Add Batch bb64c846-ae9e-4bd2-8061-08f84bae28ff data

.gitattributes

@@ -2741,3 +2741,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/T-REG_[[:space:]]Preference[[:space:]]Optimization[[:space:]]with[[:space:]]Token-Level[[:space:]]Reward[[:space:]]Regularization/fed22e5e-5a1b-40b5-8fb6-1526941f903a_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/T2A-Feedback_[[:space:]]Improving[[:space:]]Basic[[:space:]]Capabilities[[:space:]]of[[:space:]]Text-to-Audio[[:space:]]Generation[[:space:]]via[[:space:]]Fine-grained[[:space:]]AI[[:space:]]Feedback/4172e218-60f8-4fb5-b2d9-f899a2be6a02_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/T2I-FactualBench_[[:space:]]Benchmarking[[:space:]]the[[:space:]]Factuality[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Models[[:space:]]with[[:space:]]Knowledge-Intensive[[:space:]]Concepts/435f957e-e61f-401d-b1c3-262c1ca9ea55_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TACLR_[[:space:]]A[[:space:]]Scalable[[:space:]]and[[:space:]]Efficient[[:space:]]Retrieval-based[[:space:]]Method[[:space:]]for[[:space:]]Industrial[[:space:]]Product[[:space:]]Attribute[[:space:]]Value[[:space:]]Identification/939f3cd5-94e9-46cf-8def-cb09a1fd834c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TARGA_[[:space:]]Targeted[[:space:]]Synthetic[[:space:]]Data[[:space:]]Generation[[:space:]]for[[:space:]]Practical[[:space:]]Reasoning[[:space:]]over[[:space:]]Structured[[:space:]]Data/9e5b56f3-c8f5-4eb7-be25-daa4af06471e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TC–RAG_[[:space:]]Turing–Complete[[:space:]]RAG’s[[:space:]]Case[[:space:]]study[[:space:]]on[[:space:]]Medical[[:space:]]LLM[[:space:]]Systems/08530356-6e71-43b5-b2a0-ddb4582c96bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TEACH_[[:space:]]A[[:space:]]Contrastive[[:space:]]Knowledge[[:space:]]Adaptive[[:space:]]Distillation[[:space:]]Framework[[:space:]]for[[:space:]]Classical[[:space:]]Chinese[[:space:]]Understanding/aecc946d-77dc-4482-aa73-533ce4e2f50d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TESS[[:space:]]2_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Generalist[[:space:]]Diffusion[[:space:]]Language[[:space:]]Model/cc3a7cf6-9b0a-4a89-a779-2019421a1045_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TETRIS_[[:space:]]Optimal[[:space:]]Draft[[:space:]]Token[[:space:]]Selection[[:space:]]for[[:space:]]Batch[[:space:]]Speculative[[:space:]]Decoding/d8ef7f64-1cd8-42f9-93b1-aca035fdf587_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/THOR-MoE_[[:space:]]Hierarchical[[:space:]]Task-Guided[[:space:]]and[[:space:]]Context-Responsive[[:space:]]Routing[[:space:]]for[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/f6520cf3-9239-4e76-8a25-131bf94b31d7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TRACT_[[:space:]]Regression-Aware[[:space:]]Fine-tuning[[:space:]]Meets[[:space:]]Chain-of-Thought[[:space:]]Reasoning[[:space:]]for[[:space:]]LLM-as-a-Judge/142e9116-0830-41df-88fb-0f4abce5fead_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TRIDENT_[[:space:]]Enhancing[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Safety[[:space:]]with[[:space:]]Tri-Dimensional[[:space:]]Diversified[[:space:]]Red-Teaming[[:space:]]Data[[:space:]]Synthesis/f7ba882e-850d-4f46-b2a6-c075e0c61de6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TROVE_[[:space:]]A[[:space:]]Challenge[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Text[[:space:]]Provenance[[:space:]]via[[:space:]]Source[[:space:]]Sentence[[:space:]]Tracing[[:space:]]and[[:space:]]Relationship[[:space:]]Classification/0070d482-101d-4114-9d35-65f77b3fcfaa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TST_[[:space:]]A[[:space:]]Schema-Based[[:space:]]Top-Down[[:space:]]and[[:space:]]Dynamic-Aware[[:space:]]Agent[[:space:]]of[[:space:]]Text-to-Table[[:space:]]Tasks/70927308-302c-47c0-89de-e5f8e5f9c9ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TUMLU_[[:space:]]A[[:space:]]Unified[[:space:]]and[[:space:]]Native[[:space:]]Language[[:space:]]Understanding[[:space:]]Benchmark[[:space:]]for[[:space:]]Turkic[[:space:]]Languages/e19285ce-f019-47db-bc4e-ca86890ea20b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TUNA_[[:space:]]Comprehensive[[:space:]]Fine-grained[[:space:]]Temporal[[:space:]]Understanding[[:space:]]Evaluation[[:space:]]on[[:space:]]Dense[[:space:]]Dynamic[[:space:]]Videos/a4359d10-5e61-4364-9c79-6f90b4b9987e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TWIST_[[:space:]]Text-encoder[[:space:]]Weight-editing[[:space:]]for[[:space:]]Inserting[[:space:]]Secret[[:space:]]Trojans[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Models/eef8989b-066a-44df-88d9-cf70a3684ac2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Table-Critic_[[:space:]]A[[:space:]]Multi-Agent[[:space:]]Framework[[:space:]]for[[:space:]]Collaborative[[:space:]]Criticism[[:space:]]and[[:space:]]Refinement[[:space:]]in[[:space:]]Table[[:space:]]Reasoning/57614786-45cd-4dd8-82f1-0f480483559a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TableLoRA_[[:space:]]Low-rank[[:space:]]Adaptation[[:space:]]on[[:space:]]Table[[:space:]]Structure[[:space:]]Understanding[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/95f9f758-2327-40ad-b5ce-2cc3f82cf2e8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Takin-VC_[[:space:]]Expressive[[:space:]]Zero-Shot[[:space:]]Voice[[:space:]]Conversion[[:space:]]via[[:space:]]Adaptive[[:space:]]Hybrid[[:space:]]Content[[:space:]]Encoding[[:space:]]and[[:space:]]Enhanced[[:space:]]Timbre[[:space:]]Modeling/0df3e726-f843-4654-8e90-45e292537d22_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Taming[[:space:]]LLMs[[:space:]]with[[:space:]]Gradient[[:space:]]Grouping/9cf2debc-6405-42bb-8932-6ac33661dede_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Taming[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Text-attributed[[:space:]]Graph[[:space:]]Learning[[:space:]]with[[:space:]]Decoupled[[:space:]]Aggregation/4fbd717e-ceea-4a86-b9bf-174a48710094_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Targeted[[:space:]]Syntactic[[:space:]]Evaluation[[:space:]]for[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction/44ec2577-dd22-4ff9-849c-cdcaa1be17c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Task-Specific[[:space:]]Information[[:space:]]Decomposition[[:space:]]for[[:space:]]End-to-End[[:space:]]Dense[[:space:]]Video[[:space:]]Captioning/b6dd80bf-53bc-423a-a51a-0d93afde763a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TaxoAdapt_[[:space:]]Aligning[[:space:]]LLM-Based[[:space:]]Multidimensional[[:space:]]Taxonomy[[:space:]]Construction[[:space:]]to[[:space:]]Evolving[[:space:]]Research[[:space:]]Corpora/1dd7375e-4dba-496a-9fac-a82596391dbc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TeRDy_[[:space:]]Temporal[[:space:]]Relation[[:space:]]Dynamics[[:space:]]through[[:space:]]Frequency[[:space:]]Decomposition[[:space:]]for[[:space:]]Temporal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Completion/966c115d-5a5d-40cb-8b8f-cc6b3f9b4a01_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Teaching[[:space:]]Text[[:space:]]Agents[[:space:]]to[[:space:]]Learn[[:space:]]Sequential[[:space:]]Decision[[:space:]]Making[[:space:]]from[[:space:]]Failure/e2c046f7-2bbe-4990-9af2-b11184614735_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Teaching[[:space:]]Vision-Language[[:space:]]Models[[:space:]]to[[:space:]]Ask_[[:space:]]Resolving[[:space:]]Ambiguity[[:space:]]in[[:space:]]Visual[[:space:]]Questions/86ef37d4-a907-4041-92ba-a45bd87bf065_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Teaching[[:space:]]an[[:space:]]Old[[:space:]]LLM[[:space:]]Secure[[:space:]]Coding_[[:space:]]Localized[[:space:]]Preference[[:space:]]Optimization[[:space:]]on[[:space:]]Distilled[[:space:]]Preferences/a90efc6c-04b6-4f30-b21a-96c02b75a26b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TeamLoRA_[[:space:]]Boosting[[:space:]]Low-Rank[[:space:]]Adaptation[[:space:]]with[[:space:]]Expert[[:space:]]Collaboration[[:space:]]and[[:space:]]Competition/fdd0a622-7cd8-4c40-b323-b47314dc4392_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Temporal[[:space:]]Relation[[:space:]]Extraction[[:space:]]in[[:space:]]Clinical[[:space:]]Texts_[[:space:]]A[[:space:]]Span-based[[:space:]]Graph[[:space:]]Transformer[[:space:]]Approach/3bb3f196-68bc-4189-9b96-f216d84250e8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Temporal[[:space:]]reasoning[[:space:]]for[[:space:]]timeline[[:space:]]summarisation[[:space:]]in[[:space:]]social[[:space:]]media/6c718f78-4ea5-42e7-ad1f-bcb4d3d3362f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TestNUC_[[:space:]]Enhancing[[:space:]]Test-Time[[:space:]]Computing[[:space:]]Approaches[[:space:]]and[[:space:]]Scaling[[:space:]]through[[:space:]]Neighboring[[:space:]]Unlabeled[[:space:]]Data[[:space:]]Consistency/601e447f-2bde-485c-b1db-94d3d0dc96de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Text[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need_[[:space:]]LLM-enhanced[[:space:]]Incremental[[:space:]]Social[[:space:]]Event[[:space:]]Detection/c9a296b5-8c82-4f86-821a-b6a8c99cb62c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Text-to-ES[[:space:]]Bench_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]for[[:space:]]Converting[[:space:]]Natural[[:space:]]Language[[:space:]]to[[:space:]]Elasticsearch[[:space:]]Query/f11c7154-3746-4822-9f90-7b27e7c3e093_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/That[[:space:]]is[[:space:]]Unacceptable_[[:space:]]the[[:space:]]Moral[[:space:]]Foundations[[:space:]]of[[:space:]]Canceling/158cf7c0-79cf-4487-9e55-178d74dea12c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]AI[[:space:]]Gap_[[:space:]]How[[:space:]]Socioeconomic[[:space:]]Status[[:space:]]Affects[[:space:]]Language[[:space:]]Technology[[:space:]]Interactions/9b68bb85-f6dc-424a-8d6a-ba2f55d24621_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Alternative[[:space:]]Annotator[[:space:]]Test[[:space:]]for[[:space:]]LLM-as-a-Judge_[[:space:]]How[[:space:]]to[[:space:]]Statistically[[:space:]]Justify[[:space:]]Replacing[[:space:]]Human[[:space:]]Annotators[[:space:]]with[[:space:]]LLMs/91c1ef4d-2216-493f-b9d0-d3f9d6585aac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Cross-linguistic[[:space:]]Role[[:space:]]of[[:space:]]Animacy[[:space:]]in[[:space:]]Grammar[[:space:]]Structures/91a93f02-0ea5-4a12-8530-71fc051c3166_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Distracting[[:space:]]Effect_[[:space:]]Understanding[[:space:]]Irrelevant[[:space:]]Passages[[:space:]]in[[:space:]]RAG/314fd01e-2c04-4cf0-9bb8-1f0574c22cbc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Efficiency[[:space:]]vs.[[:space:]]Accuracy[[:space:]]Trade-off_[[:space:]]Optimizing[[:space:]]RAG-Enhanced[[:space:]]LLM[[:space:]]Recommender[[:space:]]Systems[[:space:]]Using[[:space:]]Multi-Head[[:space:]]Early[[:space:]]Exit/77e98168-bba4-4c26-ae9b-11516a742066_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Esethu[[:space:]]Framework_[[:space:]]Reimagining[[:space:]]Sustainable[[:space:]]Dataset[[:space:]]Governance[[:space:]]and[[:space:]]Curation[[:space:]]for[[:space:]]Low-Resource[[:space:]]Languages/58e334dd-25d6-4075-a2f8-88accb213a83_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Essence[[:space:]]of[[:space:]]Contextual[[:space:]]Understanding[[:space:]]in[[:space:]]Theory[[:space:]]of[[:space:]]Mind_[[:space:]]A[[:space:]]Study[[:space:]]on[[:space:]]Question[[:space:]]Answering[[:space:]]with[[:space:]]Story[[:space:]]Characters/5a2038c4-0354-4b58-9821-ef8891da5ac3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Harmonic[[:space:]]Structure[[:space:]]of[[:space:]]Information[[:space:]]Contours/40f942f2-5503-44eb-8bb0-8ad3899776df_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Hidden[[:space:]]Attention[[:space:]]of[[:space:]]Mamba[[:space:]]Models/357a4739-9796-4837-85cc-7f5e37144bd9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Impact[[:space:]]of[[:space:]]Auxiliary[[:space:]]Patient[[:space:]]Data[[:space:]]on[[:space:]]Automated[[:space:]]Chest[[:space:]]X-Ray[[:space:]]Report[[:space:]]Generation[[:space:]]and[[:space:]]How[[:space:]]to[[:space:]]Incorporate[[:space:]]It/10a9a6c7-806c-40f9-83da-343ca66a15e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Impact[[:space:]]of[[:space:]]Token[[:space:]]Granularity[[:space:]]on[[:space:]]the[[:space:]]Predictive[[:space:]]Power[[:space:]]of[[:space:]]Language[[:space:]]Model[[:space:]]Surprisal/e4e2a3c4-737a-4677-a966-d9329ac45925_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Impossibility[[:space:]]of[[:space:]]Fair[[:space:]]LLMs/3ece0b06-d3cb-4900-8046-100726f767c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Invisible[[:space:]]Hand_[[:space:]]Unveiling[[:space:]]Provider[[:space:]]Bias[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Code[[:space:]]Generation/58e28e99-f37d-483d-8bd2-c61b9a341141_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Knowledge[[:space:]]Microscope_[[:space:]]Features[[:space:]]as[[:space:]]Better[[:space:]]Analytical[[:space:]]Lenses[[:space:]]than[[:space:]]Neurons/39cb7702-b5de-4bd5-9f09-444ef0915060_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Lawyer[[:space:]]That[[:space:]]Never[[:space:]]Thinks_[[:space:]]Consistency[[:space:]]and[[:space:]]Fairness[[:space:]]as[[:space:]]Keys[[:space:]]to[[:space:]]Reliable[[:space:]]AI/25154214-228c-4ffe-9629-9d6af8d88b6a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Male[[:space:]]CEO[[:space:]]and[[:space:]]the[[:space:]]Female[[:space:]]Assistant_[[:space:]]Evaluation[[:space:]]and[[:space:]]Mitigation[[:space:]]of[[:space:]]Gender[[:space:]]Biases[[:space:]]in[[:space:]]Text-To-Image[[:space:]]Generation[[:space:]]of[[:space:]]Dual[[:space:]]Subjects/29eb85c1-4277-4be1-bb2e-e2e9b672fbda_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Mirage[[:space:]]of[[:space:]]Model[[:space:]]Editing_[[:space:]]Revisiting[[:space:]]Evaluation[[:space:]]in[[:space:]]the[[:space:]]Wild/d7b05e12-ebd1-4f8a-826a-2a0f9ec35d08_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Nature[[:space:]]of[[:space:]]NLP_[[:space:]]Analyzing[[:space:]]Contributions[[:space:]]in[[:space:]]NLP[[:space:]]Papers/356fd903-0d78-4562-b61a-c5f560d3b98a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Noisy[[:space:]]Path[[:space:]]from[[:space:]]Source[[:space:]]to[[:space:]]Citation_[[:space:]]Measuring[[:space:]]How[[:space:]]Scholars[[:space:]]Engage[[:space:]]with[[:space:]]Past[[:space:]]Research/71b3db0c-db8b-4e0c-a9d2-6335d1cbd90f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Role[[:space:]]of[[:space:]]Deductive[[:space:]]and[[:space:]]Inductive[[:space:]]Reasoning[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/a2241299-17e4-4c52-9ae8-e45ed2b00bb9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Role[[:space:]]of[[:space:]]Visual[[:space:]]Modality[[:space:]]in[[:space:]]Multimodal[[:space:]]Mathematical[[:space:]]Reasoning_[[:space:]]Challenges[[:space:]]and[[:space:]]Insights/e3d6e248-c8db-4d35-8c7f-154a4823fb52_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]TIP[[:space:]]of[[:space:]]the[[:space:]]Iceberg_[[:space:]]Revealing[[:space:]]a[[:space:]]Hidden[[:space:]]Class[[:space:]]of[[:space:]]Task-in-Prompt[[:space:]]Adversarial[[:space:]]Attacks[[:space:]]on[[:space:]]LLMs/3840bbbf-57c6-4849-bc8e-0cda9a8422c9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Task[[:space:]]Shield_[[:space:]]Enforcing[[:space:]]Task[[:space:]]Alignment[[:space:]]to[[:space:]]Defend[[:space:]]Against[[:space:]]Indirect[[:space:]]Prompt[[:space:]]Injection[[:space:]]in[[:space:]]LLM[[:space:]]Agents/01290c52-9bbd-4c1c-b62b-6ca046db1bcf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]Tug[[:space:]]of[[:space:]]War[[:space:]]Within_[[:space:]]Mitigating[[:space:]]the[[:space:]]Fairness-Privacy[[:space:]]Conflicts[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/d015762d-b7c7-489c-9494-7ad9a3af97f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]UD-NewsCrawl[[:space:]]Treebank_[[:space:]]Reflections[[:space:]]and[[:space:]]Challenges[[:space:]]from[[:space:]]a[[:space:]]Large-scale[[:space:]]Tagalog[[:space:]]Syntactic[[:space:]]Annotation[[:space:]]Project/2999d476-e72b-4898-9d87-2328219323b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/The[[:space:]]time[[:space:]]scale[[:space:]]of[[:space:]]redundancy[[:space:]]between[[:space:]]prosody[[:space:]]and[[:space:]]linguistic[[:space:]]context/d3312170-d2cd-48d7-a05d-158473ed3905_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Theorem[[:space:]]Prover[[:space:]]as[[:space:]]a[[:space:]]Judge[[:space:]]for[[:space:]]Synthetic[[:space:]]Data[[:space:]]Generation/7706a68a-e369-4e81-b540-b86dad5ab89d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/TheoremExplainAgent_[[:space:]]Towards[[:space:]]Video-based[[:space:]]Multimodal[[:space:]]Explanations[[:space:]]for[[:space:]]LLM[[:space:]]Theorem[[:space:]]Understanding/82226f43-5575-455b-8bde-7976df337ae9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Theoretical[[:space:]]Analysis[[:space:]]of[[:space:]]Hierarchical[[:space:]]Language[[:space:]]Recognition[[:space:]]and[[:space:]]Generation[[:space:]]by[[:space:]]Transformers[[:space:]]without[[:space:]]Positional[[:space:]]Encoding/7a8fb798-15b2-42e7-b7c2-aac02636fc72_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Theoretical[[:space:]]Guarantees[[:space:]]for[[:space:]]Minimum[[:space:]]Bayes[[:space:]]Risk[[:space:]]Decoding/df47d240-dd06-448a-a178-630e18bed0dc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Theory[[:space:]]of[[:space:]]Mind[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]Assessment[[:space:]]and[[:space:]]Enhancement/d0723338-4675-4710-925a-332bf35a49c0_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/TACLR_ A Scalable and Efficient Retrieval-based Method for Industrial Product Attribute Value Identification/939f3cd5-94e9-46cf-8def-cb09a1fd834c_content_list.json

@@ -0,0 +1,1854 @@
+[
+    {
+        "type": "text",
+        "text": "TACLR: A Scalable and Efficient Retrieval-Based Method for Industrial Product Attribute Value Identification",
+        "text_level": 1,
+        "bbox": [
+            196,
+            89,
+            801,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yindu Su $^{1,2}$ , Huike Zou $^{1}$ , Lin Sun $^{3}$ , Ting Zhang $^{2}$ , Haiyang Yang $^{1}$ , Liu Chen $^{1}$ , David Lo $^{2}$ , Qingheng Zhang $^{1}$ , Shuguang Han $^{1}$ , Jufeng Chen $^{1}$ ,",
+        "bbox": [
+            166,
+            145,
+            836,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Xianyu of Alibaba  $^{2}$ Singapore Management University  $^{3}$ Hangzhou City University",
+        "bbox": [
+            141,
+            180,
+            855,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "yindusu@foxmail.com, tingzhang.2019@phdcs.smu.edu.sg",
+        "bbox": [
+            240,
+            197,
+            757,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "qingheng.zqh@alibaba-inc.com，shuguang.sh@alibaba-inc.com",
+        "bbox": [
+            211,
+            214,
+            786,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Product Attribute Value Identification (PAVI) involves identifying attribute values from product profiles, a key task for improving product search, recommendation, and business analytics on e-commerce platforms. However, existing PAVI methods face critical challenges, such as inferring implicit values, handling out-of-distribution (OOD) values, and producing normalized outputs. To address these limitations, we introduce Taxonomy-Aware Contrastive Learning Retrieval (TACLR), the first retrieval-based method for PAVI. TACLR formulates PAVI as an information retrieval task by encoding product profiles and candidate values into embeddings and retrieving values based on their similarity. It leverages contrastive training with taxonomy-aware hard negative sampling and employs adaptive inference with dynamic thresholds. TACLR offers three key advantages: (1) it effectively handles implicit and OOD values while producing normalized outputs; (2) it scales to thousands of categories, tens of thousands of attributes, and millions of values; and (3) it supports efficient inference for high-load industrial deployment. Extensive experiments on proprietary and public datasets validate the effectiveness and efficiency of TACLR. Further, it has been successfully deployed on the real-world e-commerce platform Xianyu, processing millions of product listings daily with frequently updated, large-scale attribute taxonomies. We release the code to facilitate reproducibility and future research at https://github.com/SuYindu/TACLR.",
+        "bbox": [
+            141,
+            287,
+            460,
+            771
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            783,
+            258,
+            797
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Product attribute values are key components that support the operation of e-commerce platforms. They provide essential structural information, aiding customers in making informed purchasing decisions and enabling product listing (Chen et al., 2024), recommendation (Truong et al., 2022; Sun et al., 2020), retrieval (Magnani et al., 2019; Huang",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2a3df4355fa37b992a862a63ad8fedfc58e8f8c71ba29f228a10c8b5d1373f3b.jpg",
+        "image_caption": [
+            "Figure 1: Illustration of the PAVE and PAVI tasks. Unlike PAVE, which extracts raw value spans from product profiles, PAVI requires outputs to be normalized and supports both the identification of implicit values and the assignment of null to unavailable attribute values."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            258,
+            877,
+            479
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "et al., 2014), and question answering (Kulkarni et al., 2019; Gao et al., 2019).",
+        "bbox": [
+            507,
+            587,
+            880,
+            618
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, seller-provided attribute values are often incomplete or even inaccurate—an issue that is particularly severe on second-hand e-commerce platforms such as  $Xianyu^{1}$ . This undermines the effectiveness of downstream applications, making the automatic identification of product attribute values a fundamental requirement. Researchers have explored Product Attribute Value Extraction (PAVE), which involves extracting spans from product profiles using Named Entity Recognition (NER) (Zheng et al., 2018) or Question Answering (QA) (Wang et al., 2020) models. The upper part of Figure 1 illustrates an example of NER-based PAVE.",
+        "bbox": [
+            507,
+            620,
+            882,
+            828
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Although these approaches effectively extract value spans, the outputs remain raw text subsequences. Presenting attribute values in a standardized format is crucial for facilitating data aggrega",
+        "bbox": [
+            507,
+            829,
+            884,
+            894
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1https://www.goofish.com",
+        "bbox": [
+            531,
+            906,
+            715,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "31526",
+        "bbox": [
+            473,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 31526-31538",
+        "bbox": [
+            82,
+            945,
+            912,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 27 - August 1, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            268,
+            957,
+            727,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tion in business analytics and enhancing the user experience by providing clear and consistent information. To produce standardized values, a normalization step (Putthividhya and Hu, 2011; Zhang et al., 2021) is required to map these spans to predefined formats, as shown in the lower part of Figure 1. However, implicit values, such as Apple, cannot be directly extracted and must instead be inferred from context or prior knowledge.",
+        "bbox": [
+            112,
+            84,
+            489,
+            228
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Therefore, in this work, we focus on Product Attribute Value Identification (PAVI) (Shinzato et al., 2023), which aims to associate predefined attribute values from attribute taxonomy (illustrated in Figure 2) with products. The input for PAVI includes the product category and profile, where the profile includes textual data, such as the title and description, and may optionally incorporate visual information, such as images or videos. The output is a dictionary with predefined attributes as keys and their inferred values as corresponding entries. In addition, PAVI requires determining when attribute values are missing. For instance, as shown in Figure 1, the value for Version is unavailable and is therefore assigned an empty result or null value.",
+        "bbox": [
+            115,
+            230,
+            489,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Beyond adapting PAVE approaches, researchers have investigated classification-based (Chen et al., 2022) and generation-based paradigms (Sabeh et al., 2024b) for PAVI. Classification-based methods treat each value as a class; while this approach is straightforward, it is fundamentally limited by the inability to identify out-of-distribution (OOD) values not present in the training data, making such methods impractical for the continuously evolving nature of e-commerce platforms. In contrast, generation-based methods reformulate PAVI as a sequence-to-sequence task. Although these methods can handle implicit and OOD values, they face significant challenges, such as generating uncontrollable outputs and incurring substantial computational costs in high-load scenarios due to their reliance on Large Language Models (LLMs). In summary, existing approaches face distinct challenges, including difficulties in identifying implicit values, generalizing to OOD values, producing normalized outputs, and ensuring scalability and efficiency for large-scale industrial applications.",
+        "bbox": [
+            115,
+            470,
+            489,
+            824
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these limitations, we propose a novel retrieval-based method, Taxonomy-Aware Contrastive Learning Retrieval (TACLR). Our approach formulates PAVI as an information retrieval task: the product item serves as the query, and the attribute taxonomy acts as the corpus, enabling",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/d000e51ed45ea60b8795afc47bb3adc8796089f8fb316f60104054e705f50170.jpg",
+        "image_caption": [
+            "Figure 2: An illustration of a portion of the attribute taxonomy. Each category, such as Phone, is linked to multiple attributes, including Brand, Model, and Capacity, with standardized values enumerated for each attribute (e.g., Apple, Huawei, and Samsung for Brand)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            547,
+            83,
+            847,
+            204
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "efficient retrieval of relevant attribute values as matched documents. We use a shared encoder to generate embeddings for both the input product and candidate values from the attribute taxonomy. Our method adopts a contrastive learning framework inspired by CLIP (Radford et al., 2021). Rather than relying on in-batch negatives, we implement taxonomy-aware negative sampling, which selects hard negative values from the same category and attribute to generate a more challenging and precise contrastive signal. Additionally, learnable null values are explicitly incorporated as the ground truth for product-attribute pairs without associated values. During inference, we address the limitations of static thresholds by introducing dynamic thresholds derived from the relevance score of null values. This adaptive inference mechanism improves the model's ability to generalize across a large-scale attribute taxonomy.",
+        "bbox": [
+            507,
+            332,
+            884,
+            636
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are threefold: (1) We propose a novel retrieval-based paradigm for PAVI, introducing a scalable and efficient framework capable of handling implicit values, generalizing to OOD values, and producing normalized outputs. (2) We incorporate contrastive training into TACLR, using a taxonomy-aware negative sampling strategy to improve representation discrimination, and introduce an adaptive inference mechanism that dynamically balances precision and recall in large-scale industrial applications. (3) We validate the effectiveness of TACLR through extensive experiments on proprietary and public datasets, and demonstrate its successful deployment in a real-world industrial environment, processing millions of product listings across thousands of categories and millions of attribute values.",
+        "bbox": [
+            507,
+            646,
+            884,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "31527",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            270,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.1 Product Attribute Value Extraction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            110,
+            438,
+            124
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "PAVE as Named Entity Recognition. PAVE can be formulated as NER by identifying subsequences in product texts as entity spans and associating them with attributes as entity types. Early methods, such as OpenTag (Zheng et al., 2018), trained individual models for each category-attribute pair. Subsequent efforts generalized this approach to support multiple attributes or categories. For instance, SUOpenTag (Xu et al., 2019) incorporated attribute embeddings into an attention layer to handle multiple attributes, while AdaTag (Yan et al., 2021) used attribute embeddings to parameterize the decoder. TXtract (Karamanolakis et al., 2020) introduced a category encoder and a category attention mechanism to tackle various categories effectively. Additionally, M-JAVE (Zhu et al., 2020) jointly modeled attribute prediction and value extraction tasks while also incorporating visual information. More recently, Chen et al. (2023) scaled BERTNER by expanding the number of entity types to support a broader range of attributes.",
+        "bbox": [
+            112,
+            131,
+            489,
+            468
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "PAVE as Question Answering. The QA framework can also be adapted for PAVE by treating the product profile as context, attributes as questions, and value spans extracted from the context as answers. Wang et al. (2020) first introduced AVEQA for QA-based PAVE. Subsequent work extended this framework by incorporating multi-source information (Yang et al., 2022), multi-modal feature (Wang et al., 2022), and trainable prompts (Yang et al., 2023). Moreover, the question can be extended by appending candidate values as demonstrated by (Shinzato et al., 2022). Combining NER and QA paradigms, Ding et al. (2022) proposed a two-stage framework, which first identifies candidate values and then filters them.",
+        "bbox": [
+            112,
+            470,
+            489,
+            709
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "While NER- and QA-based paradigms have proven effective for PAVE, they struggle to identify implicit attribute values. Additionally, both paradigms rely on post-extraction normalization to standardize values, using either lexical (Putthividhya and Hu, 2011) or semantic methods (Zhang et al., 2021). Furthermore, extraction-based models require token-level annotations (e.g., BIO tags) for training and evaluation. Producing such annotations is significantly more resource-intensive than generating the value-level annotations used by TACLR, further limiting the scalability of these extraction-based methods.",
+        "bbox": [
+            112,
+            711,
+            489,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/f5fca30ada4d67d53736efda575284d7a62e894867e78482db0b7a6de142ed8d.jpg",
+        "table_caption": [
+            "Table 1: Comparison of different paradigms for identifying implicit, OOD, and normalized values."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Paradigm</td><td>Implicit</td><td>OOD</td><td>Normalized</td></tr><tr><td>Extraction</td><td>X</td><td>✓</td><td>X</td></tr><tr><td>Classification</td><td>✓</td><td>X</td><td>✓</td></tr><tr><td>Generation</td><td>✓</td><td>✓</td><td>X</td></tr><tr><td>Retrieval</td><td>✓</td><td>✓</td><td>✓</td></tr></table>",
+        "bbox": [
+            534,
+            121,
+            853,
+            205
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Product Attribute Value Identification",
+        "text_level": 1,
+        "bbox": [
+            507,
+            233,
+            855,
+            247
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Classification-Based PAVI. A straightforward approach is to frame PAVI as a multi-label classification problem over a finite set of values. Chen et al. (2022) trained a unified classification model that masks invalid labels based on the product category. However, a significant limitation of this classification-based paradigm is its inability to recognize OOD values not included in the training set. Consequently, its practicality is limited in dynamic e-commerce environments, where new categories and values frequently emerge.",
+        "bbox": [
+            505,
+            256,
+            884,
+            434
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Generation-Based PAVI. Recent advancements in LLM have spurred the exploration of generation-based PAVI methods (Sabeh et al., 2024b). Some methods (Roy et al., 2021; Nikolakopoulos et al., 2023; Blume et al., 2023) construct attribute-aware prompts to generate values for each attribute individually. In contrast, others generate values for multiple attributes simultaneously, either in a linearized sequence format (Shinzato et al., 2023) or as a hierarchical tree structure (Li et al., 2023). Multimodal information has also been integrated into LLMs to identify implicit attribute values from product images (Lin et al., 2021; Khandelwal et al., 2023). More recently, Brinkmann et al. (2024) explored the use of LLMs for both the extraction and normalization of attribute values. Additionally, Zou et al. (2024) introduced the learning-by-comparison technique to reduce model confusion, and Sabeh et al. (2024a) investigated Retrieval-Augmented Generation (RAG) technologies for PAVI.",
+        "bbox": [
+            507,
+            436,
+            884,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Although generation-based methods can infer implicit and OOD attribute values from product profiles, they face several challenges in real-world scenarios. A key issue is the potential for the LLMs to produce uncontrollable or hallucinated outputs, a known limitation of LLMs (Huang et al., 2024). Additionally, these methods often rely on large, computationally intensive models to achieve strong performance, making them inefficient and costly for large-scale industrial deployment.",
+        "bbox": [
+            507,
+            760,
+            885,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "31528",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/c4a96c6d30ecc506333876c61811570c3dcc294f57f1c80df239ba8d3ab13cc9.jpg",
+        "image_caption": [
+            "Figure 3: Overview of the training and inference process of TACLR, our retrieval-based method for the PAVI task. The left section illustrates contrastive training with taxonomy-aware negative sampling, while the right section demonstrates adaptive inference with pre-computed value embeddings."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            83,
+            467,
+            233
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/1d853a5f379894a50e542bf2833f9e7d16c22c18c3f8100744a99aaceccedbe3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            475,
+            83,
+            880,
+            233
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 Taxonomy-Aware Contrastive Learning Retrieval",
+        "text_level": 1,
+        "bbox": [
+            112,
+            319,
+            485,
+            353
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This section defines the PAVI task with an attribute taxonomy (§3.1) and presents our retrieval-based paradigm for PAVI (§3.2). We then detail the use of contrastive training with taxonomy-aware negative sampling (§3.3) and an adaptive inference mechanism with dynamic thresholds (§3.4). Figure 3 provides an overview of our approach TACLR.",
+        "bbox": [
+            112,
+            363,
+            487,
+            476
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 PAVI Task Definition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            488,
+            326,
+            504
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "PAVI is grounded in an attribute taxonomy that encompasses numerous product categories. For each category  $c$ , the taxonomy specifies a set of attributes  $\\mathcal{A}_c = \\{a_1, a_2, \\ldots\\}$  relevant to products in that category. For each attribute  $a \\in \\mathcal{A}_c$ , it provides a predefined set of standard values  $\\mathcal{V}_a = \\{v_1, v_2, \\ldots\\}$ . Figure 2 illustrates this structure<sup>2</sup>.",
+        "bbox": [
+            112,
+            510,
+            487,
+            621
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For a given product item  $i$ , with its title  $t$  and description  $d$ , the item is assigned to a category  $c$  with associated attributes  $\\mathcal{A}_c$ . The objective of the PAVI task is to identify a relevant set of values  $\\mathcal{V}_a^+\\subseteq \\mathcal{V}_a$  for each attribute  $a\\in \\mathcal{A}_c$ . The set  $\\mathcal{V}_a^+$  can take one of three forms: a singleton  $(\\{v\\})$ , multiple values  $(\\{v_{1},v_{2},\\ldots \\})$ , or an empty set  $(\\varnothing)$  if no information about  $a$  is available in the product profile. Notably, a standard value may not always appear verbatim as a text span in  $t$  or  $d$ ; it may instead be present in a paraphrased or synonymous form, which we refer to as an unnormalized value (e.g., the standard value iPhone 12 Pro Max expressed as iphone12pm in Figure 1). In other",
+        "bbox": [
+            112,
+            624,
+            489,
+            848
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "cases, a value may not be explicitly mentioned in the product profile but can be inferred from the context; these are referred to as implicit values (e.g., Apple in Figure 1).",
+        "bbox": [
+            507,
+            321,
+            884,
+            385
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Retrieval-Based PAVI",
+        "text_level": 1,
+        "bbox": [
+            507,
+            398,
+            726,
+            413
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In a standard information retrieval setting, given a query, the objective is to retrieve a list of relevant documents from a corpus. Similarly, for PAVI, we treat the input item as the query and the attribute taxonomy as the corpus, aiming to retrieve relevant attribute values as the output documents.",
+        "bbox": [
+            505,
+            420,
+            882,
+            517
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Encoding of items and values. We preprocess both item profiles and candidate values as textual inputs, utilizing a shared text encoder. For each item, we concatenate its title  $(t)$  and description  $(d)$  into a single input sentence formatted as: title: {title} description: {description}. Each candidate value is represented as a context-rich prompt, structured as: A {category} with {attribute} being {value}, e.g., A phone with brand being Apple. We explore the impact of various prompt templates in §5.3<sup>3</sup>.",
+        "bbox": [
+            505,
+            519,
+            882,
+            696
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Inference pipeline. During the deployment process, all value embeddings within the attribute taxonomy are pre-computed offline and indexed using the Faiss library4. During online inference, each item is encoded into an embedding, which is then compared against groups of indexed candidate value embeddings for various attributes. For each attribute  $a\\in \\mathcal{A}_c$  , the top-  $k$  most similar values are retrieved whose similarity scores exceed an adaptive threshold (3.4).",
+        "bbox": [
+            507,
+            697,
+            884,
+            858
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2Our approach focuses on attribute value identification, leveraging an existing attribute taxonomy as input rather than constructing or updating the taxonomy itself. In the Xianyu platform, a dedicated team and supporting system are responsible for maintaining the taxonomy (i.e., \"attribute mining\").",
+        "bbox": [
+            112,
+            858,
+            487,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "This framework can be extended to multimodal scenarios by replacing the text encoder with a multimodal encoder to incorporate features such as images.",
+        "bbox": [
+            507,
+            870,
+            882,
+            906
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>4</sup>https://github.com/facebookresearch/faiss",
+        "bbox": [
+            532,
+            906,
+            850,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "31529",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Contrastive Training",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            326,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Inspired by CLIP (Radford et al., 2021), we employ contrastive learning to train the shared encoder. Rather than relying on in-batch negatives, we compare each positive value with hard negative values from the same category and attribute in the taxonomy, providing a more challenging and precise training signal.",
+        "bbox": [
+            112,
+            105,
+            489,
+            217
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Formally, the subset of values matched with the item is referred to as the ground truth value set,  $\\mathcal{V}_a^+ \\subseteq \\mathcal{V}_a$ . If no matched values exist for a given attribute, i.e.,  $\\mathcal{V}_a^+ = \\emptyset$ , we assign a specific null value  $v_0^a$  for this attribute as the positive value, i.e.  $v_a^+ = v_0^a$ . Otherwise, a positive value is randomly drawn from the ground truth value set, i.e.  $v_a^+ \\sim \\mathcal{V}_a^+$ . For negative sampling, we select values as  $\\mathcal{V}_a^- = \\{v_1^-, v_2^-, \\ldots\\} \\subseteq \\mathcal{V}_a - \\mathcal{V}_a^+$ , ensuring a maximum of  $K$  values. The contrastive loss is then computed as follows:",
+        "bbox": [
+            112,
+            217,
+            489,
+            394
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {a} = - \\log \\left(\\frac {\\exp (\\frac {s (i , v _ {a} ^ {+})}{\\tau})}{\\exp (\\frac {s (i , v _ {a} ^ {+})}{\\tau}) + \\sum_ {v \\in \\mathcal {V} _ {a} ^ {-}} \\exp (\\frac {s (i , v)}{\\tau})}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            117,
+            404,
+            480,
+            472
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $s(i,v) = \\frac{I\\cdot V}{\\|I\\|\\|V\\|}$  denotes the cosine similarity between the item embedding  $I$  and the value embedding  $V$ , and  $\\tau$  is the temperature hyperparameter. It is important to note that each item typically includes multiple attributes, all of which share the same item embedding  $I$  while being individually compared against corresponding values. Therefore, the loss for item  $i$  is the sum of losses over all attributes from  $\\mathcal{A}_c$ :",
+        "bbox": [
+            112,
+            483,
+            489,
+            627
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {i} = \\sum_ {a \\in \\mathcal {A} _ {c}} \\mathcal {L} _ {a}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            245,
+            637,
+            356,
+            671
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "An example logit matrix is depicted on the left side of Figure 3. Note that the item embedding  $I_{1}$  contributes to the loss computations of  $\\mathcal{L}_1^1$ ,  $\\mathcal{L}_1^2$ , and  $\\mathcal{L}_1^3$ , which correspond to the attributes  $a_1, a_2$ , and  $a_3$  within the same product category. We also pad the logit matrix with negative infinity for batched computation if fewer than  $K$  values are available.",
+        "bbox": [
+            112,
+            681,
+            487,
+            793
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Adaptive Inference",
+        "text_level": 1,
+        "bbox": [
+            112,
+            804,
+            310,
+            819
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During retrieval, relevance scores are assigned to every candidate values. To filter output values, a static threshold  $T$  can be applied to these scores. However, in real-world e-commerce platforms with a vast number of category-attribute pairs, using a single threshold across all pairs is often suboptimal.",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Moreover, defining a unique threshold for each pair is tedious or even impractical.",
+        "bbox": [
+            507,
+            84,
+            880,
+            115
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To address this, we introduce an adaptive inference method that uses dynamic thresholds to make cutoff decisions. As discussed in §3.3, we add an explicit null value  $v_0^a$  for each category-attribute pair, with its embedding learned during training. In the inference phase, we compute the similarity  $s(i,v_0^a)$  between the item and the null value, using it as a dynamic threshold  $T_{a}^{\\prime}$  to exclude candidate values for attribute  $a$  that have lower scores:",
+        "bbox": [
+            507,
+            116,
+            884,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {V} _ {a} ^ {\\mathrm {p r e d}} = \\left\\{v \\mid s (i, v) > T _ {a} ^ {\\prime} \\right\\}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            594,
+            271,
+            794,
+            292
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Since most category-attribute pairs have exclusive values, meaning that each product can have at most one value for a given attribute, we focus on the top-1 predicted value in this work. The output can be further simplified as follows:",
+        "bbox": [
+            507,
+            302,
+            882,
+            382
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nv _ {a} ^ {\\text {p r e d}} = \\left\\{ \\begin{array}{l l} \\arg \\max  _ {v \\in \\mathcal {V} _ {a}} s (i, v) & \\text {i f} \\max  _ {v \\in \\mathcal {V} _ {a}} s (i, v) > T _ {a} ^ {\\prime} \\\\ n u l l & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            392,
