization for large language models. In Proceedings of the 61st ACM/IEEE Design Automation Conference, pages 1-6.  
+Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and 1 others. 2020. Conformer: Convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100.  
+Jahid Hasan. 2024. Optimizing large language models through quantization: A comparative analysis of ptq and qat techniques. arXiv preprint arXiv:2411.06084.  
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778.  
+Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29:3451-3460.  
+Nam Joon Kim and Hyun Kim. 2022. Fp-agl: Filter pruning with adaptive gradient learning for accelerating deep convolutional neural networks. IEEE Transactions on Multimedia, 25:5279-5290.  
+Nam Joon Kim and Hyun Kim. 2025. Hardware-friendly quantization via outlier scaling in convolution-attention-based hybrid networks. In 2025 IEEE International Conference on Consumer Electronics (ICCE), pages 1-3. IEEE.  
+Nam Joon Kim, Jongho Lee, and Hyun Kim. 2024. Hyq: hardware-friendly post-training quantization for cnn-transformer hybrid networks. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 4291-4299.
+
+Sehoon Kim, Amir Gholami, Zhewei Yao, Nicholas Lee, Patrick Wang, Aniruddha Nrusimha, Bohan Zhai, Tianren Gao, Michael W Mahoney, and Kurt Keutzer. 2022. Integer-only zero-shot quantization for efficient speech recognition. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4288-4292. IEEE.  
+Sehoon Kim, Coleman Hooper, Amir Gholami, Zhen Dong, Xiuyu Li, Sheng Shen, Michael W Mahoney, and Kurt Keutzer. 2023. Squeezeellm: Dense-and-sparse quantization. arXiv preprint arXiv:2306.07629.  
+Eliska Koberdanz and Wei Le. 2023. Mixquant: Mixed precision quantization with a bit-width optimization search. arXiv preprint arXiv:2309.17341.  
+Seung Il Lee, Kwanghyun Koo, Jong Ho Lee, Gilha Lee, Sangbeom Jeong, Seongjun O, and Hyun Kim. 2024. Vision transformer models for mobile/edge devices: a survey. Multimedia Systems, 30(2):109.  
+Haiguang Li, Usama Pervaiz, Joseph Antognini, Michał Matuszak, Lawrence Au, Gilles Roux, and Trausti Thormundsson. 2024a. Enhancing user experience in on-device machine learning with gated compression layers. arXiv preprint arXiv:2405.01739.  
+Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. 2021. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426.  
+Zhikai Li, Xianlei Long, Junrui Xiao, and Qingyi Gu. 2024b. Htq: Exploring the high-dimensional tradeoff of mixed-precision quantization. Pattern Recognition, 156:110788.  
+Chris J Maddison, Andriy Mnih, and Yee Whye Teh. 2016. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712.  
+Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. 2020. Up or down? adaptive rounding for post-training quantization. In International conference on machine learning, pages 7197-7206. PMLR.  
+Patrick K O'Neill, Vitaly Lavrukhin, Somshubra Majumdar, Vahid Noroozi, Yuekai Zhang, Oleksii Kuchaiev, Jagadeesh Balam, Yuliya Dovzhenko, Keenan Freyberg, Michael D Shulman, and 1 others. 2021. Spgispeech: 5,000 hours of transcribed financial audio for fully formatted end-to-end speech recognition. arXiv preprint arXiv:2104.02014.  
+Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE.
+
+Yifan Peng, Siddharth Dalmia, Ian Lane, and Shinji Watanabe. 2022. Branchformer: Parallel mlp-attention architectures to capture local and global context for speech recognition and understanding. In International Conference on Machine Learning, pages 17627-17643. PMLR.  
+Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. arXiv preprint arXiv:2012.03411.  
+Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR.  
+Navin Ranjan and Andreas Savakis. 2025. Mixqvit: Mixed-precision vision transformer quantization driven by layer importance and quantization sensitivity. arXiv preprint arXiv:2501.06357.  
+Anthony Rousseau, Paul Deléglise, and Yannick Esteve. 2012. Ted-lium: an automatic speech recognition dedicated corpus. In LREC, pages 125-129.  
+Oleg Rybakov, Phoenix Meadowlark, Shaojin Ding, David Qiu, Jian Li, David Rim, and Yanzhang He. 2023. 2-bit conformer quantization for automatic speech recognition. arXiv preprint arXiv:2305.16619.  
+Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. 2018. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520.  
+Yu-Shan Tai and 1 others. 2024. Mptq-vit: Mixed-precision post-training quantization for vision transformer. arXiv preprint arXiv:2401.14895.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.  
+Dominik Wagner, Ilja Baumann, Korbinian Riedhammer, and Tobias Bocklet. 2024. Outlier reduction with gated attention for improved post-training quantization in large sequence-to-sequence speech foundation models. arXiv preprint arXiv:2406.11022.  
+Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. arXiv preprint arXiv:2101.00390.  
+Shinji Watanabe, Takaaki Hori, Shigeki Karita, Tomoki Hayashi, Jiro Nishitoba, Yuya Unno, Nelson Enrique Yalta Soplin, Jahn Heymann, Matthew Wiesner, Nanxin Chen, Adithya Renduchintala, and Tsubasa
+
+Ochiai. 2018. ESPnet: End-to-end speech processing toolkit. In Proceedings of Interspeech, pages 2207-2211.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick 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, and 3 others. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Haoning Xu, Zhaoqing Li, Zengrui Jin, Huimeng Wang, Youjun Chen, Guinan Li, Mengzhe Geng, Shujie Hu, Jiajun Deng, and Xunying Liu. 2025. Effective and efficient mixed precision quantization of speech foundation models. arXiv preprint arXiv:2501.03643.  
+Qiuyu Zheng, Zengzhao Chen, Zhifeng Wang, Hai Liu, and Mengting Lin. 2024. Meconformer: Highly representative embedding extractor for speaker verification via incorporating selective convolution into deep speaker encoder. Expert Systems with Applications, 244:123004.
+
+# A Appendix
+
+# A.1 Detailed Hyperparameter Setting
+
+Table 9 summarizes the hyperparameter settings used in Algorithm 1. Specifically, the target bit-precision constraint  $b_{tg}$  is set to 2.5, and the initial bit-precision  $b$  is initialized to 4. Optimization-related hyperparameters (e.g., learning rate, number of iterations) follow the configuration described in Section 4.1.
+
+<table><tr><td>Hyperparameter</td><td>Value</td></tr><tr><td>b_tg</td><td>2.5</td></tr><tr><td>b</td><td>4</td></tr><tr><td>λ1</td><td>1</td></tr><tr><td>λ2</td><td>1</td></tr><tr><td>η</td><td>0.1</td></tr><tr><td>Iteration</td><td>150</td></tr></table>
+
+Table 9: GenPTQ Hyperparameter Setting
+
+# A.2 Detailed Bit Allocation per Sensitivity in Whisper with GenPTQ
+
+In Table 10, we present the bit-precision and corresponding sensitivity values of the fully-connected layer weights in the encoder and decoder blocks (up to the 11th block) of the Whisper-medium model with GenPTQ applied. The results are shown under a bit budget constraint of  $b_{tq} = 2.5$ . Due to the large number of layers, encoder and decoder blocks beyond the 11th are omitted for brevity. Notably, layers exhibiting high sensitivity—such as the first and third layers (i.e., encoder.layers.0.self_attn.k_prog and encoder.layers.0.self_attn.q_prog)—are assigned relatively high bit-precision (e.g., INT8), whereas most of the remaining layers are allocated low bit-precision values (e.g., INT2 or INT3). Given that this setting leads to only minimal increases in WER and CER, the results suggest that aggressively reducing the bit-precision of low-sensitivity layers while preserving high-precision in sensitive layers is an effective strategy for mitigating recognition performance degradation under tight precision constraints.
+
+# A.3 Recognition Example
+
+We present recognition examples obtained by applying both GenPTQ and a MinMax-based PTQ method to the Whisper-medium model across various datasets. As shown in Table 11, we report the
+
+prediction results on multiple benchmarks, including LibriSpeech, VoxPopuli, Tedmium, CommonVoice, GigaSpeech, MLS-German, MLS-Spanish, and MLS-Polish. For each method, we compare the predicted transcription (Pred) against the ground truth reference (Ref) in terms of WER and CER.
+
+<table><tr><td>Encoder Layer</td><td>Prec.</td><td>Sensitivity</td><td>Decoder Layer</td><td>Prec.</td><td>Sensitivity</td></tr><tr><td>encoder.layers.0.self_attn.k_proj</td><td>INT8</td><td>6.547e-04</td><td>decoder.layers.0.self_attn.k_proj</td><td>INT3</td><td>1.068e-07</td></tr><tr><td>encoder.layers.0.self_attn.v_proj</td><td>INT3</td><td>3.891e-06</td><td>decoder.layers.0.self_attn.v_proj</td><td>INT2</td><td>2.012e-08</td></tr><tr><td>encoder.layers.0.self_attn.q_proj</td><td>INT8</td><td>5.493e-04</td><td>decoder.layers.0.self_attn.q_proj</td><td>INT2</td><td>1.705e-13</td></tr><tr><td>encoder.layers.0.self_attn.out_proj</td><td>INT3</td><td>1.148e-06</td><td>decoder.layers.0.self_attn.out_proj</td><td>INT3</td><td>7.046e-07</td></tr><tr><td>encoder.layers.0.fc1</td><td>INT2</td><td>2.784e-12</td><td>decoder.layers.0.fc1</td><td>INT2</td><td>8.527e-14</td></tr><tr><td>encoder.layers.0.fc2</td><td>INT3</td><td>1.935e-06</td><td>decoder.layers.0.fc2</td><td>INT2</td><td>5.698e-08</td></tr><tr><td>encoder.layers.1.self_attn.k_proj</td><td>INT3</td><td>2.164e-05</td><td>decoder.layers.1.self_attn.k_proj</td><td>INT2</td><td>1.347e-08</td></tr><tr><td>encoder.layers.1.self_attn.v_proj</td><td>INT3</td><td>1.441e-07</td><td>decoder.layers.1.self_attn.v_proj</td><td>INT2</td><td>5.193e-08</td></tr><tr><td>encoder.layers.1.self_attn.q_proj</td><td>INT3</td><td>4.118e-06</td><td>decoder.layers.1.self_attn.q_proj</td><td>INT3</td><td>1.685e-07</td></tr><tr><td>encoder.layers.1.self_attn.out_proj</td><td>INT2</td><td>2.266e-09</td><td>decoder.layers.1.self_attn.out_proj</td><td>INT2</td><td>3.979e-08</td></tr><tr><td>encoder.layers.1.fc1</td><td>INT3</td><td>1.930e-07</td><td>decoder.layers.1.fc1</td><td>INT3</td><td>9.779e-07</td></tr><tr><td>encoder.layers.1.fc2</td><td>INT2</td><td>7.034e-13</td><td>decoder.layers.1.fc2</td><td>INT2</td><td>3.386e-08</td></tr><tr><td>encoder.layers.2.self_attn.k_proj</td><td>INT3</td><td>2.087e-06</td><td>decoder.layers.2.self_attn.k_proj</td><td>INT3</td><td>1.600e-06</td></tr><tr><td>encoder.layers.2.self_attn.v_proj</td><td>INT4</td><td>6.465e-07</td><td>decoder.layers.2.self_attn.v_proj</td><td>INT2</td><td>3.644e-08</td></tr><tr><td>encoder.layers.2.self_attn.q_proj</td><td>INT4</td><td>7.497e-05</td><td>decoder.layers.2.self_attn.q_proj</td><td>INT2</td><td>8.798e-09</td></tr><tr><td>encoder.layers.2.self_attn.out_proj</td><td>INT3</td><td>3.103e-07</td><td>decoder.layers.2.self_attn.out_proj</td><td>INT2</td><td>3.553e-15</td></tr><tr><td>encoder.layers.2.fc1</td><td>INT3</td><td>1.842e-07</td><td>decoder.layers.2.fc1</td><td>INT3</td><td>1.885e-07</td></tr><tr><td>encoder.layers.2.fc2</td><td>INT2</td><td>4.761e-13</td><td>decoder.layers.2.fc2</td><td>INT3</td><td>9.683e-08</td></tr><tr><td>encoder.layers.3.self_attn.k_proj</td><td>INT3</td><td>1.008e-06</td><td>decoder.layers.3.self_attn.k_proj</td><td>INT2</td><td>2.185e-08</td></tr><tr><td>encoder.layers.3.self_attn.v_proj</td><td>INT3</td><td>1.586e-07</td><td>decoder.layers.3.self_attn.v_proj</td><td>INT2</td><td>1.221e-09</td></tr><tr><td>encoder.layers.3.self_attn.q_proj</td><td>INT3</td><td>3.700e-07</td><td>decoder.layers.3.self_attn.q_proj</td><td>INT2</td><td>2.842e-14</td></tr><tr><td>encoder.layers.3.self_attn.out_proj</td><td>INT3</td><td>1.868e-06</td><td>decoder.layers.3.self_attn.out_proj</td><td>INT3</td><td>1.022e-07</td></tr><tr><td>encoder.layers.3.fc1</td><td>INT3</td><td>2.692e-07</td><td>decoder.layers.3.fc1</td><td>INT2</td><td>0.000e+00</td></tr><tr><td>encoder.layers.3.fc2</td><td>INT3</td><td>6.765e-07</td><td>decoder.layers.3.fc2</td><td>INT3</td><td>1.332e-07</td></tr><tr><td>encoder.layers.4.self_attn.k_proj</td><td>INT3</td><td>6.858e-08</td><td>decoder.layers.4.self_attn.k_proj</td><td>INT2</td><td>4.251e-08</td></tr><tr><td>encoder.layers.4.self_attn.v_proj</td><td>INT2</td><td>1.984e-08</td><td>decoder.layers.4.self_attn.v_proj</td><td>INT3</td><td>9.400e-08</td></tr><tr><td>encoder.layers.4.self_attn.q_proj</td><td>INT2</td><td>6.253e-13</td><td>decoder.layers.4.self_attn.q_proj</td><td>INT3</td><td>1.735e-07</td></tr><tr><td>encoder.layers.4.self_attn.out_proj</td><td>INT3</td><td>1.818e-07</td><td>decoder.layers.4.self_attn.out_proj</td><td>INT2</td><td>9.603e-10</td></tr><tr><td>encoder.layers.4.fc1</td><td>INT3</td><td>6.583e-08</td><td>decoder.layers.4.fc1</td><td>INT3</td><td>2.469e-07</td></tr><tr><td>encoder_layers.4.fc2</td><td>INT3</td><td>2.810e-06</td><td>decoder_layers.4.fc2</td><td>INT2</td><td>1.810e-08</td></tr><tr><td>encoder.layers.5.self_attn.k_proj</td><td>INT3</td><td>1.684e-07</td><td>decoder.layers.5.self_attn.k_proj</td><td>INT3</td><td>9.476e-06</td></tr><tr><td>encoder.layers.5.self_attn.v_proj</td><td>INT3</td><td>1.126e-07</td><td>decoder.layers.5.self_attn.v_proj</td><td>INT3</td><td>1.057e-07</td></tr><tr><td>encoder.layers.5.self_attn.q_proj</td><td>INT2</td><td>2.842e-13</td><td>decoder.layers.5.self_attn.q_proj</td><td>INT2</td><td>3.596e-08</td></tr><tr><td>encoder.layers.5.self_attn.out_proj</td><td>INT3</td><td>4.630e-07</td><td>decoder.layers.5.self_attn.out_proj</td><td>INT2</td><td>4.974e-14</td></tr><tr><td>encoder_layers.5.fc1</td><td>INT3</td><td>1.493e-07</td><td>decoder_layers.5.fc1</td><td>INT3</td><td>1.114e-07</td></tr><tr><td>encoder_layers.5.fc2</td><td>INT2</td><td>2.224e-08</td><td>decoder_layers.5.fc2</td><td>INT3</td><td>1.411e-07</td></tr><tr><td>encoder.layers.6.self_attn.k_proj</td><td>INT3</td><td>4.897e-06</td><td>decoder.layers.6.self_attn.k_proj</td><td>INT2</td><td>3.834e-08</td></tr><tr><td>encoder.layers.6.self_attn.v_proj</td><td>INT2</td><td>2.309e-08</td><td>decoder.layers.6.self_attn.v_proj</td><td>INT2</td><td>9.836e-09</td></tr><tr><td>encoder.layers.6.self_attn.q_proj</td><td>INT3</td><td>2.297e-07</td><td>decoder.layers.6.self_attn.q_proj</td><td>INT2</td><td>8.882e-14</td></tr><tr><td>encoder.layers.6.self_attn.out_proj</td><td>INT2</td><td>9.653e-09</td><td>decoder.layers.6.self_attn.out_proj</td><td>INT3</td><td>1.151e-07</td></tr><tr><td>encoder_layers.6.fc1</td><td>INT2</td><td>5.761e-08</td><td>decoder_layers.6.fc1</td><td>INT2</td><td>9.948e-14</td></tr><tr><td>encoder_layers.6.fc2</td><td>INT2</td><td>2.132e-13</td><td>decoder_layers.6.fc2</td><td>INT2</td><td>4.219e-08</td></tr><tr><td>encoder.layers.7.self_attn.k_proj</td><td>INT2</td><td>2.658e-08</td><td>decoder_layers.7.self_attn.k_proj</td><td>INT2</td><td>1.019e-08</td></tr><tr><td>encoder.layers.7.self_attn.v_proj</td><td>INT2</td><td>9.815e-10</td><td>decoder_layers.7.self_attn.v_proj</td><td>INT3</td><td>1.126e-07</td></tr><tr><td>encoder.layers.7.self_attn.q_proj</td><td>INT3</td><td>2.658e-06</td><td>decoder_layers.7.self_attn.q_proj</td><td>INT2</td><td>5.779e-09</td></tr><tr><td>encoder.layers.7.self_attn.out_proj</td><td>INT2</td><td>5.088e-08</td><td>decoder_layers.7.self_attn.out_proj</td><td>INT3</td><td>1.191e-07</td></tr><tr><td>encoder_layers.7.fc1</td><td>INT2</td><td>2.226e-08</td><td>decoder_layers.7.fc1</td><td>INT3</td><td>2.236e-07</td></tr><tr><td>encoder_layers.7.fc2</td><td>INT2</td><td>2.700e-13</td><td>decoder_layers.7.fc2</td><td>INT2</td><td>1.407e-08</td></tr><tr><td>encoder.layers.8.self_attn.k_proj</td><td>INT3</td><td>2.861e-07</td><td>decoder_layers.8.self_attn.k_proj</td><td>INT3</td><td>7.417e-06</td></tr><tr><td>encoder_layers.8.self_attn.v_proj</td><td>INT2</td><td>3.497e-08</td><td>decoder_layers.8.self_attn.v_proj</td><td>INT2</td><td>5.029e-08</td></tr><tr><td>encoder_layers.8.self_attn.q_proj</td><td>INT3</td><td>6.827e-07</td><td>decoder_layers.8.self_attn.q_proj</td><td>INT2</td><td>3.796e-08</td></tr><tr><td>encoder_layers.8.self_attn.out_proj</td><td>INT3</td><td>1.654e-06</td><td>decoder_layers.8.self_attn.out_proj</td><td>INT2</td><td>9.592e-14</td></tr><tr><td>encoder_layers.8.fc1</td><td>INT3</td><td>1.026e-07</td><td>decoder_layers.8.fc1</td><td>INT3</td><td>2.081e-07</td></tr><tr><td>encoder_layers.8.fc2</td><td>INT2</td><td>4.553e-08</td><td>decoder_layers.8.fc2</td><td>INT3</td><td>4.428e-07</td></tr><tr><td>encoder_layers.9.self_attn.k_proj</td><td>INT2</td><td>5.095e-08</td><td>decoder_layers.9.self_attn.k_proj</td><td>INT2</td><td>4.474e-08</td></tr><tr><td>encoder_layers.9.self_attn.v_proj</td><td>INT2</td><td>2.995e-08</td><td>decoder_layers.9.self_attn.v_proj</td><td>INT2</td><td>5.124e-08</td></tr><tr><td>encoder_layers.9.self_attn.q_proj</td><td>INT2</td><td>2.345e-13</td><td>decoder_layers.9.self_attn.q_proj</td><td>INT2</td><td>9.948e-14</td></tr><tr><td>encoder_layers.9.self_attn.out_proj</td><td>INT3</td><td>2.033e-07</td><td>decoder_layers.9.self_attn.out_proj</td><td>INT2</td><td>1.279e-13</td></tr><tr><td>encoder_layers.9.fc1</td><td>INT2</td><td>2.079e-08</td><td>decoder_layers.9.fc1</td><td>INT2</td><td>3.613e-08</td></tr><tr><td>encoder_layers.9.fc2</td><td>INT3</td><td>1.513e-06</td><td>decoder_layers.9.fc2</td><td>INT2</td><td>3.613e-08</td></tr><tr><td>encoder_layers.10.self_attn.k_proj</td><td>INT3</td><td>9.569e-08</td><td>decoder_layers.10.self_attn.k_proj</td><td>INT2</td><td>1.901e-08</td></tr><tr><td>encoder_layers.10.self_attn.v_proj</td><td>INT2</td><td>2.481e-08</td><td>decoder_layers.10.self_attn.v_proj</td><td>INT3</td><td>1.222e-07</td></tr><tr><td>encoder_layers.10.self_attn.q_proj</td><td>INT2</td><td>1.634e-13</td><td>decoder_layers.10.self_attn.q_proj</td><td>INT3</td><td>1.156e-06</td></tr><tr><td>encoder_layers.10.self_attn.out_proj</td><td>INT3</td><td>9.149e-08</td><td>decoder_layers.10.self_attn.out_proj</td><td>INT3</td><td>1.043e-07</td></tr><tr><td>encoder_layers.10.fc1</td><td>INT2</td><td>2.838e-08</td><td>decoder_layers.10.fc1</td><td>INT2</td><td>2.810e-09</td></tr><tr><td>encoder_layers.10.fc2</td><td>INT3</td><td>9.042e-07</td><td>decoder_layers.10.fc2</td><td>INT3</td><td>8.704e-08</td></tr></table>
+
+Table 10: Layer-wise sensitivity and assigned bit-precision in the first 11 encoder/decoder blocks of the Whisper-medium model under the 2.5-bit constraint.
+
+<table><tr><td>Dataset</td><td>Method</td><td colspan="2">Prec.</td><td>Results</td><td>WER</td><td>CER</td></tr><tr><td rowspan="6">Libirspeech</td><td rowspan="4">MinMax</td><td rowspan="2">INT4</td><td>Ref :</td><td>out in the woods stood a nice little fir tree</td><td rowspan="2">0.0</td><td rowspan="2">0.0</td></tr><tr><td>Pred :</td><td>out in the woods stood a nice little fir tree</td></tr><tr><td rowspan="2">INT2</td><td>Ref :</td><td>out in the woods stood a nice little fir tree</td><td rowspan="2">30.0</td><td rowspan="2">13.3</td></tr><tr><td>Pred :</td><td>out in the woods to the nice little fur tree</td></tr><tr><td rowspan="2">GenPTQ</td><td rowspan="2">INT2.5</td><td>Ref :</td><td>out in the woods stood a nice little fir tree</td><td rowspan="2">0.0</td><td rowspan="2">0.0</td></tr><tr><td>Pred :</td><td>out in the woods stood a nice little fir tree</td></tr><tr><td rowspan="6">Voxpopuli</td><td rowspan="4">MinMax</td><td rowspan="2">INT4</td><td>Ref :</td><td>i know it means as much to them as it means to me</td><td rowspan="2">0.0</td><td rowspan="2">0.0</td></tr><tr><td>Pred :</td><td>i know it means as much to them as it means to me</td></tr><tr><td rowspan="2">INT2</td><td>Ref :</td><td>i know it means as much to them as it means to me</td><td rowspan="2">30.8</td><td rowspan="2">30.7</td></tr><tr><td>Pred :</td><td>it means as much as it means to me</td></tr><tr><td rowspan="2">GenPTQ</td><td rowspan="2">INT2.5</td><td>Ref :</td><td>i know it means as much to them as it means to me</td><td rowspan="2">0.0</td><td rowspan="2">0.0</td></tr><tr><td>Pred :</td><td>i know it means as much to them as it means to me</td></tr><tr><td rowspan="6">Tedlium</td><td rowspan="4">MinMax</td><td rowspan="2">INT4</td><td>Ref :</td><td>senile decrepit laid up done up done for done in cracked up counted out</td><td rowspan="2">0.0</td><td rowspan="2">0.0</td></tr><tr><td>Pred :</td><td>senile decrepit laid up done up done for done in cracked up counted out</td></tr><tr><td rowspan="2">INT2</td><td>Ref :</td><td>senile decrepit laid up done up done for done in cracked up counted out</td><td rowspan="2">42.9</td><td rowspan="2">16.9</td></tr><tr><td>Pred :</td><td>seen off the crap it laid up done up done for done in crapped up counted out</td></tr><tr><td rowspan="2">GenPTQ</td><td rowspan="2">INT2.5</td><td>Ref :</td><td>senile decrepit laid up done up done for done in cracked up counted out</td><td rowspan="2">0.0</td><td rowspan="2">0.0</td></tr><tr><td>Pred :</td><td>senile decrepit laid up done up done for done in cracked up counted out</td></tr><tr><td rowspan="6">CommonVoice</td><td rowspan="4">MinMax</td><td rowspan="2">INT4</td><td>Ref :</td><td>it is a busy market town that serves a large surrounding area</td><td rowspan="2">8.3</td><td rowspan="2">4.9</td></tr><tr><td>Pred :</td><td>it is a busy market town that serves a large surrounded area</td></tr><tr><td rowspan="2">INT2</td><td>Ref :</td><td>it is a busy market town that serves a large surrounding area</td><td rowspan="2">91.7</td><td rowspan="2">63.9</td></tr><tr><td>Pred :</td><td>its a piece of market and thats around it here yeah</td></tr><tr><td rowspan="2">GenPTQ</td><td rowspan="2">INT2.5</td><td>Ref :</td><td>it is a busy market town that serves a large surrounding area</td><td rowspan="2">8.3</td><td rowspan="2">4.9</td></tr><tr><td>Pred :</td><td>it is a busy market town that serves a large surrounded area</td></tr><tr><td rowspan="6">GigaSpeech</td><td rowspan="4">MinMax</td><td rowspan="2">INT4</td><td>Ref :</td><td>i mean i showed you a short clip and you may have seen about half of them</td><td rowspan="2">5.9</td><td rowspan="2">2.7</td></tr><tr><td>Pred :</td><td>mean i showed you a short clip and you may have seen about half of them</td></tr><tr><td rowspan="2">INT2</td><td>Ref :</td><td>i mean i showed you a short clip and you may have seen about half of them</td><td rowspan="2">11.8</td><td rowspan="2">13.7</td></tr><tr><td>Pred :</td><td>i mean i showed you a clip and you may have seen about half of it</td></tr><tr><td rowspan="2">GenPTQ</td><td rowspan="2">INT2.5</td><td>Ref :</td><td>i mean i showed you a short clip and you may have seen about half of them</td><td rowspan="2">0.0</td><td rowspan="2">0.0</td></tr><tr><td>Pred :</td><td>i mean i showed you a short clip and you may have seen about half of them</td></tr></table>
+
+Table 11: Recognition examples generated by the Whisper-medium model quantized with GenPTQ across multiple speech datasets.

2025/GenPTQ_ Green Post-Training Quantization for Large-Scale ASR Models with Mixed-Precision Bit Allocation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b3993febda30217efef121cf70cf2ab998d2023682da13c96d0c8948bbcbde0
+size 1176385

2025/GenPilot_ A Multi-Agent System for Test-Time Prompt Optimization in Image Generation/5428225a-34af-4b71-95e2-69e7be23818e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27c9eefa9a0a04fdf378ca3abb7566f3c56e990a3e76a850acdf4adf6ea386d8
+size 11927186

2025/GenPilot_ A Multi-Agent System for Test-Time Prompt Optimization in Image Generation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc7147dc4f8cec6beed3e75845cac6abfefa871cdc236ea84a66a5ee039670b2
+size 1645933

2025/GenPoE_ Generative Passage-level Mixture of Experts for Knowledge Enhancement of LLMs/0fe8ac59-edc9-41a5-8924-a4c69cfc95df_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b862ce0458215a5d1b728ac78e6d9d70638868c08c998ff1e668887f52983473
+size 2428124

2025/GenPoE_ Generative Passage-level Mixture of Experts for Knowledge Enhancement of LLMs/full.md

@@ -0,0 +1,492 @@
+# GenPoE: Generative Passage-level Mixture of Experts for Knowledge Enhancement of LLMs
+
+Xuebing Liu $^{1}$ , Shanbao Qiao $^{1}$ , Seung-Hoon Na $^{2*}$
+
+$^{1}$ Department of Computer Science and Artificial Intelligence, Jeonbuk National University
+
+2Graduate School of Artificial Intelligence,
+
+Ulsan National Institute of Science and Technology
+
+{liuxuebing, Joe}@jbnu.ac.kr, nash@unist.ac.kr
+
+# Abstract
+
+Typically, parametric adaptation methods such as domain-adaptive pretraining (DAP) and retrieval-augmented generation (RAG) have been considered effective approaches for adapting large language models (LLMs) to new knowledge or domains. To unify positive effects of parametric adaptation and RAG, this paper proposes GenPoE, i.e., "generative" passage-level mixture of experts (MoEs) for enhancing knowledge of LLMs. The key component is its novel MoE-generating hypernetwork which takes in-context retrieved passages and generates their "expert" parameters, where these generated parameters are then integrated into LLMs by forming expert networks. With its use of "generated" parameters, GenPoE does not require a separate parameter training or finetuning stage, which is often costly. By parameterizing passages into expert networks, GenPoE likely exhibits robustness even when the retrieved passages are irrelevant. Experiment results in two open-domain question answering (QA) tasks present that GenPoE shows improved performances over other passage-level knowledge editing, and its combination of RAG produces superior performances over RAG. Our data and code will be available at https://github.com/Liu-Xuebing/GenPoE.
+
+# 1 Introduction
+
+Large Language Models (LLMs) have achieved remarkable success across a wide range of natural language processing tasks, especially in knowledge-intensive applications due to their vast pretrained knowledge (Zhao et al., 2023; Hadi et al., 2023). However, the dynamic nature of real-world knowledge necessitates continuous and efficient updates to these models.
+
+Existing methods for enhancing knowledge of LLMs are broadly categorized to two approaches - parametric adaptation and RAG (Lewis et al., 2020;
+
+Pan et al., 2024); 1) Parametric adaptation – such as domain-adaptive pretraining (DAP) (Ke et al., 2023) or parametric knowledge editing (De Cao et al., 2021; Meng et al., 2022a) $^{1}$  – modifies or updates the parameters of large language models (LLMs). However, it often requires substantial computational and memory cost due to LLM's huge parameters and the training time, and may suffer from the catastrophic forgetting, which can degrade previously learned knowledge and task performance. 2) RAG retrieves relevant passages and uses them as additional in-context prompts before generating responses. However, RAG likely suffers from performance degradation due to noisy or irrelevant retrieved content (Yoran et al., 2023; Tu et al., 2025), and its lack of parameter-level integration may limit the model's ability to accumulate and evolve knowledge over time.
+
+Given these characteristics, we assume that parametric adaptation and RAG serve as complementary solutions, each addressing the limitations of the other. RAG is less likely to suffer from computational overhead and catastrophic forgetting, while parametric adaptation tends to avoid the influence of noisy or irrelevant retrieved content due to its access to internalized parametric knowledge. To unify these complementary effects, this paper proposes GenPoE, i.e., "generative" passage-level MoEs for enhancing knowledge of LLMs, by internalizing passages into MoE parameters based on our novel MoE-generating hypernetwork. More specifically, the MoE-generating hypernetwork is trained via meta-learning framework: given a question and its corresponding gold passage, it learns to produce passage-level MoE parameters such that the resulting MoE-equipped LLMs maximize the
+
+autoregressive likelihood of generating the correct answer to the question. During inference, as an alternative to RAG, GenPoE employs the MoE-generating hypernetwork through a passage injection stage to incorporate retrieved passages. Once the top- $k$  passages are retrieved, the hypernetwork dynamically constructs  $k$  passage-level MoEs, which are then integrated with the original LLM parameters, as illustrated in Figure 1.
+
+Experimental results on standard open-domain question answering (QA) datasets—including Natural Questions (NQ) (Kwiatkowski et al., 2019), and TriviaQA (TQA) (Joshi et al., 2017)—demonstrate that GenPoE achieves a stronger passage injection effect by outperforming baseline parametric knowledge editing methods. When combined with RAG, GenPoE gives rise to a more powerful RAG variant that surpasses the performance of existing advanced RAG approaches, confirming the complementary relationship between RAG and parametric adaptation. Our main contributions are summarized as follows:
+
+1. We propose GenPoE, a novel passage injection framework for enhancing the knowledge of large language models (LLMs), which leverages hypernetwork to generate passage-level Mixture-of-Experts (MoE) parameters for open-domain question answering tasks.  
+2. We present a novel meta-learning framework for training a MoE-generating hypernetwork, where the hypernetwork is trained in an end-to-end manner to minimize the loss of generating correct answers, given a question, its corresponding gold passage, and the ground-truth answer.  
+3. We present comprehensive empirical results on standard QA benchmarks, demonstrating that GenPoE consistently outperforms conventional parametric knowledge editing approaches and parameter-efficient tuning methods, and further enhances performance when combined with retrieval-based generation (RAG), forming a more powerful RAG variant.
+
+# 2 Related Work
+
+# 2.1 Knowledge Enhancement
+
+Knowledge injection has emerged as a crucial technique for enhancing the performance of language
+
+![](images/3b6639dd6dca386a19aa70347868492e182473315610f47fd013496aa1a85a0a.jpg)  
+Model Single layer  
+Figure 1: Once the top- $k$  retrieved passages are obtained, GenPoE transforms this knowledge into passage-level MoEs using the MoE-generating hypernetwork, which is pretrained via a meta-learning framework. The resulting passage-level MoEs are then combined with the original feed-forward network (FFN) parameters – without modifying them – to generate a response to the given question.
+
+models by efficiently integrating external knowledge. Two primary approaches have gained significant attention: parameter-efficient tuning and model editing.
+
+Parameter-efficient tuning enhances large pretrained models by introducing small trainable modules while keeping most parameters frozen. Key approaches include adapter (Houlsby et al., 2019), which adds neural network bottlenecks to transformer blocks; Prompt Tuning (Li and Liang, 2021), which optimizes the appended prompts for task adaptation; and LoRA (Hu et al., 2021), which updates rank decomposition matrices. Recent advances, such as DyLoRA (Valipour et al., 2022), improve efficiency by selectively updating partial parameters. Building on DyLoRA, MELO (Yu et al., 2024) introduces a neuron-indexed dynamic LoRA mechanism.