+            875,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Equivalently, we can include the null value as an explicit candidate and retrieve the top-1 value as:",
+        "bbox": [
+            507,
+            451,
+            880,
+            483
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nv _ {a} ^ {\\text {p r e d}} = \\arg \\max  _ {v \\in \\mathcal {V} _ {a} \\cup \\{v _ {0} ^ {a} \\}} s (i, v)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            586,
+            494,
+            803,
+            521
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this case, selecting the  $v_{0}^{a}$  corresponds to the scenario where none of the specific candidate values surpass the dynamic threshold.",
+        "bbox": [
+            507,
+            532,
+            882,
+            579
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The inference process is illustrated on the right side of Figure 3. In this example, the predictions for  $a_3$  and  $a_5$  are determined to be empty because the highest-scoring value for these attributes is the null value.",
+        "bbox": [
+            507,
+            580,
+            880,
+            659
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            671,
+            719,
+            688
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            697,
+            621,
+            711
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To evaluate PAVI under the settings described in §3.1, we compare our proposed method against baselines on both proprietary and public datasets with normalized values<sup>5</sup>. Table 2 presents statistics of the attribute taxonomies and datasets.",
+        "bbox": [
+            507,
+            718,
+            880,
+            797
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Xianyu-PAVI. This dataset, derived from the e-commerce platform Xianyu, is constructed to evaluate the scalability and generalization of PAVI methods. The platform's attribute taxonomy comprises",
+        "bbox": [
+            507,
+            797,
+            884,
+            863
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "5Other popular benchmarks, such as AE-110k (Xu et al., 2019) and MAVE (Yang et al., 2022), provide only unnormalized values as spans extracted from product profiles, making them unsuitable for our experiments.",
+        "bbox": [
+            507,
+            871,
+            882,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "31530",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/1d3fa069674a848d507b1ba5cb40731d4b0f9704152ca65c551ac7ba0b3a12a1.jpg",
+        "table_caption": [
+            "Table 2: Statistics of the attribute taxonomies and dataset splits from Xianyu-PAVI and WDC-PAVE. \"CA Pairs\" refers to category-attribute pairs, \"CAV Tuples\" denotes category-attribute-value tuples, \"PA Pairs\" represents product-attribute pairs, and \"Null Pairs\" indicates product-attribute pairs with null values. \"Excl.\" refers to the test split excluding measurement attributes.",
+            "(a) Statistics of the attribute taxonomies."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Statistic</td><td>Xianyu</td><td>WDC</td></tr><tr><td># Categories</td><td>8,803</td><td>5</td></tr><tr><td># Attributes</td><td>3,326</td><td>24</td></tr><tr><td># CA Pairs</td><td>26,645</td><td>37</td></tr><tr><td># CAV Tuples</td><td>6,302,220</td><td>2,297</td></tr></table>",
+        "bbox": [
+            122,
+            168,
+            369,
+            247
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/67ed7ca0b2773058a6f14d87cd9203927e2b39d454555eb8870b88d3db2cc7b3.jpg",
+        "table_caption": [
+            "(b) Statistics of the datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Statistic</td><td colspan=\"3\">Xianyu-PAVI</td><td colspan=\"3\">WDC-PAVE</td></tr><tr><td>Train</td><td>Valid</td><td>Test</td><td>Train</td><td>Test</td><td>Excl.</td></tr><tr><td># Products</td><td>809,528</td><td>81,699</td><td>85,024</td><td>1,066</td><td>354</td><td>354</td></tr><tr><td># PA Pairs</td><td>3,584,462</td><td>358,582</td><td>458,954</td><td>8,832</td><td>2,937</td><td>2,285</td></tr><tr><td># Null Pairs</td><td>2,345,577</td><td>228,534</td><td>272,285</td><td>3,973</td><td>1,330</td><td>916</td></tr></table>",
+        "bbox": [
+            401,
+            168,
+            880,
+            253
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "8,803 product categories, 26,645 category-attribute pairs, and 6.3 million category-attribute-value tuples. On average, each category is associated with 3 attributes, each attribute has 237 possible values, and there are 716 candidate values per category.",
+        "bbox": [
+            112,
+            275,
+            487,
+            356
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For our experiments, we randomly sampled 1 million product items for training, 10,000 for validation, and 10,000 for testing. Each item is annotated with a category label and multiple attribute-value pair labels. Category labels were generated through a multi-step process involving automated classification, seller feedback, and annotator review, during which misclassified samples were discarded. Attribute-value pair labels were obtained through a multi-stage manual annotation process: a pool of annotators conducted the initial labeling, followed by quality checks and a second round of review, with items reassigned to different annotators after shuffling.",
+        "bbox": [
+            115,
+            357,
+            489,
+            581
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "WDC-PAVE (Brinkmann et al., 2024). This dataset contains 1,066 training and 354 test product items across 5 categories, with 8,832 and 2,937 product attribute pairs (3,973 and 1,330 nulls), respectively. On average, each category is associated with 7.4 attributes, each attribute has 62 values, and there are 459 attribute-value pairs per category. We conduct two evaluations: the first on the original test set, and the second on a test split that excludes measurement attributes, which require complex reasoning for unit conversion.",
+        "bbox": [
+            112,
+            583,
+            489,
+            758
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            770,
+            221,
+            784
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Since most attributes in the taxonomy are exclusive, i.e., each product can have at most one value per attribute, we evaluate PAVI methods using microaveraged precision@1, recall@1, and F1 score@1.",
+        "bbox": [
+            112,
+            791,
+            487,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For each attribute, the ground truth is a set of values  $\\mathcal{V}$  from the taxonomy. If the ground truth set is empty  $(\\varnothing)$ , a correct prediction (True Negative, TN) occurs when the model also predicts an empty",
+        "bbox": [
+            112,
+            857,
+            487,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/fe079ca9cae7c0aac574ab036e56e3264ca32a84e8b8dd2572b98a78af138991.jpg",
+        "table_caption": [
+            "Table 3: Confusion matrix comparing ground truth value set with predicted top-1 value."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Label</td><td>Prediction</td><td>Outcome</td></tr><tr><td>∅</td><td>∅</td><td>True Negative (TN)</td></tr><tr><td>∅</td><td>v</td><td>False Positive (FP)</td></tr><tr><td>V</td><td>v ∈ V</td><td>True Positive (TP)</td></tr><tr><td>V</td><td>∅</td><td>False Negative (FN)</td></tr><tr><td>V</td><td>v&#x27;∉ V</td><td>FP &amp; FN</td></tr></table>",
+        "bbox": [
+            554,
+            312,
+            836,
+            401
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "set; otherwise, it is a False Positive (FP). When the ground truth set is not empty, the model's top-1 output is a True Positive (TP) if it matches any ground truth value. Predicting an empty set in this case results in a False Negative (FN), while mismatched predictions are both False Positives (FP) and False Negatives (FN), as it simultaneously introduces an error and misses the correct value. Table 3 summarizes these outcomes<sup>6</sup>. Final precision, recall, and F1 scores are computed by aggregating TP, FP, and FN counts across the dataset, providing a comprehensive performance evaluation.",
+        "bbox": [
+            507,
+            426,
+            882,
+            619
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            631,
+            628,
+            645
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our retrieval-based method, TACLR, against classification and generation baselines<sup>7</sup>. For implementation details, refer to Appendix A. BERT-CLS. This baseline frames PAVI as a multi-label classification task, treating each category-attribute-value tuple as an independent label. The model is fine-tuned to predict matches, with label masking applied to exclude irrelevant labels for each category, following (Chen et al., 2022). The model outputs a probability distribution over val",
+        "bbox": [
+            507,
+            652,
+            884,
+            814
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "In prior work (Shinzato et al., 2023), metrics did not account for the FP case, and FP & FN cases were counted as FP only. We adopt more stringent metrics in this paper.",
+        "bbox": [
+            507,
+            822,
+            882,
+            860
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "7Extraction baselines, such as NER and QA models, are not included because (1) they produce unnormalized outputs, requiring an established normalization strategy for fair comparison, and (2) the large-scale Xianyu-PAVI lacks token-level annotations (e.g., BIO tags) necessary for these methods.",
+        "bbox": [
+            507,
+            860,
+            882,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "31531",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/b55cb6f89f0ed09e0a3ef1c1197222eaf6851c8ca154e62a743adef57e046316.jpg",
+        "table_caption": [
+            "Table 4: Performance comparison of classification-, generation-, and retrieval-based methods on Xianyu-PAVI and WDC-PAVE. \"F1 Excl.\" denotes the F1 score computed excluding measurement attributes (e.g., width and height), which require complex unit normalization reasoning."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Paradigm</td><td rowspan=\"2\">Method</td><td colspan=\"3\">Xianyu-PAVI</td><td colspan=\"4\">WDC-PAVE</td></tr><tr><td>Precision</td><td>Recall</td><td>F1</td><td>Precision</td><td>Recall</td><td>F1</td><td>F1 Excl.</td></tr><tr><td>Classification</td><td>BERT-CLS</td><td>50.9</td><td>50.1</td><td>50.5</td><td>68.9</td><td>12.0</td><td>20.5</td><td>23.4</td></tr><tr><td rowspan=\"8\">Generation</td><td>Llama3.1 (zero-shot)</td><td>29.1</td><td>46.2</td><td>35.7</td><td>56.6</td><td>60.8</td><td>58.6</td><td>64.6</td></tr><tr><td>Llama3.1 (few-shot)</td><td>31.0</td><td>51.1</td><td>38.6</td><td>76.0</td><td>74.1</td><td>75.0</td><td>79.0</td></tr><tr><td>Llama3.1 (RAG)</td><td>40.8</td><td>57.2</td><td>47.6</td><td>78.2</td><td>76.3</td><td>77.2</td><td>80.1</td></tr><tr><td>Llama3.1 (fine-tune)</td><td>86.9</td><td>82.7</td><td>84.7</td><td>57.7</td><td>60.4</td><td>59.0</td><td>64.5</td></tr><tr><td>Qwen2.5 (zero-shot)</td><td>42.7</td><td>55.7</td><td>48.4</td><td>51.9</td><td>60.3</td><td>55.8</td><td>60.8</td></tr><tr><td>Qwen2.5 (few-shot)</td><td>45.8</td><td>58.6</td><td>51.4</td><td>72.2</td><td>72.3</td><td>72.2</td><td>76.2</td></tr><tr><td>Qwen2.5 (RAG)</td><td>58.3</td><td>69.1</td><td>63.2</td><td>75.1</td><td>73.4</td><td>74.2</td><td>78.3</td></tr><tr><td>Qwen2.5 (fine-tune)</td><td>84.5</td><td>79.1</td><td>81.7</td><td>54.1</td><td>60.0</td><td>56.9</td><td>61.7</td></tr><tr><td>Retrieval</td><td>TACLR</td><td>85.4</td><td>87.1</td><td>86.2</td><td>74.3</td><td>70.9</td><td>72.6</td><td>80.3</td></tr></table>",
+        "bbox": [
+            154,
+            135,
+            843,
+            323
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ues and selects the highest-probability value for each attribute. If no probability exceeds a specified threshold, the prediction is set to empty.",
+        "bbox": [
+            112,
+            346,
+            487,
+            394
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "LLM. For generation-based baselines, we utilize state-of-the-art open-source LLMs, including Llama3.1-7B (Llama Team, 2024) and Qwen2.5-7B (Qwen Team, 2024). These models are evaluated in zero-shot and few-shot settings using a template adapted from (Brinkmann et al., 2024), which incorporates the category, attribute, product profile, and detailed value normalization guidelines. We further fine-tune the LLMs using LoRA (Hu et al., 2022) to improve performance.",
+        "bbox": [
+            112,
+            395,
+            489,
+            554
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "RAG. We implement RAG baselines based on both Qwen2.5 and Llama3.1, using BGE embeddings (Xiao et al., 2023) to retrieve the top-5 most similar product items from the training set. The LLM prompt is augmented with the profiles and structured attribute-value pair outputs of these retrieved examples. For all generation baselines, we apply greedy decoding to ensure reproducibility, and model outputs are formatted in JSON.",
+        "bbox": [
+            112,
+            556,
+            489,
+            701
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            713,
+            213,
+            727
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            739,
+            265,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4 presents the performance comparison between our retrieval-based method TACLR and classification- and generation-based baselines. On Xianyu-PAVI, TACLR achieves the highest F1 score of  $86.2\\%$ , surpassing fine-tuned Llama3.1, which obtains  $84.7\\%$ . Notably, TACLR excels in recall, achieving  $87.1\\%$  compared to Llama3.1's  $82.7\\%$ . On WDC-PAVE, TACLR achieves the highest F1 Excl. score of  $80.3\\%$ , which excludes measurement attributes requiring unit normalization",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "reasoning. These results highlight TACLR's effectiveness and robustness in addressing general PAVI across diverse datasets.",
+        "bbox": [
+            507,
+            346,
+            884,
+            393
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The classification-based baseline, BERT-CLS, shows the weakest performance on both datasets. It achieves an F1 score of  $50.5\\%$  on Xianyu-PAVI, but its performance drops drastically to  $20.5\\%$  on WDC-PAVE, underscoring the limitations of classification approaches in generalization. One contributing factor is the extreme label sparsity in this formulation: there are over 6.3M category-attribute-value labels, but the training set contains fewer than 3.6M instances, making it difficult for the model to learn a reliable classification head.",
+        "bbox": [
+            507,
+            399,
+            885,
+            576
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Among generation-based methods, performance improves steadily from zero-shot prompting to few-shot prompting, retrieval-augmented generation, and fine-tuning on the large-scale Xianyu-PAVI dataset. Llama3.1 progresses from  $35.7\\%$  (zero-shot) to  $38.6\\%$  (few-shot),  $47.6\\%$  (RAG), and  $84.7\\%$  (fine-tune), while Qwen2.5 improves from  $48.4\\%$  to  $51.4\\%$ ,  $63.2\\%$ , and  $81.7\\%$ , respectively. A similar trend holds on WDC-PAVE, though fine-tuning is less effective due to limited supervision. In this setting, RAG outperforms few-shot prompting for both Llama3.1 ( $77.2\\%$  vs.  $75.0\\%$ ) and Qwen2.5 ( $74.2\\%$  vs.  $72.2\\%$ ), offering a viable alternative when labeled data is limited. Nonetheless, TACLR consistently surpasses RAG, particularly on Xianyu-PAVI ( $86.2\\%$  vs.  $47.6\\%/63.2\\%$ ), demonstrating better scalability for large-scale industrial applications. On WDC-PAVE, TACLR matches or exceeds RAG and retains a slight advantage in non-measurement attributes (F1 Excl.:  $80.3\\%$  vs.  $80.1\\%/78.3\\%$ ).",
+        "bbox": [
+            507,
+            583,
+            885,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "31532",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/144c6fbaea4af2a825499897712fe32f60168a216b118155d4b28cf61dd44848.jpg",
+        "table_caption": [
+            "Table 5: Inference efficiency comparison on Xianyu-PAVI (Throughput in samples/second)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Time (ms)</td><td>Throughput</td></tr><tr><td>BERT-CLS</td><td>8.6</td><td>930</td></tr><tr><td>Llama3.1 (zero-shot)</td><td>101.3</td><td>80</td></tr><tr><td>Llama3.1 (few-shot)</td><td>124.8</td><td>64</td></tr><tr><td>Llama3.1 (RAG)</td><td>137.9</td><td>58</td></tr><tr><td>Qwen2.5 (zero-shot)</td><td>84.0</td><td>95</td></tr><tr><td>Qwen2.5 (few-shot)</td><td>98.4</td><td>81</td></tr><tr><td>Qwen2.5 (RAG)</td><td>108.3</td><td>74</td></tr><tr><td>TACLR</td><td>12.7</td><td>630</td></tr></table>",
+        "bbox": [
+            137,
+            121,
+            463,
+            259
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/ca355ec1d83184e2cff7ff1c004f12c5ea952d0f41c052d7c6426a995bfbe2ef.jpg",
+        "table_caption": [
+            "Table 6: F1 scores on product-attribute pairs with normalized vs. unnormalized and implicit values."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Normalized</td><td>Unnorm. &amp; Implicit</td></tr><tr><td>Llama3.1</td><td>83.2</td><td>79.4</td></tr><tr><td>Qwen2.5</td><td>82.6</td><td>78.6</td></tr><tr><td>TACLR</td><td>87.9</td><td>82.9</td></tr></table>",
+        "bbox": [
+            144,
+            312,
+            458,
+            385
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 Inference Efficiency",
+        "text_level": 1,
+        "bbox": [
+            112,
+            410,
+            315,
+            426
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5 presents a comparison of inference efficiency across different PAVI paradigms under identical conditions, using an unoptimized PyTorch implementation on a single NVIDIA V100 GPU.",
+        "bbox": [
+            112,
+            432,
+            489,
+            495
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "TACLR achieves a strong balance between model capacity and efficiency, processing each sample in  $12.7\\mathrm{ms}$  and achieving a throughput of 630 samples per second. In contrast, generation-based methods using Llama3.1 and Qwen2.5 exhibit substantially higher inference times (over  $100\\mathrm{ms}$  per sample) and lower throughput (below 100 samples per second), primarily due to the overhead of autoregressive decoding and reliance on large-scale models. The BERT-CLS classification baseline offers the highest raw efficiency  $(8.6\\mathrm{ms},$  930 samples per second), but its inability to handle out-of-distribution values and limited generalization capacity reduce its practical applicability.",
+        "bbox": [
+            112,
+            498,
+            489,
+            722
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In summary, among these methods, TACLR provides the best trade-off between identification capability and inference efficiency, making it well-suited for scalable industrial deployment.",
+        "bbox": [
+            112,
+            724,
+            489,
+            788
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            802,
+            226,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This section analyzes our method on the Xianyu-PAVI dataset, selected for its large scale and diverse attribute taxonomy. In contrast, the WDC-PAVE dataset is substantially smaller in both size and coverage, making it insufficient for robust evaluation of scalability and generalization.",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/110c7d7373481da84de0515654173d3e760e0f586de458ce05e148a74a873920.jpg",
+        "image_caption": [
+            "Figure 4: Comparison of negative sampling strategies with increasing number of samples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            578,
+            84,
+            816,
+            193
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Robustness Evaluation on Diverse Values. To better assess model robustness under various conditions, we partition the test set into three subsets based on the nature of the ground truth: (1) normalized values, appear verbatim in the product profile; (2) null values, where the attribute is marked as unavailable; and (3) unnormalized or implicit values, which either appear in a lexically varied form, or must be inferred. In our Xianyu-PAVI dataset, normalized values account for approximately  $15.2\\%$ , unnormalized or implicit values for  $27.2\\%$ , and null values for  $57.6\\%$ .",
+        "bbox": [
+            507,
+            260,
+            884,
+            451
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6 reports the F1 scores on the normalized and unnormalized/implicit subsets. TACLR achieves the highest performance on both subsets, with an F1 score of  $87.9\\%$  on normalized values and  $82.9\\%$  on unnormalized or implicit values. Compared to the fine-tuned LLM baselines, TACLR demonstrates stronger ability to recognize standardized values and better robustness to lexical variation and implicit reasoning.",
+        "bbox": [
+            507,
+            454,
+            882,
+            598
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of Taxonomy-Aware Negative Sampling. Figure 4 compares the proposed taxonomy-aware negative sampling (§3.3) with in-batch negative sampling across varying sample sizes. As the number of sampled values increases, the F1 score improves consistently, aligning with findings from (Chen et al., 2020). Using in-batch sampling as the baseline, the model achieves an F1 score of  $53.3\\%$  with a sample size of 128. In contrast, taxonomy-aware sampling yields substantial improvements, boosting the F1 score from  $84.0\\%$  to  $86.2\\%$  as the sample size grows from 16 to 128.",
+        "bbox": [
+            507,
+            599,
+            882,
+            791
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "These results demonstrate that taxonomy-aware sampling provides more effective supervision, encouraging the model to distinguish between fine-grained, semantically similar values within the same category and attribute. In contrast, in-batch sampling often yields random negatives that are less relevant and frequently trivial, limiting the efficacy of contrastive learning for PAVI. For example,",
+        "bbox": [
+            507,
+            793,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "31533",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3db2449ecd459fc902de0035a6c462124b8ceafa2409d8df5b1468c7a031adcc.jpg",
+        "image_caption": [
+            "(a) Comparison of inference thresholds."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            82,
+            371,
+            215
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/6501c664b39e2dc3593dde2ff818a5aa3d26d6d994642c74e858a119266379ba.jpg",
+        "image_caption": [
+            "(b) Comparison of prompt templates."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            374,
+            82,
+            623,
+            215
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/8ff9f6144c165ae41cebd84daf7487fc9dc8843d2e17d33bc01e1b10859eacdb.jpg",
+        "image_caption": [
+            "(c) Comparison of data domains.",
+            "Figure 5: Performance analysis across inference thresholds, prompt templates, and data domains."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            628,
+            82,
+            877,
+            215
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "in Figure 1, negatives such as iPhone 12 Pro or iPhone 13 Pro Max, which belong to the model attribute under the phone category, provide more challenging and informative supervision than random values such as L for T-shirt size.",
+        "bbox": [
+            112,
+            285,
+            487,
+            365
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Comparison of Dynamic and Static Thresholds. Figure 5a compares our dynamic thresholding approach (§3.4) with static thresholds of 0.6, 0.65, and 0.7, selected via validation. The dynamic threshold achieves the highest F1 (86.2%), outperforming the static baselines (75.5%, 80.2%, and 78.2%). Static thresholds exhibit the typical precision-recall trade-off: higher thresholds increase precision (65.1%→84.5%) but reduce recall (90.0%→72.8%). In contrast, the dynamic threshold balances precision (85.4%) and recall (87.1%) without manual tuning. This gain stems from a scalable design: instead of relying on fixed, hand-tuned cutoffs per category-attribute pair, our method learns null value embeddings whose similarity scores act as adaptive thresholds.",
+        "bbox": [
+            112,
+            368,
+            489,
+            626
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Performance Gains from Context-Rich Prompts. The influence of varying value prompt templates on the PAVI task is shown in Figure 5b. Using only the value as a prompt achieves an F1 score of  $83.2\\%$ . Adding category information raises the F1 score to  $83.9\\%$ , while incorporating attribute information further improves it to  $85.4\\%$ . The most comprehensive template, combining category, attribute, and value information (i.e., A {category} with {attribute} being {value}), achieves the highest F1 score of  $86.2\\%$ . These results are consistent with prior work (Radford et al., 2021), highlighting that context-rich prompts enhance the model's discriminative performance.",
+        "bbox": [
+            112,
+            629,
+            489,
+            853
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Zero-Shot Generalization Across Data Domains. Figure 5c presents zero-shot transfer results on unseen categories and values. The in-domain split achieves an F1 score of  $88.7\\%$ , while cross",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "category and cross-value splits decline to  $80.2\\%$  and  $78.2\\%$ , respectively, reflecting the challenges of adapting to evolving attribute taxonomies in out-of-distribution domains. Despite this, TACLR maintains a strong overall F1 of  $86.2\\%$ , demonstrating robust generalization. The model's generalization relies on the shared textual encoder's semantic understanding and the retrieval-based approach's capacity to leverage these embeddings for zero-shot matching. Such adaptability is critical for dynamic e-commerce platforms, where new products and attribute-value pairs continuously emerge, reducing the need for frequent retraining and lowering maintenance costs.",
+        "bbox": [
+            507,
+            285,
+            884,
+            510
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            524,
+            640,
+            539
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we present TACLR, a novel retrieval-based approach for PAVI. By formulating PAVI as an information retrieval problem, TACLR enables the inference of implicit values, generalization to OOD values, and the production of normalized outputs. Building on this framework, TACLR employs contrastive training with taxonomy-aware sampling and adaptive inference with dynamic thresholds to enhance retrieval performance and scalability.",
+        "bbox": [
+            507,
+            550,
+            884,
+            695
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Comprehensive experiments on proprietary and public datasets demonstrated TACLR's superiority over classification- and generation-based baselines. Notably, TACLR achieved an F1 score of  $86.2\\%$  on the large-scale Xianyu-PAVI dataset. Our efficiency analysis further highlighted its advantage, achieving significantly faster inference speeds than generation-based methods. Beyond these experimental results, TACLR has been successfully deployed on the real-world e-commerce platform Xianyu, processing millions of product listings daily and seamlessly adapting to dynamic attribute taxonomies, making it a practical solution for large-scale industrial applications.",
+        "bbox": [
+            507,
+            696,
+            884,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "31534",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            250,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "TACLR assumes access to a predefined attribute taxonomy, which serves as a foundation for accurate value identification. While this setup is realistic in many e-commerce platforms, where dedicated systems and teams maintain evolving taxonomies, it does require ongoing manual updates to incorporate new categories, attributes, and values. Automating product attribute mining remains an open area for future research (Ghani et al., 2006; Zhang et al., 2022; Xu et al., 2023).",
+        "bbox": [
+            112,
+            109,
+            489,
+            268
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Another limitation is that TACLR currently operates on textual product profiles and does not incorporate multimodal information, such as images or videos. Multimodal inputs could provide complementary signals for attributes that are difficult to infer from text alone (e.g., color, material, or shape). Extending the method to support multimodal input could further improve its coverage and accuracy in practical applications.",
+        "bbox": [
+            112,
+            269,
+            489,
+            413
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A further limitation of TACLR, as a retrieval-based approach, is its difficulty in handling measurement attributes that require unit conversion or numerical reasoning. To address this, future work could explore hybrid methods that combine TACLR's retrieval strengths with generative techniques to improve performance in such cases.",
+        "bbox": [
+            112,
+            414,
+            489,
+            526
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            552,
+            213,
+            568
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ansel Blume, Nasser Zalmout, Heng Ji, and Xian Li. 2023. Generative models for product attribute extraction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 575-585, Singapore. Association for Computational Linguistics.",
+            "Alexander Brinkmann, Nick Baumann, and Christian Bizer. 2024. Using llms for the extraction and normalization of product attribute values. In European Conference on Advances in Databases and Information Systems, pages 217-230. Springer.",
+            "Kang Chen, Qing Heng Zhang, Chengbao Lian, Yixin Ji, Xuwei Liu, Shuguang Han, Guoqiang Wu, Fei Huang, and Jufeng Chen. 2024. IPL: Leveraging multimodal large language models for intelligent product listing. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 697-711, Miami, Florida, US. Association for Computational Linguistics.",
+            "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings"
+        ],
+        "bbox": [
+            115,
+            576,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "of Machine Learning Research, pages 1597-1607.  \nPMLR.",
+            "Wei-Te Chen, Keiji Shinzato, Naoki Yoshinaga, and Yandi Xia. 2023. Does named entity recognition truly not scale up to real-world product attribute extraction? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 152-159, Singapore. Association for Computational Linguistics.",
+            "Wei-Te Chen, Yandi Xia, and Keiji Shinzato. 2022. Extreme multi-label classification with label masking for product attribute value extraction. In Proceedings of the Fifth Workshop on e-Commerce and NLP (EC-NLP 5), pages 134–140, Dublin, Ireland. Association for Computational Linguistics.",
+            "Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, Shijin Wang, and Guoping Hu. 2020. Revisiting pre-trained models for Chinese natural language processing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 657-668, Online. Association for Computational Linguistics.",
+            "Yifan Ding, Yan Liang, Nasser Zalmout, Xian Li, Christan Grant, and Tim Weninger. 2022. Ask-and-verify: Span candidate generation and verification for attribute value extraction. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 110–110, Abu Dhabi, UAE. Association for Computational Linguistics.",
+            "Shen Gao, Zhaochun Ren, Yihong Zhao, Dongyan Zhao, Dawei Yin, and Rui Yan. 2019. Product-aware answer generation in e-commerce question-answering. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM '19, page 429-437, New York, NY, USA. Association for Computing Machinery.",
+            "Rayid Ghani, Katharina Probst, Yan Liu, Marko Krema, and Andrew Fano. 2006. Text mining for product attribute extraction. SIGKDD Explor. Newsl., 8(1):41-48.",
+            "Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735.",
+            "Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.",
+            "Junshi Huang, Si Liu, Junliang Xing, Tao Mei, and Shuicheng Yan. 2014. *Circle & search: Attribute-aware shoe retrieval*. ACM Transactions on Multimedia Computing, Communications, and Applications, 11(1)."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "31535",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2024. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst. Just Accepted.",
+            "Giannis Karamanolakis, Jun Ma, and Xin Luna Dong. 2020. TXtract: Taxonomy-aware knowledge extraction for thousands of product categories. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8489-8502, Online. Association for Computational Linguistics.",
+            "Anant Khandelwal, Happy Mittal, Shreyas Kulkarni, and Deepak Gupta. 2023. Large scale generative multimodal attribute extraction for E-commerce attributes. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track), pages 305-312, Toronto, Canada. Association for Computational Linguistics.",
+            "Ashish Kulkarni, Kartik Mehta, Shweta Garg, Vidit Bansal, Nikhil Rasiwasia, and Srinivasan Sengamedu. 2019. Productqna: Answering user questions on e-commerce product pages. In Companion Proceedings of The 2019 World Wide Web Conference, WWW '19, page 354-360, New York, NY, USA. Association for Computing Machinery.",
+            "Yanzeng Li, Bingcong Xue, Ruoyu Zhang, and Lei Zou. 2023. AtTGen: Attribute tree generation for real-world attribute joint extraction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2139-2152, Toronto, Canada. Association for Computational Linguistics.",
+            "Rongmei Lin, Xiang He, Jie Feng, Nasser Zalmout, Yan Liang, Li Xiong, and Xin Luna Dong. 2021. Pam: Understanding product images in cross product category attribute extraction. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD '21, page 3262-3270, New York, NY, USA. Association for Computing Machinery.",
+            "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692.",
+            "Meta Llama Team. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.",
+            "Alessandro Magnani, Feng Liu, Min Xie, and Somnath Banerjee. 2019. Neural product retrieval at Walmart.com. In Companion Proceedings of The 2019 World Wide Web Conference, WWW '19, page 367-372, New York, NY, USA. Association for Computing Machinery.",
+            "Athanasios N. Nikolakopoulos, Swati Kaul, Siva Karthik Gade, Bella Dubrov, Umit Batur, and Suleiman Ali Khan. 2023. Sage: Structured"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "attribute value generation for billion-scale product catalogs. Preprint, arXiv:2309.05920.",
+            "Duangmanee Putthividhya and Junling Hu. 2011. Bootstrapped named entity recognition for product attribute extraction. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1557-1567, Edinburgh, Scotland, UK. Association for Computational Linguistics.",
+            "Alibaba Qwen Team. 2024. Qwen2 technical report. Preprint, arXiv:2407.10671.",
+            "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.",
+            "Kalyani Roy, Pawan Goyal, and Manish Pandey. 2021. Attribute value generation from product title using language models. In Proceedings of the 4th Workshop on e-Commerce and NLP, pages 13-17, Online. Association for Computational Linguistics.",
+            "Kassem Sabeh, Mouna Kacimi, Johann Gamper, Robert Litschko, and Barbara Plank. 2024a. Exploring large language models for product attribute value identification. Preprint, arXiv:2409.12695.",
+            "Kassem Sabeh, Robert Litschko, Mouna Kacimi, Barbara Plank, and Johann Gamper. 2024b. An empirical comparison of generative approaches for product attribute-value identification. Preprint, arXiv:2407.01137.",
+            "Keiji Shinzato, Naoki Yoshinaga, Yandi Xia, and Wei-Te Chen. 2022. Simple and effective knowledge-driven query expansion for QA-based product attribute extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 227-234, Dublin, Ireland. Association for Computational Linguistics.",
+            "Keiji Shinzato, Naoki Yoshinaga, Yandi Xia, and Wei-Te Chen. 2023. A unified generative approach to product attribute-value identification. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 6599–6612, Toronto, Canada. Association for Computational Linguistics.",
+            "Changfeng Sun, Han Liu, Meng Liu, Zhaochun Ren, Tian Gan, and Liqiang Nie. 2020. Lara: Attribute-to-feature adversarial learning for new-item recommendation. In Proceedings of the 13th International Conference on Web Search and Data Mining, WSDM '20, page 582-590, New York, NY, USA. Association for Computing Machinery.",
+            "Quoc-Tuan Truong, Tong Zhao, Changhe Yuan, Jin Li, Jim Chan, Soo-Min Pantel, and Hady W. Lauw. 2022."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "31536",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ampsum: Adaptive multiple-product summarization towards improving recommendation captions. In Proceedings of the ACM Web Conference 2022, WWW '22, page 2978-2988, New York, NY, USA. Association for Computing Machinery.",
+            "Qifan Wang, Li Yang, Bhargav Kanagal, Sumit Sanghai, D. Sivakumar, Bin Shu, Zac Yu, and Jon Elsas. 2020. Learning to extract attribute value from product via question answering: A multi-task approach. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '20, page 47-55, New York, NY, USA. Association for Computing Machinery.",
+            "Qifan Wang, Li Yang, Jingang Wang, Jitin Krishnan, Bo Dai, Sinong Wang, Zenglin Xu, Madian Khabsa, and Hao Ma. 2022. SMARTAVE: Structured multimodal transformer for product attribute value extraction. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 263-276, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-pack: Packaged resources to advance general chinese embedding. Preprint, arXiv:2309.07597.",
+            "Huimin Xu, Wenting Wang, Xin Mao, Xinyu Jiang, and Man Lan. 2019. Scaling up open tagging from tens to thousands: Comprehension empowered attribute value extraction from product title. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5214-5223, Florence, Italy. Association for Computational Linguistics.",
+            "Liyan Xu, Chenwei Zhang, Xian Li, Jingbo Shang, and Jinho D. Choi. 2023. Towards open-world product attribute mining: A lightly-supervised approach. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12223-12239, Toronto, Canada. Association for Computational Linguistics.",
+            "Jun Yan, Nasser Zalmout, Yan Liang, Christian Grant, Xiang Ren, and Xin Luna Dong. 2021. AdaTag: Multi-attribute value extraction from product profiles with adaptive decoding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4694–4705, Online. Association for Computational Linguistics.",
+            "Li Yang, Qifan Wang, Jingang Wang, Xiaojun Quan, Fuli Feng, Yu Chen, Madian Khabsa, Sinong Wang, Zenglin Xu, and Dongfang Liu. 2023. MixPAVE: Mix-prompt tuning for few-shot product attribute value extraction. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9978-9991, Toronto, Canada. Association for Computational Linguistics.",
+            "Li Yang, Qifan Wang, Zac Yu, Anand Kulkarni, Sumit Sanghai, Bin Shu, Jon Elsas, and Bhargav Kanagal."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "2022. Mave: A product dataset for multi-source attribute value extraction. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, WSDM '22, page 1256-1265, New York, NY, USA. Association for Computing Machinery.",
+            "Danqing Zhang, Zheng Li, Tianyu Cao, Chen Luo, Tony Wu, Hanqing Lu, Yiwei Song, Bing Yin, Tuo Zhao, and Qiang Yang. 2021. Queaco: Borrowing treasures from weakly-labeled behavior data for query attribute value extraction. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, CIKM '21, page 4362-4372, New York, NY, USA. Association for Computing Machinery.",
+            "Xinyang Zhang, Chenwei Zhang, Xian Li, Xin Luna Dong, Jingbo Shang, Christos Faloutsos, and Jiawei Han. 2022. Oa-min: Open-world attribute mining for e-commerce products with weak supervision. In Proceedings of the ACM Web Conference 2022, WWW '22, page 3153-3161, New York, NY, USA. Association for Computing Machinery.",
+            "Guineng Zheng, Subhabrata Mukherjee, Xin Luna Dong, and Feifei Li. 2018. Opentag: Open attribute value extraction from product profiles. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '18, page 1049-1058, New York, NY, USA. Association for Computing Machinery.",
+            "Tiangang Zhu, Yue Wang, Haoran Li, Youzheng Wu, Xiaodong He, and Bowen Zhou. 2020. Multimodal joint attribute prediction and value extraction for E-commerce product. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2129-2139, Online. Association for Computational Linguistics.",
+            "Henry Zou, Gavin Yu, Ziwei Fan, Dan Bu, Han Liu, Peng Dai, Dongmei Jia, and Cornelia Caragea. 2024. EIVEN: Efficient implicit attribute value extraction using multimodal LLM. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 6: Industry Track), pages 453-463, Mexico City, Mexico. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            722
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            510,
+            734,
+            749,
+            751
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "For both the BERT-CLS baseline and our TACLR method, we utilize pre-trained RoBERTa-base (Liu et al., 2019; Cui et al., 2020) as the backbone. For TACLR, we augment the model with a linear projection head to map the embedding dimension to 256. For each product-attribute pair, we sample up to 128 values for the contrastive learning setup, including a null value, an optional positive value (if present for the attribute), and negative values sampled from the value set with the same category and",
+        "bbox": [
+            510,
+            760,
+            884,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "31537",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "attribute. In the case of the null value, we replace the value slot in the prompt template with null (e.g., A phone with capacity being null.).",
+        "bbox": [
+            112,
+            84,
+            485,
+            131
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The temperature parameter is fixed at 0.05. This choice is motivated by both prior research (e.g., SimCLR (Chen et al., 2020) and MoCo (He et al., 2020)) and empirical tuning on our validation set. The encoder is fine-tuned using the AdamW optimizer, with a batch size of 32, a learning rate of  $2 \\times 10^{-5}$ , and a maximum of 5 epochs.",
+        "bbox": [
+            112,
+            133,
+            487,
+            243
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "For LLM fine-tuning, we employ LoRA (Hu et al., 2022) for efficient adaptation. The core hyperparameters are as follows: 3 training epochs, batch size of 128, AdamW optimizer, maximum learning rate of  $5 \\times 10^{-5}$ ,  $1\\%$  warmup steps, cosine learning rate scheduler, LoRA rank of 8, LoRA alpha of 16, and LoRA dropout rate of 0.1.",
+        "bbox": [
+            112,
+            244,
+            487,
+            356
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "All hyperparameters and model checkpoints are selected to maximize the F1 score on the validation set. For further details, please refer to our codebase: https://github.com/SuYindu/TACLR.",
+        "bbox": [
+            112,
+            359,
+            487,
+            420
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Deployment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            432,
+            257,
+            450
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The proposed TACLR has been successfully integrated into key functionalities of the e-commerce platform Xianyu, including product listing, search, recommendation, and price estimation. The system is designed to be highly scalable, efficiently processing millions of products daily.",
+        "bbox": [
+            112,
+            458,
+            487,
+            554
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "During the product listing process, TACLR automatically identifies attribute-value pairs from user-provided titles and descriptions. This automation significantly reduces manual effort and errors, while enhancing the quality of structured product information.",
+        "bbox": [
+            112,
+            556,
+            487,
+            649
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "For product search, the improved structured information directly supports more effective lexical retrieval and enables structured search, where items are filtered based on attributes. This leads to more accurate matching with user queries. In addition, the enriched product features enhance the quality of personalized recommendations.",
+        "bbox": [
+            112,
+            653,
+            487,
+            763
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "In the context of price estimation, TACLR identifies key attributes that influence pricing, enabling more accurate price predictions. This functionality provides sellers and buyers with reliable, market-aligned information in the context of second-hand transactions.",
+        "bbox": [
+            112,
+            766,
+            487,
+            859
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "31538",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2025/TACLR_ A Scalable and Efficient Retrieval-based Method for Industrial Product Attribute Value Identification/939f3cd5-94e9-46cf-8def-cb09a1fd834c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9870e4c422db5ca147dfdef7cf9030ebb1c7f0a69f5fad2c86f1cdb43510579
+size 1252476

2025/TACLR_ A Scalable and Efficient Retrieval-based Method for Industrial Product Attribute Value Identification/full.md

@@ -0,0 +1,341 @@
+# TACLR: A Scalable and Efficient Retrieval-Based Method for Industrial Product Attribute Value Identification
+
+Yindu Su $^{1,2}$ , Huike Zou $^{1}$ , Lin Sun $^{3}$ , Ting Zhang $^{2}$ , Haiyang Yang $^{1}$ , Liu Chen $^{1}$ , David Lo $^{2}$ , Qingheng Zhang $^{1}$ , Shuguang Han $^{1}$ , Jufeng Chen $^{1}$ ,
+
+$^{1}$ Xianyu of Alibaba  $^{2}$ Singapore Management University  $^{3}$ Hangzhou City University
+
+yindusu@foxmail.com, tingzhang.2019@phdcs.smu.edu.sg
+
+qingheng.zqh@alibaba-inc.com，shuguang.sh@alibaba-inc.com
+
+# Abstract
+
+Product Attribute Value Identification (PAVI) involves identifying attribute values from product profiles, a key task for improving product search, recommendation, and business analytics on e-commerce platforms. However, existing PAVI methods face critical challenges, such as inferring implicit values, handling out-of-distribution (OOD) values, and producing normalized outputs. To address these limitations, we introduce Taxonomy-Aware Contrastive Learning Retrieval (TACLR), the first retrieval-based method for PAVI. TACLR formulates PAVI as an information retrieval task by encoding product profiles and candidate values into embeddings and retrieving values based on their similarity. It leverages contrastive training with taxonomy-aware hard negative sampling and employs adaptive inference with dynamic thresholds. TACLR offers three key advantages: (1) it effectively handles implicit and OOD values while producing normalized outputs; (2) it scales to thousands of categories, tens of thousands of attributes, and millions of values; and (3) it supports efficient inference for high-load industrial deployment. Extensive experiments on proprietary and public datasets validate the effectiveness and efficiency of TACLR. Further, it has been successfully deployed on the real-world e-commerce platform Xianyu, processing millions of product listings daily with frequently updated, large-scale attribute taxonomies. We release the code to facilitate reproducibility and future research at https://github.com/SuYindu/TACLR.
+
+# 1 Introduction
+
+Product attribute values are key components that support the operation of e-commerce platforms. They provide essential structural information, aiding customers in making informed purchasing decisions and enabling product listing (Chen et al., 2024), recommendation (Truong et al., 2022; Sun et al., 2020), retrieval (Magnani et al., 2019; Huang
+
+![](images/2a3df4355fa37b992a862a63ad8fedfc58e8f8c71ba29f228a10c8b5d1373f3b.jpg)  
+Figure 1: Illustration of the PAVE and PAVI tasks. Unlike PAVE, which extracts raw value spans from product profiles, PAVI requires outputs to be normalized and supports both the identification of implicit values and the assignment of null to unavailable attribute values.
+
+et al., 2014), and question answering (Kulkarni et al., 2019; Gao et al., 2019).
+
+However, seller-provided attribute values are often incomplete or even inaccurate—an issue that is particularly severe on second-hand e-commerce platforms such as  $Xianyu^{1}$ . This undermines the effectiveness of downstream applications, making the automatic identification of product attribute values a fundamental requirement. Researchers have explored Product Attribute Value Extraction (PAVE), which involves extracting spans from product profiles using Named Entity Recognition (NER) (Zheng et al., 2018) or Question Answering (QA) (Wang et al., 2020) models. The upper part of Figure 1 illustrates an example of NER-based PAVE.
+
+Although these approaches effectively extract value spans, the outputs remain raw text subsequences. Presenting attribute values in a standardized format is crucial for facilitating data aggrega
+
+tion in business analytics and enhancing the user experience by providing clear and consistent information. To produce standardized values, a normalization step (Putthividhya and Hu, 2011; Zhang et al., 2021) is required to map these spans to predefined formats, as shown in the lower part of Figure 1. However, implicit values, such as Apple, cannot be directly extracted and must instead be inferred from context or prior knowledge.
+
+Therefore, in this work, we focus on Product Attribute Value Identification (PAVI) (Shinzato et al., 2023), which aims to associate predefined attribute values from attribute taxonomy (illustrated in Figure 2) with products. The input for PAVI includes the product category and profile, where the profile includes textual data, such as the title and description, and may optionally incorporate visual information, such as images or videos. The output is a dictionary with predefined attributes as keys and their inferred values as corresponding entries. In addition, PAVI requires determining when attribute values are missing. For instance, as shown in Figure 1, the value for Version is unavailable and is therefore assigned an empty result or null value.
+
+Beyond adapting PAVE approaches, researchers have investigated classification-based (Chen et al., 2022) and generation-based paradigms (Sabeh et al., 2024b) for PAVI. Classification-based methods treat each value as a class; while this approach is straightforward, it is fundamentally limited by the inability to identify out-of-distribution (OOD) values not present in the training data, making such methods impractical for the continuously evolving nature of e-commerce platforms. In contrast, generation-based methods reformulate PAVI as a sequence-to-sequence task. Although these methods can handle implicit and OOD values, they face significant challenges, such as generating uncontrollable outputs and incurring substantial computational costs in high-load scenarios due to their reliance on Large Language Models (LLMs). In summary, existing approaches face distinct challenges, including difficulties in identifying implicit values, generalizing to OOD values, producing normalized outputs, and ensuring scalability and efficiency for large-scale industrial applications.
+
+To address these limitations, we propose a novel retrieval-based method, Taxonomy-Aware Contrastive Learning Retrieval (TACLR). Our approach formulates PAVI as an information retrieval task: the product item serves as the query, and the attribute taxonomy acts as the corpus, enabling
+
+![](images/d000e51ed45ea60b8795afc47bb3adc8796089f8fb316f60104054e705f50170.jpg)  
+Figure 2: An illustration of a portion of the attribute taxonomy. Each category, such as Phone, is linked to multiple attributes, including Brand, Model, and Capacity, with standardized values enumerated for each attribute (e.g., Apple, Huawei, and Samsung for Brand).
+
+efficient retrieval of relevant attribute values as matched documents. We use a shared encoder to generate embeddings for both the input product and candidate values from the attribute taxonomy. Our method adopts a contrastive learning framework inspired by CLIP (Radford et al., 2021). Rather than relying on in-batch negatives, we implement taxonomy-aware negative sampling, which selects hard negative values from the same category and attribute to generate a more challenging and precise contrastive signal. Additionally, learnable null values are explicitly incorporated as the ground truth for product-attribute pairs without associated values. During inference, we address the limitations of static thresholds by introducing dynamic thresholds derived from the relevance score of null values. This adaptive inference mechanism improves the model's ability to generalize across a large-scale attribute taxonomy.
+
+Our contributions are threefold: (1) We propose a novel retrieval-based paradigm for PAVI, introducing a scalable and efficient framework capable of handling implicit values, generalizing to OOD values, and producing normalized outputs. (2) We incorporate contrastive training into TACLR, using a taxonomy-aware negative sampling strategy to improve representation discrimination, and introduce an adaptive inference mechanism that dynamically balances precision and recall in large-scale industrial applications. (3) We validate the effectiveness of TACLR through extensive experiments on proprietary and public datasets, and demonstrate its successful deployment in a real-world industrial environment, processing millions of product listings across thousands of categories and millions of attribute values.
+
+# 2 Related Work
+
+# 2.1 Product Attribute Value Extraction
+
+PAVE as Named Entity Recognition. PAVE can be formulated as NER by identifying subsequences in product texts as entity spans and associating them with attributes as entity types. Early methods, such as OpenTag (Zheng et al., 2018), trained individual models for each category-attribute pair. Subsequent efforts generalized this approach to support multiple attributes or categories. For instance, SUOpenTag (Xu et al., 2019) incorporated attribute embeddings into an attention layer to handle multiple attributes, while AdaTag (Yan et al., 2021) used attribute embeddings to parameterize the decoder. TXtract (Karamanolakis et al., 2020) introduced a category encoder and a category attention mechanism to tackle various categories effectively. Additionally, M-JAVE (Zhu et al., 2020) jointly modeled attribute prediction and value extraction tasks while also incorporating visual information. More recently, Chen et al. (2023) scaled BERTNER by expanding the number of entity types to support a broader range of attributes.
+
+PAVE as Question Answering. The QA framework can also be adapted for PAVE by treating the product profile as context, attributes as questions, and value spans extracted from the context as answers. Wang et al. (2020) first introduced AVEQA for QA-based PAVE. Subsequent work extended this framework by incorporating multi-source information (Yang et al., 2022), multi-modal feature (Wang et al., 2022), and trainable prompts (Yang et al., 2023). Moreover, the question can be extended by appending candidate values as demonstrated by (Shinzato et al., 2022). Combining NER and QA paradigms, Ding et al. (2022) proposed a two-stage framework, which first identifies candidate values and then filters them.
+
+While NER- and QA-based paradigms have proven effective for PAVE, they struggle to identify implicit attribute values. Additionally, both paradigms rely on post-extraction normalization to standardize values, using either lexical (Putthividhya and Hu, 2011) or semantic methods (Zhang et al., 2021). Furthermore, extraction-based models require token-level annotations (e.g., BIO tags) for training and evaluation. Producing such annotations is significantly more resource-intensive than generating the value-level annotations used by TACLR, further limiting the scalability of these extraction-based methods.
+
+Table 1: Comparison of different paradigms for identifying implicit, OOD, and normalized values.  
+
+<table><tr><td>Paradigm</td><td>Implicit</td><td>OOD</td><td>Normalized</td></tr><tr><td>Extraction</td><td>X</td><td>✓</td><td>X</td></tr><tr><td>Classification</td><td>✓</td><td>X</td><td>✓</td></tr><tr><td>Generation</td><td>✓</td><td>✓</td><td>X</td></tr><tr><td>Retrieval</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+# 2.2 Product Attribute Value Identification
+
+Classification-Based PAVI. A straightforward approach is to frame PAVI as a multi-label classification problem over a finite set of values. Chen et al. (2022) trained a unified classification model that masks invalid labels based on the product category. However, a significant limitation of this classification-based paradigm is its inability to recognize OOD values not included in the training set. Consequently, its practicality is limited in dynamic e-commerce environments, where new categories and values frequently emerge.
+
+Generation-Based PAVI. Recent advancements in LLM have spurred the exploration of generation-based PAVI methods (Sabeh et al., 2024b). Some methods (Roy et al., 2021; Nikolakopoulos et al., 2023; Blume et al., 2023) construct attribute-aware prompts to generate values for each attribute individually. In contrast, others generate values for multiple attributes simultaneously, either in a linearized sequence format (Shinzato et al., 2023) or as a hierarchical tree structure (Li et al., 2023). Multimodal information has also been integrated into LLMs to identify implicit attribute values from product images (Lin et al., 2021; Khandelwal et al., 2023). More recently, Brinkmann et al. (2024) explored the use of LLMs for both the extraction and normalization of attribute values. Additionally, Zou et al. (2024) introduced the learning-by-comparison technique to reduce model confusion, and Sabeh et al. (2024a) investigated Retrieval-Augmented Generation (RAG) technologies for PAVI.
+
+Although generation-based methods can infer implicit and OOD attribute values from product profiles, they face several challenges in real-world scenarios. A key issue is the potential for the LLMs to produce uncontrollable or hallucinated outputs, a known limitation of LLMs (Huang et al., 2024). Additionally, these methods often rely on large, computationally intensive models to achieve strong performance, making them inefficient and costly for large-scale industrial deployment.
+
+![](images/c4a96c6d30ecc506333876c61811570c3dcc294f57f1c80df239ba8d3ab13cc9.jpg)  
+Figure 3: Overview of the training and inference process of TACLR, our retrieval-based method for the PAVI task. The left section illustrates contrastive training with taxonomy-aware negative sampling, while the right section demonstrates adaptive inference with pre-computed value embeddings.
+
+![](images/1d853a5f379894a50e542bf2833f9e7d16c22c18c3f8100744a99aaceccedbe3.jpg)
+
+# 3 Taxonomy-Aware Contrastive Learning Retrieval
+
+This section defines the PAVI task with an attribute taxonomy (§3.1) and presents our retrieval-based paradigm for PAVI (§3.2). We then detail the use of contrastive training with taxonomy-aware negative sampling (§3.3) and an adaptive inference mechanism with dynamic thresholds (§3.4). Figure 3 provides an overview of our approach TACLR.
+
+# 3.1 PAVI Task Definition
+
+PAVI is grounded in an attribute taxonomy that encompasses numerous product categories. For each category  $c$ , the taxonomy specifies a set of attributes  $\mathcal{A}_c = \{a_1, a_2, \ldots\}$  relevant to products in that category. For each attribute  $a \in \mathcal{A}_c$ , it provides a predefined set of standard values  $\mathcal{V}_a = \{v_1, v_2, \ldots\}$ . Figure 2 illustrates this structure<sup>2</sup>.
+
+For a given product item  $i$ , with its title  $t$  and description  $d$ , the item is assigned to a category  $c$  with associated attributes  $\mathcal{A}_c$ . The objective of the PAVI task is to identify a relevant set of values  $\mathcal{V}_a^+\subseteq \mathcal{V}_a$  for each attribute  $a\in \mathcal{A}_c$ . The set  $\mathcal{V}_a^+$  can take one of three forms: a singleton  $(\{v\})$ , multiple values  $(\{v_{1},v_{2},\ldots \})$ , or an empty set  $(\varnothing)$  if no information about  $a$  is available in the product profile. Notably, a standard value may not always appear verbatim as a text span in  $t$  or  $d$ ; it may instead be present in a paraphrased or synonymous form, which we refer to as an unnormalized value (e.g., the standard value iPhone 12 Pro Max expressed as iphone12pm in Figure 1). In other
+
+cases, a value may not be explicitly mentioned in the product profile but can be inferred from the context; these are referred to as implicit values (e.g., Apple in Figure 1).
+
+# 3.2 Retrieval-Based PAVI
+
+In a standard information retrieval setting, given a query, the objective is to retrieve a list of relevant documents from a corpus. Similarly, for PAVI, we treat the input item as the query and the attribute taxonomy as the corpus, aiming to retrieve relevant attribute values as the output documents.
+
+Encoding of items and values. We preprocess both item profiles and candidate values as textual inputs, utilizing a shared text encoder. For each item, we concatenate its title  $(t)$  and description  $(d)$  into a single input sentence formatted as: title: {title} description: {description}. Each candidate value is represented as a context-rich prompt, structured as: A {category} with {attribute} being {value}, e.g., A phone with brand being Apple. We explore the impact of various prompt templates in §5.3<sup>3</sup>.
+
+Inference pipeline. During the deployment process, all value embeddings within the attribute taxonomy are pre-computed offline and indexed using the Faiss library4. During online inference, each item is encoded into an embedding, which is then compared against groups of indexed candidate value embeddings for various attributes. For each attribute  $a\in \mathcal{A}_c$  , the top-  $k$  most similar values are retrieved whose similarity scores exceed an adaptive threshold (3.4).
+
+# 3.3 Contrastive Training
+
+Inspired by CLIP (Radford et al., 2021), we employ contrastive learning to train the shared encoder. Rather than relying on in-batch negatives, we compare each positive value with hard negative values from the same category and attribute in the taxonomy, providing a more challenging and precise training signal.
+
+Formally, the subset of values matched with the item is referred to as the ground truth value set,  $\mathcal{V}_a^+ \subseteq \mathcal{V}_a$ . If no matched values exist for a given attribute, i.e.,  $\mathcal{V}_a^+ = \emptyset$ , we assign a specific null value  $v_0^a$  for this attribute as the positive value, i.e.  $v_a^+ = v_0^a$ . Otherwise, a positive value is randomly drawn from the ground truth value set, i.e.  $v_a^+ \sim \mathcal{V}_a^+$ . For negative sampling, we select values as  $\mathcal{V}_a^- = \{v_1^-, v_2^-, \ldots\} \subseteq \mathcal{V}_a - \mathcal{V}_a^+$ , ensuring a maximum of  $K$  values. The contrastive loss is then computed as follows:
+
+$$
+\mathcal {L} _ {a} = - \log \left(\frac {\exp (\frac {s (i , v _ {a} ^ {+})}{\tau})}{\exp (\frac {s (i , v _ {a} ^ {+})}{\tau}) + \sum_ {v \in \mathcal {V} _ {a} ^ {-}} \exp (\frac {s (i , v)}{\tau})}\right)
+$$
+
+where  $s(i,v) = \frac{I\cdot V}{\|I\|\|V\|}$  denotes the cosine similarity between the item embedding  $I$  and the value embedding  $V$ , and  $\tau$  is the temperature hyperparameter. It is important to note that each item typically includes multiple attributes, all of which share the same item embedding  $I$  while being individually compared against corresponding values. Therefore, the loss for item  $i$  is the sum of losses over all attributes from  $\mathcal{A}_c$ :
+
+$$
+\mathcal {L} _ {i} = \sum_ {a \in \mathcal {A} _ {c}} \mathcal {L} _ {a}
+$$
+
+An example logit matrix is depicted on the left side of Figure 3. Note that the item embedding  $I_{1}$  contributes to the loss computations of  $\mathcal{L}_1^1$ ,  $\mathcal{L}_1^2$ , and  $\mathcal{L}_1^3$ , which correspond to the attributes  $a_1, a_2$ , and  $a_3$  within the same product category. We also pad the logit matrix with negative infinity for batched computation if fewer than  $K$  values are available.
+
+# 3.4 Adaptive Inference
+
+During retrieval, relevance scores are assigned to every candidate values. To filter output values, a static threshold  $T$  can be applied to these scores. However, in real-world e-commerce platforms with a vast number of category-attribute pairs, using a single threshold across all pairs is often suboptimal.
+
+Moreover, defining a unique threshold for each pair is tedious or even impractical.
+
+To address this, we introduce an adaptive inference method that uses dynamic thresholds to make cutoff decisions. As discussed in §3.3, we add an explicit null value  $v_0^a$  for each category-attribute pair, with its embedding learned during training. In the inference phase, we compute the similarity  $s(i,v_0^a)$  between the item and the null value, using it as a dynamic threshold  $T_{a}^{\prime}$  to exclude candidate values for attribute  $a$  that have lower scores:
+
+$$
+\mathcal {V} _ {a} ^ {\mathrm {p r e d}} = \left\{v \mid s (i, v) > T _ {a} ^ {\prime} \right\}
+$$
+
+Since most category-attribute pairs have exclusive values, meaning that each product can have at most one value for a given attribute, we focus on the top-1 predicted value in this work. The output can be further simplified as follows:
+
+$$
+v _ {a} ^ {\text {p r e d}} = \left\{ \begin{array}{l l} \arg \max  _ {v \in \mathcal {V} _ {a}} s (i, v) & \text {i f} \max  _ {v \in \mathcal {V} _ {a}} s (i, v) > T _ {a} ^ {\prime} \\ n u l l & \text {o t h e r w i s e} \end{array} \right.
+$$
+
+Equivalently, we can include the null value as an explicit candidate and retrieve the top-1 value as:
+
+$$
+v _ {a} ^ {\text {p r e d}} = \arg \max  _ {v \in \mathcal {V} _ {a} \cup \{v _ {0} ^ {a} \}} s (i, v)
+$$
+
+In this case, selecting the  $v_{0}^{a}$  corresponds to the scenario where none of the specific candidate values surpass the dynamic threshold.
+
+The inference process is illustrated on the right side of Figure 3. In this example, the predictions for  $a_3$  and  $a_5$  are determined to be empty because the highest-scoring value for these attributes is the null value.
+
+# 4 Experiment Settings
+
+# 4.1 Datasets
+
+To evaluate PAVI under the settings described in §3.1, we compare our proposed method against baselines on both proprietary and public datasets with normalized values<sup>5</sup>. Table 2 presents statistics of the attribute taxonomies and datasets.
+
+Xianyu-PAVI. This dataset, derived from the e-commerce platform Xianyu, is constructed to evaluate the scalability and generalization of PAVI methods. The platform's attribute taxonomy comprises
+
+Table 2: Statistics of the attribute taxonomies and dataset splits from Xianyu-PAVI and WDC-PAVE. "CA Pairs" refers to category-attribute pairs, "CAV Tuples" denotes category-attribute-value tuples, "PA Pairs" represents product-attribute pairs, and "Null Pairs" indicates product-attribute pairs with null values. "Excl." refers to the test split excluding measurement attributes.  
+(a) Statistics of the attribute taxonomies.  
+
+<table><tr><td>Statistic</td><td>Xianyu</td><td>WDC</td></tr><tr><td># Categories</td><td>8,803</td><td>5</td></tr><tr><td># Attributes</td><td>3,326</td><td>24</td></tr><tr><td># CA Pairs</td><td>26,645</td><td>37</td></tr><tr><td># CAV Tuples</td><td>6,302,220</td><td>2,297</td></tr></table>
+
+(b) Statistics of the datasets.  
+
+<table><tr><td rowspan="2">Statistic</td><td colspan="3">Xianyu-PAVI</td><td colspan="3">WDC-PAVE</td></tr><tr><td>Train</td><td>Valid</td><td>Test</td><td>Train</td><td>Test</td><td>Excl.</td></tr><tr><td># Products</td><td>809,528</td><td>81,699</td><td>85,024</td><td>1,066</td><td>354</td><td>354</td></tr><tr><td># PA Pairs</td><td>3,584,462</td><td>358,582</td><td>458,954</td><td>8,832</td><td>2,937</td><td>2,285</td></tr><tr><td># Null Pairs</td><td>2,345,577</td><td>228,534</td><td>272,285</td><td>3,973</td><td>1,330</td><td>916</td></tr></table>
+
+8,803 product categories, 26,645 category-attribute pairs, and 6.3 million category-attribute-value tuples. On average, each category is associated with 3 attributes, each attribute has 237 possible values, and there are 716 candidate values per category.
+
+For our experiments, we randomly sampled 1 million product items for training, 10,000 for validation, and 10,000 for testing. Each item is annotated with a category label and multiple attribute-value pair labels. Category labels were generated through a multi-step process involving automated classification, seller feedback, and annotator review, during which misclassified samples were discarded. Attribute-value pair labels were obtained through a multi-stage manual annotation process: a pool of annotators conducted the initial labeling, followed by quality checks and a second round of review, with items reassigned to different annotators after shuffling.
+
+WDC-PAVE (Brinkmann et al., 2024). This dataset contains 1,066 training and 354 test product items across 5 categories, with 8,832 and 2,937 product attribute pairs (3,973 and 1,330 nulls), respectively. On average, each category is associated with 7.4 attributes, each attribute has 62 values, and there are 459 attribute-value pairs per category. We conduct two evaluations: the first on the original test set, and the second on a test split that excludes measurement attributes, which require complex reasoning for unit conversion.
+
+# 4.2 Metrics
+
+Since most attributes in the taxonomy are exclusive, i.e., each product can have at most one value per attribute, we evaluate PAVI methods using microaveraged precision@1, recall@1, and F1 score@1.
+
+For each attribute, the ground truth is a set of values  $\mathcal{V}$  from the taxonomy. If the ground truth set is empty  $(\varnothing)$ , a correct prediction (True Negative, TN) occurs when the model also predicts an empty
+
+Table 3: Confusion matrix comparing ground truth value set with predicted top-1 value.  
+
+<table><tr><td>Label</td><td>Prediction</td><td>Outcome</td></tr><tr><td>∅</td><td>∅</td><td>True Negative (TN)</td></tr><tr><td>∅</td><td>v</td><td>False Positive (FP)</td></tr><tr><td>V</td><td>v ∈ V</td><td>True Positive (TP)</td></tr><tr><td>V</td><td>∅</td><td>False Negative (FN)</td></tr><tr><td>V</td><td>v&#x27;∉ V</td><td>FP &amp; FN</td></tr></table>
+
+set; otherwise, it is a False Positive (FP). When the ground truth set is not empty, the model's top-1 output is a True Positive (TP) if it matches any ground truth value. Predicting an empty set in this case results in a False Negative (FN), while mismatched predictions are both False Positives (FP) and False Negatives (FN), as it simultaneously introduces an error and misses the correct value. Table 3 summarizes these outcomes<sup>6</sup>. Final precision, recall, and F1 scores are computed by aggregating TP, FP, and FN counts across the dataset, providing a comprehensive performance evaluation.
+
+# 4.3 Baselines
+
+We evaluate our retrieval-based method, TACLR, against classification and generation baselines<sup>7</sup>. For implementation details, refer to Appendix A. BERT-CLS. This baseline frames PAVI as a multi-label classification task, treating each category-attribute-value tuple as an independent label. The model is fine-tuned to predict matches, with label masking applied to exclude irrelevant labels for each category, following (Chen et al., 2022). The model outputs a probability distribution over val
+
+Table 4: Performance comparison of classification-, generation-, and retrieval-based methods on Xianyu-PAVI and WDC-PAVE. "F1 Excl." denotes the F1 score computed excluding measurement attributes (e.g., width and height), which require complex unit normalization reasoning.  
+
+<table><tr><td rowspan="2">Paradigm</td><td rowspan="2">Method</td><td colspan="3">Xianyu-PAVI</td><td colspan="4">WDC-PAVE</td></tr><tr><td>Precision</td><td>Recall</td><td>F1</td><td>Precision</td><td>Recall</td><td>F1</td><td>F1 Excl.</td></tr><tr><td>Classification</td><td>BERT-CLS</td><td>50.9</td><td>50.1</td><td>50.5</td><td>68.9</td><td>12.0</td><td>20.5</td><td>23.4</td></tr><tr><td rowspan="8">Generation</td><td>Llama3.1 (zero-shot)</td><td>29.1</td><td>46.2</td><td>35.7</td><td>56.6</td><td>60.8</td><td>58.6</td><td>64.6</td></tr><tr><td>Llama3.1 (few-shot)</td><td>31.0</td><td>51.1</td><td>38.6</td><td>76.0</td><td>74.1</td><td>75.0</td><td>79.0</td></tr><tr><td>Llama3.1 (RAG)</td><td>40.8</td><td>57.2</td><td>47.6</td><td>78.2</td><td>76.3</td><td>77.2</td><td>80.1</td></tr><tr><td>Llama3.1 (fine-tune)</td><td>86.9</td><td>82.7</td><td>84.7</td><td>57.7</td><td>60.4</td><td>59.0</td><td>64.5</td></tr><tr><td>Qwen2.5 (zero-shot)</td><td>42.7</td><td>55.7</td><td>48.4</td><td>51.9</td><td>60.3</td><td>55.8</td><td>60.8</td></tr><tr><td>Qwen2.5 (few-shot)</td><td>45.8</td><td>58.6</td><td>51.4</td><td>72.2</td><td>72.3</td><td>72.2</td><td>76.2</td></tr><tr><td>Qwen2.5 (RAG)</td><td>58.3</td><td>69.1</td><td>63.2</td><td>75.1</td><td>73.4</td><td>74.2</td><td>78.3</td></tr><tr><td>Qwen2.5 (fine-tune)</td><td>84.5</td><td>79.1</td><td>81.7</td><td>54.1</td><td>60.0</td><td>56.9</td><td>61.7</td></tr><tr><td>Retrieval</td><td>TACLR</td><td>85.4</td><td>87.1</td><td>86.2</td><td>74.3</td><td>70.9</td><td>72.6</td><td>80.3</td></tr></table>
+
+ues and selects the highest-probability value for each attribute. If no probability exceeds a specified threshold, the prediction is set to empty.
+
+LLM. For generation-based baselines, we utilize state-of-the-art open-source LLMs, including Llama3.1-7B (Llama Team, 2024) and Qwen2.5-7B (Qwen Team, 2024). These models are evaluated in zero-shot and few-shot settings using a template adapted from (Brinkmann et al., 2024), which incorporates the category, attribute, product profile, and detailed value normalization guidelines. We further fine-tune the LLMs using LoRA (Hu et al., 2022) to improve performance.
+
+RAG. We implement RAG baselines based on both Qwen2.5 and Llama3.1, using BGE embeddings (Xiao et al., 2023) to retrieve the top-5 most similar product items from the training set. The LLM prompt is augmented with the profiles and structured attribute-value pair outputs of these retrieved examples. For all generation baselines, we apply greedy decoding to ensure reproducibility, and model outputs are formatted in JSON.
+
+# 5 Results
+
+# 5.1 Main Results
+
+Table 4 presents the performance comparison between our retrieval-based method TACLR and classification- and generation-based baselines. On Xianyu-PAVI, TACLR achieves the highest F1 score of  $86.2\%$ , surpassing fine-tuned Llama3.1, which obtains  $84.7\%$ . Notably, TACLR excels in recall, achieving  $87.1\%$  compared to Llama3.1's  $82.7\%$ . On WDC-PAVE, TACLR achieves the highest F1 Excl. score of  $80.3\%$ , which excludes measurement attributes requiring unit normalization
+
+reasoning. These results highlight TACLR's effectiveness and robustness in addressing general PAVI across diverse datasets.
+
+The classification-based baseline, BERT-CLS, shows the weakest performance on both datasets. It achieves an F1 score of  $50.5\%$  on Xianyu-PAVI, but its performance drops drastically to  $20.5\%$  on WDC-PAVE, underscoring the limitations of classification approaches in generalization. One contributing factor is the extreme label sparsity in this formulation: there are over 6.3M category-attribute-value labels, but the training set contains fewer than 3.6M instances, making it difficult for the model to learn a reliable classification head.
+
+Among generation-based methods, performance improves steadily from zero-shot prompting to few-shot prompting, retrieval-augmented generation, and fine-tuning on the large-scale Xianyu-PAVI dataset. Llama3.1 progresses from  $35.7\%$  (zero-shot) to  $38.6\%$  (few-shot),  $47.6\%$  (RAG), and  $84.7\%$  (fine-tune), while Qwen2.5 improves from  $48.4\%$  to  $51.4\%$ ,  $63.2\%$ , and  $81.7\%$ , respectively. A similar trend holds on WDC-PAVE, though fine-tuning is less effective due to limited supervision. In this setting, RAG outperforms few-shot prompting for both Llama3.1 ( $77.2\%$  vs.  $75.0\%$ ) and Qwen2.5 ( $74.2\%$  vs.  $72.2\%$ ), offering a viable alternative when labeled data is limited. Nonetheless, TACLR consistently surpasses RAG, particularly on Xianyu-PAVI ( $86.2\%$  vs.  $47.6\%/63.2\%$ ), demonstrating better scalability for large-scale industrial applications. On WDC-PAVE, TACLR matches or exceeds RAG and retains a slight advantage in non-measurement attributes (F1 Excl.:  $80.3\%$  vs.  $80.1\%/78.3\%$ ).
+
+Table 5: Inference efficiency comparison on Xianyu-PAVI (Throughput in samples/second).  
+
+<table><tr><td>Method</td><td>Time (ms)</td><td>Throughput</td></tr><tr><td>BERT-CLS</td><td>8.6</td><td>930</td></tr><tr><td>Llama3.1 (zero-shot)</td><td>101.3</td><td>80</td></tr><tr><td>Llama3.1 (few-shot)</td><td>124.8</td><td>64</td></tr><tr><td>Llama3.1 (RAG)</td><td>137.9</td><td>58</td></tr><tr><td>Qwen2.5 (zero-shot)</td><td>84.0</td><td>95</td></tr><tr><td>Qwen2.5 (few-shot)</td><td>98.4</td><td>81</td></tr><tr><td>Qwen2.5 (RAG)</td><td>108.3</td><td>74</td></tr><tr><td>TACLR</td><td>12.7</td><td>630</td></tr></table>
+
+Table 6: F1 scores on product-attribute pairs with normalized vs. unnormalized and implicit values.  
+
+<table><tr><td>Method</td><td>Normalized</td><td>Unnorm. &amp; Implicit</td></tr><tr><td>Llama3.1</td><td>83.2</td><td>79.4</td></tr><tr><td>Qwen2.5</td><td>82.6</td><td>78.6</td></tr><tr><td>TACLR</td><td>87.9</td><td>82.9</td></tr></table>
+
+# 5.2 Inference Efficiency
+
+Table 5 presents a comparison of inference efficiency across different PAVI paradigms under identical conditions, using an unoptimized PyTorch implementation on a single NVIDIA V100 GPU.
+
+TACLR achieves a strong balance between model capacity and efficiency, processing each sample in  $12.7\mathrm{ms}$  and achieving a throughput of 630 samples per second. In contrast, generation-based methods using Llama3.1 and Qwen2.5 exhibit substantially higher inference times (over  $100\mathrm{ms}$  per sample) and lower throughput (below 100 samples per second), primarily due to the overhead of autoregressive decoding and reliance on large-scale models. The BERT-CLS classification baseline offers the highest raw efficiency  $(8.6\mathrm{ms},$  930 samples per second), but its inability to handle out-of-distribution values and limited generalization capacity reduce its practical applicability.
+
+In summary, among these methods, TACLR provides the best trade-off between identification capability and inference efficiency, making it well-suited for scalable industrial deployment.
+
+# 5.3 Analysis
+
+This section analyzes our method on the Xianyu-PAVI dataset, selected for its large scale and diverse attribute taxonomy. In contrast, the WDC-PAVE dataset is substantially smaller in both size and coverage, making it insufficient for robust evaluation of scalability and generalization.
+
+![](images/110c7d7373481da84de0515654173d3e760e0f586de458ce05e148a74a873920.jpg)  
+Figure 4: Comparison of negative sampling strategies with increasing number of samples.
+
+Robustness Evaluation on Diverse Values. To better assess model robustness under various conditions, we partition the test set into three subsets based on the nature of the ground truth: (1) normalized values, appear verbatim in the product profile; (2) null values, where the attribute is marked as unavailable; and (3) unnormalized or implicit values, which either appear in a lexically varied form, or must be inferred. In our Xianyu-PAVI dataset, normalized values account for approximately  $15.2\%$ , unnormalized or implicit values for  $27.2\%$ , and null values for  $57.6\%$ .
+
+Table 6 reports the F1 scores on the normalized and unnormalized/implicit subsets. TACLR achieves the highest performance on both subsets, with an F1 score of  $87.9\%$  on normalized values and  $82.9\%$  on unnormalized or implicit values. Compared to the fine-tuned LLM baselines, TACLR demonstrates stronger ability to recognize standardized values and better robustness to lexical variation and implicit reasoning.
+
+Impact of Taxonomy-Aware Negative Sampling. Figure 4 compares the proposed taxonomy-aware negative sampling (§3.3) with in-batch negative sampling across varying sample sizes. As the number of sampled values increases, the F1 score improves consistently, aligning with findings from (Chen et al., 2020). Using in-batch sampling as the baseline, the model achieves an F1 score of  $53.3\%$  with a sample size of 128. In contrast, taxonomy-aware sampling yields substantial improvements, boosting the F1 score from  $84.0\%$  to  $86.2\%$  as the sample size grows from 16 to 128.
+
+These results demonstrate that taxonomy-aware sampling provides more effective supervision, encouraging the model to distinguish between fine-grained, semantically similar values within the same category and attribute. In contrast, in-batch sampling often yields random negatives that are less relevant and frequently trivial, limiting the efficacy of contrastive learning for PAVI. For example,
+
+![](images/3db2449ecd459fc902de0035a6c462124b8ceafa2409d8df5b1468c7a031adcc.jpg)  
+(a) Comparison of inference thresholds.
+
+![](images/6501c664b39e2dc3593dde2ff818a5aa3d26d6d994642c74e858a119266379ba.jpg)  
+(b) Comparison of prompt templates.
+
+![](images/8ff9f6144c165ae41cebd84daf7487fc9dc8843d2e17d33bc01e1b10859eacdb.jpg)  
+(c) Comparison of data domains.  
+Figure 5: Performance analysis across inference thresholds, prompt templates, and data domains.
+
+in Figure 1, negatives such as iPhone 12 Pro or iPhone 13 Pro Max, which belong to the model attribute under the phone category, provide more challenging and informative supervision than random values such as L for T-shirt size.
+
+Comparison of Dynamic and Static Thresholds. Figure 5a compares our dynamic thresholding approach (§3.4) with static thresholds of 0.6, 0.65, and 0.7, selected via validation. The dynamic threshold achieves the highest F1 (86.2%), outperforming the static baselines (75.5%, 80.2%, and 78.2%). Static thresholds exhibit the typical precision-recall trade-off: higher thresholds increase precision (65.1%→84.5%) but reduce recall (90.0%→72.8%). In contrast, the dynamic threshold balances precision (85.4%) and recall (87.1%) without manual tuning. This gain stems from a scalable design: instead of relying on fixed, hand-tuned cutoffs per category-attribute pair, our method learns null value embeddings whose similarity scores act as adaptive thresholds.
+
+Performance Gains from Context-Rich Prompts. The influence of varying value prompt templates on the PAVI task is shown in Figure 5b. Using only the value as a prompt achieves an F1 score of  $83.2\%$ . Adding category information raises the F1 score to  $83.9\%$ , while incorporating attribute information further improves it to  $85.4\%$ . The most comprehensive template, combining category, attribute, and value information (i.e., A {category} with {attribute} being {value}), achieves the highest F1 score of  $86.2\%$ . These results are consistent with prior work (Radford et al., 2021), highlighting that context-rich prompts enhance the model's discriminative performance.
+
+Zero-Shot Generalization Across Data Domains. Figure 5c presents zero-shot transfer results on unseen categories and values. The in-domain split achieves an F1 score of  $88.7\%$ , while cross
+
+category and cross-value splits decline to  $80.2\%$  and  $78.2\%$ , respectively, reflecting the challenges of adapting to evolving attribute taxonomies in out-of-distribution domains. Despite this, TACLR maintains a strong overall F1 of  $86.2\%$ , demonstrating robust generalization. The model's generalization relies on the shared textual encoder's semantic understanding and the retrieval-based approach's capacity to leverage these embeddings for zero-shot matching. Such adaptability is critical for dynamic e-commerce platforms, where new products and attribute-value pairs continuously emerge, reducing the need for frequent retraining and lowering maintenance costs.
+
+# 6 Conclusion
+
+In this work, we present TACLR, a novel retrieval-based approach for PAVI. By formulating PAVI as an information retrieval problem, TACLR enables the inference of implicit values, generalization to OOD values, and the production of normalized outputs. Building on this framework, TACLR employs contrastive training with taxonomy-aware sampling and adaptive inference with dynamic thresholds to enhance retrieval performance and scalability.
+
+Comprehensive experiments on proprietary and public datasets demonstrated TACLR's superiority over classification- and generation-based baselines. Notably, TACLR achieved an F1 score of  $86.2\%$  on the large-scale Xianyu-PAVI dataset. Our efficiency analysis further highlighted its advantage, achieving significantly faster inference speeds than generation-based methods. Beyond these experimental results, TACLR has been successfully deployed on the real-world e-commerce platform Xianyu, processing millions of product listings daily and seamlessly adapting to dynamic attribute taxonomies, making it a practical solution for large-scale industrial applications.
+
+# 7 Limitations
+
+TACLR assumes access to a predefined attribute taxonomy, which serves as a foundation for accurate value identification. While this setup is realistic in many e-commerce platforms, where dedicated systems and teams maintain evolving taxonomies, it does require ongoing manual updates to incorporate new categories, attributes, and values. Automating product attribute mining remains an open area for future research (Ghani et al., 2006; Zhang et al., 2022; Xu et al., 2023).
+
+Another limitation is that TACLR currently operates on textual product profiles and does not incorporate multimodal information, such as images or videos. Multimodal inputs could provide complementary signals for attributes that are difficult to infer from text alone (e.g., color, material, or shape). Extending the method to support multimodal input could further improve its coverage and accuracy in practical applications.
+
+A further limitation of TACLR, as a retrieval-based approach, is its difficulty in handling measurement attributes that require unit conversion or numerical reasoning. To address this, future work could explore hybrid methods that combine TACLR's retrieval strengths with generative techniques to improve performance in such cases.
+
+# References
+
+Ansel Blume, Nasser Zalmout, Heng Ji, and Xian Li. 2023. Generative models for product attribute extraction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 575-585, Singapore. Association for Computational Linguistics.  
+Alexander Brinkmann, Nick Baumann, and Christian Bizer. 2024. Using llms for the extraction and normalization of product attribute values. In European Conference on Advances in Databases and Information Systems, pages 217-230. Springer.  
+Kang Chen, Qing Heng Zhang, Chengbao Lian, Yixin Ji, Xuwei Liu, Shuguang Han, Guoqiang Wu, Fei Huang, and Jufeng Chen. 2024. IPL: Leveraging multimodal large language models for intelligent product listing. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 697-711, Miami, Florida, US. Association for Computational Linguistics.  
+Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings
+
+of Machine Learning Research, pages 1597-1607.  
+PMLR.  
+Wei-Te Chen, Keiji Shinzato, Naoki Yoshinaga, and Yandi Xia. 2023. Does named entity recognition truly not scale up to real-world product attribute extraction? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 152-159, Singapore. Association for Computational Linguistics.  
+Wei-Te Chen, Yandi Xia, and Keiji Shinzato. 2022. Extreme multi-label classification with label masking for product attribute value extraction. In Proceedings of the Fifth Workshop on e-Commerce and NLP (EC-NLP 5), pages 134–140, Dublin, Ireland. Association for Computational Linguistics.  
+Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, Shijin Wang, and Guoping Hu. 2020. Revisiting pre-trained models for Chinese natural language processing. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 657-668, Online. Association for Computational Linguistics.  
+Yifan Ding, Yan Liang, Nasser Zalmout, Xian Li, Christan Grant, and Tim Weninger. 2022. Ask-and-verify: Span candidate generation and verification for attribute value extraction. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 110–110, Abu Dhabi, UAE. Association for Computational Linguistics.  
+Shen Gao, Zhaochun Ren, Yihong Zhao, Dongyan Zhao, Dawei Yin, and Rui Yan. 2019. Product-aware answer generation in e-commerce question-answering. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM '19, page 429-437, New York, NY, USA. Association for Computing Machinery.  
+Rayid Ghani, Katharina Probst, Yan Liu, Marko Krema, and Andrew Fano. 2006. Text mining for product attribute extraction. SIGKDD Explor. Newsl., 8(1):41-48.  
+Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.  
+Junshi Huang, Si Liu, Junliang Xing, Tao Mei, and Shuicheng Yan. 2014. *Circle & search: Attribute-aware shoe retrieval*. ACM Transactions on Multimedia Computing, Communications, and Applications, 11(1).
+
+Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2024. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst. Just Accepted.  
+Giannis Karamanolakis, Jun Ma, and Xin Luna Dong. 2020. TXtract: Taxonomy-aware knowledge extraction for thousands of product categories. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8489-8502, Online. Association for Computational Linguistics.  
+Anant Khandelwal, Happy Mittal, Shreyas Kulkarni, and Deepak Gupta. 2023. Large scale generative multimodal attribute extraction for E-commerce attributes. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track), pages 305-312, Toronto, Canada. Association for Computational Linguistics.  
+Ashish Kulkarni, Kartik Mehta, Shweta Garg, Vidit Bansal, Nikhil Rasiwasia, and Srinivasan Sengamedu. 2019. Productqna: Answering user questions on e-commerce product pages. In Companion Proceedings of The 2019 World Wide Web Conference, WWW '19, page 354-360, New York, NY, USA. Association for Computing Machinery.  
+Yanzeng Li, Bingcong Xue, Ruoyu Zhang, and Lei Zou. 2023. AtTGen: Attribute tree generation for real-world attribute joint extraction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2139-2152, Toronto, Canada. Association for Computational Linguistics.  
+Rongmei Lin, Xiang He, Jie Feng, Nasser Zalmout, Yan Liang, Li Xiong, and Xin Luna Dong. 2021. Pam: Understanding product images in cross product category attribute extraction. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD '21, page 3262-3270, New York, NY, USA. Association for Computing Machinery.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692.  
+Meta Llama Team. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Alessandro Magnani, Feng Liu, Min Xie, and Somnath Banerjee. 2019. Neural product retrieval at Walmart.com. In Companion Proceedings of The 2019 World Wide Web Conference, WWW '19, page 367-372, New York, NY, USA. Association for Computing Machinery.  
+Athanasios N. Nikolakopoulos, Swati Kaul, Siva Karthik Gade, Bella Dubrov, Umit Batur, and Suleiman Ali Khan. 2023. Sage: Structured
+
+attribute value generation for billion-scale product catalogs. Preprint, arXiv:2309.05920.  
+Duangmanee Putthividhya and Junling Hu. 2011. Bootstrapped named entity recognition for product attribute extraction. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1557-1567, Edinburgh, Scotland, UK. Association for Computational Linguistics.  
+Alibaba Qwen Team. 2024. Qwen2 technical report. Preprint, arXiv:2407.10671.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.  
+Kalyani Roy, Pawan Goyal, and Manish Pandey. 2021. Attribute value generation from product title using language models. In Proceedings of the 4th Workshop on e-Commerce and NLP, pages 13-17, Online. Association for Computational Linguistics.  
+Kassem Sabeh, Mouna Kacimi, Johann Gamper, Robert Litschko, and Barbara Plank. 2024a. Exploring large language models for product attribute value identification. Preprint, arXiv:2409.12695.  
+Kassem Sabeh, Robert Litschko, Mouna Kacimi, Barbara Plank, and Johann Gamper. 2024b. An empirical comparison of generative approaches for product attribute-value identification. Preprint, arXiv:2407.01137.  
+Keiji Shinzato, Naoki Yoshinaga, Yandi Xia, and Wei-Te Chen. 2022. Simple and effective knowledge-driven query expansion for QA-based product attribute extraction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 227-234, Dublin, Ireland. Association for Computational Linguistics.  
+Keiji Shinzato, Naoki Yoshinaga, Yandi Xia, and Wei-Te Chen. 2023. A unified generative approach to product attribute-value identification. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 6599–6612, Toronto, Canada. Association for Computational Linguistics.  
+Changfeng Sun, Han Liu, Meng Liu, Zhaochun Ren, Tian Gan, and Liqiang Nie. 2020. Lara: Attribute-to-feature adversarial learning for new-item recommendation. In Proceedings of the 13th International Conference on Web Search and Data Mining, WSDM '20, page 582-590, New York, NY, USA. Association for Computing Machinery.  
+Quoc-Tuan Truong, Tong Zhao, Changhe Yuan, Jin Li, Jim Chan, Soo-Min Pantel, and Hady W. Lauw. 2022.
+
+Ampsum: Adaptive multiple-product summarization towards improving recommendation captions. In Proceedings of the ACM Web Conference 2022, WWW '22, page 2978-2988, New York, NY, USA. Association for Computing Machinery.  
+Qifan Wang, Li Yang, Bhargav Kanagal, Sumit Sanghai, D. Sivakumar, Bin Shu, Zac Yu, and Jon Elsas. 2020. Learning to extract attribute value from product via question answering: A multi-task approach. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '20, page 47-55, New York, NY, USA. Association for Computing Machinery.  
+Qifan Wang, Li Yang, Jingang Wang, Jitin Krishnan, Bo Dai, Sinong Wang, Zenglin Xu, Madian Khabsa, and Hao Ma. 2022. SMARTAVE: Structured multimodal transformer for product attribute value extraction. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 263-276, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-pack: Packaged resources to advance general chinese embedding. Preprint, arXiv:2309.07597.  
+Huimin Xu, Wenting Wang, Xin Mao, Xinyu Jiang, and Man Lan. 2019. Scaling up open tagging from tens to thousands: Comprehension empowered attribute value extraction from product title. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5214-5223, Florence, Italy. Association for Computational Linguistics.  
+Liyan Xu, Chenwei Zhang, Xian Li, Jingbo Shang, and Jinho D. Choi. 2023. Towards open-world product attribute mining: A lightly-supervised approach. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12223-12239, Toronto, Canada. Association for Computational Linguistics.  
+Jun Yan, Nasser Zalmout, Yan Liang, Christian Grant, Xiang Ren, and Xin Luna Dong. 2021. AdaTag: Multi-attribute value extraction from product profiles with adaptive decoding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4694–4705, Online. Association for Computational Linguistics.  
+Li Yang, Qifan Wang, Jingang Wang, Xiaojun Quan, Fuli Feng, Yu Chen, Madian Khabsa, Sinong Wang, Zenglin Xu, and Dongfang Liu. 2023. MixPAVE: Mix-prompt tuning for few-shot product attribute value extraction. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9978-9991, Toronto, Canada. Association for Computational Linguistics.  
+Li Yang, Qifan Wang, Zac Yu, Anand Kulkarni, Sumit Sanghai, Bin Shu, Jon Elsas, and Bhargav Kanagal.
+
+2022. Mave: A product dataset for multi-source attribute value extraction. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, WSDM '22, page 1256-1265, New York, NY, USA. Association for Computing Machinery.  
+Danqing Zhang, Zheng Li, Tianyu Cao, Chen Luo, Tony Wu, Hanqing Lu, Yiwei Song, Bing Yin, Tuo Zhao, and Qiang Yang. 2021. Queaco: Borrowing treasures from weakly-labeled behavior data for query attribute value extraction. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, CIKM '21, page 4362-4372, New York, NY, USA. Association for Computing Machinery.  
+Xinyang Zhang, Chenwei Zhang, Xian Li, Xin Luna Dong, Jingbo Shang, Christos Faloutsos, and Jiawei Han. 2022. Oa-min: Open-world attribute mining for e-commerce products with weak supervision. In Proceedings of the ACM Web Conference 2022, WWW '22, page 3153-3161, New York, NY, USA. Association for Computing Machinery.  
+Guineng Zheng, Subhabrata Mukherjee, Xin Luna Dong, and Feifei Li. 2018. Opentag: Open attribute value extraction from product profiles. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '18, page 1049-1058, New York, NY, USA. Association for Computing Machinery.  
+Tiangang Zhu, Yue Wang, Haoran Li, Youzheng Wu, Xiaodong He, and Bowen Zhou. 2020. Multimodal joint attribute prediction and value extraction for E-commerce product. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2129-2139, Online. Association for Computational Linguistics.  
+Henry Zou, Gavin Yu, Ziwei Fan, Dan Bu, Han Liu, Peng Dai, Dongmei Jia, and Cornelia Caragea. 2024. EIVEN: Efficient implicit attribute value extraction using multimodal LLM. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 6: Industry Track), pages 453-463, Mexico City, Mexico. Association for Computational Linguistics.
+
+# A Implementation Details
+
+For both the BERT-CLS baseline and our TACLR method, we utilize pre-trained RoBERTa-base (Liu et al., 2019; Cui et al., 2020) as the backbone. For TACLR, we augment the model with a linear projection head to map the embedding dimension to 256. For each product-attribute pair, we sample up to 128 values for the contrastive learning setup, including a null value, an optional positive value (if present for the attribute), and negative values sampled from the value set with the same category and
+
+attribute. In the case of the null value, we replace the value slot in the prompt template with null (e.g., A phone with capacity being null.).
+
+The temperature parameter is fixed at 0.05. This choice is motivated by both prior research (e.g., SimCLR (Chen et al., 2020) and MoCo (He et al., 2020)) and empirical tuning on our validation set. The encoder is fine-tuned using the AdamW optimizer, with a batch size of 32, a learning rate of  $2 \times 10^{-5}$ , and a maximum of 5 epochs.
+
+For LLM fine-tuning, we employ LoRA (Hu et al., 2022) for efficient adaptation. The core hyperparameters are as follows: 3 training epochs, batch size of 128, AdamW optimizer, maximum learning rate of  $5 \times 10^{-5}$ ,  $1\%$  warmup steps, cosine learning rate scheduler, LoRA rank of 8, LoRA alpha of 16, and LoRA dropout rate of 0.1.
+
+All hyperparameters and model checkpoints are selected to maximize the F1 score on the validation set. For further details, please refer to our codebase: https://github.com/SuYindu/TACLR.
+
+# B Deployment
+
+The proposed TACLR has been successfully integrated into key functionalities of the e-commerce platform Xianyu, including product listing, search, recommendation, and price estimation. The system is designed to be highly scalable, efficiently processing millions of products daily.
+
+During the product listing process, TACLR automatically identifies attribute-value pairs from user-provided titles and descriptions. This automation significantly reduces manual effort and errors, while enhancing the quality of structured product information.
+
+For product search, the improved structured information directly supports more effective lexical retrieval and enables structured search, where items are filtered based on attributes. This leads to more accurate matching with user queries. In addition, the enriched product features enhance the quality of personalized recommendations.
+
+In the context of price estimation, TACLR identifies key attributes that influence pricing, enabling more accurate price predictions. This functionality provides sellers and buyers with reliable, market-aligned information in the context of second-hand transactions.