+
+Model editing focuses on maintaining the reliability of edited knowledge, ensuring that the changes successfully address the target queries. Additionally, it emphasizes enhancing the generality, allowing the edited model to generalize the new knowledge to related queries effectively. Furthermore, it seeks to preserve locality, ensuring that the modifications do not interfere with the retention of unrelated original knowledge. It is categorized into three types: Meta-learning editors (De Cao et al., 2021; Mitchell et al., 2021; Tan et al., 2023), which use hyper-networks to adjust gradients; Locate-then-edit editors (Meng et al., 2022a,b; Li et al., 2024), which identify and update relevant parameters; and Memory-based editors, where (Zheng et al., 2023b; Zhong et al., 2023; Gu et al., 2023; Cheng et al., 2023) update knowledge from prompts using in-context learning without gradient updates or parameter modifications, where other approaches like T-Patcher (Huang et al.) and MEMoE (Wang and Li, 2024) store the memory of edited facts using additional parameters.
+
+Further, PRAG (Su et al., 2025) is a novel RAG paradigm that integrates external knowledge into a LLM's FFN layers by parameterizing passage representations into LoRA weights. The recently proposed DyPRAG (Tan et al., 2025) extends PRAG by introducing a hypernetwork component that enables dynamic generation of passage-specific LoRA parameters at inference time. Moreover, DyPRAG employs a two-stage training procedure: it first learns passage-specific LoRA weights, which are then used as supervision to train the hypernetwork for dynamic LoRA parameter generation during inference. By comparison, our proposed method, GenPoE, is an end-to-end framework that trains the hypernetwork using only the answer as supervision, enabling dynamic MoE parameter generation at inference time without requiring any additional training stage.
+
+# 2.2 Mixture of Experts
+
+In transformer-based Large Language Models (LLMs), Mixture-of-Experts (MoE) layers utilize a set of expert networks and a gating mechanism to route inputs to the most suitable experts (Shazeer et al., 2017; Antoniak et al., 2023). These layers are strategically positioned after the self-attention sublayer to optimize feed-forward network (FFN) selection, significantly reducing computational overhead in large models like PaLM (Chowdhery et al., 2023), where FFN layers account for the majority
+
+of parameters.
+
+Dense MoE approaches activate all available experts simultaneously, which enhances predictive accuracy but demands substantial computational resources. Early implementations (Jacobs et al., 1991; Rasmussen and Ghahramani, 2001; Aljundi et al., 2017) demonstrated this effectiveness, and more recent methods like EvoMoE (Nie et al., 2021), MoLE (Wu et al., 2024), LoRAMoE (Dou et al., 2023), and DSMoE (Pan et al., 2024) have refined the dense MoE structure to balance performance and efficiency.
+
+Sparse MoE improves computational efficiency by selecting only the top-k experts for each input, thereby maintaining accuracy while reducing processing demands (Shazeer et al., 2017). However, this selective activation can cause load imbalances, where certain experts are overused while others are underutilized. To counter this, auxiliary loss functions are introduced to distribute tokens more evenly across experts, as seen in (Lepikhin et al., 2020; Jiang et al., 2024a; Du et al., 2022; Fedus et al., 2022). This strategy allows sparse MoE models to scale effectively by expanding parameter capacity without a corresponding increase in computational cost.
+
+# 3 Methodology
+
+Figure 2 provides an overview of GenPoE, highlighting its two key stages: (1) the meta-learning stage for training the hypernetwork (Section 3.2), and (2) the passage injection via MoEs during inference (Section 3.3.1). In the following sections, we first introduce our MoE framework that integrates passage-level experts.
+
+# 3.1 Passage-level MoE
+
+In GenPoE, passages are parameterized to passage-level experts, referred to as passage-level MoEs, and original parameters of LLMs are not modified. More specifically, each passage-level expert module is constructed as an auxiliary FFN that operates in parallel with the original FFN layer of the base transformer model. Formally, we denote the original FNN block as  $\mathrm{FFN}^l (\mathbf{x}_t^l)$ , defined as follows:
+
+$$
+\mathrm {F F N} ^ {l} \left(\mathbf {x} _ {t} ^ {l}\right) = \operatorname {R e L U} \left(\mathbf {x} _ {t} ^ {l} \mathbf {W} _ {K} ^ {l}\right) \mathbf {W} _ {V} ^ {l} \tag {1}
+$$
+
+where  $\mathbf{x}_t^l\in \mathbb{R}^{d_m}$  is the input representation of  $t$  -th token at  $l$  -th FFN layer, and  $\mathbf{W}_K\in \mathbb{R}^{d_m\times d}$  and  $\mathbf{W}_V\in \mathbb{R}^{d\times d_m}$  represent the up and down projection matrices at the original FFN layer, respectively.
+
+![](images/30c462314e406e7002a0546c4b5e69809470cc751fd14086456f9aec5c8e3df0.jpg)  
+Figure 2: An overview of the proposed GenPoE framework: 1) Meta-learning for training the MoE-generating hypernetwork: For each tuple (query, gold passage, answer) in the dataset  $\mathcal{D}$ , the hypernetwork  $\mathcal{H}_{\phi}$  is trained to generate passage-specific expert parameters. This ensures that contextual information from each passage is effectively incorporated into the model to support reasoning, by optimizing the meta-loss defined in Eq. (5) (see §3.2.2). 2) Inference via Passage Injection: At inference time, given the top-retrieved passages for a test query  $q_{\mathrm{test}}$ , obtained through a retrieval and reranking process (see § 3.3.2), GenPoE applies the hypernetwork  $\mathcal{H}_{\phi}$  to generate passage-specific expert parameters. These parameters are then dynamically integrated into the sparse mixture-of-experts (MoE) architecture (see § 3.3.1), enabling the model to perform query-aware reasoning by incorporating contextual knowledge from the retrieved content.
+
+In addition to the original FFN, a passage-level expert  $\mathrm{E}^l (\mathbf{x}_t^l)$  is incorporated. To reduce parameter overhead and enable efficient training, we apply low-rank matrix factorization to the expert's up and down projection matrices, resulting in a low-rank approximated FFN:
+
+$$
+\mathrm {E} _ {\boldsymbol {\theta} _ {E}} ^ {l} \left(\mathbf {x} _ {t} ^ {l}\right) = \operatorname {R e L U} \left(\mathbf {x} _ {t} ^ {l} \mathbf {W} _ {u p 1} ^ {l} \mathbf {W} _ {u p 2} ^ {l}\right) \mathbf {W} _ {d w 1} ^ {l} \mathbf {W} _ {d w 2} ^ {l} \tag {2}
+$$
+
+where  $\theta_{E} = \{\mathbf{W}_{up1}^{l},\mathbf{W}_{up2}^{l},\mathbf{W}_{dw1}^{l},\mathbf{W}_{dw2}^{l}\}$ ,  $\mathbf{W}_{up1}^{l}\in \mathbb{R}^{d_m\times k}$ ,  $\mathbf{W}_{up2}^{l}\in \mathbb{R}^{k\times d_m}$ ,  $\mathbf{W}_{dw1}^{l}\in \mathbb{R}^{d_m\times k}$ , and  $\mathbf{W}_{dw2}^{l}\in \mathbb{R}^{k\times d_m}$  represent the low-rank approximations of the up and down projection metrics for the passage-level expert.
+
+The expert is inserted alongside the original FFN block, forming the following "passage-level" MoE layer:
+
+$$
+\mathbf {y} _ {t} ^ {l} = \operatorname {F F N} ^ {l} \left(\mathbf {x} _ {t} ^ {l}\right) + \lambda^ {l} \mathrm {E} _ {\boldsymbol {\theta} _ {E}} ^ {l} \left(\mathbf {x} _ {t} ^ {l}\right) \tag {3}
+$$
+
+where  $\lambda^t$  is the mixing parameter, set to 1 for a specific layer  $l$  and 0 otherwise, depending on the layer index (i.e.  $\lambda^l = 0$ , no expert module is inserted at the current layer), enabling sparse expert selec
+
+tion and targeted knowledge editing at the passage level.
+
+# 3.2 Meta Learning for Training MoE-generating Hypernetwork
+
+The expert parameters  $\theta_{E}$  need to be specifically determined for each passage. Instead of applying parametric adaptation methods such as fine-tuning for individual passages, GenPoE "generates" these parameters of the passage-specific expert using the MoE-generating hypernetwork, inspired by the generative adaptor proposed in (Chen et al., 2025).
+
+# 3.2.1 Hypernetwork Architecture
+
+The structure of the MoE-generating hypernetwork is illustrated in upper part of Figure 2; it takes a passage representation as input and produces the expert parameters  $\theta_{E}$  as output.
+
+Suppose a passage  $psg = w_1 \cdots w_T$  of length  $T$  is given. The passage  $psg$  is fed into the word embedding layer of the LLM, and mean pooling is applied to produce a passage embedding  $\operatorname{Emd}(psg) \in \mathbb{R}^{1 \times d_m}$ . The hypernetwork, consisting of two linear layers with a ReLU activation
+
+function, takes  $\operatorname{Emd}(psg)$  as input and generates the expert parameters, formulated as follows:
+
+$$
+\begin{array}{l} \mathcal {H} _ {\phi} (\operatorname {E m d} (p s g)) = \\ \operatorname {L i n e a r} \left(\operatorname {R e L U} \left(\operatorname {L i n e a r} (\operatorname {E m d} (p s g)))\right) \in \mathbb {R} ^ {d _ {\text {o u t}}} \right. \tag {4} \\ \end{array}
+$$
+
+where  $d_{out}$  corresponds to the concatenated dimension of the expert parameters, i.e.,  $2 \times 2 \times (d_m \times k)$ . We reshape the output  $\mathcal{H}_{\phi}(\operatorname{Emd}(psg))$  to obtain  $\theta_E$ , which represents the expert parameters for the given passage  $psg$ .
+
+# 3.2.2 Meta Learning for Training Hypernetwork
+
+To inject passage-specific knowledge into the expert modules, we train a hypernetwork to dynamically generate the parameters of each expert based on its corresponding passage. To this end, we first prepare a training dataset  $\mathcal{D} = (q_i,p_g^{q_i},a^{q_i})_{i = 1}^N$ , collected from QA tasks, where  $q_{i}$  denotes the  $i$ th question,  $p_{g}^{q_{i}}$  is its corresponding gold passage, and  $a^{q_i}$  is the ground-truth answer. Suppose that  $P(y\mid x,\pmb{\theta}_E)$  denotes the probability of generating a sequence  $y$  given input  $x$ , using Eq. (3), where  $\pmb{\theta}_{E}$  represents the expert parameters used in  $\mathrm{E}_{\pmb{\theta}_E}^l$ .
+
+For designing the meta loss, which serves as the objective function for training the parameters  $\phi$  that generate the expert parameters, we use a simple "autoregressive" language modeling loss over the question-answer text, ensuring that the generated expert enhances the model's response capability for a given question, formally defined as follows:
+
+$$
+\mathcal {L} (\phi) = - \sum_ {(q, p, a) \in \mathcal {D}} \log P (a \mid q, \mathcal {H} _ {\phi} (p)) \tag {5}
+$$
+
+# 3.3 Inference via Passage Injection
+
+During inference, GenPoE first retrieves the top passages for a given test query  $q_{\mathrm{test}}$ , using dense retrieval followed by a reranking step to refine the results. Based on the top-retrieved passages  $Top_K(q_{\mathrm{test}})$ , GenPoE then performs "passage injection", which dynamically generates the corresponding expert parameters by  $\mathcal{H}_{\phi}$ , allowing the model to adapt its reasoning process to the contextualized knowledge retrieved at inference time.
+
+# 3.3.1 Passage Injection to MoE
+
+Given  $Top_{K}(q_{\mathrm{test}})$ , the meta-learned MoE-generating hypernetwork  $\mathcal{H}_{\phi}$  generates expert-specific parameter residuals  $\pmb{\theta}_{j}$  for each selected passage  $p_j \in Top_K(q_{\mathrm{test}})$ , which are then injected into the expert modules for inference.
+
+Single-Expert Mode: In the single-expert setting  $(K = 1)$ , only the top passage is selected, where the inference is conducted using Eq. (3), where the parameters for the expert is generated according to Eq. (4).
+
+Multi-Expert Mode: Generalizing to the multiexpert setting  $(K > 1)$ , as shown in the Figure 1, the model incorporates information from knowledge in top  $K$ -retrieved passages  $Top_{K}(q_{\mathrm{test}})$ . The Eq. (3) is then further extended to the following:
+
+$$
+\mathbf {y} _ {t} ^ {l} = \mathrm {F F N} ^ {l} (\mathbf {x} _ {t} ^ {l}) + \lambda_ {l} ^ {t} \sum_ {k = 1} ^ {K} \mathbf {r} _ {k} \cdot \mathrm {E} _ {\mathcal {H} _ {\phi} (p _ {k})} ^ {l} (\mathbf {x} _ {t} ^ {l}) (6)
+$$
+
+where  $\mathrm{E}_{\mathcal{H}_{\phi}(p_k)}^l$  is the expert module using the generated parameters for  $p_k$  by the hypernetwork  $\mathcal{H}_{\phi}$  and the relevance vector  $\mathbf{r} \in \mathbb{R}^{K}$  is computed according to the relevance distribution over top  $k$  passages, which will be presented in Section 3.3.2.
+
+The multi-expert setting enhances the model's ability to capture diverse contextual information by aggregating knowledge from multiple source passages, thereby improving the robustness and accuracy of response generation.
+
+# 3.3.2 Retrieval-Reranker
+
+Given a test query  $q_{test}$ , Dense Passage Retrieval (DPR) (Karpukhin et al., 2020) uses both the query vector  $\mathrm{Emd}(q_{test})$  and the passage vector  $\mathrm{Emd}(p_i)$  for  $p_i$  in WiKipage database (Kwiatkowski et al., 2019), and computes inner product similarity between them:
+
+$$
+s c o r e \left(q _ {t e s t}, p _ {i}\right) = \mathsf {E m d} \left(q _ {t e s t}\right) ^ {\top} \mathsf {E m d} \left(p _ {i}\right) \tag {7}
+$$
+
+For the reranking, we incorporate BAAI/bgeranker-  $\nu 2$  -gemma2 (Xiao et al., 2024) as a reranker model and fine-tune it to improve passage relevance. For the finetuning the reranker, we additionally construct a training dataset passages retrieved by DPR. To optimize the reranker, we adopt a contrastive loss function (Sohn, 2016), ensuring that relevant passages receive higher scores than irrelevant ones. For a given question  $q$  in the training set, let  $p_{+}$  be a positive passage, and let  $(p_{-}^{(1)},\dots ,p_{-}^{(L)})$  be  $L$  negative passages for  $q$  .The loss function is defined as:
+
+$$
+\mathcal {L} _ {\text {r e r a n k e r}} = - \log \frac {\exp \left(\operatorname {s c o r e} _ {\text {r e r a n k}} (q , p _ {+})\right)}{\sum_ {i = 1} ^ {L + 1} \exp \left(\operatorname {s c o r e} _ {\text {r e r a n k}} (q , p ^ {(i)})\right)} \tag {8}
+$$
+
+The details and results of the re-ranker can be found in Appendix B.
+
+The relevance vector  $\mathbf{r} \in \mathbb{R}^K$  in Eq. (6) is obtained after reranking step, defined as:
+
+$$
+\mathbf {r} = \operatorname {s o f t m a x} \left(\operatorname {t o p} K \left(\left[ \operatorname {s c o r e} _ {\text {r e r a n k}} \left(q _ {\text {t e s t}}, p _ {j}\right) \right] _ {j = 1} ^ {K}\right)\right) \tag {9}
+$$
+
+# 4 Experiments
+
+In this part, we explain the experimental setup and present the key results of our experiments.
+
+# 4.1 Experimental Setup
+
+Dataset: To evaluate the proposed method's effectiveness in paragraph-level editing, we conduct experiments using the open-domain question answering datasets Natural Questions (NQ) (Kwiatkowski et al., 2019) and TriviaQA (TQA) (Joshi et al., 2017). These datasets provide a diverse set of challenging questions across various domains, enabling a comprehensive assessment of model performance.
+
+Backbone Model: We adopt Llama2 $_{7B}$  (Touvron et al., 2023), Qwen2.5 $_{7B}$  (Yang et al., 2024; Team, 2024) and the more advanced Baichuan2 $_{7B}$  (Baichuan, 2023) as base models, upon which expert modules are integrated to realize our proposed framework.
+
+Methods: We evaluate several widely-used knowledge editing methods for comparison, including MEMIT (Meng et al., 2022b), MEND (Mitchell et al., 2021), MALMEN (Tan et al., 2023) and Finetuning with LoRA (FT-LoRA) (Hu et al., 2021). All methods are implemented using the standardized configurations provided by the EasyEdit framework (Wang et al., 2024), a unified library for model editing. See the Appendix A for more details.
+
+Metric: We use Exact Match (EM) and F1-score (F1), which are standard metrics (Rajpurkar, 2016) in question answering for evaluating answer accuracy and completeness.
+
+Implementation Details: All experiments, including data construction, knowledge editing, and evaluation, were conducted on workstations with 8×NVIDIA RTX A6000 GPUs. For training the hypernetwork, we used the AdamW (Loshchilov and Hutter, 2019) optimizer for 1 epoch, with the learning rate decaying from 1e-4 to 1e-6 using cosine annealing. The expert module, implemented as a
+
+![](images/a02a7a4098ec041176df08d5266b63aab27e745d742067ad6cb93c9ce6a9e803.jpg)  
+Figure 3: Normalized EM for expert insertion at different layers using Llama2 $_{7B}$  on the NQ dataset.
+
+low-rank adaptation with  $k = 512$ , is integrated into the  $12_{th}$  layer of the transformer.
+
+# 4.2 Selection of Effective Layer
+
+To identify the most effective layer for expert insertion, we conducted a comprehensive layer-wise analysis by injecting experts at each transformer layer using QA pairs from the validation set, as illustrated in Figure 3. This experiment allows us to systematically examine how well injected knowledge is utilized across different depths of the model.
+
+For the Llama2 $_{7B}$  model, we observe that inserting experts at the  $12_{th}$  layer leads to the best performance. Since Baichuan $_{27B}$  adopts an architecture that is structurally consistent with Llama2, we follow the same choice for expert insertion to ensure comparability. Similarly, we extend the analysis to Qwen $_{2.5_{7B}}$ , which also shares a comparable transformer structure, and therefore apply the same  $12_{th}$  layer insertion strategy for consistency across models.
+
+Notably, the superior performance of the  $12_{th}$  layer aligns with the intuition that middle layers of transformer models are particularly effective at integrating semantic information. Prior work (Hendel et al., 2023) has also highlighted that intermediate layers capture task-specific representations more robustly, making them a natural fit for injecting external knowledge. This consistency between our empirical findings and theoretical insights further validates the choice of middle-layer insertion as a principled design decision.
+
+# 4.3 Main Results and Analysis
+
+We evaluate the effectiveness of GenPoE under two distinct evaluation paradigms: the W/o-RAG Setting and the Vanilla-RAG Setting, as summarized in Table 1.
+
+W/o-RAG Setting: In the W/o-RAG Setting, models are directly evaluated on the updated pa
+
+<table><tr><td rowspan="3">Model</td><td rowspan="3">Methods</td><td colspan="4">W/o-RAG Setting</td><td colspan="4">Vanilla-RAG Setting</td></tr><tr><td colspan="2">NQ</td><td colspan="2">TQA</td><td colspan="2">NQ</td><td colspan="2">TQA</td></tr><tr><td>EM</td><td>F1</td><td>EM</td><td>F1</td><td>EM</td><td>F1</td><td>EM</td><td>F1</td></tr><tr><td rowspan="6">Llama27B</td><td>Base</td><td>7.80</td><td>16.89</td><td>49.47</td><td>59.38</td><td>13.60</td><td>23.46</td><td>54.40</td><td>67.49</td></tr><tr><td>MEMIT</td><td>1.01</td><td>1.97</td><td>6.45</td><td>8.62</td><td>1.01</td><td>6.36</td><td>8.60</td><td>14.45</td></tr><tr><td>MEND</td><td>0.13</td><td>0.17</td><td>0.07</td><td>0.24</td><td>0.13</td><td>0.20</td><td>0.13</td><td>0.40</td></tr><tr><td>MALMEN</td><td>24.53</td><td>33.18</td><td>34.73</td><td>45.35</td><td>27.07</td><td>36.43</td><td>30.80</td><td>41.92</td></tr><tr><td>FT-LoRA</td><td>19.53</td><td>29.33</td><td>54.60</td><td>63.96</td><td>21.20</td><td>31.02</td><td>58.53</td><td>67.98</td></tr><tr><td>GenPoE</td><td>33.60</td><td>42.05</td><td>60.73</td><td>66.98</td><td>45.13</td><td>54.82</td><td>70.07</td><td>76.61</td></tr><tr><td rowspan="6">Baichuan27B</td><td>Base</td><td>6.07</td><td>14.54</td><td>43.47</td><td>53.07</td><td>10.47</td><td>19.81</td><td>50.47</td><td>60.18</td></tr><tr><td>MEMIT</td><td>2.50</td><td>3.50</td><td>0.84</td><td>2.47</td><td>1.67</td><td>3.86</td><td>0.84</td><td>2.01</td></tr><tr><td>MEND</td><td>15.33</td><td>22.70</td><td>24.60</td><td>32.97</td><td>17.27</td><td>25.97</td><td>27.06</td><td>36.21</td></tr><tr><td>MALMEN</td><td>22.93</td><td>31.97</td><td>30.93</td><td>42.23</td><td>20.53</td><td>29.66</td><td>28.00</td><td>38.86</td></tr><tr><td>FT-LoRA</td><td>19.87</td><td>28.56</td><td>49.73</td><td>56.80</td><td>38.07</td><td>48.78</td><td>64.67</td><td>72.47</td></tr><tr><td>GenPoE</td><td>24.53</td><td>32.12</td><td>52.80</td><td>58.44</td><td>42.33</td><td>52.96</td><td>69.80</td><td>75.56</td></tr><tr><td rowspan="6">Qwen2.57B</td><td>Base</td><td>4.20</td><td>10.43</td><td>36.60</td><td>43.96</td><td>28.80</td><td>40.08</td><td>61.67</td><td>70.41</td></tr><tr><td>MEMIT</td><td>0.98</td><td>1.94</td><td>5.56</td><td>7.78</td><td>0.96</td><td>5.43</td><td>7.70</td><td>12.33</td></tr><tr><td>MEND</td><td>0.00</td><td>0.00</td><td>0.00</td><td>0.09</td><td>0.00</td><td>0.00</td><td>0.00</td><td>0.23</td></tr><tr><td>MALMEN</td><td>21.40</td><td>28.90</td><td>19.00</td><td>26.28</td><td>40.67</td><td>50.74</td><td>21.93</td><td>31.08</td></tr><tr><td>FT-LoRA</td><td>21.66</td><td>31.43</td><td>47.33</td><td>54.46</td><td>40.07</td><td>51.61</td><td>64.93</td><td>73.32</td></tr><tr><td>GenPoE</td><td>24.53</td><td>32.84</td><td>51.13</td><td>56.68</td><td>45.20</td><td>55.52</td><td>70.73</td><td>76.91</td></tr></table>
+
+Table 1: Performance of Llama2 $_{7B}$ , Baichuan2 $_{7B}$  and Qwen2.5 $_{7B}$  on the NQ and TQA datasets under W/o-RAG and Vanilla-RAG settings. In the W/o-RAG setting, only the test question  $q_{test}$  is provided as input to the model, while in the Vanilla-RAG setting, the input consists of a concatenation of the retrieved passage and  $q_{test}$ , i.e., [retrieved passage,  $q_{test}$ ].
+
+rameters using only the test question  $q_{test}$ , without access to retrieved passages. This setup isolates the effect of model editing alone. Results indicate that GenPoE significantly outperforms existing editing baselines (e.g., MEMIT, MEND, MALMEN, and FT-LoRA) across all datasets and both model backbones. The results demonstrate that GenPoE can more effectively incorporate external knowledge at the passage level, leading to improved performance. In contrast, other editing methods show minimal or negligible gains, highlighting their limited applicability in real-world QA scenarios.
+
+Vanilla-RAG Setting: Models receive both the retrieved passages and the test question as input, allowing them to leverage explicit external knowledge during inference. Even in this more favorable setting, GenPoE continues to outperform both the base RAG models and fine-tuning baselines. This suggests that GenPoE not only enhances the model's internal knowledge via editing but also synergizes effectively with external retrieval, yielding substantial improvements over standard retrieval-augmented approaches.
+
+# 4.4 Effect of the Number of Passage for Enhancement
+
+Based on the trends illustrated in the Figure 4, we can observe the influence of passage quantity on model performance. As the number of retrieved passages increases during inference, Vanilla-RAG shows a noticeable decline in both EM and F1 scores. This degradation suggests that longer input contexts, caused by more passages, make it increasingly difficult for the model to focus on the correct evidence, thereby hampering answer accuracy.
+
+In contrast, GenPoE demonstrates greater robustness to the increasing number of passages. This is largely due to its multi-expert architecture, which distributes the processing passages load across multiple specialized experts. Each expert is exposed to one passage, which helps maintain context clarity and ensures the preservation of answer quality even when more passages are retrieved. These results highlight the scalability advantage of GenPoE in handling expanded input without suffering from the noise and distraction often introduced by lengthy contexts in conventional RAG approaches.
+
+![](images/0e1f62c0a46addcd216044fc4918acd688ee38d4eea351039bff5cbf76b3182a.jpg)  
+Figure 4: Ablation study on inference with varying numbers of passages, using Llama2 $_{7B}$  on the NQ dataset.
+
+# 4.5 Sequential Enhancement Progress
+
+Recent works (Gupta et al., 2024; Jiang et al., 2024b) has highlighted the limitations of current knowledge editing methods in continual learning or editing settings. Specifically, sequential edits often lead to catastrophic forgetting of previously injected knowledge and may cause interference or conflicts between newly added and existing information. In contrast, our proposed method demonstrates superior robustness in continual editing scenarios.
+
+As shown in Figure 5, we divide 1500 editing instances into batches of 250 and compute the average EM and F1 scores after each batch. Each diagonal block in the heatmap represents the model's performance on a batch of edited samples immediately after their insertion. The results indicate that our method maintains strong performance across editing steps. This advantage primarily stems from the design of our method, where each passage is associated with an independent and pluggable expert module. This modular structure enables new knowledge to be integrated without overwriting previously acquired information, thereby alleviating forgetting and enhancing stability throughout the continual editing process.
+
+# 4.6 Robustness Analysis
+
+To assess robustness under challenging conditions, we evaluate models in the Oracle-negative setting, where retrieved passages do not contain the ground truth answer. This simulates a realistic open-domain QA scenario with incomplete information. As shown in Table 2, Vanilla-RAG performs similarly to the Base W/o-RAG baseline, indicating limited benefit without answer-containing passages. This suggests Vanilla-RAG heavily depends on explicit answer spans. In contrast, Gen-
+
+![](images/bb2d874e000ce0cf5cec36e54461ae3c1f21c6afb0ccf79e1a190507dc7b9dfc.jpg)
+
+![](images/fbc8e1e20c2ba3495b7eb7280c6d23845387d89cc9f272f1d31dcb4707a8b832.jpg)  
+Figure 5: Continuous learning study of using Llama2 $_{7B}$  on the NQ dataset. Continuous learning study of using Llama2 $_{7B}$  on the NQ dataset. The upper part shows results under the W/o-RAG setting, while the lower part corresponds to the W/ RAG setting.
+
+![](images/b9a8cdb0285ea54952659549325611e1046a24ceaeb33a959069aaaf332942e6.jpg)
+
+![](images/f8dad9165428b9c46b81a4f760e23bba5947aecc95122c16a645077d0a5ab301.jpg)
+
+<table><tr><td>Method</td><td>EM</td><td>F1</td></tr><tr><td>Base W/o-RAG</td><td>7.80</td><td>16.89</td></tr><tr><td>Base Vanilla-RAG</td><td>7.33</td><td>16.12</td></tr><tr><td>GenPoE (W/o RAG)</td><td>32.13</td><td>40.98</td></tr><tr><td>GenPoE (W/ RAG)</td><td>29.53</td><td>38.24</td></tr></table>
+
+Table 2: Robustness analysis under retrieval failure: results of Llama2 $_{7B}$  on the NQ dataset.
+
+PoE remains robust. Even without RAG, it achieves strong results, significantly outperforming baselines. With RAG on non-answer passages, GenPoE still performs competitively, showing its ability to leverage partial or noisy content through its expert-based design.
+
+These results demonstrate that GenPoE not only excels in standard settings but also preserves its advantage in robustness under conditions where conventional retrieval-augmented methods fail.
+
+# 4.7 Efficiency Analysis
+
+The efficiency comparison in Table 3 demonstrates the superior speed of our method. Benefiting from the lightweight design of the hypernetwork, GenPoE integrates paragraph-level information into the model with minimal injection time (0.006s), significantly outperforming other editing methods. Additionally, since expert modules are only inserted at a single transformer layer, the increase in inference
+
+<table><tr><td>Method</td><td>Injection</td><td>Inference</td><td>All</td></tr><tr><td>Base</td><td>-</td><td>0.427s</td><td>0.427s</td></tr><tr><td>MEMIT</td><td>52.494s</td><td>0.427s</td><td>52.921s</td></tr><tr><td>MEND</td><td>1.258s</td><td>0.427s</td><td>1.685s</td></tr><tr><td>MALMEN</td><td>1.717s</td><td>0.427s</td><td>2.144s</td></tr><tr><td>FT-LoRa</td><td>7.646s</td><td>0.427s</td><td>8.073s</td></tr><tr><td>GenPoE</td><td>0.006s</td><td>0.447s</td><td>0.453s</td></tr></table>
+
+Table 3: Efficiency analysis on the NQ dataset using the Llama2 $_{7B}$  model
+
+time is negligible (from 0.427s to 0.447s), maintaining efficiency comparable to the base model.
+
+Another advantage of our design lies in its suitability for real-world deployment. Unlike approaches that require persistent storage of newly learned experts, GenPoE employs a lightweight low-rank decomposition, representing each expert as a compact parameter delta dynamically generated by the hypernetwork based on the retrieved passage. The expert module is ephemeral—created on-the-fly for the current forward pass and discarded immediately afterward—ensuring no long-term accumulation of experts or memory growth. Since generation is conditioned on top-retrieved passages (typically top-1), the number of active experts per query is strictly limited, which avoids runtime slowdown or memory bottlenecks even in knowledge-intensive scenarios. This dynamic and stateless design not only guarantees minimal computational overhead but also enables scalable, efficient, and practical deployment in real applications.
+
+# 5 Conclusion
+
+In this work, we propose GenPoE, a hypernetwork-driven mixture-of-experts framework for passage-level knowledge enhancement, designed to address the limitations of large language models in effectively leveraging external prompts. Given a query, relevant passages are dynamically encoded into expert modules via a hypernetwork, allowing knowledge to be injected into the model on-the-fly. This design ensures that the generated experts are ephemeral and lightweight, leading to minimal computational overhead and negligible memory growth. Such properties make GenPoE not only effective in enhancing factual knowledge utilization but also highly suitable for real-world deployment scenarios, where scalability and efficiency are critical.
+
+Extensive experiments on NQ and TQA demonstrate that GenPoE successfully incorporates knowledge while avoiding the degradation typically associated with general parameter-update-based methods. The framework achieves strong editing effectiveness without sacrificing inference efficiency, validating its practicality for knowledge-intensive applications.
+
+For future work, we plan to explore more efficient and expressive hypernetwork and mixture-of-experts architectures, as well as stable continual learning techniques to further improve adaptability and long-term scalability. Moreover, we aim to extend the framework to more challenging multi-hop reasoning tasks, where multiple passages must be integrated and jointly reasoned over. This would push GenPoE beyond single-hop factual injection toward a more general paradigm for dynamic knowledge integration in large language models.
+
+# Limitations
+
+In the current setup, while GenPoE demonstrates strong performance in passage-level knowledge injection, it relies on a hypernetwork to dynamically generate expert parameters conditioned on retrieved passages. The effectiveness of this generation process is inherently constrained by the representational capacity of the hypernetwork. In more complex tasks involving nuanced or diverse knowledge, the current design may struggle to produce accurate and generalizable experts, potentially limiting the model's ability to fully exploit the retrieved information. Future work should explore more expressive and robust hypernetwork architectures to improve expert generation stability and scalability under challenging settings.
+
+# Acknowledgments
+
+This work was supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No.RS-2020-II201336, Artificial Intelligence Graduate School Program(UNIST)) and IITP grant funded by the Korea government(MSIT) (No.RS-2023-00216011, Development of artificial complex intelligence for conceptually understanding and inferring like human). Xuebing Liu and Shanbao Qiao were also supported by China Scholarship Council (CSC).
+
+# 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.  
+Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. 2017. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3366-3375.  
+Szymon Antoniak, Sebastian Jasczur, Michal Krutul, Maciej Pioro, Jakub Krajewski, Jan Ludziejewski, Tomasz Odrzygoźdź, and Marek Cygan. 2023. Mixture of tokens: Efficient llms through cross-example aggregation. arXiv preprint arXiv:2310.15961.  
+Baichuan. 2023. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305.  
+Tong Chen, Hao Fang, Patrick Xia, Xiaodong Liu, Benjamin Van Durme, Luke Zettlemoyer, Jianfeng Gao, and Hao Cheng. 2025. Generative adapter: Contextualizing language models in parameters with a single forward pass. In The Thirteenth International Conference on Learning Representations.  
+Siyuan Cheng, Bozhong Tian, Qingbin Liu, Xi Chen, Yongheng Wang, Huajun Chen, and Ningyu Zhang. 2023. Can we edit multimodal large language models? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13877-13888, Singapore. Association for Computational Linguistics.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1-113.  
+Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164.  
+Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. 2023. Loramoe: Revolutionizing mixture of experts for maintaining world knowledge in language model alignment. arXiv preprint arXiv:2312.09979, 4(7).  
+Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. 2022. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR.  
+William Fedus, Barret Zoph, and Noam Shazeer. 2022. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39.
+
+Hengrui Gu, Kaixiong Zhou, Xiaotian Han, Ninghao Liu, Ruobing Wang, and Xin Wang. 2023. Pokemqa: Programmable knowledge editing for multi-hop question answering. arXiv preprint arXiv:2312.15194.  
+Akshat Gupta, Anurag Rao, and Gopala Anumanchipalli. 2024. Model editing at scale leads to gradual and catastrophic forgetting. arXiv preprint arXiv:2401.07453.  
+Muhammad Usman Hadi, Rizwan Qureshi, Abbas Shah, Muhammad Irfan, Anas Zafar, Muhammad Bilal Shaikh, Naveed Akhtar, Jia Wu, Seyedali Mirjalili, et al. 2023. A survey on large language models: Applications, challenges, limitations, and practical usage. Authorea Preprints.  
+Roee Hendel, Mor Geva, and Amir Globerson. 2023. In-context learning creates task vectors. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 9318-9333, Singapore. Association for Computational Linguistics.  
+Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.  
+Zeyu Huang, Yikang Shen, Xiaofeng Zhang, Jie Zhou, Wenge Rong, and Zhang Xiong. Transformer-patcher: One mistake worth one neuron. In The Eleventh International Conference on Learning Representations.  
+Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. 1991. Adaptive mixtures of local experts. Neural computation, 3(1):79-87.  
+Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. 2024a. Mixtral of experts. arXiv preprint arXiv:2401.04088.  
+Houcheng Jiang, Junfeng Fang, Tianyu Zhang, An Zhang, Ruipeng Wang, Tao Liang, and Xi-ang Wang. 2024b. Neuron-level sequential editing for large language models. arXiv preprint arXiv:2410.04045.  
+Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551.  
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and
+
+Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781, Online. Association for Computational Linguistics.  
+Zixuan Ke, Yijia Shao, Haowei Lin, Tatsuya Konishi, Gyuhak Kim, and Bing Liu. 2023. Continual pretraining of language models. In The Eleventh International Conference on Learning Representations.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453-466.  
+Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2020. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668.  
+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.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190.  
+Xiaopeng Li, Shasha Li, Shezheng Song, Jing Yang, Jun Ma, and Jie Yu. 2024. Pmet: Precise model editing in a transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18564-18572.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022a. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359-17372.  
+Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2022b. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229.  
+Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2021. Fast model editing at scale. arXiv preprint arXiv:2110.11309.  
+Xiaonan Nie, Xupeng Miao, Shijie Cao, Lingxiao Ma, Qibin Liu, Jilong Xue, Youshan Miao, Yi Liu, Zhi Yang, and Bin Cui. 2021. Evomoe: An evolutionary mixture-of-experts training framework via dense-tosparse gate. arXiv preprint arXiv:2112.14397.
+
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744.  
+Bowen Pan, Yikang Shen, Haokun Liu, Mayank Mishra, Gaoyuan Zhang, Aude Oliva, Colin Raffel, and Rameswar Panda. 2024. Dense training, sparse inference: Rethinking training of mixture-of-experts language models. arXiv preprint arXiv:2404.05567.  
+P Rajpurkar. 2016. Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250.  
+Carl Rasmussen and Zoubin Ghahramani. 2001. Infinite mixtures of gaussian process experts. Advances in neural information processing systems, 14.  
+Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538.  
+Kihyuk Sohn. 2016. Improved deep metric learning with multi-class n-pair loss objective. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc.  
+Weihang Su, Yichen Tang, Qingyao Ai, Junxi Yan, Changyue Wang, Hongning Wang, Ziyi Ye, Yu-jia Zhou, and Yiqun Liu. 2025. Parametric retrieval augmented generation. arXiv preprint arXiv:2501.15915.  
+Chenmien Tan, Ge Zhang, and Jie Fu. 2023. Massive editing for large language models via meta learning. arXiv preprint arXiv:2311.04661.  
+Yuqiao Tan, Shizhu He, Huanxuan Liao, Jun Zhao, and Kang Liu. 2025. Better wit than wealth: Dynamic parametric retrieval augmented generation for test-time knowledge enhancement. arXiv preprint arXiv:2503.23895.  
+Qwen Team. 2024. Qwen2.5: A party of foundation models.  
+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.  
+Yiteng Tu, Weihang Su, Yujia Zhou, Yiqun Liu, and Qingyao Ai. 2025. Rbft: Robust fine-tuning for retrieval-augmented generation against retrieval defects. arXiv preprint arXiv:2501.18365.  
+Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. 2022. Dylora: Parameter efficient tuning of pre-trained models using dynamic
+
+search-free low-rank adaptation. arXiv preprint arXiv:2210.07558.  
+Peng Wang, Ningyu Zhang, Bozhong Tian, Zekun Xi, Yunzhi Yao, Ziwen Xu, Mengru Wang, Shengyu Mao, Xiaohan Wang, Siyuan Cheng, Kangwei Liu, Yansheng Ni, Guozhou Zheng, and Huajun Chen. 2024. EasyEdit: An easy-to-use knowledge editing framework for large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 82-93, Bangkok, Thailand. Association for Computational Linguistics.  
+Renzhi Wang and Piji Li. 2024. Memoe: Enhancing model editing with mixture of experts adaptors. arXiv preprint arXiv:2405.19086.  
+Xun Wu, Shaohan Huang, and Furu Wei. 2024. Mixture of lora experts. arXiv preprint arXiv:2404.13628.  
+Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, pages 641-649.  
+An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Hao ran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671.  
+Wanli Yang, Fei Sun, Jiajun Tan, Xinyu Ma, Qi Cao, Dawei Yin, Huawei Shen, and Xueqi Cheng. 2025. The mirage of model editing: Revisiting evaluation in the wild. arXiv preprint arXiv:2502.11177.  
+Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. Editing large language models: Problems, methods, and opportunities. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10222-10240, Singapore. Association for Computational Linguistics.  
+Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2023. Making retrieval-augmented language models robust to irrelevant context. arXiv preprint arXiv:2310.01558.  
+Lang Yu, Qin Chen, Jie Zhou, and Liang He. 2024. Melo: Enhancing model editing with neuron-indexed
+
+dynamic lora. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19449-19457.  
+Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223.  
+Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023a. Can we edit factual knowledge by in-context learning? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4862-4876, Singapore. Association for Computational Linguistics.  
+Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. 2023b. Can we edit factual knowledge by in-context learning? arXiv preprint arXiv:2305.12740.  
+Zexuan Zhong, Zhengxuan Wu, Christopher D Manning, Christopher Potts, and Danqi Chen. 2023. Mquake: Assessing knowledge editing in language models via multi-hop questions. arXiv preprint arXiv:2305.14795.
+
+# A Limited Editing Approaches
+
+Most existing knowledge editing or augmentation methods are designed and evaluated under idealized settings, making them difficult to generalize or apply directly to real-world scenarios (Yang et al., 2025) (See the Table 4). In this section, we categorize and analyze current editing approaches by their limitations in three key dimensions: Editing Input, Generation Strategy, and Output Truncation.
+
+Editing Input: Some representative methods, such as ROME and MEMIT, impose strict constraints on the editing input format. Specifically, they require a structured triple input, typically represented as (subject, relation, target). This restricts their applicability in open-domain or flexible editing tasks where knowledge is expressed more freely or contextually, such as full passages or natural questions. In contrast, MEND and MALMEN relax this constraint by allowing more flexible QA-style inputs. However, these methods still require that the editing and testing queries follow the same format and linguistic style. As a result, they struggle to support use cases like updating passage-level knowledge based on question-answer inputs. Our proposed method, GenPoE, overcomes these limitations by leveraging a hypernetwork that generates expert parameters from diverse types of knowledge input. This flexibility allows GenPoE to support
+
+<table><tr><td>Method</td><td>Input Type</td><td>Generation Strategy</td><td>Output Truncation</td><td>Real-world Friendliness</td></tr><tr><td>ROME</td><td>Triple data</td><td>teacher forcing</td><td>ground truth length</td><td>⊙</td></tr><tr><td>MEMIT</td><td>Triple data</td><td>teacher forcing</td><td>ground truth length</td><td>⊙</td></tr><tr><td>MEND</td><td>Context-free</td><td>teacher forcing</td><td>ground truth length</td><td>⊙</td></tr><tr><td>MALMEN</td><td>Context-free</td><td>teacher forcing</td><td>ground truth length</td><td>⊙</td></tr><tr><td>GenPoE</td><td>Context-guided</td><td>autoregressive decoding</td><td>natural stopping criteria</td><td>⊙</td></tr></table>
+
+editing at the passage, sentence, or even entity level. To ensure fair comparison during experiments, we convert the passage-level knowledge into formats compatible with MEMIT, MEND, and MALMEN. Details of this transformation process can be found in Appendix D.
+
+Generation Strategy: Another major limitation of prior work lies in the generation strategy. Most existing methods (ROME (Meng et al., 2022a), MEMIT (Meng et al., 2022b), MEND (Mitchell et al., 2021), MALMEN (Tan et al., 2023)) use teacher forcing during generation. While this simplifies training and evaluation, it leads to overly optimistic results, as the model is always conditioned on the ground truth tokens. This setting hides generation errors that may occur in actual deployment, such as token drift or hallucination. In contrast, GenPoE adopts a fully autoregressive decoding process, better simulating real inference scenarios. This strategy allows us to assess the true quality and robustness of the edited model outputs.
+
+Output Truncation: Output truncation strategies in prior works further exaggerate their performance under unrealistic assumptions. Existing methods commonly truncate model outputs to the same length as the gold answer. While this creates a cleaner evaluation setup, it masks issues like repetition, irrelevance, or incomplete answers that naturally arise when relying on the model's own stopping behavior. Our method avoids this artificial constraint by allowing natural stopping criteria during decoding. This setup more accurately reflects real-world model usage, where answers are generated until a learned end-of-sequence signal is reached.
+
+As summarized in Table 4, these limitations collectively make many prior methods less suitable for real-world deployment. Our method, GenPoE, addresses all three limitations, making it significantly more adaptable and effective in practical knowledge editing scenarios.
+
+Table 4: Comparison of current editing methods in real-world applications. Ⓢ: unfriendly, Ⓞ: moderately friendly, and ⓥ: highly friendly.  
+
+<table><tr><td>Top-K</td><td>Type</td><td>NQ</td><td>TQA</td></tr><tr><td rowspan="2">@1</td><td>Before</td><td>44.60</td><td>56.53</td></tr><tr><td>After</td><td>64.67</td><td>76.33</td></tr><tr><td rowspan="2">@2</td><td>Before</td><td>55.73</td><td>65.27</td></tr><tr><td>After</td><td>71.73</td><td>79.60</td></tr><tr><td rowspan="2">@4</td><td>Before</td><td>64.47</td><td>72.07</td></tr><tr><td>After</td><td>77.87</td><td>82.80</td></tr><tr><td rowspan="2">@8</td><td>Before</td><td>72.93</td><td>76.73</td></tr><tr><td>After</td><td>81.47</td><td>84.53</td></tr></table>
+
+Table 5: Comparison of DPR retrieval accuracy results before and after applying the re-ranker we trained.
+
+# B Details of Reranker
+
+This section presents the performance of applying a re-ranker to refine the retrieval results of DPR (Karpukhin et al., 2020), comparing the accuracy before and after re-ranking. In a question-answering system, the accuracy of the retrieval source plays a crucial role in the overall result. Additionally, the scoring produced by the re-ranker is important for merging between experts. The improvements were evaluated using Top-K accuracy, which measures the presence of a golden passage, ensuring a stable document source for responding answer.
+
+# C Case Study
+
+Table 6 presents a case study illustrating the effectiveness of different methods in answering the question "Who produced A Change Is Gonna Come?" The target answer is "Hugo & Luigi".
+
+The retrieved top-1 document provides relevant information, stating that the song was produced by "Hugo & Luigi". However, both the Base and Vanilla-RAG methods incorrectly answer "Sam Cooke", indicating a failure to extract the precise producer information from the retrieved passage.
+
+In contrast, the GenPoE method, whether applied without retrieval augmentation (W/o RAG) or with retrieval augmentation (W/ RAG), successfully generates the correct answer "Hugo & Luigi",
+
+Question: Who produced a change is gonna come?
+
+# Target answer: Hugo & Luigi
+
+Retrieved Top-1 Documents: A Change Is Gonna Come" is a song by American recording artist Sam Cooke. It initially appeared on Cooke's album "Ain't That Good News", released March 1, 1964 by RCA Victor; a slightly edited version of the recording was released as a single on December 22, 1964. Produced by Hugo & Luigi and arranged and conducted by René Hall, ...
+
+<table><tr><td>Method</td><td>Answer</td><td>Status</td></tr><tr><td>Base</td><td>Sam Cooke</td><td>X</td></tr><tr><td>Vanilla-RAG</td><td>Sam Cooke</td><td>X</td></tr><tr><td>GenPoE (W/o RAG)</td><td>Hugo &amp; Luigi</td><td>✓</td></tr><tr><td>GenPoE (W/ RAG)</td><td>Hugo &amp; Luigi</td><td>✓</td></tr></table>
+
+Table 6: Case Study Example
+
+demonstrating its superior ability to leverage the retrieved evidence for precise question answering.
+
+This case exemplifies the advantage of Gen-PoE in extracting accurate factual knowledge from retrieved documents compared to baseline approaches.
+
+# D Generation of Editing Facts
+
+As shown in Table 7, we first adopt template to generate synthetic QA pairs tailored for the MEND, MALMEN, and FT-LoRA methods. Specifically, the template incorporates a structured instruction along with the retrieved passage, which is inserted into the placeholder " $\{\text{paragraph}\}$ ", and is then fed into a large language model "gpt4o-mini" (Ouyang et al., 2022; Achiam et al., 2023) to produce concise, exact-match question-answer pairs suitable for knowledge editing supervision.
+
+Building on these QA pairs, we further apply template of Table 8 to transform them into structured (subject, relation, object), which are used as training instances for the MEMIT method (Meng et al., 2022b). To generate data suitable for MEMIT's knowledge injection mechanism, the template places the generated question into the {"prompt"} placeholder, ensuring alignment with the required input format.
+
+# Prompt template for self-generating synthetic knowledge
+
+You are an assistant who is good at organizing questions and answers from paragraphs. Here is an example.
+
+Paragraph: "on death row in the United States on January 1, 2013. Since 1977, the states of Texas (464), Virginia (108) and Oklahoma (94) have executed the most death row inmates. , California (683), Florida (390), Texas (330) and Pennsylvania (218) housed more than half of all inmates pending on death row. , the longest-serving prisoner on death row in the US who has been executed was Jack Alderman who served over 33 years. He was executed in Georgia in 2008. However, Alderman only holds the distinction of being the longest-serving executed inmate so far. A Florida inmate, Gary Alvord, arrived"
+
+1. Q: How many death row inmates did Texas execute since 1977?
+
+A:464
+
+2. Q: Which state executed 108 death row inmates since 1977?
+
+A: Virginia
+
+3. Q: How many death row inmates did Oklahoma execute since 1977?
+
+A:94
+
+4. Q: Which state housed 683 death row inmates as of January 1, 2013?
+
+A: California
+
+5. Q: How many inmates did Florida house on death row?
+
+A:390
+
+6. Q: How many death row inmates did Texas have pending?
+
+A:330
+
+7. Q: How many death row inmates did Pennsylvania house?
+
+A:218
+
+8. Q: Who was the longest-serving prisoner on death row who was executed?
+
+A: Jack Alderman
+
+9. Q: How many years did Jack Alderman serve on death row?
+
+A: over 33 years
+
+10. Q: In which year was Jack Alderman executed?
+
+A:2008
+
+11. Q: Which state executed Jack Alderman?
+
+A: Georgia
+
+12. Q: Who is noted as the longest-serving "executed" inmate?
+
+A: Jack Alderman
+
+13. Q: Which inmate arrived in Florida?
+
+A: Gary Alvord
+
+14. Q: What is the date referenced for death row statistics in the passage?
+
+A:January 1,2013
+
+15. Q: Since when has the execution data been tracked in this passage?
+
+A:1977
+
+16. Q: What constitutes more than half of all inmates pending on death row?
+
+A: California, Florida, Texas, and Pennsylvania
+
+Please follow the format of the example above to generate sixteen questions and corresponding answers for the following Paragraph. The format of answers should be a very short phrase from paragraph, such as "464", "2008", "May 16th, 1931", or "Jack Alderman", to meet the criteria of exact match Paragraph.
+
+Paragraph:" $\{\text{paragraph}\}$ "
+
+Table 7: This prompt template is designed to generate synthetic question-answer pairs from a passage. It includes clear instructions outlining the requirements, along with an example paragraph and its corresponding question-answer pairs. The model is then expected to create similar question-answer pairs for new paragraphs following this format.
+
+# Prompt template for generating subject extraction
+
+Read the following prompt carefully. Identify the subject of the sentence. Output must be only the subject word, exactly as it appears in the 'prompt' — preserving the original capitalization and formatting.
+
+Here are some examples for guidance:
+
+'prompt': 'Which religion is noted as the fourth largest after Christianity, Islam, and Hinduism?', 'subject': 'religion'
+
+'prompt': 'Who was mainly responsible for the design of Abney Park Chapel?', 'subject': 'design of Abney Park Chapel'
+
+'prompt': "What layer of skin is directly below the dermis and epidermis?", 'subject': 'layer of skin'
+
+'prompt': "What mountain range is Aconcagua Provincial Park part of?", 'subject': "mountain range"
+
+Based on the examples, for 'prompt':  $\{\mathrm{prompt}\}$ , subject:
+
+Table 8: This prompt template is used to extract subjects and generate (subject, relation, target) triples for editing with the MEMIT method.

2025/GenPoE_ Generative Passage-level Mixture of Experts for Knowledge Enhancement of LLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0a14c32c0fba8934bcc75c78fb4dac8f54571ebd614712ff7cc0cd29491ec19
+size 538724

2025/Generation-Augmented Retrieval_ Rethinking the Role of Large Language Models in Zero-Shot Relation Extraction/06f5d856-0d14-45d9-9bc3-7006f452ae3b_content_list.json

@@ -0,0 +1,2267 @@
+[
+    {
+        "type": "text",
+        "text": "Generation-Augmented Retrieval: Rethinking the Role of Large Language Models in Zero-Shot Relation Extraction",
+        "text_level": 1,
+        "bbox": [
+            115,
+            85,
+            880,
+            124
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zehan Li, Fu Zhang*, Tianyue Peng, He Liu, Jingwei Cheng",
+        "bbox": [
+            220,
+            137,
+            769,
+            154
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "School of Computer Science and Engineering, Northeastern University, China Key Laboratory of Intelligent Computing in Medical Image of Ministry of Education Northeastern University, China",
+        "bbox": [
+            154,
+            154,
+            843,
+            203
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{lizehan1999,pengtianyue0716,liuhe0607}@163.com,",
+        "bbox": [
+            208,
+            204,
+            784,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{zhangfu,chengjingwei}@neu.edu.cn",
+        "bbox": [
+            300,
+            221,
+            699,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in Relation Extraction (RE) emphasize Zero-Shot methodologies, aiming to recognize unseen relations between entities with no annotated data. Although Large Language Models (LLMs) have demonstrated outstanding performance in many NLP tasks, their performance in Zero-Shot RE (ZSRE) without entity type constraints still lags behind Small Language Models (SLMs). LLM-based ZSRE often involves manual interventions and significant computational overhead, especially when scaling to large-scale multi-choice data.",
+        "bbox": [
+            141,
+            287,
+            460,
+            458
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To this end, we introduce RE-GAR-AD, which not only leverages the generative capability of LLMs but also utilizes their representational power without tuning LLMs. We redefine LLM-based ZSRE as a retrieval challenge, utilizing a Generation-Augmented Retrieval framework coupled with a retrieval ADjuster. Specifically, our approach guides LLMs through crafted prompts to distill sentence semantics and enrich relation labels. We encode sentences and relation labels using LLMs and match their embeddings in a triplet fashion. This retrieval technique significantly reduces token input requirements. Additionally, to further optimize embeddings, we propose a plug-in retrieval adjuster with only 2M parameters, which allows rapid fine-tuning without accessing LLMs' parameters. Our LLM-based model demonstrates comparable performance on multiple benchmarks.",
+        "bbox": [
+            141,
+            462,
+            460,
+            747
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            769,
+            258,
+            785
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Relation Extraction (RE) aims to select the relation from a predefined set for two target entities in a given sentence. Traditional supervised RE methods (Wu and He, 2019; Zheng et al., 2021) demand significant annotated data and struggle to handle",
+        "bbox": [
+            112,
+            795,
+            487,
+            876
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/08876b2a495cb73d83ab72adf7c7c4f85ae289f6983313f0e91ee9588478b53d.jpg",
+        "image_caption": [
+            "Figure 1: Comparison between (a) QA4RE (Zhang et al., 2023), (b) SUMASK (Li et al., 2023), and (c) our approach."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            259,
+            873,
+            428
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "unseen relations effectively. Recently, there has been a growing interest in Few-Shot RE (FSRE) (Yang et al., 2020) and Zero-Shot RE (ZSRE) (Levy et al., 2017; Gong and Eldardiry, 2021), which have limited or no labeled data for novel relations.",
+        "bbox": [
+            507,
+            491,
+            882,
+            570
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing research on ZSRE typically involves fine-tuning Small Language Models (SLMs)  $^{2}$  on predicting unseen relations (Zhao et al., 2023a; Li et al., 2024c). Large Language Models (LLMs) have demonstrated impressive performance in various downstream natural language processing (NLP) tasks (Zhao et al., 2023b), leveraging techniques such as In-Context Learning (ICL) (Ma et al., 2023c) and Chain of Thought (CoT) (Ma et al., 2023a) to address FSRE. Recently, LLM-based ZSRE studies have also made significant progress. QA4RE (Zhang et al., 2023) redefines ZSRE as a Question Answering (QA) task and manually construct templates for relation labels. It requires token inputs at the  $n \\times r$  level for  $n$  input sentences and  $r$  candidate relations (Figure 1 (a)). However, the inherent selection bias of LLMs may result in poor robustness (Zheng et al., 2024). SUMASK (Li et al., 2023) determines relations by calculating",
+        "bbox": [
+            507,
+            571,
+            884,
+            879
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "2We define SLMs as language models with less than 1B parameters that can be easily fine-tuned (Ma et al., 2023c).",
+        "bbox": [
+            507,
+            894,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "17928",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 17928-17941",
+        "bbox": [
+            208,
+            945,
+            786,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            280,
+            959,
+            715,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "uncertainty across  $k$  responses of LLMs. The recursive input leads to a significant  $n \\times k \\times r$  token input overhead (Figure 1 (b)), restricting its scalability on large-scale data. Therefore, making ZSRE more efficient and accurate while benefiting from LLMs is a key concern for the current system.",
+        "bbox": [
+            112,
+            84,
+            489,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we explore new possibilities for LLM-based ZSRE and propose the RE-GAR-AD framework, which is based on the Generation-Augmented Retrieval (GAR) paradigm and incorporates a retrieval ADjuster. Considering the potential of LLMs as encoders for obtaining embeddings (BehnamGhader et al., 2024), we rethink and redefine LLM-based ZSRE as an embedding-based retrieval problem. We input generation-augmented sentences and relation labels separately to LLMs to obtain their embeddings and conduct sentence and relation retrieval through matching scores. All candidate relation labels only need to be input once and kept fixed, requiring only token inputs at  $n + r$  level as shown in Figure 1 (c). This not only reduces the inference cost of large-scale multi-choice ZSRE data for LLMs but also demonstrates performance comparable to that of existing methods.",
+        "bbox": [
+            115,
+            187,
+            490,
+            479
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we leverage the generation capability of LLMs to improve retrieval efficiency. Considering that a sentence may contain more than one relation pattern, we distill valuable information for RE from sentences using Sentence Distillation Prompt, which includes the types of two target entities and their possible semantic relation. Additionally, we guide LLMs to utilize their internal general knowledge to generate deep information about relation labels through the Relation Enrichment Prompt, which explores possible entity types relevant to the relation. We then use LLMs to encode and obtain embeddings of sentences and relation labels, and achieve more accurate relation retrieval through Triplet-based Retrieval. Furthermore, SLMs-based ZSRE fine-tuning typically involves learning prior knowledge from seen data, which may be computationally expensive for LLMs and impractical for API-based LLMs (Yoon et al., 2023; Li et al., 2024b). Therefore, we carefully design a plug-in Retrieval Adjuster for the framework. It aims to learn the ability to align sentence and relation label representations from seen data, and further adjust and optimize the embeddings of unseen data without accessing the full parameters of LLMs, thereby improving retrieval efficiency. Our contributions are summarized as follows:",
+        "bbox": [
+            115,
+            487,
+            490,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We redefine LLM-based ZSRE as a retrieval task and propose a novel Generation-Augmented Retrieval (GAR) framework, significantly reducing the token inputs while ensuring accuracy.",
+            "- We carefully design two prompts further to process the semantics of sentences and relation labels using the generative ability of LLMs, improving retrieval efficiency.",
+            "- We introduce a retrieval adjuster plugin for the GAR framework, with only 2M parameters, which can enhance retrieval performance without accessing the parameters of LLMs.",
+            "- Experimental results show that RE-GAR-AD achieves competitive performance across five ZSRE benchmark datasets. Without involving fine-tuning, our approach outperforms the state-of-the-art LLM-based method by 6.93."
+        ],
+        "bbox": [
+            531,
+            84,
+            884,
+            409
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            422,
+            665,
+            437
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 LLMs for Language Representation",
+        "text_level": 1,
+        "bbox": [
+            509,
+            448,
+            838,
+            464
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models (LLMs) are typically trained on large corpora and have demonstrated advanced performance across many downstream NLP tasks (Zhao et al., 2023b). For tasks such as retrieval and clustering that require sentence embeddings, a common approach involves leveraging the representation of the end-of-sequence token of LLMs as sentence embeddings (Ma et al., 2023b). Considering that causal attention might limit the representation capability of decoder-only LLMs, recent works introduce bidirectional attention (Muennighoff et al., 2024) and echo mechanisms (Springer et al., 2024) to overcome limitations. These works demonstrate that generative LLMs are also proficient at language representation (Wang et al., 2023), and any robust decoder-only LLMs can produce universal text embeddings with minor adaptation (BehnamGhader et al., 2024).",
+        "bbox": [
+            507,
+            469,
+            884,
+            759
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 LLMs for Relation Extraction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            771,
+            793,
+            785
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The existing work on LLM-based RE can be categorized into three paradigms: Tuning-based methods involve fine-tuning LLMs with accessible parameters (Kim et al., 2023; Sun et al., 2024). Recent works adapt LLMs to RE tasks through annotation guidelines (Sainz et al., 2024) or tabular prompting (Li et al., 2024a). ICL-based methods facilitate In-Context Learning (ICL) (Min",
+        "bbox": [
+            507,
+            791,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "17929",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/e13709f946280b437af3d15445ba77cfc2b7514e1355f40fad8b2725695d88ca.jpg",
+        "image_caption": [
+            "Figure 2: An overview of the proposed RE-GAR-AD framework. It consists of three key components: (1) Generation-Augmented Embeddings, where Generative LLMs distill semantic information from sentences and relation descriptions using two prompts; (2) Retrieval Adjuster, which calibrates embeddings using relation patterns learned from seen data; and (3) Triplet-based Retrieval, where embedding triplets are matched to infer the final relation label."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            84,
+            880,
+            284
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "et al., 2022) by observing demonstrations within prompts without updating the LLMs parameters, which is applied in FSRE. They use the Retrieval-Augmented Generation (RAG) paradigm (Rubin et al., 2022) to retrieve suitable demonstrations, which may include quality analysis (Wan et al., 2023) or logical reasoning (Ma et al., 2023c). Instruction-based methods execute tasks through explicit instructions such as multiple-choice QA (Zhang et al., 2023) or uncertainty estimation (Li et al., 2023) without demonstrations, making them applicable to ZSRE.",
+        "bbox": [
+            112,
+            368,
+            487,
+            558
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A common limitation of the above methods is the increase in candidate relations along with a large number of token inputs, which consumes time and resources in large-scale ZSRE. Additionally, post-processing may be required to address output format issues. To address these challenges, we propose a novel LLM-based paradigm called Generative-Augmented Retrieval (GAR). Unlike RAG, GAR directly identifies relations through retrieval, fundamentally avoiding format issues and reducing computational costs by separately modeling sentences and relations. Furthermore, we devise a retrieval adjuster that can be plugged into LLMs to further improve ZSRE performance.",
+        "bbox": [
+            112,
+            563,
+            489,
+            788
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            804,
+            263,
+            821
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            833,
+            329,
+            848
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a predefined set of seen relations  $R_{s}$  and unseen relations  $R_{u}$ , with no overlap between them. Given seen data  $D_{s}$  and unseen data  $D_{u}$ . The ZSRE's objective is to predict  $r_{u} \\in R_{u}$",
+        "bbox": [
+            112,
+            857,
+            487,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "between the head entity  $e^1$  and the tail entity  $e^2$  in the input sentence  $x_u$  for unseen instance  $I_u = (x_u, e_u^1, e_u^2) \\in D_u$ . Before prediction, learning can be conducted from seen instances  $I_s = (x_s, e_s^1, e_s^2, r_s) \\in D_s, r_s \\in R_s$ , and the knowledge can be transferred to predict.",
+        "bbox": [
+            507,
+            368,
+            882,
+            464
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Overview of RE-GAR-AD",
+        "text_level": 1,
+        "bbox": [
+            507,
+            476,
+            759,
+            491
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The proposed RE-GAR-AD framework, which redefines LLM-based ZSRE as an embedding-based retrieval problem, consists of three modules as shown in Figure 2:",
+        "bbox": [
+            507,
+            499,
+            882,
+            563
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(1) Generation-Augmented Embeddings involves the use of Generative LLMs to distill and enrich sentences and relation labels through two carefully designed prompts. These are then encoded by Embedding LLMs to obtain fixed-dimensional embedding triplets. (2) Retrieval Adjuster learns from the relation patterns of the seen dataset  $D_{s}$  and is used to adjust the embedding triplets corresponding to the unseen sentences and relations, thereby improving retrieval performance. It is worth noting that we use the same data processing method in the testing phase as in the training phase, applying both Generation Augmentation and Retrieval Adjustment. (3) Through Triplet-based Retrieval, the adjusted sentence embedding triplets are matched with the adjusted relation embedding triplets, and the highest scoring relation is chosen as the final relation.",
+        "bbox": [
+            507,
+            564,
+            882,
+            852
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 Generation-Augmented Embeddings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            842,
+            882
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Sentence Distillation Prompt. To distill crucial content for RE from sentences with redundant infor",
+        "bbox": [
+            507,
+            889,
+            882,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "17930",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "mation, we propose a Sentence Distillation Prompt comprising three carefully crafted instructions. For a given input sentence  $x_{u}$  with two target entities  $e^{1}, e^{2}$ , the Relation Summarization instruction requires LLMs to encapsulate the semantic relation between the annotated entities within one sentence, denoted as  $x_{u}^{*}$ . This requires LLMs to utilize their summarization abilities. The Type Identification instruction aims to infer the types of the head and tail entities that are crucial for RE (Lyu and Chen, 2021), represented as  $T_{u}^{1}$  and  $T_{u}^{2}$ , respectively. The final triplet output is organized into a JSON object using the Triplet Combination instruction, described as  $\\{x_{u}^{*}, T_{u}^{1}, T_{u}^{2}\\}$ .",
+        "bbox": [
+            112,
+            84,
+            489,
+            310
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Relation Enrichment Prompt. To leverage the general knowledge of LLMs for mining deep information from relation labels and descriptions, we introduce a Relation Enrichment Prompt. It aims to instruct LLMs to freely generate all possible entity types  $\\hat{T_u}$  associated with relation labels. This automated process effectively alleviates the manual labor involved in annotating head entity types and tail entity types for relations. Subsequently, we organize them into triplets in JSON format:  $\\{r_u,\\{T_{u,i}^1\\}_{i = 1}^n,\\{T_{u,j}^2\\}_{j = 1}^m\\}$ .",
+        "bbox": [
+            112,
+            311,
+            489,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Embedding with LLMs. To match sentence triplets and relation triplets, we embed them into the same embedding space. We obtain an embedding  $h \\in \\mathbb{R}^d$  for each element of the sentence triplets and relation triplets by calling the open API of the latest GPT series model \"text-embedding-3-small\", which is trained via matryoshka representation learning (Kusupati et al., 2022). It is noteworthy that for  $\\hat{T}_u$ , we conduct average pooling for each set to obtain  $h_{\\hat{T}_u}$  of entity types. Each sentence embedding triplet and relation embedding triplet can be represented respectively as  $\\mathrm{Tri}_x = \\{h_{x_u^*}, h_{T_u^1}, h_{T_u^2}\\}$  and  $\\mathrm{Tri}_r = \\{h_{r_u}, h_{\\hat{T}_u^1}, h_{\\hat{T}_u^2}\\}$ .",