2025/TACLR_ A Scalable and Efficient Retrieval-based Method for Industrial Product Attribute Value Identification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e19c009a7a03a5b6f96d3da3311d28223c693cee7ecd09e026e4740eab003114
+size 390514

2025/TARGA_ Targeted Synthetic Data Generation for Practical Reasoning over Structured Data/9e5b56f3-c8f5-4eb7-be25-daa4af06471e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:091d481e638d0f03b16a473754e9d3bd02c6220b20a87992910222f44f89dc6c
+size 1089376

2025/TARGA_ Targeted Synthetic Data Generation for Practical Reasoning over Structured Data/full.md

@@ -0,0 +1,671 @@
+# TARGA: Targeted Synthetic Data Generation for Practical Reasoning over Structured Data
+
+Xiang Huang $^{1}$ , Jiayu Shen $^{1*}$ , Shanshan Huang $^{1}$ , Sitao Cheng $^{2}$ , Xiaxia Wang $^{3}$ , Yuzhong Qu $^{1}$
+
+<sup>1</sup>State Key Laboratory for Novel Software Technology, Nanjing University, China
+
+$^{2}$ University of California, Santa Barbara
+
+$^{3}$ University of Oxford
+
+xianghuang@smail.nju.edu.cn, yzqu@nju.edu.cn
+
+# Abstract
+
+Semantic parsing, which converts natural language questions into logic forms, plays a crucial role in reasoning within structured environments. However, existing methods encounter two significant challenges: reliance on extensive manually annotated datasets and limited generalization capability to unseen examples. To tackle these issues, we propose Targeted Synthetic Data Generation (TARGA), a practical framework that dynamically generates high-relevance synthetic data without manual annotation. Starting from the pertinent entities and relations of a given question, we probe for the potential relevant queries through layer-wise expansion and cross-layer combination. Then we generate corresponding natural language questions for these constructed queries to jointly serve as the synthetic demonstrations for in-context learning. Experiments on multiple knowledge base question answering (KBQA) datasets demonstrate that TARGA, using only a 7B-parameter model, substantially outperforms existing non-finetuned methods that utilize close-sourced model, achieving notable improvements in F1 scores on GrailQA (+7.7) and KBQA-Agent (+12.2). Furthermore, TARGA also exhibits superior sample efficiency, robustness, and generalization capabilities under non-I.I.D. settings.
+
+# 1 Introduction
+
+Reasoning over structured environments, such as Knowledge Base (KB), Database, and Web, has emerged as a crucial ability for large language models (LLMs) (Liu et al., 2024; Gu et al., 2023). Among various methods for structural reasoning, semantic parsing stands out as a mainstream and has garnered increasing attention from researchers. By translating natural language questions (NLQ) into logic forms, semantic parsing enables seamless interaction with structured environments, thereby enhancing user experience and accessibility.
+
+![](images/bb5917509ec42aaaf4d74ff84694c4f13a92efaa79e484ba551f0f7901493c89.jpg)  
+Figure 1: Compared with previous methods, TARGA aims to mitigate the reliance on large amounts of manually labeled data and enhance generalization capabilities in non-i.i.d. scenarios.
+
+However, current semantic paring methods typically face two significant challenges:
+
+1) Dependence on annotation. Previous methods usually rely on extensive amounts of manually annotated data. By training (Ye et al., 2022; Shu et al., 2022; Huang et al., 2023b) or retrieval (Li et al., 2023b; Nie et al., 2023) based on large-scale annotations, existing works have made remarkable progress. Unfortunately, collecting manual annotations in specific environments is labor-intensive and time-consuming. In real-world scenarios, a large pre-collected annotated dataset is often unavailable, limiting the scalability of these methods.  
+2) Limited generalization capability. Even with access to large annotated datasets, previous methods still struggle with generalizing to unseen examples. Regardless of the paradigm (e.g., in-context learning or fine-tuning), any method relying on a static, offline-collected dataset is inevitably influenced by the dataset's distribution. Specifically, these methods tend to perform well on examples encountered in the dataset (the I.I.D. setting) but exhibit weaker generalization when faced with unseen environmental items or query structures (the non-
+
+I.I.D. settings), as shown in Figure 1. In complex environments, such as Freebase (Bollacker et al., 2008) with over three billion triples, it is nearly impossible for a pre-collected static dataset to cover the full scope of the environment. Additionally, as the coverage of annotations increases, so do the costs associated with training and retrieval, which further limits the scalability and generalization.
+
+To address the aforementioned challenges, in this work, we propose a practical semantic parsing framework called Targeted Synthetic Data Generation (TARGA), which does not need any manually annotated data and can efficiently work on a 7B model. Specifically, TARGA addresses these challenges by dynamically synthesizing highly relevant examples of a test question as demonstrations for in-context learning. Starting from the KB items (entity, relation) that may be related to the given question, we construct logic forms through layer-wise expansion (extend a new edge for a sub-structure) and cross-layer combination (combine different sub-structures), gradually evolving from simple to complex structures. To further enhance relevance, we re-rank the synthetic logic forms to select the most pertinent ones and generate their corresponding natural language questions, which are then used as demonstrations for reasoning. Through this automatic data synthesis, TARGA free annotators from the heavy burden of labeling tasks. Additionally, the demonstrations are generated based on the given question, thus naturally avoiding the challenge of generalization.
+
+Without any data annotation, TARGA significantly outperforms all non-fine-tuned approaches across multiple complex KBQA datasets, particularly excelling in non-I.I.D. settings. Remarkably, TARGA achieves this with only a 7B-parameter model, whereas most baselines rely on advanced closed-source models, such as gpt-3.5-turbo, enabling faster and more cost-efficient inference. On the GrailQA dataset, we improve the performance of non-fine-tuned methods from an F1 score of 61.3 to 69.0. On KBQA-Agent, the most challenging dataset, we elevate the SOTA performance from 34.3 to 46.5 F1 scores. Further analyses highlight the high quality of the data generated by TARGA. Even with a single demonstration, TARGA still surpasses all non-fine-tuned methods on GrailQA. Additionally, TARGA exhibits remarkable robustness in adversarial settings<sup>1</sup>.
+
+# 2 Related Works
+
+# 2.1 Few-shot KBQA with LLMs
+
+With the advancement of large language models, recent works have adopted LLMs as the backend for KBQA (Cheng et al., 2024). In particular, In-Context Learning (Brown et al., 2020) requires dozens of demonstrations to guide the model's responses. To achieve competitive performance, existing ICL-based KBQA works (Li et al., 2023b; Nie et al., 2023) typically retrieve the most similar examples from a manually annotated training set as demonstrations. However, this strategy often results in performance degradation on non-I.I.D. questions involving unseen structures or KB items. For example, Nie et al. (2023); Li et al. (2023b) reported that the performance in zero-shot settings can be up to  $20\%$  lower compared to I.I.D. settings.
+
+Another line of KBQA methods, agent-based methods (Liu et al., 2024; Huang et al., 2024; Gu et al., 2024), decomposes questions into individual steps to solve. While step-by-step solving aligns with human intuition and demonstrates remarkable generalization ability, it incurs high computational costs and presents challenges in constructing trajectories. Moreover, the effectiveness of the agent-based paradigm relies heavily on the planning and generalization abilities of advanced LLMs, leading to subpar performance when using weaker models, such as some open-source variants. Such dependency underscores the limitation of agent-based approaches when superior LLMs are unavailable or impractical to use due to resource constraints.
+
+# 2.2 Synthetic Data Generation
+
+Instead of relying solely on human annotation for training data, recent works have leveraged LLMs to generate synthetic data, thereby reducing the burden on human annotators. For instance, Chiang et al. (2023); Taori et al. (2023) use instructions to generate training data as a supplement to manual annotation via self-instruct techniques (Wang et al., 2023). However, this approach still requires human-annotated seed examples to ensure high-quality demonstrations, which entails significant demand for LLM usage. Rather than directly prompting LLMs to generate training data, Cao et al. (2022); Huang et al. (2023a) address this problem by first sampling structured queries from the environment and then converting these queries into natural language using LLMs. Nevertheless, obtaining meaningful structured queries remains a non-trivial task.
+
+Other works similar to ours include FlexKBQA (Li et al., 2023c) and BYOKG (Agarwal et al., 2024). FlexKBQA relies on predefined templates and model training stages to automatically annotate data, while BYOKG synthesizes data from scratch. However, both approaches require a time-consuming offline data collection phase. More importantly, like other methods, they rely on reasoning over a pre-collected static dataset, which still suffers from generalization issues. In TARGA, we systematically design a framework to dynamically synthesize relevant examples in an online manner. It avoids the need for a lengthy data collection process, enabling us to dynamically obtain the most relevant examples for each test case without being constrained by a static dataset.
+
+# 3 Methods
+
+# 3.1 Overview
+
+As shown in Figure 2, our targeted data synthesis framework, TARGA, consists of four parts. Given a natural language question  $nlq$ , we first collect candidate KB items such as entities  $E_{nlq}$  and relations  $R_{nlq}$  as initialization. Then, we explore valid query graphs from simple to complex structures to construct synthetic queries  $Q$ . Next, we filter high-relevance candidate queries by ranking. Finally, we use these high-quality synthetic data as demonstrations for QA reasoning.
+
+# 3.2 Candidate KB Items Retrieval
+
+For candidate entities, we adopt the linking result provided by Gu et al. (2023) for a fair comparison (detail in Appendix E.2). For candidate relations, we compute the similarity of question and freebase relations based on text-embedding-ada-002² and retain the top 20 most similar candidates. Different from previous fine-tuned methods (Ye et al., 2022; Hu et al., 2022), which typically require training a relation linking model for higher precision and recall, our method does not rely on the precision of these items. This is because, although all retrieved KB items are relevant to the question, they do not necessarily form valid combinations within a specific graph structure. In this way, the subsequent query construction steps in Section 3.3 can be viewed as a joint entity-relation disambiguation process, thus significantly reducing the number of invalid queries.
+
+# 3.3 Synthetic Query Construction
+
+This stage aims to construct question-targeted queries to facilitate subsequent QA reasoning. Given a knowledge base  $\mathcal{G}$  and the set of retrieved KB items relevant to the question  $nlq$ , we explore the possible query structures  $Q$  that are valid (i.e., yield non-empty execution results). Specifically, we used PyQL (Huang et al., 2023a) to represent logic form during construction. However, enumerating all possible structures may lead to an unmanageable combinatorial explosion. To mitigate this, our exploration of candidate queries follows a simple-to-complex manner, where we only further explore new structures that are derived from the sub-structures already verified as valid. Starting from the simplest structure  $(\mathcal{L}_1$  in Figure 2), we progressively search for more complex query structures through Layer-wise Expansion (for multi-hop structures) and Cross-layer Combination (for multi-constraint structures), gradually extending the obtained query graphs until a desired complexity is achieved.
+
+Layer-wise Expansion is utilized to model multi-hop structures (the depth of the query graph), which are chain-like, non-branching query structures originating from a single entity. We define  $\mathcal{L}_k$  as the set of queries in which the distance from the entity to the farthest variable in the chain-like query structure is  $k$ . Specifically, we first identify all possible connections between  $E_{nlq}$  and  $R_{nlq}$ , forming the simplest query structure  $\mathcal{L}_1$ , where an entity  $s$  is connected to a variable  $o$  through a single relation  $p$ .  $\operatorname{EXEC}(q, \mathcal{G})$  indicates the execution results of query  $q$  against  $\mathcal{G}$ .
+
+$$
+\begin{array}{l} \mathcal {L} _ {1} = \left\{\left(s, p, o\right) | s \in E _ {n l q}, p \in R _ {n l q}, \right. \tag {1} \\ \operatorname {E X E C} \left((s, p, o), \mathcal {G}\right) \neq \emptyset \}. \\ \end{array}
+$$
+
+We then progressively expand outward from the terminal variable nodes by connecting them to a new variable through another relation to construct  $\mathcal{L}_2$ , and so forth. Generally,  $\mathcal{L}_{k + 1}$  is formed by expanding the valid queries from the previous layer  $(\mathcal{L}_k)$  with an additional edge:
+
+$$
+\begin{array}{l} \mathcal {L} _ {k + 1} = \left\{q \cup \left(o _ {i}, p ^ {\prime}, o _ {j}\right) \mid q \in \mathcal {L} _ {k}, o _ {i} \in \mathcal {O} (q), \right. \\ p ^ {\prime} \in R _ {n l q}, \operatorname {E X E C} \left(q \cup \left(o _ {i}, p ^ {\prime}, o _ {j}\right), \mathcal {G}\right) \neq \emptyset \}. \tag {2} \\ \end{array}
+$$
+
+where  $\mathcal{O}(q)$  represents the set of variables in  $q$  and  $o_j$  is a newly introduced variable. The expansion process stops when the complexity threshold
+
+![](images/84532b8f5709b5a05fd29ef6d592f6893319eac45e771ede90d1bf8248e27564.jpg)  
+Figure 2: Overview of TARGA.
+
+(e.g., 3 hops) is reached, since for a coherent and reasonable question, the distance between a specific entity and the final answer typically does not exceed three hops.
+
+Cross-layer Combination models multi-constraint structures (the width of the query graph) by merging two queries, thereby applying multiple constraints to the same variable. Given two queries  $q$  and  $q'$ , we choose one of the variables from each query ( $o_i$  for  $q$  and  $o_j$  for  $q'$ ) as the common variable of them, then combine these two queries into a more complex query through this shared variable. We define  $\mathcal{L}_{x \times y}$  as the set of queries formed by combining a query from  $\mathcal{L}_x$  and a query from  $\mathcal{L}_y$ . Specifically, we start from the simplest combinations, such as merging two queries in  $\mathcal{L}_1$ , and gradually explore more complex combination patterns, such as merging  $\mathcal{L}_2$  with  $\mathcal{L}_3$  or merging  $\mathcal{L}_1$  with  $\mathcal{L}_{1 \times 2}$ . This combination process can be formally expressed as:
+
+$$
+\mathcal {L} _ {x \times y} = \left\{q \cup q ^ {\prime} \mid q \in \mathcal {L} _ {x}, q ^ {\prime} \in \mathcal {L} _ {y}, \right.
+$$
+
+$$
+\exists o _ {i} \in \mathcal {O} (q), o _ {j} \in \mathcal {O} \left(q ^ {\prime}\right), \tag {3}
+$$
+
+$$
+\left. \mathcal {E} \left(o _ {i}\right) \cap \mathcal {E} \left(o _ {j}\right) \neq \emptyset , \operatorname {E X E C} \left(q \cup q ^ {\prime}, \mathcal {G}\right) \neq \emptyset \right\},
+$$
+
+where  $\mathcal{E}(o_i)$  refers to the set of entities corresponding to the variable  $o_i$  in the execution result of query  $q$  on  $\mathcal{G}$ .  $o_i$  and  $o_j$  serve as the shared variable. This combination terminates once the query structure reaches five edges, which is sufficient to model most questions in current datasets.
+
+In this manner, we circumvented a significant number of invalid queries, thus obtaining most of the potentially relevant queries with relatively lower query overhead. We also provide statistical data regarding these synthetic queries in Appendix
+
+E.5. The average number of valid candidate queries per question is only in the range of several dozen, which is well within the contextual length limits manageable by an ICL model.
+
+# 3.4 Synthetic Query Re-ranking
+
+To obtain the most relevant examples for the subsequent QA task, we re-rank all valid queries using the bge-reranker-v2-m3 model (Chen et al., 2024) based on their similarity to the question. Additionally, we employ a process called Query Textification, where the synthesized query is transformed into a format closer to natural language through heuristic rules. This step helps bridge the gap between the text embedding model and the query, further improving the quality of the ranking. Detail and examples of textification process are provided in Appendix E.4 and Table 12.
+
+To address the imbalance caused by the exponential growth of complex queries, we implement a Hierarchical Ranking strategy. For all queries derived from the same parent query (the sub-query that this query is derived from), we retain only the top  $n$  candidates. The final candidate query pool is the union of all top tanked candidates:
+
+$$
+Q _ {\text {r a n k e d}} = \bigcup_ {a \in \mathcal {A}} \underset {q \in Q, \mathrm {P A R E N T} (q) = a} {\operatorname {A R G M A X} ^ {(n)}} \operatorname {S C O R E} (Q T (q), n l q). \tag {4}
+$$
+
+where  $Q$  denotes the set of queries generated during query construction, SCORE measures similarity and  $QT$  is Query Textification.  $\mathrm{PARENT}(q)$  indicates the parent query of  $q$ , and  $\mathcal{A}$  refers to the set of parent queries that have child queries. This approach ensures that the size of the candidate pool grows at a manageable rate, while preserving high
+
+quality queries for downstream processing.
+
+# 3.5 Question Answering
+
+To help the LLM understand the semantics of the provided query, we equip each generated query with its corresponding natural language questions (NLQ), forming (NLQ, Query) pairs. Specifically, we directly utilize the textification results mentioned in Section 3.4 as the NLQ, ensuring both efficiency and the preservation of information integrity. Then, we adopt the In-context Learning paradigm to generate the target query. Finally, we parse and execute the output query from the LLM to obtain the answer.
+
+# 4 Evaluation
+
+# 4.1 Setup
+
+We experiment with four complex KBQA datasets, i.e., GrailQA (Gu et al., 2021), GraphQ (Su et al., 2016), KBQA-Agent (Gu et al., 2024), MetaQA (Zhang et al., 2018) and a Text2SQL dataset, i.e., WikiSQL (Zhong et al., 2017). We use the F1 scores as the evaluation metric for KBQA and denotation accuracy for Text2SQL. We compare TARGA with various paradigms of baselines, including finetuning, ICL, and Agent, where we report performance in the original paper. For experiments with other settings, we copy the re-implemented result from Gu et al. (2024). By default, we use Qwen-2.5-7B-Instruct as the base LLM in our experiments with 10 demonstrations for all datasets. Detailed introduction of datasets and baselines are available in Appendices B and C.
+
+# 4.2 Main Result
+
+Table 1 illustrates the main result for KBQA. we compare TARGA with methods that require different amounts of annotation. For the relatively challenging datasets, i.e., GrailQA, GraphQ, and KBQA-Agent, based on a 7B model, TARGA achieves the best performance among all non-fine-tuned methods which are based on advanced close-sourced LLMs. On GrailQA and KBQA-Agent, TARGA surpasses previous SOTA non-fine-tuned methods by 8.7 and 13.0 F1. On GraphQ, TARGA even beats some fine-tuned methods and achieves similar performance with the best non-fine-tuned method.
+
+When compared to methods with a similar paradigm (ICL-based), TARGA outperform previous methods by 14.0 and 28.4 in F1 on GraphQ
+
+![](images/d286d6a44b087959b304b4e235214c7916b3360e13f133c57fba011f8c237ee2.jpg)  
+Figure 3: Performance with various numbers of demonstrations on GrailQA (1,000 randomly sampled questions).
+
+and KBQA-Agent, respectively. It is worth noting that our method requires neither any manually annotated corpus nor the expensive close-sourced model. Besides, we have not incorporated self-consistency to boost the performance. This can be attributed to the high quality of the synthetic data, which has led to a reduction in task difficulty and a decreased reliance on the capabilities of strong LLMs. Moreover, compared with other ICL-based methods which include 40-100 demonstrations, TARGA uses only 10 demonstrations but still achieves the best performance, demonstrating notable data efficiency.
+
+Compared to BYOKG which also works without annotated data, TARGA achieves approximately  $1.5 \times$  performance on GrailQA and MetaQA-3Hop. More importantly, TARGA dynamically synthesizes the most relevant data for different questions, enabling seamless adaptation to questions from any distribution. Besides, the synthetic data by TARGA is generated online, eliminating the need for a time-consuming offline data collection phase.
+
+# 4.3 Detailed Analyses
+
+To gain more insights, we conduct detailed experiments to illustrate some favorable practical characteristics of TARGA on: sample efficiency, robustness, generalization ability, efficiency, model size requirements, and transferability.
+
+# 4.3.1 Sample Efficiency
+
+In this section, we analyze how the number of demonstrations impacts the performance. We experiment on GrailQA, with the number of demonstrations ranging from 1 to 20. Based on our QA framework, we compare three distinct sampling settings: Random, Similarity-based, and Ours (synthetic), corresponding to examples randomly sampled from the training set, retrieved by similarity
+
+<table><tr><td>Methods</td><td>Models</td><td>GrailQA</td><td>GraphQ</td><td>KBQA-Agent</td><td>MetaQA</td></tr><tr><td colspan="6">full training set (Seq2Seq Fine-tuning / ICL / Agent Fine-tuning)</td></tr><tr><td>ArcaneQA (Gu and Su, 2022)</td><td>T5-base</td><td>73.7</td><td>31.8</td><td>-</td><td>-</td></tr><tr><td>Pangu (Gu et al., 2023)</td><td>T5-3B</td><td>83.4</td><td>57.7</td><td>-</td><td>-</td></tr><tr><td>KB-Binder-R (Li et al., 2023b)</td><td>GPT-3.5-turbo</td><td>58.5</td><td>32.5</td><td>-</td><td>99.5</td></tr><tr><td>KB-Coder-R (Nie et al., 2023)</td><td>GPT-3.5-turbo</td><td>61.3</td><td>36.6</td><td>-</td><td>-</td></tr><tr><td>KG-Agent (Jiang et al., 2024)</td><td>Llama2-7B</td><td>86.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DARA* (Fang et al., 2024)</td><td>Llama2-7B</td><td>77.7</td><td>62.7</td><td>-</td><td>-</td></tr><tr><td colspan="6">dozens of annotations (ICL)</td></tr><tr><td>KB-Binder (Li et al., 2023b)</td><td>GPT-3.5-turbo</td><td>50.8</td><td>34.5</td><td>4.2</td><td>96.4</td></tr><tr><td>KB-Coder (Nie et al., 2023)</td><td>GPT-3.5-turbo</td><td>51.7</td><td>35.8</td><td>-</td><td>-</td></tr><tr><td>Pangu (ICL) (Li et al., 2023b)</td><td>Codex</td><td>53.5</td><td>35.4</td><td>18.1</td><td>-</td></tr><tr><td colspan="6">one annotation (Agent Training-free)</td></tr><tr><td>AgentBench (Liu et al., 2024)</td><td>GPT-3.5-turbo</td><td>30.5</td><td>25.1</td><td>25.9</td><td>-</td></tr><tr><td>MIDDLEWARE (Gu et al., 2024)</td><td>GPT-3.5-turbo</td><td>-</td><td>-</td><td>34.3</td><td>-</td></tr><tr><td>QueryAgent (Huang et al., 2024)</td><td>GPT-3.5-turbo</td><td>60.5</td><td>50.8</td><td>-</td><td>98.5</td></tr><tr><td colspan="6">zero annotation (ICL)</td></tr><tr><td>BYOKG (Agarwal et al., 2024)</td><td>MPT-7B</td><td>46.5</td><td>-</td><td>-</td><td>56.5</td></tr><tr><td>TARGA (Ours)</td><td>QWen-2.5-7B-Instruct</td><td>69.0</td><td>50.6</td><td>46.5</td><td>85.7</td></tr><tr><td></td><td>QWen-2.5-72B-Instruct</td><td>70.6</td><td>54.1</td><td>57.3</td><td>99.8</td></tr><tr><td></td><td>GPT-3.5-turbo</td><td>68.9</td><td>51.0</td><td>52.7</td><td>96.5</td></tr><tr><td></td><td>GPT-4-turbo</td><td>69.8</td><td>52.5</td><td>51.4</td><td>99.9</td></tr></table>
+
+Table 1: Main results of KBQA performance, categorized by the amount of required annotated data example. Seq2Seq Fine-tuning / ICL / Agent Fine-tuning indicates different reasoning paradigms (split by the dashed line). Bold values highlight the best among non-fine-tuned models. * indicates using golden entity linking result.
+
+from the training set, and retrieved by similarity from the synthetic data by TARGA, respectively. The random and similarity settings can be viewed as reflections of the previous ICL-based and the retrieval-augmented ICL-based methods. Results are illustrated in Figure 3. With only one demonstration, our synthetic setting significantly outperforms the random and retrieval settings with 20 shots, suggesting the high quality of our synthetic data. Moreover, the growth curve in the synthetic setting (blue line) is relatively flat as the number of demonstrations increases. After reaching 7 shots, the synthetic setting exhibits almost no further improvement, while the other two settings continue to show growth even after reaching 20 shots, highlighting the data efficiency of our methods.
+
+# 4.3.2 Robustness Analyses
+
+To further validate the robustness of our approach in real-world scenarios, we conduct an adversarial experiment designed to simulate conditions of poor synthetic data quality. Specifically, the attack involves randomly replacing one relation in a candidate query. We compare the same three settings as in Section 4.3.1. As in Figure 4, our method exhibits significantly stronger robustness under adversarial conditions. Even when all demonstrations
+
+![](images/b08576b6a939d754ab60ed1d4ea2015b281f6e46a506d957c93c6afe425a7293.jpg)  
+Figure 4: Performance under attack setting on 1,000 randomly sampled GrailQA questions. Attack level indicates how many demonstrations have been corrupted.
+
+were compromised, the performance degradation of TARGA was only around  $25\%$ . In contrast, the other setups experience a sharp decline: the F1 scores of similarity-based setup drop by about  $40\%$ , and the random setting even falls by approximately  $75\%$ . This further demonstrates the superior robustness of our method compared to other approaches. We also provide another analysis about when corrupt entities in demonstrations in Appendix F.2
+
+# 4.3.3 Performance on Different Generalization Levels
+
+We experiment on GrailQA to compare the performance on different generalization levels (preliminary in Appendix D). To make a fair comparison
+
+<table><tr><td>Methods</td><td>I.I.D.</td><td>Comp.</td><td>Zero.</td><td>Avg.</td></tr><tr><td>KB-BINDER</td><td>48.3</td><td>48.8</td><td>41.8</td><td>50.8</td></tr><tr><td>KB-CODER</td><td>49.3</td><td>49.6</td><td>43.2</td><td>51.7</td></tr><tr><td>TARGA</td><td>68.4</td><td>62.2</td><td>71.7</td><td>69.0</td></tr><tr><td>KB-BINDER-R</td><td>80.6</td><td>53.6</td><td>50.7</td><td>58.5</td></tr><tr><td>KB-CODER-R</td><td>81.0</td><td>57.8</td><td>54.1</td><td>61.3</td></tr><tr><td>TARGA-R</td><td>80.8</td><td>63.6</td><td>71.6</td><td>71.9</td></tr></table>
+
+with the “-R” setting of previous methods, we also implemented a “-R” version of TARGA where the entire training set was incorporated into our demonstration sampling pool. Specifically, we retain the top 5 most similar examples from the training set and the top 5 most similar synthetic data instances jointly as demonstrations.
+
+As shown in Table 2, under the I.I.D setting, TARGA-R benefits from the inclusion of high-quality annotations (training set), achieving similar performance to previous “-R” methods. However, for the more challenging Compositional and Zeroshot settings, where similar questions are absent from the pre-collected training set, the performance of previous methods in the “-R” setup dramatically decreases by approximately 30 in F1. In contrast, TARGA shows no significant decline, demonstrating its strong generalization ability in scenarios that more closely resemble real-world situations where relevant corpora are unavailable. Notably, under the zero-shot setting, TARGA derives minimal improvements from the training set, suggesting that pre-collecting a substantial corpus of examples is ultimately an incomplete solution and tends to fail when confronted with real, unseen problems.
+
+# 4.3.4 Efficiency analysis
+
+For a practical QA system, high efficiency is also a key characteristic. Following Huang et al. (2024), we analyzed three efficiency metrics: TPQ, QPQ, and CPQ, as shown in Table 3. Regarding TPQ, our method significantly outperforms previous methods with only 4.5 seconds response time on GrailQA (detailed in Appendix F.4). Regarding QPQ, Agent-based methods have an inherent advantage. However, comparatively speaking, the overhead of QPQ remains relatively inexpensive than the other two metrics. Compared to KB-
+
+Table 2: Results of different generalization levels on GrailQA. “-R” indicates a version accessing the whole training set for similarity retrieval. Comp. and Zero. indicates the Compositional and Zero-shot setting on GrailQA, respectively. Avg. denotes the average F1.  
+
+<table><tr><td rowspan="2">Methods</td><td colspan="3">GrailQA</td><td colspan="3">GraphQ</td></tr><tr><td>TPQ</td><td>QPQ</td><td>CPQ</td><td>TPQ</td><td>QPQ</td><td>CPQ</td></tr><tr><td>KB-BINDER</td><td>51.2</td><td>3,297.7</td><td>0.010</td><td>84.0</td><td>2,113.8</td><td>0.024</td></tr><tr><td>AgentBench</td><td>40.0</td><td>7.4</td><td>0.034</td><td>65.1</td><td>7.2</td><td>0.035</td></tr><tr><td>QueryAgent</td><td>16.6</td><td>5.2</td><td>0.019</td><td>15.3</td><td>6.2</td><td>0.021</td></tr><tr><td>TARGA</td><td>4.5</td><td>256.8</td><td>0.000</td><td>13.0</td><td>1,094.6</td><td>0.000</td></tr></table>
+
+Table 3: Efficiency analysis with ICL-based (KB-Binder) and Agent-based methods (AgentBench and QueryAgent). TPQ, QPQ, and CPQ denote the time cost (seconds), number of SPARQL query times, and open-source model invocation cost ($) per question.
+
+Binder, which also employs the ICL paradigm, our approach demonstrates a marked superiority on QPQ. This is primarily because our synthetic demonstrations are highly aligned with the target question, enabling the generated logic forms to be executable without any post-processing in most cases. Conversely, since previous methods can not always retrieve the relevant candidate query in the training set, the generated logic forms are often not executable. Consequently, the logic form necessitates stepwise binding to valid KB items, which leads to a large demand for queries.
+
+In terms of CPQ, agent-based methods inherently face challenges due to the lengthy trajectory of demonstrations and the need for multiple calls of LLM. Since our method does not rely on close-sourced LLM, the CPQ is zero. If compared with the consumed tokens, TARGA uses significantly fewer tokens because it requires fewer examples as demonstrations and does not use self-consistency. As a result, the token cost is only about 1/10 that of other ICL-based methods. We also provide a detailed analysis of token consumption in the Appendix F.6 and more detail in Appendix F.5.
+
+# 4.3.5 Performance on different sizes of LLM
+
+In real-world applications, large and powerful LLMs are not always accessible or affordable. Therefore, we further analyze the performance of various methods across different model sizes. We compare the Agent method (QueryAgent), ICL method (KB-Binder), and the retrieval-augmented ICL method (KB-BINDER-R). As shown in Table 4, our approach demonstrates remarkable adaptability from qwen-1.5B-instruct to gpt-4o-mini. With just the 1.5B model, our method already surpasses the previous best-performing method, while at 7B, it only slightly lags behind the closed-source model. The Agent method has strong generalization capabilities but is heavily reliant on the planning and
+
+<table><tr><td></td><td>1.5B</td><td>7B</td><td>32B</td><td>72B</td><td>4o-mini</td><td>Δ</td></tr><tr><td>QueryAgent</td><td>10.3</td><td>16.1</td><td>50.8</td><td>58.5</td><td>62.3</td><td>52.0</td></tr><tr><td>KB-Binder</td><td>20.2</td><td>39.8</td><td>51.3</td><td>50.6</td><td>47.0</td><td>31.1</td></tr><tr><td>KB-Binder-R</td><td>27.6</td><td>55.0</td><td>59.4</td><td>63.8</td><td>58.0</td><td>36.3</td></tr><tr><td>TARGA</td><td>61.3</td><td>65.3</td><td>67.4</td><td>67.5</td><td>67.7</td><td>9.4</td></tr></table>
+
+Table 4: The performance on GrailQA with different base model sizes. The 1.5B, 7B, 32B, and 72B represent the Qwen 2.5 instruct models family, while 4o-mini indicates GPT-4o-mini. We experiment on 500 random sampling questions.  $\Delta$  indicates the max performance gap between different models.  
+
+<table><tr><td></td><td>GrailQA</td><td>GraphQ</td><td>KBQA-Agent</td></tr><tr><td>TARGA</td><td>69.0</td><td>50.6</td><td>46.5</td></tr><tr><td>w/o Query Textification</td><td>64.9</td><td>46.6</td><td>39.6</td></tr><tr><td>w/o Re-Ranking</td><td>59.9</td><td>38.7</td><td>19.3</td></tr><tr><td>w/o Synthetic Question</td><td>67.3</td><td>49.5</td><td>45.5</td></tr></table>
+
+self-correction abilities of the most advanced LLM, which smaller models do not excel at. For models below 72B, the performance of the QueryAgent is essentially unusable. For the 72B model, the performance of QueryAgent is still inferior to that of the ICL method using a model of the same size and ultimately failing to exceed closed-source model performance. Regarding the ICL methods, previous works typically experiment on the strongest closed-source models without testing their performance on open-source models. We demonstrate here the performance of ICL on open-source models, revealing that the latest open-source models can reach or even surpass the capabilities of the GPT series models in certain tasks. This provides a feasible assurance for continuing research in semantic parsing based on closed-source models.
+
+# 4.3.6 Ablation Study
+
+Table 5 presents the impact of distinct components on model performance across three datasets. Compared to the full model, removing the query text- tification component leads to a noticeable drop, particularly on KBQA-Agent (-6.9), highlighting the importance of bridging the gap between the text embedding model and logic form. The removal of the re-ranking component results in the largest performance decrease, with reductions of 9.1, 11.9, and 27.2 on GrailQA, GraphQ, and KBQA-Agent, respectively, underscoring the importance of the re-ranking step. In contrast, excluding synthetic question generation yields more modest declines, suggesting it is less critical than the other compo
+
+Table 5: Ablation study of each component on GrailQA.  
+
+<table><tr><td>Methods</td><td>Acc.</td></tr><tr><td>RESDNSL-3B + NATSQL* (Li et al., 2023a)</td><td>79.9</td></tr><tr><td>T5-3B+PICARD* (Scholak et al., 2021)</td><td>75.1</td></tr><tr><td>StructGPT (ChatGPT) (Jiang et al., 2023)</td><td>65.6</td></tr><tr><td>Readi (Cheng et al., 2024)</td><td>66.2</td></tr><tr><td>AgentBench (Liu et al., 2024)</td><td>57.6</td></tr><tr><td>QueryAgent (Huang et al., 2024)</td><td>72.5</td></tr><tr><td>TARGA</td><td>75.5</td></tr></table>
+
+Table 6: Results on WikiSQL. * indicates fine-tuned.
+
+nents but still beneficial for KBQA-Agent. It is unexpected, but from another aspect, it indicates that even only using the synthetic query as the demonstration the performance is also competitive.
+
+# 4.3.7 Transferability to Text2SQL
+
+We adapted our framework to the Text2SQL task to demonstrate the generality of our approach in other semantic parsing tasks. Employing the WikisSQL dataset, we compare TARGA with both the fine-tuned and non-fine-tuned methods. Among them, StructGPT and Readi are 32-shot and 7-shot methods, respectively. AgentBench and QueryAgent both use 1 shot. As shown in Table 6, with merely 10 synthetic examples as demonstrations, our method surpasses prior methods with 32 manually annotated examples and also outperforms the best 1-shot method, all while incurring a lower cost and smaller model. Besides, TARGA can even surpass a fine-tuned method with 3B model.
+
+# 5 Conclusion
+
+In this paper, we explore two critical challenges in the semantic parsing task: reliance on annotated data and poor generalization on non-I.I.D. cases. We proposed a novel method called TARGA, which automatically synthesizes examples that are most relevant to the test data and utilizes them as demonstrations for in-context learning. Remarkably, TARGA achieves the best performance among all non-fine-tuned methods across three complex KBQA datasets and one Text2SQL dataset, especially on GrailQA and KBQA-Agent (7.7 and 12.2 F1 points, respectively). While achieving impressive performance, TARGA also exhibits the following practical properties: 1) It does not require any annotated data. 2) It is effective even with a model size of just 7B parameters. 3) The synthetic data is generated online. 4) It exhibits superior generalization, robustness, and speed. This work highlights the potential of leveraging synthetic data in seman
+
+tic parsing, and we hope that TARGA can serve as a valuable foundation for developing more practical systems in this field.
+
+# Limitations
+
+We would like to discuss some limitations of our work. First, in this paper, we validate TARGA on two specific semantic parsing tasks: KBQA and Text2SQL. While these tasks demonstrate the potential of our approach, further exploration across a broader range of tasks that involve transforming natural language into logical forms could strengthen the generalizability of TARGA. Additionally, we have not yet investigated the feasibility of our synthetic data generation method in other paradigms, such as agent-based or fine-tuned models. We would like to adapt TARGA to these paradigms in future work.
+
+# Acknowledgements
+
+This work is supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62072224. The authors would like to thank all anonymous reviewers for their advice.
+
+# References
+
+Dhruv Agarwal, Rajarshi Das, Sopan Khosla, and Rashmi Gangadharaiah. 2024. Bring your own KG: Self-supervised program synthesis for zero-shot KGQA. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 896-919.  
+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.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901.  
+Shulin Cao, Jiaxin Shi, Liangming Pan, Lunyiu Nie, Yutong Xiang, Lei Hou, Juanzi Li, Bin He, and Hanwang Zhang. 2022. KQA pro: A dataset with explicit
+
+compositional programs for complex question answering over knowledge base. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6101-6119.  
+Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. Preprint, arXiv:2402.03216.  
+Sitao Cheng, Ziyuan Zhuang, Yong Xu, Fangkai Yang, Chaoyun Zhang, Xiaoting Qin, Xiang Huang, Ling Chen, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan, and Qi Zhang. 2024. Call me when necessary: Llms can efficiently and faithfully reason over structured environments. Preprint, arXiv:2403.08593.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+Haishuo Fang, Xiaodan Zhu, and Iryna Gurevych. 2024. DARA: Decomposition-alignment-reasoning autonomous language agent for question answering over knowledge graphs. In Findings of the Association for Computational Linguistics: ACL 2024, pages 3406-3432, Bangkok, Thailand. Association for Computational Linguistics.  
+Williams Gregory. 2013. Sparql 1.1 service description. w3c recommendation. World Wide Web Consortium.  
+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.  
+Yu Gu, Sue Kase, Michelle Vanni, Brian Sadler, Percy Liang, Xifeng Yan, and Yu Su. 2021. Beyond I.I.D.: Three levels of generalization for question answering on knowledge bases. The Web Conference 2021 - Proceedings of the World Wide Web Conference, WWW 2021, 2021:3477-3488.  
+Yu Gu, Yiheng Shu, Hao Yu, Xiao Liu, Yuxiao Dong, Jie Tang, Jayanth Srinivasa, Hugo Latapie, and Yu Su. 2024. Middleware for llms: Tools are instrumental for language agents in complex environments. Preprint, arXiv:2402.14672.  
+Yu Gu and Yu Su. 2022. ArcaneQA: Dynamic program induction and contextualized encoding for knowledge base question answering. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1718-1731.  
+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.  
+Xiang Huang, Sitao Cheng, Yuheng Bao, Shanshan Huang, and Yuzhong Qu. 2023a. MarkQA: A large scale KBQA dataset with numerical reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10241-10259, Singapore. Association for 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. Preprint, arXiv:2403.11886.  
+Xiang Huang, Sitao Cheng, Yiheng Shu, Yuheng Bao, and Yuzhong Qu. 2023b. Question decomposition tree for answering complex questions over knowledge bases. Preprint, arXiv:2306.07597.  
+Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Xin Zhao, and Ji-Rong Wen. 2023. StructGPT: A general framework for large language model to reason over structured data. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9237-9251.  
+Jinhao Jiang, Kun Zhou, Wayne Xin Zhao, Yang Song, Chen Zhu, Hengshu Zhu, and Ji-Rong Wen. 2024. Kg-agent: An efficient autonomous agent framework for complex reasoning over knowledge graph. Preprint, arXiv:2402.11163.  
+Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. Preprint, arXiv:2309.06180.  
+Haoyang Li, Jing Zhang, Cuiping Li, and Hong Chen. 2023a. Resdnsq: Decoupling schema linking and skeleton parsing for text-to-sqI. In AAAI.  
+Tianle Li, Xueguang Ma, Alex Zhuang, Yu Gu, Yu Su, and Wenhu Chen. 2023b. Few-shot in-context learning on knowledge base question answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6966-6980.  
+Zhenyu Li, Sunqi Fan, Yu Gu, Xiuxing Li, Zhichao Duan, Bowen Dong, Ning Liu, and Jianyong Wang. 2023c. Flexkbq: A flexible llm-powered framework for few-shot knowledge base question answering. Preprint, arXiv:2308.12060.
+
+Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie Huang, Yuxiao Dong, and Jie Tang. 2024. Agent-bench: Evaluating LLMs as agents. In The Twelfth International Conference on Learning Representations.  
+Zhijie Nie, Richong Zhang, Zhongyuan Wang, and Xudong Liu. 2023. Code-style in-context learning for knowledge-based question answering. Preprint, arXiv:2309.04695.  
+Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. 2021. PICARD: Parsing incrementally for constrained auto-regressive decoding from language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9895-9901, 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.  
+Yu Su, Huan Sun, Brian Sadler, Mudhakar Srivatsa, Izzeddin Gur, Zenghui Yan, and Xifeng Yan. 2016. On generating characteristic-rich question sets for QA evaluation. In Empirical Methods in Natural Language Processing (EMNLP), Austin, Texas, USA.  
+Alon Talmor and Jonathan Berant. 2018. The web as a knowledge-base for answering complex questions. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 641-651, New Orleans, Louisiana.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.  
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508.  
+Xi Ye, Semih Yavuz, Kazuma Hashimoto, Yingbo Zhou, Caiming Xiong, and Computer Science. 2022. RNGKBQA: Generation Augmented Iterative Ranking for Knowledge Base Question Answering. In Proceedings of the 60th Annual Meeting of the Association
+
+for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 6032-6043.  
+Wen-tau Yih, Matthew Richardson, Chris Meek, Ming-Wei Chang, and Jina Suh. 2016. The value of semantic parse labeling for knowledge base question answering. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 201–206.  
+Yuyu Zhang, Hanjun Dai, Zornitsa Kozareva, Alexander Smola, and Le Song. 2018. Variational reasoning for question answering with knowledge graph. In Proceedings of the AAAI conference on artificial intelligence.  
+Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2sql: Generating structured queries from natural language using reinforcement learning. CoRR, abs/1709.00103.
+
+# A Preliminary
+
+# A.1 Knowledge Base Question Answering
+
+We introduce related concepts and the task of knowledge base question answering as follows. Let  $E$  be a set of entities,  $P$  be a set of relations,  $C$  be a set of classes, and  $I$  be a set of literals. A knowledge base  $\mathcal{K} \subseteq E \times P \times (E \cup C \cup I)$  is a set of triples  $(s, p, o)$ , where  $s \in E$  is a subject entity,  $p \in P$  is a predicate, and  $o \in (E \cup C \cup I)$  is an object entity, class or literal value.
+
+The task of knowledge base question answering can be formalized as learning a function  $f$  that takes a natural language question  $q$  as input, and outputs a structured query  $q' = f(q)$  in a formal language such as SPARQL (Gregory, 2013). The structured query  $q'$  should ideally encode the entities, relations, and constraints specified by the input question  $q$ , such that executing  $q'$  over the knowledge base  $\mathcal{K}$  yields the correct answer of  $q$ .
+
+# A.2 In-Context Learning
+
+In-context learning (Brown et al., 2020) allows LLMs to perform new tasks by simply observing examples provided within the input, without updating their internal parameters. Intuitively, the model "learns" from the context and uses it to generate appropriate responses for similar tasks, relying on patterns it recognizes from the given examples.
+
+More precisely, let  $[x_1, x_2, \dots, x_n]$  be a sequence of input tokens representing natural language texts, and let  $[y_1, y_2, \dots, y_n]$  be a corresponding sequence of output tokens representing the desired task, which in our context are structured queries. A LLM, denoted as  $f_{\theta}$ , is a function parameterized by  $\theta$ , which takes an input sequence and predicts the next token or sequence of tokens. In-context learning refers to the ability of a pretrained LLM to learn and adapt to a specific task purely by conditioning on a sequence of examples  $S = [(x_1, y_1), (x_2, y_2), \dots, (x_n, y_n)]$ , provided as part of the input context, without updating the model parameter  $\theta$ . In contrast to traditional learning paradigms that require parameter updates via gradient descent, the LLM uses the provided examples to infer the underlying task and generate predictions for a new input  $x_{n+1}$ .
+
+# B Datasets
+
+We conduct experiments on four KBQA datasets and one Text2SQL dataset, their statistics are shown in Table 7.
+
+<table><tr><td>Datasets</td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td>GRAILQA</td><td>44,337</td><td>6,763</td><td>13,231</td></tr><tr><td>GRAPHQ</td><td>2,381</td><td>-</td><td>2,395</td></tr><tr><td>KBQA-Agent</td><td>-</td><td>-</td><td>500</td></tr><tr><td>METAQA-3HOP</td><td>114,196</td><td>14,274</td><td>14,274</td></tr><tr><td>WIKISQL</td><td>56,355</td><td>8,421</td><td>15,878</td></tr></table>
+
+Table 7: Statistics of datasets.
+
+- GrailQA (Gu et al., 2021) is one of the most popular complex KBQA datasets. It divides the dataset into three levels of generalization, i.e., I.I.D, Compositional, and Zero-shot.  
+- GraphQ (Su et al., 2016) is a challenging dataset that only consists of Compositional questions.  
+- KBQA-Agent (Gu et al., 2024) is a mixed dataset of the most difficult questions from four datasets (i.e., GrailQA (Gu et al., 2021), ComplexWebQuesiton (Talmor and Berant, 2018), GraphQ (Su et al., 2016), and WebQSP (Yih et al., 2016)).  
+MetaQA-3Hop (Zhang et al., 2018) is the most difficult 3-hop subset of a large-scale KBQA dataset based on Wiki-Movies KG.  
+- WikiSQL (Zhong et al., 2017) is a large-scale complex Text2SQL dataset which requiring comparison, aggregation and arithmetic operations.
+
+# C Baselines
+
+# C.1 Seq2Seq Fine-tuning Methods
+
+- ArcaneQA (Gu and Su, 2022) is a generation-based method that incrementally synthesizes a program by dynamically predicting a sequence of subprograms. It prunes the search space by constrained decoding.  
+- Pangu (Gu et al., 2023) leverages the discrimination ability of language models to build queries in an incremental manner. It consists of a symbolic agent to collect valid candidate plans and an LM to select the most likely one.
+
+# C.2 ICL-based Method
+
+- Pangu-ICL (Gu et al., 2023) is a ICL version of Pangu. The result is based on output distribution of the LLMs, which require access to model parameters.
+
+- KB-Binder (Li et al., 2023b) propose an ICL-based method for few-shot KBQA by feeding LLM with some (question, S-expression) pairs.  
+- KB-Coder (Nie et al., 2023) further optimizes KB-Binder by changing the target format from S-expression to code-style logic form.  
+- Readi (Cheng et al., 2024) propose a Reasoning-then-Editing framework that initially generates a reasoning path given a query, then instantiates the path. Editing is triggered only when necessary.  
+- BYOKG (Agarwal et al., 2024) does not need manually annotated dataset. It first explores the KB to collect synthetic datasets within a day (e.g., 10K data in 10 hours), then uses the synthetic dataset for bottom-up reasoning.
+
+# C.3 Agent Fine-tuning Methods
+
+- KG-Agent (Jiang et al., 2024) enables a small LLM to actively make decisions until finishing the reasoning process over KGs through fine-tuning an Agent.  
+- DARA (Fang et al., 2024) propose a Decomposition-Alignment-Reasoning Autonomous Language Agent which can be efficiently trained with a small number of high-quality reasoning trajectories.
+
+# C.4 Agent-based Method (Training-free)
+
+- AgentBench (Liu et al., 2024) model KBQA as a tool learning task and outfitting LLM with an array of KG-querying tools such as "get Relation", "argmax" and "intersection".  
+- MIDDLEWARE (Gu et al., 2024) design customized tools acting as middleware between LLMs and complex environments. They also incorporate decoupled generation and error feedback to boost performance.  
+- QueryAgent (Huang et al., 2024) step-by-step build the target query and use an environmental feed-based self-correction method to reduce hallucination.
+
+# D Levels of Generalization
+
+In the context of KBQA, the three levels of generalization—I.I.D. generalization, compositional generalization, and zero-shot generalization—refer to the capability of models to handle increasingly challenging and diverse types of questions (Gu et al., 2021). In particular, I.I.D. generalization refers to the model's ability to correctly answer questions that are sampled from the same distribution as the training set, which assumes that the test data follow similar patterns and schema. On top of that, compositional generalization refers to the model's ability to handle novel combinations of KB items (e.g., entities and relations) that were seen during training, but in configurations that the model has not encountered before. Finally, zero-shot generalization refers to the model's ability to answer questions involving entirely new KB items, such as unseen entities and relations that were never presented in the training set.
+
+Gu et al. (2021) argue that relying solely on datasets under I.I.D. settings limits the practical use of KBQA models, as real-world questions often involve unfamiliar entities or require novel reasoning. This limitation is evident in the degraded performance of existing KBQA methods under compositional and zero-shot settings compared to I.I.D. settings. Therefore, practical KBQA models should be equipped with built-in generalization capabilities across all three levels to better handle diverse, real-world questions.
+
+# E More Details
+
+# E.1 Other Experiment Settings
+
+TARGA and Pangu used 10 demonstrations across all datasets. KB-Binder and KB-Coder used 40 shots for GrailQA, 100 shots for GraphQ, and 5 shots for MetaQA, with KB-Binder employing 20 shots for KBQA-Agent. AgentBench, MIDDLE-WARE, and QueryAgent are all 1-shot methods. The experiments using models with 1.5B, 7B, 32B, and 72B parameters were conducted on 1, 1, 4, and 8 A100 GPUs, respectively.
+
+# E.2 Entity Linking Detail
+
+The detailed entity linking results of all compared methods are listed in Table 8. For KBQA-Agent and MetaQA-3Hop, all compared methods use golden linking results. For GrailQA and GraphQ, most compared methods use the entity linking re
+
+<table><tr><td>Method</td><td>GrailQA</td><td>GraphQ</td><td>KBQA-Agent</td><td>MetaQA-3Hop</td></tr><tr><td>KB-Binder</td><td>FACC1+BM25</td><td>FACC1+BM25</td><td>Golden</td><td>Exact match</td></tr><tr><td>KB-Coder</td><td>FACC1+SimCSE</td><td>FACC1+SimCSE</td><td>-</td><td>-</td></tr><tr><td>Pangu(ICL)</td><td>Pangu</td><td>Pangu</td><td>Golden</td><td>-</td></tr><tr><td>MIDDLEWARE</td><td>-</td><td>-</td><td>Golden</td><td>-</td></tr><tr><td>AgentBench</td><td>Pangu</td><td>Pangu</td><td>Golden</td><td>-</td></tr><tr><td>QueryAgent</td><td>Pangu</td><td>Pangu</td><td>-</td><td>Exact match</td></tr><tr><td>TARGA</td><td>Pangu</td><td>Pangu</td><td>Golden</td><td>Exact match</td></tr></table>
+
+sult by Pangu (Gu et al., 2023). Therefore, we follow their setting to make a fair comparison.
+
+Among all the methods compared, KB-Binder and KB-Coder chose a pipeline that was different from the other methods. Other methods first get the entity linking results and then generate logic forms based on them. KB-Binder and KB-Coder first generate a logic form draft and then bind each KB item in the logic form draft to the knowledge base. For KB-Binder and KB-Coder, the elements requiring linking depend on the generated draft, making it less suitable for directly using the entity linking results as other methods.
+
+# E.3 Logic Form Design
+
+Inspired by PyQL (Huang et al., 2023a), we designed a simplified logical form for constructing queries, which uses a series of functions for semantic expression, making it easier for LLMs to learn and enabling seamless translation into SPARQL. Table 10 shows all functions that we used.
+
+# E.4 Query Textification Detail
+
+To improve the performance of the text embedding model, we transform the synthesized query into a format closer to natural language by implementing a simple yet effective rule-based parsing program. An example of query textification is provided in Table 12.
+
+Specifically, we first replace the entities and relations in the triples with their labels, concatenating the subject, predicate, and object to create a description. For two nested descriptions (multi-hop), we represent them in a hierarchical form using "a have b". For two conjunction descriptions (multi-constraint), we connect the two parts with "and." We also applied some simple processing to the filters, such as using "more than" and "less than" to
+
+connect comparison objects.
+
+# E.5 Statistics of Query Construction
+
+As shown in Table 9, we present the statistics of the synthetic query in Section 3.3.
+
+Table 8: The detailed source of entity linking data for compared method. The original performance of AgentBench is based on the Golden entity linking result. For a fair comparison, we report the performance with the Pangu entity linking result re-implemented by QueryAgent (Huang et al., 2024) in Table 1.  
+
+<table><tr><td>Datasets</td><td>Size</td><td>Cvg.</td><td>TPQ</td><td>QPQ</td></tr><tr><td>GrailQA</td><td>17.4</td><td>0.79</td><td>2.99</td><td>256.8</td></tr><tr><td>GraphQ</td><td>89.3</td><td>0.69</td><td>11.41</td><td>1,094.6</td></tr><tr><td>KBQA-Agent</td><td>173.1</td><td>0.78</td><td>20.07</td><td>2,270.3</td></tr><tr><td>MetaQA-3Hop</td><td>25.9</td><td>1.00</td><td>0.05</td><td>56.1</td></tr><tr><td>WikiSQL</td><td>18.1</td><td>0.75</td><td>0.16</td><td>88.6</td></tr></table>
+
+Table 9: Statistics of query construction. Size represents the average number of valid queries per question. Cvg. refers to coverage, indicating the proportion of questions with at least one correct  $(\mathrm{F}1 = 1)$  in synthetic queries. TPQ and QPQ denote the running time (in seconds) and the number of query attempts per question, respectively.
+
+# E.6 Natural Language Question Generation Detail
+
+We directly use the pseudo-questions generated by query textification for building (query, question) pairs as demonstrations in QA phase, and the performance is good enough. We also experimented with using an LLM (GPT or Qwen) to generate questions during this phase, but the downstream QA performance remained nearly unchanged, with a difference of less than 0.5 F1. Moreover, leveraging an LLM to generate corresponding questions requires labeling a few examples as demonstrations and incurs an extra time cost (less than 1 second).
+
+In fact, even without generating questions and using only the queries as demonstrations, the performance showed only a slight decline (see the ablation study in Table 6). We recommend choosing
+
+between using rules or an LLM based on specific scenarios:
+
+- If you want to use TARGA for ICL-style QA, the quality of question generation at this stage does not need to be particularly high. Heuristic rules are sufficient enough.  
+- If you want to construct standard parallel corpora for fine-tuning or other purposes, we recommend you use an LLM along with some demonstrations to generate higher-quality natural language questions.
+
+# E.7 Prompt for Question Answering
+
+We provide example prompts for TARGA on the KBQA and Text2SQL tasks In Tables 16 and 17. It consists of some synthetic demonstrations (NLQ-Query pairs) and the test question. For the KBQA task, we provide the entity linking results. For the Text2SQL task, we provide the headers of the table.
+
+# F Further Analysis
+
+# F.1 The performance difference between Compositional and zero-shot setting
+
+In Table 2, we notice that the performance of the compositional setting is lower than the zero-shot setting for TARGA, but this trend is reversed for other compared methods. Our perspective is outlined as follows:
+
+- This is a matter of relative performance: If the performance of TARGA on zero-shot performance is above the normal level, then compositional will appear lower than zero-shot, and vice versa for comparative methods. If the performance of other methods on zero-shot setting is very poor, the performance of compositional setting will appear relatively high.  
+- Setting aside whether the test questions are seen in the training set or not, there are significant inherent difficulty differences among the three generalization levels of GrailQA: compositional is the hardest, followed by i.i.d., and zero-shot is the easiest.
+
+Firstly, we explain why other methods exhibit lower zero-shot performance:
+
+The zero-shot setting implies that the given question contains entities or relationships not seen in the training set. The compositional setting involves combinations of entities and relationships that have
+
+not been seen together, though each individual element exists in the training set. Note that the comparative methods we used are, to some degree, influenced by the training set. Retrieving several examples from the training set can still essentially cover all the elements needed to solve a question, but this is not the case with zero-shot. This may explain why compared methods exhibit lower zero-shot performance than compositional settings. Besides, for GraphQ, the  $100\%$  Compositional dataset, TARGA exhibit more significant improvement compared to KB-Binder and KB-Coder than GrailQA (25% Compositional), indicating that TARGA does not harm the performance of the compositional setting.
+
+Secondly, we explain why our compositional performance is lower than zero-shot:
+
+We believe that the fundamental reason for this phenomenon lies in how the dataset is split and the inherent difficulty of the problems themselves. For TARGA, there is no substantial distinction between input questions, as each one is treated as an unseen question. However, the generalization level of a question is determined solely by whether it appears in the training set, irrespective of the question's inherent difficulty. As a result, TARGA serves as a fair model for evaluating a question's inherent difficulty, as it is unaffected by any static dataset distribution. We hypothesize that TARGA's performance across the three generalization levels reflects significant differences in difficulty within the GrailQA dataset, with the compositional setting being the most challenging, followed by i.i.d., and zero-shot being the easiest.
+
+To support this hypothesis, Table 11 reports the average length of the final SPARQL queries for questions across the three generalization levels. In general, longer SPARQL queries correspond to more complex questions (i.e., higher difficulty). Compositional questions yield the longest queries and the lowest performance, whereas zero-shot questions are the simplest and achieve the highest performance. These statistical results strongly validate our earlier conjecture.
+
+# F.2 Further Analysis on Robustness
+
+In Section 4.3.2, we have analyzed the effect of randomly replacing relations in demonstrations. Here, we present the performance of randomly replacing entities in the demonstrations. Similarly to 4.3.2, we corrupt N demonstrations by replacing one entity with another randomly selected entity from the
+
+<table><tr><td>Functions</td><td>Brief Description</td></tr><tr><td>triplet(s, p, o)</td><td>Add a condition that the subject s is linked to the object o via a relation p.</td></tr><tr><td>argmax(v)</td><td>Add a condition that the variable v must have the maximum value.</td></tr><tr><td>argmin(v)</td><td>Add a condition that the variable v must have the minimum value.</td></tr><tr><td>filter(v, op, value)</td><td>Add a condition that the variable v must meet, where the operator op is one of [&lt;, &gt;, &lt;=, &gt;=].</td></tr><tr><td>type(v, t)</td><td>Specify that the variable v belongs to the class (entity type) t.</td></tr><tr><td>count(v)</td><td>Set the count of variable v as the final answer.</td></tr><tr><td>answer(v)</td><td>Set the variable v as the final answer.</td></tr></table>
+
+Table 10: Functions used in our logic form.  
+
+<table><tr><td>Generalization level</td><td>Difficulty</td><td>Performance</td></tr><tr><td>Compositional</td><td>5.1</td><td>62.2</td></tr><tr><td>I.I.D.</td><td>4.8</td><td>68.4</td></tr><tr><td>Zero-shot</td><td>4.5</td><td>71.7</td></tr></table>
+
+Table 11: Difficulty and performance of TARGA across three generalization settings on GrailQA. The difficulty is measured by the average length of SPARQL queries, and the performance shows TARGA's F1 score.
+
+![](images/5c8a3d4de429cbaaff18be752ae9353d6f6734da33c7e02df731fc57ea47a528.jpg)  
+Figure 5: Performance under attack setting(entity) on 1,000 randomly sampled GrailQA questions.
+
+training set (with a candidate pool of 32K entities). The results in Figure 5 indicate that replacing the entity has a minimal impact across all three settings.
+
+# F.3 Further analysis on ranking
+
+Considering that a poor ranking might incorrectly identify the top 1 or top 2 results, while the top 10 overall might still be generally correct. Therefore, we conducted a more convincing experiment using only the top 1 data as the demonstrations to better illustrate the impact of query textification on ranking quality.
+
+The results are shown in Table 13: we found that in this setting, the decline without query textification was generally more pronounced (9.9, 7.6, and
+
+3.1 on three datasets) than Table 5 (4.1, 4.0 and 6.9 on three datasets), indicating that query textification can enhance ranking quality, thereby further improving QA performance.
+
+# F.4 Runtime Breakdown
+
+We present a detailed runtime breakdown to provide a clearer understanding of the time costs associated with each stage, using GrailQA as an example.
+
+(1). Entity linking: nearly negligible (0 s): We use the cache provided by previous work for fair comparison. Most previous non-finetuning methods either rely on cached results or the golden linking results. Details are provided in Section E.2.  
+(2). Relation linking: very fast (0.60 s): Relation linking consists of three steps:
+
+a). Obtain the relation embedding (0 s, cached in advance). Note that the relations in Freebase are fixed and limited, so they only need to be cached once in advance.  
+b). Obtain the question embedding (0.52 s).  
+c). Calculate the embedding similarity between question embedding and all relation embeddings (0.08s). We utilized FAISS (Johnson et al., 2019) to accelerate similarity computations.  
+(3). Query construction: the slowest part (2.99 s for GrailQA):
+
+This is the most time-consuming stage, which is described in detail in Section 3.3.
+
+(4). Reranking: very fast (0.11 s): This is implemented by bge-reranker. Since we carefully controlled the number of candidate queries during the construction process, the final set for ranking is relatively small. Details about the size of the candidate queries are provided in Table 9. For instance, only an average of 17.4 candidate queries were involved in the reranking stage, which contributed to its speed.
+
+# Queries with textification results
+
+# Query 1:
+
+triplet(?v0, measurement_unit MASS_unit.weightmass_in_kilograms, ?v1)  
+argmin(?v1)  
+answer(?v0)
+
+# Textification Result:
+
+what mass_unit, mass_unit has weightmass_in_kilograms, when weightmass_in_kilograms is the smallest
+
+# Query 2:
+
+triplet(?v0, boats ship_class.date-designed, ?v1)  
+argmax(?v1)  
+answer(?v0)  
+Textification Result:  
+what ship_class, ship_class has date-designed, when date-designed is the largest
+
+# Query 3:
+
+triplet(?v0, spaceflight.rocket_engine.designed_by, [rocketdyne])  
+triplet(?v0, spaceflight.rocket.engine.isp_sea_level, ?v1)  
+filter(?v1, <=, 260.0)  
+answer(?v0)  
+Textification Result:  
+what rocket ENGINE, rocket ENGINE has rocketdyne, rocket ENGINE has isp_sea_level, when isp_sea_level no more than 260.0
+
+Table 12: Examples of Query Textification.  
+
+<table><tr><td></td><td>GrailQA</td><td>GraphQ</td><td>KBQA-Agent</td></tr><tr><td>w/ QT</td><td>61.46</td><td>44.31</td><td>37.53</td></tr><tr><td>w/o QT</td><td>52.55</td><td>36.76</td><td>34.48</td></tr></table>
+
+Table 13: Abalation study of Query Textification (QT) when only using Top 1 candidate as the demonstration (experiment on 500 randomly selected examples).
+
+(5). Question generation: nearly negligible (0 s): Since we use the pseudo-questions generated by query textification (implemented by simple rules) as the corresponding question of a candidate query, this time cost is negligible.  
+(6). In-Context Learning QA (0.78 s): This is a simple ICL request with about 700 input tokens; a normal inference speed is generally less than 1 second. The time difference between using locally deployed open-source models (Qwen) and online closed-source models (GPT) is not significant. For closed-source models (GPT), a normal request speed is generally on the order of seconds. For open-source models, the specific speed depends on your hardware configuration. We deployed a 7B model using a single 80GB A100 GPU and
+
+a 72B model using eight A100 GPUs, leveraging vLLM (Kwon et al., 2023) for deployment. Using more devices can further speed up.
+
+# F.5 Clarifications on the Computational Details of CPQ
+
+We follow the OpenAI official price calculation method  ${}^{3}$  as reported in QueryAgent (Huang et al., 2024) to calculate CPQ in Table 3. TARGA only relying on an open-source model, except for Section 3.2 use ada-v2 for obtain the embedding and question. However, this part of the cost is less than 1e-5 $ per question, therefore, the CPQ is 0 $ when retaining three decimal places in Table 3.
+
+The cost in Section 3.2 consists of two parts: relation embedding and question embedding. The cost of caching embeddings for all relations in Freebase (20K relations) is less than  $0.01 \, \text{s}$ , and this cost is one-time; once cached, it does not increase regardless of the number of questions solved. For obtaining the embeddings for the test question itself, the average cost per question is less than 1e-5  $\text{s}$ , which rounds to approximately  $0.000 \, \text{s}$  when
+
+retaining three decimal places in Table 3. In summary, the close-source model invocations cost of TARGA is negligible.
+
+# F.6 Analysis of Token Consumption
+
+Since TARGA basically does not rely on closed-source models, to make the inference cost more comparable, we additionally provide a comparison of the total tokens consumed (input tokens + output tokens) for a more comprehensive evaluation. The result is shown in Table 14 TARGA has significant advantages in terms of token usage, with inference costs being less than one-tenth of those of other methods. Note that the substantial difference in token consumption by KB-Binder across the two datasets primarily stems from the use of more demonstration: 100 shots for GraphQ compared to 40 shots for GrailQA.
+
+<table><tr><td></td><td>GrailQA</td><td>GraphQ</td></tr><tr><td>KB-Binder</td><td>6,138 + 189</td><td>15, 166 + 206</td></tr><tr><td>AgentBench</td><td>19, 783 + 934</td><td>20, 143 + 968</td></tr><tr><td>QueryAgent</td><td>12, 277 + 391</td><td>13, 453 + 391</td></tr><tr><td>TARGA</td><td>638 + 36</td><td>734 + 38</td></tr></table>
+
+Table 14: Average consumed tokens (input + output).
+
+# F.7 Case Study
+
+Table 15 presents a detailed demonstration of how TARGA operates with a new question.
+
+<table><tr><td>Pipeline</td><td>Details</td></tr><tr><td>Question</td><td>the camera with a sensor from canon and a compression format of JPEG (exif 2.21) uses which viewfinder?</td></tr><tr><td>Candidate Entities</td><td>{canon: m.01bvx1,JPEG (exif 2.21): m.03h4lt3}</td></tr><tr><td>Candidates Relations</td><td>[digicams.digital_camera.viewfinder_type,digicams.camera_compressed_format.cameras,digicams.digital_camera(sensor_type,digicams.camera_uncompressed_format.cameras,digicams.cameraIso.cameras,digicams.camera_format.cameras,digicams.camera_sensor_manufacturer.cameras,...,digicams.camera_white_balance.cameras]</td></tr><tr><td>Query Construction</td><td>Query 1 triplet([jpeg (exif 2.21)], digicams.camera_compressed_format.cameras, ?v0) answer(?v0)Query 2 triplet([jpeg (exif 2.21)], digicams.camera_compressed_format.cameras, ?v0) triplet(?v0, digicams.digital-camera.viewfinder_type, ?v1) answer(?v1)......Query 56 triplet([jpeg (exif 2.21)], digicams.camera_compressed_format.cameras, ?v0) triplet(?v0, digicams.digital-camera.viewfinder_type, ?v1) triplet([canon], digicams.camera SENSOR_manufacturer.cameras, ?v0) answer(?v1)......Query 71 triplet([canon], digicams.digital-camera_manufacturer.cameras, ?v0) triplet(?v0, digicams.camera_storage_typeCompatible Cameras, ?v1) answer(?v1)</td></tr><tr><td>Query Re-ranking</td><td>Rank 1 triplet([jpeg (exif 2.21)], digicams.camera_compressed_format.cameras, ?v0) triplet(?v0, digicams.digitalcamera.viewfinder_type, ?v1) triplet([canon], digicams.camera SENSOR_manufacturer.cameras, ?v0) answer(?v1)Rank 2 triplet([jpeg (exif 2.21)], digicams.camera_compressed_format.cameras, ?v0) triplet(?v0, digicams.digitalcamera.viewfinder_type, ?v1) triplet([canon], digicams.digital-camera_manufacturer.cameras, ?v0) answer(?v1)......Rank 71 triplet([canon], digicams.digitalcamera_manufacturer.cameras, ?v0) triplet(?v0, digicams.camera(storage_typeCompatible Cameras, ?v1) answer(?v1)</td></tr><tr><td>Question Generation</td><td>Query 1
+ triplet([jpeg (exif 2.21)], digicamslicer_compressed_format.cameras, ?v0)
+ triplet(?v0, digicams.digital Cameras.viewfinder_type, ?v1)
+ triplet([canon], digicams.camerasicosensor_manufacturer.cameras, ?v0)
+ answer(?v1)
+Generated Question
+what viewfinder_type, jpeg (exif 2.21) has cameras, cameras has viewfinder_type, canon has cameras</td></tr><tr><td></td><td>Query 2
+ triplet([jpeg (exif 2.21)], digicamslicer_compressed_format.cameras, ?v0)
+ triplet(?v0, digicams.digital Cameras.viewfinder_type, ?v1)
+ triplet([canon], digicams.digital Camerasmanufacturer.cameras, ?v0)
+ answer(?v1)
+Generated Question
+what viewfinder_type, jpeg (exif 2.21) has cameras, cameras has viewfinder_type, canon has cameras</td></tr><tr><td></td><td>......</td></tr><tr><td></td><td>Query 10
+ triplet([jpeg (exif 2.21)], digicams.cameras_compressed_format.cameras, ?v0)
+ triplet(?v0, digicams.digital Camerassensor_resolution, ?v1)
+ triplet([canon], digicams.digital Camerasmanufacturer.cameras, ?v0)
+ answer(?v1)
+Generated Question
+what sensor_resolution, jpeg (exif 2.21) has cameras, cameras has sen-sor_resolution, canon has cameras</td></tr><tr><td>QA Prompt</td><td>You are a powerful model for generating PgQL queries to answer natural language questions. Here are some exemplars: 
+#####Question 
+what viewfinder_type,.jpeg (exif 2.21) has cameras, cameras has viewfinder_type, canon has cameras 
+#####PgQL 
+triplet([jpeg (exif 2.21)], digicamslicer_compressed_format.cameras, ?v0) 
+triplet(?v0, digicams.digital Cameras.viewfinder_type, ?v1) 
+triplet([canon], digicams&amp;sensor_manufacturer.cameras, ?v0) 
+answer(?v1) 
+#####Question 
+what viewfinder_type,.jpeg (exif 2.21) has cameras, cameras has viewfinder_type, canon has cameras 
+#####PgQL 
+triplet([jpeg (exif 2.21)], digicamslicer_compressed_format.cameras, ?v0) 
+triplet(?v0, digicams.digital Cameras.viewfinder_type, ?v1) 
+triplet([canon], digicamsdigital Camerasmanufacturer.cameras, ?v0) 
+answer(?v1) 
+……(8 demonstrations omitted here.) 
+Please follow the format of exemplars and output PgQL query for the following question. No explanation or questioning allowed. 
+Entity List: ['jpeg (exif 2.21)', 'canon'] 
+#####Question 
+the camera with a sensor from canon and a compression format of JPEG (exif 2.21) uses which viewfinder? 
+#####PgQL</td></tr><tr><td>LLM Output</td><td>triplet([jpeg (exif 2.21)], digicamslicer_compressed_format.cameras, ?v0) 
+triplet(?v0, digicams.digital Cameras.viewfinder_type, ?v1) 
+triplet([canon], digicams&amp;sensor_manufacturer.cameras, ?v0) 
+answer(?v1)</td></tr><tr><td>Parsing SPARQL</td><td>SELECT DISTINCT ?v1 WHERE { 
+:m.03h4lt3 :digicamslicer_compressed_format.cameras ?v0 . 
+?v0 :digicams.digital Cameras.viewfinder_type ?v1 . 
+:m.01bvx1 :digicams&amp;sensor_manufacturer.cameras ?v0 . }</td></tr><tr><td>Execution Result</td><td>{m.01xrg1f: Optical, m.01xrg6z: TTL}</td></tr><tr><td>F1 Score</td><td>1.0</td></tr></table>
+
+Table 15: A case study of TARGA from GrailQA.
+
+# Prompt for KBQA
+
+You are a powerful model for generating PgQL queries to answer natural language questions. Here are some exemplars:
+
+Question
+
+what football_league_system, football_league_system has conference premier
+
+##PyQL
+
+triplet(?v0, soccer.football_league_system.leagues, [conference premier])
+
+answer(?v0)
+
+Question
+
+what leagues, football_league_system has conference premier, football_league_system has leagues
+
+##PyQL
+
+triplet(?v0, soccer.football_league_system.leagues, [conference premier])
+
+triplet(?v0, soccer.football_league_system.leagues, ?v1)
+
+answer(?v1)
+
+Question
+
+what sport, sport has conference premier
+
+##PyQL
+
+triplet(?v0, sports.sport.leagues, [conference premier])
+
+answer(?v0)
+
+(5 demonstrations omitted here.)
+
+Question
+
+what positions, sport has conference premier, sport has positions
+
+##PyQL
+
+triplet(?v0, sports.sport.leagues, [conference premier])
+
+triplet(?v0, sports.sport.positions, ?v1)
+
+answer(?v1)
+
+Question
+
+what team_coaches, sport has conference premier, sport has team_coaches
+
+```python
+>>>PyQL
+
+triplet(?v0, sports.sport.leagues, [conference premier])
+
+triplet(?v0, sports.sport.team_coaches, ?v1)
+
+answer(?v1)
+
+Please follow the format of exemplars and output PgQL query for the following question. No explanation or questioning allowed.
+
+Entity List: ['conference premier']
+
+Question
+
+what are the names of the football leagues that are in the same football league system with conference premier?
+
+##PyQL
+
+Table 16: Prompt for KBQA.
+
+# Prompt for Text2SQL
+
+You are a powerful model for generating SQL queries to answer natural language questions.
+
+Here are some related exemplars you can learn from:
+
+SELECT MIN(Magnitude (M bol )) FROM TABLE WHERE Radius  $(\mathbb{R}\odot) = '10'$
+
+SELECT MIN(Mass (M ⊙)) FROM TABLE WHERE Radius (R ⊙) = '10'
+
+SELECT MIN(Spectral type) FROM TABLE WHERE Radius  $(\mathbb{R}\odot) = '10'$
+
+SELECT MIN(Star(Pismis24-#)) FROM TABLE WHERE Radius  $(\mathbb{R}\odot) = '10'$
+
+SELECT MIN(Radius (R ⊙)) FROM TABLE WHERE Radius (R ⊙) = '10'
+
+SELECT MIN(Mass (M  $\odot$ ) FROM TABLE WHERE Star (Pismis24-#) = '10'
+
+SELECT MIN(Temperature (K)) FROM TABLE WHERE Radius  $(\mathrm{R}\odot) = '10'$
+
+SELECT MIN(Radius (R  $\odot$  )) FROM TABLE WHERE Star (Pismis24-#) = '10'
+
+SELECT MIN(Magnitude (M bol )) FROM TABLE WHERE Star (Pismis24-#) = '10'
+
+SELECT MIN(Spectral type) FROM TABLE WHERE Star (Pismis24-#) = '10'
+
+Please output SQL query for the following question. No explanation or questioning allowed. Note that there is no need to use the LIKE keyword. And table name is TABLE for all questions.
+
+Question
+
+If a radius is 10, what is the lowest possible mass?
+
+```header
+
+['Star (Pismis24-#)', 'Spectral type', 'Magnitude (M bol)', 'Temperature (K)', 'Radius (R ⊙)', 'Mass (M ⊙)']
+
+SQL
+
+Table 17: Prompts for Text2SQL.