+        "bbox": [
+            112,
+            495,
+            489,
+            709
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Retrieval Adjuster",
+        "text_level": 1,
+        "bbox": [
+            112,
+            720,
+            307,
+            736
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Embeddings from LLMs have a potential anisotropy problem due to common-word biases (Li and Li, 2024), causing the embeddings to occupy a narrow cone. Therefore, contrastive learning-based fine-tuning is required to distribute the embeddings uniformly (Gao et al., 2021).",
+        "bbox": [
+            112,
+            743,
+            487,
+            839
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "However, fine-tuning LLMs for specific tasks is resource-intensive and necessitates full access to the parameters of the LLMs, posing challenges with API-based LLMs. Therefore, we propose an independent retrieval adjuster (AD) for adapting",
+        "bbox": [
+            112,
+            841,
+            485,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ZSRE, aimed at learning the ability to align sentence and relation label representations from seen data and further adjust the unseen embeddings. AD does not require access to LLMs' parameters and can easily plug into LLMs.",
+        "bbox": [
+            507,
+            84,
+            884,
+            165
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To adapt to the task while retaining the LLM-based embedding semantics, AD integrates residual features learned by network  $\\mathrm{F}(\\cdot)$  with the original embeddings through skip connections to produce the final embedding  $\\tilde{h}$ :",
+        "bbox": [
+            507,
+            165,
+            882,
+            244
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\tilde {h} = h + \\mathrm {F} (h). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            636,
+            252,
+            880,
+            269
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We employ a straightforwardly connected network as  $\\mathrm{F}(\\cdot)$ , with detailed operations as follows:",
+        "bbox": [
+            507,
+            279,
+            882,
+            312
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {F} (h) = W _ {2} \\left(\\operatorname {R e L U} \\left(W _ {1} (h) + b _ {1}\\right)\\right) + b _ {2}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            536,
+            319,
+            880,
+            337
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $W_{1}\\in \\mathbb{R}^{d\\times n}$ ,  $W_{2}\\in \\mathbb{R}^{n\\times d}$  are parameter matrices, and  $b_{1},b_{2}$  are bias. Notably, inspired by Yoon et al. (2023), we employ a shared  $\\mathrm{F}(\\cdot)$  for all embeddings, thereby maintaining a unified embedding space to achieve better retrieval.",
+        "bbox": [
+            507,
+            344,
+            882,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.5 Triplet-based Retrieval",
+        "text_level": 1,
+        "bbox": [
+            507,
+            436,
+            737,
+            451
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To enable  $\\mathrm{Tri}_x$  to retrieve the corresponding  $\\mathrm{Tri}_r$ , we introduce triplet score  $c$  to compute the similarity coefficient:",
+        "bbox": [
+            507,
+            457,
+            882,
+            505
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc \\left(\\operatorname {T r i} _ {x}, \\operatorname {T r i} _ {r}\\right) = \\alpha \\left\\langle h _ {x ^ {*}}, h _ {r} \\right\\rangle + \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            514,
+            880,
+            537
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left(1 - \\alpha\\right) \\left(\\langle h _ {T ^ {1}}, h _ {\\hat {T} ^ {1}} \\rangle + \\langle h _ {T ^ {2}}, h _ {\\hat {T} ^ {2}} \\rangle\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            533,
+            831,
+            551
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha$  is hyperparameter, and  $\\langle ,\\rangle$  denotes the cosine similarity. We train AD on seen data  $D_{s}$  to align the embeddings of  $\\mathrm{Tri}_{x_s}$  and  $\\mathrm{Tri}_{r_s}$ . The training objective of  $\\mathrm{F}(\\cdot)$  is to maximize the triplet score between  $\\tilde{\\mathrm{Tri}}_{x_s}^i$  and positive  $\\tilde{\\mathrm{Tri}}_{r_s}^i$ , while minimizing its score with negative  $\\tilde{\\mathrm{Tri}}_{r_s}^j$ :",
+        "bbox": [
+            507,
+            558,
+            882,
+            663
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\max  \\left\\{0, \\max  \\left(c \\left(\\tilde {\\operatorname {T r i}} _ {x _ {s}} ^ {i}, \\tilde {\\operatorname {T r i}} _ {r _ {s}} ^ {j}\\right)\\right) \\right. \\tag {4} \\\\ - c \\left(\\tilde {\\operatorname {T r i}} _ {x _ {s}} ^ {i}, \\tilde {\\operatorname {T r i}} _ {r _ {s}} ^ {i}\\right) + \\Delta \\}, i \\neq j, \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            534,
+            671,
+            880,
+            738
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $N$  denotes the batch size, and  $\\Delta$  represents a predefined threshold. Negative samples are randomly selected within the same batch, maintaining a 1:3 ratio of positive to negative samples. This approach can promote the uniform distribution of sentence embeddings and relation embeddings, thereby alleviating the anisotropy problem.",
+        "bbox": [
+            507,
+            744,
+            882,
+            856
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The trained AD is utilized to adjust the embeddings of unseen  $\\mathrm{Tri}_{x_u}$  and  $\\mathrm{Tri}_{r_u}$  during the testing phase. The highest computed  $c(\\tilde{\\mathrm{Tri}}_{x_u},\\tilde{\\mathrm{Tri}}_{r_u})$  is chosen as the final relation.",
+        "bbox": [
+            507,
+            857,
+            882,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "17931",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            312,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            110,
+            226,
+            124
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments on five benchmarks. Following established setups (Zhao et al., 2023a), we conduct our primary experiments on two widely used datasets: Wiki-ZSL (Chen and Li, 2021) and FewRel (Han et al., 2018). Wiki-ZSL contains 94,383 instances and 113 relations, and is acquired through distant supervision from Wiki-KB, characterized by imbalanced samples and heightened noise. FewRel contains 56,000 instances and 80 relations, and is meticulously annotated, ensuring equal samples per relation. Consistent with prior works, we divide the datasets into a training set containing seen relations and a validation set containing unseen relations, focusing on evaluating under unseen relations  $m = 5, 10, 15$ . To ensure the reliability of results, we perform 10-fold cross-validation and report the average results. Notably, the entity types are not provided in Wiki-ZSL and FewRel, posing challenges to ZSRE as candidate relations cannot be filtered based on entity types.",
+        "bbox": [
+            112,
+            131,
+            489,
+            453
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Moreover, following the setups of other LLM-based methods (Li et al., 2023), we also select three datasets where entity types are given: (1) TACRED (Zhang et al., 2017) contains 42 relations and 17 entity types, (2) TACREV (Alt et al., 2020) contains 42 relations and 17 entity types, and (3) Re-TACRED (Stoica et al., 2021) contains 40 relations and 17 entity types. We further evaluate on their test sets with the none-of-the-above (NoTA) relation excluded.",
+        "bbox": [
+            112,
+            454,
+            489,
+            613
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            626,
+            344,
+            640
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We set temperature  $= 0$  for GPT series models to improve the reproducibility. We apply the same prompt for Wiki-ZSL and FewRel. Since the entity types are seen in TACRED, TACREV, and Re-TACRED, we remove instructions for generating entity types from prompt. Following previous work, we filter a subset of candidate relations based on the entity types. For retrieval adjuster AD, we set the same hyperparameters for all experiments. We use AdamW with a learning rate of  $1e - 5$ , 10 epochs, and a batch size of 64. We set the dimension of the hidden layer  $n$  to  $d / 2$ , resulting in approximately 2.36M parameters for AD, which is much smaller than BERT-based SLMs. We conducted a grid search on FewRel and determined hyperparameters  $\\alpha = 0.75, \\Delta = 0.1$ , negative ratio  $= 1:3$  when the model performed the best.",
+        "bbox": [
+            112,
+            646,
+            489,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baseline Models",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            684,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "SLMs-based methods. We compare: embedding-based methods R-BERT (Wu and He, 2019), ZSBERT (Chen and Li, 2021), and RE-Matching (Zhao et al., 2023a); text entailment-based methods ESIM (Chen et al., 2017), LaVeEntail (Sainz et al., 2021); and generation-based methods RE-Prompt (Chia et al., 2022), SuRE (Lu et al., 2022). LLM-based methods. We compare previous competitive baselines, including VanillaPrompt, QA4RE, SUMASK, and MICRE. VanillaPrompt and QA4RE are from Zhang et al. (2023), respectively guide LLMs to generate relation labels or multiple-choice options, where QA4RE involves manually building relation templates. As they did not evaluate on Wiki-ZSL and FewRel, we rerun the results based on the original paper and available code. SUMASK (Li et al., 2023) classifies relations by generating multiple responses for all candidate relations and estimating uncertainty from LLMs. MICRE (Li et al., 2024a) uses 12 public datasets to fine-tune LLaMA (Touvron et al., 2023a) through tabular prompting, achieving performance competitive with SLMs. Following the previous approach, we randomly sample 1,000 examples as test data from each dataset to control the cost of comparison with the LLM-based baselines. Since using seen data for training the AD incurs additional costs and may not be available in some datasets, we provide two sets of results indicating whether seen data are used to train AD, denoted as RE-GAR and RE-GAR-AD, respectively.",
+        "bbox": [
+            507,
+            105,
+            884,
+            605
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Analysis and Discussion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            615,
+            751,
+            633
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Main Result",
+        "text_level": 1,
+        "bbox": [
+            509,
+            642,
+            653,
+            657
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We report the main results on Wiki-ZSL and FewRel in Table 1. Overall, our approach demonstrates competitive performance across all settings on both datasets.",
+        "bbox": [
+            507,
+            663,
+            882,
+            726
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "(1) Compared to LLM-based methods, our approach ensures strong performance even without considering seen data (RE-GAR's average F1 is 6.93 higher than QA4RE across six settings). We attribute this to the novel GAR framework, which effectively utilizes both the generative and representational capabilities of LLMs, allowing LLMs to unleash more potential on ZSRE tasks. Additionally, we observe that when considering seen data, RE-GAR-AD achieves even stronger performance with fewer training samples (with an F1 advantage of 3.71 over MICRE).",
+        "bbox": [
+            507,
+            728,
+            884,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "17932",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/8efe8386e61c4939310c544c1bdafe0d91ab26a5049635ef6e56052b63b5fb06.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"11\">Wiki-ZSL</td><td colspan=\"7\">FewRel</td></tr><tr><td colspan=\"3\">m=5</td><td colspan=\"3\">m=10</td><td colspan=\"3\">m=15</td><td colspan=\"3\">m=5</td><td colspan=\"3\">m=10</td><td colspan=\"3\">m=15</td></tr><tr><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td></tr><tr><td colspan=\"19\">- SLMs-based</td></tr><tr><td>R-BERT*</td><td>39.22</td><td>43.27</td><td>41.15</td><td>26.18</td><td>29.69</td><td>27.82</td><td>17.31</td><td>18.82</td><td>18.03</td><td>42.19</td><td>48.61</td><td>45.17</td><td>25.52</td><td>33.02</td><td>25.52</td><td>16.95</td><td>19.37</td><td>18.08</td></tr><tr><td>ESIM*</td><td>48.58</td><td>47.74</td><td>48.16</td><td>44.12</td><td>45.46</td><td>44.78</td><td>27.31</td><td>29.62</td><td>28.42</td><td>56.27</td><td>58.44</td><td>57.33</td><td>42.89</td><td>44.17</td><td>43.52</td><td>29.15</td><td>31.59</td><td>30.32</td></tr><tr><td>ZS-BERT*</td><td>71.54</td><td>72.39</td><td>71.96</td><td>60.51</td><td>60.98</td><td>60.74</td><td>34.12</td><td>34.38</td><td>34.25</td><td>76.96</td><td>78.86</td><td>77.90</td><td>56.92</td><td>57.59</td><td>57.25</td><td>35.54</td><td>38.19</td><td>36.82</td></tr><tr><td>LaVeEntail*</td><td>77.39</td><td>75.90</td><td>76.63</td><td>71.86</td><td>71.14</td><td>71.50</td><td>62.13</td><td>61.76</td><td>61.95</td><td>91.14</td><td>90.86</td><td>91.00</td><td>83.05</td><td>82.55</td><td>82.80</td><td>72.83</td><td>72.10</td><td>72.46</td></tr><tr><td>REPrompt*</td><td>70.66</td><td>83.75</td><td>76.63</td><td>68.51</td><td>74.76</td><td>71.50</td><td>63.69</td><td>67.93</td><td>65.74</td><td>90.15</td><td>88.50</td><td>89.30</td><td>80.33</td><td>79.62</td><td>79.96</td><td>74.33</td><td>72.51</td><td>73.40</td></tr><tr><td>RE-Matching*</td><td>78.19</td><td>78.41</td><td>78.30</td><td>74.39</td><td>73.54</td><td>73.96</td><td>67.31</td><td>67.33</td><td>67.32</td><td>92.82</td><td>92.34</td><td>92.58</td><td>83.21</td><td>82.64</td><td>82.93</td><td>73.80</td><td>73.52</td><td>73.66</td></tr><tr><td colspan=\"19\">- LLM-based</td></tr><tr><td>VanillaPrompt</td><td>79.83</td><td>56.56</td><td>66.21</td><td>62.98</td><td>58.05</td><td>60.41</td><td>58.54</td><td>51.26</td><td>54.66</td><td>74.87</td><td>51.15</td><td>60.77</td><td>65.28</td><td>46.46</td><td>54.29</td><td>59.70</td><td>50.56</td><td>54.76</td></tr><tr><td>QA4RE</td><td>87.20</td><td>74.97</td><td>80.62</td><td>65.80</td><td>58.25</td><td>61.80</td><td>62.17</td><td>57.80</td><td>59.90</td><td>88.00</td><td>75.66</td><td>81.37</td><td>79.00</td><td>73.26</td><td>76.02</td><td>65.33</td><td>67.45</td><td>66.37</td></tr><tr><td>SUMASK†</td><td>75.64</td><td>70.96</td><td>73.23</td><td>62.31</td><td>61.08</td><td>61.69</td><td>43.55</td><td>40.27</td><td>41.85</td><td>78.27</td><td>72.55</td><td>75.30</td><td>64.77</td><td>60.94</td><td>62.80</td><td>44.76</td><td>41.13</td><td>42.87</td></tr><tr><td>MICRE†</td><td>76.46</td><td>78.53</td><td>77.48</td><td>72.36</td><td>74.88</td><td>73.60</td><td>67.14</td><td>68.87</td><td>67.99</td><td>89.34</td><td>91.88</td><td>90.59</td><td>80.67</td><td>82.31</td><td>81.48</td><td>73.74</td><td>75.83</td><td>74.77</td></tr><tr><td colspan=\"19\">- Ours (LLM-based)</td></tr><tr><td>RE-GAR</td><td>85.13</td><td>83.85</td><td>84.44</td><td>74.98</td><td>73.32</td><td>74.12</td><td>68.45</td><td>66.35</td><td>67.37</td><td>90.75</td><td>89.40</td><td>90.07</td><td>80.89</td><td>78.76</td><td>79.81</td><td>73.48</td><td>70.26</td><td>71.82</td></tr><tr><td>RE-GAR-AD</td><td>88.96</td><td>88.42</td><td>88.63</td><td>77.63</td><td>77.85</td><td>77.70</td><td>70.80</td><td>69.26</td><td>70.02</td><td>93.03</td><td>92.16</td><td>92.59</td><td>83.48</td><td>82.10</td><td>82.78</td><td>77.69</td><td>75.31</td><td>76.47</td></tr></table>",
+        "bbox": [
+            126,
+            80,
+            878,
+            329
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "(2) Compared to SLMs-based methods trained on seen data, RE-GAR-AD achieves up to 5.59 higher F1 on Wiki-ZSL than the existing SOTA model, RE-Matching. Interestingly, our method outperforms other baselines by a large margin at  $m = 5$  on Wiki-ZSL, and we attribute this improvement partly to the effective noise reduction in the distant supervision dataset through Sentence Distillation. It is worth noting that the parameter of the fine-tuned retrieval adjuster is much smaller than BERT, which is adopted as the backbone for most SLM-based methods.",
+        "bbox": [
+            112,
+            423,
+            487,
+            615
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "(3) Additionally, in Table 2, we report the micro-averaged F1 on TACRED, TACREV, and Re-TACRED. Since entity types are provided, we follow the previous works (Zhang et al., 2023; Li et al., 2023) by discarding irrelevant relations. The results indicate that RE-GAR achieves comparable performance on three datasets.",
+        "bbox": [
+            112,
+            617,
+            487,
+            728
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Ablation Experiment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            739,
+            326,
+            755
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We report the impact of different components on RE-GAR-AD, as shown in Table 3.",
+        "bbox": [
+            112,
+            760,
+            485,
+            790
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "1. We use Embedding LLMs to encode the raw sentences when we remove the Sentence Distillation. The average performance drops by 7.74. This indicates that the summarization capability of LLMs can effectively reduce irrelevant information in sentences, allowing them to focus more on the semantic association between entities, thus gaining an advantage in relation matching.",
+        "bbox": [
+            112,
+            791,
+            487,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2. We manually annotate possible entity types for relations when we remove the Relation Enrichment. We observe a slight decrease in average performance by 1.67, indicating that LLMs have the potential to generate data of comparable quality to manually annotated data, and that richer information may yield greater benefits.",
+        "bbox": [
+            507,
+            423,
+            884,
+            536
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3. We solely match sentences and relation labels",
+        "bbox": [
+            527,
+            537,
+            880,
+            552
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/6f614037a62fd1fe945ac24da70615dd33b5d5c3054d8219ff5e4ff5414c40ad.jpg",
+        "table_caption": [
+            "Table 1: Experimental results on Wiki-ZSL and FewRel. We mark the best results in bold and the second-best underlined. All generation based on API are run on GPT-3.5-turbo. RE-GAR and RE-GAR-AD respectively represent the results without using seen data, and the results of fine-tuning our retrieval adapter with only seen data. The results indicated by  $\\dagger$  and  $*$  are reported in MICRE (Li et al., 2024a) and RE-Matching (Zhao et al., 2023a). P., R., and F1 denote Precision, Recall, and F1 score, respectively."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>TACRED</td><td>TACREV</td><td>Re-TACRED</td></tr><tr><td>LaVeEntail_BART‡</td><td>51.4</td><td>55.3</td><td>44.0</td></tr><tr><td>LaVeEntail_ DeBERTa‡</td><td>55.1</td><td>57.2</td><td>64.3</td></tr><tr><td>SuRE_BART‡</td><td>20.4</td><td>22.2</td><td>23.6</td></tr><tr><td>SuRE_PEGASUS‡</td><td>21.8</td><td>21.6</td><td>22.4</td></tr><tr><td>VanillaPrompt‡</td><td>48.1</td><td>51.0</td><td>55.3</td></tr><tr><td>QA4RE‡</td><td>59.4</td><td>59.4</td><td>61.2</td></tr><tr><td>SUMASK†</td><td>57.7</td><td>-</td><td>-</td></tr><tr><td>RE-GAR</td><td>64.8</td><td>63.4</td><td>53.8</td></tr></table>",
+        "bbox": [
+            519,
+            576,
+            878,
+            722
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/e889ee3a3360129817d21c3ebcad4a2807553c69b7f1fa47ad4d3756ba444576.jpg",
+        "table_caption": [
+            "Table 2: Experimental results on TACRED, TACREV, and Re-TACRED. The results indicated by  $\\dagger$  and  $\\ddagger$  are reported in SUMASK (Li et al., 2023) and QA4RE (Zhang et al., 2023)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>m=5</td><td>m=10</td><td>m=15</td></tr><tr><td>RE-GAR-AD</td><td>92.59</td><td>82.78</td><td>76.47</td></tr><tr><td>w/o. Sentence Distillation</td><td>88.45</td><td>73.06</td><td>67.12</td></tr><tr><td>w/o. Relation Enrichment</td><td>92.06</td><td>80.13</td><td>74.64</td></tr><tr><td>w/o. Triplet-based Retrieval</td><td>82.18</td><td>68.44</td><td>60.32</td></tr><tr><td>w/o. Retrieval Adjuster</td><td>90.07</td><td>79.81</td><td>71.82</td></tr></table>",
+        "bbox": [
+            517,
+            781,
+            878,
+            883
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Ablation study of RE-GAR-AD. We report the F1 performance on FewRel under  $m = 5,10,15$",
+        "bbox": [
+            507,
+            891,
+            880,
+            916
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "17933",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/5a55b687451ed87757f7afa80f1a91694fb7b0cebc91c3dcc827bdfcea3229fb.jpg",
+        "image_caption": [
+            "Figure 3: Effect of data scale, data quality, and residual connections on retrieval adjuster AD."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            82,
+            473,
+            206
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d9ead2937ebf408a3305a3f47400452dd7b0bcad783309f35977344745acbf2b.jpg",
+        "image_caption": [
+            "Figure 4: Comparison of API overhead (Bar) and F1 score (Line) for 8,000 instances. To control the cost, we randomly sample 100 samples for each relation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            252,
+            478,
+            363
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "when we remove the Triplet-based Retrieval (i.e., without considering entity type information in the retrieval), there is a dramatic decline in average performance (-13.63), which proves the crucial role of entity information in relation retrieval.",
+        "bbox": [
+            112,
+            434,
+            487,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4. When Retrieval Adjuster AD is removed, it means that we do not benefit from fine-tuning, leading to an average performance drop of 3.42.",
+        "bbox": [
+            112,
+            514,
+            489,
+            564
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Analysis on Retrieval Adjuster",
+        "text_level": 1,
+        "bbox": [
+            112,
+            576,
+            401,
+            592
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Retrieval adjuster AD needs to be trained on seen data, which may introduce additional generation costs and training overhead. Therefore, we further investigate the impact of training data of varying scales and qualities on the performance of AD under the setting of  $m = 15$  on the FewRel dataset. Specifically, we randomly extract  $\\{1\\%, 10\\%, 50\\% \\}$  of the samples from the seen data to construct subsets, maintaining an even balance of samples per relation.",
+        "bbox": [
+            112,
+            598,
+            487,
+            758
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Data scale. As shown in Figure 3, more training samples can bring greater benefits to AD, which is consistent with intuition. Notably, only  $10\\%$  of the distilled seen data can lead to a 2.48 F1 score increase, while increasing the data scale tenfold only further improves the F1 score by 2.17.",
+        "bbox": [
+            112,
+            760,
+            487,
+            854
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Data quality. Even if the data is not distilled (raw data), AD can still enhance performance without negatively impacting the model. Since this strategy does not require LLMs to generate, it can",
+        "bbox": [
+            112,
+            856,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "be more cost-effective in some scenarios.",
+        "bbox": [
+            507,
+            84,
+            815,
+            98
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Residual connection. We observe that removing residual connection when the training data scale is small leads to a performance decrease, indicating that residual connection can effectively improve information flow and reduce overfitting. However, with sufficient data scale, the additional benefits of residual connection may not be as apparent.",
+        "bbox": [
+            507,
+            101,
+            884,
+            212
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Model Efficiency",
+        "text_level": 1,
+        "bbox": [
+            507,
+            227,
+            687,
+            242
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To evaluate our approach from both performance and cost perspectives, we compare it with LLM-based baselines on FewRel with settings of  $m = 5,10,15,20,30,40$  and report the results in Figure 4. For a fair comparison, we only report the performance of RE-GAR, which does not benefit from fine-tuning on seen data. Moreover, as described in Section 1, SUMASK requires an input overhead of  $n\\times k\\times r$  tokens. Therefore, we choose QA4RE as the baseline, which has a more efficient input overhead  $n\\times r$ . The results are derived from GPT-3.5-turbo and text-embedding-3-small. The estimation of API costs is sourced from OpenAI",
+        "bbox": [
+            507,
+            248,
+            884,
+            457
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The results indicate that RE-GAR consistently demonstrates advantages in both F1 score and cost compared to QA4RE across all settings of  $m$ . Moreover, as  $m$  increases, RE-GAR's performance is only mildly affected. Additionally, since relation labels serving as retrieval targets are only pre-embedded once and fixed, the average cost of RE-GAR in large-scale ZSRE scenarios only undergoes slight changes as  $m$  increases.",
+        "bbox": [
+            507,
+            458,
+            885,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.5 Scalability Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            615,
+            710,
+            631
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct scalability research on SLMs, open-source LLMs, and API-based LLMs, and report F1 score on FewRel with  $m = 15$ . We select 100 samples for each relation to build subsets to control costs. Note that the generation-augmentation and retrieval are executed in a pipelined manner, and their combination is not fixed.",
+        "bbox": [
+            507,
+            637,
+            882,
+            749
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In Figure 5, we compare different embedding models. For SLMs, we choose SROBERTa (Reimers and Gurevych, 2019), PromptROBERTa (Jiang et al., 2022), SimCSE-Roberta (Gao et al., 2021), and BGE-en (Xiao et al., 2023). For open-source LLMs, we introduce LLM2Vec (BehnamGhader et al., 2024), which converts LLaMA2 (Touvron et al., 2023b) and Mistral (Jiang et al., 2023) into general text encoders. In addition,",
+        "bbox": [
+            507,
+            751,
+            885,
+            895
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://openai.com/api/pricing/",
+        "bbox": [
+            529,
+            904,
+            727,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "17934",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4117becbccd735b45b70a8a16096d0b035647d6eb82dd807eb42becf6f054744.jpg",
+        "image_caption": [
+            "Figure 5: Scalability analysis: Generate with GPT-3.5-turbo. Embedding with different SLMs and LLMs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            489,
+            197
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/9f0a43d085b2c3e226ea233ad9cbbaf4f7ad1223fce239dfbcfe8a23bbd26933.jpg",
+        "image_caption": [
+            "Figure 6: Scalability analysis: Embedding with text-embedding-3-small. Generate with different LLMs.",
+            "Figure 7: Case study on FewRel with  $m = 15$ . Sentence, Head Entity, and Tail Entity are derived from the raw data. Distilled Sentence represents the data distilled through the RE-GAR-AD framework."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            132,
+            242,
+            467,
+            328
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "we explore different API-based LLMs (OpenAI text-embedding (Neelakantan et al., 2022)). We observe that the performance of SLMs generally does not exceed 61, which may indicate a significant gap in representation capabilities between SLMs and LLMs on ZSRE, revealing the potential of LLMs in representation. RE-GAR-AD demonstrates excellent performance on all LLMs, proving its robustness.",
+        "bbox": [
+            112,
+            387,
+            487,
+            530
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In Figure 6, we compare different generative models, including LLaMA3, Mistral, GPT-3.5-turbo, and GPT-4-turbo, which achieve remarkable performance with their strong generation capabilities. We observe that more powerful LLMs such as text-embedding-3-large and GPT-4-turbo can provide sustained benefits in RE-GAR-AD.",
+        "bbox": [
+            112,
+            533,
+            489,
+            645
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we notice that our proposed plugin in retrieval adjuster consistently shows improvement, indicating that it is model-agnostic.",
+        "bbox": [
+            112,
+            646,
+            489,
+            694
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.6 Error Analysis and Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            707,
+            403,
+            722
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To investigate the effectiveness of our RE-GAR framework, we provide a typical example as shown in Case 1 of Figure 7. For QA4RE, we construct relation-entailed templates as label verbalization: \"P27: country of citizenship\" is processed into \"[Andrew Goldberg] is a citizen of [United States]\", and \"P361: part of\" is processed into \"[Andrew Goldberg] is part of [United States]\". Due to the semantic similarity between the two options, generative LLMs might confuse them and output the incorrect option. RE-GAR refines the semantic relationship between Andrew Goldberg and the United",
+        "bbox": [
+            112,
+            728,
+            489,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/62a9206f2d6747d86ff4454ccf85d502a934030aff2fe439110bdbcaa2663a79.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Sentence</td><td colspan=\"2\">After leaving Island Records due to a disagreement over the band&#x27;s artistic direction, Fightstar signed with the independent label Institute Records for their second album.</td></tr><tr><td colspan=\"2\">Head Entity: Fightstar</td><td>Tail Entity: Island Records</td></tr><tr><td>Distilled Sentence</td><td colspan=\"2\">Fightstar left Island Records due to a disagreement over artistic direction. Band. Record Label.</td></tr><tr><td>Enriched relation (P175)</td><td>Actor, Musician, Band</td><td>Musical work</td></tr><tr><td>Enriched relation (P264)</td><td>Music recordings, Music videos</td><td>Record label, Brand, Trademark</td></tr><tr><td>Ground Truth</td><td>P264: record label</td><td>✔</td></tr><tr><td>QA4RE</td><td>P175: performer</td><td>✘</td></tr><tr><td>RE-GAR</td><td>P175: performer</td><td>✘</td></tr><tr><td>RE-GAR-AD</td><td>P264: record label</td><td>✔</td></tr></table>",
+        "bbox": [
+            521,
+            82,
+            870,
+            254
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "States, yielding the accurate information: \"Andrew Goldberg is a writer from the United States.\" Additionally, it pays extra attention to the type information of the two entities (Person and Country).",
+        "bbox": [
+            507,
+            338,
+            882,
+            401
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Case 2 of Figure 7, when LLMs fail to capture meaningful semantic connections between Fightstar and Island records, it can lead to ineffective retrieval. Our proposed plug-in retrieval adjuster addresses this by learning relation patterns from seen data and subsequently adjusting when retrieving unseen relation labels, successfully categorizing the relation as \"record label\".",
+        "bbox": [
+            507,
+            401,
+            882,
+            530
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we report the performance of our framework on 80 relation categories in Appendix C. We observe that although RE-GAR-AD can distinguish similar relations better with the retrieval adjuster, challenges still exist when co-occurring relations are present in sentences. This will be one of our future optimization directions.",
+        "bbox": [
+            507,
+            531,
+            882,
+            643
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            655,
+            640,
+            669
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We present RE-GAR-AD, a simple and effective ZSRE framework based on LLMs. We introduce a novel LLM-based ZSRE paradigm of Generation-Augmented Retrieval, leveraging LLMs' generative and representational capabilities. Additionally, we train a plug-in retrieval adjuster to further optimize embeddings. Extensively evaluated on common ZSRE benchmarks, RE-GAR-AD demonstrates comparable performance across various settings, especially in cases where entity types are excluded. Furthermore, our method significantly improves the efficiency of LLMs in multi-choice relation extraction, which is advantageous for real-world ZSRE and opens up interesting avenues for future zero-shot classification efforts.",
+        "bbox": [
+            507,
+            678,
+            884,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "17935",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Although the proposed framework demonstrates advanced performance and cost-effectiveness on large-scale ZSRE, its cost-effectiveness is not evident when the number of input sentences  $n$  is small. Additionally, due to API limitations, we are unable to obtain complete results from more powerful models such as GPT-4 and others. Furthermore, we employ only a straightforward structure for the retrieval adjuster, and we believe that further research on LLM-embedding-based adjusters would be an interesting direction. Finally, since our method encounters difficulties in handling the none-of-the-above (NoTA) relation, it may not be suitable for cases where the dataset contains the NoTA relation.",
+        "bbox": [
+            112,
+            111,
+            492,
+            336
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            351,
+            278,
+            368
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The authors sincerely thank the reviewers for their valuable comments, which improved the paper. The work is supported by the National Natural Science Foundation of China (62276057).",
+        "bbox": [
+            112,
+            378,
+            489,
+            443
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            472,
+            213,
+            488
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Christoph Alt, Aleksandra Gabryszak, and Leonhard Hennig. 2020. Tacred revisited: A thorough evaluation of the tacred relation extraction task. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1558-1569.",
+            "Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961.",
+            "Chih Yao Chen and Cheng-Te Li. 2021. Zs-bert: Towards zero-shot relation extraction with attribute representation learning. In NAACL 2021, pages 3470-3479.",
+            "Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced LSTM for natural language inference. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1657-1668.",
+            "Yew Ken Chia, Lidong Bing, Soujanya Poria, and Luo Si. 2022. Relationprompt: Leveraging prompts to generate synthetic data for zero-shot relation triplet extraction. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 45-57.",
+            "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP 2021, pages 6894-6910."