2025/TARGA_ Targeted Synthetic Data Generation for Practical Reasoning over Structured Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:169b001747ff919d2f216524037878bc201fd8d2cfc9c3b828038396ea9c79c5
+size 1280308

2025/TC–RAG_ Turing–Complete RAG’s Case study on Medical LLM Systems/08530356-6e71-43b5-b2a0-ddb4582c96bf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f1e749f98a8ee9135b9dac6a3b3701cc08a4432b79f88150e2a96bc2a77bc71
+size 1839549

2025/TC–RAG_ Turing–Complete RAG’s Case study on Medical LLM Systems/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42c74143e9f4d3f18a75f0b85417a10a68bb1c8449aa5f5d58e1906a29b3df09
+size 1100131

2025/TEACH_ A Contrastive Knowledge Adaptive Distillation Framework for Classical Chinese Understanding/aecc946d-77dc-4482-aa73-533ce4e2f50d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07dc650faffe7470508bf2bdd1fe9af14506dfce401b6d415c9f62d7b36a8c02
+size 1427639

2025/TEACH_ A Contrastive Knowledge Adaptive Distillation Framework for Classical Chinese Understanding/full.md

@@ -0,0 +1,462 @@
+# TEACH: A Contrastive Knowledge Adaptive Distillation Framework for Classical Chinese Understanding
+
+Yuting Wei, Qi Meng, Yuanxing Xu, Bin Wu*
+
+Beijing Key Laboratory of Intelligent Telecommunication Software and Multimedia,
+
+Beijing University of Posts and Telecommunications
+
+{yuting_wei, mikemeng09, xyx, wubin} @bupt.edu.cn
+
+# Abstract
+
+Traditional methods for processing classical Chinese typically segment language understanding into discrete tasks, which overlook crucial background information and reduce user engagement. Large language models (LLMs) provide integrated solutions, yet they entail high computational costs and risks of generating inaccurate historical information. To tackle these challenges, we propose a novel framework, TEACH (conTrastive knowlEdge Adaptive distillation with enhanced Historical interpretability), which focuses on classical Chinese understanding by integrating word sense disambiguation with sentence translation. This integration leverages a confidence-annotated knowledge base and a step-by-step Chain-of-Thought prompting mechanism to minimize hallucinations and improve semantic analysis. Moreover, TEACH employs contrastive distillation learning to efficiently transfer capabilities from larger models to smaller ones (e.g., Qwen2-1.5B), addressing overly liberal translations. Additionally, we introduce an innovative generation evaluation metric using iterative word alignment, enhancing LLM performance assessments by distinguishing additional information and addressing excessive translation issues. Experiments conducted on real-world datasets validate TEACH's efficacy in classical Chinese educational scenarios. $^{1}$
+
+# 1 Introduction
+
+Classical Chinese literature, as a crucial carrier of millennia-old cultural heritage, plays an irreplaceable role in historical understanding and cultural transmission. However, understanding classical Chinese poses numerous challenges, such as scarce corpora, complex linguistic structures, variable word meanings, and rich historical and cultural contexts, all of which increase the difficulty
+
+of comprehension (Zhang et al., 2024; Wei et al., 2024b). With the rapid development of natural language processing technologies, more researchers are focusing on classical Chinese to address these challenges (Li et al., 2022; Liu et al., 2022; Xiang et al., 2024).
+
+The key tasks in understanding classical Chinese include word sense disambiguation (WSD) and translation. The WSD aims to select the most fitting meaning of a polysemous word within its context (Shu et al., 2021; Pan et al., 2022). For translation, traditional methods employ encoder-decoder approaches to achieve literal translation of sentences (Chang et al., 2021; Wang et al., 2023). However, the segmentation of these tasks hinders holistic analysis. Furthermore, traditional models frequently suffer from semantic inaccuracies due to inadequate historical background. Additionally, these models typically require coding skills, limiting their accessibility and reducing their suitability for educational applications.
+
+LLMs (Wei et al., 2022a; Zhao et al., 2023; Wei et al., 2022b) can easily engage with users and enhance language understanding by providing rich historical context, thereby improving sentence interpretability. This capability extends beyond basic word sense analysis and sentence translation, potentially increasing their interest in classical Chinese and historical studies (Wei et al., 2024a). Despite their strong generative capabilities and broad knowledge coverage, LLMs face challenges such as the occasional generation of hallucinations (Chang et al., 2024). This misleads users in educational scenarios. Additionally, LLMs entail high inference costs and substantial computational resources for training and deployment (Yang et al., 2024b). Consequently, it is a significant challenge to develop a smaller model that requires less data while ensuring efficient inference speed (Hu et al., 2022; Ding et al., 2023).
+
+To address these problems, we propose TEACH,
+
+a novel framework for classical Chinese understanding that integrates conTrastive knowlEdge Adaptive distillation with enhanCed Historical interpretability. This framework unifies word sense analysis and sentence comprehension, facilitating the transfer of historical knowledge and reasoning abilities from large to smaller models. Specifically, we first construct a confidence-annotated knowledge base including historical information and word annotations for retrieval augmentation. Then we design a step-by-step Chain-of-Thought (CoT) prompt to guide the large model in minimizing hallucinations and generating comprehensive contextual and semantic analysis. Through the distillation process, the small student model, with only  $0.8 - 4\%$  of the parameters of the large teacher model, acquires step-by-step thinking capabilities and becomes a proficient classical Chinese interpreter. To address the tendency of large models to produce liberal translations, we employ contrastive learning to train the model in the precise and literal style essential for classical Chinese understanding. Moreover, we introduce a generation evaluation metric based on iterative word alignment for LLMs, enabling a more accurate assessment of their performance by distinguishing additional information and addressing excessive translation issues. Our key contributions are as follows:
+
+- To the best of our knowledge, it is the first educational model for classical Chinese that combines word sense disambiguation and sentence translation tasks, while also providing historical interpretability.  
+- We propose TEACH, a novel contrastive knowledge adaptive distillation framework, which transfers reasoning capabilities from large to small models by step-by-step CoT prompting, significantly reducing hallucinations and liberal translations.  
+- We improve a generation evaluation metric based on iterative word alignment for LLMs, which considers the additional information and excessive translation.  
+- Experiments on real datasets demonstrate the effectiveness of our model, showing that it generates high-quality historical context analysis while also accurately understanding classical Chinese semantics.
+
+# 2 Related Work
+
+# 2.1 Classical Chinese Understanding
+
+Classical Chinese WSD and translation are two crucial tasks for understanding classical Chinese texts (Pan et al., 2022; Shu et al., 2021). WSD is typically approached as a classification challenge, where traditional methodologies deploy word embeddings that integrate contextual or annotated data, utilizing algorithms such as KNN to deduce word meanings (Loureiro and Jorge, 2019). Regarding translation, methods often leverage pretrained models specific to classical Chinese, such as SikuBERT (Wang et al., 2022) and AnchiBERT (Tian et al., 2021), which employ various encoder-decoder frameworks. For example, Guo et al. (2023) features the translation task by dual-syllable alignment substitution and dual-mask language modeling. Evidently, existing research isolates the two tasks. Moreover, traditional methods provide only answers, lacking interpretability, which is crucial for educational applications. Consequently, we leverage the extensive knowledge and robust reasoning abilities of LLMs to simultaneously address both tasks, generating interpretable historical background and semantic analyses.
+
+# 2.2 LLMs with RAG
+
+LLMs have demonstrated significant potential in downstream tasks such as recommendation systems (Zhu et al., 2024) and question answering (Li et al., 2024), thanks to their advanced reasoning capabilities. However, their deployment in sensitive areas is not without risks, notably the propensity to generate misleading or incorrect information, often referred to as "hallucinations". To address these challenges, Retrieval-Augmented Generation (RAG) technology has been employed to enhance the reliability of LLM outputs by incorporating domain-specific knowledge directly into the generation process (Lewis et al., 2020; Gao et al., 2023; Zhao et al., 2024). Further advancements in RAG technology include the development of graph-based approaches (Peng et al., 2024), which enhance the model's ability to handle complex datasets with interlinked information. For instance, in healthcare, RAG can leverage up-to-date research findings and patient-specific data to produce accurate diagnostic recommendations.
+
+For classical Chinese understanding, LLMs also face hallucinations, such as incorrect historical information and the misinterpretation of archaic
+
+![](images/ef524795da63d80c9fa1e5e8075b4c47c3c91e0fecfe986ffd7c3166df3a6589.jpg)  
+Figure 1: The overview of the proposed framework.
+
+terms. Hence, we propose a straightforward method for constructing a knowledge base to efficiently address this problem.
+
+# 2.3 Generation Evaluation Metrics
+
+Evaluating the generative capabilities of large models is crucial for assessing their effectiveness. Metrics utilized to assess the quality of generated content generally fall into three categories: those based on word overlap, word embeddings, and language models (Lee et al., 2023). Word overlap-based metrics such as BLEU (Papineni et al., 2002; Chen and Cherry, 2014) and ROUGE (Lin, 2004) measure quality by assessing the n-gram overlap between candidate and reference texts. Word embedding-based metrics convert sentences into vector representations and compare these within low-dimensional semantic spaces, using techniques such as Greedy Matching (Rus and Lintean, 2012) and Vector Extrema (Forgues et al., 2014). Language model-based metrics utilize pre-trained models to evaluate semantic and syntactic congruence between translated and reference texts (Zhang et al., 2020; Shin et al., 2024). Despite advancements, traditional metrics face challenges with LLM-generated classical Chinese translations, notably the inclusion of useful yet additional information and overly detailed explanations of original words. To overcome these issues, an iterative self-labeled word alignment method is employed to refine evaluation metrics.
+
+# 3 Methodology
+
+In this section, we introduce TEACH, an innovative contrastive knowledge adaptive distillation framework tailored for classical Chinese education.
+
+This approach integrates the reasoning capabilities of LLMs into classical Chinese language comprehension while enhancing historical explainability and optimizing resource efficiency. An overview of the framework is provided in Figure 1.
+
+# 3.1 Interpretable Knowledge Construction
+
+LLMs often suffer from hallucination problems, which can undermine their utility in educational applications. To mitigate this, we construct a knowledge base that integrates confidence-annotated historical information and authoritative annotations of classical Chinese word meanings, which assists in accurate and interpretable analysis and reasoning.
+
+# 3.1.1 Historical Information
+
+To obtain relevant historical information for a given sentence, we extract its surrounding context, defined as the ten sentences before and after the target sentence. Specifically, we utilize the complete text version of the Twenty-Four Histories, supplemented by novels, poetry, and other genres, to construct a retrieval database powered by Milvus. Details of this database are provided in Appendix A.1.
+
+To prevent misleading the model's analysis, we introduce a confidence metric for historical information:
+
+$$
+C o n f i d e n c e = 1 - \mathrm {B P} (t, q) \cdot \frac {L 2 (t , q)}{\sqrt {d}}, \tag {1}
+$$
+
+$$
+\mathrm {B P} (t, q) = \left\{ \begin{array}{l l} 1, & l _ {t} > l _ {q} \\ \exp \left(1 - l _ {q} / l _ {t}\right), & l _ {t} \leq l _ {q} \end{array} , \right. \tag {2}
+$$
+
+where BP is length penalty factor,  $l_{t}$  and  $l_{q}$  denote the lengths of the retrieved sentence  $t$  and the query  $q$ , respectively.  $L2$  is the Euclidean distance,  $d$  represents the dimension of the vectors.
+
+# 3.1.2 Word Sense Annotation
+
+Despite significant changes in the meanings of classical Chinese words compared to modern Chinese, annotations provide a bridge between them. We extract all words with a frequency greater than five from the classical Chinese corpus and collect their annotations from Handian $^2$ , ultimately obtaining 169,742 annotated classical words (see Appendix A.2 for details).
+
+# 3.2 Contrastive Knowledge Adaptive Distillation
+
+Our contrastive distillation strategy comprises two steps. First, it leverages CoT prompts enriched with interpretable knowledge to guide the LLM (the teacher) through step-by-step reasoning. Its multi-granularity reasoning strategy—ranging from macro-level historical and cultural context to micro-level linguistic features—further enhances the model's understanding and generalization. Then, contrastive learning with negative samples is employed for adaptive knowledge transfer, enabling the smaller models (the student) to effectively approximate the reasoning capabilities of the teacher. Beyond reasoning, TEACH uniquely incorporates style-preservation constraints tailored to the stylistic characteristics of classical Chinese translations, ensuring that the distilled model retains key cultural and literary features.
+
+# 3.2.1 Generating Step-by-step Reasoning by the Teacher
+
+We utilize zero-shot CoT prompting to elicit reasoning information from LLMs. Specifically, given the sentence  $s$  to be analyzed, we obtain the corresponding historical information  $\mathcal{H}$  and its related annotated word set  $\mathcal{A}$  from the constructed knowledge base. Then, we design a prompt template  $\mathcal{T}$  consisting of three progressive steps as follows:
+
+- Step1. Analyze the historical background and persons based on the given historical information  $\mathcal{H}$  with the relevance Confidence.  
+- Step2. Select the most appropriate word sense from the provided annotations  $\mathcal{A}$ .  
+- Step3. Provide a literal translation by combining the analyses from Steps1 and 2.
+
+The template starts with a macro-level paragraph perspective and gradually zooms into a micro-level
+
+word perspective, guiding LLMs in a step-by-step thinking process (see Figure 5 for details). Subsequently, we extract a very small sentence subdataset  $S'$  from the training set  $S$  by random sampling, where  $S' \subset S$  and  $|S'| \ll |S|$ . We further fill the template  $T$  with  $S'$  to generate corresponding CoT prompts  $\mathcal{P} = \{p_s | s \in S'\}$  for the teacher models. With these knowledge-augmented prompts  $\mathcal{P}$ , LLMs will generate corresponding step-by-step interpretable analysis  $o_s$  for each input  $p_s$ .
+
+# 3.2.2 Training the student by contrastive distillation
+
+High-quality analysis can be generated from LLMs. However, their extensive computational demands make them unsuitable for low-latency educational applications. Additionally, LLMs often produce liberal translations that do not suit the needs for literal translations essential in understanding classical Chinese texts. To overcome these challenges, we employ contrastive knowledge distillation, adaptively transferring the comprehension and reasoning capabilities of larger teacher models to more computationally efficient smaller student models.
+
+Specifically, we fine-tune a small model using the same prompts  $p_{s}$  as input. For the output, we have two types: 1) the original response  $o_{s}$ , which includes the liberal translation  $t_{\text{liberal}}$ , and 2) the expected literal response  $e_{s}$ , where  $t_{b}$  is replaced with the standard literal translation  $t_{\text{literal}}$ . Our goal is to generate outputs that align more closely with the latter.
+
+The training comprises two parts. First, we use  $e_{s}$  as the output labels and optimize the negative log-likelihood of the conditional language modeling objective:
+
+$$
+\mathcal {L} _ {d i s} = \sum_ {i = 1} ^ {| e |} \log \left(P _ {\theta} \left(e _ {s, i} \mid p _ {s}, e _ {s, <   i}\right)\right), \tag {3}
+$$
+
+where  $e_{s,i}$  is the  $i$ -th token of  $e_s$ , and  $e_{s, < i}$  represents the tokens before  $e_{s,i}$ . To conserve resources, we employ LoRA (Hu et al., 2022) for parameter-efficient model fine-tuning with parameters  $\theta$ . Second, we adopt a contrastive learning-based adaptive distillation technique. We utilize  $o_s$  as negative samples and  $e_s$  as the positive sample. Similar to Zeng et al. (2024), we directly tune the language model using a token-level contrastive loss:
+
+$$
+\mathcal {L} _ {\text {c o n}} = \frac {1}{N - M} \sum_ {m = M} ^ {N} \max  \left(0, - f _ {\beta} \left(h _ {m} ^ {e _ {s}}\right) + f _ {\beta} \left(h _ {m} ^ {o _ {s}}\right) + 1\right), \tag {4}
+$$
+
+where  $N$  is the maximum length of two sequences, and  $h_m^{e_s}$  and  $h_m^{o_s}$  are the hidden states of the  $m$ -th token of the preferred literal output  $e_s$  and the comparison liberal output  $o_s$ , respectively.  $M$  is the index starting from the segments that differ (i.e., translation parts) between  $e_s$  and  $o_s$ .  $f_\beta$  is a linear head that takes the hidden state of the top layer and returns a scalar.
+
+The overall loss function for tuning the model is:
+
+$$
+\mathcal {L} = - \sum_ {s \in \mathcal {S} ^ {\prime}} \left(\mathcal {L} _ {d i s} + \mathcal {L} _ {c o n}\right). \tag {5}
+$$
+
+# 4 Refined Evaluation Metric for LLMs
+
+When evaluating the generated translations of LLMs, traditional metrics often overlook two key aspects: the overly detailed explanation of original terms (term1) and the useful but additional information (word2), as shown in Figure 2. These issues frequently arise in LLMs. To address them, we develop an iterative self-labeled word alignment method to refine BLEU and ROUGE, enabling more accurate assessments of translation quality.
+
+# 4.1 Word Alignment Self-Labeling (WAS)
+
+The goal of WAS is to establish word alignment between a reference text  $R = (r_1, r_2, \ldots, r_n)$  and a candidate text  $C = (c_1, c_2, \ldots, c_m)$ . We define the alignment relationship between  $R$  and  $C$  as  $A_{ij}$ , which represents the probability of alignment between  $r_i$  and  $c_j$ . The word alignment challenge is formulated as identifying matrix  $A$  that maximizes sentence similarity:
+
+$$
+\max  _ {A} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {m} A _ {i j} f _ {\text {s i m}} \left(r _ {i}, c _ {j}\right), \tag {6}
+$$
+
+where  $f_{sim}(\cdot)$  denotes the similarity function. To solve this equation, we reformulate it as an optimal transport problem, referencing Peyre et al. (2019) and Chi et al. (2021). Then, we obtain the solved initial matrix  $A'$ . For the detailed solving process, please refer to Appendix B.
+
+The alignment labels  $\mathcal{B}_{ij}$  are extracted using the following formula:
+
+$$
+\mathcal {B} _ {i j} = \left\{ \begin{array}{l l} 1 & i = \left(\arg \max  _ {k} A _ {i k} ^ {\prime}\right) \cap (j = \arg \max  _ {k} A _ {k j} ^ {\prime}) \\ 0 & \text {o t h e r w i s e} \end{array} . \right. \tag {7}
+$$
+
+To maximize the identification of alignment labels, we employ an iterative approach to update  $A^{\prime}$  and  $\mathcal{B}$ , following Sabet et al. (2020). Initially,  $A^{\prime}$  is
+
+![](images/999801ff8010b821b1a935233a67df0b6e12e8ec12d8d6bace749d6c83b2167a.jpg)  
+Figure 2: An example of translation outputs from LLMs.
+
+updated according to the initial state of  $\mathcal{B}$  using the following formula:
+
+$$
+A _ {i j} ^ {\prime} = \left\{ \begin{array}{l l} 0, & \mathcal {B} _ {i j} = 1 \\ \alpha A _ {i j} ^ {\prime}, & \exists k (\mathcal {B} _ {i k} = 1 \vee \mathcal {B} _ {k j} = 1), \\ A _ {i j} ^ {\prime}, & \text {o t h e r w i s e} \end{array} \right. \tag {8}
+$$
+
+where  $\alpha$  is a discount factor. Then,  $\mathcal{B}$  is updated according to Eq.equation 7 for those entries where preceding  $\mathcal{B}_{ij} = 0$ . After multiple iterations, this method yields high-precision self-labeled alignment labels  $\mathcal{B}$ .
+
+To address both one-to-many and many-to-one alignments, we implement a merging strategy. Specifically, when a token  $j$  in the translated text  $C$  aligns with multiple consecutive tokens in the reference text  $R$ , or vice versa, these tokens are merged. Furthermore, tokens in  $C$  without correspondences in  $R$  are considered additional information and are thus concealed (e.g.,  $\text{word}_2$ ). This process results in the merged token sequences  $C'$  and  $R'$ , alongside one-to-one alignment labels  $\mathcal{B}'$ .
+
+# 4.2 Refined BLEU based on WAS
+
+In the adaptation of the BLEU metric for the WASText, we introduce  $\mathrm{BLEU}_{\mathrm{WAS}}$  defined as follows:
+
+$$
+\mathrm {B L E U} _ {\mathrm {W A S}} = \mathrm {B P} \left(C ^ {\prime}, R ^ {\prime}\right) \cdot \exp \left(\sum_ {n = 1} ^ {N} \frac {1}{N} \log p _ {n}\right), \tag {9}
+$$
+
+$$
+p _ {n} = \frac {\sum_ {c \in C ^ {\prime}} \operatorname {C o u n t} \left(c , R ^ {\prime} , \mathcal {B} ^ {\prime} , n\right)}{\left| C ^ {\prime} \right| - (n - 1)}, \tag {10}
+$$
+
+$$
+\mathrm {B P} \left(C ^ {\prime}, R ^ {\prime}\right) = \exp \left(1 - \frac {\max \left(l _ {C ^ {\prime}} , l _ {R ^ {\prime}}\right)}{\min \left(l _ {C ^ {\prime}} , l _ {R ^ {\prime}}\right)}\right), \tag {11}
+$$
+
+where,  $\mathrm{BP}(C', R')$  represents the sentence length penalty factor, which penalizes both overly short (incomplete translation) and overly long (excessive translation) sentences.  $\mathrm{Count}(c, R', \mathcal{B}', n)$  quantifies the occurrences of the  $n$ -gram of  $c$  within  $R'$ , as dictated by the alignment labels  $\mathcal{B}'$ .
+
+# 4.3 Refined ROUGE based on WAS
+
+The ROUGE metric is designed to assess recall in translation evaluations. In our adaptation, we introduce ROUGE- $\mathrm{N}_{\mathrm{WAS}}$ , which is formulated as follows:
+
+$$
+\mathrm {R O U G E} - \mathrm {N} _ {\mathrm {W A S}} = \frac {\sum_ {c \in C ^ {\prime} , r \in R ^ {\prime}} \mathrm {B P} (c , r) \cdot \operatorname {C o u n t} \left(c , r , \mathcal {B} ^ {\prime} , n\right)}{\left| R ^ {\prime} \right| - (N - 1)}, \tag {12}
+$$
+
+where  $\mathrm{BP}(c,r)$  represents the token length penalty factor, calculated similar to Eq. equation 11.  $\mathrm{Count}(c,r,\mathcal{B}',n)$  quantifies the occurrences where the  $n$ -grams of  $c$  and  $r$  coincide, as dictated by the alignment labels  $\mathcal{B}'$ .
+
+# 5 Experiments
+
+# 5.1 Experiment Settings
+
+# 5.1.1 Datasets
+
+Despite the absence of datasets that jointly address translation and semantic analysis, the accuracy of translations can indirectly reflect the precision of WSD. Therefore, we focus on the Erya dataset, which is the largest dataset for classical Chinese translation tasks. It covers a comprehensive range of historical periods and genres, including poetry, prose, philosophy works, and literary criticism.
+
+# 5.1.2 Baselines
+
+We compare against two types of baseline models: the classical classical Chinese translation model Erya, and various LLMs. Specifically, Erya is the most advanced classical Chinese translation model available. For large models, we select popular Chinese large models in both base and chat versions, such as Yi1.5-6B (Young et al., 2024), ChatGLM3-6B (Du et al., 2022; Zeng et al., 2023), GLM-4-9B (GLM et al., 2024), Qwen2-1.5B/7B (Yang et al., 2024a). We also evaluate Xunzi-Qwen2<sup>3</sup> which is finetuned on the Qwen2 model using classical Chinese corpus.
+
+# 5.1.3 Evaluation Metrics
+
+We employ three categories of evaluation metrics to comprehensively assess translation quality. First, classic token-based metrics, including BLEU and ROUGE-N, focus on surface-form overlap. Second, semantic-based metrics, such as BERTScore and METEOR, evaluate the preservation of meaning and semantic similarity. Third, we use refined metrics  $\mathrm{BLEU}_{\mathrm{WAS}}$  and ROUGE-  $\mathrm{N_{WAS}}$  , which incorporate contextual word alignment signals to better align with human judgment. Specifically, we
+
+set  $N = 4$  for BLEU and  $\mathrm{BLEU}_{\mathrm{WAS}}$ , while  $N = 1$  for ROUGE-N, ROUGE- $\mathrm{N}_{\mathrm{WAS}}$ . Human evaluation is also provided in Table 4 to further analyze the quality of word sense selection and translation.
+
+# 5.1.4 Implementation Details
+
+For the teacher model, we employ the ERNIE-Bot4 $^{4}$  API from Baidu, which has been demonstrated by Wei et al. (2024a) to achieve state-of-the-art performance in classical Chinese language understanding. We randomly select 3,000 entries from the training dataset to generate high-quality reasoning data using a zero-shot, step-by-step CoT, as proven optimal in Appendix C.1. This data is further refined and corrected by experts to ensure its accuracy. For the student model, we select 10,000 entries from the validation set for testing. All models are fine-tuned for 10 epochs with a batch size of 64, a maximum input length of 1024, and an output length of 512. The learning rate is set at  $5e - 5$  for LoRA. Fine-tuning is conducted on 1 A100 GPU with 40G of memory. For comparison, we also employ the minimal prompt template as the Normal method: "Translate from classical Chinese to modern Chinese. Classical Chinese: src. Modern Chinese:" extracting  $trg$  from their outputs using regular expressions. Both training and testing set sizes are consistent with those used for TEACH. For the refined evaluation metric, aligning 10,000 sentence pairs using AnchiBERT takes approximately 2-3 minutes on an NVIDIA 3090 GPU, with a batch size of 8 and 2 iterations, without any additional pre-training.
+
+Additionally, Appendix C.2 offers a detailed comparison of the teacher and student models, focusing on model size, deployment complexity, cost, and functionality to assess their real-world applicability.
+
+# 5.2 Main Results
+
+Our TEACH framework is highly flexible and can be seamlessly integrated with any LLM backbone. To validate the effectiveness of TEACH, we first present the results from classic translation models and LLM trained with Normal prompt as baselines for comparison. Then, we evaluate the performance of TEACH across various LLMs. Results are shown in Table 1. Based on these evaluations, we make the following observations: (1) TEACH significantly outperforms Normal prompts, achieving performance improvements ranging from 4-
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Type</td><td rowspan="2">Method</td><td colspan="7">Metrics</td></tr><tr><td>BLEU</td><td>ROUGE</td><td>BERTScore</td><td>METEOR</td><td>BLEUWAS</td><td>ROUGEWAS</td><td>Avg</td></tr><tr><td>Erya</td><td>-</td><td>-</td><td>18.39</td><td>50.15</td><td>76.11</td><td>42.39</td><td>39.68</td><td>71.18</td><td>49.65</td></tr><tr><td rowspan="4">Yi1.5-6B</td><td rowspan="2">Base</td><td>Normal</td><td>16.81</td><td>47.96</td><td>73.97</td><td>41.36</td><td>42.69</td><td>74.13</td><td>49.49</td></tr><tr><td>TEACH</td><td>19.23 (14.40%)</td><td>51.91 (8.24%)</td><td>76.80 (3.83%)</td><td>45.62 (10.30%)</td><td>44.06 (3.21%)</td><td>74.82 (0.93%)</td><td>51.77 (4.61%)</td></tr><tr><td rowspan="2">Chat</td><td>Normal</td><td>15.71</td><td>46.68</td><td>73.32</td><td>40.09</td><td>40.81</td><td>73.31</td><td>48.49</td></tr><tr><td>TEACH</td><td>17.70 (12.67%)</td><td>50.44 (8.05%)</td><td>76.01 (3.67%)</td><td>43.81 (9.28%)</td><td>42.21 (3.43%)</td><td>73.89 (0.79%)</td><td>50.76 (4.69%)</td></tr><tr><td rowspan="4">ChatGLM3-6B</td><td rowspan="2">Base</td><td>Normal</td><td>15.62</td><td>46.55</td><td>73.32</td><td>39.66</td><td>40.11</td><td>72.93</td><td>47.92</td></tr><tr><td>TEACH</td><td>18.77 (20.17%)</td><td>51.25 (10.08%)</td><td>76.10 (3.79%)</td><td>44.43 (12.03%)</td><td>42.93 (7.03%)</td><td>74.01 (1.48%)</td><td>50.86 (6.14%)</td></tr><tr><td rowspan="2">Chat</td><td>Normal</td><td>14.27</td><td>44.59</td><td>72.37</td><td>37.48</td><td>37.48</td><td>71.36</td><td>46.33</td></tr><tr><td>TEACH</td><td>16.90 (18.43%)</td><td>48.72 (9.26%)</td><td>75.59 (4.45%)</td><td>39.83 (6.27%)</td><td>39.88 (6.40%)</td><td>73.15 (2.51%)</td><td>49.29 (6.39%)</td></tr><tr><td rowspan="4">GLM-4-9B</td><td rowspan="2">Base</td><td>Normal</td><td>17.99</td><td>49.56</td><td>74.82</td><td>42.77</td><td>43.06</td><td>74.35</td><td>50.43</td></tr><tr><td>TEACH</td><td>20.46 (13.73%)</td><td>53.15 (7.24%)</td><td>76.86 (2.73%)</td><td>46.78 (9.38%)</td><td>46.41 (7.78%)</td><td>75.98 (2.19%)</td><td>52.94 (4.99%)</td></tr><tr><td rowspan="2">Chat</td><td>Normal</td><td>17.11</td><td>49.22</td><td>74.74</td><td>42.11</td><td>42.82</td><td>73.99</td><td>49.83</td></tr><tr><td>TEACH</td><td>19.73 (15.31%)</td><td>50.59 (2.78%)</td><td>75.12 (0.51%)</td><td>44.50 (5.68%)</td><td>43.03 (0.49%)</td><td>73.64 (0.48%)</td><td>52.37 (5.10%)</td></tr><tr><td rowspan="4">Qwen2-1.5B</td><td rowspan="2">Base</td><td>Normal</td><td>14.98</td><td>46.04</td><td>72.84</td><td>37.93</td><td>36.53</td><td>70.35</td><td>48.50</td></tr><tr><td>TEACH</td><td>18.08 (20.69%)</td><td>50.26 (9.17%)</td><td>75.65 (3.86%)</td><td>43.38 (14.37%)</td><td>41.32 (13.10%)</td><td>71.49 (1.62%)</td><td>51.67 (6.54%)</td></tr><tr><td rowspan="2">Chat</td><td>Normal</td><td>14.16</td><td>44.23</td><td>71.82</td><td>37.20</td><td>37.39</td><td>71.49</td><td>46.33</td></tr><tr><td>TEACH</td><td>17.76 (25.42%)</td><td>50.02 (13.09%)</td><td>75.47 (5.08%)</td><td>43.16 (16.02%)</td><td>41.32 (10.51%)</td><td>73.19 (2.38%)</td><td>49.77 (7.44%)</td></tr><tr><td rowspan="4">Qwen2-7B</td><td rowspan="2">Base</td><td>Normal</td><td>17.06</td><td>48.35</td><td>74.29</td><td>41.86</td><td>43.13</td><td>74.31</td><td>50.40</td></tr><tr><td>TEACH</td><td>21.88 (28.25%)</td><td>54.41 (12.53%)</td><td>77.45 (4.25%)</td><td>47.87 (14.36%)</td><td>45.93 (15.75%)</td><td>75.59 (1.72%)</td><td>53.74 (6.64%)</td></tr><tr><td rowspan="2">Chat</td><td>Normal</td><td>17.13</td><td>48.39</td><td>74.28</td><td>41.84</td><td>43.10</td><td>74.45</td><td>49.64</td></tr><tr><td>TEACH</td><td>20.09 (17.28%)</td><td>52.66 (8.82%)</td><td>76.74 (3.31%)</td><td>46.07 (10.11%)</td><td>45.00 (4.41%)</td><td>75.19 (0.99%)</td><td>52.40 (5.56%)</td></tr><tr><td rowspan="2">Xunzi-Qwen2-1.5B</td><td rowspan="2">Base</td><td>Normal</td><td>14.75</td><td>44.94</td><td>72.32</td><td>38.13</td><td>38.63</td><td>71.86</td><td>46.57</td></tr><tr><td>TEACH</td><td>19.70 (33.56%)</td><td>51.88 (15.44%)</td><td>76.22 (5.39%)</td><td>44.93 (17.83%)</td><td>43.04 (11.42%)</td><td>73.76 (2.64%)</td><td>51.76 (11.16%)</td></tr><tr><td rowspan="2">Xunzi-Qwen2-7B</td><td rowspan="2">Base</td><td>Normal</td><td>18.78</td><td>50.21</td><td>75.13</td><td>43.72</td><td>44.48</td><td>74.91</td><td>51.51</td></tr><tr><td>TEACH</td><td>23.23 (23.70%)</td><td>55.51 (10.56%)</td><td>78.14 (4.01%)</td><td>49.25 (12.65%)</td><td>46.31 (4.11%)</td><td>75.75 (1.12%)</td><td>54.70 (6.19%)</td></tr></table>
+
+$11\%$  in average metrics. This substantial increase validates the efficacy of our framework. (2) While LLMs with Normal prompts generally do not surpass classic models in traditional metrics, they perform competitively or even better on refined metrics. This suggests that despite different from expected literal translations, LLMs can accurately interpret sentence meanings. (3) Our refined metrics outperform both token-based and semantic-based ones, showing stronger ability to assess additional knowledge conveyed by LLMs. (4) Base models usually outperform chat models, as the former are more tailored to specific tasks. (5) The Qwen series models exhibit superior performance. Among them, the Qwen-1.5B-Base with TEACH, which has a smaller parameter size for easier deployment, remains competitive across all metrics. (6) While Xunzi-Qwen2-7B benefits from pretraining on classical Chinese corpus, our TEACH framework further improves its performance by leveraging a macro-to-micro prompting strategy, enabling more accurate and interpretable outputs. (7) Beyond mere translation outputs, LLMs with TEACH offer macro historical and micro word sense analyses. This enhances interpretability and practicality, making LLMs particularly suitable for educational scenarios.
+
+# 5.3 Analysis
+
+To comprehensively examine our framework and mitigate the effects of data leakage from the XunziLLMs, we utilize its original Qwen2 as backbones.
+
+Table 1: Comparison of performance among different backbones fine-tuning with Normal/TEACH method. "Avg" indicates the average of four metrics, while "(.)" show the relative improvement of TEACH compared to Normal.  
+
+<table><tr><td>ID</td><td>Model Type</td><td>Trad. Avg</td><td>Refined Avg</td></tr><tr><td>1</td><td>Qwen2-7B (TEACH CoT, trained)</td><td>37.95</td><td>61.09</td></tr><tr><td>2</td><td>Qwen2-7B (TEACH CoT, untrained)</td><td>29.97</td><td>55.97</td></tr><tr><td>3</td><td>Qwen2-7B (Normal CoT, trained)</td><td>31.36</td><td>57.44</td></tr><tr><td>4</td><td>Qwen2-7B (Normal CoT, untrained)</td><td>34.40</td><td>57.08</td></tr><tr><td>5</td><td>ERNIE-Bot4 (TEACH CoT)</td><td>32.61</td><td>65.82</td></tr><tr><td>6</td><td>ERNIE-Bot4 (Normal CoT)</td><td>35.14</td><td>64.30</td></tr></table>
+
+Table 2: Comparison of prompting and training strategies. Models are evaluated on both traditional and refined metrics.
+
+# 5.3.1 Ablation Study I: Training and Prompting Strategies
+
+To examine the overall effectiveness of our TEACH framework, we evaluate the individual and joint contributions of TEACH-style CoT prompting and fine-tuning. Table 2 presents results on 500 entries from validation set using both traditional metrics (BLEU, ROUGE, BERTScore, METEOR) and our refined metrics  $(\mathrm{BLEU}_{\mathrm{WAS}}, \mathrm{ROUGE}_{\mathrm{WAS}})$ .
+
+The results in Table 2 reveal three key findings. First, the untrained Qwen2-7B model with TEACH CoT (ID 2) yields the lowest scores, confirming the necessity of TEACH framework for effective semantic and stylistic generation. Second, among the trained models, TEACH CoT (ID 1) outperforms Normal CoT (ID 3), indicating that our step-by-step prompting provides more informative guidance. Third, on the teacher model (ERNIE-Bot4), TEACH prompts (ID 5) demonstrated superior performance in improved metrics but slightly lagged in traditional ones. This can be attributed to our metrics' enhanced ability to process additional information, aligning more closely with human judg
+
+<table><tr><td>Method</td><td>B</td><td>R</td><td>BWAS</td><td>RWAS</td><td>Avg</td></tr><tr><td>TEACH</td><td>21.88</td><td>54.41</td><td>45.93</td><td>75.59</td><td>49.45</td></tr><tr><td>w/o H</td><td>18.86</td><td>51.19</td><td>43.46</td><td>74.26</td><td>46.94</td></tr><tr><td>w/o A</td><td>20.37</td><td>52.99</td><td>45.19</td><td>75.39</td><td>48.49</td></tr><tr><td>w/o H&amp;A</td><td>18.88</td><td>51.19</td><td>43.18</td><td>74.29</td><td>46.89</td></tr><tr><td>w/o Lcon</td><td>20.59</td><td>53.16</td><td>47.13</td><td>76.51</td><td>49.35</td></tr><tr><td>Normal</td><td>17.06</td><td>48.35</td><td>43.13</td><td>74.31</td><td>45.71</td></tr></table>
+
+Table 3: Component-wise ablation study on Qwen2-7B-Base.
+
+ment.
+
+# 5.3.2 Ablation Study II: Component Analysis
+
+To further analyze the internal structure of the TEACH framework, we ablate specific components of the prompt and loss design: historical context  $(\mathcal{H})$ , word-level annotations  $(\mathcal{A})$ , both  $\mathcal{H} \& \mathcal{A}$ , and the style-preserving contrastive loss  $(\mathcal{L}_{con})$ . Table 3 reports results using BLEU, ROUGE,  $\mathrm{BLEU}_{\mathrm{WAS}}$ , ROUGE $_{\mathrm{WAS}}$ , and their average.
+
+The results, detailed in Table 3, offer profound insights. The absence of historical contexts  $\mathcal{H}$  significantly impairs performance, underscoring their essential role in comprehending classical Chinese. The removing of word annotations  $\mathcal{A}$  decreases performance, yet closely matching results of w/o  $\mathcal{H} \& \mathcal{A}$ . This indicates that combining  $\mathcal{H}$  and  $\mathcal{A}$  synergistically amplifies their effects, yielding a greater overall impact. Interestingly, removing  $\mathcal{L}_{con}$  results in a noticeable decline in traditional metrics; however, it achieves peak performance on refined metrics. This suggests that without the contrastive constraints imposed on translation style, while LLMs can translate sentences with higher semantic accuracy, there is a significant discrepancy between their formatting and the literal translation style required for classical texts. This highlights the effectiveness of  $\mathcal{L}_{con}$  in helping models learn translation styles specific to classical Chinese texts.
+
+# 5.3.3 Metrics Evaluation
+
+To validate our proposed evaluation metric, we present a comparative example in Figure 3. Based on traditional metrics, Erya clearly outperforms the LLM. However, from a more intuitive perspective, the LLM's translations are more detailed and accurate, whereas Erya occasionally fails to translate certain parts. This discrepancy arises because, while the LLM's output is semantically aligned, it does not match the reference at the character level.
+
+Our refined metric takes semantic similarity into account and also addresses the LLM's tendencies
+
+![](images/d0c68d7aa40a192e59bfc0555f547bfb2901a3bdc2de8647fa735506d997f5d1.jpg)  
+Figure 3: Comparative evaluation of metric quality. Underlines in the same color indicate character-level alignment, and words in the same color box indicate semantic alignment. Greyed-out text represents elements hidden during computing  $\mathrm{BLEU}_{\mathrm{WAS}}$  and ROUGE $_{\mathrm{WAS}}$ .
+
+![](images/32559ffb6f4ac5cde41212f2e40cf389e14dcba42a012c842d41444c12e6b857.jpg)  
+Figure 4: Case study. Comparing outputs from traditional WSD and translation Models, and TEACH-trained LLM.
+
+to over-translate and add extra information. Under this new metric, the performance gap between the LLM and Erya is noticeably reduced. Still, over-translation of LLM, such as turning "神机" into "机智神明", leads to a slightly lower overall score compared to Erya. In contrast, when measured by ROUGE, the improved metric surpasses the traditional one, reflecting the LLM's higher recall-oriented accuracy.
+
+# 5.3.4 Case Study
+
+In Figure 4, we illustrate the inference results of traditional models compared to those trained with TEACH-trained Qwen-7B-Base. Traditional models process two tasks separately and misinterpret word meanings due to a lack of historical context. In contrast, for TEACH-trained LLM, users simply input a classical text sentence, and the model automatically provides an analysis of the relevant historical background, selections for word meanings,
+
+<table><tr><td rowspan="2">Metrics</td><td colspan="2">ERNIE-Bot4</td><td colspan="2">Qwen2-TEACH</td></tr><tr><td>w/o</td><td>w/</td><td>1.5B</td><td>7B</td></tr><tr><td>Hist. Comp.</td><td>2.97</td><td>4.40</td><td>3.67</td><td>3.96</td></tr><tr><td>Hist. Acc.</td><td>2.89</td><td>4.24</td><td>3.29</td><td>3.82</td></tr><tr><td>Word Anal. Acc.</td><td>3.61</td><td>4.85</td><td>4.06</td><td>4.37</td></tr><tr><td>Trans. Flu.</td><td>3.51</td><td>4.21</td><td>3.88</td><td>4.18</td></tr><tr><td>Trans. Acc.</td><td>3.44</td><td>4.41</td><td>3.82</td><td>4.15</td></tr><tr><td>Trans. Sty. Cons.</td><td>3.00</td><td>3.57</td><td>3.77</td><td>4.05</td></tr><tr><td>Avg</td><td>3.24</td><td>4.28</td><td>3.75</td><td>4.09</td></tr></table>
+
+Table 4: Manual quality assessment of LLMs' outputs.
+
+and the final sentence translation. This approach offers a multi-dimensional perspective compared to traditional translation methods. It enhances the user's understanding of the details, making it ideal for educational settings where depth of knowledge is essential.
+
+# 5.3.5 Human Evaluation
+
+Due to the limitations of automatic metrics, we conduct a manual quality assessment in Table 4. We select 100 random entries and gather their real labels. These are compared against outputs from ERNIE-Bot4 using the CoT template that w/ and w/o  $\mathcal{H}$  &  $\mathcal{A}$ , as well as outputs from TEACH-trained Qwen-series LLMs. We assess these outputs on a 5-point scale, focusing on the comprehensiveness and accuracy of historical information, word analysis accuracy, and the fluency, accuracy, and stylistic consistency of translations (see Appendix C.3 for more details). The results reveal that ERNIE-Bot4, w/o  $\mathcal{H}$  and  $\mathcal{A}$ , generally scores extremely lower compared to w/, demonstrating serious historical hallucinations which our template effectively mitigates. Meanwhile, the Qwen2-7B model approaches the performance of ERNIE-Bot4 w/  $\mathcal{H}$  &  $\mathcal{A}$ , and even surpasses it in translation stylistic consistency. This underscores the effectiveness of our TEACH framework and external knowledge integration.
+
+# 6 Conclusion
+
+We introduce TEACH, a contrastive knowledge distillation approach with interpretability features tailored for training LLMs in classical Chinese educational scenarios. TEACH enhances the model's ability to process historical and semantic information, applying a step-by-step CoT prompt that progresses from paragraph to character-level analysis. Additionally, we propose refined metrics to evaluate translation by LLMs, addressing overtranslation and unnecessary additions. Experimental results confirm the effectiveness of TEACH and
+
+our metrics. Future work could expand LLM capabilities to include more tasks related to classical Chinese, improving their educational utility.
+
+# 7 Limitations
+
+Although the TEACH framework significantly enhances classical Chinese understanding, extending it to other languages involves certain adaptations. In practice, this process focuses on two primary modifications: first, constructing a tailored corpus and lexical knowledge base; and second, adjusting the CoT prompts to match the linguistic characteristics of the target language.
+
+For the first task, building a language-specific knowledge base and extracting related word meanings may require domain expertise, especially in cultures without well-established historical references. Fortunately, modern NLP techniques and large-scale web corpora can automate much of this data collection and annotation, thereby reducing the need for extensive manual efforts. As for the CoT prompts, while they may not transfer seamlessly due to differences in syntax and cultural context, only moderate prompt adjustments are generally needed to accommodate new grammatical structures and stylistic conventions. In sum, although some degree of careful customization is required, the availability of robust NLP resources and the inherent flexibility of prompt design greatly simplify the adaptation of TEACH to other languages.
+
+# 8 Ethical Considerations
+
+The corpora used for constructing our interpretable knowledge base are sourced exclusively from publicly available websites that offer free access for academic research purposes.
+
+# Acknowledgment
+
+This work is supported by the National Natural Science Foundations of China under Grant (62372060).
+
+# References
+
+Ernie Chang, Yow-Ting Shiue, Hui-Syuan Yeh, and Vera Demberg. 2021. Time-aware Ancient Chinese text translation and inference. In Proceedings of the 2nd International Workshop on Computational Approaches to Historical Language Change 2021, pages 1-6. Association for Computational Linguistics.
+
+Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. 2024. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45.  
+Boxing Chen and Colin Cherry. 2014. A systematic comparison of smoothing techniques for sentence-level bleu. In Proceedings of the ninth workshop on statistical machine translation, pages 362-367.  
+Zewen Chi, Li Dong, Bo Zheng, Shaohan Huang, XianLing Mao, He-Yan Huang, and Furu Wei. 2021. Improving pretrained cross-lingual language models via self-labeled word alignment. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3418-3430.  
+Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. 2023. Parameter-efficient fine-tuning of large-scale pretrained language models. Nature Machine Intelligence, 5(3):220-235.  
+Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335.  
+Gabriel Forgues, Joelle Pineau, Jean-Marie Larchevêque, and Réal Tremblay. 2014. Bootstrapping dialog systems with word embeddings. In Nips, modern machine learning and natural language processing workshop, volume 2, page 168.  
+Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997.  
+Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, et al. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793.  
+Geyang Guo, Jiarong Yang, Fengyuan Lu, Jiaxin Qin, Tianyi Tang, and Wayne Xin Zhao. 2023. Towards effective ancient chinese translation: Dataset, model, and evaluation. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 416-427.  
+Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations.
+
+Seungjun Lee, Jungseob Lee, Hyeonseok Moon, Chanjun Park, Jaehyung Seo, Sugyeong Eo, Seonmin Koo, and Heuseok Lim. 2023. A survey on evaluation metrics for machine translation. Mathematics, 11(4):1006.  
+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.  
+Yuqing Li, Yuxin Zhang, Bin Wu, Ji-Rong Wen, Ruihua Song, and Ting Bai. 2022. A multi-modal knowledge graph for classical Chinese poetry. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2318-2326. Association for Computational Linguistics.  
+Zhenyu Li, Sunqi Fan, Yu Gu, Xiuxing Li, Zhichao Duan, Bowen Dong, Ning Liu, and Jianyong Wang. 2024. Flexkbq: A flexible llm-powered framework for few-shot knowledge base question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18608-18616.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Maofu Liu, Junyi Xiang, Xu Xia, and Huijun Hu. 2022. Contrastive learning between classical and modern chinese for classical chinese machine reading comprehension. ACM Trans. Asian Low-Resour. Lang. Inf. Process., 22(2).  
+Yutong Liu, Tao Xie, Bin Wu, and Bai Wang. 2019. New word detection in ancient chinese corpus (in Chinese). In Journal of Chinese Information Processing, pages 46-55.  
+Daniel Loureiro and Alipio Jorge. 2019. Language modelling makes sense: Propagating representations through WordNet for full-coverage word sense disambiguation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5682-5691. Association for Computational Linguistics.  
+Xiaomeng Pan, Hongfei Wang, Teruaki Oka, and Mamoru Komachi. 2022. Zuo zhuan Ancient Chinese dataset for word sense disambiguation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 129-135. Association for Computational Linguistics.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318.
+
+Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph retrieval-augmented generation: A survey.  
+Gabriel Peyré, Marco Cuturi, et al. 2019. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607.  
+Vasile Rus and Mihai Lintean. 2012. A comparison of greedy and optimal assessment of natural language student input using word-to-word similarity metrics. In Intelligent Tutoring Systems: 11th International Conference, ITS 2012, Chania, Crete, Greece, June 14-18, 2012. Proceedings 11, pages 675-676. Springer.  
+Masoud Jalili Sabet, Philipp Dufter, François Yvon, and Hinrich Schütze. 2020. Simalign: High quality word alignments without parallel training data using static and contextualized embeddings. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1627-1643.  
+Jiho Shin, Hadi Hemmati, Moshi Wei, and Song Wang. 2024. Assessing evaluation metrics for neural test oracle generation. IEEE Transactions on Software Engineering, pages 1-13.  
+Lei Shu, Yiluan Guo, Huiping Wang, Xuetao Zhang, and Renfen Hu. 2021. The construction and application of Ancient Chinese corpus with word sense annotation. In Proceedings of the 20th Chinese National Conference on Computational Linguistics, pages 549-563. (in Chinese).  
+Huishuang Tian, Kexin Yang, Dayiheng Liu, and Jiancheng Lv. 2021. Archibert: a pre-trained model for ancient chinese language understanding and generation. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8.  
+Dongbo Wang, Litao Lin, Zhixiao Zhao, Wenhao Ye, Kai Meng, Wenlong Sun, Lianzhen Zhao, Xue Zhao, Si Shen, Wei Zhang, and Bin Li. 2023. EvaHan2023: Overview of the first international Ancient Chinese translation bakeoff. In Proceedings of ALT2023: Ancient Language Translation Workshop, pages 1-14, Macau SAR, China. Asia-Pacific Association for Machine Translation.  
+Dongbo Wang, Chang Liu, Zihe Zhu, Jiangfeng Liu, Haotian Hu, Si Shen, and Bin Li. 2022. Construction and application of pre-training model of "Si ku Quan shu" oriented to digital humanities (in Chinese). Library Tribune, 42(6):14.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. 2022a. Emergent abilities of large language models. Transactions on Machine Learning Research.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou,
+
+et al. 2022b. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Yuting Wei, Yuanxing Xu, Xinru Wei, Simin Yang, Yangfu Zhu, Yuqing Li, Di Liu, and Bin Wu. 2024a. AC-EVAL: Evaluating Ancient Chinese language understanding in large language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2024, pages 1600–1617. Association for Computational Linguistics.  
+Yuting Wei, Yangfu Zhu, Ting Bai, and Bin Wu. 2024b. A cross-temporal contrastive disentangled model for ancient chinese understanding. *Neural Networks*, 179:106559.  
+Junyi Xiang, Maofu Liu, Qiyuan Li, Chen Qiu, and Huijun Hu. 2024. A cross-guidance cross-lingual model on generated parallel corpus for classical Chinese machine reading comprehension. Information Processing & Management, 61(2):103607.  
+An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024a. Qwen2 technical report. arXiv preprint arXiv:2407.10671.  
+Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Shaochen Zhong, Bing Yin, and Xia Hu. 2024b. Harnessing the power of llms in practice: A survey on chatgpt and beyond. ACM Trans. Knowl. Discov. Data, 18(6).  
+Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. 2024. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652.  
+Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. 2023. Glm-130b: An open bilingual pre-trained model. In The Eleventh International Conference on Learning Representations.  
+Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2024. Teaching large language models to translate with comparison. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19488-19496.  
+Shitou Zhang, Ping Wang, Zuchao Li, Jingrui Hou, and Qibiao Hu. 2024. Confidence-based syntax encoding network for better ancient chinese understanding. Information Processing & Management, 61(3):103616.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations.  
+Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, and Bin Cui. 2024. Retrieval-augmented generation for ai-generated content: A survey. CoRR.
+
+Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223.
+
+Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2024. Collaborative large language model for recommender systems. In Proceedings of the ACM on Web Conference 2024, pages 3162-3172.
+
+# A Knowledge-Augmented Template
+
+The illustration of our knowledge-augmented template is shown in Figure 5, where the user only needs to input a query sentence to automatically retrieve relevant historical context and annotations, with the LLM providing a step-by-step reasoning analysis. Below, we detail the process of retrieving the context and annotations.
+
+# A.1 Historical Database Construction
+
+Our corpus of historical knowledge contains the complete classical text version of the Twenty-Four Histories, supplemented by novels, poetry, and other literary genres. We pre-process this corpus by cleaning, segmenting texts into sentences, and filtering out short and duplicate sentences. Utilizing the pre-trained classical Chinese model AnchiBERT, we vectorize all sentences and record their IDs along with the names of the books from which they originate. These vectors are then inserted into the Milvus vector database, which contains 4.32 million entries and boasts an average retrieval speed of 0.35 seconds.
+
+For a query sentence, we vectorize it and apply a re-ranking method to retrieve the most relevant vector ID, along with its associated confidence score. We then extract the surrounding context—defined as the ten sentences before and after the matched ID—to obtain contextual information, with the corresponding book name cited as the source.
+
+To evaluate the retrieval accuracy of queries in both Simplified and Traditional Chinese, we randomly selected 1,000 sentences for assessment. Our findings indicate that the database can accurately retrieve relevant contexts for both text forms. As demonstrated in Table 5, the accuracy rate for Traditional Chinese queries is slightly higher than that for Simplified Chinese. This discrepancy is likely due to a larger proportion of Traditional Chi
+
+<table><tr><td>Metrics\Query</td><td>Trad.</td><td>Simp.</td></tr><tr><td>Accuracy</td><td>95.70%</td><td>95.20%</td></tr><tr><td>R to Trad.</td><td>58.60%</td><td>39.70%</td></tr><tr><td>R to Simp.</td><td>41.40%</td><td>60.30%</td></tr></table>
+
+Table 5: Retrieval (R) Accuracy and Distribution for Traditional (Trad.) and Simplified (Simp.) Chinese Queries.
+
+nese texts within our corpus. Moreover, the retrieval results encompass both Simplified and Traditional texts, confirming that our vector database is effectively adapted to both. Notably, even when faced with misspelled sentences, our database still manages to identify relevant contexts, albeit the relationships may be weaker.
+
+# A.2 Word Annotation Collection
+
+We use the word segmentation algorithm in Liu et al. (2019) for classical Chinese to segment all classical texts, filter out low-frequency words, and crawl annotations from the Dictionary of Classical Chinese (ZDIC). If a classical Chinese word  $w$  has  $K$  annotated definitions, its corresponding annotation set is defined as  $S_{w} = \{s_{k} \mid k = 1, \dots, K\}$ , where each  $s_{k}$  denotes the annotation text associated with the  $k$ -th sense. Each annotation  $s_{k}$  can be further tokenized into a sequence of words  $\{s_{k,i} \mid i = 1, \dots, N_{k}\}$ , where  $N_{k}$  denotes the number of tokens in  $s_{k}$ . The word  $w$  and its annotation set  $S_{w}$  are added to the candidate set  $C = \{(w : S_{w})\}$ . This process resulted in a total of 169,742 annotated classical words.
+
+For a given query sentence, we segment it and obtain all corresponding annotated words. In practical applications, annotations of certain common but politically related words can lead to erroneous outputs. Therefore, we excluded sensitive words (e.g., place names, kings, emperors, prime ministers, countries, generals, Qin Shi Huang, monarchs) to avoid such issues.
+
+# B Solved initial matrix for word alignment
+
+The word alignment challenge is formulated as identifying matrix  $A$  that maximizes sentence similarity:
+
+$$
+\max  _ {A} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {m} A _ {i j} f _ {\text {s i m}} \left(r _ {i}, c _ {j}\right), \tag {13}
+$$
+
+![](images/17aad5ee31b098970d1aee5e92c35cc244af2b2b411d64190bab6d9f963b0170.jpg)  
+Figure 5: The illustration of our knowledge-augmented template.
+
+where  $f_{\mathrm{sim}}(\cdot)$  denotes the similarity function. To solve this equation, we reformulate it as an optimal transport problem, referencing Peyre et al. (2019) and Chi et al. (2021).
+
+Specifically, an entropic regularization is added to  $A$ :
+
+$$
+\max  _ {A} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {m} A _ {i j} f _ {\operatorname {s i m}} \left(r _ {i}, c _ {j}\right) - \mu A _ {i j} \log A _ {i j}. \tag {14}
+$$
+
+Then, Eq. (14) has a unique solution  $A^{\prime}$ :
+
+$$
+A ^ {\prime} = \operatorname {d i a g} (\boldsymbol {u}) K \operatorname {d i a g} (\boldsymbol {v}), \tag {15}
+$$
+
+$$
+K _ {i j} = e ^ {f _ {\operatorname {s i m}} \left(r _ {i}, c _ {j}\right) / \mu}, \tag {16}
+$$
+
+where  $\pmb{u} \in \mathbb{R}_{+}^{n}$ ,  $\pmb{v} \in \mathbb{R}_{+}^{m}$ ,  $K \in \mathbb{R}_{+}^{n \times m}$ . According to Sinkhorn's algorithm (Peyre et al., 2019), the variables  $\pmb{u}$  and  $\pmb{v}$  can be calculated by the following iterations:
+
+$$
+\boldsymbol {u} ^ {t + 1} = \frac {\mathbf {1} _ {n}}{K \boldsymbol {v} ^ {t}}, \quad \boldsymbol {v} ^ {t + 1} = \frac {\mathbf {1} _ {m}}{K ^ {\top} \boldsymbol {u} ^ {t + 1}}, \tag {17}
+$$
+
+where  $\pmb{v}^t$  can be initialized by  $\pmb{v}^{t = 0} = \mathbf{1}_m$ .
+
+# C Analysis
+
+# C.1 Effect of Training Corpus Size
+
+We evaluate the performance of the Qwen2-series models across various sizes of high-quality training data. Figure 6 illustrates that as the corpus size increases, the model's performance progressively improves but eventually stabilizes. Notably, even with as few as 1000 training entries, the model shows significant improvement. Considering the costs, we
+
+![](images/dbde3ebefa4e413485347cb82b7b7e7423993a1096cd82c2eafbd0c06b472306.jpg)  
+Figure 6: Compare the performance of finetuning student models with different corpus sizes. The average (avg) score is computed over four metrics: BLEU, ROUGE,  $\mathrm{BLEU}_{\mathrm{WAS}}$ , and ROUGE $_{\mathrm{WAS}}$ .
+
+![](images/49a5c1a786108dea316a4f9a3bfa5da64f9548ba3d067dfe1299ffa9677a5555.jpg)
+
+randomly select 3,000 high-quality entries to finetune the LLMs. The experiments reveal that with these entries, the model outperforms the traditional Erya model, which was trained on 300,000 entries. Additionally, it provides comprehensive historical context and semantic analysis. This highlights the effectiveness of our interpretable knowledge and the TEACH framework.
+
+# C.2 Efficiency and Applicability Compare
+
+In Table 6, we examine the efficiency and applicability of TEACH-trained Qwen-series models in comparison to two baselines: the closed-source ERNIE-Bot4 and Erya. We focus on model size, deployment complexity, cost, and functionality.
+
+Firstly, ERNIE-Bot4, with its larger parameter size, requires substantial resources for deployment or high API costs. In contrast, TEACH trains smaller models effectively on just one Nvidia A100 GPU, with capabilities for inference on
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Erya</td><td rowspan="2">ERNIE-Bot4</td><td colspan="2">Qwen2-TEACH</td></tr><tr><td>1.5B</td><td>7B</td></tr><tr><td>Model Size (B)</td><td>0.15</td><td>Closed</td><td>1.5</td><td>7</td></tr><tr><td>Training GPU</td><td>13090</td><td>Hard</td><td>1A100</td><td>1A100</td></tr><tr><td>Deployment GPU</td><td>13090</td><td>Hard</td><td>13090</td><td>13090</td></tr><tr><td>Retrieval Time (s)</td><td>X</td><td>0.35</td><td>0.35</td><td>0.35</td></tr><tr><td>Reasoning Time (s)</td><td>0.15</td><td>22.51</td><td>0.87</td><td>2.59</td></tr><tr><td>API Costs/Input</td><td>X</td><td>0.04</td><td>X</td><td>X</td></tr><tr><td>API Costs/Output</td><td>X</td><td>0.12</td><td>X</td><td>X</td></tr><tr><td>Historical Info</td><td>X</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>Word Analysis</td><td>X</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+Table 6: The comparison of efficiency and applicability. API Costs are denoted in ¥/1K tokens.  
+
+<table><tr><td>Metric</td><td>Description</td></tr><tr><td>Hist. Comp.</td><td>Assesses the exhaustiveness of the historical content, ensuring that all pertinent facts are thoroughly in- 
+cluded.</td></tr><tr><td>Hist. Acc.</td><td>Evaluates the correctness of the his-torical content, verifying that all in- 
+cluded facts are accurately depicted.</td></tr><tr><td>Word Anal. Acc.</td><td>Examines the precision in analyzing and interpreting individual words, 
+focusing on their meanings and con-
+texts within the text.</td></tr><tr><td>Trans. Flu.</td><td>Assesses the smoothness and natu-
+ral flow of translations, checking for 
+natural readability and the absence 
+of awkward phrasing.</td></tr><tr><td>Trans. Acc.</td><td>Focuses on the correctness of trans-
+lations, particularly how faithfully 
+the text conveys the semantic intent 
+of the original language.</td></tr><tr><td>Trans. Sty.
+Cons.</td><td>Evaluates how consistently the 
+translation maintains the literal 
+style of the answers.</td></tr></table>
+
+Table 7: Definition of Evaluation Metrics Used to Assess Quality of Outputs
+
+a single Nvidia GeForce RTX 3090 GPU. Secondly, ERNIE-Bot4 experiences prolonged inference times due to API dependency, while smaller models like the 7B deliver results in less than one-tenth of that time. Lastly, although TEACH may extend reasoning times compared to Erya, it substantially enriches outputs with detailed historical context and semantic analysis. This enhancement not only boosts user engagement but also strengthens its educational utility.
+
+# C.3 Details of Human Evaluation
+
+Evaluation Metrics Defined: We introduce six evaluation metrics used to assess the quality of outputs, detailed in Table 7. These metrics collectively help in providing a thorough evaluation of the linguistic and historical accuracy of our model outputs.
+
+Expert Panel Composition and Reliability: To
+
+ensure the reliability and reduce individual biases in our assessment, we invite five experts in classical Chinese language, all holding either Master's or Ph.D. degrees. The experts, ranging in age from 25 to 60, include three males and two females, representing diverse academic institutions and possessing extensive experience in historical linguistics. We average the ratings from all experts, and the Cronbach's Alpha of these annotations is 0.89, indicating a high level of inter-annotator agreement.

2025/TEACH_ A Contrastive Knowledge Adaptive Distillation Framework for Classical Chinese Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:706231dd9a7028d737b0e3d82453cd0357869bd56b843281816031a40b3edfa1
+size 751980

2025/TESS 2_ A Large-Scale Generalist Diffusion Language Model/cc3a7cf6-9b0a-4a89-a779-2019421a1045_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63fcef0296f89aa4c6c2d13dfcc9e125c808fd9c9ff32b99c0b5701af876ef97
+size 670161

2025/TESS 2_ A Large-Scale Generalist Diffusion Language Model/full.md

@@ -0,0 +1,503 @@
+# TESS 2: A Large-Scale Generalist Diffusion Language Model
+
+Jaesung Tae $^{1*}$ , Hamish Ivison $^{2,3*}$ , Sachin Kumar $^{3,4}$ , Arman Cohan $^{1}$
+
+$^{1}$ Yale University,  $^{2}$ University of Washington,
+
+3Allen Institute for AI, 4The Ohio State University
+
+# Abstract
+
+We introduce TESS 2, a general instruction-following diffusion language model that outperforms contemporary instruction-tuned diffusion models, as well as matches and sometimes exceeds strong autoregressive (AR) models. We train TESS 2 by first adapting an AR model via continued pretraining with the usual cross-entropy as diffusion loss, and then performing further instruction tuning. We find that adaptation training as well as the choice of the base model is crucial for training good instruction-following diffusion models. Furthermore, we propose reward guidance, a novel and modular inference-time guidance procedure to align model outputs without needing to train the underlying model. Finally, we show that TESS 2 further improves with increased inference-time compute, highlighting the utility of diffusion LMs in having fine-grained controllability over the amount of compute used at inference time. Code and models are available at https://github.com/hamishivi/tess-2.
+
+# 1 Introduction
+
+Existing language models are predominantly trained using an autoregressive (AR) paradigm. Prior works have suggested that AR models may have a number of limitations, especially in planning and self-correction (Lin et al., 2021; Bachmann and Nagarajan, 2024; Huang et al., 2024). While recent work has made large strides in improving AR model reasoning, this hinges on increasing test-time compute using specialized reinforcement learning-based training procedures (Kumar et al., 2024; Laskin et al., 2022) and costly long generations (DeepSeek-AI, 2025; Min et al., 2024).
+
+An alternative paradigm lies in diffusion language models, which have shown promise in addressing some of the limitations of AR models (Ye et al., 2024; Zhang et al., 2023). Diffusion models
+
+![](images/8a7adffa3d4bc6812aa5cc1aa88cc94c25a01356d18f265f784a6e48fc6b18c1.jpg)  
+Figure 1: Overview of inference with TESS 2. We provide a chat-template formatted query and iteratively denoise over the simplex space for a number of diffusion steps (typically 100). Optionally, we also incorporate reward guidance at each diffusion step by applying the gradient of the predicted reward to the intermediate logits, i.e., gradient ascent on reward.
+
+provide plug-and-play controllability at inference time (Li et al., 2022), are naturally capable of flexible generation such as infilling, and enable precise control of inference compute, more so than recent long chain-of-thought approaches. However, diffusion LMs have remained relatively small scale and focused on improving intrinsic metrics such as perplexity (Lou et al., 2024; Gulrajani and Hashimoto, 2023), rather than examining common downstream tasks used to evaluate AR LMs. This constrains their broader impact and practical applicability.
+
+In this work, we aim to close this gap by introducing TESS 2, a large-scale diffusion language model. We show that strong diffusion models can be trained by adapting from existing AR models, and that these models can be further instruction-tuned to serve as strong generalist models across varied downstream settings. We ablate a number of configurations and present a recipe for adapting existing AR LMs to diffusion models. Our final model outperforms contemporary diffusion LMs and performs close to or exceeds AR baselines, showing that continuous diffusion models can serve as generalist instruction-tuned LMs.
+
+To the best of our knowledge, we are the first work to provide a full recipe for training generalist instruction-following fully non-autoregressive continuous diffusion models, with prior work only examining performance after task-specific finetuning (Karimi Mahabadi et al., 2024; Han et al., 2024a), using semi-autoregressive setups and only testing chat settings (Han et al., 2023), or using a discrete diffusion approach instead of a continuous one (Gong et al., 2024). We also underscore the unique advantages of diffusion LMs over AR models: the scalability of inference-time compute with increased diffusion steps and the ability to supplement the backward diffusion process with reward models. We summarize our contributions.
+
+1. We propose a recipe for adapting LLMs into instruction-following diffusion LMs that significantly outperform prior text diffusion models and perform similarly to or better than AR models on QA and general instruction-following tasks.  
+2. We ablate components of the recipe and highlight salient elements, such as the choice of the base model, where possible.  
+3. We show that our resulting model can further improve with increased test-time compute by employing more diffusion steps.  
+4. We introduce reward guidance, a novel technique for steering diffusion LMs to generate text aligned with user preferences, enhancing downstream chat performance without any additional training.
+
+# 2 Related Work
+
+Diffusion Language Models With the popularization of diffusion models in domains such as image (Nichol and Dhariwal, 2021), audio (Kong et al., 2020; Tae et al., 2022; Shen et al., 2023), video (Ho et al., 2022), and text-to-image generation (Saharia et al., 2022; Ramesh et al., 2022), there have been growing attempts to adapt and improve diffusion methods for language generation. One line of diffusion LMs proposes a diffusion process that adds Gaussian noise to word embeddings (Li et al., 2022; Gulrajani and Hashimoto, 2023). Another approach uses discrete diffusion based on categorical jump probabilities and continuous-time Markov chains (Lou et al., 2024; Richemond et al., 2022; Gong et al., 2024; Nie et al., 2025), removing the need for mapping diffused embeddings to discrete inputs or outputs.
+
+In contrast, simplex diffusion language models maintain the continuity of the diffusion process, instead aiming to learn discrete data by modelling over a continuous simplex. We build off TESS (Karimi Mahabadi et al., 2024), which diffuses over the probability simplex and itself is a fully-non-autoregressive version of SSD-LM (Han et al., 2022). While Han et al. (2024b) explored scaling SSD-LM for instruction-following, it remained semi-autoregressive and focused on a novel collaborative decoding strategy without releasing code, models, or data. In contrast, we show that fully non-autoregressive generation is possible at context lengths of up to 2048 tokens while only using publicly available data and model checkpoints. We also take a different approach from LLaDA (Nie et al., 2025), a contemporary work that focuses on pretraining from scratch at scale, and instead explore how we can leverage already pretrained AR models to reformulate them as diffusion LMs.
+
+Adapting Diffusion Models Concurrent and prior work has also examined instruction-tuning diffusion models and adapting them from existing AR models. Han et al. (2024a) builds off SUN-DAE (Savinov et al., 2022) and shows that AR models can be successfully converted into diffusion models for further downstream finetuning. They focus on training diffusion checkpoints that can then be finetuned to perform specific tasks, as opposed to examining instruction tuning and general performance as done in this work.
+
+Ye et al. (2023) examines scaling reparameterized discrete diffusion models (Zheng et al., 2024) by adapting XLM-R checkpoints (Conneau et al., 2020), but report that they are unable to adapt more modern LMs and find their models fall short on more complex reasoning tasks.
+
+Concurrent to our work, Gong et al. (2024) proposes DiffuLlama, an absorbing discrete diffusion model (Austin et al., 2021) adapted from Llama (Touvron et al., 2023). However, they focus on discrete diffusion and pretraining without exploring instruction tuning. In comparison, we show that simpler continuous diffusion algorithms from SSD-LM and TESS can be straightforwardly translated to adapt diffusion LMs. We additionally explore instruction tuning adapted diffusion LMs and evaluate the model on common instruction tuning benchmarks, which better reflect real-life applications that require the ability to provide long-form generations that engage with user queries.
+
+# 3 TESS 2
+
+# 3.1 Simplex Diffusion Language Models
+
+For the model architecture of TESS 2, we largely follow TESS (Karimi Mahabadi et al., 2024), a fully non-autoregressive diffusion model with self-conditioning. We provide a brief recap.
+
+Simplex-based Representation Let  $\mathcal{V}$  denote the vocabulary space. We map the index of each token to be generated  $w\in \mathcal{V}$  to a  $k$  -logit simplex to produce  $\mathbf{s}^w\in \{\pm k\}^{|\mathcal{V}|}$ , whose  $i$  -th component satisfies
+
+$$
+s _ {(i)} ^ {w} = \left\{ \begin{array}{l l} k, & \text {i f} \quad i = w, \\ - k, & \text {o t h e r w i s e}, \end{array} \right. \tag {1}
+$$
+
+with a hyperparameter  $k\in \mathbb{R}^{+}$ . We then produce a probability simplex over  $\nu$  via  $\mathbf{p}^w = \mathrm{softmax}(\mathbf{s}^w)$ . Finally, we compute the weighted sum of word embeddings to obtain a continuous embedding vector,  $\mathbf{h}^w = \mathbf{E}\mathbf{p}^w$ , where  $\mathbf{E}\in \mathbb{R}^{d\times |\mathcal{V}|}$  is the word embedding matrix,  $d$  denotes the size of the hidden dimension, and  $\mathbf{h}^w\in \mathbb{R}^d$ .
+
+Timestep Embeddings We use a single linear layer without bias to produce timestep embeddings. These embeddings are directly added to the token embeddings before being fed into the first transformer block to inform the model of the current timestep in the diffusion process.
+
+Forward Diffusion Let  $\mathbf{w} = (w_{1},\dots ,w_{L})$  be a sentence of  $L$  tokens such that  $w_{i}\in \mathcal{V}$ , and  $\mathbf{S}_0 = (\mathbf{s}^{w_1},\ldots ,\mathbf{s}^{w_L})\in \{\pm k\}^{L\times |\mathcal{V}|}$  be the  $k$ -logit simplex representation of  $\mathbf{w}$ . Following the standard Denoising Diffusion Probabilistic Models (DDPM) formulation (Ho et al., 2020), we add noise to the  $k$ -logit simplex representation during training according to
+
+$$
+\mathbf {S} _ {t} = \sqrt {\bar {\alpha} _ {t}} \mathbf {S} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon} _ {t}, \tag {2}
+$$
+
+where  $t\in \{0,1,\dots ,T\}$  is the timestep,  $T\in \mathbb{N}$  is the total number of diffusion steps,  $\epsilon_t\sim \mathcal{N}(0,k^2\mathbf{I})$  and  $\bar{\alpha}_t$  follows the cosine noise schedule (Nichol and Dhariwal, 2021).
+
+Training We train the model by computing the usual cross-entropy loss between the ground-truth tokens  $\mathbf{w}$  and the model prediction given a noisy logit simplex  $\mathbf{S}_t$  at timestep  $t$ .
+
+$$
+\mathcal {L} = \mathbb {E} _ {t, q (\mathbf {S} _ {0}), q (\mathbf {S} _ {t} | \mathbf {S} _ {0})} \left[ - \sum_ {i = 1} ^ {L} \log p _ {\boldsymbol {\theta}} \left(w _ {i} | \mathbf {S} _ {t}, t\right) \right]. \tag {3}
+$$
+
+This is in contrast to the usual mean squared error loss used in standard diffusion models, and is found to be stable in training simplex-based diffusion language models (Han et al., 2022; Karimi Mahabadi et al., 2024).
+
+Sampling During inference, we sample  $\mathbf{S}_T$  from the prior  $\mathcal{N}(0,k^2\mathbf{I})$  and run the reverse process for  $t = T,\ldots ,1$  on the noisy  $k$ -logit simplex. The reverse process can be approximated via
+
+$$
+\mathbf {S} _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \hat {\mathbf {S}} _ {\boldsymbol {\theta}} \left(\mathbf {S} _ {t}, t\right) + \sqrt {1 - \bar {\alpha} _ {t - 1}} \boldsymbol {\epsilon} _ {t}. \tag {4}
+$$
+
+Thus, to reverse one step from  $t$ , we take the model prediction  $\hat{\mathbf{S}}_{\theta}$  as the hypothetical ground truth, then corrupt it by  $(t - 1)$  timesteps. To construct the model prediction, we project the logits predicted by the underlying encoder model via argmax as a pseudo-inverse of Eq. (1) to match the initial  $k$ -logit representation:
+
+$$
+\hat {s} _ {(i)} ^ {w} = \left\{ \begin{array}{l l} k, & \text {i f} \quad i = \operatorname {a r g m a x} \left(\mathbf {s} ^ {w}\right), \\ - k, & \text {o t h e r w i s e .} \end{array} \right. \tag {5}
+$$
+
+# 3.2 Adapting AR models to Diffusion
+
+Our approach to adapting AR models to diffusion consists of three key elements: UL2 masking, label shifting, and full bidirectional attention.
+
+UL2 Masking Inspired by Tay et al. (2023), we train our model with a mixture of span infilling and prefix completion training objectives. For span infilling, we mask out random spans (with lengths randomly sampled from a pre-specified range) within the text and train the model to predict them. For prefix completion, we mask the last  $n$  tokens, where  $n$  is randomly chosen, and ask the model to predict the rest. The latter objective aligns with how we use the model for downstream tasks, but we found training with both objectives to work best for overall performance in pilot experiments. We provide further details in App. A.
+
+Label Shifting Given some noised input, a typical diffusion LM predicts the less noised version of the same token at each position in the sequence. That is, given a noised simplex  $\mathbf{S}_t = (\mathbf{s}_t^{w_1},\dots,\mathbf{s}_t^{w_L})$ , the model predicts  $\mathbf{S}_{t - 1} = (\mathbf{s}_{t - 1}^{w_1},\dots,\mathbf{s}_{t - 1}^{w_L})$ . In our approach, we instead train the model to predict the token at the next index position to align with the next token prediction objective used for AR models, i.e., the model predicts  $(\mathbf{s}_{t - 1}^{w_2},\dots,\mathbf{s}_{t - 1}^{w_L + 1})$ . This does not affect the validity of the denoising diffusion setup, but we find that it
+
+helps the model converge faster. Concurrent to our work, DiffuLlama (Gong et al., 2024) also reported positive results with label shifting.
+
+Full Bidirectional Attention Finally, we disable causal masking and train our model with full bidirectional attention masks, as we found this led to better performance in pilot experiments. Fortunately, existing optimized model components such as Flash Attention (Dao et al., 2022; Dao, 2024) are fully compatible with this change. This enables fast and efficient diffusion adaptation training. Concurrently, DiffuLlama (Gong et al., 2024) also reported using full bidirectional attention.
+
+# 3.3 Instruction Tuning
+
+Once we have a diffusion-adapted model, we further put it through a stage of instruction tuning, in which we train the model on a smaller set of carefully curated instruction data. For this stage, we apply the same label shifting and full bidirectional attention mask as before, but switch from UL2 to only using a prefix-LM (Liu et al., 2018) objective for training. Specifically, we format our dataset to contain only single-turn instances, mask out the assistant response, and train the model to predict it. We also explored multi-turn training but observed marginal gains. We leave improved multi-turn training for future work.
+
+# 3.4 Reward-based Classifier Guidance
+
+Finally, we explore reward guidance, a novel method to improve model outputs without further training. We extend the guidance method proposed in Han et al. (2022) to be used with a scalar reward model. Intuitively, reward guidance works by adjusting the in-progress generation at each diffusion step using gradients from an off-the-shelf reward model, which is used to estimate the potential reward of the generation. This ultimately allows us to guide the generation process towards sequences with higher rewards.
+
+Specifically, at each diffusion step, we take the output from the model at a given step,  $\hat{\mathbf{S}}_{\theta}$ , and convert it to continuous token embeddings that we can then feed into an off-the-shelf reward model (with a matching tokenizer):
+
+$$
+\mathbf {p} _ {t} = \operatorname {s o f t m a x} (\hat {\mathbf {S}} _ {\boldsymbol {\theta}}) \tag {6}
+$$
+
+$$
+\mathbf {c} _ {w} = \mathbf {E} \mathbf {p} _ {t}, \tag {7}
+$$
+
+where  $\mathbf{E}$  is the embedding matrix of the classifier. We then pass  $\mathbf{c}_w$  as input into the reward model,
+
+adding any required positional information before feeding it into the first hidden layer. This results in a predicted reward  $R \in \mathbb{R}$  for the given timestep. Since we want to increase this, we perform gradient ascent on the output reward, computing the gradient of  $\hat{\mathbf{S}}_{\theta}$  with respect to the reward via backpropagation and then adding the resulting gradient to the model prediction by
+
+$$
+\hat {\mathbf {S}} _ {\boldsymbol {\theta}} := \hat {\mathbf {S}} _ {\boldsymbol {\theta}} + \boldsymbol {\eta} \cdot \nabla_ {\boldsymbol {\theta}} R, \tag {8}
+$$
+
+where  $\eta$  is the guidance coefficient that controls how much guidance to apply at each timestep. Note that this is an inference-time algorithm that is applied at each timestep during the backward process; we do not apply any guidance during training.
+
+# 4 Experiments
+
+# 4.1 Setup
+
+We evaluate our model in three settings: directly after diffusion adaptive pretraining, after instruction tuning, and with reward guidance applied.
+
+Diffusion Adaptation For diffusion adaptation training, we train the model on Dolma 1.7 (Soldaini et al., 2024), an open pretraining dataset, for up to 200,000 train steps, or approximately 45 billion tokens. We use a constant learning rate of  $1 \times 10^{-5}$  with a linear warmup of 5000 train steps, and a batch size of 112 samples. We adapt the Mistral-7B-v0.1 model (Jiang et al., 2023), based on model ablations explored in Sec. 4.2. We use a simplex value of 5 and self-conditioning as done in Karimi Mahabadi et al. (2024). We provide additional details in App. A.
+
+Instruction Tuning For instruction tuning, we use the Tulu 2 SFT mixture (Ivison et al., 2023), which consists of roughly 326k samples. We fine-tune for 3 epochs with a learning rate of  $1 \times 10^{-5}$  with linear warmup for the first  $3\%$  of training steps and linear cooldown for the rest of training, which we found worked best in pilot experiments.
+
+Reward Guidance For reward guidance, we train a standard reward model on human preferences using Mistral-7B-v0.1 and the data and hyperparameters provided by Dong et al. (2023); Xiong et al. (2023). We provide training details in App. C.
+
+Adaptation Evaluation Following Han et al. (2022, 2023); Richemond et al. (2022), we evaluate our adaptation using perplexity, average distinct  $n$ -grams in the output samples (d-1/2/3/4), unigram (per-token) entropy, and Mauve score (Pillutla et al., 2021). Perplexity is measured using a held-out model, $^{3}$  following Han et al. (2022, 2023).
+
+Downstream Evaluation In addition to the adaptation evaluation, we evaluate our models on a diverse set of downstream tasks: (1) AlpacaEval (Li et al., 2023), which tests general instruction following capabilities; (2) GSM8k (Cobbe et al., 2021), which tests simple math word problems; (3) SQuAD (Rajpurkar et al., 2016) (with context), which tests general question-answering abilities; (4) TriviaQA (Joshi et al., 2017), a reading comprehension dataset; (5) IFEval (Zhou et al., 2023), which tests exact instruction-following; and (6) Big Bench Hard (BBH) (Suzgun et al., 2022), which tests general reasoning. For GSM8k, we additionally test the performance of our model after finetuning on the GSM8k symbolic data provided by Ye et al. (2024), following DiffuLLaMA (Gong et al., 2024). We provide further details in App. D.
+
+# 4.2 Diffusion Adaptation
+
+<table><tr><td></td><td>Rand.</td><td>RoBERTa</td><td>Llama 2</td><td>Llama 3</td><td>Mistral</td></tr><tr><td>Perp.</td><td>54.4</td><td>20.2</td><td>3619.2</td><td>880.4</td><td>24.3</td></tr><tr><td>d-1</td><td>0.55</td><td>0.36</td><td>0.94</td><td>0.97</td><td>0.62</td></tr><tr><td>d-2</td><td>0.82</td><td>0.59</td><td>0.99</td><td>0.99</td><td>0.87</td></tr><tr><td>d-3</td><td>0.91</td><td>0.69</td><td>0.98</td><td>0.98</td><td>0.93</td></tr><tr><td>d-4</td><td>0.94</td><td>0.75</td><td>0.98</td><td>0.98</td><td>0.94</td></tr><tr><td>Mauve</td><td>0.92</td><td>0.93</td><td>0.01</td><td>0.93</td><td>0.95</td></tr><tr><td>Entropy</td><td>5.7</td><td>4.8</td><td>7.7</td><td>7.8</td><td>6.3</td></tr></table>
+
+Table 1: Performance of diffusion adaptation after 35,000 steps of pretraining over varied base models. In all cases except perplexity, higher is better. Rand inicidates initializing from scratch.
+
+While TESS used a RoBERTa model (Liu et al., 2019) as its backbone, recently there has been a wave of increasingly larger and more powerful open-weight pretrained models (Jiang et al., 2023; Touvron et al., 2023; Llama Team, 2024). However, unlike RoBERTa, these models are typically decoder-only causal LMs, which makes it unclear how well they transfer to the full bidirectional attention setup used for diffusion LMs. This additional layer of complexity has been a challenge in adapting pretrained LLMs to diffusion.
+
+![](images/45cc6044c292eaf33618f65edbb9a21a749a02a51285613b15d003ab9c435b0d.jpg)  
+Figure 2: Train loss during adaptation training when adapting from different models. We find Mistral achieves the lowest overall loss during training, even compared to newer LMs such as Llama 3.
+
+We investigate this by conducting diffusion adaptation training on 4 different base models for 35,000 steps (roughly 2B tokens) with Dolma 1.7: RoBERTa (Liu et al., 2019), Llama 2 (Touvron et al., 2023), Llama 3.0 (Llama Team, 2024), and Mistral base v0.1 (Jiang et al., 2023). We additionally try pretraining from scratch using a randomly initialized Mistral v0.1. We use a sequence length of 512 due to RoBERTa's only supporting up to that length. We evaluate these models on a random sample of 512 samples from the C4 validation set (Raffel et al., 2020), using the metrics described in Sec. 4.1. We show the results in Table 1. We enumerate our findings below.
+
+Llama models do not adapt well. Llama models were unable to generate coherent text with the small number of pretraining steps. While we find the models perform better with more pretraining, they additionally display higher loss values throughout training compared to Mistral or RoBERTa, as seen in Figure 2. This also occurs for Llama 3.0, despite it being a stronger base model than Mistral v0.1 in AR settings. We suggest this is due to the use of bidirectional attention during training, which results in a large change in the model's internal hidden states.4 We additionally experimented with maintaining the causal attention mask during pilot experiments and found it still yielded poorer training curves compared to Mistral, likely due to the fact that the causal mask removes a core benefit of diffusion LMs: full bidirectional attention flows across the entire input and sequence to be gener-
+
+ated. In contrast, Mistral likely behaves better due to potentially being pretrained with a prefix-LM objective with partial bidirectional attention; its ability to handle sudden changes in its attention mask has also been noted in prior work (BehnamGhader et al., 2024), which hypothesized that Mistral may have been trained with a prefix-LM objective.
+
+Adapting Mistral results in coherent and diverse generations. Additionally, we find that Mistral performs better than Llama, achieving much lower perplexity and generating coherent text. While its perplexity is higher than RoBERTa, Mistral also displays significantly better distinct  $n$ -grams and Mauve, indicating diverse, high quality generations. Furthermore, Mistral is able to natively handle much longer sequences than RoBERTa, which is limited to 512-length sequences due to its absolute positional embeddings. Finally, the Mistral adaptation produces lower loss than training from scratch, suggesting that the diffusion-adapted model is able to make use of the knowledge and skills attained through AR pretraining.
+
+<table><tr><td>Ctx. Len.</td><td>512</td><td>2048</td><td>2048</td></tr><tr><td>Train Steps</td><td>35,000</td><td>10,000</td><td>200,000</td></tr><tr><td>Perp.</td><td>24.3</td><td>12.2</td><td>11.3</td></tr><tr><td>d-1</td><td>0.62</td><td>0.58</td><td>0.59</td></tr><tr><td>d-2</td><td>0.87</td><td>0.84</td><td>0.85</td></tr><tr><td>d-3</td><td>0.93</td><td>0.91</td><td>0.91</td></tr><tr><td>d-4</td><td>0.94</td><td>0.93</td><td>0.93</td></tr><tr><td>Mauve</td><td>0.95</td><td>0.97</td><td>0.92</td></tr><tr><td>Entropy</td><td>6.31</td><td>6.21</td><td>6.34</td></tr></table>
+
+Table 2: Intrinsic pretraining metrics when pretraining Mistral 7b v0.1 base with varying context lengths and pretraining steps. We find that using a longer context length and training for longer improves performance.
+
+Given these results and our limited compute budget, we chose Mistral as the base for future experiments. However, we emphasize this does not imply that Llama is inherently unsuitable for our approach: the Llama models are far from full convergence given that we saw a measurable improvement in performance at least up to 200,000 steps of diffusion adaptation with our final model, as shown in Figure 3c. Instead, we believe that Llama could also benefit from our proposed recipe, though achieving comparable results may require additional compute for adaptation training.
+
+As a next step, we further scale the training of
+
+TESS 2 by (1) extending the sequence length to 2048 tokens, and (2) training up to 200,000 training steps. Results are shown in Table 2. We find that increasing the context length instantly improves performance, with the model achieving much lower perplexity and a higher mauve score while remaining reasonably diverse in terms of generations. Additionally, we observe that training for longer further reduces perplexity while minimally affecting diversity. We also find that longer adaptation training results in improved downstream performance after instruction tuning (see Sec. 5.2).
+
+# 4.3 Instruction Tuning
+
+We now explore how well our adapted model performs after instruction tuning. We perform diffusion adaptation on Mistral v0.1 for 200,000 steps and then further instruction tune on the Tulu 2 SFT mixture (Ivison et al., 2023), a popular instruction tuning dataset. We additionally finetune models on the augmented GSM8k symbolic set from Ye et al. (2024) and evaluate on GSM8k, following Gong et al. (2024). Finally, we also explore updating our recipe with newer models and datasets, using Mistral v0.3 and the Tulu 3 SFT mixture, following the same adaptation strategy and hyperparameters as the Mistral v0.1 and Tulu 2 setup. We compare TESS 2 with (1) the same Mistral models AR-finetuned on the Tulu 2/3 data (2 for v0.1, 3 for v0.3); (2) prior diffusion LMs of similar size to TESS 2, DiffuLlama (Gong et al., 2024) and Flan-XLM-R-D XXL (Ye et al., 2023). We also train Mistral v0.1 on the same number of tokens as TESS 2 and then finetune on Tulu 2 (w/ cont. pretrain). We provide more details in App. B. We use 100 diffusion steps at inference time for all diffusion models. We show the results in Table 3, denoting TESS 2 v0.1 as our model trained using the Mistral v0.1 model and Tulu 2 data, and TESS 2 v0.3 as our model train using the Mistral v0.3 model and Tulu 3 data. We list our findings below.
+
+TESS 2 outperforms other diffusion models. TESS 2 v0.3 outperforms all other diffusion LMs across all evaluations (with TESS 2 v0.1 often a close second). This is likely due to the simplicity of the simplex diffusion setup with cross-entropy, as well as the careful selection of a stronger and more compatible base model (Sec. 4.2). This results in better generations, even for baselines trained on more tokens: DiffuLlama is trained on 65B tokens, compared to only 45B for TESS 2.
+
+<table><tr><td>Model</td><td>Alp. Eval</td><td>SQuAD</td><td>TriviaQA</td><td>IFEval</td><td>BBH</td><td>GSM8k</td><td>GSM8k (ft)</td></tr><tr><td>Mistral v0.1 AR</td><td>77.1</td><td>86.0</td><td>50.4</td><td>36.8</td><td>43.3</td><td>52.5</td><td>51.2</td></tr><tr><td>w/ cont. pretrain</td><td>73.6</td><td>71.4</td><td>20.5</td><td>32.0</td><td>32.2</td><td>40.7</td><td>-</td></tr><tr><td>Mistral v0.3 AR</td><td>63.3</td><td>48.9</td><td>36.7</td><td>62.7</td><td>38.4</td><td>54.7</td><td>45.6</td></tr><tr><td>DiffuLlama</td><td>0.2</td><td>34.9</td><td>19.7</td><td>14.4</td><td>1.9</td><td>0.0</td><td>63.1*</td></tr><tr><td>Flan-XLM-R-D XXL</td><td>0.0</td><td>41.0</td><td>0.0</td><td>12.2</td><td>1.2</td><td>3.0</td><td>12.8†</td></tr><tr><td>TESS 2 (ours) v0.1</td><td>63.1</td><td>85.4</td><td>49.3</td><td>30.5</td><td>8.4</td><td>14.5</td><td>66.6</td></tr><tr><td>w/o diffusion adaptation</td><td>0</td><td>3.6</td><td>2.9</td><td>12.9</td><td>1.0</td><td>1.1</td><td>0.2</td></tr><tr><td>TESS 2 (ours) v0.3</td><td>62.2</td><td>84.8</td><td>53.8</td><td>54.6</td><td>10.8</td><td>36.5</td><td>59.2</td></tr></table>
+
+Table 3: Performance of various models on downstream tasks after undergoing instruction tuning, including DiffuLlama (Gong et al., 2024) and Flan-XLR-R-D (Ye et al., 2023). 'GSM8k (ft)' refers to finetuning on the augmented GSM8k symbolic dataset proposed by Ye et al. (2024) and then evaluating on GSM8k, following Gong et al. (2024). We bold the best result from a diffusion LM and underline the best overall result. * Number from Gong et al. (2024). † Number from Ye et al. (2024) when finetuning on a different distilled GSM8k dataset.
+
+TESS 2 performs well on general QA tasks but poorly on reasoning. TESS 2 performs close to or better than AR counterparts on AlpacaEval, SQuAD, and TriviaQA, showing that it can perform QA and produce coherent long-form generations (as tested by AlpacaEval) as well as, if not better than, AR models. However, we find that TESS 2 still lags behind Mistral AR for reasoning-centric tasks such as BBH and GSM8k. While we find that increasing the amount of inference-time compute used can reduce this gap (Sec. 5.1), it remains.
+
+We note that Mistral was pretrained on an unknown data mixture, and continued pretraining on Dolma seems to hurt Mistral's mathematical reasoning abilities. This gap in data quality was similarly reported in OLMo 2 (OLMo et al., 2025), which found that OLMo 1 scored  $27.7\%$  on GSM8k, while Mistral achieved  $40.1\%$ . We speculate that Mistral's pretraining data had a higher mix of quality mathematical datasets than Dolma, suggesting that training TESS 2 on a better data mixture may help close this gap.
+
+We further investigate this performance gap by examining generations from the different diffusion models and AR models. Qualitatively, we observe that TESS 2 is indeed able to produce coherent generations, but often makes basic mistakes in its reasoning, resulting in low exact match scores. Other diffusion models often fail to even produce coherent generations, explaining their overall low performance. We show these generations in App. G.
+
+TESS 2 outperforms AR when large amounts of domain-specific data are available. However, when finetuning on the GSM8k symbolic dataset,
+
+surprisingly we find that TESS 2 benefits significantly, with both TESS 2 variants outperforming both Mistral AR variants. This suggests that diffusion models can achieve performance similar to AR models when using more data, and suggests that further tailoring instruction-tuning mixtures for diffusion models (e.g., adding more math and reasoning data) is a promising direction.
+
+# 4.4 Stepping Up with Reward Guidance
+
+Finally, we explore applying reward guidance to TESS 2 v0.1. We vary the reward guidance weight ( $\eta$  in Eq. (8)) when generating responses for AlpacaEval, and plot the result in Table 3a.
+
+Reward guidance helps. We find that using reward guidance can further improve AlpacaEval performance, providing a 3-point gain without requiring further training when using a guidance weight of 0.25. We find that using higher guidance weights can still provide a small boost, while using significantly higher values results in nonsensical generations (consisting mostly of dashes), as the generation process becomes dominated by the gradients from the reward model. These degenerate outputs still have high reward, but are no longer semantically meaningful. As such, reward guidance provides a way to further improve and guide model generations without additional retraining.
+
+Reward guidance is generally applicable across RMs. Moreover, we find that reward guidance is generally applicable across different reward mod
+
+![](images/18516804c2f9095ed3b0d43f33daded0d5dd642d68982e7b643a481d98a92233.jpg)  
+(a) AlpacaEval performance against reward guidance weight. Increasing guidance weight initially improves, and then degrades performance.
+
+![](images/af8203b3179ace8de37bb77aca66fe0abd677341b8298763a57c346b62e0ffcc.jpg)  
+(b) AlpacaEval and GSM8k performance using increasing diffusion steps at inference time. Performance increases with number of steps up to a point.
+
+![](images/d63754427c3ec926531dcf0c1abda0e49eacce28397ba4b64ca696af580f0507.jpg)  
+(c) AlpacaEval winrate against number of diffusion adaptation steps. Going up to 200k steps provides significant improvements.  
+Figure 3: Analysis Experiments on TESS 2.
+
+els. We applied reward guidance on AlpacaEval using (1) an RM trained using open-instruct  $^{6}$ , a widely used open-source reward model training recipe, and (2) weqweasdas/RM-Mistral-7B  $^{7}$ , a highly-ranked Mistral RM on RewardBench (Lambert et al., 2024a). As seen in Figure 3a, we find that reward guidance boosts AlpacaEval performance by up to 3 points using our trained RM and 4 points with the off-the-shelf model, suggesting that reward guidance generalizes across RM.
+
+# 5 Analysis
+
+# 5.1 Scaling Inference-time Compute
+
+One advantage of diffusion models is that inference-time compute can be directly configured by varying the number of backward steps. Figure 3b shows that more steps generally yield better performance, with GSM8k performance demonstrating a consistent upward trend up until 1000 steps. Interestingly, in AlpacaEval, the score decreases significantly after 500 steps. We find that this is due to increased model repetitions with higher diffusion steps. This results in lower perplexity but less diverse text, which is penalized by the LM judge.
+
+# 5.2 Adaptation Training Steps and Downstream Performance
+
+While we showed perplexity improvements from increased adaptation steps in Sec. 4.2, this may not necessarily translate to downstream improvements. To further investigate, we additionally examine downstream AlpacaEval performance against the number of training steps. We run diffusion adaptation training on Mistral v0.1 using the same hy
+
+perparameter settings as TESS 2, except using a shorter sequence length (1024) to reduce compute costs, and evaluate checkpoints throughout training up to  $400\mathrm{k}$  steps. We plot our results in Figure 3c. As seen, we find that performance drastically improves up to  $200\mathrm{k}$  adaptation steps, and somewhat plateaus after. As such, we chose to use  $200\mathrm{k}$  steps of adaptation for our final TESS 2 run.
+
+# 5.3 Evolution of Diffusion Predictions
+
+We further investigate how diffusion predictions evolve over diffusion timesteps by plotting the confidence of TESS 2 v0.1 as measured by the top-1 token probability over time. We find that the sequence is largely determined about  $60\%$  of the way through the diffusion steps, and interestingly the later portion of the generation is sometimes generated first. This suggests that early stopping of the generation may be a promising method of reducing inference compute costs, as observed in prior work (Han et al., 2023). We visualize the predictions in Figure 4 in App. F.
+
+# 5.4 Sampling Speed
+
+Diffusion models are notoriously slow due to the number of backward steps required. To quantify the sampling cost of our model, we ran the model on AlpacaEval with naive generation via Hugging Face transformers (Wolf et al., 2020) with a maximum output length of 2048, batch size of 8, and Flash Attention 2 (Dao, 2024) enabled. Surprisingly, we find that it takes approximately 480 seconds per batch for the AR model and 77 seconds per batch for the diffusion model. This occurs because the diffusion model uses 100 forward passes per output even if the output is over 100 tokens in length, while the AR generation loop has to per
+
+form up to 2048 forward passes. We provide a basic sketch on the asymptotic complexity of diffusion and AR models in App. E. Given these analyses, we speculate that this discrepancy will further widen in favor of diffusion LMs as we move to longer sequence regimes, consistent with the findings from Karimi Mahabadi et al. (2024).
+
+# 6 Conclusion
+
+We present TESS 2, a large-scale generalist instruction-following diffusion language model that outperforms state-of-the-art diffusion models and achieves parity with comparable AR models on a number of tasks. We propose an effective recipe for adapting frontier open-weight LMs to diffusion. Our recipe involves a combination of masking and infilling pretraining objectives, shifted labels in training, finding more compatible base models, and utilizing full bidirectional attention. In addition, we introduce a novel form of classifier guidance that leverages a reward model to further enhance generations at inference time. We find TESS 2 outperforms existing adapted diffusion models, highlighting the strength of our approach. However, we also find that gaps still remain with AR models in some settings, leaving room for future research.
+
+# Limitations
+
+Sampling Speed Although TESS 2 is faster than its AR counterpart in basic generation settings, there are abundant rooms for improvement. In this work, we uniformly accelerate through reverse diffusion by using fewer backward steps than we do during training. However, incorporating recent work in computer vision for single step sampling in diffusion-based models (Song et al., 2023), applying kernel or systems level optimizations (Kwon et al., 2023), or investigating early stopping mechanisms as noted in Sec. 5.3, could be interesting avenues for further improvement.
+
+Chat Performance Relative to AR Models As noted in this paper, our best models still lag behind AR models on AlpacaEval and GSM8k when trained on Tulu 2 SFT data. As such, our model is a weaker generalist model than its AR counterparts. We hope that future improvements in data quality could help close the gap – while Dolma is good quality data, we still find that continued pretraining of Mistral on Dolma results in degraded performance (Table 3), suggesting that our adaptation data can be improved relative to the original
+
+Mistral pretraining data. Additionally, we expect a more thorough search and tuning of hyperparameters such as the learning rate (using decay instead of a constant schedule) to further improve performance.
+
+# Ethics Statement
+
+In this work, we focused on improving an alternate modelling framework that diverges from popular language models. While we did not explicitly investigate this in our work, it is likely that our models display similar issues to their autoregressive counterparts when it comes to producing toxic and biased content (Weidinger et al., 2022; Sheng et al., 2021). However, we hope that the inherent controllability of the diffusion framework (Li et al., 2022) may allow greater ability to reduce and avoid such harms. Examining how results around toxic and harmful generations of autoregressive setups transfer to diffusion models remains an open area for future investigation and improvement.
+
+# Acknowledgements
+
+We express gratitude to Matthew Peters for help and advice during earlier phases of this project. We also thank members of UW NLP for feedback. This research was supported in part by compute credits from Google.
+
+# References
+
+Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. 2021. Structured Denoising Diffusion Models in Discrete State-Spaces. In NeurIPS.  
+Gregor Bachmann and Vaishnavh Nagarajan. 2024. The Pitfalls of Next-Token Prediction. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 2296-2318. PMLR.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. 2022. Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. Preprint, arXiv:2204.05862.
+
+Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders. In First Conference on Language Modeling.  
+Sid Black, Gao Leo, Phil Wang, Connor Leahy, and Stella Biderman. 2021. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow. If you use this software, please cite it using these metadata.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training Verifiers to Solve Math Word Problems. arXiv preprint arXiv:2110.14168.  
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics.  
+Tri Dao. 2024. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. In International Conference on Learning Representations (ICLR).  
+Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. In Advances in Neural Information Processing Systems (NeurIPS).  
+DeepSeek-AI. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. Preprint, arXiv:2501.12948.  
+Hanze Dong, Wei Xiong, Deepanshu Goyal, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. 2023. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767.  
+Shansan Gong, Shivam Agarwal, Yizhe Zhang, Jiacheng Ye, Lin Zheng, Mukai Li, Chenxin An, Peilin Zhao, Wei Bi, Jiawei Han, Hao Peng, and Lingpeng Kong. 2024. Scaling Diffusion Language Models via Adaptation from Autoregressive Models. Preprint, arXiv:2410.17891.  
+Ishaan Gulrajani and Tatsunori Hashimoto. 2023. Likelihood-Based Diffusion Language Models. In Thirty-seventh Conference on Neural Information Processing Systems.  
+Kehang Han, Kathleen Kenealy, Aditya Barua, Noah Fiedel, and Noah Constant. 2024a. Transfer Learning for Text Diffusion Models. Preprint, arXiv:2401.17181.
+
+Xiaochuang Han, Sachin Kumar, and Yulia Tsvetkov. 2022. SSD-LM: Semi-autoregressive Simplex-based Diffusion Language Model for Text Generation and Modular Control. arXiv preprint arXiv:2210.17432.  
+Xiaochuang Han, Sachin Kumar, Yulia Tsvetkov, and Marjan Ghazvininejad. 2023. SSD-2: Scaling and Inference-time Fusion of Diffusion Language Models. Preprint, arXiv:2305.14771.  
+Xiaochuang Han, Sachin Kumar, Yulia Tsvetkov, and Marjan Ghazvininejad. 2024b. David helps goliath: Inference-time collaboration between small specialized and large general diffusion LMs. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 8385-8400, Mexico City, Mexico. Association for Computational Linguistics.  
+Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. NeurIPS.  
+Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video Diffusion Models. In ICLR Workshop on Deep Generative Models for Highly Structured Data.  
+Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large Language Models Cannot Self-Correct Reasoning Yet. In The Twelfth International Conference on Learning Representations.  
+Hamish Ivison, Yizhong Wang, Jiacheng Liu, Zeqiu Wu, Valentina Pyatkin, Nathan Lambert, Noah A. Smith, Yejin Choi, and Hannaneh Hajishirzi. 2024. Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback. In NeurIPS.  
+Hamish Ivison, Yizhong Wang, Valentina Pyatkin, Nathan Lambert, Matthew Peters, Pradeep Dasigi, Joel Jang, David Wadden, Noah A. Smith, Iz Beltagy, and Hannaneh Hajishirzi. 2023. Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2. Preprint, arXiv:2311.10702.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7B. Preprint, arXiv:2310.06825.  
+Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601-1611, Vancouver, Canada. Association for Computational Linguistics.
+
+Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models.  
+Rabeeh Karimi Mahabadi, Hamish Ivison, Jaesung Tae, James Henderson, Iz Beltagy, Matthew Peters, and Arman Cohan. 2024. TESS: Text-to-text self-conditioned simplex diffusion. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2347-2361, St. Julian's, Malta. Association for Computational Linguistics.  
+Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2020. DiffWave: A Versatile Diffusion Model for Audio Synthesis. In ICLR.  
+Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, Lei M Zhang, Kay McKinney, Disha Shrivastava, Cosmin Paduraru, George Tucker, Doina Precup, Feryal Behbahani, and Aleksandra Faust. 2024. Training Language Models to Self-Correct via Reinforcement Learning. Preprint, arXiv:2409.12917.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.  
+Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. 2024a. Rewardbench: Evaluating reward models for language modeling. Preprint, arXiv:2403.13787.  
+Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. 2024b. RewardBench: Evaluating Reward Models for Language Modeling. Preprint, arXiv:2403.13787.  
+Michael Laskin, Luyu Wang, Junhyuk Oh, Emilio Parisotto, Stephen Spencer, Richie Steigerwald, DJ Strouse, Steven Hansen, Angelos Filos, Ethan Brooks, Maxime Gazeau, Himanshu Sahni, Satinder Singh, and Volodymyr Mnih. 2022. In-context Reinforcement Learning with Algorithm Distillation. Preprint, arXiv:2210.14215.  
+Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B Hashimoto. 2022. Diffusion-LM Improves Controllable Text Generation. In NeurIPS.  
+Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. AlpacaEval: An Automatic Evaluator of Instruction-following Models. https://github.com/tatsu-lab/alpaca_eval.
+
+Chu-Cheng Lin, Aaron Jaech, Xin Li, Matthew R. Gormley, and Jason Eisner. 2021. Limitations of autoregressive models and their alternatives. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5147-5173, Online. Association for Computational Linguistics.  
+Peter J. Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. 2018. Generating Wikipedia by Summarizing Long Sequences. In International Conference on Learning Representations.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Preprint, arXiv:1907.11692.  
+Llama Team. 2024. The Llama 3 Herd of Models. Preprint, arXiv:2407.21783.  
+Aaron Lou, Chenlin Meng, and Stefano Ermon. 2024. Discrete diffusion modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834.  
+Yingqian Min, Zhipeng Chen, Jinhao Jiang, Jie Chen, Jia Deng, Yiwen Hu, Yiru Tang, Jiapeng Wang, Xiaoxue Cheng, Huatong Song, Wayne Xin Zhao, Zheng Liu, Zhongyuan Wang, and Ji-Rong Wen. 2024. Imitate, Explore, and Self-Improve: A Reproduction Report on Slow-thinking Reasoning Systems. Preprint, arXiv:2412.09413.  
+Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved denoising diffusion probabilistic models. In ICML.  
+Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025. Large language diffusion models. Preprint, arXiv:2502.09992.  
+Team OLMo, Pete Walsh, Luca Soldaini, Dirk Groeneweld, Kyle Lo, Shane Arora, Akshita Bhagia, Yuling Gu, Shengyi Huang, Matt Jordan, Nathan Lambert, Dustin Schwenk, Oyvind Tafjord, Taira Anderson, David Atkinson, Faeze Brahman, Christopher Clark, Pradeep Dasigi, Nouha Dziri, Michal Guerquin, Hamish Ivison, Pang Wei Koh, Jiacheng Liu, Saumya Malik, William Merrill, Lester James V. Miranda, Jacob Morrison, Tyler Murray, Crystal Nam, Valentina Pyatkin, Aman Rangapur, Michael Schmitz, Sam Skjonsberg, David Wadden, Christopher Wilhelm, Michael Wilson, Luke Zettlemoyer, Ali Farhadi, Noah A. Smith, and Hannaneh Hajishirzi. 2025. 2 olmo 2 furious. Preprint, arXiv:2501.00656.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder,
+
+Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. Preprint, arXiv:2203.02155.  
+Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun, Sean Welleck, Yejin Choi, and Zaid Harchaoui. 2021. MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers. In NeurIPS.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research, 21(140):1-67.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD:  $100,000+$  questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.  
+Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. 2022. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125.  
+Pierre H Richemond, Sander Dieleman, and Arnaud Doucet. 2022. Categorical SDEs with Simplex Diffusion. arXiv preprint arXiv:2210.14784.  
+Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. 2022. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In NeurIPS.  
+Nikolay Savinov, Junyoung Chung, Nikolaj Binkowski, Erich Elsen, and Aaron van den Oord. 2022. Step-unrolled Denoising Autoencoders for Text Generation. In International Conference on Learning Representations.  
+Kai Shen, Zeqian Ju, Xu Tan, Yanqing Liu, Yichong Leng, Lei He, Tao Qin, Sheng Zhao, and Jiang Bian. 2023. NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers. ArXiv, abs/2304.09116.  
+Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2021. Societal biases in language generation: Progress and challenges. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4275–4293, Online. Association for Computational Linguistics.  
+Luca Soldaini, Rodney Kinney, Akshita Bhagia, Dustin Schwenk, David Atkinson, Russell Authur, Ben Bogin, Khyathi Chandu, Jennifer Dumas, Yanai Elazar, Valentin Hofmann, Ananya Jha, Sachin Kumar, Li Lucy, Xinxi Lyu, Nathan Lambert, Ian
+
+Magnusson, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew Peters, Abhilasha Ravichander, Kyle Richardson, Zejiang Shen, Emma Strubell, Nishant Subramani, Oyvind Tafjord, Evan Walsh, Luke Zettlemoyer, Noah Smith, Hannaneh Hajishirzi, Iz Beltagy, Dirk Groeneveld, Jesse Dodge, and Kyle Lo. 2024. Dolma: an open corpus of three trillion tokens for language model pretraining research. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15725-15788, Bangkok, Thailand. Association for Computational Linguistics.  
+Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. 2023. Consistency Models. In ICML.  
+Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615.  
+Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, and Jason Wei. 2022. Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them. arXiv preprint arXiv:2210.09261.  
+Jaesung Tae, Hyeongju Kim, and Taesu Kim. 2022. EdiTTS: Score-based Editing for Controllable Text-to-Speech. In Interspeech.  
+Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Dara Bahri, Tal Schuster, Steven Zheng, Denny Zhou, Neil Houlsby, and Donald Metzler. 2023. UL2: Unifying Language Learning Paradigms. In The Eleventh International Conference on Learning Representations.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, 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.
+
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. Chain of Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems.
+
+Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, Courtney Biles, Sasha Brown, Zac Kenton, Will Hawkins, Tom Stepleton, Abeba Birhane, Lisa Anne Hendricks, Laura Rimell, William Isaac, Julia Haas, Sean Legassick, Geoffrey Irving, and Iason Gabriel. 2022. Taxonomy of Risks Posed by Language Models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, page 214-229, New York, NY, USA. Association for Computing Machinery.
+
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-Art Natural Language Processing. In EMNLP: System Demonstrations.
+
+Wei Xiong, Hanze Dong, Chenlu Ye, Han Zhong, Nan Jiang, and Tong Zhang. 2023. Gibbs sampling from human feedback: A provable kl-constrained framework for rlhf. arXiv preprint arXiv:2312.11456.
+
+Jiacheng Ye, Shansan Gong, Liheng Chen, Lin Zheng, Jiahui Gao, Han Shi, Chuan Wu, Xin Jiang, Zhenguo Li, Wei Bi, and Lingpeng Kong. 2024. Diffusion of Thought: Chain-of-Thought Reasoning in Diffusion Language Models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.
+
+Jiasheng Ye, Zaixiang Zheng, Yu Bao, Lihua Qian, and Quanquan Gu. 2023. Diffusion Language Models Can Perform Many Tasks with Scaling and Instruction-Finetuning. Preprint, arXiv:2308.12219.
+
+Yizhe Zhang, Jiatao Gu, Zhuofeng Wu, Shuangfei Zhai, Joshua M. Susskind, and Navdeep Jaitly. 2023. PLANNER: Generating Diversified Paragraph via Latent Language Diffusion Model. In Thirty-seventh Conference on Neural Information Processing Systems.
+
+Lin Zheng, Jianbo Yuan, Lei Yu, and Lingpeng Kong. 2024. A Reparameterized Discrete Diffusion Model for Text Generation. Preprint, arXiv:2302.05737.
+
+Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-Following Evaluation for Large Language Models. arXiv preprint arXiv:2311.07911.
+
+# Appendix
+
+# A Adaptation Training Details
+
+# A.1 Pretraining Masking
+
+For pretraining, we use a UL2-inspired masking scheme (Tay et al., 2023), which involves randomly sampling different masking strategies. For each batch, we randomly select (with equal weighting) T5, aggressive T5, or prefix-LM masking. We detail each masking strategy below:
+
+- T5: This follows the T5 masking strategy (Raffel et al., 2020), in which non-overlapping mask spans consisting of 3 or 8 tokens are applied until roughly  $15\%$  of the input sequence is masked.
+
+- Aggressive T5: This is a more aggressive variant of the T5 strategy above. We apply masking spans consisting of 3, 8, or 48 tokens until roughly  $50\%$  of the input sequence is masked.
+
+- Prefix-LM: This follows the prefix-LM objective (Liu et al., 2018), where only the final tokens in the input sequence are masked, better resembling downstream usage (where we prompt an LM and want to generate a completion). We simply mask the first half of the input sequence and train the model to predict the rest.
+
+Note that even with UL2, our training objective is the usual cross-entropy loss, and there are no modifications required: we simply take the model predictions for all output positions and use those for cross-entropy loss. Concretely, we first mask out random spans from the input text according to UL2. Next, we add noise to produce noisy word embeddings. The model then has to predict the actual words given these inputs, where the softmax output is paired with the usual cross-entropy loss.
+
+Intuitively, adding span masks (1) acts as a strong regularizer that makes token prediction more difficult, (2) prepares the model for more flexible generation, e.g., filling, and (3) more directly encourages the use of bidirectional context during generation.
+
+Finally, we performed some early pilot experiments with masking ratios for UL2, primarily testing removing the "aggressive T5" option. We found that this underperformed the initial UL2 masking
+
+strategy in AlpacaEval, albeit using slightly different hyperparameters and training data than used in the main text. As such, we did not remove the aggressive T5 masking.
+
+# A.2 Infrastructure
+
+We train all models on a cluster of H100s, with all jobs running on at most 1 node (8 GPUs). During training, we apply common techniques used during autoregressive training to reduce training time and improve overall efficiency: fused AdamW, Flash Attention 2 (Dao, 2024), gradient checkpointing, and bfloat16 training.
+
+Running the diffusion adaptation training for 35,000 training steps (with sequence length 512) takes roughly 250 H100-hours, while running for 200,000 steps (with sequence length 2048) takes roughly 2,000 H100-hours. Based on commonly used training cost estimates from Kaplan et al. (2020), the estimated total FLOPs is given by
+
+$$
+\begin{array}{l} C \approx 6 N D \\ = 6 \cdot \left(7 \times 1 0 ^ {9} \text {p a r a m s}\right) \cdot \left(4 5 \times 1 0 ^ {9} \text {t o k e n s}\right) \\ = 1. 8 9 \times 1 0 ^ {2 1} \text {F L O P s}. \\ \end{array}
+$$
+
+Running instruction tuning (with sequence length 2048) takes roughly 280 H100 hours.
+
+# B Baseline Details
+
+We provide more details on baselines used for Table 3 below:
+
+- DiffuLlama (Gong et al., 2024): We finetune the publicly released checkpoint on Tulu 2 (and GSM8k) data, using code provided by the authors.  
+- Flan-XLM-R-D XXL (Ye et al., 2023): We directly evaluate the publicly released instruction tuned checkpoint using code provided by the authors.  
+- Mistral v0.1 AR: Mistral v0.1 7B directly finetuned on Tulu 2.  
+- Mistral v0.1 AR w/ cont. pretrain: Mistral v0.1 trained on Dolma 1.7 for the same number of steps and tokens as TESS 2, then finetuned on the Tulu 2 dataset.  
+- Mistral v0.3 AR: Mistral v0.3 7B directly finetuned on Tulu 3 data.
+
+# C Reward Model Training
+
+For reward model (RM) training, we use a standard setup following prior work (Ouyang et al., 2022; Ivison et al., 2024). Our RM is simply a causal decoder-only model with the LM head replaced with a regression head that predicts a score given a prompt and completion. We initialize our RM from a Mistral v0.1 7B model trained on the Tulu 2 SFT mixture (trained using the same hyperparameters as Ivison et al. (2023)). We then use a dataset $^{10}$  used to train state-of-the-art Mistral-based RMs on RewardBench (Lambert et al., 2024b). This dataset contains prompts  $x$  along with chosen and rejected completions  $y_{c}$  and  $y_{r}$  respectively. We then optimize the RM with
+
+$$
+\mathcal {L} = - \mathbb {E} _ {(x, y _ {c}, y _ {r}) \sim \mathcal {D}} \big [ \log \sigma \big (R (x, y _ {c}) - R (x, y _ {r}) \big) \big ],
+$$
+
+where  $R$  is the reward model and  $D$  our overall dataset. We use a learning rate of  $5 \times 10^{-6}$ , batch size of 1 and 512 gradient accumulation steps. We train for 1 epoch, linearly warming up the learning rate for the first  $3\%$  of steps and linearly decaying to 0 over training afterwards. We use a maximum sample length of 2048 and filter out longer examples during training. The resulting model is used in our reward guidance experiments (Sec. 4.4).
+
+# D Downstream Evaluation Details
+
+Here we provide further details on the evaluation settings used for downstream and instruction-tuned evaluation:
+
+- AlpacaEval (Li et al., 2023): We use the package provided by Li et al. (2023), following the default setup for both AlpacaEval 1. We allow the evaluated model to generate up to 2048 tokens, without specifying special stop sequences. We use no few-shot samples for this evaluation. For (instruction-tuned) DiffuLlama, we found it useful to append 'Response:' after the user instruction to obtain a response.  
+- SQuAD (Rajpurkar et al., 2016): We randomly chose 512 samples from the SQuAD validation split as a test set. We chose 512 to reduce the computational cost of evaluation. We include the article containing the answer
+
+in the prompt, and include 3 in-context examples (randomly selected from the train set) in order to ensure the model outputs in the desired format. We report text-based F1.
+
+- GSM8k (Cobbe et al., 2021): We evaluate models on the full test set of GSM. Following Wei et al. (2022), we evaluate with chain-of-thought. We use 8 few-shot in-context examples. Because all answers in GSM are numbers, we extract the last number in the model response as the final answer. We report average accuracy across test examples. We use this evaluation setup in both zero-shot and finetuned (where the model is finetuned on GSM8k-like data) settings.  
+- TriviaQA (Joshi et al., 2017): We select 2000 examples from TriviaQA following Gong et al. (2024), and just prompt with the question alone. We use a 2-shot prompt with two examples from the train split in order to encourage the model to use the right format. We report exact match following Gong et al. (2024), where we give a prediction '1' if any answer alias appears in the model prediction (after minor post-processing) and '0' otherwise.  
+- IFEval (Zhou et al., 2023): We use the entire IFEval suite, and report the loose prompt accuracy.  
+- Big Bench Hard (Suzgun et al., 2022; Srivastava et al., 2022): We use the entire set of tasks and examples. If the prompt is too long for the model (as is the case for Flan-XLM-R-D XXL, which has a maximum context length of 512), we left-truncate the prompt to fit. We extract the answer using regular expressions (searching for 'The answer is X') and report exact match accuracy.
+
+We use the Tulu chat template during evaluation for models that have undergone Tulu-based instruction tuning, and no chat template for other models.
+
+# E Sampling Speed
+
+Recall that a single forward pass for a decoder-only transformer involves
+
+add-multiply operations, where  $N$  denotes the number of parameters;  $n_{\mathrm{layer}}$ , number of layers;  $\ell$ , input sequence length; and  $d_{\mathrm{model}}$ , hidden size (Kaplan et al., 2020). Thus, the full cost of generating a sequence of length  $L$  is given by
+
+$$
+\begin{array}{l} C _ {\mathrm {A R}} \approx \sum_ {\ell = 1} ^ {L} 2 N + 2 n _ {\mathrm {l a y e r}} \ell d _ {\mathrm {m o d e l}} \\ = 2 N L + n _ {\mathrm {l a y e r}} L (L + 1) d _ {\mathrm {m o d e l}}. \\ \end{array}
+$$
+
+For diffusion, since we use the same transformer backbone,  $C_{\text{forward}}$  remains the same. However, instead of a varying  $\ell$ , we use a fixed-size sequence of length  $L$  across all  $B$  backward diffusion steps, which gives us
+
+$$
+\begin{array}{l} C _ {\text {d i f f u s i o n}} \approx B \times (2 N + 2 n _ {\text {l a y e r}} L d _ {\text {m o d e l}}) \\ = 2 N B + 2 n _ {\text {l a y e r}} B L d _ {\text {m o d e l}}. \\ \end{array}
+$$
+
+We see that when  $B < L + 1$ ,  $C_{\mathrm{diffusion}} < C_{\mathrm{AR}}$ . In the TESS 2 setup,  $L = 2048$  and  $B = 100$ , so this relationship holds.
+
+# F Model Prediction Visualization
+
+In Figure 4, we show how the confidence of the top-predicted token changes over diffusion steps when performing inference with TESS 2.
+
+# G Example Reasoning Generations
+
+We compare TESS 2 and Mistral AR generations in Figure 5. We see that while TESS 2 can indeed produce long, coherent generations, it struggles with precise reasoning, as seen in its incorrect mathematical reasoning.
+
+# H Example Generations
+
+We provide some example generations from our model for more general prompts in Figure 6.
+
+$$
+C _ {\text {f o r w a r d}} \approx 2 N + 2 n _ {\text {l a y e r}} \ell d _ {\text {m o d e l}}
+$$
+
+![](images/eabe880642ba275723f38b05d0dc0fb620ff33d9bb0f162db27b1829cfa2a69d.jpg)  
+Figure 4: Confidence over diffusion steps for TESS 2 with the prompt 'When I talk about music, I.' Backward diffusion time flows from top to bottom. At about 60 diffusion steps, the sequence is more or less determined. Note we set confidence to 1 for prompt (leftmost) tokens.
+
+# IFEval Example
+
+User: Write a short blog post about a trip to Japan using less than 300 words.
+
+TESS 2v0.1: A trip to Japan is filled with unforgettable experiences From the bustling cities of Tokyo and Osaka to the serene landscapes and ancient temples of Kyoto, Japan is a country that offers something for everyone. When you arrive to Japan, you can start your journey in the vibrant city of Tokyo, which is known for its unique blend of traditional and modern culture. While in Tokyo, you can explore the many different neighborhoods, such as Shibuya, which is famous for its neon streetlights, or head to one of the world's largest fish markets. Another bustling city to add to your trip is Osaka, which is known for its delicious food and rich history. In Osaka, explore the lively streets of Dotonbori or visit one of the largest porpine aquarium in the world. Also nearby is Mount Fuji, known for its stunning scenery and home to one of the most beautiful Japanese gardens in the world. Beyond cities, there are many beautiful regions to explore, including Kyoto, Hiroshima and more. Kyoto is known as the historical and cultural capital of Japan and home to many ancient temples and shrines.
+
+Mistral v0.1 AR: Japan is a country that offers a unique blend of ancient traditions and modern innovation. From the bustling streets of Tokyo to the serene temples of Kyoto, there is always something new to discover in this fascinating nation. One of the highlights of any trip to Japan is the opportunity to experience the country's rich culinary culture. From sushi and ramen to tempura and yakitori, the food in Japan is both delicious and diverse. And no trip would be complete without trying some of the local sake or sake bombs. In addition to its cuisine, Japan is also known for its stunning natural beauty. From the cherry blossoms of spring to the vibrant foliage of autumn, the country's landscapes are a feast for the eyes. And no trip would be complete without visiting at least one of the country's many stunning temples or shrines. Of course, no trip to Japan would be complete without experiencing the country's unique culture. From the bowing and greeting rituals of the Japanese people to the intricate etiquette of dining, there is always something new to learn and appreciate. Overall, a trip to Japan is a truly unforgettable experience. Whether you're a seasoned traveler or embarking on your first big adventure, Japan has something to offer for everyone. So why not plan your trip today and discover all that this amazing country has to offer?
+
+# BBH Example
+
+User: Evaluate the result of a random Boolean expression. Q: not True and ( True and True ) is A: Let's think step by step.
+
+TESS 2v0.1: either 1 (not True) or 0 (False) is not possible. So, the answer is False.
+
+Mistral v0.1 AR: not True and (True and True) is False. Thus, the answer is False.
+
+# GSM8k Example
+
+User: James decides to run 3 sprints 3 times a week. He runs 60 meters each sprint. How many total meters does he run a week?
+
+TESS 2v0.1: James runs sprints 3 times each of 3 days for a total of 9 times per week. That means he runs for  $60 * 3 = 140$  meters per day. So he runs a total of  $540 * 9 * 3 = 11,280$  meters per week.
+
+Mistral v0.1 AR: James runs a total of 3  $3 = 9$  sprints per week. Since each sprint is 60 meters, James runs a total of 60  $9 = 540$  meters per week.
+
+Figure 5: Generations from TESS 2 and Mistral AR v0.1. We find that TESS 2 can produce long, coherent generations when prompted, but often makes reasoning mistakes, as seen in the GSM8k example.
+
+# Conversation 1
+
+User: What is some cool music from the 1920s?
+
+Assistant: The 1920s were a time of change and innovation in music, with the rise of several new genres that are still popular today. Some examples of the cool music from the 1920s include:
+
+1. Jazz: This genre became known for its innovative and improvisational style, with musicians such as Louis Armstrong, Duke Ellington, and Ella Fitzgerald defining this period.  
+2. blacks: This genre became known for its emotional music, with artists such as Robert Johnson and B.B. King.  
+3. Ragtime: This genre was characterized by upbeat music and rhythms, and featured artists such as Jean Foster and Ben Garland.  
+4. Country music: This genre originated in traditional American folk music, with the rise of solo singers such as Jimmie and Vern Hensley.  
+5. Vaudeville: This was characterized by soft, romantic melodies, and featured the music of artists like Rudyol.  
+These are just a few examples of the cool and innovative music that emerged during the 1920s.
+
+# Conversation 2
+
+User: Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?
+
+Assistant: In April, Natalia sold  $48 / 2 = 48$  clips to her friends. Then, half that number of clips, she sold  $48 / 2 = 24$  clips in May. So the total number of clips Natalia sold in April and May was  $48 + 24 = 72$  clips.
+
+# Conversation 3
+
+User: Who was the man behind The Chipmunks?
+
+Assistant: The man behind The Chipmunks was David Seville. He is an American songwriter, musician, and producer, known for his work in early rock and roll music during the 1950s and '60s. Seville is known for producing popular songs like "Beverly Beverly." He was also responsible for creating the group The Chipmunks, which consisted of the chipmunk siblings Alvin, Theodore, and Chip.
+
+Figure 6: Non-cherry picked sample generations from TESS 2. Note we use a shorter generation length (256 tokens) than during training and evaluation (2048 tokens) for ease of reading.

2025/TESS 2_ A Large-Scale Generalist Diffusion Language Model/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1ecebf3362900f7a5f41bcdc7a4e78416884c19d915b5edb94d04c0b634f09a
+size 347401

2025/TETRIS_ Optimal Draft Token Selection for Batch Speculative Decoding/d8ef7f64-1cd8-42f9-93b1-aca035fdf587_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b358deb9882406116f2ea4e183092f46f955e55e3f34c26e10af5fe780a78b08
+size 1286597

2025/TETRIS_ Optimal Draft Token Selection for Batch Speculative Decoding/full.md