+        ],
+        "bbox": [
+            115,
+            495,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jiaying Gong and Hoda Eldardiry. 2021. Zero-shot relation classification from side information. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 576-585.",
+            "Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. Fewrel: A large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. In EMNLP 2018, pages 4803-4809.",
+            "Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, and 1 others. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.",
+            "Ting Jiang, Jian Jiao, Shaohan Huang, Zihan Zhang, Deqing Wang, Fuzhen Zhuang, Furu Wei, Haizhen Huang, Denvy Deng, and Qi Zhang. 2022. Promptbert: Improving bert sentence embeddings with prompts. In EMNLP 2022, pages 8826-8837.",
+            "Bosung Kim, Hayate Iso, Nikita Bhatani, Estevam Hruschka, Ndapandula Nakashole, and Tom Mitchell. 2023. Zero-shot triplet extraction by template infilling. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 272-284.",
+            "Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, and 1 others. 2022. Matryoshka representation learning. Advances in Neural Information Processing Systems, 35:30233-30249.",
+            "Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. 2017. Zero-shot relation extraction via reading comprehension. In CoNLL 2017, pages 333-342.",
+            "Guozheng Li, Peng Wang, and Wenjun Ke. 2023. Revisiting large language models as zero-shot relation extractors. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 6877-6892.",
+            "Guozheng Li, Peng Wang, Jiajun Liu, Yikai Guo, Ke Ji, Ziyu Shang, and Zijie Xu. 2024a. Meta in-context learning makes large language models better zero and few-shot relation extractors. arXiv preprint arXiv:2404.17807.",
+            "Haitao Li, Qingyao Ai, Jia Chen, Qian Dong, Zhijing Wu, Yiqun Liu, Chong Chen, and Qi Tian. 2024b. Blade: Enhancing black-box large language models with small domain-specific models. arXiv preprint arXiv:2403.18365.",
+            "Xianming Li and Jing Li. 2024. Bellm: Backward dependency enhanced large language model for sentence embeddings. In *NAACL* 2024, pages 792-804."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "17936",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zehan Li, Fu Zhang, and Jingwei Cheng. 2024c. Alignre: An encoding and semantic alignment approach for zero-shot relation extraction. In *Findings of the Association for Computational Linguistics ACL* 2024, pages 2957-2966.",
+            "Keming Lu, I-Hung Hsu, Wenxuan Zhou, Mingyu Derek Ma, and Muhao Chen. 2022. Summarization as indirect supervision for relation extraction. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6575-6594.",
+            "Shengfei Lyu and Huanhuan Chen. 2021. Relation classification with entity type restriction. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 390-395.",
+            "Xilai Ma, Jing Li, and Min Zhang. 2023a. Chain of thought with explicit evidence reasoning for few-shot relation extraction. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 2334-2352.",
+            "Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2023b. Fine-tuning llama for multi-stage text retrieval. arXiv preprint arXiv:2310.08319.",
+            "Yubo Ma, Yixin Cao, Yong Hong, and Aixin Sun. 2023c. Large language model is not a good few-shot information extractor, but a good reranker for hard samples! In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 10572-10601.",
+            "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In EMNLP 2022, pages 11048-11064.",
+            "Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2024. Generative representational instruction tuning. arXiv preprint arXiv:2402.09906.",
+            "Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, and 1 others. 2022. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005.",
+            "Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In EMNLP-IJCNLP 2019, pages 3982-3992.",
+            "Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2022. Learning to retrieve prompts for in-context learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2655-2671.",
+            "Oscar Sainz, Oier Lopez de Lacalle, Gorka Labaka, Ander Barrena, and Eneko Agirre. 2021. Label verbalization and entailment for effective zero and few-shot"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "relation extraction. In EMNLP 2021, pages 1199-1212.",
+            "Oscar Sainz, Iker García-Ferrero, Rodrigo Agerri, Oier Lopez de Lacalle, German Rigau, and Eneko Agirre. 2024. Gollie: Annotation guidelines improve zero-shot information-extraction. In *ICLR* 2024.",
+            "Jacob Mitchell Springer, Suhas Kotha, Daniel Fried, Graham Neubig, and Aditi Raghunathan. 2024. Repetition improves language model embeddings. arXiv preprint arXiv:2402.15449.",
+            "George Stoica, Emmanouil Antonios Platanios, and Barnabás Póczos. 2021. Re-tacred: Addressing shortcomings of the tacred dataset. In Proceedings of the AAAI conference on artificial intelligence, pages 13843-13850.",
+            "Qi Sun, Kun Huang, Xiaocui Yang, Rong Tong, Kun Zhang, and Soujanya Poria. 2024. Consistency guided knowledge retrieval and denoising in llms for zero-shot document-level relation triplet extraction. In Proceedings of the ACM on Web Conference 2024, pages 4407-4416.",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023a. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.",
+            "Zhen Wan, Fei Cheng, Zhuoyuan Mao, Qianying Liu, Haiyue Song, Jiwei Li, and Sadao Kurohashi. 2023. Gpt-re: In-context learning for relation extraction using large language models. In EMNLP 2023, pages 3534-3547.",
+            "Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368.",
+            "Shanchan Wu and Yifan He. 2019. Enriching pretrained language model with entity information for relation classification. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 2361-2364.",
+            "Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighof. 2023. C-pack: Packaged resources to advance general chinese embedding. arXiv preprint arXiv:2309.07597.",
+            "Kaijia Yang, Nantao Zheng, Xinyu Dai, Liang He, Shujian Huang, and Jiajun Chen. 2020. Enhance prototypical network with text descriptions for few-shot relation classification. In Proceedings of the 29th ACM"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "17937",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "International Conference on Information & Knowledge Management, pages 2273-2276.",
+        "bbox": [
+            132,
+            85,
+            487,
+            112
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jinsung Yoon, Sercan O Arik, Yanfei Chen, and Tomas Pfister. 2023. Search-adaptor: Text embedding customization for information retrieval. arXiv preprint arXiv:2310.08750.",
+        "bbox": [
+            115,
+            123,
+            487,
+            175
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Kai Zhang, Bernal Jiménez Gutiérrez, and Yu Su. 2023. Aligning instruction tasks unlocks large language models as zero-shot relation extractors. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 794-812.",
+        "bbox": [
+            115,
+            186,
+            487,
+            253
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D Manning. 2017. Position-aware attention and supervised data improve slot filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 35-45.",
+        "bbox": [
+            115,
+            263,
+            487,
+            341
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jun Zhao, Wenyu Zhan, Wayne Xin Zhao, Qi Zhang, Tao Gui, Zhongyu Wei, Junzhe Wang, Minlong Peng, and Mingming Sun. 2023a. Re-matching: A fine-grained semantic matching method for zero-shot relation extraction. In ACL 2023, pages 6680-6691.",
+        "bbox": [
+            115,
+            353,
+            487,
+            419
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, and 1 others. 2023b. A survey of large language models. arXiv preprint arXiv:2303.18223.",
+        "bbox": [
+            115,
+            430,
+            487,
+            495
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024. Large language models are not robust multiple choice selectors. In *ICLR* 2024.",
+        "bbox": [
+            115,
+            506,
+            487,
+            546
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Hengyi Zheng, Rui Wen, Xi Chen, Yifan Yang, Yunyan Zhang, Ziheng Zhang, Ningyu Zhang, Bin Qin, Xu Ming, and Yefeng Zheng. 2021. Prgc: Potential relation and global correspondence based joint relational triple extraction. In ACL-IJCNLP 2021, pages 6225-6235.",
+        "bbox": [
+            115,
+            558,
+            487,
+            634
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Details of Generation Augmentation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            663,
+            463,
+            680
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We provide several examples of running Sentence Distillation and Relation Enrichment on WikiZSL in Tables 4 and 5, respectively. These examples all run on GPT-3.5-turbo-0125. It is noteworthy that for all LLMs, the JSON parsing rate can reach over  $99.2\\%$ . For responses that could not be parsed, we directly use the sentences and entities from the raw data as the \"relation_sentence\" and \"entity_type.\"",
+        "bbox": [
+            112,
+            689,
+            487,
+            816
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Details of VanillaPrompt and QA4RE",
+        "text_level": 1,
+        "bbox": [
+            112,
+            831,
+            475,
+            847
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "To investigate the performance of VanillaPrompt and QA4RE on Wiki-ZSL and FewRel, we follow the approach of Zhang et al. (2023), manually crafting templates for all relations. Examples of",
+        "bbox": [
+            112,
+            857,
+            487,
+            921
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "VanillaPrompt and QA4RE running on GPT-3.5-turbo-0125 are provided in Table 6. It is worth noting that the quality of manually constructed templates may affect performance, thus the results we report may contain slight errors.",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Error Analysis",
+        "text_level": 1,
+        "bbox": [
+            509,
+            175,
+            678,
+            193
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "In Figure 8, we report the accuracy of our framework across 80 categories of relations in FewRel. We observe that, after plugging the retrieval adjuster, the retrieval performance for most relation categories improved (indicated in green). For instance, similar relations such as \"P131: located in the administrative territorial entity,\" \"P276: location,\" \"P740: location of formation,\" and \"P937: work location\" can often be confused during retrieval. The retrieval adjuster can learn the relation patterns from seen data and effectively distinguish between similar relations. However, the retrieval adjuster might decrease the performance for a few relations (indicated in red). For example, for the relation \"P400: platform,\" some instances might be predicted as \"P710: participant\" after adjustment. Optimizing the retrieval adjuster to address these challenges will be part of our future work.",
+        "bbox": [
+            505,
+            200,
+            884,
+            491
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Additionally, we observe that our framework performs poorly on the relation \"P750: distributor\" because these instances are often misclassified as other relations such as \"P123: publisher\" or \"P57: director.\" We suspect this happens because they are frequently mentioned in similar contexts. (e.g., in discussions about the commercialization of films or TV shows, these relation patterns might coexist). Therefore, due to this co-occurrence pattern, LLMs may easily focus on incorrect information during sentence distillation or semantic representation, leading to failures. Therefore, we believe that further enriching the semantics of relation labels and improving their distinctiveness will be the direction for future optimization.",
+        "bbox": [
+            507,
+            492,
+            882,
+            732
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "17938",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/f8ef26ed20c6afacf710d550f72f694e6a45068bf7d5b117346b13054a912849.jpg",
+        "image_caption": [
+            "Figure 8: We conduct 10 runs on the FewRel dataset with the setting of  $m = 15$ . Each relation is covered at least once in the unseen relations set  $R_{u}$ . We obtain the average accuracy across all 80 relations. Green and red represent the performance increase or decrease on each relation after adding the retrieval adjuster, respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            142,
+            884,
+            239
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/e1bbaa16296dc7fcc9afe82ad2aab1fe405421d9796d6cfa3eeeae05eec0ed01.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Examples of Sentence Distillation</td></tr><tr><td>Prompt:Provide a summary of the relation between head entity &quot;\\{Head Entity\\}&quot; and tail entity &quot;\\{Tail Entity\\}&quot; in one sentence. Make sure that these are the only two entities in this sentence, and the head entity is in front of the tail entity. Give the entity types of the head entity and tail entity, and finally present them in JSON format.Desired format:\\{&quot;relation}sentence&quot;:&quot;&quot;,&quot;head_type&quot;:&quot;&quot;,&quot;tail_type&quot;:&quot;&quot;} Text:&quot;&quot;\\{Sentence\\}&quot;&quot;</td></tr><tr><td>Input:{Sentence}: Robin Rand (born c. 1956) is a United States Air Force general and commander of Air Force Global Strike Command. \\{Head Entity\\}: Air Force Global Strike Command. \\{Tail Entity\\}: United States Air Force.Output:&quot;relation}sentence&quot;:&quot;Air Force Global Strike Command is commanded by a general from the United States Air Force.&quot;, &quot;head_type&quot;:&quot;Military Unit&quot;, &quot;tail_type&quot;:&quot;Armed Forces&quot;)}.</td></tr><tr><td>Input:{Sentence}: Eye to Ear II is a studio album by English guitarist , composer and improvisor Fred Frith. \\{Head Entity\\}: Fred Frith. \\{Tail Entity\\}: composer.Output:&quot;relation_sentence&quot;:&quot;Fred Frith is a composer.&quot;, &quot;head_type&quot;:&quot;Person&quot;, &quot;tail_type&quot;:&quot;Occupation)}.</td></tr><tr><td>Input:{Sentence}: Guy Rodgers from Temple University was selected before the draft as Philadelphia Warriors territorial pick. \\{Head Entity\\}: Guy Rodgers. \\{Tail Entity\\}: Philadelphia Warriors.Output:&quot;relation_sentence&quot;:&quot;Guy Rodgers was selected as a territorial pick by the Philadelphia Warriors.&quot;, &quot;head_type&quot;:&quot;Person&quot;, &quot;tail_type&quot;:&quot;Sports Team)}.</td></tr><tr><td>Input:{Sentence}: His grandfather Uma Shankar Dikshit was a part of the freedom struggle for thirty years who was closely associated with Pandit Jawahar Lal Nehru , Indira Gandhi and Rajiv Gandhi. \\{Head Entity\\}: Rajiv Gandhi. \\{Tail Entity\\}: Indira Gandhi.Output:&quot;relation_sentence&quot;:&quot;Rajiv Gandhi is the son of Indira Gandhi.&quot;, &quot;head_type&quot;:&quot;Person&quot;, &quot;tail_type&quot;:&quot;Person&quot;.</td></tr><tr><td>Input:{Sentence}: Developed by SanDisk, the Cruiser Enterprise is an encrypted USB flash drive. \\{Head Entity\\}: SanDisk. \\{Tail Entity\\}: USB flash drive.Output:&quot;relation&quot;sentence&quot;:&quot;SanDisk developed the Cruiser Enterprise.&quot;, &quot;head_type&quot;:&quot;Company&quot;, &quot;tail_type&quot;:&quot;Product&quot;).</td></tr></table>",
+        "bbox": [
+            127,
+            414,
+            872,
+            834
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Examples of Sentence Distillation on WikiZSL.",
+        "bbox": [
+            322,
+            843,
+            672,
+            857
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "17939",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/7f0455bd2d3bcf7cc56b4dc96f8a87c49c78ec1249211532843884b59331ee22.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Examples of Relation Enrichment</td></tr><tr><td>Prompt:Please provide all possible types of head entity and tail entity for the given relation, and finally present them in JSON format.Desired format: {&quot;relation&quot;:&quot;&quot;,&quot;head_types&quot;:[], &quot;tail_types&quot;:[]} Relation:&quot;】【Label): {Description}&quot;&quot;</td></tr><tr><td>Input:{Label}: country of citizenship. {Description}: the object is a country that recognizes the subject as its citizenOutput: {&quot;relation&quot;: &quot;country of citizenship&quot;, &quot;head_types&quot;:&quot;Person&quot;], &quot;tail_types&quot;:&quot;Country&quot;]}</td></tr><tr><td>Input:{Label}: director. {Description}: director(s) of film, TV-series, stageplay, video game or similarOutput: {&quot;relation&quot;: &quot;director&quot;, &quot;head_types&quot;:&quot;Film&quot;, &quot;TV-series&quot;, &quot;Stageplay&quot;, &quot;Video game&quot;], &quot;tail_types&quot;:&quot;Person&quot;]}</td></tr><tr><td>Input:{Label}: voice type. {Description}: person&#x27;s voice type. expected values: soprano, mezzo-soprano, contralto, countertenor, tenor, baritone, bass (and derivatives)Output: {&quot;relation&quot;: &quot;voice type&quot;, &quot;head_types&quot;:&quot;person&quot;], &quot;tail_types&quot;:&quot;soprano&quot;, &quot;mezzo-soprano&quot;, &quot;contralto&quot;, &quot;cun-tertenor&quot;, &quot;tenor&quot;, &quot;baritone&quot;, &quot;bass&quot;]}</td></tr><tr><td>Input:{Label}: conflict. {Description}: battles, wars or other military engagements in which the person or item participatedOutput: {&quot;relation&quot;: &quot;conflict&quot;, &quot;head_types&quot;:&quot;person&quot;, &quot;organization&quot;], &quot;tail_types&quot;:&quot;battle&quot;, &quot;war&quot;, &quot;military engage-ment&quot;]}</td></tr><tr><td>Input:{Label}: said to be the same as. {Description}: this item is said to be the same as that item, but the statement is disputedOutput: {&quot;relation&quot;: &quot;said to be the same as&quot;, &quot;head_types&quot;:&quot;Concept&quot;, &quot;Object&quot;], &quot;tail_types&quot;:&quot;Concept&quot;, &quot;Object&quot;]}</td></tr></table>",
+        "bbox": [
+            127,
+            324,
+            872,
+            651
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Examples of Relation Enrichment on WikiZSL.",
+        "bbox": [
+            324,
+            661,
+            670,
+            674
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "17940",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Example of VanillaPrompt",
+        "text_level": 1,
+        "bbox": [
+            129,
+            126,
+            304,
+            139
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Input:",
+        "text_level": 1,
+        "bbox": [
+            134,
+            145,
+            174,
+            156
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Given a sentence and two entities within the sentence, classify the relationship between the two entities based on the provided sentence. Please give the relationship only. All possible relationships are listed below:",
+        "bbox": [
+            134,
+            158,
+            862,
+            183
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- architect",
+            "- main subject",
+            "- winner",
+            "- said to be the same as",
+            "- participant",
+            "-nominated for",
+            "- field of work",
+            "- place served by transport hub",
+            "- tributary",
+            "- military rank",
+            "-league",
+            "- heritage designation",
+            "- member of political party",
+            "- location of formation",
+            "- follows"
+        ],
+        "bbox": [
+            136,
+            185,
+            317,
+            374
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Sentence: California longhorn cattle were given as a gift to Hawaiian King Kamehameha I by British Captain George Vancouver in 1793.",
+        "bbox": [
+            134,
+            376,
+            860,
+            400
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Entity 1: george vancouver",
+        "bbox": [
+            136,
+            403,
+            297,
+            414
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Entity 2: captain",
+        "bbox": [
+            136,
+            416,
+            235,
+            426
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Relationship:",
+        "bbox": [
+            136,
+            429,
+            213,
+            439
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "text_level": 1,
+        "bbox": [
+            136,
+            447,
+            183,
+            458
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Entity 1: George Vancouver",
+        "bbox": [
+            136,
+            460,
+            302,
+            470
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Entity 2: Captain",
+        "bbox": [
+            136,
+            473,
+            238,
+            483
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Relationship: Military rank",
+        "bbox": [
+            136,
+            486,
+            297,
+            495
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Example of QA4RE",
+        "text_level": 1,
+        "bbox": [
+            136,
+            506,
+            262,
+            518
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Input:",
+        "text_level": 1,
+        "bbox": [
+            136,
+            525,
+            174,
+            536
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Determine which option can be inferred from the given Sentence. You can only choose one of the following options.",
+        "bbox": [
+            136,
+            538,
+            818,
+            549
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Sentence: California longhorn cattle were given as a gift to Hawaiian King Kamehameha I by British Captain George Vancouver in 1793.",
+        "bbox": [
+            136,
+            550,
+            860,
+            573
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Options:",
+        "bbox": [
+            136,
+            576,
+            186,
+            587
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A. [george vancouver] is the architect of [captain]",
+        "bbox": [
+            136,
+            589,
+            428,
+            601
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B. [george vancouver] is the main subject of [captain]",
+        "bbox": [
+            136,
+            602,
+            452,
+            614
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C. [george vancouver] is the winner of [captain]",
+        "bbox": [
+            136,
+            615,
+            418,
+            626
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D. [george vancouver] is said to be the same as [captain]",
+        "bbox": [
+            136,
+            627,
+            468,
+            639
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "E. [george vancouver] is a participant in [cap",
+        "bbox": [
+            136,
+            640,
+            400,
+            652
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "F. [george vancouver] was nominated for [captain]",
+        "bbox": [
+            136,
+            653,
+            433,
+            665
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "G. [george vancouver] works in the field of [captain]",
+        "bbox": [
+            136,
+            665,
+            445,
+            678
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "H. [george vancouver] serves as the place served by [captain] transport hub",
+        "bbox": [
+            136,
+            678,
+            578,
+            690
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "I. [george vancouver] is a tributary of [captain]",
+        "bbox": [
+            136,
+            692,
+            411,
+            703
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "J. [george vancouver] holds the military rank of [captain]",
+        "bbox": [
+            136,
+            705,
+            472,
+            717
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "K. [george vancouver] is a member of the league [captain]",
+        "bbox": [
+            136,
+            718,
+            478,
+            728
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "L. [george vancouver] has the heritage designation [captain]",
+        "bbox": [
+            136,
+            731,
+            489,
+            741
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "M. [george vancouver] is a member of the political party [captain]",
+        "bbox": [
+            136,
+            743,
+            524,
+            755
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "N. [george vancouver] is the location where [captain] was formed",
+        "bbox": [
+            136,
+            756,
+            522,
+            766
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "O. [george vancouver] follows [captain]",
+        "bbox": [
+            136,
+            769,
+            371,
+            780
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Which option can be inferred from the given sentence?",
+        "bbox": [
+            136,
+            782,
+            458,
+            793
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Option:",
+        "bbox": [
+            136,
+            795,
+            181,
+            806
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "text_level": 1,
+        "bbox": [
+            136,
+            814,
+            183,
+            825
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "J",
+        "bbox": [
+            136,
+            827,
+            142,
+            835
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Several running examples of VanillaPrompt and QA4RE based on the descriptions in the original paper (Zhang et al., 2023).",
+        "bbox": [
+            115,
+            854,
+            880,
+            878
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "17941",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2025/Generation-Augmented Retrieval_ Rethinking the Role of Large Language Models in Zero-Shot Relation Extraction/06f5d856-0d14-45d9-9bc3-7006f452ae3b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd789d688e87ee528b22615eaac9cfdf77795cd1b81d232461ef8c10645d91ce
+size 851228

2025/Generation-Augmented Retrieval_ Rethinking the Role of Large Language Models in Zero-Shot Relation Extraction/full.md

@@ -0,0 +1,421 @@
+# Generation-Augmented Retrieval: Rethinking the Role of Large Language Models in Zero-Shot Relation Extraction
+
+Zehan Li, Fu Zhang*, Tianyue Peng, He Liu, Jingwei Cheng
+
+School of Computer Science and Engineering, Northeastern University, China Key Laboratory of Intelligent Computing in Medical Image of Ministry of Education Northeastern University, China
+
+{lizehan1999,pengtianyue0716,liuhe0607}@163.com,
+
+{zhangfu,chengjingwei}@neu.edu.cn
+
+# Abstract
+
+Recent advances in Relation Extraction (RE) emphasize Zero-Shot methodologies, aiming to recognize unseen relations between entities with no annotated data. Although Large Language Models (LLMs) have demonstrated outstanding performance in many NLP tasks, their performance in Zero-Shot RE (ZSRE) without entity type constraints still lags behind Small Language Models (SLMs). LLM-based ZSRE often involves manual interventions and significant computational overhead, especially when scaling to large-scale multi-choice data.
+
+To this end, we introduce RE-GAR-AD, which not only leverages the generative capability of LLMs but also utilizes their representational power without tuning LLMs. We redefine LLM-based ZSRE as a retrieval challenge, utilizing a Generation-Augmented Retrieval framework coupled with a retrieval ADjuster. Specifically, our approach guides LLMs through crafted prompts to distill sentence semantics and enrich relation labels. We encode sentences and relation labels using LLMs and match their embeddings in a triplet fashion. This retrieval technique significantly reduces token input requirements. Additionally, to further optimize embeddings, we propose a plug-in retrieval adjuster with only 2M parameters, which allows rapid fine-tuning without accessing LLMs' parameters. Our LLM-based model demonstrates comparable performance on multiple benchmarks.
+
+# 1 Introduction
+
+Relation Extraction (RE) aims to select the relation from a predefined set for two target entities in a given sentence. Traditional supervised RE methods (Wu and He, 2019; Zheng et al., 2021) demand significant annotated data and struggle to handle
+
+![](images/08876b2a495cb73d83ab72adf7c7c4f85ae289f6983313f0e91ee9588478b53d.jpg)  
+Figure 1: Comparison between (a) QA4RE (Zhang et al., 2023), (b) SUMASK (Li et al., 2023), and (c) our approach.
+
+unseen relations effectively. Recently, there has been a growing interest in Few-Shot RE (FSRE) (Yang et al., 2020) and Zero-Shot RE (ZSRE) (Levy et al., 2017; Gong and Eldardiry, 2021), which have limited or no labeled data for novel relations.
+
+Existing research on ZSRE typically involves fine-tuning Small Language Models (SLMs)  $^{2}$  on predicting unseen relations (Zhao et al., 2023a; Li et al., 2024c). Large Language Models (LLMs) have demonstrated impressive performance in various downstream natural language processing (NLP) tasks (Zhao et al., 2023b), leveraging techniques such as In-Context Learning (ICL) (Ma et al., 2023c) and Chain of Thought (CoT) (Ma et al., 2023a) to address FSRE. Recently, LLM-based ZSRE studies have also made significant progress. QA4RE (Zhang et al., 2023) redefines ZSRE as a Question Answering (QA) task and manually construct templates for relation labels. It requires token inputs at the  $n \times r$  level for  $n$  input sentences and  $r$  candidate relations (Figure 1 (a)). However, the inherent selection bias of LLMs may result in poor robustness (Zheng et al., 2024). SUMASK (Li et al., 2023) determines relations by calculating
+
+uncertainty across  $k$  responses of LLMs. The recursive input leads to a significant  $n \times k \times r$  token input overhead (Figure 1 (b)), restricting its scalability on large-scale data. Therefore, making ZSRE more efficient and accurate while benefiting from LLMs is a key concern for the current system.
+
+In this work, we explore new possibilities for LLM-based ZSRE and propose the RE-GAR-AD framework, which is based on the Generation-Augmented Retrieval (GAR) paradigm and incorporates a retrieval ADjuster. Considering the potential of LLMs as encoders for obtaining embeddings (BehnamGhader et al., 2024), we rethink and redefine LLM-based ZSRE as an embedding-based retrieval problem. We input generation-augmented sentences and relation labels separately to LLMs to obtain their embeddings and conduct sentence and relation retrieval through matching scores. All candidate relation labels only need to be input once and kept fixed, requiring only token inputs at  $n + r$  level as shown in Figure 1 (c). This not only reduces the inference cost of large-scale multi-choice ZSRE data for LLMs but also demonstrates performance comparable to that of existing methods.
+
+Specifically, we leverage the generation capability of LLMs to improve retrieval efficiency. Considering that a sentence may contain more than one relation pattern, we distill valuable information for RE from sentences using Sentence Distillation Prompt, which includes the types of two target entities and their possible semantic relation. Additionally, we guide LLMs to utilize their internal general knowledge to generate deep information about relation labels through the Relation Enrichment Prompt, which explores possible entity types relevant to the relation. We then use LLMs to encode and obtain embeddings of sentences and relation labels, and achieve more accurate relation retrieval through Triplet-based Retrieval. Furthermore, SLMs-based ZSRE fine-tuning typically involves learning prior knowledge from seen data, which may be computationally expensive for LLMs and impractical for API-based LLMs (Yoon et al., 2023; Li et al., 2024b). Therefore, we carefully design a plug-in Retrieval Adjuster for the framework. It aims to learn the ability to align sentence and relation label representations from seen data, and further adjust and optimize the embeddings of unseen data without accessing the full parameters of LLMs, thereby improving retrieval efficiency. Our contributions are summarized as follows:
+
+- We redefine LLM-based ZSRE as a retrieval task and propose a novel Generation-Augmented Retrieval (GAR) framework, significantly reducing the token inputs while ensuring accuracy.  
+- We carefully design two prompts further to process the semantics of sentences and relation labels using the generative ability of LLMs, improving retrieval efficiency.  
+- We introduce a retrieval adjuster plugin for the GAR framework, with only 2M parameters, which can enhance retrieval performance without accessing the parameters of LLMs.  
+- Experimental results show that RE-GAR-AD achieves competitive performance across five ZSRE benchmark datasets. Without involving fine-tuning, our approach outperforms the state-of-the-art LLM-based method by 6.93.
+
+# 2 Related Work
+
+# 2.1 LLMs for Language Representation
+
+Large Language Models (LLMs) are typically trained on large corpora and have demonstrated advanced performance across many downstream NLP tasks (Zhao et al., 2023b). For tasks such as retrieval and clustering that require sentence embeddings, a common approach involves leveraging the representation of the end-of-sequence token of LLMs as sentence embeddings (Ma et al., 2023b). Considering that causal attention might limit the representation capability of decoder-only LLMs, recent works introduce bidirectional attention (Muennighoff et al., 2024) and echo mechanisms (Springer et al., 2024) to overcome limitations. These works demonstrate that generative LLMs are also proficient at language representation (Wang et al., 2023), and any robust decoder-only LLMs can produce universal text embeddings with minor adaptation (BehnamGhader et al., 2024).
+
+# 2.2 LLMs for Relation Extraction
+
+The existing work on LLM-based RE can be categorized into three paradigms: Tuning-based methods involve fine-tuning LLMs with accessible parameters (Kim et al., 2023; Sun et al., 2024). Recent works adapt LLMs to RE tasks through annotation guidelines (Sainz et al., 2024) or tabular prompting (Li et al., 2024a). ICL-based methods facilitate In-Context Learning (ICL) (Min
+
+![](images/e13709f946280b437af3d15445ba77cfc2b7514e1355f40fad8b2725695d88ca.jpg)  
+Figure 2: An overview of the proposed RE-GAR-AD framework. It consists of three key components: (1) Generation-Augmented Embeddings, where Generative LLMs distill semantic information from sentences and relation descriptions using two prompts; (2) Retrieval Adjuster, which calibrates embeddings using relation patterns learned from seen data; and (3) Triplet-based Retrieval, where embedding triplets are matched to infer the final relation label.
+
+et al., 2022) by observing demonstrations within prompts without updating the LLMs parameters, which is applied in FSRE. They use the Retrieval-Augmented Generation (RAG) paradigm (Rubin et al., 2022) to retrieve suitable demonstrations, which may include quality analysis (Wan et al., 2023) or logical reasoning (Ma et al., 2023c). Instruction-based methods execute tasks through explicit instructions such as multiple-choice QA (Zhang et al., 2023) or uncertainty estimation (Li et al., 2023) without demonstrations, making them applicable to ZSRE.
+
+A common limitation of the above methods is the increase in candidate relations along with a large number of token inputs, which consumes time and resources in large-scale ZSRE. Additionally, post-processing may be required to address output format issues. To address these challenges, we propose a novel LLM-based paradigm called Generative-Augmented Retrieval (GAR). Unlike RAG, GAR directly identifies relations through retrieval, fundamentally avoiding format issues and reducing computational costs by separately modeling sentences and relations. Furthermore, we devise a retrieval adjuster that can be plugged into LLMs to further improve ZSRE performance.
+
+# 3 Methodology
+
+# 3.1 Problem Formulation
+
+Given a predefined set of seen relations  $R_{s}$  and unseen relations  $R_{u}$ , with no overlap between them. Given seen data  $D_{s}$  and unseen data  $D_{u}$ . The ZSRE's objective is to predict  $r_{u} \in R_{u}$
+
+between the head entity  $e^1$  and the tail entity  $e^2$  in the input sentence  $x_u$  for unseen instance  $I_u = (x_u, e_u^1, e_u^2) \in D_u$ . Before prediction, learning can be conducted from seen instances  $I_s = (x_s, e_s^1, e_s^2, r_s) \in D_s, r_s \in R_s$ , and the knowledge can be transferred to predict.
+
+# 3.2 Overview of RE-GAR-AD
+
+The proposed RE-GAR-AD framework, which redefines LLM-based ZSRE as an embedding-based retrieval problem, consists of three modules as shown in Figure 2:
+
+(1) Generation-Augmented Embeddings involves the use of Generative LLMs to distill and enrich sentences and relation labels through two carefully designed prompts. These are then encoded by Embedding LLMs to obtain fixed-dimensional embedding triplets. (2) Retrieval Adjuster learns from the relation patterns of the seen dataset  $D_{s}$  and is used to adjust the embedding triplets corresponding to the unseen sentences and relations, thereby improving retrieval performance. It is worth noting that we use the same data processing method in the testing phase as in the training phase, applying both Generation Augmentation and Retrieval Adjustment. (3) Through Triplet-based Retrieval, the adjusted sentence embedding triplets are matched with the adjusted relation embedding triplets, and the highest scoring relation is chosen as the final relation.
+
+# 3.3 Generation-Augmented Embeddings
+
+Sentence Distillation Prompt. To distill crucial content for RE from sentences with redundant infor
+
+mation, we propose a Sentence Distillation Prompt comprising three carefully crafted instructions. For a given input sentence  $x_{u}$  with two target entities  $e^{1}, e^{2}$ , the Relation Summarization instruction requires LLMs to encapsulate the semantic relation between the annotated entities within one sentence, denoted as  $x_{u}^{*}$ . This requires LLMs to utilize their summarization abilities. The Type Identification instruction aims to infer the types of the head and tail entities that are crucial for RE (Lyu and Chen, 2021), represented as  $T_{u}^{1}$  and  $T_{u}^{2}$ , respectively. The final triplet output is organized into a JSON object using the Triplet Combination instruction, described as  $\{x_{u}^{*}, T_{u}^{1}, T_{u}^{2}\}$ .
+
+Relation Enrichment Prompt. To leverage the general knowledge of LLMs for mining deep information from relation labels and descriptions, we introduce a Relation Enrichment Prompt. It aims to instruct LLMs to freely generate all possible entity types  $\hat{T_u}$  associated with relation labels. This automated process effectively alleviates the manual labor involved in annotating head entity types and tail entity types for relations. Subsequently, we organize them into triplets in JSON format:  $\{r_u,\{T_{u,i}^1\}_{i = 1}^n,\{T_{u,j}^2\}_{j = 1}^m\}$ .
+
+Embedding with LLMs. To match sentence triplets and relation triplets, we embed them into the same embedding space. We obtain an embedding  $h \in \mathbb{R}^d$  for each element of the sentence triplets and relation triplets by calling the open API of the latest GPT series model "text-embedding-3-small", which is trained via matryoshka representation learning (Kusupati et al., 2022). It is noteworthy that for  $\hat{T}_u$ , we conduct average pooling for each set to obtain  $h_{\hat{T}_u}$  of entity types. Each sentence embedding triplet and relation embedding triplet can be represented respectively as  $\mathrm{Tri}_x = \{h_{x_u^*}, h_{T_u^1}, h_{T_u^2}\}$  and  $\mathrm{Tri}_r = \{h_{r_u}, h_{\hat{T}_u^1}, h_{\hat{T}_u^2}\}$ .
+
+# 3.4 Retrieval Adjuster
+
+Embeddings from LLMs have a potential anisotropy problem due to common-word biases (Li and Li, 2024), causing the embeddings to occupy a narrow cone. Therefore, contrastive learning-based fine-tuning is required to distribute the embeddings uniformly (Gao et al., 2021).
+
+However, fine-tuning LLMs for specific tasks is resource-intensive and necessitates full access to the parameters of the LLMs, posing challenges with API-based LLMs. Therefore, we propose an independent retrieval adjuster (AD) for adapting
+
+ZSRE, aimed at learning the ability to align sentence and relation label representations from seen data and further adjust the unseen embeddings. AD does not require access to LLMs' parameters and can easily plug into LLMs.
+
+To adapt to the task while retaining the LLM-based embedding semantics, AD integrates residual features learned by network  $\mathrm{F}(\cdot)$  with the original embeddings through skip connections to produce the final embedding  $\tilde{h}$ :
+
+$$
+\tilde {h} = h + \mathrm {F} (h). \tag {1}
+$$
+
+We employ a straightforwardly connected network as  $\mathrm{F}(\cdot)$ , with detailed operations as follows:
+
+$$
+\mathrm {F} (h) = W _ {2} \left(\operatorname {R e L U} \left(W _ {1} (h) + b _ {1}\right)\right) + b _ {2}, \tag {2}
+$$
+
+where  $W_{1}\in \mathbb{R}^{d\times n}$ ,  $W_{2}\in \mathbb{R}^{n\times d}$  are parameter matrices, and  $b_{1},b_{2}$  are bias. Notably, inspired by Yoon et al. (2023), we employ a shared  $\mathrm{F}(\cdot)$  for all embeddings, thereby maintaining a unified embedding space to achieve better retrieval.
+
+# 3.5 Triplet-based Retrieval
+
+To enable  $\mathrm{Tri}_x$  to retrieve the corresponding  $\mathrm{Tri}_r$ , we introduce triplet score  $c$  to compute the similarity coefficient:
+
+$$
+c \left(\operatorname {T r i} _ {x}, \operatorname {T r i} _ {r}\right) = \alpha \left\langle h _ {x ^ {*}}, h _ {r} \right\rangle + \tag {3}
+$$
+
+$$
+\left(1 - \alpha\right) \left(\langle h _ {T ^ {1}}, h _ {\hat {T} ^ {1}} \rangle + \langle h _ {T ^ {2}}, h _ {\hat {T} ^ {2}} \rangle\right),
+$$
+
+where  $\alpha$  is hyperparameter, and  $\langle ,\rangle$  denotes the cosine similarity. We train AD on seen data  $D_{s}$  to align the embeddings of  $\mathrm{Tri}_{x_s}$  and  $\mathrm{Tri}_{r_s}$ . The training objective of  $\mathrm{F}(\cdot)$  is to maximize the triplet score between  $\tilde{\mathrm{Tri}}_{x_s}^i$  and positive  $\tilde{\mathrm{Tri}}_{r_s}^i$ , while minimizing its score with negative  $\tilde{\mathrm{Tri}}_{r_s}^j$ :
+
+$$
+\begin{array}{l} \mathcal {L} = \frac {1}{N} \sum_ {i = 1} ^ {N} \max  \left\{0, \max  \left(c \left(\tilde {\operatorname {T r i}} _ {x _ {s}} ^ {i}, \tilde {\operatorname {T r i}} _ {r _ {s}} ^ {j}\right)\right) \right. \tag {4} \\ - c \left(\tilde {\operatorname {T r i}} _ {x _ {s}} ^ {i}, \tilde {\operatorname {T r i}} _ {r _ {s}} ^ {i}\right) + \Delta \}, i \neq j, \\ \end{array}
+$$
+
+where  $N$  denotes the batch size, and  $\Delta$  represents a predefined threshold. Negative samples are randomly selected within the same batch, maintaining a 1:3 ratio of positive to negative samples. This approach can promote the uniform distribution of sentence embeddings and relation embeddings, thereby alleviating the anisotropy problem.
+
+The trained AD is utilized to adjust the embeddings of unseen  $\mathrm{Tri}_{x_u}$  and  $\mathrm{Tri}_{r_u}$  during the testing phase. The highest computed  $c(\tilde{\mathrm{Tri}}_{x_u},\tilde{\mathrm{Tri}}_{r_u})$  is chosen as the final relation.
+
+# 4 Experiments Setup
+
+# 4.1 Datasets
+
+We conduct experiments on five benchmarks. Following established setups (Zhao et al., 2023a), we conduct our primary experiments on two widely used datasets: Wiki-ZSL (Chen and Li, 2021) and FewRel (Han et al., 2018). Wiki-ZSL contains 94,383 instances and 113 relations, and is acquired through distant supervision from Wiki-KB, characterized by imbalanced samples and heightened noise. FewRel contains 56,000 instances and 80 relations, and is meticulously annotated, ensuring equal samples per relation. Consistent with prior works, we divide the datasets into a training set containing seen relations and a validation set containing unseen relations, focusing on evaluating under unseen relations  $m = 5, 10, 15$ . To ensure the reliability of results, we perform 10-fold cross-validation and report the average results. Notably, the entity types are not provided in Wiki-ZSL and FewRel, posing challenges to ZSRE as candidate relations cannot be filtered based on entity types.
+
+Moreover, following the setups of other LLM-based methods (Li et al., 2023), we also select three datasets where entity types are given: (1) TACRED (Zhang et al., 2017) contains 42 relations and 17 entity types, (2) TACREV (Alt et al., 2020) contains 42 relations and 17 entity types, and (3) Re-TACRED (Stoica et al., 2021) contains 40 relations and 17 entity types. We further evaluate on their test sets with the none-of-the-above (NoTA) relation excluded.
+
+# 4.2 Implementation Details
+
+We set temperature  $= 0$  for GPT series models to improve the reproducibility. We apply the same prompt for Wiki-ZSL and FewRel. Since the entity types are seen in TACRED, TACREV, and Re-TACRED, we remove instructions for generating entity types from prompt. Following previous work, we filter a subset of candidate relations based on the entity types. For retrieval adjuster AD, we set the same hyperparameters for all experiments. We use AdamW with a learning rate of  $1e - 5$ , 10 epochs, and a batch size of 64. We set the dimension of the hidden layer  $n$  to  $d / 2$ , resulting in approximately 2.36M parameters for AD, which is much smaller than BERT-based SLMs. We conducted a grid search on FewRel and determined hyperparameters  $\alpha = 0.75, \Delta = 0.1$ , negative ratio  $= 1:3$  when the model performed the best.
+
+# 4.3 Baseline Models
+
+SLMs-based methods. We compare: embedding-based methods R-BERT (Wu and He, 2019), ZSBERT (Chen and Li, 2021), and RE-Matching (Zhao et al., 2023a); text entailment-based methods ESIM (Chen et al., 2017), LaVeEntail (Sainz et al., 2021); and generation-based methods RE-Prompt (Chia et al., 2022), SuRE (Lu et al., 2022). LLM-based methods. We compare previous competitive baselines, including VanillaPrompt, QA4RE, SUMASK, and MICRE. VanillaPrompt and QA4RE are from Zhang et al. (2023), respectively guide LLMs to generate relation labels or multiple-choice options, where QA4RE involves manually building relation templates. As they did not evaluate on Wiki-ZSL and FewRel, we rerun the results based on the original paper and available code. SUMASK (Li et al., 2023) classifies relations by generating multiple responses for all candidate relations and estimating uncertainty from LLMs. MICRE (Li et al., 2024a) uses 12 public datasets to fine-tune LLaMA (Touvron et al., 2023a) through tabular prompting, achieving performance competitive with SLMs. Following the previous approach, we randomly sample 1,000 examples as test data from each dataset to control the cost of comparison with the LLM-based baselines. Since using seen data for training the AD incurs additional costs and may not be available in some datasets, we provide two sets of results indicating whether seen data are used to train AD, denoted as RE-GAR and RE-GAR-AD, respectively.
+
+# 5 Analysis and Discussion
+
+# 5.1 Main Result
+
+We report the main results on Wiki-ZSL and FewRel in Table 1. Overall, our approach demonstrates competitive performance across all settings on both datasets.
+
+(1) Compared to LLM-based methods, our approach ensures strong performance even without considering seen data (RE-GAR's average F1 is 6.93 higher than QA4RE across six settings). We attribute this to the novel GAR framework, which effectively utilizes both the generative and representational capabilities of LLMs, allowing LLMs to unleash more potential on ZSRE tasks. Additionally, we observe that when considering seen data, RE-GAR-AD achieves even stronger performance with fewer training samples (with an F1 advantage of 3.71 over MICRE).
+
+<table><tr><td rowspan="3">Methods</td><td colspan="11">Wiki-ZSL</td><td colspan="7">FewRel</td></tr><tr><td colspan="3">m=5</td><td colspan="3">m=10</td><td colspan="3">m=15</td><td colspan="3">m=5</td><td colspan="3">m=10</td><td colspan="3">m=15</td></tr><tr><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td><td>P.</td><td>R.</td><td>F1</td></tr><tr><td colspan="19">- SLMs-based</td></tr><tr><td>R-BERT*</td><td>39.22</td><td>43.27</td><td>41.15</td><td>26.18</td><td>29.69</td><td>27.82</td><td>17.31</td><td>18.82</td><td>18.03</td><td>42.19</td><td>48.61</td><td>45.17</td><td>25.52</td><td>33.02</td><td>25.52</td><td>16.95</td><td>19.37</td><td>18.08</td></tr><tr><td>ESIM*</td><td>48.58</td><td>47.74</td><td>48.16</td><td>44.12</td><td>45.46</td><td>44.78</td><td>27.31</td><td>29.62</td><td>28.42</td><td>56.27</td><td>58.44</td><td>57.33</td><td>42.89</td><td>44.17</td><td>43.52</td><td>29.15</td><td>31.59</td><td>30.32</td></tr><tr><td>ZS-BERT*</td><td>71.54</td><td>72.39</td><td>71.96</td><td>60.51</td><td>60.98</td><td>60.74</td><td>34.12</td><td>34.38</td><td>34.25</td><td>76.96</td><td>78.86</td><td>77.90</td><td>56.92</td><td>57.59</td><td>57.25</td><td>35.54</td><td>38.19</td><td>36.82</td></tr><tr><td>LaVeEntail*</td><td>77.39</td><td>75.90</td><td>76.63</td><td>71.86</td><td>71.14</td><td>71.50</td><td>62.13</td><td>61.76</td><td>61.95</td><td>91.14</td><td>90.86</td><td>91.00</td><td>83.05</td><td>82.55</td><td>82.80</td><td>72.83</td><td>72.10</td><td>72.46</td></tr><tr><td>REPrompt*</td><td>70.66</td><td>83.75</td><td>76.63</td><td>68.51</td><td>74.76</td><td>71.50</td><td>63.69</td><td>67.93</td><td>65.74</td><td>90.15</td><td>88.50</td><td>89.30</td><td>80.33</td><td>79.62</td><td>79.96</td><td>74.33</td><td>72.51</td><td>73.40</td></tr><tr><td>RE-Matching*</td><td>78.19</td><td>78.41</td><td>78.30</td><td>74.39</td><td>73.54</td><td>73.96</td><td>67.31</td><td>67.33</td><td>67.32</td><td>92.82</td><td>92.34</td><td>92.58</td><td>83.21</td><td>82.64</td><td>82.93</td><td>73.80</td><td>73.52</td><td>73.66</td></tr><tr><td colspan="19">- LLM-based</td></tr><tr><td>VanillaPrompt</td><td>79.83</td><td>56.56</td><td>66.21</td><td>62.98</td><td>58.05</td><td>60.41</td><td>58.54</td><td>51.26</td><td>54.66</td><td>74.87</td><td>51.15</td><td>60.77</td><td>65.28</td><td>46.46</td><td>54.29</td><td>59.70</td><td>50.56</td><td>54.76</td></tr><tr><td>QA4RE</td><td>87.20</td><td>74.97</td><td>80.62</td><td>65.80</td><td>58.25</td><td>61.80</td><td>62.17</td><td>57.80</td><td>59.90</td><td>88.00</td><td>75.66</td><td>81.37</td><td>79.00</td><td>73.26</td><td>76.02</td><td>65.33</td><td>67.45</td><td>66.37</td></tr><tr><td>SUMASK†</td><td>75.64</td><td>70.96</td><td>73.23</td><td>62.31</td><td>61.08</td><td>61.69</td><td>43.55</td><td>40.27</td><td>41.85</td><td>78.27</td><td>72.55</td><td>75.30</td><td>64.77</td><td>60.94</td><td>62.80</td><td>44.76</td><td>41.13</td><td>42.87</td></tr><tr><td>MICRE†</td><td>76.46</td><td>78.53</td><td>77.48</td><td>72.36</td><td>74.88</td><td>73.60</td><td>67.14</td><td>68.87</td><td>67.99</td><td>89.34</td><td>91.88</td><td>90.59</td><td>80.67</td><td>82.31</td><td>81.48</td><td>73.74</td><td>75.83</td><td>74.77</td></tr><tr><td colspan="19">- Ours (LLM-based)</td></tr><tr><td>RE-GAR</td><td>85.13</td><td>83.85</td><td>84.44</td><td>74.98</td><td>73.32</td><td>74.12</td><td>68.45</td><td>66.35</td><td>67.37</td><td>90.75</td><td>89.40</td><td>90.07</td><td>80.89</td><td>78.76</td><td>79.81</td><td>73.48</td><td>70.26</td><td>71.82</td></tr><tr><td>RE-GAR-AD</td><td>88.96</td><td>88.42</td><td>88.63</td><td>77.63</td><td>77.85</td><td>77.70</td><td>70.80</td><td>69.26</td><td>70.02</td><td>93.03</td><td>92.16</td><td>92.59</td><td>83.48</td><td>82.10</td><td>82.78</td><td>77.69</td><td>75.31</td><td>76.47</td></tr></table>
+
+(2) Compared to SLMs-based methods trained on seen data, RE-GAR-AD achieves up to 5.59 higher F1 on Wiki-ZSL than the existing SOTA model, RE-Matching. Interestingly, our method outperforms other baselines by a large margin at  $m = 5$  on Wiki-ZSL, and we attribute this improvement partly to the effective noise reduction in the distant supervision dataset through Sentence Distillation. It is worth noting that the parameter of the fine-tuned retrieval adjuster is much smaller than BERT, which is adopted as the backbone for most SLM-based methods.
+
+(3) Additionally, in Table 2, we report the micro-averaged F1 on TACRED, TACREV, and Re-TACRED. Since entity types are provided, we follow the previous works (Zhang et al., 2023; Li et al., 2023) by discarding irrelevant relations. The results indicate that RE-GAR achieves comparable performance on three datasets.
+
+# 5.2 Ablation Experiment
+
+We report the impact of different components on RE-GAR-AD, as shown in Table 3.
+
+1. We use Embedding LLMs to encode the raw sentences when we remove the Sentence Distillation. The average performance drops by 7.74. This indicates that the summarization capability of LLMs can effectively reduce irrelevant information in sentences, allowing them to focus more on the semantic association between entities, thus gaining an advantage in relation matching.
+
+2. We manually annotate possible entity types for relations when we remove the Relation Enrichment. We observe a slight decrease in average performance by 1.67, indicating that LLMs have the potential to generate data of comparable quality to manually annotated data, and that richer information may yield greater benefits.
+
+3. We solely match sentences and relation labels
+
+Table 1: Experimental results on Wiki-ZSL and FewRel. We mark the best results in bold and the second-best underlined. All generation based on API are run on GPT-3.5-turbo. RE-GAR and RE-GAR-AD respectively represent the results without using seen data, and the results of fine-tuning our retrieval adapter with only seen data. The results indicated by  $\dagger$  and  $*$  are reported in MICRE (Li et al., 2024a) and RE-Matching (Zhao et al., 2023a). P., R., and F1 denote Precision, Recall, and F1 score, respectively.  
+
+<table><tr><td>Methods</td><td>TACRED</td><td>TACREV</td><td>Re-TACRED</td></tr><tr><td>LaVeEntail_BART‡</td><td>51.4</td><td>55.3</td><td>44.0</td></tr><tr><td>LaVeEntail_ DeBERTa‡</td><td>55.1</td><td>57.2</td><td>64.3</td></tr><tr><td>SuRE_BART‡</td><td>20.4</td><td>22.2</td><td>23.6</td></tr><tr><td>SuRE_PEGASUS‡</td><td>21.8</td><td>21.6</td><td>22.4</td></tr><tr><td>VanillaPrompt‡</td><td>48.1</td><td>51.0</td><td>55.3</td></tr><tr><td>QA4RE‡</td><td>59.4</td><td>59.4</td><td>61.2</td></tr><tr><td>SUMASK†</td><td>57.7</td><td>-</td><td>-</td></tr><tr><td>RE-GAR</td><td>64.8</td><td>63.4</td><td>53.8</td></tr></table>
+
+Table 2: Experimental results on TACRED, TACREV, and Re-TACRED. The results indicated by  $\dagger$  and  $\ddagger$  are reported in SUMASK (Li et al., 2023) and QA4RE (Zhang et al., 2023).  
+
+<table><tr><td>Methods</td><td>m=5</td><td>m=10</td><td>m=15</td></tr><tr><td>RE-GAR-AD</td><td>92.59</td><td>82.78</td><td>76.47</td></tr><tr><td>w/o. Sentence Distillation</td><td>88.45</td><td>73.06</td><td>67.12</td></tr><tr><td>w/o. Relation Enrichment</td><td>92.06</td><td>80.13</td><td>74.64</td></tr><tr><td>w/o. Triplet-based Retrieval</td><td>82.18</td><td>68.44</td><td>60.32</td></tr><tr><td>w/o. Retrieval Adjuster</td><td>90.07</td><td>79.81</td><td>71.82</td></tr></table>
+
+Table 3: Ablation study of RE-GAR-AD. We report the F1 performance on FewRel under  $m = 5,10,15$
+
+![](images/5a55b687451ed87757f7afa80f1a91694fb7b0cebc91c3dcc827bdfcea3229fb.jpg)  
+Figure 3: Effect of data scale, data quality, and residual connections on retrieval adjuster AD.
+
+![](images/d9ead2937ebf408a3305a3f47400452dd7b0bcad783309f35977344745acbf2b.jpg)  
+Figure 4: Comparison of API overhead (Bar) and F1 score (Line) for 8,000 instances. To control the cost, we randomly sample 100 samples for each relation.
+
+when we remove the Triplet-based Retrieval (i.e., without considering entity type information in the retrieval), there is a dramatic decline in average performance (-13.63), which proves the crucial role of entity information in relation retrieval.
+
+4. When Retrieval Adjuster AD is removed, it means that we do not benefit from fine-tuning, leading to an average performance drop of 3.42.
+
+# 5.3 Analysis on Retrieval Adjuster
+
+Retrieval adjuster AD needs to be trained on seen data, which may introduce additional generation costs and training overhead. Therefore, we further investigate the impact of training data of varying scales and qualities on the performance of AD under the setting of  $m = 15$  on the FewRel dataset. Specifically, we randomly extract  $\{1\%, 10\%, 50\% \}$  of the samples from the seen data to construct subsets, maintaining an even balance of samples per relation.
+
+Data scale. As shown in Figure 3, more training samples can bring greater benefits to AD, which is consistent with intuition. Notably, only  $10\%$  of the distilled seen data can lead to a 2.48 F1 score increase, while increasing the data scale tenfold only further improves the F1 score by 2.17.
+
+Data quality. Even if the data is not distilled (raw data), AD can still enhance performance without negatively impacting the model. Since this strategy does not require LLMs to generate, it can
+
+be more cost-effective in some scenarios.
+
+Residual connection. We observe that removing residual connection when the training data scale is small leads to a performance decrease, indicating that residual connection can effectively improve information flow and reduce overfitting. However, with sufficient data scale, the additional benefits of residual connection may not be as apparent.
+
+# 5.4 Model Efficiency
+
+To evaluate our approach from both performance and cost perspectives, we compare it with LLM-based baselines on FewRel with settings of  $m = 5,10,15,20,30,40$  and report the results in Figure 4. For a fair comparison, we only report the performance of RE-GAR, which does not benefit from fine-tuning on seen data. Moreover, as described in Section 1, SUMASK requires an input overhead of  $n\times k\times r$  tokens. Therefore, we choose QA4RE as the baseline, which has a more efficient input overhead  $n\times r$ . The results are derived from GPT-3.5-turbo and text-embedding-3-small. The estimation of API costs is sourced from OpenAI
+
+The results indicate that RE-GAR consistently demonstrates advantages in both F1 score and cost compared to QA4RE across all settings of  $m$ . Moreover, as  $m$  increases, RE-GAR's performance is only mildly affected. Additionally, since relation labels serving as retrieval targets are only pre-embedded once and fixed, the average cost of RE-GAR in large-scale ZSRE scenarios only undergoes slight changes as  $m$  increases.
+
+# 5.5 Scalability Analysis
+
+We conduct scalability research on SLMs, open-source LLMs, and API-based LLMs, and report F1 score on FewRel with  $m = 15$ . We select 100 samples for each relation to build subsets to control costs. Note that the generation-augmentation and retrieval are executed in a pipelined manner, and their combination is not fixed.
+
+In Figure 5, we compare different embedding models. For SLMs, we choose SROBERTa (Reimers and Gurevych, 2019), PromptROBERTa (Jiang et al., 2022), SimCSE-Roberta (Gao et al., 2021), and BGE-en (Xiao et al., 2023). For open-source LLMs, we introduce LLM2Vec (BehnamGhader et al., 2024), which converts LLaMA2 (Touvron et al., 2023b) and Mistral (Jiang et al., 2023) into general text encoders. In addition,
+
+![](images/4117becbccd735b45b70a8a16096d0b035647d6eb82dd807eb42becf6f054744.jpg)  
+Figure 5: Scalability analysis: Generate with GPT-3.5-turbo. Embedding with different SLMs and LLMs.
+
+![](images/9f0a43d085b2c3e226ea233ad9cbbaf4f7ad1223fce239dfbcfe8a23bbd26933.jpg)  
+Figure 6: Scalability analysis: Embedding with text-embedding-3-small. Generate with different LLMs.  
+Figure 7: Case study on FewRel with  $m = 15$ . Sentence, Head Entity, and Tail Entity are derived from the raw data. Distilled Sentence represents the data distilled through the RE-GAR-AD framework.
+
+we explore different API-based LLMs (OpenAI text-embedding (Neelakantan et al., 2022)). We observe that the performance of SLMs generally does not exceed 61, which may indicate a significant gap in representation capabilities between SLMs and LLMs on ZSRE, revealing the potential of LLMs in representation. RE-GAR-AD demonstrates excellent performance on all LLMs, proving its robustness.
+
+In Figure 6, we compare different generative models, including LLaMA3, Mistral, GPT-3.5-turbo, and GPT-4-turbo, which achieve remarkable performance with their strong generation capabilities. We observe that more powerful LLMs such as text-embedding-3-large and GPT-4-turbo can provide sustained benefits in RE-GAR-AD.
+
+Furthermore, we notice that our proposed plugin in retrieval adjuster consistently shows improvement, indicating that it is model-agnostic.
+
+# 5.6 Error Analysis and Case Study
+
+To investigate the effectiveness of our RE-GAR framework, we provide a typical example as shown in Case 1 of Figure 7. For QA4RE, we construct relation-entailed templates as label verbalization: "P27: country of citizenship" is processed into "[Andrew Goldberg] is a citizen of [United States]", and "P361: part of" is processed into "[Andrew Goldberg] is part of [United States]". Due to the semantic similarity between the two options, generative LLMs might confuse them and output the incorrect option. RE-GAR refines the semantic relationship between Andrew Goldberg and the United
+
+<table><tr><td>Sentence</td><td colspan="2">After leaving Island Records due to a disagreement over the band&#x27;s artistic direction, Fightstar signed with the independent label Institute Records for their second album.</td></tr><tr><td colspan="2">Head Entity: Fightstar</td><td>Tail Entity: Island Records</td></tr><tr><td>Distilled Sentence</td><td colspan="2">Fightstar left Island Records due to a disagreement over artistic direction. Band. Record Label.</td></tr><tr><td>Enriched relation (P175)</td><td>Actor, Musician, Band</td><td>Musical work</td></tr><tr><td>Enriched relation (P264)</td><td>Music recordings, Music videos</td><td>Record label, Brand, Trademark</td></tr><tr><td>Ground Truth</td><td>P264: record label</td><td>✔</td></tr><tr><td>QA4RE</td><td>P175: performer</td><td>✘</td></tr><tr><td>RE-GAR</td><td>P175: performer</td><td>✘</td></tr><tr><td>RE-GAR-AD</td><td>P264: record label</td><td>✔</td></tr></table>
+
+States, yielding the accurate information: "Andrew Goldberg is a writer from the United States." Additionally, it pays extra attention to the type information of the two entities (Person and Country).
+
+As shown in Case 2 of Figure 7, when LLMs fail to capture meaningful semantic connections between Fightstar and Island records, it can lead to ineffective retrieval. Our proposed plug-in retrieval adjuster addresses this by learning relation patterns from seen data and subsequently adjusting when retrieving unseen relation labels, successfully categorizing the relation as "record label".
+
+Additionally, we report the performance of our framework on 80 relation categories in Appendix C. We observe that although RE-GAR-AD can distinguish similar relations better with the retrieval adjuster, challenges still exist when co-occurring relations are present in sentences. This will be one of our future optimization directions.
+
+# 6 Conclusion
+
+We present RE-GAR-AD, a simple and effective ZSRE framework based on LLMs. We introduce a novel LLM-based ZSRE paradigm of Generation-Augmented Retrieval, leveraging LLMs' generative and representational capabilities. Additionally, we train a plug-in retrieval adjuster to further optimize embeddings. Extensively evaluated on common ZSRE benchmarks, RE-GAR-AD demonstrates comparable performance across various settings, especially in cases where entity types are excluded. Furthermore, our method significantly improves the efficiency of LLMs in multi-choice relation extraction, which is advantageous for real-world ZSRE and opens up interesting avenues for future zero-shot classification efforts.
+
+# Limitations
+
+Although the proposed framework demonstrates advanced performance and cost-effectiveness on large-scale ZSRE, its cost-effectiveness is not evident when the number of input sentences  $n$  is small. Additionally, due to API limitations, we are unable to obtain complete results from more powerful models such as GPT-4 and others. Furthermore, we employ only a straightforward structure for the retrieval adjuster, and we believe that further research on LLM-embedding-based adjusters would be an interesting direction. Finally, since our method encounters difficulties in handling the none-of-the-above (NoTA) relation, it may not be suitable for cases where the dataset contains the NoTA relation.
+
+# Acknowledgments
+
+The authors sincerely thank the reviewers for their valuable comments, which improved the paper. The work is supported by the National Natural Science Foundation of China (62276057).
+
+# References
+
+Christoph Alt, Aleksandra Gabryszak, and Leonhard Hennig. 2020. Tacred revisited: A thorough evaluation of the tacred relation extraction task. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1558-1569.  
+Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961.  
+Chih Yao Chen and Cheng-Te Li. 2021. Zs-bert: Towards zero-shot relation extraction with attribute representation learning. In NAACL 2021, pages 3470-3479.  
+Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced LSTM for natural language inference. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1657-1668.  
+Yew Ken Chia, Lidong Bing, Soujanya Poria, and Luo Si. 2022. Relationprompt: Leveraging prompts to generate synthetic data for zero-shot relation triplet extraction. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 45-57.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP 2021, pages 6894-6910.
+
+Jiaying Gong and Hoda Eldardiry. 2021. Zero-shot relation classification from side information. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 576-585.  
+Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. Fewrel: A large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. In EMNLP 2018, pages 4803-4809.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, and 1 others. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Ting Jiang, Jian Jiao, Shaohan Huang, Zihan Zhang, Deqing Wang, Fuzhen Zhuang, Furu Wei, Haizhen Huang, Denvy Deng, and Qi Zhang. 2022. Promptbert: Improving bert sentence embeddings with prompts. In EMNLP 2022, pages 8826-8837.  
+Bosung Kim, Hayate Iso, Nikita Bhatani, Estevam Hruschka, Ndapandula Nakashole, and Tom Mitchell. 2023. Zero-shot triplet extraction by template infilling. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 272-284.  
+Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, and 1 others. 2022. Matryoshka representation learning. Advances in Neural Information Processing Systems, 35:30233-30249.  
+Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. 2017. Zero-shot relation extraction via reading comprehension. In CoNLL 2017, pages 333-342.  
+Guozheng Li, Peng Wang, and Wenjun Ke. 2023. Revisiting large language models as zero-shot relation extractors. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 6877-6892.  
+Guozheng Li, Peng Wang, Jiajun Liu, Yikai Guo, Ke Ji, Ziyu Shang, and Zijie Xu. 2024a. Meta in-context learning makes large language models better zero and few-shot relation extractors. arXiv preprint arXiv:2404.17807.  
+Haitao Li, Qingyao Ai, Jia Chen, Qian Dong, Zhijing Wu, Yiqun Liu, Chong Chen, and Qi Tian. 2024b. Blade: Enhancing black-box large language models with small domain-specific models. arXiv preprint arXiv:2403.18365.  
+Xianming Li and Jing Li. 2024. Bellm: Backward dependency enhanced large language model for sentence embeddings. In *NAACL* 2024, pages 792-804.
+
+Zehan Li, Fu Zhang, and Jingwei Cheng. 2024c. Alignre: An encoding and semantic alignment approach for zero-shot relation extraction. In *Findings of the Association for Computational Linguistics ACL* 2024, pages 2957-2966.  
+Keming Lu, I-Hung Hsu, Wenxuan Zhou, Mingyu Derek Ma, and Muhao Chen. 2022. Summarization as indirect supervision for relation extraction. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6575-6594.  
+Shengfei Lyu and Huanhuan Chen. 2021. Relation classification with entity type restriction. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 390-395.  
+Xilai Ma, Jing Li, and Min Zhang. 2023a. Chain of thought with explicit evidence reasoning for few-shot relation extraction. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 2334-2352.  
+Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2023b. Fine-tuning llama for multi-stage text retrieval. arXiv preprint arXiv:2310.08319.  
+Yubo Ma, Yixin Cao, Yong Hong, and Aixin Sun. 2023c. Large language model is not a good few-shot information extractor, but a good reranker for hard samples! In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 10572-10601.  
+Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In EMNLP 2022, pages 11048-11064.  
+Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2024. Generative representational instruction tuning. arXiv preprint arXiv:2402.09906.  
+Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, and 1 others. 2022. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005.  
+Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In EMNLP-IJCNLP 2019, pages 3982-3992.  
+Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2022. Learning to retrieve prompts for in-context learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2655-2671.  
+Oscar Sainz, Oier Lopez de Lacalle, Gorka Labaka, Ander Barrena, and Eneko Agirre. 2021. Label verbalization and entailment for effective zero and few-shot
+
+relation extraction. In EMNLP 2021, pages 1199-1212.  
+Oscar Sainz, Iker García-Ferrero, Rodrigo Agerri, Oier Lopez de Lacalle, German Rigau, and Eneko Agirre. 2024. Gollie: Annotation guidelines improve zero-shot information-extraction. In *ICLR* 2024.  
+Jacob Mitchell Springer, Suhas Kotha, Daniel Fried, Graham Neubig, and Aditi Raghunathan. 2024. Repetition improves language model embeddings. arXiv preprint arXiv:2402.15449.  
+George Stoica, Emmanouil Antonios Platanios, and Barnabás Póczos. 2021. Re-tacred: Addressing shortcomings of the tacred dataset. In Proceedings of the AAAI conference on artificial intelligence, pages 13843-13850.  
+Qi Sun, Kun Huang, Xiaocui Yang, Rong Tong, Kun Zhang, and Soujanya Poria. 2024. Consistency guided knowledge retrieval and denoising in llms for zero-shot document-level relation triplet extraction. In Proceedings of the ACM on Web Conference 2024, pages 4407-4416.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023a. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Zhen Wan, Fei Cheng, Zhuoyuan Mao, Qianying Liu, Haiyue Song, Jiwei Li, and Sadao Kurohashi. 2023. Gpt-re: In-context learning for relation extraction using large language models. In EMNLP 2023, pages 3534-3547.  
+Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368.  
+Shanchan Wu and Yifan He. 2019. Enriching pretrained language model with entity information for relation classification. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 2361-2364.  
+Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighof. 2023. C-pack: Packaged resources to advance general chinese embedding. arXiv preprint arXiv:2309.07597.  
+Kaijia Yang, Nantao Zheng, Xinyu Dai, Liang He, Shujian Huang, and Jiajun Chen. 2020. Enhance prototypical network with text descriptions for few-shot relation classification. In Proceedings of the 29th ACM
+
+International Conference on Information & Knowledge Management, pages 2273-2276.
+
+Jinsung Yoon, Sercan O Arik, Yanfei Chen, and Tomas Pfister. 2023. Search-adaptor: Text embedding customization for information retrieval. arXiv preprint arXiv:2310.08750.
+
+Kai Zhang, Bernal Jiménez Gutiérrez, and Yu Su. 2023. Aligning instruction tasks unlocks large language models as zero-shot relation extractors. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 794-812.
+
+Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D Manning. 2017. Position-aware attention and supervised data improve slot filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 35-45.
+
+Jun Zhao, Wenyu Zhan, Wayne Xin Zhao, Qi Zhang, Tao Gui, Zhongyu Wei, Junzhe Wang, Minlong Peng, and Mingming Sun. 2023a. Re-matching: A fine-grained semantic matching method for zero-shot relation extraction. In ACL 2023, pages 6680-6691.
+
+Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, and 1 others. 2023b. A survey of large language models. arXiv preprint arXiv:2303.18223.
+
+Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024. Large language models are not robust multiple choice selectors. In *ICLR* 2024.
+
+Hengyi Zheng, Rui Wen, Xi Chen, Yifan Yang, Yunyan Zhang, Ziheng Zhang, Ningyu Zhang, Bin Qin, Xu Ming, and Yefeng Zheng. 2021. Prgc: Potential relation and global correspondence based joint relational triple extraction. In ACL-IJCNLP 2021, pages 6225-6235.
+
+# A Details of Generation Augmentation
+
+We provide several examples of running Sentence Distillation and Relation Enrichment on WikiZSL in Tables 4 and 5, respectively. These examples all run on GPT-3.5-turbo-0125. It is noteworthy that for all LLMs, the JSON parsing rate can reach over  $99.2\%$ . For responses that could not be parsed, we directly use the sentences and entities from the raw data as the "relation_sentence" and "entity_type."
+
+# B Details of VanillaPrompt and QA4RE
+
+To investigate the performance of VanillaPrompt and QA4RE on Wiki-ZSL and FewRel, we follow the approach of Zhang et al. (2023), manually crafting templates for all relations. Examples of
+
+VanillaPrompt and QA4RE running on GPT-3.5-turbo-0125 are provided in Table 6. It is worth noting that the quality of manually constructed templates may affect performance, thus the results we report may contain slight errors.
+
+# C Error Analysis
+
+In Figure 8, we report the accuracy of our framework across 80 categories of relations in FewRel. We observe that, after plugging the retrieval adjuster, the retrieval performance for most relation categories improved (indicated in green). For instance, similar relations such as "P131: located in the administrative territorial entity," "P276: location," "P740: location of formation," and "P937: work location" can often be confused during retrieval. The retrieval adjuster can learn the relation patterns from seen data and effectively distinguish between similar relations. However, the retrieval adjuster might decrease the performance for a few relations (indicated in red). For example, for the relation "P400: platform," some instances might be predicted as "P710: participant" after adjustment. Optimizing the retrieval adjuster to address these challenges will be part of our future work.
+
+Additionally, we observe that our framework performs poorly on the relation "P750: distributor" because these instances are often misclassified as other relations such as "P123: publisher" or "P57: director." We suspect this happens because they are frequently mentioned in similar contexts. (e.g., in discussions about the commercialization of films or TV shows, these relation patterns might coexist). Therefore, due to this co-occurrence pattern, LLMs may easily focus on incorrect information during sentence distillation or semantic representation, leading to failures. Therefore, we believe that further enriching the semantics of relation labels and improving their distinctiveness will be the direction for future optimization.
+
+![](images/f8ef26ed20c6afacf710d550f72f694e6a45068bf7d5b117346b13054a912849.jpg)  
+Figure 8: We conduct 10 runs on the FewRel dataset with the setting of  $m = 15$ . Each relation is covered at least once in the unseen relations set  $R_{u}$ . We obtain the average accuracy across all 80 relations. Green and red represent the performance increase or decrease on each relation after adding the retrieval adjuster, respectively.
+
+<table><tr><td>Examples of Sentence Distillation</td></tr><tr><td>Prompt:Provide a summary of the relation between head entity &quot;\{Head Entity\}&quot; and tail entity &quot;\{Tail Entity\}&quot; in one sentence. Make sure that these are the only two entities in this sentence, and the head entity is in front of the tail entity. Give the entity types of the head entity and tail entity, and finally present them in JSON format.Desired format:\{&quot;relation}sentence&quot;:&quot;&quot;,&quot;head_type&quot;:&quot;&quot;,&quot;tail_type&quot;:&quot;&quot;} Text:&quot;&quot;\{Sentence\}&quot;&quot;</td></tr><tr><td>Input:{Sentence}: Robin Rand (born c. 1956) is a United States Air Force general and commander of Air Force Global Strike Command. \{Head Entity\}: Air Force Global Strike Command. \{Tail Entity\}: United States Air Force.Output:&quot;relation}sentence&quot;:&quot;Air Force Global Strike Command is commanded by a general from the United States Air Force.&quot;, &quot;head_type&quot;:&quot;Military Unit&quot;, &quot;tail_type&quot;:&quot;Armed Forces&quot;)}.</td></tr><tr><td>Input:{Sentence}: Eye to Ear II is a studio album by English guitarist , composer and improvisor Fred Frith. \{Head Entity\}: Fred Frith. \{Tail Entity\}: composer.Output:&quot;relation_sentence&quot;:&quot;Fred Frith is a composer.&quot;, &quot;head_type&quot;:&quot;Person&quot;, &quot;tail_type&quot;:&quot;Occupation)}.</td></tr><tr><td>Input:{Sentence}: Guy Rodgers from Temple University was selected before the draft as Philadelphia Warriors territorial pick. \{Head Entity\}: Guy Rodgers. \{Tail Entity\}: Philadelphia Warriors.Output:&quot;relation_sentence&quot;:&quot;Guy Rodgers was selected as a territorial pick by the Philadelphia Warriors.&quot;, &quot;head_type&quot;:&quot;Person&quot;, &quot;tail_type&quot;:&quot;Sports Team)}.</td></tr><tr><td>Input:{Sentence}: His grandfather Uma Shankar Dikshit was a part of the freedom struggle for thirty years who was closely associated with Pandit Jawahar Lal Nehru , Indira Gandhi and Rajiv Gandhi. \{Head Entity\}: Rajiv Gandhi. \{Tail Entity\}: Indira Gandhi.Output:&quot;relation_sentence&quot;:&quot;Rajiv Gandhi is the son of Indira Gandhi.&quot;, &quot;head_type&quot;:&quot;Person&quot;, &quot;tail_type&quot;:&quot;Person&quot;.</td></tr><tr><td>Input:{Sentence}: Developed by SanDisk, the Cruiser Enterprise is an encrypted USB flash drive. \{Head Entity\}: SanDisk. \{Tail Entity\}: USB flash drive.Output:&quot;relation&quot;sentence&quot;:&quot;SanDisk developed the Cruiser Enterprise.&quot;, &quot;head_type&quot;:&quot;Company&quot;, &quot;tail_type&quot;:&quot;Product&quot;).</td></tr></table>
+
+Table 4: Examples of Sentence Distillation on WikiZSL.
+
+<table><tr><td>Examples of Relation Enrichment</td></tr><tr><td>Prompt:Please provide all possible types of head entity and tail entity for the given relation, and finally present them in JSON format.Desired format: {&quot;relation&quot;:&quot;&quot;,&quot;head_types&quot;:[], &quot;tail_types&quot;:[]} Relation:&quot;】【Label): {Description}&quot;&quot;</td></tr><tr><td>Input:{Label}: country of citizenship. {Description}: the object is a country that recognizes the subject as its citizenOutput: {&quot;relation&quot;: &quot;country of citizenship&quot;, &quot;head_types&quot;:&quot;Person&quot;], &quot;tail_types&quot;:&quot;Country&quot;]}</td></tr><tr><td>Input:{Label}: director. {Description}: director(s) of film, TV-series, stageplay, video game or similarOutput: {&quot;relation&quot;: &quot;director&quot;, &quot;head_types&quot;:&quot;Film&quot;, &quot;TV-series&quot;, &quot;Stageplay&quot;, &quot;Video game&quot;], &quot;tail_types&quot;:&quot;Person&quot;]}</td></tr><tr><td>Input:{Label}: voice type. {Description}: person&#x27;s voice type. expected values: soprano, mezzo-soprano, contralto, countertenor, tenor, baritone, bass (and derivatives)Output: {&quot;relation&quot;: &quot;voice type&quot;, &quot;head_types&quot;:&quot;person&quot;], &quot;tail_types&quot;:&quot;soprano&quot;, &quot;mezzo-soprano&quot;, &quot;contralto&quot;, &quot;cun-tertenor&quot;, &quot;tenor&quot;, &quot;baritone&quot;, &quot;bass&quot;]}</td></tr><tr><td>Input:{Label}: conflict. {Description}: battles, wars or other military engagements in which the person or item participatedOutput: {&quot;relation&quot;: &quot;conflict&quot;, &quot;head_types&quot;:&quot;person&quot;, &quot;organization&quot;], &quot;tail_types&quot;:&quot;battle&quot;, &quot;war&quot;, &quot;military engage-ment&quot;]}</td></tr><tr><td>Input:{Label}: said to be the same as. {Description}: this item is said to be the same as that item, but the statement is disputedOutput: {&quot;relation&quot;: &quot;said to be the same as&quot;, &quot;head_types&quot;:&quot;Concept&quot;, &quot;Object&quot;], &quot;tail_types&quot;:&quot;Concept&quot;, &quot;Object&quot;]}</td></tr></table>
+
+Table 5: Examples of Relation Enrichment on WikiZSL.