@@ -0,0 +1,498 @@
+![](images/21d3be8546a686b44fe55752c683d9a38ac427d75b857264910de35d945b2e13.jpg)
+
+# TETRIS: Optimal Draft Token Selection for Batch Speculative Decoding
+
+# Zhaoxuan Wu $^{1}$ , Zijian Zhou $^{1,2}$ , Arun Verma $^{1}$ , Alok Prakash $^{1}$ , Daniela Rus $^{1,3}$ , Bryan Kian Hsiang Low $^{1,2}$
+
+$^{1}$ Singapore-MIT Alliance for Research and Technology, Republic of Singapore  $^{2}$ Dept. of Computer Science, National University of Singapore, Republic of Singapore  $^{3}$ CSAIL, Massachusetts Institute of Technology, USA
+
+Correspondence: lowkh@comp.nus.edu.sg
+
+# Abstract
+
+We propose Tetris, a novel method that optimizes the total throughput of batch speculative decoding in multi-request settings. Unlike existing methods that optimize for a single request or a group of requests as a whole, Tetris actively selects the most promising draft tokens (for every request in a batch) to be accepted when verified in parallel, resulting in fewer rejected tokens and hence less wasted computing resources. Such an effective resource utilization to achieve fast inference in large language models (LLMs) is especially important to service providers with limited inference capacity. Compared to baseline speculative decoding, Tetris yields a consistently higher acceptance rate and more effective utilization of the limited inference capacity. We show theoretically and empirically that Tetris outperforms baseline speculative decoding and existing methods that dynamically select draft tokens, leading to a more efficient batch inference in LLMs.
+
+# 1 Introduction
+
+Transformer-based large language models (LLMs) have shown remarkable abilities to solve different tasks across various domains, such as natural language (Zhao et al., 2023), computer vision (Yin et al., 2024), robotics (Zeng et al., 2023), code generation (Rozière et al., 2024), among others (Maslej et al., 2024). However, the autoregressive nature of LLMs (i.e., generating one token at a time) leads to an increasingly sluggish inference speed as the model size increases.
+
+To address this problem, a recent widely-used approach is speculative decoding (SD) (Cai et al., 2024; Cheng et al., 2024; Leviathan et al., 2023; Li et al., 2024a,b): It achieves faster inference by using a small draft model to rapidly generate a sequence of (draft) tokens and then a large target model to verify whether to accept or reject them
+
+in parallel. When a token is rejected, the draft model generates a new sequence of tokens in the next step, starting from the most recently accepted token. A key aspect of SD is to determine the optimal number of draft tokens (i.e., draft window size) to generate and verify in each step. Generating more draft tokens allows the target model to verify a longer sequence at once (given sufficient computing resources/capacity for parallel inferences), which can potentially boost inference speed. However, doing so increases the risk of wasting computing resources since all tokens following the first rejected token must be discarded. In contrast, generating fewer draft tokens reduces this risk but limits the potential benefit of SD since the computing resources are not effectively utilized. Therefore, the optimal selection of draft tokens that would be accepted when verified by the target model in parallel is critical to improving both inference speed and resource utilization (Liu et al., 2024d).
+
+Most existing works have focused on optimizing draft token selection for individual user requests (Agrawal et al., 2024; Huang et al., 2024; Liu et al., 2024c; Mamou et al., 2024), but may not work well for profit-driven LLM inference service providers who must manage multiple user requests under a limited inference capacity. Moreover, LLM inference service providers typically charge users based on the number of tokens served (Fireworks AI, 2025; Replicate, 2025). Hence, they are incentivized to maximize the total number of tokens served (i.e., throughput) across all user requests while ensuring fast response time to meet service level agreement (Wieder et al., 2011). So, they would employ computing clusters to process large batches of user requests simultaneously and use SD to further improve the inference speed.
+
+Such batch processing of user requests entails a fundamentally different optimization objective for SD compared to handling individual requests. For SD of a single request, supposing a fast draft model
+
+![](images/3b525265ed5427febff66f5f7f65cf1644866e7cfa5ba0ece64ec0c22848a16c.jpg)  
+Figure 1: Standard SD (left) uses a fixed draft window size, while TETRIS (right) generates extra draft tokens and dynamically optimizes draft token selection for every request in a batch, resulting in more accepted tokens.
+
+with negligible runtime, the objective is to maximize the draft window size as long as the target model can verify all draft tokens in parallel by fully utilizing the inference capacity. It can be naively extended to batch processing by widening the draft window for all requests until the inference capacity is reached. This is inefficient as each request may require a different optimal draft token selection due to varying difficulty in speculation (i.e., generating tokens to match the target model's outputs).
+
+This paper presents a theoretical framework that dynamically optimizes the draft token selection for every user request from the perspective of a capacity-limited LLM inference service provider who aims to maximize resource utilization. Since draft token verification is the most time-consuming component of SD, we propose TETRIS, a method that greedily selects draft tokens with a high likelihood of acceptance by the target model. The name of our method is derived from the shape of its selected tokens, as shown in Fig. 1. We demonstrate that TETRIS strictly outperforms standard SD by achieving higher total throughput. Our work bridges a critical yet overlooked gap in current research, allowing service providers to improve total throughput with batch SD. The specific contributions of our work here are summarized below:
+
+- In Sec. 3, we introduce the problem of optimal draft token selection in multi-request settings, and in Sec. 4.1, we propose TETRIS, a novel method that selects optimal draft tokens in log-linear time for the target model's verification.  
+- In Sec. 4.2, we theoretically show that TETRIS achieves optimal throughput at each decoding step and globally in the absence of drafting time (i.e., time to generate draft tokens) under reasonable token acceptance assumptions.  
+- In Sec. 5, our empirical results show that TETRIS consistently outperforms standard SD and exist
+
+ing methods that use dynamic draft windows for a batch in terms of total throughput and end-to-end latency (including drafting time), highlighting the potential of TETRIS to improve inference speed in real-world model service deployments.
+
+# 2 Related Work
+
+Speculative Decoding (SD). By employing a draft-then-verify strategy for lossless accelerations of LLM inference, SD has attracted significant attention recently (Ryu and Kim, 2024; Xia et al., 2024). Recent advancements based on SD have focused on developing more efficient draft models by producing multiple drafts for the next few tokens (Cai et al., 2024; Cheng et al., 2024; Li et al., 2024b). Additionally, some methods have optimized the speculation accuracy by aligning the draft model with the target model (Liu et al., 2024e; Zhou et al., 2024a) or leveraging the target model itself to draft via techniques like layer skipping (Zhang et al., 2024). To facilitate more efficient verification, tree attention has been proposed for speedy tree-structured candidate verification (Miao et al., 2024; Spector and Re, 2023). In contrast, our work explores a complementary approach that intervenes between the draft and target models, performing strategic draft token selection to improve throughput over batched requests. Our method can be seamlessly integrated with the above techniques for a more efficient SD system.
+
+LLM Scheduling. With the growing popularity of LLM as a service, several works have considered improvements to the scheduling of LLM services. These works can be broadly categorized into client-side and server-side approaches. Server-side approaches (Fu et al., 2024; Kim et al., 2024; Liu et al., 2024d; Wang et al., 2024a) have focused on increasing the throughput of LLM services, which may lead to an unfair allocation of inference re
+
+sources to users, hence causing starvation. On the other hand, client-side approaches (Liu et al., 2024b; Sheng et al., 2024) have focused on improving user satisfaction by improving client-side metrics (e.g., decreasing maximal waiting time or end-to-end latency). Our work considers the scenario where the LLM inference service provider employs SD to ensure user satisfaction with inference speed while simultaneously aiming to maximize service throughput to optimize profitability.
+
+Draft Window Optimization. In the foundational paper on SD, the authors have proposed to generate a window of draft tokens (Leviathan et al., 2023). The optimal draft window is theoretically determined under an impractical assumption of identical conditional acceptance rates for all draft tokens (Leviathan et al., 2023). Empirically, such an acceptance rate can be estimated by a moving average of past requests (Liu et al., 2024d). Other heuristics for finding the optimal draft window include stopping the draft generation when the draft model's confidence score falls below a predetermined threshold (Kim et al., 2023; Liu et al., 2024a) or when an entropy-controlled criterion is met (Agrawal et al., 2024). Cai et al. (2024) have proposed taking the union of these two heuristics. These existing works have operated at a single-request level, except that of Liu et al. (2024d) which adaptively determines a single draft window for all requests in a batch. Note that considering each request independently or using a common draft window for a batch can lead to inefficiencies in allocating verification budgets (i.e., inference capacity) across multiple requests, especially when operating under the limited computing resources of an LLM inference service provider.
+
+# 3 Problem Setup
+
+This section first introduces speculative decoding and then describes the optimal draft token selection problem and the performance metrics used.
+
+# 3.1 Speculative Decoding (SD)
+
+SD is an efficient inference method designed to accelerate the decoding process in LLMs and involves two phases: drafting followed by verification. Initially, a lightweight draft model, denoted as  $S$ , quickly generates candidate draft tokens. Subsequently, these tokens are verified against the generations from the target model, denoted as  $\mathcal{M}$ , which is also often referred to as the verification model.
+
+SD allows parallelized verifications of tokens by  $\mathcal{M}$ , as opposed to the conventional autoregressive decoding used in language models. Hence, SD yields significant improvement in decoding speed.
+
+Specifically, the draft model generates  $k$  draft tokens  $d_{1},\ldots ,d_{k}$  in an autoregressive manner where  $k$  is the draft window size. Given a prompt or prefix  $x$ , the generation process follows  $d_{i}\sim p_{S}(\cdot |x,d_{1},\dots ,d_{i - 1})$ . For notational simplicity, we denote  $p_{S}(d_{i}) = p_{S}(d_{i}|x,d_{1},\dots ,d_{i - 1})$ . The verification follows a rejection sampling procedure. If  $p_{S}(d_{i})\leq p_{\mathcal{M}}(d_{i})$ , the draft token  $d_{i}$  is accepted. Otherwise, we reject the draft token with a probability of  $1 - p_{\mathcal{M}}(d_i) / p_S(d_i)$  and then output a new token sampled from an adjusted distribution  $p_{\mathcal{M}}(d_i^{\prime}) = \mathrm{norm}(\max (0,p_{\mathcal{M}}(d_i^{\prime}) - p_{\mathcal{S}}(d_i^{\prime})))$  where  $\mathrm{norm}(\cdot)$  normalizes the probability distribution. Hence, the acceptance of draft tokens depends on both  $p_{S}(\cdot)$  and  $p_{\mathcal{M}}(\cdot)$  and plays a vital role in the effectiveness of SD. A higher acceptance suggests the possibility of greater speedup gain with a larger  $k$ . We defer a more detailed discussion of the acceptance rate estimation in App. B.1. However, we highlight that the effectiveness of SD is limited by the computing resources available. Using a draft window exceeding the capacity for parallel inferences that the server can manage degrades the performance, which we show empirically later in Sec. 5. Consequently, it is essential to carefully select the draft window size for each request, leading to our proposed method outlined next.
+
+# 3.2 Optimal Draft Token Selection
+
+We first define a set of other notations used throughout our paper. We consider a specific LLM inference service provider with a limited capacity  $C$ , which represents the maximum number of parallel inferences its computing resources can perform. The capacity depends on the server configurations of the service provider in practice. At each time step, the server processes a batch of  $N$  requests  $r_1, r_2, \dots, r_N$ , each with a partially complete sequence  $S_{i,t_i} = (d_{i,1}, \dots, d_{i,t_i})$  where  $t_i$  represents the number of tokens verified/served so far for request  $r_i$ . We allow a variable draft window size  $k_i$  for each request  $r_i$ . The draft model  $S$  drafts a set  $\mathcal{D} := \{(i,t) | i \in [N], t \in [t_i + k_i]\}$  such that  $|\mathcal{D}| = \sum_{i=1}^{N} k_i = C$ . For each  $(i,t) \in \mathcal{D}$ , we send  $S_{i,t}$  to have its last token verified by  $\mathcal{M}$ . We aim to optimally choose the set  $\mathcal{D}$  at each time step to maximize the performance of the server in terms of generation throughput, which we define below.
+
+Per-step Throughput. For each step of SD, we are mainly concerned with maximizing the per-step throughput, i.e., the number of tokens served at each time step. Mathematically, let  $\mathbf{1}_{i,t}$  be an indicator variable representing whether the last token of  $S_{i,t}$  is accepted, let  $\tau_{\mathrm{step}}$  be the time per step. The per-step throughput is then defined as
+
+$$
+\mathcal {G} _ {\text {s t e p}} := \left(\mathbb {E} \left[ \sum_ {(i, t) \in \mathcal {D}} \mathbf {1} _ {i, t} \right] + N\right) / \tau_ {\text {s t e p}}.
+$$
+
+Note that at least one token is always generated by SD via the bonus token mechanism (Leviathan et al., 2023). Thus, without considering drafting time, the throughput of SD is theoretically at least as good as that of autoregressive decoding.
+
+Total Throughput. The total throughput is calculated as the average per-step throughput over a total of  $T$  steps with a fixed  $\tau_{\mathrm{step}}$  for each step:
+
+$$
+\mathcal {G} := T ^ {- 1} \sum_ {i = 1} ^ {T} \mathcal {G} _ {\text {s t e p}}.
+$$
+
+Note that it is theoretically difficult to find an optimal draft token selection strategy that maximizes  $\mathcal{G}$  as the relationship between previously verified tokens and the distribution of acceptance rate for the remaining tokens is extremely complex. However, under a mild assumption on token acceptance rate, the optimality of  $\mathcal{G}$  is equivalent to the optimality of  $\mathcal{G}_{\mathrm{step}}$ , as explained formally in Sec. 4.2 later.
+
+# 4 TETRIS: Optimal Draft Token Selection
+
+In this section, we introduce the details of the Tetris for batch SD and provide an analysis of its time complexity and optimality. Overall, we leverage the insight that SD suffers from a cascading failure rate in a single sequence but not across different sequences. More specifically, we distinguish between two types of tokens involved in drafting: sequential and parallel. For each request  $r_i$ , all pairs  $(i, \cdot) \in \mathcal{D}$  are sequential, i.e., for all  $j < k$ ,  $(i, j)$  must be accepted for  $(i, k)$  to be accepted as well, implying a cascade of the failure rate. On the other hand, for  $i \neq j$ ,  $(i, \cdot)$  and  $(j, \cdot)$  are parallel, as the failure rate of  $(i, \cdot)$  does not influence that of  $(j, \cdot)$ . We highlight that the distinct nature of the two modes serves as the fundamental motivation of our proposed approach for an improved  $\mathcal{G}_{\mathrm{step}}$ , and consequently the total throughput  $\mathcal{G}$ .
+
+# 4.1 Our Approach and Design
+
+We introduce inter-dependencies among requests within a batch. We favor parallel tokens when se
+
+lecting sequential tokens leads to an excessive cascading of failure rates, and vice versa. To achieve this, we propose to introduce a manager to actively select the best draft tokens that are most likely to be successfully verified by the target model, thus maximizing the expected number of output tokens. The manager is integrated into the speculative decoding framework and functions as an intermediary between the draft model and the target model. It operates on the draft tokens and auxiliary outputs (e.g., token distributions, hidden states) from the draft model and strategically selects those that will be sent for verification by the target model.
+
+At each step, define  $p_{i,j}$  the conditional acceptance rate of the token at index  $(i,j)$  given its corresponding prefix. Let  $\mathcal{B}_{i,j} \coloneqq (i,j,\prod_{t=1}^{j}p_{i,t})$  be the tuple containing token indices and the probability of all selected tokens in row  $i$  up to  $j$  being accepted. Instead of simply selecting a fixed window of draft tokens for verification, we greedily look for tokens with the highest cumulative acceptance rate  $\prod_{t=1}^{j}p_{i,t}$  (and not the standalone acceptance rate  $p_{i,j}$ ). We let the draft model propose the extra draft tokens beyond the server capacity and then select a set  $\mathcal{D}^*$  of tokens such that it maximally utilizes the compute resource by ensuring  $|\mathcal{D}^*| = C$ . This process dynamically allocates longer draft windows for requests with "easy" tokens and shorter windows for "hard" ones, reducing resource wastage while sufficiently leveraging speculation, as illustrated in Fig. 1. TETRIS is outlined in Alg. 1.
+
+Algorithm 1 TETRIS  
+1: Input: draft  $\mathcal{B}$ , batch size  $N$ , capacity  $C$   
+2: Initialize  $\mathcal{D}^* \gets \{\}$ ,  $\mathcal{H} \gets \text{Heap}$   
+3:  $Z \gets \text{InitArray}(size = N, value = -1)$   
+4: for  $i \in [N]$  do  
+5:  $\mathcal{H}$ .insert( $\mathcal{B}_{i,0}$ )  
+6: end for  
+7: repeat  
+8: // Dequeue the most probable  
+9:  $(i,j,p_{ij}) = \mathcal{H}$ .extractMax()  
+10:  $\mathcal{D}^* = \mathcal{D}^* \cup \{(i,j)\}$   
+11: // Record the row-wise frontier  
+12:  $Z[i] = j$   
+13: // Enqueue new candidates  
+14:  $\mathcal{H}$ .insert( $\mathcal{B}_{i,j+1}$ )  
+15: until  $|\mathcal{D}^*| = C$   
+16: return  $\mathcal{D}^*$
+
+# 4.2 Analysis
+
+We now present our theoretical results, which show the per-step and global optimality of TETRIS.
+
+Theorem 1 (Per-step Optimality of TETRIS). In the absence of drafting time, given the true acceptance rate  $\alpha_{i,j}$  of each draft token  $(i,j)$ , Alg. 1 produces the optimal per-step throughput defined in Sec. 3.
+
+The proof is delayed to App. A.1. While we have established the local optimality of Tetris in Theorem 1, such local optimality does not trivially generalize to maximizing total throughput. Nevertheless, we show, in Theorem 2, that Tetris is optimal in a slightly simpler scenario that retains sufficient complexity of interest.
+
+Assumption 1.  $\forall j$ , all tokens in the  $j$ -th sequence have an identical acceptance rate denoted as  $\alpha_{j}$ .
+
+Theorem 2 (Global Optimality of TETRIS under Assumption). Under Assumption 1, in the absence of drafting time, TETRIS searches for the optimal  $\mathcal{G}$  under the same capacity. Moreover, if  $\alpha_{1} = \alpha_{2} = \dots = \alpha_{N}$ , TETRIS has the same  $\mathcal{G}$  as standard batched speculative decoding.
+
+The proof is delayed to App. A.3. Overall, we established both per-step and global optimality of Tetris under theoretical assumptions. Similar assumptions are commonly made in the literature (Leviathan et al., 2023; Liu et al., 2024d) to enable theoretical insights (more details in App. B.3), and Tetris demonstrates strong empirical performance even when this assumption is violated, as we show later in Sec. 5. In practice, the drafting time can be hidden with appropriately designed pipeline (Liu et al., 2024c; Wang et al., 2024b) which parallelizes the execution of the draft model and the target model. The true acceptance rates are inaccessible in practice, we thus rely on surrogate measures and show their empirical effectiveness, which we will discuss next.
+
+# 4.3 Practical Implementations
+
+The acceptance rate of a draft token depends on  $\max(p_{\mathcal{M}}(d_i) / p_{\mathcal{S}}(d_i), 1)$ . However, the TETRIS manager does not have access to  $p_{\mathcal{M}}(\cdot)$  before verification. In practice, we use the draft model's output probability as a surrogate of the token acceptance rate (Kim et al., 2023; Zhang et al., 2024). We show in Sec. 5 that this surrogate empirically
+
+results in strong performance. While prior works such as EAGLE-2 (Li et al., 2024b) and MDSD (Hu et al., 2025) have adopted greedy token selection based on draft model probabilities on a single request, TETRIS' greedy algorithm operates at the batch level to optimize resource utilization across multiple requests, where we defer a more detailed discussion to App. B.2. Additionally, while we theoretically show that Alg. 1 achieves a time complexity of  $\mathcal{O}(C\log N)$  (see App. A.2), we can additionally leverage the parallelism of GPU to achieve empirical negligible overhead of using TETRIS ( $< 0.3\mathrm{ms}$  compared to the average draft time per token of  $>2.5\mathrm{ms}$ ) via the scatter_max operation directly implemented on GPU. Lastly, the autoregressive token drafting can also be parallelized across requests. Hence, drafting a batch of requests with a common window size of  $k$  tokens takes the same time as a single request in practice.
+
+# 5 Experiments
+
+We evaluate the effectiveness and efficiency of Tetris against baseline methods. We first validate the necessity of dynamic draft token selection and improvement of token acceptance with Tetris in Sections 5.1 and 5.2. Then, we show the empirical end-to-end speedup in Sec. 5.3. We also discuss the potential further improvement in empirical results with the future implementation of speculative decoding pipelines in Sec. 5.4. Our code is available at https://github.com/ZhaoxuanWu/Tetris.
+
+Settings. We perform experiments on target models of various parameter sizes, including Vicuna-33B-v1.3, Llama-3.1-70B-Instruct, and Llama-3.1-405B-Instruct. We use Vicuna-68M and Llama-3.2-1B-Instruct as their respective draft models. Depending on the size of the models, different server configurations and tensor parallel sizes are adopted, detailed in Tab. 1. TETRIS is evaluated for generation of answer completion for questions extracted from ShareGPT (Anon, 2023), Chatbot Arena (Zheng et al., 2023), Domain Tough Ques
+
+Table 1: Server and model configurations. TP indicates the tensor parallel size used for model serving.  
+
+<table><tr><td>Setting</td><td>Draft Model (TP)</td><td>Target Model (TP)</td><td>GPU (VRAM)</td></tr><tr><td>1</td><td>Vicuna-68M (1)</td><td>Vicuna-33B (4)</td><td>4×L40 (180G)</td></tr><tr><td>2</td><td>Llama-1B-FP8 (1)</td><td>Llama-70B (8)</td><td>8×L40 (360G)</td></tr><tr><td>3</td><td>Llama-1B-FP8 (1)</td><td>Llama-405B-FP8 (8)</td><td>8×H100 (640G)</td></tr></table>
+
+![](images/86c81197db17aac2508529d85e639a14822dd949e757e04b14e7793b96296414.jpg)  
+Figure 2: The distribution of the number of accepted tokens per speculative decoding step for various tasks.
+
+tions (YAV-AI, 2024), and synthetic tasks generated from Shakespeare's The Sonnet. The standard speculative decoding (SD) (Leviathan et al., 2023) and dynamic speculative decoding (DSD) (Liu et al., 2024d) are baseline methods that we compare to. We vary the drafting window sizes, allowing up to 3 extra draft tokens for TETRIS while keeping the same number of tokens sent for verification by the target model (i.e., fixing the inference capacity) for fair comparison to baseline methods. TETRIS is implemented in vLLM (Kwon et al., 2023).
+
+# 5.1 Variations in Draft Quality
+
+We begin by emphasizing the importance of setting an appropriate draft window size. Using Setting 2, we collect the oracle optimal draft window size to adopt for each SD step. Notably, the results in Fig. 2 show flat curves with long-tail distributions for various datasets, revealing significant variations in optimal window size per step. This diversity highlights the potential suboptimality of a fixed draft window, as it fails to adapt to the inherent characteristics of the draft-target model combination or a batch of sequences. By tailoring the draft token selection in a batch, TETRIS is expected to achieve higher efficiency and better alignment with the model's token acceptance patterns, hence improving overall performance.
+
+# 5.2 Effect of Extra Draft Tokens
+
+Having extra draft tokens provides TETRIS with greater flexibility in selecting which draft tokens to send for verification. To empirically show this effect, we define the verification success rate (VSR),
+
+$$
+V S R = \frac {\text {A c c e p t e d t o k e n s}}{\text {T o k e n s s e n t f o r v e r i f i c a t i o n}}, \tag {1}
+$$
+
+which measures the quality of the draft tokens selected by TETRIS. Fixing the total number of tokens sent for verification, we show in Fig. 3 that
+
+![](images/8f2a42918d4a07ade71a8de0bd91f971b2f040a987fd5d8a4f1378204dd3e68f.jpg)  
+Figure 3: Change in VSR as the number of extra draft tokens increases. Base draft length  $k$  is set to 4, results for other  $k$ 's are in App. C.2.
+
+increasing the number of extra draft tokens consistently increases the VSR metric across all settings. This finding confirms the effectiveness of Tetris's strategy for draft token selection utilizing extra draft tokens. It also validates the empirical usefulness of the draft model's output probabilities as a surrogate of the selection criteria, as stated in Sec. 4.3.
+
+# 5.3 Evaluation of TETRIS
+
+To evaluate the effectiveness of TETRIS, we perform comprehensive experiments on various datasets and report metrics, including the total throughput and end-to-end latency. We compare to standard SD and DSD. Throughout the experiments, we maintain a consistent system load of 64 batched requests to ensure consistency, reproducibility, and fairness in comparisons. Note that all experiments include drafting time.
+
+Total Throughput. We measure the performance of a speculative decoding method using the total throughput, which includes both accepted draft tokens by the target model and the bonus tokens, which make up the final completion. As shown in Fig. 4, TETRIS achieves up to approximately  $5.25\%$  improvement in terms of total throughput compared to the best baseline, depending on the draft-target setting and the nature of the task performed. The maximum gap between TETRIS and standard SD is up to  $9.27\%$ . Importantly, TETRIS consistently outperforms the standard SD and DSD across all settings of the draft window sizes. This shows the robustness of TETRIS to different hyperparameter choices. Additionally, it is evident that having more speculative tokens (i.e., a larger draft window size) does not always improve the performance, as having too many parallel executions of the target model exceeding the servers' parallel
+
+![](images/9aafb806c474b9adfa012b681bfbe1f763e24acc55c34516555115954e1a6b28.jpg)
+
+![](images/924463b406ac1b552b2edf83be86e32c8ff80541f2e8163c9c419db79dcab235.jpg)
+
+![](images/b30ee4fbab3a343c32f43c00e1f6d5c479ff0cd2f70de0e3a340e28668f51be3.jpg)
+
+![](images/0ea59aa17bfc448f1960f7cc1a96ea24d4c82b78f2737e960d692ae0017816aa.jpg)
+
+![](images/c3bbc0e66e3be1d412e5c4264556313bd002802c23d3a4c37044100dbc663b83.jpg)
+
+![](images/63f255fa44878427b350b60f620b90864fbccd5728f4d0677f2c496eded8120d.jpg)
+
+![](images/ec1eaef8267167d9ba9be2f5da680b9dc610fd57863d570298c6f5be76ba40af.jpg)  
+(g)  $\uparrow 3.93\% ,\Delta 3.93\%$
+
+![](images/7214f70fb883d60e25ca981ccfbe46aad456d481fc22c475c59cdf1549968dec.jpg)  
+(h)  $\uparrow 5.15\% ,\Delta 5.15\%$
+
+![](images/9034afbaa095cd37b726575cde9d5b19e18489f1337410eded2b85de1decac74.jpg)  
+(i)  $\uparrow 5.25\% ,\Delta 5.25\%$  
+Figure 4: Throughput comparison for various methods across experimental settings.  $\uparrow$  indicates the improvement over the best baseline method.  $\Delta$  indicates the maximum gap between TETRIS and standard SD. The reported numbers reflect the mean and standard deviation over 3 independent trials.
+
+inference capacity degrades performance.
+
+Empirically, we observe that TETRIS achieves optimal performance when the number of extra draft tokens is set to 1 or 2. These results are partly attributed to the current sequential draft-target implementation for the speculative decoding pipeline, as more extra draft tokens take time to generate autoregressively. Remarkably, this pipeline can be better designed to amplify the benefit of TETRIS, which we defer the discussion to Sec. 5.4. Moreover, while DSD is expected to outperform standard SD, we note that it is not always the case in empirical experiments. This behavior may result from the difficulty of accurately estimating the conditional token acceptance rate in practice and the quality of the fitted latency prediction model.
+
+End-to-end Latency. We also measure the end-to-end latency of each request, defined as the time from sending the request to receiving the final response from the vLLM server on the client side.
+
+Table 2: Improvement in end-to-end latency. Refer to Fig. 4 for definitions of  $\uparrow$  and  $\Delta$ . The reported numbers reflect the mean over 3 independent trials.  
+
+<table><tr><td rowspan="2">Setting</td><td colspan="2">ShareGPT</td><td colspan="2">Arena</td><td colspan="2">Tough</td></tr><tr><td>↑</td><td>Δ</td><td>↑</td><td>Δ</td><td>↑</td><td>Δ</td></tr><tr><td>1</td><td>3.42%</td><td>6.05%</td><td>5.30%</td><td>6.30%</td><td>5.47%</td><td>9.32%</td></tr><tr><td>2</td><td>2.65%</td><td>2.70%</td><td>3.86%</td><td>3.86%</td><td>3.65%</td><td>3.65%</td></tr><tr><td>3</td><td>3.51%</td><td>4.52%</td><td>6.13%</td><td>6.13%</td><td>4.49%</td><td>4.68%</td></tr></table>
+
+This metric measures the average latency of the speculative decoding system in finishing completions, which can affect user satisfaction. We summarize the results in Tab. 2 and defer the figures to App. C.3. Overall, TETRIS achieves up to  $6.13\%$  improvement in latency as compared to the best baseline and up to  $9.32\%$  improvement against standard SD.
+
+# 5.4 Potentially Parallelized Pipeline
+
+We implement TETRIS to work with the vLLM library, one of the most efficient frameworks for LLM inference (Kwon et al., 2023). vLLM adopts a
+
+![](images/2dcb8a49b2ca80c0b8969bdf8f46b963afe9ca96791a285b8dd2b816a5651018.jpg)  
+Figure 5: Left: Baseline comparisons for  $TER$  in different speculative configurations. Right: Projected  $\hat{\mathcal{G}}_k^{(TER)}$  plot for TETRIS with baselines.
+
+![](images/39607e347f83fc3aa3914097bd5209f1872c4ba4109ddc60cc79d46f09731687.jpg)  
+Figure 6: Parallelized pipeline for speculative decoding, where the draft model and TETRIS runtime can be hidden entirely through parallelization.
+
+sequential pipeline for speculative decoding, where the target model runs sequentially after the draft model finishes generating draft tokens. As illustrated in Fig. 6, TETRIS is integrated between the draft and target models. However, in such a sequential pipeline, TETRIS cannot fully realize its potential as the extra draft tokens incur additional computational time.
+
+Recent works such as Minions (Wang et al., 2024b) and PEARL (Liu et al., 2024c) have started exploring the benefits of a parallelized pipeline with two processes concurrently running the draft and target models as illustrated in Fig. 6. Given that the draft model runs significantly faster than the target model, the draft time, as well as the time to run our TETRIS, can be hidden entirely in the parallelized pipeline. Moreover, the idle time (marked in green) of Process 2 between steps can be utilized to draft more extra tokens of TETRIS or to run more complex algorithms.
+
+Under the constraint of sequential pipelines in vLLM, we instead adopt an alternative performance metric that better captures the potential advantages of TETRIS in parallelized pipelines. We use the target efficiency rate (TER) defined as follows,
+
+$$
+T E R = \frac {\text {A c c e p t e d t o k e n s} + \text {B o u n s t o k e n s}}{\text {M a x p o s s i b l e n u m b e r o f t o k e n s i f a l l a c c e p t e d}}. \tag {2}
+$$
+
+As  $TER$  measures the efficiency of target model verifications and is unaffected by the drafting pro
+
+Table 3: Projected throughput  $\hat{\mathcal{G}}^{(TER)}$  improvement based on TER metric improvement, realizable under a parallelized speculative decoding pipeline.  
+
+<table><tr><td>Setting</td><td>Dataset</td><td>G↑</td><td>G(TER)↑</td></tr><tr><td rowspan="3">1</td><td>ShareGPT</td><td>3.50%</td><td>9.70%</td></tr><tr><td>Arena</td><td>5.17%</td><td>7.79%</td></tr><tr><td>Tough</td><td>4.85%</td><td>8.92%</td></tr><tr><td rowspan="3">2</td><td>ShareGPT</td><td>2.01%</td><td>11.70%</td></tr><tr><td>Arena</td><td>2.71%</td><td>11.17%</td></tr><tr><td>Tough</td><td>3.43%</td><td>11.91%</td></tr><tr><td rowspan="3">3</td><td>ShareGPT</td><td>3.93%</td><td>11.67%</td></tr><tr><td>Arena</td><td>5.15%</td><td>10.53%</td></tr><tr><td>Tough</td><td>5.25%</td><td>12.04%</td></tr></table>
+
+cess and TETRIS runtime, it provides an accurate indication of the net benefit of TETRIS. In Fig. 5, we demonstrate a case study for Setting 3 on Tough dataset: The improvement of TER is first calculated from the left figure, and is then used to compute the projected throughput  $\hat{\mathcal{G}}_k^{(TER)}$ , following
+
+$$
+\hat {\mathcal {G}} _ {k} ^ {(T E R)} = \mathcal {G} _ {\mathrm {S D}, k} \times \frac {(T E R _ {\mathrm {T E T R I S} , k} - T E R _ {\mathrm {S D} , k})}{T E R _ {\mathrm {S D} , k}},
+$$
+
+where  $k$  is the number of speculative tokens (i.e., draft window size) and  $\mathcal{G}$  represents throughput. Consequently, using TETRIS is projected (i.e., not realized in the current implementation) to achieve  $12.04\%$  improvement for this setting under parallelized pipeline. The full results are shown in Tab. 3 and the figures are shown in App. C.4.
+
+# 5.5 Ablation Study
+
+Robustness to Variations in Draft Quality. We artificially introduce additional variations in draft quality by mixing datasets of different difficulty levels. We create synthetic prompts designed for models to repeat lines from a poem named Sonnet. Since Sonnet is relatively easy for the small draft model, it achieves a high rate of successful verification by the target model. We then construct
+
+Table 4: Tetris improvement in throughput for ablation study of robustness to variations in draft quality.  
+
+<table><tr><td>Setting</td><td>Sonnet</td><td>Tough</td><td>Mix</td></tr><tr><td>1</td><td>2.46%</td><td>4.85%</td><td>4.12%</td></tr><tr><td>2</td><td>-0.81%</td><td>3.43%</td><td>3.48%</td></tr><tr><td>3</td><td>2.07%</td><td>5.25%</td><td>4.24%</td></tr></table>
+
+a new dataset, Mix, by randomly mixing Sonnet and a more challenging dataset, Tough, in equal proportions. As shown Tab. 4, the performance improvement of TETRIS over the best baseline suffers only a marginal or no decline, indicating its robustness to substantial variations in draft quality.
+
+Extension to Medusa. The Medusa model generates multiple subsequent draft tokens using a single forward pass (as opposed to autoregressive generation) through multiple decoding heads (Cai et al., 2024). Leveraging Medusa, it is possible to generate extra draft tokens for TETRIS at minimal marginal computational cost. We show in App. C.5 that integrating TETRIS to Medusa achieves a  $3.19\%$  improvement in total throughput.
+
+Other Ablations. We also include ablations on TETRIS's improvement in verification success rate (VSR) in App. C.6, and the effect of batch size on the performance in App. C.7.
+
+# 6 Conclusion and Future Work
+
+In this paper, we study the problem of optimizing batch speculative decoding to maximize throughput in multi-request settings, such as those faced by model service providers. To this end, we propose TETRIS, a novel method that efficiently selects optimal draft tokens for the LLM verification in log-linear time. We have theoretically shown that, in the absence of drafting time, TETRIS achieves optimal throughput both at each decoding step and globally under reasonable assumptions about token acceptance rates. Our empirical results further validate that TETRIS consistently outperforms standard speculative decoding and existing dynamic draft window selection methods, even when accounting for the extra time required for drafting extra tokens. These results highlight the potential of TETRIS to improve inference efficiency in real-world model service deployments. A key future direction is adapting TETRIS to tree decoding, a notable feature in recent advancements in speculative decod
+
+ing (Cai et al., 2024; Li et al., 2024a,b). Another interesting direction to explore is to design better draft token selection techniques. Insights can be derived from existing works on data selection (Lin et al., 2024; Zhou et al., 2023; Xu et al., 2024; Zhou et al., 2024b) that find important data points. Similarly, further research can investigate how to leverage the probability distribution between the draft model and the target model to improve the selection efficiency.
+
+# 7 Limitations
+
+In this paper, our empirical experiments only demonstrate results using the current sequential speculative decoding pipeline implemented on vLLM. That is, the target model stays idle while waiting for draft tokens from the draft model. Consequently, the performance improvement of TETRIS is heavily dependent on the trade-off between the additional runtime required to generate extra draft tokens and the gain in token acceptance achieved through TETRIS. Such trade-off limits the practical effectiveness of TETRIS, especially when a slow draft model is required. We anticipate that future implementations of a parallelized pipeline could potentially reveal greater speedups with TETRIS. However, we have not yet integrated such features into vLLM for testing in empirical experiments.
+
+# Acknowledgments
+
+This research is supported by the National Research Foundation (NRF), Prime Minister's Office, Singapore under its Campus for Research Excellence and Technological Enterprise (CREATE) programme. The Mens, Manus, and Machina (M3S) is an interdisciplinary research group (IRG) of the Singapore MIT Alliance for Research and Technology (SMART) centre.
+
+# References
+
+Sudhanshu Agrawal, Wonseok Jeon, and Mingu Lee. 2024. Adaedl: Early draft stopping for speculative decoding of large language models via an entropy-based lower bound on token acceptance probability. In NeurIPS Workshop on Efficient Natural Language and Signal Processing.  
+anon8231489123 Anon. 2023. Sharegpt dataset. https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered.  
+Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, Deming Chen, and Tri Dao. 2024.
+
+Medusa: Simple llm inference acceleration framework with multiple decoding heads. In Proc. ICML, pages 5209-5235.  
+Zhuoming Chen, Avner May, Ruslan Svirschevski, YuHsun Huang, Max Ryabinin, Zhihao Jia, and Beidi Chen. 2024. Sequoia: Scalable and robust speculative decoding. In Proc. NeurIPS, pages 129531-129563.  
+Yunfei Cheng, Aonan Zhang, Xuanyu Zhang, Chong Wang, and Yi Wang. 2024. Recurrent drafter for fast speculative decoding in large language models. arXiv:2403.09919.  
+DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and et al. 2024. Deepseek-v3 technical report. arXiv:2412.19437.  
+Fireworks AI. 2025. Fireworks ai: High-performance inference platform. https://fireworks.ai/. Accessed: 2025-02-10.  
+Yichao Fu, Siqi Zhu, Runlong Su, Aurick Qiao, Ion Stoica, and Hao Zhang. 2024. Efficient LLM scheduling by learning to rank. In Proc. NeurIPS, pages 59006-59029.  
+Zhengmian Hu, Tong Zheng, Vignesh Viswanathan, Ziyi Chen, Ryan A. Rossi, Yihan Wu, Dinesh Manocha, and Heng Huang. 2025. Towards optimal multi-draft speculative decoding. In Proc. ICLR.  
+Kaixuan Huang, Xudong Guo, and Mengdi Wang. 2024. Specdec++: Boosting speculative decoding via adaptive candidate lengths. arXiv:2405.19715.  
+Sehoon Kim, Karttikeya Mangalam, Suhong Moon, Jitendra Malik, Michael W. Mahoney, Amir Gholami, and Kurt Keutzer. 2023. Speculative decoding with big little decoder. In Proc. NeurIPS, pages 39236-39256.  
+Taehyeon Kim, Ananda Theertha Suresh, Kishore A Papineni, Michael Riley, Sanjiv Kumar, and Adrian Benton. 2024. Accelerating blockwise parallel language models with draft refinement. In Proc. NeurIPS, pages 34294-34321.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proc. SOSP.  
+Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023. Fast inference from transformers via speculative decoding. In Proc. ICML, pages 19274-19286.  
+Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2024a. EAGLE-2: Faster inference of language models with dynamic draft trees. In Proc. EMNLP, pages 7421-7432.
+
+Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang. 2024b. EAGLE: Speculative sampling requires rethinking feature uncertainty. In Proc. ICML, pages 28935-28948.  
+Xiaoqiang Lin, Xinyi Xu, Zhaoxuan Wu, See-Kiong Ng, and Bryan Kian Hsiang Low. 2024. Distributionally robust data valuation. In Proc. ICML.  
+Fangcheng Liu, Yehui Tang, Zhenhua Liu, Yunsheng Ni, Duyu Tang, Kai Han, and Yunhe Wang. 2024a. Kangaroo: Lossless self-speculative decoding for accelerating LLMs via double early exiting. In Proc. NeurIPS, pages 11946-11965.  
+Jiachen Liu, Zhiyu Wu, Jae-Won Chung, Fan Lai, Myungjin Lee, and Mosharaf Chowdhury. 2024b. Andes: Defining and Enhancing Quality-of-Experience in LLM-Based Text Streaming Services. arXiv:2404.16283.  
+Tianyu Liu, Yun Li, Qitan Lv, Kai Liu, Jianchen Zhu, and Winston Hu. 2024c. Parallel speculative decoding with adaptive draft length. arXiv:2408.11850.  
+Xiaoxuan Liu, Cade Daniel, Langxiang Hu, Woosuk Kwon, Zhuohan Li, Xiangxi Mo, Alvin Cheung, Zhijie Deng, Ion Stoica, and Hao Zhang. 2024d. Optimizing speculative decoding for serving large language models using goodput. arXiv:2406.14066.  
+Xiaoxuan Liu, Lanxiang Hu, Peter Bailis, Alvin Cheung, Zhijie Deng, Ion Stoica, and Hao Zhang. 2024e. Online speculative decoding. In Proc. ICML, pages 31131-31146.  
+Jonathan Mamou, Oren Pereg, Daniel Korat, Moshe Berchansky, Nadav Timor, Moshe Wasserblat, and Roy Schwartz. 2024. Dynamic speculation looks at head accelerates speculative decoding of large language models. arXiv:2405.04304.  
+Nestor Masleh, Loredana Fattorini, Raymond Perrault, Vanessa Parli, Anka Reuel, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Russell Wald, and Jack Clark. 2024. The AI index 2024 annual report. Technical report, AI Index Steering Committee, Institute for Human-Centered AI, Stanford University.  
+Xupeng Miao, Gabriele Oliaro, Zhihao Zhang, Xinhao Cheng, Zeyu Wang, Zhengxin Zhang, Rae Ying Yee Wong, Alan Zhu, Lijie Yang, Xiaoxiang Shi, Chunan Shi, Zhuoming Chen, Daiyaan Arfeen, Reyna Abhyankar, and Zhihao Jia. 2024. Specinfer: Accelerating large language model serving with tree-based speculative inference and verification. In Proc. ASPLOS, page 932-949.  
+Replicate. 2025. Replicate: Run AI with an API. https://replicate.com/. Accessed: 2025-02-10.  
+Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, Jérémy
+
+Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafori, Wenhan Xiong, Alexandre Defossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. 2024. Code llama: Open foundation models for code. arXiv:2308.12950.  
+Hyun Ryu and Eric Kim. 2024. Closer look at efficient inference methods: A survey of speculative decoding. arXiv:2411.13157.  
+Ying Sheng, Shiyi Cao, Dacheng Li, Banghua Zhu, Zhuohan Li, Danyang Zhuo, Joseph E. Gonzalez, and Ion Stoica. 2024. Fairness in serving large language models. In Proc. OSDI.  
+Benjamin Spector and Chris Re. 2023. Accelerating lIm inference with staged speculative decoding. arXiv:2308.04623.  
+Jikai Wang, Yi Su, Juntao Li, Qingrong Xia, Zi Ye, Xinyu Duan, Zhefeng Wang, and Min Zhang. 2024a. Opt-tree: Speculative decoding with adaptive draft tree structure. arXiv:2406.17276.  
+Siqi Wang, Hailong Yang, Xuezhu Wang, Tongxuan Liu, Pengbo Wang, Xuning Liang, Kejie Ma, Tianyu Feng, Xin You, Yongjun Bao, Yi Liu, Zhongzhi Luan, and Depei Qian. 2024b. Minions: Accelerating large language model inference with aggregated speculative execution. arXiv:2402.15678.  
+P. Wieder, J.M. Butler, W. Theilmann, and R. Yahyapour. 2011. Service Level Agreements for Cloud Computing. SpringerLink: Bucher. Springer New York.  
+Heming Xia, Zhe Yang, Qingxiu Dong, Peiyi Wang, Yongqi Li, Tao Ge, Tianyu Liu, Wenjie Li, and Zhifang Sui. 2024. Unlocking efficiency in large language model inference: A comprehensive survey of speculative decoding. In Proc. ACL Findings, pages 7655-7671.  
+Xinyi Xu, Zhaoxuan Wu, Rui Qiao, Arun Verma, Yao Shu, Jingtan Wang, Xinyuan Niu, Zhenfeng He, Jiangwei Chen, Zijian Zhou, Gregory Kang Ruey Lau, Hieu Dao, Lucas Agussurja, Rachael Hwee Ling Sim, Xiaogiang Lin, Wenyang Hu, Zhongxiang Dai, Pang Wei Koh, and Bryan Kian Hsiang Low. 2024. Data-centric AI in the age of large language models. In Proc. EMNLP Findings.  
+YAV-AI. 2024. Llm-tough-questions dataset: A collection of complex questions across 100 domains. https://huggingface.co/datasets/YAV-AI/11m-domain-specific-tough-questions.  
+Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. 2024. A survey on multimodal large language models. National Science Review, 11(12).  
+Fanlong Zeng, Wensheng Gan, Yongheng Wang, Ning Liu, and Philip S. Yu. 2023. Large language models for robotics: A survey. arXiv:2311.07226.
+
+Jun Zhang, Jue Wang, Huan Li, Lidan Shou, Ke Chen, Gang Chen, and Sharad Mehrotra. 2024. Draft & verify: Lossless large language model acceleration via self-speculative decoding. In Proc. ACL, pages 11263-11282.  
+Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. 2023. A survey of large language models. arXiv:2303.18223.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. In Proc. NeurIPS (Datasets and Benchmarks Track), pages 46595-46623.  
+Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, Clark Barrett, and Ying Sheng. 2024. Sglang: Efficient execution of structured language model programs. arXiv:2312.07104.  
+Yongchao Zhou, Kaifeng Lyu, Ankit Singh Rawat, Aditya Krishna Menon, Afshin Rostamizadeh, Sanjiv Kumar, Jean-François Kagy, and Rishabh Agarwal. 2024a. Distillspec: Improving speculative decoding via knowledge distillation. In Proc. ICLR.  
+Zijian Zhou, Xiaoqiang Lin, Xinyi Xu, Alok Prakash, Daniela Rus, and Bryan Kian Hsiang Low. 2024b. Detail: Task demonstration attribution for interpretable in-context learning. In Proc. NeurIPS.  
+Zijian Zhou, Xinyi Xu, Rachael Hwee Ling Sim, Chuan Sheng Foo, and Kian Hsiang Low. 2023. Probably approximate shapley fairness with applications in machine learning. In Proc. AAAI.
+
+# A Leftover Proofs
+
+# A.1 Proof of Theorem 1.
+
+Theorem 1 (Per-step Optimality of TETRIS). In the absence of drafting time, given the true acceptance rate  $\alpha_{i,j}$  of each draft token  $(i,j)$ , Alg. 1 produces the optimal per-step throughput defined in Sec. 3.
+
+Proof. We prove it by contradiction. Let the selection of Alg. 1 be  $\mathcal{D}^*$ . Suppose the actual optimal solution is  $\mathcal{D}' \neq \mathcal{D}^*$ . Let  $\tilde{\mathcal{D}} = \mathcal{D}' \cap \mathcal{D}^*$  be the overlapping tokens selected by both Alg. 1 and the actual optimal solution. Note that the tokens in each row are selected sequentially (i.e., tokens cannot be skipped in a row).
+
+Case 1: TETRIS selects some token  $d \in \mathcal{D}^* \setminus \tilde{\mathcal{D}}$  before selecting  $\tilde{\mathcal{D}}$ . In this case, the  $\mathbb{E}[1]$  of the token  $d$  is higher than the token last selected in  $\tilde{\mathcal{D}}$ . This suggests that the optimal selection should include  $d$ . However, it can be observed that  $d \notin D'$  since otherwise  $d \in \tilde{\mathcal{D}}$ . This contradicts the fact that  $\mathcal{D}'$  is optimal.
+
+Case 2: TETRIS selects  $\tilde{\mathcal{D}}$  first before selecting other tokens. Since Alg. 1 always selects the token with the highest  $\mathbb{E}[\mathbf{1}]$ , every element in  $\mathcal{D}^* \setminus \tilde{\mathcal{D}}$  is larger than or equal to that in  $\mathcal{D}' \setminus \tilde{\mathcal{D}}$ . As such, we have  $\mathbb{E}[\sum_{p \in \mathcal{D}'} \mathbf{1}_p] \leq \mathbb{E}[\sum_{p \in \mathcal{D}^*} \mathbf{1}_p]$ . However, this contradicts the fact that  $\mathcal{D}'$  is optimal as Alg. 1 has a higher number of accepted tokens. Therefore, Alg. 1 must be optimal.
+
+Combining the two cases finishes the proof.
+
+![](images/6c018f466336f0e3c257b58386ae23d94d5f7933504af766817f5f3c8d5f2f34.jpg)
+
+# A.2 Running Time of TETRIS
+
+Lemma 1. Alg. 1 achieves a time complexity of  $\mathcal{O}(C\log N)$
+
+Proof. Note that Alg. 1 maintains a heap. The heap is initialized with  $N$  items. Since only  $C$  pairs are selected, there are  $2C$  operations of enqueue and deque. Following classic results of heap operation, each enqueue of deque operation requires  $\mathcal{O}(\log C)$  time. As such, the overall time complexity of TETRIS is  $\mathcal{O}(C\log N)$ .
+
+# A.3 Proof of Theorem 2.
+
+Theorem 2 (Global Optimality of TETRIS under Assumption). Under Assumption 1, in the absence of drafting time, TETRIS searches for the optimal  $\mathcal{G}$  under the same capacity. Moreover, if  $\alpha_{1} = \alpha_{2} = \dots = \alpha_{N}$ , TETRIS has the same  $\mathcal{G}$  as standard batched speculative decoding.
+
+Proof. The proof of global optimality is established on Theorem 1. Since all tokens in each row have the same acceptance rate. After each step, we have the same distribution of 1 no matter what tokens are accepted, where 1 is the indicator variable of whether the token is accepted. As such, at each step, performing TETRIS is per-step optimal by Theorem 1. Moreover, since the state at each step is identical, a per-step optimal strategy is also globally optimal.  $\square$
+
+# B Additional Related Work and Discussion
+
+# B.1 Acceptance Rate
+
+The acceptance rate plays a vital role in the effectiveness of speculative decoding. A higher acceptance rate should be paired with a larger draft window size  $k$  to achieve optimal speedup. In the typical rejection sampling setting of speculative decoding, the acceptance of draft tokens depends on the probability distributions of both the draft and target models. When the probability distribution of the draft model,  $p_{S}(\cdot)$ , closely approximates that of the target model,  $p_{\mathcal{M}}(\cdot)$ , a higher number of tokens are accepted on average. Since the value of  $k$  is chosen in the drafting process, we do not have access to  $p_{\mathcal{M}}(\cdot)$  and have to rely on  $p_{S}(\cdot)$  to estimate the acceptance rate.
+
+Leviathan et al. (2023) derive that the expected acceptance rate is 1 minus the KL divergence between the token distributions of the draft and the target model. Hence, the acceptance rates for all draft tokens are considered constant. Liu et al. (2024d) assume uniform token acceptance behavior across diverse requests. It proposes SmartSpec, which calculates the average acceptance rate from past generation steps. Li et al.
+
+(2024a) and Wang et al. (2024a) utilize the draft model's confidence score (i.e., the output probability of each token) to estimate the acceptance rate. Chen et al. (2024) make the positional acceptance assumption so that the acceptance rate of tokens is determined solely by their position (i.e., number of tokens away) relative to the already accepted tokens. Agrawal et al. (2024) instead consider an approximate lower bound on the expected acceptance rate of a token that depends on the entropy of prediction probabilities of the draft model  $p_{S}(\cdot)$ . Noting the acceptability of diverse tokens, especially in the real world with a high value of temperature hyperparameter, Medusa proposes to use both a hard threshold and an entropy-dependent threshold as a criterion to accept draft tokens (Cai et al., 2024). In Medusa, the first token is always accepted using greedy decoding to ensure at least one token is generated in each step.
+
+# B.2 Greedy Algorithms for Speculative Decoding
+
+Several existing approaches in speculative decoding employ greedy algorithms to improve efficiency, though they operate at different levels and with distinct objectives. EAGLE-2 (Li et al., 2024a) and MDSD (Hu et al., 2025), for instance, utilize greedy selection strategies within single-request, multi-draft settings. EAGLE-2 achieves speedups by generating multiple branches in a draft tree, which the top candidates are then greedily selected for verification. It requires substantial GPU resources for each request and may not translate to improved throughput in batch inference due to limited total GPU capacity. MDSD, on the other hand, focuses on increasing the acceptance rate in multi-draft speculative decoding through greedy draft sampling, an objective orthogonal to our work. Our TETRIS distinguishes itself by applying a greedy algorithm at the batch level, optimizing GPU resource utilization across multiple requests rather than within a single draft tree. Due to the simplicity of greedy approaches in implementation, our method can be readily adapted to major inference frameworks (e.g., vLLM (Kwon et al., 2023)) with minimal empirical overhead. Our paper also demonstrates that TETRIS is complementary to existing speculative decoding frameworks, such as Medusa (Cai et al., 2024) (see Sec. 5.5 and App. C.5), by introducing a novel axis of optimization not explored by prior works.
+
+# B.3 Assumptions for Analysis
+
+Our theoretical analysis of the global optimality of TETRIS in Theorem 2 is conditioned on Assumption 1. Similar assumptions are commonly made in the literature (Leviathan et al., 2023; Liu et al., 2024d) to analyze the length of generated outputs. Specifically, the pioneering work of Leviathan et al. (2023) assumes a constant acceptance rate  $\alpha$  and applies a capped geometric distribution to derive the expected number of generated tokens, under a single-request setting. For a single request, this assumption is equivalent to our Assumption 1. A natural extension of the constant acceptance rate assumption to the batch inference setting would be assuming the identical acceptance rate across all requests. However, we highlight that our Assumption 1 is weaker than assuming the same constant acceptance rate across all requests, as it allows the acceptance rate to vary across requests. Moreover, even when Assumption 1 is violated, our method still demonstrates strong empirical performance as shown in Sec. 5 of the main paper.
+
+![](images/eb451e62d05522ab2970d70b71b31cb524ea72ceeeb37140458311afbc58d6f3.jpg)  
+(a) Base  $k = 3$
+
+![](images/4a10b9596333cb9232d7f718c232aa46ec93369a1c9db0cee80a2808557cc485.jpg)  
+(b) Base  $k = 2$  
+Figure 7: Additional plots for the change in VSR as the number of extra draft tokens increases, supplementary to Fig. 3 where base draft length  $k = 4$ . Here, we provide results for  $k = 1,2,3$
+
+![](images/385a371f610c658206c4390b7f134a44d7b6737253caee5a03114fd48d3fbebd.jpg)  
+(c) Base  $k = 1$
+
+![](images/8d561452b6e86947d5d1e26be0bce9953717503d5947337a25182f356559f338.jpg)
+
+![](images/09f29231c4400c61e23b3e1b107aa9eacefe3a92181dc086cd7ac9f7c1c2f6b8.jpg)
+
+![](images/d86d11e534ae10e4c58f7183d682f38b661f09be8311cc9a665258c67d0c3b2e.jpg)
+
+![](images/26cc5914d8194ff44308b0332c698117d761156db192e470d7bedd045dd6912b.jpg)
+
+![](images/24931425d9c6b8909066938aa7246cf8d94b46dcb4ef220d99f4a9e9e42bacfc.jpg)
+
+![](images/683e01b9bf6033e63d63180180ec32560bea2fbbfecbc2006b47ee8a35fd3c85.jpg)
+
+![](images/e2adff4d83cc9fcac43d7645c8f0e94b4972d299d490f1b2643239fe77fdc7a4.jpg)  
+(g)  $\uparrow 3.51\% ,\Delta 4.52\%$
+
+![](images/6e1319008fbaaf6e24b614ef34ad0004f0acc052c0a30afc9f9c80a01b337ea9.jpg)  
+(h)  $\uparrow 6.13\% ,\Delta 6.13\%$
+
+![](images/045ea4c9fe0a664dc67b77b1bd608ba01acb3a342a087ccbbd9714da516766e6.jpg)  
+(i)  $\uparrow 4.49\% ,\Delta 4.68\%$  
+Figure 8: Mean end-to-end latency comparison for various methods across experimental settings.  $\uparrow$  indicates the improvement from best baseline method.  $\Delta$  indicates the maximum gap between TETRIS and standard SD. The reported numbers reflect the mean and standard deviation over 3 independent trials.
+
+# C Additional Results
+
+# C.1 Dataset License
+
+ShareGPT (Anon, 2023): Apache license 2.0; Arena (Zheng et al., 2023): CC; Domain-specific Tough Questions (YAV-AI, 2024): MIT.
+
+# C.2 Additional Plots for Effect of Extra Draft Tokens
+
+Fig. 3 demonstrates the benefit of adding extra draft tokens (from 0 to 9 extra tokens). The base draft length  $k$  serves only as a reference baseline and its value does not significantly impact the observed trends in Fig. 3. We set  $k = 4$  based on established reasonable ranges used in existing studies (Leviathan et al., 2023; Zhang et al., 2024; Wang et al., 2024b). This choice balances the trade-off between extra compute required and additional throughput gained.
+
+For completeness, we conducted additional experiments with  $k = 1,2,3$  and present the results in Fig. 7. The observed trends are consistent across different values of  $k$ : VSR generally increases with extra tokens, demonstrating the effectiveness of having extra draft tokens in our TETRIS.
+
+# C.3 Plots for End-to-end Latency
+
+We provide an extended discussion on the improvement of end-of-end latency from Sec. 5.3. In Fig. 8, we show the plots for the end-to-end latency over all speculative decoding configurations and settings used in the paper. Tetris consistently outperforms the existing baselines and achieves up to  $6.13\%$  improvement over the best baseline and up to  $9.32\%$  maximum gap over standard SD. Therefore, Tetris has demonstrated to effectively reduce end-to-end request latency, which is also essential for enhancing the user experience with LLM inference service providers.
+
+![](images/cee547eb1c0e366ffa7447f6be89047e2975850339701ff951803cc1494d3383.jpg)
+
+![](images/27df194c935debc847b2bf1d795786fdc5a9651919f20828503520924d0bdda9.jpg)
+
+![](images/3d29294e1ed4eeee8eab85770f73e034c92bbe6971ef177cb705a3662ab98f77.jpg)
+
+![](images/3c05344cd8615d19e7f16cb9b372abea6aab0d9443d3c119189a66f452cf1bd0.jpg)
+
+![](images/934c1cb686c0a6a2fd187a977f73b6fcd2ae6cd2d98e310247fea6b74bfb475b.jpg)
+
+![](images/77258bb2b23e0971fb12d770e349b19a42910b48de61d53a93f7df1866a5c620.jpg)
+
+![](images/9d7ce2d7e17a9612ae085b3fecd19be80a865c80c856c61c2b646110d7e14a52.jpg)  
+(g)  $\hat{\mathcal{G}}^{(TER)}\uparrow 11.67\%$
+
+![](images/c4b220e191d9753e20716ecc0b7fc9012a510b63d7399be8d0d90d267ff8d1c5.jpg)  
+(h)  $\hat{\mathcal{G}}^{(TER)}\uparrow 10.53\%$
+
+![](images/00d2777b611af9b34963d426f08daddabfe6ed6ea209daead978d8955ef64da5.jpg)  
+(i)  $\hat{\mathcal{G}}^{(TER)}\uparrow 12.04\%$  
+Figure 9: Mean projected throughput  $\hat{\mathcal{G}}^{(TER)}$  comparison for various methods across experimental settings.  $\uparrow$  indicates the improvement from the best baseline method. The reported numbers reflect the mean over 3 independent trials.
+
+# C.4 Plots for Projected Improvement based on TER
+
+Complementary to Tab. 3, which contains the numerical results for the projected improvement of TETRIS in terms of the projected throughput  $\hat{\mathcal{G}}^{(TER)}$ , we also show the plots in Fig. 9 to visually illustrate the effectiveness of our method. The dotted lines for TETRIS (drawn in blue, orange, and green) represent the projected throughput calculated based on the throughput of the standard SD and also the TETRIS's improvement in terms of target efficiency rate (TER, as defined in Eq. (2)). We note that these improvement numbers are theoretically computed and are not yet realizable in empirical settings due to the lack of parallelized pipeline implementations of speculative decoding in vLLM.
+
+# C.5 Extension to Medusa
+
+We evaluate the top-1 proposal version (i.e., only draft the most likely token for each position) of Medusa and its integration with TETRIS. As the Medusa model outputs multiple subsequent tokens in a single forward pass, $^{3}$  we leverage this feature to produce extra draft tokens for TETRIS. We show the results in Tab. 5. We achieved a throughput improvement of  $3.19\%$  as compared to the baseline Medusa. The development of such multi-token prediction models, including models like EAGLE (Li et al., 2024b) and DeepSeek-V3 (DeepSeek-AI et al., 2024) presents further potential for TETRIS to achieve greater speedups. Other improvements in engineering, including using tree-decoding and using a larger target model also potentially further boost the speedup.
+
+Table 5: Mean total throughput (± standard deviation) for the ablation study of TETRIS extension to Medusa over three independent trials. The integration of TETRIS with Medusa further improves the total throughput.  
+
+<table><tr><td>No. Speculative Tokens</td><td>TETRIS (extra=1)</td><td>TETRIS (extra=2)</td><td>TETRIS (extra=3)</td><td>Baseline Medusa</td></tr><tr><td>1</td><td>591.26±0.46</td><td>590.83±8.30</td><td>586.47±3.66</td><td>572.97±1.79</td></tr><tr><td>2</td><td>571.05±0.80</td><td>568.82±6.52</td><td>571.95±1.06</td><td>563.94±2.95</td></tr><tr><td>Best</td><td>591.26</td><td>590.83</td><td>586.47</td><td>572.97</td></tr></table>
+
+![](images/1a7d89e2d76acaf74deb5bbac4c5f396c1e93b97caf6831fcef0d7445949f322.jpg)
+
+![](images/4186574d880a23da9d99c8bed497e4044cc5bcf967d2ec4754d6fe1b40b7b4ec.jpg)
+
+![](images/802e622eb08c8274e0ab4e4d7a918f2fb32e85daebbfe2463e3dfe4760de03f9.jpg)
+
+![](images/4fe8050b2e20073e91211455d1b98da095d04bc1b8d0f77f1d456374af31551f.jpg)
+
+![](images/90dea63d3c9f0693db22da60141f19e9e4e2291822a171e011c93f1d09bee73b.jpg)
+
+![](images/a568f2ad059b58814601346692c69c7d631a377528cb9e6a17dcf729b22b7aa8.jpg)
+
+![](images/fd4efe32a3e50d197fc065be45e712d536ca451002931cc23d79c1545831ffc6.jpg)  
+Figure 10: The verification success rate comparison for various methods across experimental settings.  $\Delta$  indicates the maximum gap between TETRIS and standard SD. The reported numbers reflect the mean over 3 independent trials.
+
+![](images/d6a10bcf999f97128847339c1f70c12b5ec44501f7bab6b1ead69b9da62457df.jpg)
+
+![](images/3fde4a2418233dcecdebf040cdd189599be2155c640cf24bda3a1e440fe23880.jpg)
+
+# C.6 Improvement in Verification Success Rate
+
+As an ablation study, we also illustrate the improvement of TETRIS in terms of VSR (as defined in Eq. (1)), which is an important measure of the effectiveness of speculative decoding. We show in Fig. 10 that the maximum gap between TETRIS and standard SD in terms of VSR is consistently above  $20\%$  and reaching over  $30\%$  in some instances. This validates the significant effect of TETRIS in selecting draft tokens that are most likely to be accepted by the target model without exceeding the system capacity of the server. However, it is worth noting that this improvement in VSR does not translate entirely to an increment in total throughput or a reduction in end-to-end latency. This is because the throughput in practice also depends on the running time of the draft model (especially when the speculative decoding pipeline is sequential, as discussed in Sec. 5.4), and VSR does not account for the generation of the bonus token (which takes up a portion of the generated tokens).
+
+# C.7 The Effect of Batch Size on TETRIS Performance
+
+Theoretically speaking, a larger batch size creates more possible combinations for draft token selection by Tetris. Therefore, Tetris is likely to perform better in a speculative decoding server that processes a
+
+![](images/8b76770decb6b9ccf93dbb540ddf9af64cd855fab3ae2a7f28eebc8dade14f76.jpg)
+
+![](images/183173e2f4a25509f0168ad7cb46d50e563f822ff7eed655b450e9e1f83873f4.jpg)
+
+![](images/d50c3577389fbb0a5dfe0143dddcb1876944012529c9bf1ba02a07c912b76f23.jpg)  
+Figure 11: The change in the verification success rate (VSR) and target efficiency rate (TER) when we vary the batch size (BS) from 64 to 32 and 16. The reported numbers reflect the mean over 3 independent trials.
+
+![](images/165733a5e907cbedad379415e12a68d2352c6524a93748a6d555a7b72939d078.jpg)
+
+larger batch of requests concurrently. In Fig. 11, we show a visual illustration of the verification success rate (VSR) and target efficiency rate (TER) (as defined in Eq. (1) and Eq. (2), respectively).
+
+In setting 2 (draft model: Llama-1B-Instruct-FP8, target model: Llama-70B-Instruct), we observe a significant increase in VSR and TER when the batch size is increased to 64. However, batch sizes of 16 and 32 have similar VSR and TER values.
+
+In setting 3 (draft model: Llama-1B-Instruct-FP8, target model: Llama-405B-Instruct-FP8), we do not observe a significant change in VSR and TER, suggesting that the way that the batch size affects performance is highly dependent on the specific draft-target combination, too.
+
+Overall, we expect a more significant improvement in the performance of adopting TETRIS by LLM inference service providers with larger capacities to handle a larger number of concurrent requests.
+
+# D Broader Impacts
+
+While this research work is primarily foundational, focusing on computational performance, the resulting increase in inference speed and efficiency of Large Language Models (LLMs) could indirectly contribute to certain societal risks associated with LLMs. Making LLM inference faster and cheaper lowers the barrier to deploying these models at scale. Consequently, this could potentially accelerate the proliferation of LLM-generated content, increasing the risks of misuse such as the large-scale generation of disinformation, spam, or fake online personas, if the underlying models are deployed without adequate safeguards.
+
+Mitigation strategies depend on responsible deployment. Developers using TETRIS should employ robust safety measures, bias detection, and content filtering for the LLMs being served. Importantly, the efficiency gains from TETRIS could be leveraged positively to make computational overhead for safety checks, alignment techniques, or bias mitigation measures more feasible during inference.