+
+# Example of VanillaPrompt
+
+# Input:
+
+Given a sentence and two entities within the sentence, classify the relationship between the two entities based on the provided sentence. Please give the relationship only. All possible relationships are listed below:
+
+- architect  
+- main subject  
+- winner  
+- said to be the same as  
+- participant  
+-nominated for  
+- field of work  
+- place served by transport hub  
+- tributary  
+- military rank  
+-league  
+- heritage designation  
+- member of political party  
+- location of formation  
+- follows
+
+Sentence: California longhorn cattle were given as a gift to Hawaiian King Kamehameha I by British Captain George Vancouver in 1793.
+
+Entity 1: george vancouver
+
+Entity 2: captain
+
+Relationship:
+
+# Output:
+
+Entity 1: George Vancouver
+
+Entity 2: Captain
+
+Relationship: Military rank
+
+# Example of QA4RE
+
+# Input:
+
+Determine which option can be inferred from the given Sentence. You can only choose one of the following options.
+
+Sentence: California longhorn cattle were given as a gift to Hawaiian King Kamehameha I by British Captain George Vancouver in 1793.
+
+Options:
+
+A. [george vancouver] is the architect of [captain]
+
+B. [george vancouver] is the main subject of [captain]
+
+C. [george vancouver] is the winner of [captain]
+
+D. [george vancouver] is said to be the same as [captain]
+
+E. [george vancouver] is a participant in [cap
+
+F. [george vancouver] was nominated for [captain]
+
+G. [george vancouver] works in the field of [captain]
+
+H. [george vancouver] serves as the place served by [captain] transport hub
+
+I. [george vancouver] is a tributary of [captain]
+
+J. [george vancouver] holds the military rank of [captain]
+
+K. [george vancouver] is a member of the league [captain]
+
+L. [george vancouver] has the heritage designation [captain]
+
+M. [george vancouver] is a member of the political party [captain]
+
+N. [george vancouver] is the location where [captain] was formed
+
+O. [george vancouver] follows [captain]
+
+Which option can be inferred from the given sentence?
+
+Option:
+
+# Output:
+
+J
+
+Table 6: Several running examples of VanillaPrompt and QA4RE based on the descriptions in the original paper (Zhang et al., 2023).

2025/Generation-Augmented Retrieval_ Rethinking the Role of Large Language Models in Zero-Shot Relation Extraction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29f9f78fca623dc83c2344d9e8bac8d5e98a5744b0ed06307aa26359f4dd9714
+size 918356

2025/GeoChain_ Multimodal Chain-of-Thought for Geographic Reasoning/2884051f-2e1a-4d25-8d9b-40c2021e1eb9_content_list.json

@@ -0,0 +1,2140 @@
+[
+    {
+        "type": "text",
+        "text": "GeoChain: Multimodal Chain-of-Thought for Geographic Reasoning",
+        "text_level": 1,
+        "bbox": [
+            139,
+            85,
+            855,
+            108
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sahiti Yerramilli*",
+        "bbox": [
+            253,
+            118,
+            406,
+            134
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Google sahitiy@google.com",
+        "bbox": [
+            238,
+            136,
+            423,
+            168
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Nilay Pande*",
+        "bbox": [
+            606,
+            118,
+            722,
+            135
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Waymo\nnilayp@waymo.com",
+        "bbox": [
+            584,
+            136,
+            749,
+            168
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Rynaa Grover*",
+        "bbox": [
+            263,
+            187,
+            396,
+            203
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Google rynaa@google.com",
+        "bbox": [
+            248,
+            205,
+            413,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jayant Sravan Tamarapalli*",
+        "bbox": [
+            542,
+            187,
+            786,
+            204
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Google jayantsravan@google.com",
+        "bbox": [
+            549,
+            205,
+            783,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This paper introduces GeoChain, a large-scale benchmark for evaluating step-by-step geographic reasoning in multimodal large language models (MLLMs). Leveraging 1.46 million Mapillary street-level images, GeoChain pairs each image with a 21-step chain-of-thought (CoT) question sequence (over 30 million Q&A pairs). These sequences guide models from coarse attributes to fine-grained localization across four reasoning categories - visual, spatial, cultural, and precise geolocation - annotated by difficulty. Images are also enriched with semantic segmentation (150 classes) and a visual locatability score. Our benchmarking of frontier MLLMs on a diverse 2,088-image subset reveals consistent challenges: models frequently exhibit weaknesses in visual grounding, display erratic reasoning, and struggle to achieve accurate localization, especially as the reasoning complexity escalates. GeoChain offers a robust diagnostic methodology, critical for fostering significant advancements in complex geographic reasoning within MLLMs.",
+        "bbox": [
+            141,
+            291,
+            460,
+            619
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Code: GeoChain on GitHub",
+        "bbox": [
+            142,
+            625,
+            344,
+            638
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dataset: GeoChain on Hugging Face",
+        "bbox": [
+            142,
+            646,
+            403,
+            659
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            676,
+            258,
+            690
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "As large vision-language models (VLMs) continue to make rapid progress on general visual question answering and captioning tasks (Team et al., 2024; OpenAI et al., 2024; Wang et al., 2024; Dai et al., 2023), their capacity for structured geographic reasoning remains underexplored. The ability to infer a location from visual cues -such as terrain, signage, vehicles, or architecture - considered alongside spatial and cultural knowledge, is crucial for real-world applications like remote sensing, disaster response, and autonomous navigation. More broadly, geographic localization serves as a testbed",
+        "bbox": [
+            112,
+            703,
+            489,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "for grounded intelligence, requiring models to reason over subtle visual features, incorporate world knowledge, and disambiguate locations that may be visually similar. Despite this, existing benchmarks rarely probe the kind of step-by-step reasoning that such tasks demand.",
+        "bbox": [
+            507,
+            261,
+            884,
+            356
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce GeoChain, a novel multimodal benchmark for evaluating structured geographic reasoning in large language models (MLLMs). As depicted in Figure 1, each GeoChain sample features a street-level image from the Mapillary dataset (Warburg et al., 2020) paired with a 21-step chain-of-thought (CoT) question sequence. These sequences progressively guide models from coarse inferences, such as hemisphere or continent, to fine-grained predictions like city, latitude, and longitude. The complete GeoChain framework comprises 1.46 million images, each with this 21-question CoT structure, yielding over 30 million question-answer pairs. Questions span four core reasoning categories: visual cues, spatial localization, cultural inference, and precise geolocation, all annotated with difficulty levels for granular evaluation. This curriculum-style structure offers vital diagnostic insights into where and why models fail across reasoning stages, moving beyond sole reliance on final predictions.",
+        "bbox": [
+            507,
+            357,
+            884,
+            694
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To facilitate focused evaluations, we curated GeoChain Test-Mini, a diverse and challenging subset. This curation process leverages a locatability score, adapted from GeoReasoner (Li et al., 2024) and computed using features from a pretrained MaskFormer model (Cheng et al., 2021). This score quantifies the visual identifiability of a location from a single image, allowing us to stratify GeoChain Test-Mini into Easy, Medium, and Hard tiers based on thresholds in the 0.12-0.6 range. The resulting GeoChain Test-Mini contains 2,088 carefully selected images, designed to offer a representative yet manageable scale for robust MLLM assessment.",
+        "bbox": [
+            507,
+            696,
+            885,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal Contribution.",
+        "bbox": [
+            136,
+            906,
+            265,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "23624",
+        "bbox": [
+            473,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 23624-23639 November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            208,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2d06ef4dc2b4684c1e97d1229a34af48a285c3c3b819afb9af71c6d24aa189e2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            233,
+            86,
+            492,
+            225
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/68fc1d39a9796cde5af048aabfa3b4b8dd5f2abdabc23cd17e2e033a2946fad0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            505,
+            85,
+            763,
+            225
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/aa7bbda59f30d56240bac86eac7a3e39250c9b8cd0a57a86759e0bc96cf45a84.jpg",
+        "image_caption": [
+            "Figure 1: Components of a GeoChain instance: (Top-Left) Easy Mapillary Street-Level Sequences (MSLS) image with locatability score of 0.45. (Top-Right) Example chain-of-thought questions with difficulty indicators. (Bottom-Left) Derived semantic segmentation map. (Bottom-Right) Extracted key semantic labels. Together, these elements enable step-by-step diagnostic evaluation of geographic reasoning."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            233,
+            241,
+            492,
+            379
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/0f6de990bb6079f0093bf92edccf535f4fcc77a62be8a093d82813490543f895.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            505,
+            237,
+            763,
+            380
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our main contributions are as follows:",
+        "bbox": [
+            132,
+            476,
+            418,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- GeoChain Benchmark Framework: A novel benchmark for evaluating step-by-step MLLM geographic reasoning, derived from 1.46 million Mapillary street-level images and generating over 30 million Q&A pairs through 21-step chain-of-thought questions, all structured across diverse reasoning categories and difficulty levels.",
+            "- Rich Augmentation & Curated Evaluation Set: A methodology for enhancing images with semantic labels (150 classes) and a human-inspired locatability score for difficulty stratification, culminating in the GeoChain Test-Mini: a quality-controlled 2088-image evaluation set; the resulting rich semantic metadata also offers a valuable resource for broader community research and future investigations.",
+            "- Comprehensive MLLM Benchmarking & Analysis: Evaluation of leading MLLMs on GeoChain Test-Mini, providing detailed insights into their geographic reasoning capabilities, performance variations, and common failure modes."
+        ],
+        "bbox": [
+            134,
+            507,
+            489,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            476,
+            665,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Image-Based Geolocation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            508,
+            756,
+            524
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Early work in visual geolocation predominantly focused on matching query images to large, geotagged image databases, often aiming for direct coordinate prediction. For instance, Im2GPS (Hays and Efros, 2008) pioneered retrieving locations by comparing against a massive photograph dataset. Later, deep learning significantly advanced the field; PlaNet (Weyand et al., 2016) utilized convolutional neural networks (CNNs) for global location prediction, and architectures like NetVLAD (Arandjelovic et al., 2016) learned robust image representations for effective place recognition, improving upon earlier retrieval methods. Other approaches, such as those focusing on urban or cross-view settings (Tian et al., 2017), further specialized these techniques. GeoChain diverges from these paradigms, which primarily target endpoint localization accuracy or image retrieval. Instead, it introduces a structured multimodal reasoning benchmark where models must articulate a 21-step chain-of-thought (CoT) sequence of answers to geographically relevant questions, thereby enabling finer-grained diagnostic insight into their internal reasoning processes.",
+        "bbox": [
+            505,
+            535,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "23625",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Multimodal Geographic Reasoning and Benchmarks",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            470,
+            115
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "More recent efforts have begun to integrate visual understanding with language-based reasoning for complex geographic tasks. GeoReasoner (Li et al., 2024), for example, introduced a fine-tuning strategy for MLLMs using human gameplay traces, primarily to improve final location prediction by modeling human-like inference. Similarly, other recent studies (Pramanik et al., 2024; Yang et al., 2024) also concentrate on predicting precise latitude and longitude. GeoComp (Song et al., 2025) presents a large-scale dataset of geolocation gameplay data, emphasizing step-wise reasoning rooted in real human gameplay that often involves external metadata, active exploration, and dynamic information gathering. While these approaches offer valuable insights into human-like inference and gameplay dynamics, GeoChain's contribution is complementary. It does not involve model finetuning or rely on gameplay trajectories. Instead, GeoChain employs a fully static, image-grounded evaluation framework: each sample consists of a single image paired with its fixed CoT question sequence, standardized across the entire benchmark. This design facilitates direct and controlled benchmarking of different models' inherent reasoning capabilities under uniform conditions, distinct from evaluating exploratory strategies or the ability to process dynamic data.",
+        "bbox": [
+            115,
+            127,
+            490,
+            576
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Other benchmarks, such as GAEA (Campos et al., 2025), generate diverse conversational questions from detailed, place-specific metadata like OpenStreetMap attributes. While this can create rich contextual queries, it introduces challenges related to the temporal stability of dynamic data (e.g., changes in urban landscape) and complicates fair, apples-to-apples model comparisons due to nonuniform question sets. Consequently, disentangling model reasoning failures from idiosyncratic question characteristics becomes difficult. GeoChain mitigates these issues by grounding its standardized questions in more enduring visual semantics such as the presence of characteristic vegetation, architectural styles, or road infrastructure, often identifiable through image segmentation and stable general geographic facts. This focus ensures the evaluation centers on the consistency of the reasoning process itself.",
+        "bbox": [
+            115,
+            581,
+            489,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, existing geospatial benchmarks like GEO-Bench (Lacoste et al., 2023) primar",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ily target remote sensing applications, offering valuable tools for Earth monitoring with satellite imagery and tasks such as classification or segmentation. In contrast, GeoChain specifically addresses agent-level geographic reasoning from high-resolution, ground-level imagery, emphasizing natural-language understanding of spatial, cultural, and visual cues directly perceivable in such environments.",
+        "bbox": [
+            507,
+            84,
+            885,
+            229
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Mapillary Street-Level Sequences Dataset",
+        "text_level": 1,
+        "bbox": [
+            507,
+            239,
+            880,
+            255
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "GeoChain is built upon the Mapillary Street-Level Sequences (MSLS) dataset (Warburg et al., 2020), a large-scale, crowd-sourced collection of diverse, geo-tagged street-level images. MSLS's global coverage, with data from numerous cities worldwide reflecting the breadth of the MSLS ecosystem, and its varied capture conditions (diverse cameras, viewpoints, seasons, times of day) make it an ideal foundation for a benchmark aiming to evaluate generalizable geographic reasoning.",
+        "bbox": [
+            507,
+            259,
+            885,
+            420
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 GeoChain Benchmark Construction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            432,
+            855,
+            447
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The GeoChain benchmark is constructed by augmenting the Mapillary Street-Level Sequences (MSLS) dataset (Warburg et al., 2020). MSLS provides a diverse collection of geo-tagged street-level imagery (approximately 1.4 million images in its full extent, with a geographical distribution across numerous cities as illustrated in Figure 2), crucial for developing and evaluating geographic localization models. However, to facilitate fine-grained, step-by-step reasoning, we introduce several layers of annotation and metadata. Our contributions enhance the MSLS dataset in three primary ways: semantic class labeling, locatability score computation, and the design of a structured chain-of-thought question battery. These augmentations, followed by a careful test set curation process, collectively enable a more nuanced evaluation of multimodal models' geographic reasoning capabilities.",
+        "bbox": [
+            507,
+            456,
+            885,
+            746
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Semantic Class Labeling",
+        "text_level": 1,
+        "bbox": [
+            507,
+            755,
+            749,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To ground visual reasoning in explicit semantic content, each image in our benchmark is augmented with a semantic segmentation map. This map provides a detailed understanding of the scene's composition by identifying various objects and environmental features. We employed MaskFormer (Cheng et al., 2021), a state-of-the-art transformer-based architecture for semantic segmentation. Specifically, we utilized a MaskFormer",
+        "bbox": [
+            507,
+            776,
+            885,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "23626",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/0d635365e3106fe931ee375a416398a9d6c24288bb7a999da4d34e0577fac860.jpg",
+        "image_caption": [
+            "Figure 2: Count of images per city, illustrating the city distribution within the GeoChain dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            157,
+            84,
+            842,
+            350
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "model pre-trained on the ADE20K dataset (Zhou et al., 2017), which offers a rich label set of 150 distinct classes, encompassing a wide array of objects, environmental elements (e.g., \"tree\", \"sky\", \"road\"), and architectural features (e.g., \"building\", \"window\", \"door\").",
+        "bbox": [
+            110,
+            403,
+            489,
+            499
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "From the segmentation map, we calculate how much of the image is covered by each category. We do this by working out the percentage of the image's total area that each specific category takes up. For example, we might find that 'sky' covers  $30\\%$  of an image, and 'road' covers  $15\\%$ . This measurement of what's in the scene, and how much of it there is, then helps us create the correct answers for many of the visual questions in our benchmark.",
+        "bbox": [
+            112,
+            501,
+            489,
+            645
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Locatability Score Computation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            658,
+            413,
+            673
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To systematically assess model performance across varying levels of visual ambiguity, we compute a locatability score for every image considered for the GeoChain benchmark. This score, ranging from 0 to 1, quantifies how visually identifiable a location is likely to be, with higher scores indicating more distinct and easily locatable scenes. Our methodology for calculating this score is adopted from (Li et al., 2024). The distribution of these computed locatability scores across the considered images is shown in Figure 3.",
+        "bbox": [
+            112,
+            678,
+            489,
+            854
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The core idea behind this score is to leverage common visual cues that humans, particularly proficient GeoGuessr players (GeoGuessr, 2013), rely on for geolocation. The process involves sev",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/e6ea7e9d9887b835a698d72f64c1981ce88f037613469da71c30be8d5cc2ed0e.jpg",
+        "image_caption": [
+            "Figure 3: Distribution of Locatability Scores in GeoChain"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            401,
+            878,
+            554
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "eral steps:",
+        "bbox": [
+            507,
+            619,
+            589,
+            634
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Identification of Cues: A set of cues frequently used by GeoGuessr players (e.g., \"houses in central Chile are more likely to have terracotta tiled roofs\") is established.",
+            "2. Cue-to-Class Similarity: The semantic similarity between these cues and the 150 class labels produced by the MaskFormer model (as described in Section 3.1) is computed. This typically involves using text embeddings to represent both the cues and the class labels, followed by a similarity measure (e.g., cosine similarity).",
+            "3. Class Weight Derivation: The similarities are aggregated across all cues for each class and then subjected to min-max normalization to derive a set of weights  $w_{c}$  for each class c. These weights reflect the importance of each visual class for"
+        ],
+        "bbox": [
+            504,
+            648,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "23627",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "geolocation.",
+        "bbox": [
+            131,
+            85,
+            253,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4. Weighted Score Aggregation: The final locatability score for an image is computed as a weighted sum of the percentage areas of the classes present in the image.",
+        "bbox": [
+            109,
+            107,
+            489,
+            172
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This locatability score is then used to stratify the images within the GeoChain test set into three distinct tiers: Hard if score  $\\in$  [0.12, 0.22), Medium if score  $\\in$  [0.22, 0.45), and Easy if score  $\\in$  [0.45, 0.6). This stratification, based purely on visual cues inherent in the imagery, allows for a more granular analysis of model performance and helps identify specific weaknesses in reasoning about visually challenging environments.",
+        "bbox": [
+            112,
+            184,
+            490,
+            329
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Chain-of-Thought Question Design",
+        "text_level": 1,
+        "bbox": [
+            112,
+            340,
+            436,
+            357
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A central component of GeoChain is a carefully designed sequence of 21 questions that guide the model through a step-by-step reasoning process, from coarse-grained observations to fine-grained localization. This chain-of-thought (CoT) approach aims to mimic a structured human-like deduction process. The questions are ordered such that earlier questions elicit information or focus attention on attributes that can be instrumental in answering subsequent, more complex questions.",
+        "bbox": [
+            112,
+            361,
+            487,
+            521
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The full list of 21 questions, along with their rank, assigned difficulty (Easy, Medium, Hard), question type (Binary, Multiclass, Free-text), and question category (e.g., Culture/Infrastructure, Geo Localization, Terrain/Environment), is provided in Appendix A.1.1. The difficulty annotation (Easy, Medium, Hard) for each question reflects the anticipated challenge of answering that specific question in isolation, based on the type of information required.",
+        "bbox": [
+            112,
+            523,
+            489,
+            682
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The question set is designed to be static across all data points in the benchmark. This uniformity ensures a consistent evaluation framework, allowing for direct, apples-to-apples comparisons of different models' reasoning capabilities. The questions cover diverse aspects:",
+        "bbox": [
+            112,
+            684,
+            489,
+            780
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- Visual Object/Attribute Presence: Some questions directly query the presence of specific objects or attributes identifiable from the image (e.g., \"Do you see any boats or ships?\"). Ground truth for these questions is primarily derived from the semantic class labels extracted via the MaskFormer model (Section 3.1). For instance, if the class \"boat\" occupies a non-zero percent",
+        "bbox": [
+            115,
+            791,
+            490,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "age of the image, the answer would be affirmative.",
+        "bbox": [
+            524,
+            84,
+            884,
+            115
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- Inferential and Contextual Knowledge: Other questions require more derivative reasoning or contextual knowledge beyond direct object identification (e.g., \"Is this place near a coast?\", \"What side of the road do vehicles drive on here?\"). The MSLS dataset encompasses images from 24 distinct cities globally. For images originating from these locations, we manually curated ground truth answers for city-level attributes or environmental characteristics (e.g., predominant architectural styles, typical climate indicators) that apply broadly to the image's geographic area.",
+        "bbox": [
+            510,
+            122,
+            885,
+            331
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- Progressive Localization: The sequence progresses from general observations (e.g., hemisphere, continent) to specific details (e.g., country, city, precise latitude and longitude coordinates).",
+        "bbox": [
+            510,
+            337,
+            885,
+            416
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The question types include binary (Yes/No), multiclass (selection from a predefined set of options), and free-text (open-ended answers, such as country name or coordinates). This variety tests different aspects of a model's understanding and generation capabilities.",
+        "bbox": [
+            507,
+            424,
+            882,
+            519
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The semantic segmentation labels generated in Section 3.1 were instrumental in constructing several questions that directly probe the visual understanding capabilities of the models. Beyond their use in the current benchmark, this rich semantic metadata, now part of GeoChain, offers a valuable resource for the community. It can be leveraged to design new questions aimed at further investigating specific aspects of model behavior, such as tendencies towards visual hallucination (Li et al., 2023) (Rohrbach et al., 2018) or the fine-grained ability to identify a wider array of objects. The insights derived from such extended evaluations can subsequently guide targeted improvements in model development.",
+        "bbox": [
+            507,
+            521,
+            882,
+            760
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By analyzing model performance across this structured chain of questions, GeoChain aims to provide deeper insights into the strengths and weaknesses of multimodal geographic reasoning systems.",
+        "bbox": [
+            507,
+            762,
+            884,
+            841
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Test Set Curation and Sampling Strategy",
+        "text_level": 1,
+        "bbox": [
+            507,
+            852,
+            877,
+            868
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To create the \"GeoChain Test-Mini\" subset for focused evaluation, we prioritized stratification, visual quality, and diversity. We initially targeted",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "23628",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2100 images, stratified by locatability scores into 700 Easy, 700 Medium, and 700 Hard examples. A tiered, unique-sequence sampling strategy was employed: unique image sequences were randomly sampled first for the Hard tier, then for the Medium tier (from remaining unique sequences), and finally for the Easy tier, ensuring no sequence was reused across tiers. The underlying MSLS dataset exhibits a notable skew in its per-city image distribution (as highlighted by the overall dataset statistics in Figure 2). Consequently, to avoid introducing new biases that could arise from attempting to manually balance city representation or 'carefully' over/under-sample from specific locations, our approach was to randomly sample unique image sequences across all available cities within each defined locatability tier. These 2100 candidates underwent manual visual inspection, where 12 images with critical quality issues (e.g., excessive blur, poor exposure) were removed. This rigorous curation yielded a final Test-Mini set of 2088 high-quality, diverse, and appropriately challenging images.",
+        "bbox": [
+            115,
+            82,
+            490,
+            451
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            464,
+            221,
+            481
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we evaluate the performance of frontier vision-language models: GPT-4.1, GPT-4.1-mini (OpenAI et al., 2024), Claude 3.7 Sonnet (Sonnet, 2025), Gemini 2.5 Flash (Google, 2025a), Gemini 2.5 Pro (Google, 2025b), Qwen 2.5 VL (Bai et al., 2025), Pixtral Large (Mistral AI, 2024), and Gemma 3 (Team et al., 2025) on the GeoChain \"Test-Mini\" benchmark, focusing on their ability to reason accurately and consistently across a structured 21-step geographic reasoning chain.",
+        "bbox": [
+            112,
+            489,
+            489,
+            651
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            661,
+            310,
+            675
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1.1 Haversine Distance",
+        "text_level": 1,
+        "bbox": [
+            112,
+            682,
+            326,
+            696
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The final question in each GeoChain sequence (Question 21) requires the model to predict the geographic coordinates (latitude, longitude) of the depicted scene. To evaluate the accuracy of these specific predictions, we compute the Haversine distance: the shortest distance over the Earth's surface between the predicted and ground-truth coordinates, assuming a spherical Earth. A detailed explanation of the Haversine distance calculation is provided in Section A.2.",
+        "bbox": [
+            112,
+            701,
+            489,
+            860
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1.2 Pass Score",
+        "text_level": 1,
+        "bbox": [
+            112,
+            870,
+            258,
+            883
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The Pass Score is computed as the average fraction of questions correctly answered across the full 21-",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "step reasoning chain for each image. A prediction for any question is considered correct if it matches the ground-truth answer for that specific question, accounting for its type (e.g., exact match for fre-text, class match for multiclass, or binary match). Crucially, for the final latitude and longitude prediction (Question 21), a response is deemed correct contributing to the Pass Score if its Haversine distance (as defined in Section 4.1.1) from the ground truth is less than  $50\\mathrm{km}$ .",
+        "bbox": [
+            507,
+            84,
+            885,
+            244
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Overall Model Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            256,
+            774,
+            269
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Overall model performance (Table 1) offers nuanced insights into current MLLM geographic reasoning. The leading Gemini models exhibit specialized strengths: Gemini-2.5-pro excels in complex multi-step reasoning (pass score  $81.84\\%$ ), whereas Gemini-2.5-Flash achieves superior localization precision ( $445.24\\mathrm{km}$  mean error), hinting at differing architectural or training optimizations. This divergence underscores that broad inferential ability and precise geolocalization are distinct skills, likely requiring separate optimization pathways rather than being monolithic capabilities. GPT-4.1 maintains a competitive position; however, the substantial localization inaccuracies of the other models underscore that robust geospatial grounding is a significant developmental hurdle, indicating a key area for advancement in MLLM capabilities.",
+        "bbox": [
+            507,
+            277,
+            885,
+            551
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The introduction of threshold-based localization accuracies - at City  $(< 25\\mathrm{km})$ , Region  $(< 200\\mathrm{km})$ , and Country  $(< 750\\mathrm{km})$  levels further refines this performance landscape. Gemini-2.5-pro's superior performance is reinforced by its top-tier City-level precision  $(59.38\\%)$ . Complementing this, Gemini-2.5-Flash excels in broader accuracy, leading at both Region-level  $(70.02\\%)$  and Country-level  $(90.31\\%)$ . GPT-4.1 also demonstrates notable strength in City-level performance  $(57.84\\%)$ , surpassing Gemini-2.5-Flash in this specific high-precision context. The challenges in Claude 3.7 Sonnet, GPT 4.1 Mini, and other open source models are starkly emphasized by their profound difficulties at these finer scales. These granular metrics effectively highlight that achieving reliable, high-confidence city-level precision is a primary differentiator and a significant challenge across the evaluated MLLMs.",
+        "bbox": [
+            507,
+            551,
+            885,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To provide a task-specific point of comparison, we also include GeoReasoner (Li et al., 2024), a contemporary VLM-based geolocation model. Since it only predicts country and city, we evaluate",
+        "bbox": [
+            507,
+            857,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "23629",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/1cbd7820fd9763cca63c079270cc0845d62a7dff36d29ebfecf81ab7f9bdf382.jpg",
+        "table_caption": [
+            "Table 1: Overall model-level accuracy and localization metrics."
+        ],
+        "table_footnote": [
+            "*Distance is an approximation calculated from the predicted city's center."
+        ],
+        "table_body": "<table><tr><td>Model</td><td>Pass Score (%)</td><td>Mean Dist (km)</td><td>&lt; 25 km (%)</td><td>&lt; 200 km (%)</td><td>&lt; 750 km (%)</td></tr><tr><td>Gemini-2.5-pro</td><td>81.84</td><td>489.51</td><td>59.38</td><td>69.95</td><td>88.51</td></tr><tr><td>Gemini-2.5-Flash</td><td>79.77</td><td>445.24</td><td>55.71</td><td>70.02</td><td>90.31</td></tr><tr><td>GPT-4.1</td><td>79.25</td><td>611.89</td><td>57.84</td><td>67.36</td><td>86.24</td></tr><tr><td>Claude 3.7 Sonnet</td><td>76.23</td><td>1289.04</td><td>40.34</td><td>47.07</td><td>73.31</td></tr><tr><td>GPT-4.1 Mini</td><td>70.42</td><td>1194.77</td><td>48.61</td><td>52.87</td><td>72.77</td></tr><tr><td>Qwen 2.5 VL (72B)</td><td>74.6</td><td>1067.0</td><td>41.7</td><td>49.0</td><td>75.0</td></tr><tr><td>Pixtral Large (124B)</td><td>68.1</td><td>1869.8</td><td>27.9</td><td>34.2</td><td>59.8</td></tr><tr><td>Gemma 3 (27B)</td><td>59.4</td><td>1235.5</td><td>34.4</td><td>43.9</td><td>72.0</td></tr><tr><td>GeoReasoner</td><td>—</td><td>1139.57*</td><td>4.76</td><td>—</td><td>39.49</td></tr></table>",
+        "bbox": [
+            114,
+            107,
+            885,
+            284
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "it on the relevant metrics. Despite its task-specific design, its city-level accuracy is substantially lower than that of the top general-purpose models, and its mean distance error is high. This result reinforces the profound difficulty of the geographic reasoning task and validates the need for a granular, step-by-step benchmark like GeoChain to diagnose model failures.",
+        "bbox": [
+            110,
+            319,
+            489,
+            448
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Breakdown by Image Difficulty",
+        "text_level": 1,
+        "bbox": [
+            112,
+            494,
+            406,
+            510
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Analyzing model performance by image difficulty (Table 2) reveals critical operational characteristics. As expected, 'Hard' images significantly challenge all models, leading to substantial increases in mean localization errors often exceeding  $1000 - 2000\\mathrm{km}$  for several models. The Gemini models consistently lead: Gemini-2.5-pro achieves top Pass Scores across all difficulties (e.g.,  $78.0\\%$  on Hard), while Gemini-2.5-Flash generally provides superior localization on 'Easy' and 'Medium' images (e.g.,  $188.45\\mathrm{km}$  on Medium). Notably, Gemini-2.5-pro performs the best for localization precision on 'Hard' images  $(866.62\\mathrm{km})$ , possibly where its stronger inferential capacity becomes decisive. An intriguing anomaly is the better localization by some models, like Gemini-2.5-Flash, on 'Medium' versus 'Easy' images, potentially due to bias towards certain cities in pre-training data. Furthermore, Claude 3.7 Sonnet's performance is particularly interesting: despite reasonable Pass Scores (e.g.,  $73.2\\%$  on Hard), its poor localization  $(2000.14\\mathrm{km}$  on Hard) highlights a profound disconnect between understanding cues and grounding them spatially.",
+        "bbox": [
+            110,
+            533,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Breakdown by Question Category",
+        "text_level": 1,
+        "bbox": [
+            507,
+            319,
+            821,
+            336
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Analyzing Pass Scores by question category (Table 3), informed by the benchmark's diverse question structures (e.g., visual queries versus free-text specific knowledge), reveals distinct performance strata. Foundational \"Visual\" questions, focusing on direct object presence (e.g., \"Do you see any boats?\"), yield universally high scores (all models  $>82\\%$ ), suggesting robust basic visual grounding and low immediate hallucination, with Claude 3.7 Sonnet leading  $(92.8\\%)$ . Similarly, \"Terrain\" identification is generally strong. In contrast, categories like \"Geo Localization\" and \"Cultural\" show mixed results; models likely handle simpler, coarse queries (e.g., continent identification) better than challenging free-text questions requiring specific knowledge (e.g., city/state names, language identification). Unsurprisingly, \"Exact Loc\" demanding precise latitude/longitude output—is definitively the most challenging category across all models. Within this landscape, Gemini-2.5-pro consistently excels, particularly in the more demanding categories like \"Terrain\"  $(87.4\\%)$ , \"Cultural\"  $(77.9\\%)$ , and \"Exact Loc.\"  $(63.5\\%)$ . GPT-4.1 also demonstrates strong performance, notably in \"Geo Localization\"  $(76.9\\%)$  and \"Exact Loc.\"  $(61.5\\%)$ . Claude 3.7 Sonnet's profile, with its excellent \"Visual\" scores but significantly weaker \"Exact Loc.\" performance  $(51.0\\%)$ , starkly illustrates a common theme: a disconnect between initial cue processing and final, precise geospatial grounding, which remains the primary MLLM hurdle.",
+        "bbox": [
+            505,
+            341,
+            885,
+            843
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4.1 Accuracy vs. Reasoning Depth",
+        "text_level": 1,
+        "bbox": [
+            507,
+            852,
+            808,
+            868
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 reveals a typical degradation pattern: All models perform well in the initial questions (1-9), which ask about visual or global cues such as ve",
+        "bbox": [
+            507,
+            873,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "23630",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/a05c49ae183905be43900b25165922e73fb6ec115519a28b5dff3d42947b946e.jpg",
+        "table_caption": [
+            "Table 2: Performance by image difficulty. Accuracy  $(\\%)$  and Haversine distance (km) for each difficulty level."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Diff</td><td>Pass Score</td><td>M. Dist.</td></tr><tr><td>Claude 3.7</td><td>Easy</td><td>77.2</td><td>885.86</td></tr><tr><td>Sonnet</td><td>Medium</td><td>78.3</td><td>989.13</td></tr><tr><td></td><td>Hard</td><td>73.2</td><td>2000.14</td></tr><tr><td>GPT-4.1</td><td>Easy</td><td>70.8</td><td>863.19</td></tr><tr><td>Mini</td><td>Medium</td><td>73.2</td><td>827.78</td></tr><tr><td></td><td>Hard</td><td>67.3</td><td>1910.44</td></tr><tr><td>GPT-4.1</td><td>Easy</td><td>79.3</td><td>357.36</td></tr><tr><td></td><td>Medium</td><td>81.6</td><td>428.46</td></tr><tr><td></td><td>Hard</td><td>76.8</td><td>1052.13</td></tr><tr><td>Gemini-2.5</td><td>Easy</td><td>80.5</td><td>287.61</td></tr><tr><td>Flash</td><td>Medium</td><td>82.5</td><td>188.45</td></tr><tr><td></td><td>Hard</td><td>76.3</td><td>873.78</td></tr><tr><td>Gemini-2.5</td><td>Easy</td><td>83.3</td><td>300.29</td></tr><tr><td>Pro</td><td>Medium</td><td>84.2</td><td>304.32</td></tr><tr><td></td><td>Hard</td><td>78.0</td><td>866.62</td></tr><tr><td>Qwen 2.5</td><td>Easy</td><td>74.7</td><td>739.49</td></tr><tr><td>VL</td><td>Medium</td><td>77.1</td><td>853.91</td></tr><tr><td></td><td>Hard</td><td>71.8</td><td>1636.54</td></tr><tr><td>Pixtral</td><td>Easy</td><td>67.9</td><td>1514.53</td></tr><tr><td>Large</td><td>Medium</td><td>70.1</td><td>1609.17</td></tr><tr><td></td><td>Hard</td><td>66.1</td><td>2491.22</td></tr><tr><td>Gemma 3</td><td>Easy</td><td>60.3</td><td>889.06</td></tr><tr><td></td><td>Medium</td><td>61.1</td><td>1056.79</td></tr><tr><td></td><td>Hard</td><td>56.6</td><td>1766.11</td></tr></table>",
+        "bbox": [
+            139,
+            122,
+            460,
+            587
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "hicles, hemisphere, or continent. These are relatively easy to infer on the basis of surface-level features. As the questions become more complex and semantically demanding, the accuracy drops sharply, especially at questions 10 and 17. These questions requiring nuanced interpretation of environmental and infrastructure signals. In particular, we observe a performance bump around questions 12-14. Despite appearing later in the sequence, these questions ask about relatively easy visual features (e.g., desert, hills, or city size). This reinforces the value of structuring questions not just by logical sequence but also by measured difficulty, allowing finer-grained diagnostics of model capability.",
+        "bbox": [
+            112,
+            614,
+            489,
+            853
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The final steps of the chain (questions 18-21) see the steepest drop in performance, as models are asked to predict language, administrative region, city name, and exact coordinates - tasks that require",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/ae66a47fa3ea439ba9701336d3d4d3d43431bc624dda21ca0c77c3d9c9858cb6.jpg",
+        "image_caption": [
+            "Figure 4: Average pass score across the 21-step Geochain reasoning chain"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            82,
+            880,
+            230
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "multi-modal reasoning, robust world knowledge, and low-level visual grounding. This progressive breakdown highlights GeoChain's utility as a diagnostic benchmark. By tracking model accuracy at each reasoning step, researchers can understand how performance degrades under deeper spatial inference chains.",
+        "bbox": [
+            507,
+            298,
+            884,
+            411
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Fine-Tuning Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            426,
+            702,
+            443
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our primary analysis revealed a critical failure point in state-of-the-art MLLMs: their inability to consistently link visual cues to a final geographic location through sequential reasoning. This finding raised a pivotal question: is this a fundamental limitation of current models, or is it a tractable skill that can be taught?",
+        "bbox": [
+            507,
+            451,
+            882,
+            563
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To address this, we conducted a targeted fine-tuning experiment with a clear objective: to determine if training on the GeoChain framework could teach a model to forge this connection between visual evidence and final localization - the very skill our initial analysis found lacking - and then generalize this ability to entirely unseen cities. For this study, we fine-tuned the open-source Qwen-VL (9.6B) (Bai et al., 2023) model using Low-Rank Adaptation (LoRA) for 5 epochs on a modest set of 900 samples from five geographically diverse cities (Moscow, Berlin, Bangkok, Sao Paulo, and Phoenix). We then evaluated on a held-out set of 850 samples from five entirely different, unseen cities (Budapest, Ottawa, Helsinki, Melbourne, and Amsterdam).",
+        "bbox": [
+            507,
+            565,
+            882,
+            821
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The results of this brief intervention are striking. As shown in Table 4, fine-tuning boosted the overall pass score by over 20 percentage points (46.8% to 68.2%) and nearly halved the mean localization error, demonstrating a significant improvement in final prediction accuracy. Fine-tuning",
+        "bbox": [
+            507,
+            824,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "23631",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/62d5e15e3fa1732da92887e281fc61e16a52b098632d10d12fe1e2785af9afad.jpg",
+        "table_caption": [
+            "Table 3: Pass score  $(\\%)$  by question category."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Visual</td><td>Terrain</td><td>Geo Localization</td><td>Cultural</td><td>Exact Loc.</td></tr><tr><td>Claude 3.7 Sonnet</td><td>92.8</td><td>84.7</td><td>69.4</td><td>67.4</td><td>51.0</td></tr><tr><td>GPT-4.1 Mini</td><td>92.3</td><td>78.7</td><td>64.1</td><td>56.8</td><td>40.7</td></tr><tr><td>GPT-4.1</td><td>91.8</td><td>84.8</td><td>76.9</td><td>68.3</td><td>61.5</td></tr><tr><td>Gemini-2.5-Flash</td><td>92.4</td><td>86.0</td><td>73.5</td><td>75.3</td><td>59.8</td></tr><tr><td>Gemini-2.5-pro</td><td>92.1</td><td>87.4</td><td>76.8</td><td>77.9</td><td>63.5</td></tr><tr><td>Qwen 2.5 VL</td><td>92.3</td><td>81.6</td><td>70.8</td><td>62.0</td><td>45.0</td></tr><tr><td>Pixtral Large</td><td>91.0</td><td>78.7</td><td>61.1</td><td>53.9</td><td>28.2</td></tr><tr><td>Gemma 3</td><td>82.3</td><td>70.7</td><td>55.2</td><td>35.0</td><td>37.6</td></tr></table>",
+        "bbox": [
+            173,
+            107,
+            826,
+            271
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/31c8817be6e8af7c4452a5d55cfd62482285dd7047ac7fe43f8b0e22913c4219.jpg",
+        "table_caption": [
+            "Table 4: Overall performance comparison before and after fine-tuning."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Pass Score (%)</td><td>Mean Dist. (km)</td><td>&lt;25 km (%)</td><td>&lt;200 km (%)</td><td>&lt;750 km (%)</td></tr><tr><td>Base</td><td>46.8</td><td>5866.7</td><td>0.15</td><td>1.0</td><td>18.7</td></tr><tr><td>Finetuned</td><td>68.2</td><td>2963.4</td><td>6.3</td><td>15.5</td><td>40.5</td></tr></table>",
+        "bbox": [
+            166,
+            307,
+            831,
+            373
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/31bb0f2f37d11fccde6f7dd733880065c8e7eaab878863c90eced86d4af82e05.jpg",
+        "table_caption": [
+            "Table 5: City-level localization accuracy (<25 km) by image difficulty."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Easy</td><td>Medium</td><td>Hard</td></tr><tr><td>Base</td><td>0.31%</td><td>0.00%</td><td>0.00%</td></tr><tr><td>Finetuned</td><td>4.8%</td><td>7.2%</td><td>7.5%</td></tr></table>",
+        "bbox": [
+            141,
+            432,
+            460,
+            499
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/3cf2bcd212d3be0f180ac7f4ddfeb2f4f21183e0b30e635ce47ffad79f2be11b.jpg",
+        "table_caption": [
+            "Table 6: Pass score  $(\\%)$  by question difficulty."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Easy</td><td>Medium</td><td>Hard</td></tr><tr><td>Base</td><td>66.5%</td><td>40.5%</td><td>2.1%</td></tr><tr><td>Finetuned</td><td>91.1%</td><td>63.0%</td><td>12.9%</td></tr></table>",
+        "bbox": [
+            139,
+            537,
+            463,
+            602
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "proved particularly effective at addressing the base model's key weaknesses:",
+        "bbox": [
+            112,
+            627,
+            487,
+            659
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Geospatial Grounding: Localization accuracy on \"Hard\" images, a primary challenge for the base model, jumped from  $0.00\\%$  to  $7.5\\%$  (Table 5), indicating that the model learned to connect visual evidence to specific geographic contexts.",
+            "- Complex Reasoning: The model's capacity for nuanced inference saw the most dramatic gains. The pass score on \"Hard\" questions increased more than six-fold, from  $2.1\\%$  to  $12.9\\%$  (Table 6)."
+        ],
+        "bbox": [
+            136,
+            671,
+            487,
+            859
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "These findings, achieved with limited data, serve a crucial diagnostic purpose. The fact that targeted training on GeoChain so effectively remedies the",
+        "bbox": [
+            112,
+            873,
+            487,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "specific deficiencies it identifies poor grounding and brittle multi-step inference validates its utility beyond mere evaluation. It establishes GeoChain as a fine-grained diagnostic framework that provides a clear and actionable roadmap for targeted model improvement.",
+        "bbox": [
+            507,
+            395,
+            882,
+            492
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            504,
+            640,
+            519
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This paper introduced GeoChain, a large-scale, chain-of-thought benchmark designed to dissect multimodal geographic reasoning in MLLMs using street-level imagery and a 21-step diagnostic framework. Our evaluations on the curated GeoChain Test-Mini subset reveal that even leading MLLMs exhibit significant deficiencies in visual grounding, reasoning consistency, and localization accuracy, particularly as task and visual complexity escalate. By enabling a granular, step-by-step analysis, GeoChain moves beyond simple end-task accuracy to pinpoint these critical failure modes, thereby providing an essential diagnostic resource and methodology. We anticipate that GeoChain will steer future research towards developing more robust, geographically aware, and reliable AI systems capable of nuanced real-world understanding.",
+        "bbox": [
+            507,
+            530,
+            884,
+            803
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            815,
+            645,
+            829
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While GeoChain offers a novel diagnostic approach, we acknowledge several limitations. GeoChain is built upon the Mapillary Street-Level Sequences training split; consequently, while our chain-of-thought reasoning framework and the",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "23632",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "overall task are novel, there is a potential that MLLMs have encountered these specific visual scenes or highly similar ones during their extensive pre-training. Evaluating performance on truly \"unseen\" street-level imagery is an inherent challenge for the field, given the ubiquity of data from sources like Google Street View (Anguelov et al., 2010) and OpenStreetMap (Haklay and Weber, 2008), meaning that performance assessments may partly reflect familiarity with certain visual data rather than solely generalization to entirely new scenes. Additionally, the underlying geographical distribution of the images, though diverse, retains some skew, potentially affecting the generalizability of the findings in all urban contexts. Furthermore, our locatability score's precision is contingent upon the accuracy of an upstream semantic segmentation model, which could introduce noise into the difficulty stratification.",
+        "bbox": [
+            112,
+            84,
+            490,
+            391
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "7 Usage of Generative AI tools",
+        "text_level": 1,
+        "bbox": [
+            112,
+            401,
+            394,
+            418
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We utilized Generative AI tools to help improve the language, phrasing, and readability of this manuscript.",
+        "bbox": [
+            112,
+            426,
+            487,
+            475
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            502,
+            213,
+            517
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Dragomir Anguelov, Carole Dulong, Daniel Filip, Christian Frueh, Stéphane Lafon, Richard Lyon, Abhijit Ogale, Luc Vincent, and Josh Weaver. 2010. Google street view: Capturing the world at street level. Computer, 43(6):32-38.",
+            "Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. 2016. NetVLAD: CNN architecture for weakly supervised place recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5287-5297.",
+            "Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. Preprint, arXiv:2308.12966.",
+            "Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. Preprint, arXiv:2502.13923.",
+            "Ron Campos, Ashmal Vayani, Parth Parag Kulkarni, Rohit Gupta, Aritra Dutta, and Mubarak Shah. 2025. GAEA: A geolocation aware conversational model. Preprint, arXiv:2503.16423."
+        ],
+        "bbox": [
+            115,
+            524,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yuchen Cao, Nilay Pande, Ayush Jain, Shikhar Sharma, Gabriel Sarch, Nikolaos Gkanatsios, Xian Zhou, and Katerina Fragkiadaki. Embodied symbiotic assistants that see, act, infer and chat.",
+            "Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. 2021. Per-pixel classification is not all you need for semantic segmentation. In Advances in Neural Information Processing Systems 34 (NeurIPS 2021), pages 17864-17875. Curran Associates, Inc.",
+            "Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. Instructlip: Towards general-purpose vision-language models with instruction tuning. Preprint, arXiv:2305.06500.",
+            "GeoGuessr. 2013. Geoguessr - let's explore the world! Accessed: [Date you accessed the website].",
+            "Google. 2025a. Gemini 2.5 flash is now in preview. https://blog.google/products/gemini/gemini-2-5-flash-preview/. Accessed 2025-05-18.",
+            "Google. 2025b. Gemini 2.5 pro model. https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/gemini-2-5-thinking. Accessed through the Gemini API on [Date of access] or Vertex AI.",
+            "Muki Haklay and Patrick Weber. 2008. Openstreetmap: User-generated street maps. IEEE Pervasive Computing, 7(4):12-18.",
+            "James Hays and Alexei A. Efros. 2008. IM2GPS: estimating geographic information from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE.",
+            "Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan David Sherwin, Hannah Kerner, Björn Lütjens, Jeremy Andrew Irvin, David Dao, Hamed Alemohammad, Alexandre Drouin, Mehmet Gunturkun, Gabriel Huang, David Vazquez, Dava Newman, Yoshua Bengio, Stefano Ermon, and Xiao Xiang Zhu. 2023. Geo-bench: Toward foundation models for earth monitoring. Preprint, arXiv:2306.03831.",
+            "Ling Li, Yu Ye, Bingchuan Jiang, and Wei Zeng. 2024. Georeasoner: Geo-localization with reasoning in street views using a large vision-language model. In International Conference on Machine Learning (ICML).",
+            "Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Xin Zhao, and Ji-Rong Wen. 2023. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing.",
+            "Mistral AI. 2024. Introducing Pixtral Large. https://mistral.ai/news/pixtral-large. Accessed: 07/10/2025."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "23633",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, and 262 others. 2024. Gpt-4 technical report. Preprint, arXiv:2303.08774.",
+            "Nilay Pande and Suyash P. Awate. 2021. Generative deep-neural-network mixture modeling with semi-supervised minmax+em learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5666-5673.",
+            "Saksham Pramanik, Aayush Mundra, Ashutosh Mittal, Sreyas Mohan, Saim Wani, Shramay S Vernekar, Pranav M Dixit, Shanti Priya, Ankur Beniwal, Ojaswa Sharma, and Senthil Mani. 2024. Evaluating precise geolocation inference capabilities of vision language models. Preprint, arXiv:2502.14412.",
+            "Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045, Brussels, Belgium. Association for Computational Linguistics.",
+            "Zirui Song, Jingpu Yang, Yuan Huang, Jonathan Tonglet, Zeyu Zhang, Tao Cheng, Meng Fang, Iryna Gurevych, and Xiuying Chen. 2025. Geolocation with real human gameplay data: A large-scale dataset and human-like reasoning framework. Preprint, arXiv:2502.13759.",
+            "Claude 3.7 Sonnet. 2025. Claude 3.7 sonnet documentation. https://docs.anthropic.com/en/docs/overview. Anthropic AI Model.",
+            "Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, Soroosh Mariooryad, Yifan Ding, Xinyang Geng, Fred Alcober, Roy Frostig, Mark Omernick, Lexi Walker, Cosmin Paduraru, Christina Sorokin, and 1118 others. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. Preprint, arXiv:2403.05530.",
+            "Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Rame, Morgane Riviere, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, and 197 others. 2025. Gemma 3 technical report. Preprint, arXiv:2503.19786.",
+            "Yicong Tian, Chen Chen, and Mubarak Shah. 2017. Cross-view image matching for geo-localization in urban environments. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1998-2006."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Carnegie Mellon University. 2023. Embodied symbiotic assistants that see, act, infer and chat.",
+            "Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. Preprint, arXiv:2409.12191.",
+            "Frederik Warburg, Søren Hauberg, Gregory D. D. Funke, and Yoko Yuki. 2020. Mapillary street-level sequences: A dataset for lifelong place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 931-940. IEEE.",
+            "Tobias Weyand, Ilya Kostrikov, and James Philbin. 2016. PlaNet - photo geolocation with convolutional neural networks. In Computer Vision - ECCV 2016, Proceedings, Part II, volume 9906 of Lecture Notes in Computer Science, pages 37-55. Springer.",
+            "Ruixiang Yang, Cheng Zhang, Lingxi Meng, He Wang, Xiaoyan Li, Yuke Li, Shuo Wang, Haoran Wei, Yiyang Li, Wentao Qu, Pengchuan Zhang, Jiazheng Xu, Bihan Wen, Diyi Yang, Kangkang Lu, Saurabh Gupta, Guanzhong Wang, Zhiqiang Shen, Baining Guo, and 3 others. 2024. VLMs as GeoGuessr masters—exceptional performance, hidden biases, and privacy risks. Preprint, arXiv:2502.11163.",
+            "Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. 2017. Scene parsing through ADE20K dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 633-641. IEEE."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            595
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            509,
+            609,
+            633,
+            627
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.1 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            509,
+            637,
+            744,
+            653
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.1.1 Questions",
+        "text_level": 1,
+        "bbox": [
+            509,
+            659,
+            653,
+            674
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "This section details the complete 21-question sequence (Table 7) that forms the core of the GeoChain benchmark, designed to evaluate the step-by-step geographic reasoning capabilities of Multimodal Large Language Models (MLLMs). Each question in the sequence is characterized by its rank, designated difficulty level (Easy, Medium, or Hard), expected response format (Binary, Multiclass, or Free-text), and its primary Question Category (Visual Cues, Geographical localization, Culture/Infrastructure, Terrain/Environment, or Exact Location). This comprehensive listing provides a transparent foundation for understanding the specific tasks underpinning the performance evaluations discussed throughout this paper.",
+        "bbox": [
+            507,
+            680,
+            884,
+            921
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "23634",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A.1.2 System Prompt",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            299,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To guide the Multimodal Large Language Models (MLLMs) and standardize their responses for the GeoChain benchmark tasks, the following system prompt was consistently employed:",
+        "bbox": [
+            112,
+            103,
+            485,
+            167
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "System Prompt",
+        "text_level": 1,
+        "bbox": [
+            139,
+            179,
+            258,
+            195
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are an accurate geolocation model. Given the image, answer the following questions in order. Please provide your best guess. Each question is also provided with question type. For Binary questions, answer Yes/No only. For Multiclass questions, answer as one of the provided options in brackets. Final question type is a free text question, answer it as a free string text. If you are not sure about the answer, give your best guess. Answer format should be a json dict with question indices as keys (0 indexed) and values as Answer: <answer>, Reasoning: <reasoning>.",
+        "bbox": [
+            137,
+            205,
+            463,
+            430
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.1.3 Tools and Infrastructure",
+        "text_level": 1,
+        "bbox": [
+            114,
+            448,
+            371,
+            462
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The execution of model inference was managed by Promptfoo<sup>1</sup>, a platform that ensures reproducibility in benchmarking by offering versatile prompt configuration and effective API linkage. We used the transformers library in Hugging Face<sup>2</sup>; to run the MaskFormer model for computing segmentation masks. These calculations were performed on an NVIDIA GeForce RTX 3060 graphics processing unit.",
+        "bbox": [
+            112,
+            466,
+            487,
+            609
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.2 Haversine Distance",
+        "text_level": 1,
+        "bbox": [
+            114,
+            621,
+            315,
+            634
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Haversine Distance the shortest distance over the Earth's surface between the predicted and ground-truth coordinates, assuming a spherical Earth.",
+        "bbox": [
+            112,
+            642,
+            487,
+            689
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The Haversine formula is given by:",
+        "bbox": [
+            132,
+            690,
+            396,
+            705
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Delta \\phi = \\phi_ {2} - \\phi_ {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            240,
+            712,
+            357,
+            728
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Delta \\lambda = \\lambda_ {2} - \\lambda_ {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            243,
+            734,
+            357,
+            751
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} a = \\sin^ {2} \\left(\\frac {\\Delta \\phi}{2}\\right) + \\cos (\\phi_ {1}) \\cos (\\phi_ {2}) \\sin^ {2} \\left(\\frac {\\Delta \\lambda}{2}\\right) \\\\ d = 2 R \\cdot \\arcsin (\\sqrt {a}) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            117,
+            753,
+            478,
+            808
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Here,  $d$  is the Haversine distance between two points  $(\\phi_1,\\lambda_1)$  and  $(\\phi_{2},\\lambda_{2})$ . This metric provides an interpretable and robust way to measure geographic prediction error.",
+        "bbox": [
+            112,
+            810,
+            487,
+            875
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "1https://www.promptfoo.dev 2https://huggingface.co/docs/transformers/en/ index",
+        "bbox": [
+            112,
+            881,
+            477,
+            919
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.3 Additional Analysis",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            714,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.3.1 Image Difficulty vs Question Difficulty Interaction",
+        "text_level": 1,
+        "bbox": [
+            509,
+            107,
+            873,
+            137
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To analyze how visual and reasoning difficulty interact, we compute a two-dimensional pass rate matrix over question difficulty (Easy, Medium, Hard) and image difficulty (Easy, Medium, Hard). Table 8 presents this breakdown for each model.",
+        "bbox": [
+            507,
+            143,
+            882,
+            223
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We observe a consistent trend across all models: accuracy declines with both increasing image difficulty and question difficulty. Importantly, hard questions on hard images represent the most challenging setting, with pass rates often below  $40\\%$  even for state-of-the-art models.",
+        "bbox": [
+            507,
+            225,
+            882,
+            319
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Gemini-2.5-pro shows the strongest resilience across the board, maintaining high scores even on hard questions in ambiguous scenes. In contrast, Claude 3.7 Sonnet and GPT 4.1 Mini exhibit large drops in performance under compounding difficulty, confirming their brittleness in multi-factor reasoning.",
+        "bbox": [
+            507,
+            322,
+            882,
+            434
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "This matrix allows us to quantify model sensitivity to visual ambiguity and pinpoint failure modes. For example, a model that performs well on hard questions from easy images but poorly on the same questions from hard images may lack robustness in interpreting noisy visual cues. Conversely, a model that fails uniformly on hard questions indicates weaknesses in logical chaining or symbolic inference. Together, this analysis emphasizes the need for benchmarks that probe cross-modal interactions, rather than evaluating visual or linguistic difficulty in isolation.",
+        "bbox": [
+            507,
+            436,
+            882,
+            627
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.3.2 Breakdown by Question Difficulty",
+        "text_level": 1,
+        "bbox": [
+            507,
+            640,
+            840,
+            657
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To better understand how models handle increasing reasoning complexity, we group questions by their annotated difficulty levels: Easy, Medium, and Hard. These difficulty tags were assigned manually based on the subtlety, required external knowledge, and ambiguity of each question.",
+        "bbox": [
+            507,
+            662,
+            882,
+            758
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Across all models, accuracy decreases consistently with question difficulty. Gemini-2.5-pro achieves the highest pass rates at all levels, followed closely by Gemini-2.5-Flash and GPT-4.1. Interestingly, Claude 3.7 Sonnet and GPT 4.1 Mini both exhibit sharp drops on hard questions, with performance falling below  $45\\%$  and  $35\\%$ , respectively.",
+        "bbox": [
+            507,
+            759,
+            884,
+            887
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "These findings suggest that while many models can answer surface-level geographic questions",
+        "bbox": [
+            507,
+            889,
+            882,
+            921
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "23635",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/d9d8502510840228a13de80f767ee6836c143516be78b2b9bad224acdfba877a.jpg",
+        "table_caption": [
+            "Table 7: The GeoChain 21-Step Benchmark Question Set."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Rank</td><td>Difficulty</td><td>Question</td><td>Question Type</td><td>Question Category</td></tr><tr><td>1</td><td>Easy</td><td>Do you see any boats or ships?</td><td>Binary</td><td>Visual Cues</td></tr><tr><td>2</td><td>Easy</td><td>Do you see one or more of the following vehicles: Bus, Truck, Car, Van, Motorbike, Minibike, Bicycle?</td><td>Binary</td><td>Visual Cues</td></tr><tr><td>3</td><td>Easy</td><td>Can you see any traffic lights?</td><td>Binary</td><td>Visual Cues</td></tr><tr><td>4</td><td>Easy</td><td>Can you see any flag?</td><td>Binary</td><td>Visual Cues</td></tr><tr><td>5</td><td>Easy</td><td>Would you say this location is near the Equator?</td><td>Binary</td><td>Geographical localization</td></tr><tr><td>6</td><td>Easy</td><td>Does this location seem to be close to the Poles?</td><td>Binary</td><td>Geographical localization</td></tr><tr><td>7</td><td>Easy</td><td>Is this place located in the Northern Hemisphere?</td><td>Binary</td><td>Geographical localization</td></tr><tr><td>8</td><td>Easy</td><td>Which continent best describes where this location is? (7 continents: North America/South America/Europe/Africa/Asia/Oceania/Antarctica)</td><td>Multiclass</td><td>Geographical localization</td></tr><tr><td>9</td><td>Medium</td><td>What side of the road do vehicles drive on here? (Left/Right)</td><td>Multiclass</td><td>Culture/Infrastructure</td></tr><tr><td>10</td><td>Medium</td><td>What country is this place located in?</td><td>Free-text</td><td>Geographical localization</td></tr><tr><td>11</td><td>Medium</td><td>Is this place near coast?</td><td>Binary</td><td>Terrain/Environment</td></tr><tr><td>12</td><td>Medium</td><td>Does this location appear to be an island?</td><td>Binary</td><td>Terrain/Environment</td></tr><tr><td>13</td><td>Easy</td><td>Is this place located in a desert region?</td><td>Binary</td><td>Terrain/Environment</td></tr><tr><td>14</td><td>Easy</td><td>Does this location seem to be in a mountainous or hilly region?</td><td>Binary</td><td>Terrain/Environment</td></tr><tr><td>15</td><td>Medium</td><td>What is the most likely climate type for this location? (5 main climate types: Tropical/Dry/Temperate/Continental/Polar)</td><td>Multiclass</td><td>Terrain/Environment</td></tr><tr><td>16</td><td>Easy</td><td>Does this place look like a big city?</td><td>Binary</td><td>Culture/Infrastructure</td></tr><tr><td>17</td><td>Medium</td><td>Would you classify this place as a small town?</td><td>Binary</td><td>Culture/Infrastructure</td></tr><tr><td>18</td><td>Hard</td><td>What language(s) are most likely spoken at this place?</td><td>Free-text</td><td>Culture/Infrastructure</td></tr><tr><td>19</td><td>Hard</td><td>Can you name the state or province this place belongs to?</td><td>Free-text</td><td>Geographical localization</td></tr><tr><td>20</td><td>Hard</td><td>What is the name of the city, town, or village seen here?</td><td>Free-text</td><td>Geographical localization</td></tr><tr><td>21</td><td>Hard</td><td>Based on everything observed, what are the latitude and longitude coordinates of this place? Please give a tuple of float coordinates (lat, lon)</td><td>Free-text</td><td>Exact Location</td></tr></table>",
+        "bbox": [
+            147,
+            256,
+            852,
+            770
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "23636",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/5b5236975823e6c00594b47f4ddee30d16834aa075a217680e85cb8b36a7eb6b.jpg",
+        "table_caption": [
+            "Table 8: Pass score (\\%) across question difficulty and image difficulty. Each row shows performance on a given question difficulty across images of increasing ambiguity."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Question Difficulty</td><td>Easy Images</td><td>Medium Images</td><td>Hard Images</td></tr><tr><td rowspan=\"3\">Claude 3.7 Sonnet</td><td>Easy</td><td>89.3</td><td>89.1</td><td>87.7</td></tr><tr><td>Medium</td><td>76.0</td><td>75.7</td><td>72.0</td></tr><tr><td>Hard</td><td>45.8</td><td>52.7</td><td>34.8</td></tr><tr><td rowspan=\"3\">GPT 4.1 Mini</td><td>Easy</td><td>86.7</td><td>86.7</td><td>84.2</td></tr><tr><td>Medium</td><td>66.1</td><td>67.3</td><td>62.1</td></tr><tr><td>Hard</td><td>37.4</td><td>44.6</td><td>27.9</td></tr><tr><td rowspan=\"3\">GPT-4.1</td><td>Easy</td><td>87.9</td><td>87.5</td><td>84.3</td></tr><tr><td>Medium</td><td>75.5</td><td>76.5</td><td>71.8</td></tr><tr><td>Hard</td><td>51.8</td><td>60.0</td><td>43.3</td></tr><tr><td rowspan=\"3\">Gemini-2.5-Flash</td><td>Easy</td><td>90.7</td><td>91.0</td><td>89.4</td></tr><tr><td>Medium</td><td>76.7</td><td>77.8</td><td>74.2</td></tr><tr><td>Hard</td><td>47.3</td><td>54.9</td><td>38.4</td></tr><tr><td rowspan=\"3\">Gemini-2.5-pro</td><td>Easy</td><td>91.6</td><td>91.3</td><td>89.8</td></tr><tr><td>Medium</td><td>78.2</td><td>79.9</td><td>75.7</td></tr><tr><td>Hard</td><td>52.4</td><td>61.6</td><td>45.9</td></tr></table>",
+        "bbox": [
+            139,
+            122,
+            858,
+            423
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/4ca659628ad021cc5074aed59993a3dcdc1410bfd46514afca4cb046e5da78f8.jpg",
+        "table_caption": [
+            "Table 9: Pass score  $(\\%)$  by question difficulty."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Easy</td><td>Medium</td><td>Hard</td></tr><tr><td>Claude 3.7 Sonnet</td><td>88.7</td><td>74.6</td><td>44.5</td></tr><tr><td>GPT 4.1 Mini</td><td>85.9</td><td>65.2</td><td>33.4</td></tr><tr><td>GPT-4.1</td><td>87.3</td><td>75.8</td><td>54.7</td></tr><tr><td>Gemini-2.5-Flash</td><td>90.8</td><td>76.2</td><td>51.3</td></tr><tr><td>Gemini-2.5-pro</td><td>91.1</td><td>78.4</td><td>55.1</td></tr></table>",
+        "bbox": [
+            119,
+            468,
+            485,
+            583
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "accurately, their reasoning falters as complexity increases especially when fine-grained localization or symbolic inference is required. The relatively better performance of Gemini-2.5-pro on hard questions indicates more stable multi-hop reasoning or greater robustness to subtle visual signals.",
+        "bbox": [
+            112,
+            609,
+            489,
+            707
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "A.3.3 Breakdown by Question Type",
+        "text_level": 1,
+        "bbox": [
+            112,
+            721,
+            410,
+            736
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "To assess how models handle varying degrees of response constraint, we analyzed Pass Scores across three fundamental question types: Binary, Multiclass, and Free-text, with results presented in Table 10 and Figure 6. This breakdown reveals a distinct performance hierarchy directly correlated with the open-endedness of the required answer.",
+        "bbox": [
+            112,
+            741,
+            487,
+            854
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Across all evaluated MLLMs, a clear difficulty gradient was observed: Binary questions yielded the highest success rates, followed by Multiclass questions, with Free-text questions proving to be",
+        "bbox": [
+            112,
+            857,
+            487,
+            921
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "the most challenging by a substantial margin. For instance, Gemini-2.5-pro achieved  $88.9\\%$  on Binary and an exceptional  $92.9\\%$  on Multiclass questions, but its score dropped to  $56.7\\%$  for Free-text tasks. This pattern of significantly lower performance on Free-text questions was universal, underscoring the inherent difficulty in precise, open-ended generation and factual recall compared to selecting from constrained options.",
+        "bbox": [
+            507,
+            445,
+            884,
+            592
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "In the structured formats, Gemini-2.5-pro consistently led, achieving the top scores for both Binary (88.9%) and Multiclass (92.9%) questions, with Gemini-2.5-Flash also performing strongly. Notably, for the more demanding Free-text questions, GPT-4.1 emerged as the top performer with a Pass Score of 57.8%, slightly ahead of Gemini-2.5-pro (56.7%). This suggests a particular strength in GPT-4.1's generative capabilities for unconstrained answers. Claude 3.7 Sonnet demonstrated robust performance on Binary (86.2%) and Multiclass (84.5%) questions, often comparable to GPT-4.1, but its accuracy significantly declined on Free-text questions (45.5%), reaffirming its challenges with precise, unprompted generation. As anticipated, GPT-4.1 Mini generally recorded the lowest scores across all types. This analysis by question type effectively highlights that while current MLLMs are largely proficient with constrained-choice tasks, open-ended free-text responses remain a key area",
+        "bbox": [
+            507,
+            599,
+            884,
+            921
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "23637",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/be9255124a5d18bc368fec9c415a0eec1ad0bfaac179ac06c60d96ab99d6425b.jpg",
+        "image_caption": [
+            "Figure 5: Mean Haversine distance (km) by city and model. Larger values indicate poor localization precision."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            84,
+            877,
+            350
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/cc05641d2ea3cd69aa16d75fe7a9e8182fafdca321a61a81ae8c1924775fbb70.jpg",
+        "image_caption": [
+            "Figure 6: Model vs Question Type"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            403,
+            478,
+            558
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "for improvement.",
+        "bbox": [
+            114,
+            609,
+            247,
+            625
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/54dd3e2469b1d1d3f72c50513b5939a3cd9fb74fc628346f4569deaa306509b6.jpg",
+        "table_caption": [
+            "Table 10: Pass score  $(\\%)$  by question type."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Binary</td><td>Multiclass</td><td>Free-text</td></tr><tr><td>Claude 3.7 Sonnet</td><td>86.2</td><td>84.5</td><td>45.5</td></tr><tr><td>GPT-4.1 Mini</td><td>82.3</td><td>73.8</td><td>37.5</td></tr><tr><td>GPT-4.1</td><td>85.9</td><td>85.8</td><td>57.8</td></tr><tr><td>Gemini-2.5-Flash</td><td>88.5</td><td>90.9</td><td>50.5</td></tr><tr><td>Gemini-2.5-pro</td><td>88.9</td><td>92.9</td><td>56.7</td></tr></table>",
+        "bbox": [
+            115,
+            663,
+            500,
+            777
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "A.4 Breakdown by City",
+        "text_level": 1,
+        "bbox": [
+            114,
+            803,
+            317,
+            819
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "A city-level view (Fig. 7) shows that performance is far from uniform:",
+        "bbox": [
+            112,
+            824,
+            485,
+            854
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Gemini-2.5-pro is the most stable, topping the leaderboard in 20 / 24 cities and exceeding  $85\\%$  accuracy in visually distinctive urban centres such as Tokyo, Zurich and Toronto. Gemini-2.5-Flash",
+        "bbox": [
+            112,
+            857,
+            487,
+            921
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "and GPT-4.1 follow closely, maintaining more than  $75\\%$  accuracy in most regions. Performance on Claude 3.7 Sonnet and GPT 4.1 Mini fluctuate sharply: they perform competitively in cue-rich European cities (Paris, Berlin) but collapse in visually ambiguous locales (Nairobi, São Paulo, Amman). Mean Haversine error (Fig. 5) confirms the pattern: Gemini-2.5-pro keeps errors below  $300\\mathrm{km}$  in nearly every city, whereas Claude and GPT 4.1 Mini exceed  $1000\\mathrm{km}$  in several cases (Helsinki, Melbourne, São Paulo).",
+        "bbox": [
+            507,
+            403,
+            884,
+            580
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "These results highlight how regional factors such as vegetation, signage language, traffic orientation and architectural style strongly modulate geolocation accuracy.",
+        "bbox": [
+            507,
+            580,
+            882,
+            645
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "23638",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f6280791744fb4f99ffc98a3d4f8e99b11a6bbceb4bad7834a6513f6bcc9447.jpg",
+        "image_caption": [
+            "Figure 7: Pass score (\\%) by city, highlighting the influence of geographical location on model accuracy."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            354,
+            875,
+            617
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "page_number",
+        "text": "23639",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 15
+    }
+]