2025/TETRIS_ Optimal Draft Token Selection for Batch Speculative Decoding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f11ec892a1a76ee7dfaee74acd8408be6e54431d284625571490f0cb166d23cb
+size 1077789

2025/THOR-MoE_ Hierarchical Task-Guided and Context-Responsive Routing for Neural Machine Translation/f6520cf3-9239-4e76-8a25-131bf94b31d7_content_list.json

@@ -0,0 +1,2083 @@
+[
+    {
+        "type": "text",
+        "text": "THOR-MoE: Hierarchical Task-Guided and Context-Responsive Routing for Neural Machine Translation",
+        "text_level": 1,
+        "bbox": [
+            117,
+            89,
+            878,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yunlong Liang, Fandong Meng*, Jie Zhou",
+        "bbox": [
+            310,
+            158,
+            685,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Pattern Recognition Center, WeChat AI, Tencent Inc",
+        "bbox": [
+            282,
+            175,
+            712,
+            190
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{yunlonliang, fandongmeng, withtomzhou}@tencent.com",
+        "bbox": [
+            250,
+            192,
+            747,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            342,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The sparse Mixture-of-Experts (MoE) has achieved significant progress for neural machine translation (NMT). However, there exist two limitations in current MoE solutions which may lead to sub-optimal performance: 1) they directly use the task knowledge of NMT into MoE (e.g., domain/linguistics-specific knowledge), which are generally unavailable at practical application and neglect the naturally grouped domain/linguistic properties; 2) the expert selection only depends on the localized token representation without considering the context, which fully grasps the state of each token in a global view. To address the above limitations, we propose THOR-MoE via arming the MoE with hierarchical task-guided and context-responsive routing policies. Specifically, it 1) firstly predicts the domain/language label and then extracts mixed domain/language representation to allocate task-level experts in a hierarchical manner; 2) injects the context information to enhance the token routing from the pre-selected task-level experts set, which can help each token to be accurately routed to more specialized and suitable experts. Extensive experiments on multi-domain translation and multilingual translation benchmarks with different architectures consistently demonstrate the superior performance of THOR-MoE. Additionally, the THOR-MoE operates as a plug-and-play module compatible with existing Top- $k$  (Shazeer et al., 2017a) and Top $p$  (Huang et al., 2024) routing schemes, ensuring broad applicability across diverse MoE architectures. For instance, compared with vanilla Top- $p$  (Huang et al., 2024) routing, the context-aware manner can achieve an average improvement of 0.75 BLEU with less than  $22\\%$  activated parameters on multi-domain translation tasks.",
+        "bbox": [
+            142,
+            311,
+            460,
+            881
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            510,
+            260,
+            653,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The rapid advancement of neural machine translation (NMT; Zhang et al. (2020); Fan et al. (2020)) has been significantly propelled by sparse Mixture-of-Experts (MoE) architectures (Shazeer et al., 2017a; Lepikhin et al., 2021a; Fedus et al., 2022), which increase model capacity while maintaining computational efficiency through conditional computation. In the literature, existing work can be roughly classified into two categories: (1) some studies mainly focus on how to effectively incorporate task information (e.g., domain/linguistics-specific knowledge) into MoE models designing task-specific modules (Pham et al., 2023; Jiang et al., 2024); (2) another line of work concentrates on how to improve the training and inference efficiency via reducing the activated experts in MoE (Jawahar et al., 2023a; Elbayad et al., 2023; Wang et al., 2024; Zeng et al., 2024).",
+        "bbox": [
+            507,
+            291,
+            885,
+            580
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Specifically, in (1), Kudugunta et al. (2021) routes input examples to different experts based on translation language representations. Li et al. (2023c) explores multiple language-group-specific routers to incorporate language group knowledge into models. Zhao et al. (2024b) takes one step further by designing hierarchical language-guided token routing. Although intuitive and effective, they heavily rely on explicit task-specific knowledge (e.g., domain or language labels), which is often unavailable in real-world scenarios. This reliance not only restricts generalization but also overlooks the intrinsic hierarchical grouping of domain and linguistic properties inherent in multilingual or multi-domain translation tasks. In (2), Zhao et al. (2024a) distills the knowledge of unselected experts to the selected one and thus reduces the number of experts without impacting performance; Li et al. (2023a) and Huang et al. (2024) decrease the number of experts by designing a threshold based on expert probability distribution; Zhao et al.",
+        "bbox": [
+            507,
+            583,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* Corresponding author.",
+        "bbox": [
+            136,
+            906,
+            289,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "21433",
+        "bbox": [
+            473,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 21433-21445 July 27 - August 1, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            82,
+            945,
+            914,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(2024b) adapts the number of experts by the difficulty of language. Generally, the global context grasp the overall situation and might know whether each token is difficult or not (Gloeckle et al., 2024). However, previous routing strategies select experts solely based on localized token-level representations, neglecting the broader contextual dependencies that govern token interactions. This myopic view limits the model's ability to allocate experts optimally in a globally coherent manner.",
+        "bbox": [
+            110,
+            84,
+            489,
+            244
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, to address the above issues, we propose an innovative hierarchical context-responsive routing (THOR-MoE) framework for neural machine translation. Specifically, the THOR-MoE first predicts which domain/language the input belongs to and then extracts the mixed task representation to allocate experts in a hierarchical manner. It is not only flexible in practice but also effectively infuses the domain knowledge into the task-level routing via encouraging experts to specialize in certain domains/languages. Furthermore, before each token routing, the THOR-MoE injects the context information to help accurately assign the experts from the pre-selected task-level experts set for each token in a global perspective. This enables the model to adaptively capture evolving contextual dependencies, such as long-range syntactic structures or discourse coherence. Crucially, the THOR-MoE operates as a plug-and-play module compatible with existing Top- $k$  (Shazeer et al., 2017a) and Top- $p$  (Huang et al., 2024) routing schemes, ensuring broad applicability across diverse MoE architectures.",
+        "bbox": [
+            115,
+            247,
+            489,
+            613
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We validate our proposed THOR-MoE framework on the commonly-used multi-domain NMT benchmark (Aharoni and Goldberg, 2020) and multilingual NMT benchmark (Zhang et al., 2020), which contains 5 domains and 16 languages, respectively. Extensive experiments show that the hierarchical context-responsive routing is pivotal for unlocking the full potential of MoE systems.",
+        "bbox": [
+            112,
+            617,
+            489,
+            745
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, our main contributions are:",
+        "bbox": [
+            131,
+            747,
+            433,
+            762
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose two components to expand MoE: (i) a hierarchical design to enable language- and domain-specific expert routing; (ii) enabling the use of context information during expert selection, which is important for translation.",
+            "- Extensive experiments on both multi-domain and multilingual translation consistently show the ef"
+        ],
+        "bbox": [
+            112,
+            769,
+            489,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "fectiveness and generalization of THOR-MoE. For example, our model achieves consistent improvements of an average improvement of 0.75 BLEU with less than  $22\\%$  activated parameters over vanilla Top-p routing in multi-domain translation tasks.",
+        "bbox": [
+            521,
+            84,
+            884,
+            180
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Background",
+        "text_level": 1,
+        "bbox": [
+            507,
+            195,
+            650,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Neural Machine Translation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            223,
+            779,
+            237
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given an input sentence in the source language  $X = \\{x_{i}\\}_{i=1}^{|X|}$ , the goal of the model is to produce its translation in the target language  $Y = \\{y_{i}\\}_{i=1}^{|Y|}$ . The conditional distribution of the model is:",
+        "bbox": [
+            507,
+            244,
+            882,
+            313
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N M T}} = - \\sum_ {t = 1} ^ {| Y |} \\log \\left(p \\left(y _ {t} \\mid X, y _ {1: t - 1}\\right)\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            325,
+            882,
+            370
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $y_{1:t - 1}$  is the partial translation.",
+        "bbox": [
+            507,
+            384,
+            789,
+            400
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Sparse Mixture-of-Experts",
+        "text_level": 1,
+        "bbox": [
+            507,
+            414,
+            766,
+            429
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The vanilla MoE model can be seen the variant of Transformer model via replacing the feedforward (FFN) sub-block of the transformer block with MoE layer, in which per token selects fixed number of experts (Lepikhin et al., 2021b). In each MoE layer, there are  $N$  experts, denoted as  $\\mathcal{E} = \\{e_1,e_2,\\dots ,e_N\\}$ . An input  $\\mathbf{x}$  is dispatched to these experts, and the output of the MoE layer is computed as the weighted average of the outputs from the experts:",
+        "bbox": [
+            507,
+            435,
+            882,
+            596
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {M o E} (\\mathbf {x}) = \\sum_ {i = 1} ^ {N} g _ {i} (\\mathbf {x}) * e _ {i} (\\mathbf {x}), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            612,
+            882,
+            655
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $g_{*}(\\mathbf{x})$  is computed by a router that predicts the contribution of each expert to the final output. Given the computing efficiency, the MoE assigns each token to limited experts (e.g., 1 or 2).",
+        "bbox": [
+            507,
+            664,
+            884,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To derive  $g_{*}(\\mathbf{x})$ , we generally compute the probability  $\\mathbf{P}$  of each expert being selected for the input  $\\mathbf{x}$  as follows:",
+        "bbox": [
+            507,
+            728,
+            882,
+            776
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {P} = \\operatorname {S o f t m a x} \\left(\\mathbf {W} _ {\\mathbf {r}} \\cdot \\mathbf {x} ^ {T}\\right), \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            600,
+            791,
+            882,
+            810
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{W}_{\\mathbf{r}} \\in N \\times d$  serves as a learnable parameter and  $d$  denotes the dimension of the input  $\\mathbf{x}$ . The vector  $\\mathbf{P}$ , with a size of  $N$ , encapsulates the probabilities associated with the selection of each expert. And  $P_{i}$  indicates the likelihood of choosing the  $i^{th}$  expert  $e_{i}$  to calculate the input  $\\mathbf{x}$ .",
+        "bbox": [
+            507,
+            824,
+            882,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Note that the task generally denotes the domain/language-specific knowledge in this work.",
+        "bbox": [
+            112,
+            894,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "21434",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/e9a5792ab50d7a0f8f2cc6098e6ceaa16d3138cde790dd1b2357ac7a8ed36b75.jpg",
+        "image_caption": [
+            "(1) Vanilla Token Routing;"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            84,
+            440,
+            192
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/df4bed66f43b928c36629b6683b81ab72a184188338292fdadc967a04d0eaa3d.jpg",
+        "image_caption": [
+            "(3) Context-Responsive Token Routing."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            452,
+            82,
+            868,
+            193
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/8324a558aad599412debbe3de019202ac74c186c8132e29dec2106846a2765ee.jpg",
+        "image_caption": [
+            "(2) Hierarchical Task-Guided Routing;",
+            "Figure 1: The overview of the proposed THOR-MoE. 1) vanilla token routing; 2) hierarchical task-guided routing; 3) context-responsive token routing. The 'Emb.' and 'Rep.' denotes the embedding and representation, respectively. The hierarchical manner denotes that the task-guided routing firstly selects different task-specific experts sets for different queries (e.g.,  $\\mathcal{E}_t$  for query 1). Then the context-responsive token routing assigns experts from  $S^t$  for each token in query 1."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            210,
+            875,
+            349
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2.1 Top-  $k$  Routing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            470,
+            290,
+            487
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Top- $k$  routing (Shazeer et al., 2017b; Lepikhin et al., 2021b; Zuo et al., 2022) entails the selection of the top  $k$  experts, which correspond to the  $k$  highest probabilities within the vector  $\\mathbf{P}$ . Then, the probabilities of the chosen experts are re-scaled through normalization, while the probabilities of the non-selected experts are set to zero, signifying their deactivation. The resultant computation for  $g_{*}(\\mathbf{x})$  proceeds as below:",
+        "bbox": [
+            112,
+            489,
+            487,
+            634
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ng _ {i} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} \\frac {P _ {i}}{\\sum_ {j \\in \\operatorname {T o p K} (\\mathbf {P})} P _ {j}}, & i \\in \\operatorname {T o p K} (\\mathbf {P}) \\\\ 0, & i \\notin \\operatorname {T o p K} (\\mathbf {P}) \\end{array} \\right. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            134,
+            645,
+            487,
+            689
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\operatorname{TopK}(\\mathbf{P})$  returns the indices of the k largest elements in  $\\mathbf{P}$ .",
+        "bbox": [
+            112,
+            700,
+            485,
+            731
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2.2 Top-  $p$  Routing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            741,
+            290,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Top- $p$  routing is initially proposed by (Li et al., 2023a; Huang et al., 2024), which aims to overcome the drawback of Top- $k$  activating the fixed number of experts overlooking the variability in difficulty across different inputs. We refer to readers to Huang et al. (2024) for more details.",
+        "bbox": [
+            112,
+            760,
+            487,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Formally, Top-  $p$  first sorts the elements in  $\\mathbf{P}$  from highest to lowest, returning in a sorted index list  $I$ . Subsequently, we identify the minimal subset of experts for which the aggregate probability",
+        "bbox": [
+            112,
+            856,
+            487,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "surpasses the threshold  $p$ . The count of selected experts, denoted as  $t$ , is then determined by:",
+        "bbox": [
+            507,
+            470,
+            880,
+            502
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nt = \\underset {k \\in \\{1 \\dots , N \\}} {\\arg \\min } \\sum_ {j <   = k} P _ {i, j} \\geq p, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            512,
+            882,
+            546
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $p$  is the threshold that controls how confident the model should be when stopping adding more experts.  $p$  is a hyper-parameter whose range is from 0 to 1. The higher the  $p$  is, the more experts will be activated.",
+        "bbox": [
+            507,
+            556,
+            880,
+            634
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In dynamic routing mechanism, the calculation of  $g_{*}(\\mathbf{x})$  is:",
+        "bbox": [
+            507,
+            637,
+            880,
+            669
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ng _ {i} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} P _ {i} & e _ {i} \\in S \\\\ 0, & e _ {i} \\notin S \\end{array} \\right. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            606,
+            678,
+            882,
+            721
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $S$  is the set of selected experts controlled by  $t$  in Equation 5:",
+        "bbox": [
+            507,
+            728,
+            880,
+            760
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nS = \\left\\{e _ {I _ {1}}, e _ {I _ {2}} \\dots e _ {I _ {t}} \\right\\}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            616,
+            771,
+            882,
+            789
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 THOR-MoE",
+        "text_level": 1,
+        "bbox": [
+            507,
+            799,
+            653,
+            814
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the proposed hierarchical context-responsive routing framework as shown in Figure 1, including two components: hierarchical task-guided routing (§ 3.1) and context-responsive token routing (§ 3.2). Finally, we describe the training objectives (§ 3.3) in detail.",
+        "bbox": [
+            507,
+            824,
+            884,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "21435",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Hierarchical Task-Guided Routing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            433,
+            101
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our hierarchical task-guided routing consists of three stages, as shown in 2) of 1. The first stage is task prediction, which aims to automatically obtain the task knowledge by predicting which domain/language the input belongs to. The second stage is to inject the mixed task representation to allocates experts in a hierarchical task-guided manner (third stage).",
+        "bbox": [
+            112,
+            105,
+            489,
+            233
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Task Predictor. During real testing for a query, it is hard to reach its domain or language category. Therefore, to automatically obtain the domain knowledge or language knowledge, we add a special token [CLS] in front of the input. Then, it represents the global sentence information and thus can effectively reflect which domain/language the input belongs to (Devlin et al., 2019). Finally, we transform the representation  $\\mathbf{H}^{cls}$  into a fixed-size vector through Maxpooling and apply a fully-connected layer to predict suitable labels:",
+        "bbox": [
+            112,
+            234,
+            489,
+            411
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {H} ^ {m a x} = \\operatorname {M a x p o o l i n g} \\left(\\mathbf {H} ^ {c l s}\\right), \\mathbf {H} ^ {m a x} \\in \\mathbb {R} ^ {d}, \\\\ \\mathcal {P} ^ {t} = \\operatorname {S o f t m a x} (\\mathbf {W} ^ {p} \\mathbf {H} ^ {m a x}), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            418,
+            487,
+            456
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{W}^p\\in \\mathbb{R}^{K\\times d}$  is trainable weight and  $K$  denotes the number of domains or languages and  $d$  is the model dimension.",
+        "bbox": [
+            112,
+            466,
+            487,
+            513
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Mixed Task Representation. Theoretically, we can directly take the predicted task label to extract corresponding task representation. However, in some cases in practice, it is hard to judge since some sentences belongs to multiple domains. In language cases, there are some code-mixed sentences, which is also confused to decide. Therefore, directly use the given label in existing work or the predicted label cannot accurately help model select the specialized experts for such cases and thus limits their potential, as shown in § 6.2. To address this issue, we use the mixed task representation as follows:",
+        "bbox": [
+            112,
+            514,
+            489,
+            722
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {E} _ {p} = \\sum \\left(\\mathcal {P} ^ {t} \\cdot \\mathbf {E M B 1}\\right), \\mathbf {E} _ {p} \\in \\mathbb {R} ^ {d}, \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            164,
+            731,
+            487,
+            753
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{E}_p$  denotes the mixed task representation generated by the weighted sum of the task distribution  $\\mathcal{P}$  (predicted by the task predictor as Eq. 8) and the task parameter matrix EMB1. The dimension of EMB1 is num (number of tasks/language groups) * d (embedding dimension), which is randomly initialized and then optimized with the training objective.",
+        "bbox": [
+            112,
+            760,
+            489,
+            888
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Hierarchical Task-Guided Routing. With the mixed task representation, we design a task router",
+        "bbox": [
+            112,
+            889,
+            487,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "$g^{t}$  at the task level. In each MoE layer,  $g^{t}$  takes the mixed task representation  $\\mathbf{E}_p$  as input and outputs a task-dependent expert vector. Then, based on it, we use TopK function to select task-specific candidate experts as  $S^t$  from all experts. Finally, based on the selected task-specific candidate experts  $S_{t}$ , we can conduct general token routing at the token level with Top-  $k$  or Top-  $p$  routing as described in § 2.2.1 and § 2.2.2, respectively. The final routing policy can be written as  $\\sum_{i}^{|\\mathcal{S}^{t}|}\\sum_{j}^{|\\mathcal{E}|}g_{i}^{t}\\cdot g_{j}\\cdot e_{i,j}(\\mathbf{x})$ .",
+        "bbox": [
+            507,
+            83,
+            884,
+            249
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "However, current expert selection at the token level only depends on the current token representation without considering the complete word meaning and global context. Generally, the global context grasp the overall situation and might know which token is difficult or not. Thus, the context can schedules the token to route to suitable experts in a global view. To this end, we propose context-responsive experts selection at the token level.",
+        "bbox": [
+            507,
+            249,
+            885,
+            392
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Context-Responsive Routing",
+        "text_level": 1,
+        "bbox": [
+            507,
+            404,
+            781,
+            420
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown in 3) of 1, before the experts selection at the token level, we explicitly inject the context into the token representation. Since it is hard to determine for a code-mixed input or a query that can belong to several domains, which thus limits its application in practice. Therefore, we use the averaged context representation as the context representation.  $\\mathbf{H}_{ctx} = \\frac{1}{|Y_{1:t - 1}|}\\sum_{t = 1}^{|Y_{1:t - 1}|}\\mathbf{h}_t^L$  where  $\\mathbf{h}$  is the hidden state of the  $t$ -th token at the  $L$  layer. At each decoding step, the context used for each input token is the prefix of the generated next token. That is, the context is dynamically updated with decoding.",
+        "bbox": [
+            507,
+            425,
+            884,
+            634
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To effectively incorporate the context representation into the routing process, we design a gate to dynamically control the contribution of these information:",
+        "bbox": [
+            507,
+            637,
+            882,
+            699
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {x} _ {i} = g \\odot \\mathbf {x} _ {i} + (1 - g) \\odot \\mathbf {H} _ {c t x}, \\\\ g = \\sigma ([ \\mathbf {x} _ {i}; \\mathbf {H} _ {c t x} ] \\mathbf {W} ^ {g} + \\mathbf {b} ^ {g}), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            574,
+            711,
+            811,
+            747
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{W}^g$  and  $\\mathbf{b}^g$  are the trainable parameters.",
+        "bbox": [
+            507,
+            760,
+            860,
+            775
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this manner, we can inject the context information into each token to accurately assign specialized experts to suitable tokens and unlock the potential of the MoE.",
+        "bbox": [
+            507,
+            776,
+            882,
+            839
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Training Objectives",
+        "text_level": 1,
+        "bbox": [
+            507,
+            851,
+            710,
+            866
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our training objectives consist of four parts in Top $k$  routing (translation loss  $\\mathcal{L}_{\\mathrm{NMT}}$ , task prediction loss  $\\mathcal{L}_{\\mathrm{tp}}$ , load balance loss at the task level  $\\mathcal{L}_{bd}$",
+        "bbox": [
+            507,
+            873,
+            882,
+            922
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "21436",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "and token level  $\\mathcal{L}_{bt}$  and five parts in Top-  $p$  routing  $(\\mathcal{L}_{\\mathrm{NMT}},\\mathcal{L}_{\\mathrm{tp}},\\mathcal{L}_{bd},\\mathcal{L}_{bt}$  and dynamic routing loss  $\\mathcal{L}_d$  in Top-  $p$",
+        "bbox": [
+            112,
+            84,
+            487,
+            134
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Task Prediction Loss. The task prediction loss is defined as:",
+        "bbox": [
+            112,
+            143,
+            485,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {t p} = \\min  \\left(- \\log \\left(\\mathcal {P} ^ {t} \\left[ t _ {g} \\right]\\right)\\right), \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            190,
+            487,
+            209
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $t_g$  is the golden class label,  $\\mathcal{P}^t$  is calculated as in Eq. 8.",
+        "bbox": [
+            112,
+            222,
+            485,
+            255
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Load Balance Loss at the Task Level. To fully encourage the experts to specialize in certain tasks, we propose a task-guided routing loss. It aims to achieve the number of tasks processed by different experts to be roughly the same. Therefore, the task-guided routing loss is defined as:",
+        "bbox": [
+            112,
+            266,
+            487,
+            363
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {b d} = N * \\sum_ {i = 1} ^ {N} F _ {i} ^ {t} * Q _ {i} ^ {t}, \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            208,
+            376,
+            487,
+            419
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $F_{i}^{t}$  represents the proportion of tasks selecting expert  $e_{i}$ , and  $Q_{i}^{t}$  denotes the proportion of the router's probability allocated to expert  $e_{i}$ . For  $K$  tasks,  $F_{i}^{t}$  and  $Q_{i}^{t}$  are defined as:  $F_{i}^{t} = \\frac{1}{K}\\sum_{j=1}^{K} 1\\{e_{i} \\in S^{t,j}\\}$  and  $Q_{i}^{t} = \\frac{1}{K}\\sum_{j=1}^{K} P_{i}^{j}$  where  $S^{t,j}$  is the set of activated experts for task  $j$ , which is calculated by Equation 7, and  $P^{j}$  is the probability of selecting each expert for task  $j$ , calculated by Equation 3.",
+        "bbox": [
+            112,
+            429,
+            489,
+            577
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Load Balance Loss at the Token Level. This loss is the similar to the vanilla load balance loss with a slight difference that the experts candidates is pre-selected by the hierarchical task-guided routing. Formally, the load balance loss at the token level is defined as:",
+        "bbox": [
+            112,
+            588,
+            489,
+            683
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {b t} = \\left| \\mathcal {S} ^ {t} \\right| * \\sum_ {k = 1} ^ {\\left| \\mathcal {S} ^ {t} \\right|} F _ {k} ^ {b} * Q _ {k} ^ {b}, \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            201,
+            696,
+            487,
+            741
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $F_{k}^{b}$  represents the proportion of tokens selecting expert  $e_k$ , and  $Q_{k}^{b}$  denotes the proportion of the router's probability allocated to expert  $e_k$ . For a sequence comprising  $M$  tokens,  $F_{k}^{b}$  and  $Q_{k}^{b}$  are defined as:  $F_{k}^{b} = \\frac{1}{M}\\sum_{\\ell = 1}^{M}1\\{e_{k}\\in S^{t,\\ell}\\}$  and  $Q_{k} = \\frac{1}{M}\\sum_{\\ell = 1}^{M}P_{k}^{\\ell}$  where  $S^{t,\\ell}$  is the set of activated experts for token  $\\ell$  from pre-selected experts set  $S^{d,\\ell}$ , which is calculated by Equation 7, and  $P^{\\ell}$  is the probability of selecting each experts for token  $\\ell$ , calculated by Equation 3.",
+        "bbox": [
+            112,
+            756,
+            490,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Dynamic Routing Loss in Top- $p$ . The dynamic routing loss aims to prevent dynamic routing from using too many parameters to cheat and losing its ability to selectively choose experts. Therefore, Huang et al. (2024) introduce a constraint on  $\\mathbf{P}$  and minimize the entropy of the distribution  $\\mathbf{P}$ , ensuring that every token can focus on as less specific experts as possible, which is formalized as:",
+        "bbox": [
+            507,
+            84,
+            882,
+            225
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o p p}} = - \\sum_ {i = 1} ^ {N} P _ {i} * \\log \\left(P _ {i}\\right). \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            589,
+            225,
+            882,
+            267
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Final Loss. Our approach can be flexibly applied in Top- $k$  and Top- $p$  routing strategies. Therefore, we have two final loss functions for Top- $k$  and Top- $p$ , respectively. It is a combination of the translation loss, task prediction loss, load balance loss at the task level and at the token level, and dynamic loss:",
+        "bbox": [
+            507,
+            275,
+            882,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {J} _ {t o p k} = \\mathcal {L} _ {\\mathrm {N M T}} + \\alpha \\mathcal {L} _ {d p} + \\beta \\mathcal {L} _ {b d} + \\gamma \\mathcal {L} _ {b p}, \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            529,
+            385,
+            882,
+            404
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {J} _ {\\text {t o p p}} = \\mathcal {J} _ {\\text {t o p k}} + \\delta \\mathcal {L} _ {\\text {t o p p}}, \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            601,
+            414,
+            882,
+            432
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha, \\beta, \\gamma,$  and  $\\delta$  are hyper-parameters to adjust the contribution among these loss functions, respectively.",
+        "bbox": [
+            507,
+            439,
+            882,
+            488
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            500,
+            655,
+            517
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets and Metric",
+        "text_level": 1,
+        "bbox": [
+            507,
+            526,
+            714,
+            541
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We use the multi-domain translation dataset proposed by Koehn and Knowles (2017). The dataset mainly covers five diverse domains: IT, Koran, Law, Medical, and Subtitles, which are available in OPUS (Aulamo and Tiedemann, 2019). Following previous work (Gu et al., 2022; Liang et al., 2024), we use the new data splitting released by Aharoni and Goldberg (2020), and perform German to English translation  $(\\mathrm{De}\\rightarrow \\mathrm{En})$ . We use the OPUS-16 for multilingual translation following (Zhao et al., 2024b). The OPUS-16 comes from OPUS-100 (Zhang et al., 2020), which includes 16 languages (8 high resource  $(>0.9\\mathrm{M})$ , 4 medium resource, and 4 low resource  $(< 0.1\\mathrm{M})$ ). Please refer to Tab. 8 of § 8 for detailed data statistics.",
+        "bbox": [
+            507,
+            546,
+            882,
+            802
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Metric. For a fair comparison, we follow previous work (Gu et al., 2022; Zhao et al., 2024b) and adopt the 4-gram case-sensitive BLEU with the SacreBLEU tool $^{2}$  (Post, 2018) and report the statistical significance test (Koehn, 2004). For",
+        "bbox": [
+            507,
+            804,
+            882,
+            885
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ BLEU+case.mixed+numrefs.1+smooth.exp+tok.13a+version.1.4.13",
+        "bbox": [
+            507,
+            894,
+            865,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "21437",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/0b9c5d270ee10666d873e7fe9f32ad3ce3e62776e017676ade371576792cbd88.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Models</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td rowspan=\"3\">Dense</td><td>SFT-3B</td><td>40.65</td><td>20.4</td><td>51.4</td><td>54.8</td><td>28.33</td><td>39.12</td></tr><tr><td>Trim-MoE (Top-1)</td><td>40.03</td><td>19.55</td><td>51.80</td><td>55.83</td><td>25.50</td><td>38.54</td></tr><tr><td>Trim-MoE (Top-2)</td><td>45.10</td><td>22.68</td><td>51.84</td><td>57.12</td><td>29.02</td><td>41.15</td></tr><tr><td rowspan=\"4\">MoE</td><td>Trim-MoE (Top-p, p=0.5)</td><td>39.39</td><td>19.21</td><td>55.67</td><td>60.18</td><td>29.21</td><td>40.73</td></tr><tr><td>THOR-MoE (Top-1)</td><td>43.96‡</td><td>21.93‡</td><td>52.52†</td><td>56.25</td><td>28.41‡</td><td>40.61‡</td></tr><tr><td>THOR-MoE (Top-2)</td><td>46.00†</td><td>23.35</td><td>55.79‡</td><td>61.06‡</td><td>28.23</td><td>42.89‡</td></tr><tr><td>THOR-MoE (Top-p, p=0.5)</td><td>44.63‡</td><td>22.53‡</td><td>53.58</td><td>58.65</td><td>27.99</td><td>41.48†</td></tr></table>",
+        "bbox": [
+            127,
+            80,
+            870,
+            227
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: BLEU score on multi-domain translation benchmarks with decoder-only architecture. “†” and “‡” denote that statistically significant better than the best result of the counterpart (e.g., THOR-MoE (Top-2) vs. Trim-MoE (Top-2)) with t-test  $p < 0.05$  and  $p < 0.01$  hereinafter, respectively. The best and second best results are bold and underlined, respectively.",
+        "bbox": [
+            112,
+            237,
+            882,
+            297
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "multilingual translation, the scores encompass the average ratings across all language pairs, such as English  $\\rightarrow$  Any (En  $\\rightarrow$  XX), and Any  $\\rightarrow$  English (XX  $\\rightarrow$  En) on the OPUS-16 dataset.",
+        "bbox": [
+            112,
+            321,
+            487,
+            385
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            399,
+            346,
+            414
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In multi-domain translation, we use decoder-only Transformer architecture. Specifically, we use Qwen1.5-MoE-A2.7B (Team, 2024a). It has 14.3B parameters (60 non shared experts + 4 shared experts) in total and 2.7B activated parameters (4 non shared experts + 4 shared experts) during runtime. Due to limited GPU resource, we used a trimmed version, which has 3.5B parameters (8 non shared experts + 4 shared experts) in total and 2.3B activated parameters (2 non shared experts + 4 shared experts). To keep its capacity, we initialize the trimmed model with Qwen1.5-MoE-A2.7B and denote it as Trim-MoE. During training, we use Llama-Factory (Zheng et al., 2024) to instruct-tune LLMs. All LLMs are tuned on an  $8 \\times$  NVIDIA A100 GPUs (40G) with 1e-5 learning rate. We set gradient accumulation to 16 and batch size to 1, which gives us  $2^{*}8^{*}16^{*}1$  batch in total. We use the DeepSpeed optimization (Rasley et al., 2020), and set ZeRO-3 optimization. Following Qin et al. (2024), we set the number of training epochs to 3. We set hyper-parameters  $\\alpha$ ,  $\\beta$ ,  $\\gamma$ , and  $\\delta$  to 1e-2, 1e-2, 1e-2, and 1e-4, which are determined by a grid search.",
+        "bbox": [
+            112,
+            420,
+            489,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In multilingual translation, we follow (Zhao et al., 2024b) and compare our method with the Transformer-Base model (as Dense) and its MoE variants that have 6 encoder and decoder layers, 32 experts. The input and hidden dimensions of all feed-forward networks are 512 and 2048. We set the training processes to 35K iterations with",
+        "bbox": [
+            112,
+            809,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "a learning rate of 5e-4, which follows the inverse square root with 4,000 warm-up steps. To keep balanced training, we use temperature-based data sampling strategy with temperature 1.5. We set hyper-parameters  $\\alpha$ ,  $\\beta$ ,  $\\gamma$ , and  $\\delta$  to 1e-2, 5e-2, 5e-2, and 1e-4, which are determined by a grid search.",
+        "bbox": [
+            507,
+            321,
+            884,
+            418
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Comparison Models",
+        "text_level": 1,
+        "bbox": [
+            507,
+            430,
+            717,
+            445
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our comparison models mainly include two types: Decoder-only and Encoder-Decoder based.",
+        "bbox": [
+            507,
+            453,
+            882,
+            483
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Decoder-only. We fine-tune two dense models based on Qwen2.5-3B (Team, 2024b), which have similar parameters with the activated parameter and denoted as SFT-3B. Besides, based on the TrimMoE-A2.3B model, we also fine-tuned three MoE models with Top-1, Top-2, and Top- $p$  routing strategy.",
+        "bbox": [
+            507,
+            486,
+            884,
+            598
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Encoder-Decoder. There are three types of comparison models: vanilla dense model, vanilla MoE model, and enhanced MoE models via language knowledge. We use the Transformer-Base model as Dense model. For vanilla MoE model, the Switch Transformer (Fedus et al., 2022) with a top-1 token-based routing (as ST-MoE) and GShard (Lepikhin et al., 2021a) with a top-2 token-based routing (as GS-MoE) are adopted. For Language-specific MoE models, the LS-MoE with fixed routing inspired by (Pires et al., 2023), assigning 2 non-overlapping experts for tokens according to their source language in the encoder and target language in the decoder; Hybrid-MoE (Kudugunta et al., 2021), with a top-2 token routing in the encoder and a top-2 target language routing in the decoder side; Residual-MoE (Elbayad et al., 2023; Rajbhandari et al., 2022; Zhang et al., 2021) that augments each MoE layer with a shared feed-forward network through a binary gate function; Lingual",
+        "bbox": [
+            507,
+            599,
+            885,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "21438",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "MoE (Zhao et al., 2024b) stands for an MoE model with linguistic-guided routing and dynamic expert allocation, where the first-level language router selects the top 8 experts and the second-level token router activates the dynamical number of experts.",
+        "bbox": [
+            112,
+            84,
+            487,
+            164
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            179,
+            263,
+            193
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1 shows main results on the multi-domain translation benchmark with decoder-only architecture. Table 2 presents main results on the multilingual translation with encoder-decoder architecture. For a fair comparison, all models are trained and assessed on the OPUS-16 dataset with Transformer-Base as the backbone architecture.",
+        "bbox": [
+            112,
+            206,
+            489,
+            317
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Results on Multi-Domain Translation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            332,
+            453,
+            346
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1 shows that the Trim-MoE-based models generally surpass the dense one. For example, the Trim-MoE (Top-2) outperforms SFT-3B model by average 2 BLEU scores where the activated parameters is less than dense model (2.3B vs. 3B), showing the effectiveness of MoE model. Furthermore, the proposed THOR-MoE significantly and consistently surpasses the counterpart of Trim-MoE-based one. For instance, the THOR-MoE (Top-2) outperforms the Trim-MoE (Top-2) by averaged 1.74 BLEU scores. This clear advantage confirms the superiority of incorporating domain knowledge and context knowledge into the routing. What's important, the proposed approach is compatible with the Top- $k$  (Shazeer et al., 2017a) and Top $p$  (Huang et al., 2024) routing strategies, validating the generalization of THOR-MoE, which can be a plug-and-play module. Besides, we find that the scores on the same domain (e.g., Law) change largely. The reason may be that the data distribution is unbalanced. We also list the results in terms of COMET (Rei et al., 2020) score in Table 3 and we can conclude the similar findings.",
+        "bbox": [
+            112,
+            354,
+            489,
+            724
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Results on Multilingual Translation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            737,
+            440,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2 summarizes the results and we can conclude several observations:",
+        "bbox": [
+            112,
+            760,
+            489,
+            790
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "THOR-MoE vs. Dense and Vanilla MoE Baselines. The THOR-MoE substantially surpasses the Dense and vanilla MoE multilingual translation baselines with a large margin. Specifically, compared with Dense and ST-MoE, the THOR-MoE (Top-p) achieves  $\\{5.39\\%, 3.99\\%, 4.70\\% \\}$  and  $2.46\\%, 0.55\\%, 1.51\\%$  averaged improvement in terms of BLEU score for Avg1., Avg2., and All",
+        "bbox": [
+            112,
+            791,
+            489,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Avg., respectively. This significant margin demonstrates the effectiveness of hierarchical language-guided routing and context-responsive routing.",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "THOR-MoE vs. Language-Guided MoE models. The THOR-MoE also consistently outperforms language-guided MoE models, including LS-MoE, hybrid-MoE, Residual-MoE, and Linguual-MoE. It shows again that the superiority of hierarchical language-guided routing and context-responsive routing. In detail, the Linguual-MoE performs the highest in baselines and it also employs a hierarchical language-group-guided and dynamical routing, which introduces the hard language id embedding and fails to consider code-mixed cases and that the token in different language can be the same (e.g., 'Internet' in German and English). In contrast, THOR-MoE applies a mixed language representation where the language ids are predicted, which comprehensively incorporates the languages knowledge. Furthermore, the dynamical routing in Linguual-MoE does not incorporate the context information. The context generally knows which token is difficult or not in a global view. The THOR-MoE fully considers the above issues and obtains better results.",
+        "bbox": [
+            507,
+            133,
+            884,
+            485
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6 Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            500,
+            618,
+            516
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.1 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            527,
+            673,
+            542
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct ablation studies to investigate how well hierarchical context-responsive routing of THORMoE works. We conclude two findings from the results in Tab. 4.",
+        "bbox": [
+            507,
+            548,
+            882,
+            611
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) \"w/o hierarchical task-guided routing\": i.e., without using any task-related routing and decaying to vanilla manner (select experts from full set of experts), the model performance greatly degrades on both translation tasks. It shows the necessity of using task-guided routing in a hierarchical manner.",
+            "(2) \"w/o context-responsive routing\": the model performance becomes worse on both tasks when removing context. This shows that our context-responsive routing indeed can enhance the routing effectiveness and guarantee the token can be assigned to suitable and specialized experts with the indication of context, which thus benefits the model performance on both translation tasks."
+        ],
+        "bbox": [
+            507,
+            613,
+            882,
+            838
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2 Comparison among Different Task Representations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            851,
+            826,
+            882
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we aim to investigate the impact of task representation in different manners. Table 5",
+        "bbox": [
+            507,
+            889,
+            882,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "21439",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/0ba8f1edf9edcc7257ab1386c344ef1a7a96c1d1cefb03518ef6e62d5d0bf25c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Models</td><td colspan=\"4\">En→XX</td><td colspan=\"4\">XX→En</td><td rowspan=\"2\">Avg.</td></tr><tr><td>high</td><td>medium</td><td>low</td><td>Avg1.</td><td>high</td><td>medium</td><td>low</td><td>Avg2.</td></tr><tr><td rowspan=\"5\">Dense</td><td>Transformer-base</td><td>25.37</td><td>39.12</td><td>14.78</td><td>26.16</td><td>28.81</td><td>39.24</td><td>24.21</td><td>30.27</td><td>28.21</td></tr><tr><td>GS-MoE</td><td>25.55</td><td>40.71</td><td>18.97</td><td>27.70</td><td>28.77</td><td>41.70</td><td>29.25</td><td>32.12</td><td>29.91</td></tr><tr><td>ST-MoE</td><td>26.55</td><td>43.04</td><td>20.23</td><td>29.09</td><td>29.94</td><td>44.00</td><td>30.94</td><td>33.71</td><td>31.40</td></tr><tr><td>LS-MoE</td><td>20.06</td><td>37.31</td><td>11.72</td><td>22.29</td><td>23.69</td><td>42.99</td><td>32.01</td><td>30.60</td><td>26.44</td></tr><tr><td>Hybird-MoE</td><td>24.56</td><td>31.70</td><td>13.38</td><td>23.55</td><td>29.59</td><td>41.74</td><td>27.81</td><td>32.18</td><td>27.85</td></tr><tr><td rowspan=\"5\">MoE</td><td>Residual-MoE</td><td>26.52</td><td>43.49</td><td>21.58</td><td>29.53</td><td>29.84</td><td>43.94</td><td>31.25</td><td>33.72</td><td>31.62</td></tr><tr><td>Lingual-MoE</td><td>27.11</td><td>46.24</td><td>23.34</td><td>30.95</td><td>29.87</td><td>43.49</td><td>32.02</td><td>33.81</td><td>32.38</td></tr><tr><td>THOR-MoE (Top-1)</td><td>27.21</td><td>46.37</td><td>23.88</td><td>31.17</td><td>29.68</td><td>44.01</td><td>32.58</td><td>33.99</td><td>32.57</td></tr><tr><td>THOR-MoE (Top-2)</td><td>27.88†</td><td>47.29†</td><td>24.85†</td><td>31.98†</td><td>30.38†</td><td>44.52†</td><td>33.26†</td><td>34.64†</td><td>33.31†</td></tr><tr><td>THOR-MoE (Top-p)</td><td>27.63†</td><td>46.82†</td><td>24.15†</td><td>31.55†</td><td>29.97</td><td>44.18†</td><td>32.96†</td><td>34.26</td><td>32.91†</td></tr></table>",
+        "bbox": [
+            124,
+            80,
+            877,
+            266
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a698788ba33027384a9bba7da30bf785713b5691798c0ced4d56e5adae715dd5.jpg",
+        "table_caption": [
+            "Table 2: Averaged BLEU scores on multilingual translation with encoder-decoder architectures. “†” denotes that statistically significant better than Lingual-MoE with t-test  $p < 0.05$ ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td>SFT-3B</td><td>86.39</td><td>73.17</td><td>85.43</td><td>87.19</td><td>78.52</td><td>82.14</td></tr><tr><td>THOR-MoE (Top-p)</td><td>87.31</td><td>74.09</td><td>86.09</td><td>87.88</td><td>79</td><td>82.87</td></tr></table>",
+        "bbox": [
+            121,
+            328,
+            485,
+            373
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/f3907fd03eeac33e38a870283e45765e72598608ff1587049e4601b46ba62215.jpg",
+        "table_caption": [
+            "Table 3: COMET scores on multi-domain translation benchmarks with decoder-only architecture."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Multi-Domain</td><td>Multilingual</td></tr><tr><td>THOR-MoE (Top-p)</td><td>41.48</td><td>32.91</td></tr><tr><td>w/o hierarchical task-guided routing</td><td>40.16</td><td>31.57</td></tr><tr><td>w/o context-responsive routing</td><td>40.74</td><td>32.12</td></tr></table>",
+        "bbox": [
+            121,
+            429,
+            484,
+            483
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "shows the results where the 'Non-Mixed Representation' denotes that directly uses the automatically predict task label to extract corresponding task representation and 'Golden Representation' indicates using the golden task label to extract corresponding representation. We conclude that the task knowledge indeed has a positive impact on translation performance (vs. baseline). We also observe that using the automatically predicted task labels (actually the mixed task representation) shows better results than using ground truth and Non-Mixed manner in terms of Avg. BLEU scores. The reason may be that the mixed task representation has certain fault tolerance. We also analyze the accuracy of task prediction to further show its quality",
+        "bbox": [
+            112,
+            539,
+            489,
+            781
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/747dad36aff06e268e81bd6ae7f56620a7361eb3669a2ae8b51eca6f9a4700d7.jpg",
+        "table_caption": [
+            "Table 4: Ablation Study with Avg. results."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Multi-Domain</td><td>Multilingual</td></tr><tr><td>baseline (w/o any task Rep.)</td><td>40.74</td><td>32.12</td></tr><tr><td>Non-Mixed Rep.</td><td>40.95</td><td>32.44</td></tr><tr><td>Golden Rep.</td><td>41.32</td><td>32.78</td></tr><tr><td>THOR-MoE (Mixed manner)</td><td>41.48</td><td>32.91</td></tr></table>",
+        "bbox": [
+            121,
+            804,
+            484,
+            878
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/fe840274113d930ed550a515eba4c43c71e0170127e8b7cbef58e5d61f60ce4e.jpg",
+        "table_caption": [
+            "Table 5: The Avg. results of different representation (Rep.) manner during hierarchical task-guided routing."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Multi-Domain</td><td>Multilingual</td></tr><tr><td>baseline (w/o any task Rep.)</td><td>40.74</td><td>32.12</td></tr><tr><td>Infuse task Rep. into token Rep.</td><td>40.95</td><td>32.44</td></tr><tr><td>THOR-MoE (Hierarchical manner)</td><td>41.48</td><td>32.91</td></tr></table>",
+        "bbox": [
+            515,
+            328,
+            878,
+            385
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: The investigation of why hierarchical manner.",
+        "bbox": [
+            509,
+            394,
+            880,
+            409
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7c46b75344c67dc342d848a35ef6d0fc4422e4513c6c5b4e77b20107b57dddb3.jpg",
+        "image_caption": [
+            "Figure 2: Average activated experts number across training steps on the multi-domain translation task."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            510,
+            432,
+            884,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "in Appendix A.",
+        "bbox": [
+            507,
+            706,
+            628,
+            722
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.3 Why Hierarchical Design?",
+        "text_level": 1,
+        "bbox": [
+            507,
+            747,
+            763,
+            764
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we aim to investigate the manner of using mixed task representation. Table 6 shows that introducing task knowledge indeed helps tokens route well and achieve improvement (vs. baseline). The hierarchical manner greatly outperforms the direct infusion with token representation before routing, proving the effectiveness of hierarchical design, which unlock the potential of MoE with the help of the mixed task knowledge.",
+        "bbox": [
+            505,
+            776,
+            882,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "21440",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/29a0d70766864764131ef58e14cbaa2f71b428de0bf889fb282b27a4aa2b206c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Domains</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td>Top-p</td><td>1.87</td><td>1.95</td><td>1.82</td><td>1.92</td><td>1.77</td><td>1.87</td></tr><tr><td>Top-p w/ ctx</td><td>1.37</td><td>1.61</td><td>1.42</td><td>1.56</td><td>1.31</td><td>1.45</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            485,
+            134
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Average activated experts (AE) in different domain translation tasks. 'Top-  $p$  w/ctx' denotes the proposed method to infuse the context during routing.",
+        "bbox": [
+            112,
+            143,
+            487,
+            187
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6.4 Analysis of Efficient Training and Inference",
+        "text_level": 1,
+        "bbox": [
+            112,
+            214,
+            425,
+            244
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "To further explore whether our proposed method is efficient in training and inference, we calculate the average number of experts activated by the model on multi-domain translation tasks. Figure 2 and Table 7 shows the average number of experts activated per token across various translation tasks during training and inference. The result is averaged across all the layers of transformers.",
+        "bbox": [
+            112,
+            252,
+            487,
+            380
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "During training, we can see that our method converges faster to use less activated experts than the vanilla Top- $p$  (Huang et al., 2024). During inference, we can observe that across all five tasks, the number of activated experts is less than original Top- $p$  routing, averaging 1.45 activated experts (less than  $22\\%$  activated parameters) with better performance. Both findings show that the context plays a key role in guiding token routing.",
+        "bbox": [
+            112,
+            382,
+            489,
+            526
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            539,
+            270,
+            555
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Neural Machine Translation. The NMT have received remarkable attention in the era of LLMs (Jawahar et al., 2023b; Zhou et al., 2023). Previous work mainly focuses on the continual learning of new domains (Gu and Feng, 2020; Gu et al., 2022; Liang et al., 2024), introducing ready-made task-related (linguistics) knowledge via adapter (Zhang et al., 2020, 2021), parameter sharing (Aharoni et al., 2019), or task-specific modules (Pires et al., 2023), and multilingual representation learning (Yang et al., 2021).",
+        "bbox": [
+            112,
+            565,
+            489,
+            743
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Mixture-of-Experts. Jacobs et al. (1991) first proposes the concept of MoE, which consists of a series of network sub-modules. The Sparately-gated MoE, as a variant, which only activates a few expert networks for each input, has been shown its superiority in diverse NLP and computer vision applications(Shazeer et al., 2017b; Zoph et al., 2022). In the context of NMT, most of previous work focuses on addressing the over/under-fitting (Elbayad et al., 2023) via regularization strategies (Elbayad et al., 2023) or modularizing MoE (Li et al., 2023c;",
+        "bbox": [
+            112,
+            744,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Zhang et al., 2024), and incorporating task-related knowledge (language, domain etc) (Kudugunta et al., 2021; Zhao et al., 2024b; Pham et al., 2023; Gururangan et al., 2022; Li et al., 2023b), achieving impressive performance. Besides, there are some studies aim to improve the training/inference efficiency of routing via adaptive computation (Jawahar et al., 2023a) or reducing the number of activated experts (Li et al., 2023a; Huang et al., 2024).",
+        "bbox": [
+            507,
+            84,
+            884,
+            230
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Different from these prior works that directly use the task-related knowledge, we propose a hierarchical context-responsive routing method where we automatically extract corresponding domain/language knowledge and design a hierarchical network to guide the task-level routing. Besides, we aim to help each token accurately select specialized and suitable experts and thus we incorporate the context to guide the token routing rather than relying on the localized token only in existing work. In this manner, the context-responsive routing can improve the training or inference efficiency, which is compatible with previous Top- $k$  and Top- $p$  routing policies.",
+        "bbox": [
+            507,
+            231,
+            885,
+            458
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            479,
+            640,
+            494
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a new hierarchical task-gudied and context-responsive routing framework for NMT. To automatically obtain the task knowledge, we propose to predict it and then use mixed task representation. Consequently, we design a hierarchical routing at the task level and the token level. Further, we propose to inject context to enhance the effectiveness of token routing. Extensive experiments on multi-domain and multilingual translation benchmarks show the superiority and generalization of our proposed approach.",
+        "bbox": [
+            507,
+            512,
+            884,
+            690
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            711,
+            613,
+            726
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While we introduce the mixed domain/language and context knowledge into routing in hierarchical manner and achieve good results, there are some limitations worth considering to study in future work: (1) In this study, the design relies on the prior (the number of tasks and language groups), which may limit its extension to broader topics (Li and Zhou, 2025); (2) This work only conduct experiments on translation tasks and does not conduct experiments other generation or discriminative tasks.",
+        "bbox": [
+            507,
+            744,
+            885,
+            904
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "21441",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Roee Aharoni and Yoav Goldberg. 2020. Unsupervised domain clusters in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7747-7763, Online. Association for Computational Linguistics.",
+            "Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Mikko Aulamo and Jörg Tiedemann. 2019. The OPUS resource repository: An open package for creating parallel corpora and machine translation services. In Proceedings of the 22nd Nordic Conference on Computational Linguistics, NoDaLiDa 2019, Turku, Finland, September 30 - October 2, 2019, pages 389-394.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Maha Elbayad, Anna Sun, and Shruti Bhosale. 2023. Fixing MoE over-fitting on low-resource languages in multilingual machine translation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 14237-14253, Toronto, Canada. Association for Computational Linguistics.",
+            "Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, and Armand Joulin. 2020. Beyond english-centric multilingual machine translation. CoRR, abs/2010.11125.",
+            "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.",
+            "Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Roziere, David Lopez-Paz, and Gabriel Synnaeve. 2024. Better & faster large language models via multi-token prediction. Preprint, arXiv:2404.19737.",
+            "Shuhao Gu and Yang Feng. 2020. Investigating catastrophic forgetting during continual training for neural machine translation. In Proceedings of the 28th International Conference on Computational Linguistics,"
+        ],
+        "bbox": [
+            115,
+            105,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "pages 4315-4326, Barcelona, Spain (Online). International Committee on Computational Linguistics.",
+            "Shuhao Gu, Bojie Hu, and Yang Feng. 2022. Continual learning of neural machine translation within low-forgetting risk regions. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1707-1718, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Suchin Gururangan, Mike Lewis, Ari Holtzman, Noah A. Smith, and Luke Zettlemoyer. 2022. DEMix layers: Disentangling domains for modular language modeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5557-5576, Seattle, United States. Association for Computational Linguistics.",
+            "Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. 2024. Harder task needs more experts: Dynamic routing in MoE models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12883-12895, Bangkok, Thailand. Association for Computational Linguistics.",
+            "Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. 1991. Adaptive mixtures of local experts. Neural Comput., 3(1):79-87.",
+            "Ganesh Jawahar, Subhabrata Mukherjee, Xiaodong Liu, Young Jin Kim, Muhammad Abdul-Mageed, Laks Lakshmanan, V.S., Ahmed Hassan Awadallah, Sebastian Bubeck, and Jianfeng Gao. 2023a. AutoMoE: Heterogeneous mixture-of-experts with adaptive computation for efficient neural machine translation. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9116-9132, Toronto, Canada. Association for Computational Linguistics.",
+            "Ganesh Jawahar, Subhabrata Mukherjee, Xiaodong Liu, Young Jin Kim, Muhammad Abdul-Mageed, Laks V. S. Lakshmanan, Ahmed Hassan Awadallah, Sébastien Bubeck, and Jianfeng Gao. 2023b. Automoe: Heterogeneous mixture-of-experts with adaptive computation for efficient neural machine translation. In *Findings of the Association for Computational Linguistics: ACL 2023*, Toronto, Canada, July 9-14, 2023, pages 9116-9132. Association for Computational Linguistics.",
+            "Songtao Jiang, Tuo Zheng, Yan Zhang, Yeying Jin, Li Yuan, and Zuozhu Liu. 2024. Med-MoE: Mixture of domain-specific experts for lightweight medical vision-language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2024, pages 3843-3860, Miami, Florida, USA. Association for Computational Linguistics.",
+            "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of EMNLP, pages 388-395."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "21442",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, NMT@ACL 2017, Vancouver, Canada, August 4, 2017, pages 28-39.",
+            "Sneha Kudugunta, Yanping Huang, Ankur Bapna, Maxim Krikun, Dmitry Lepikhin, Minh-Thang Luong, and Orhan First. 2021. Beyond distillation: Task-level mixture-of-experts for efficient inference. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3577-3599, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2021a. {GS}hard: Scaling giant models with conditional computation and automatic sharding. In International Conference on Learning Representations.",
+            "Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2021b. Gshard: Scaling giant models with conditional computation and automatic sharding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.",
+            "Jiamin Li, Qiang Su, Yitao Yang, Yimin Jiang, Cong Wang, and Hong Xu. 2023a. Adaptive gating in mixture-of-experts based language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3577-3587, Singapore. Association for Computational Linguistics.",
+            "Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A. Smith, and Luke Zettle-moyer. 2023b. Branch-train-merge: Embarrassingly parallel training of expert language models.",
+            "Shangjie Li, Xiangpeng Wei, Shaolin Zhu, Jun Xie, Baosong Yang, and Deyi Xiong. 2023c. MMNMT: Modularizing multilingual neural machine translation with flexibly assembled MoE and dense blocks. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4978-4990, Singapore. Association for Computational Linguistics.",
+            "Ziyue Li and Tianyi Zhou. 2025. Your mixture-of-experts LLM is secretly an embedding model for free. In The Thirteenth International Conference on Learning Representations.",
+            "Yunlong Liang, Fandong Meng, Jiaan Wang, Jinan Xu, Yufeng Chen, and Jie Zhou. 2024. Continual learning with semi-supervised contrastive distillation for incremental neural machine translation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10914-10928, Bangkok, Thailand. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Hai Pham, Young Jin Kim, Subhabrata Mukherjee, David P. Woodruff, Barnabas Poczos, and Hany Hassan. 2023. Task-based MoE for multitask multilingual machine translation. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 164–172, Singapore. Association for Computational Linguistics.",
+            "Telmo Pires, Robin Schmidt, Yi-Hsiu Liao, and Stephan Peitz. 2023. Learning language-specific layers for multilingual machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14767-14783, Toronto, Canada. Association for Computational Linguistics.",
+            "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of WMT, pages 186-191.",
+            "Yiwei Qin, Xuefeng Li, Haoyang Zou, Yixiu Liu, Shijie Xia, Zhen Huang, Yixin Ye, Weizhe Yuan, Hector Liu, Yuanzhi Li, et al. 2024. O1 replication journey: A strategic progress report-part 1. arXiv preprint arXiv:2410.18982.",
+            "Samyam Rajbhandari, Conglong Li, Zhewei Yao, Minjia Zhang, Reza Yazdani Aminabadi, Ammar Ahmad Awan, Jeff Rasley, and Yuxiong He. 2022. Deepspeed-moe: Advancing mixture-of-experts inference and training to power next-generation ai scale. Preprint, arXiv:2201.05596.",
+            "Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 3505-3506.",
+            "Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702, Online. Association for Computational Linguistics.",
+            "Noam Shazeer, *Azalia Mirhoseini, *Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017a. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations.",
+            "Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. 2017b. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.",
+            "Qwen Team. 2024a. Qwen1.5-moe: Matching 7b model performance with 1/3 activated parameters\".",
+            "Qwen Team. 2024b. Qwen2.5: A party of foundation models."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "21443",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "An Wang, Xingwu Sun, Ruobing Xie, Shuaipeng Li, Ji- aqi Zhu, Zhen Yang, Pinxue Zhao, J. N. Han, Zhanhui Kang, Di Wang, Naoaki Okazaki, and Cheng zhong Xu. 2024. Hmoe: Heterogeneous mixture of experts for language modeling. Preprint, arXiv:2408.10681.",
+            "Junhong Wu, Yuchen Liu, and Chengqing Zong. 2024. F-MALLOC: Feed-forward memory allocation for continual learning in neural machine translation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7180-7192, Mexico City, Mexico. Association for Computational Linguistics.",
+            "Yilin Yang, Akiko Eriguchi, Alexandre Muzio, Prasad Tadepalli, Stefan Lee, and Hany Hassan. 2021. Improving multilingual translation by representation and gradient regularization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7266-7279, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Zihao Zeng, Yibo Miao, Hongcheng Gao, Hao Zhang, and Zhijie Deng. 2024. AdaMoE: Token-adaptive routing with null experts for mixture-of-experts language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6223-6235, Miami, Florida, USA. Association for Computational Linguistics.",
+            "Biao Zhang, Ankur Bapna, Rico Sennrich, and Orhan First. 2021. Share or not? learning to schedule language-specific capacity for multilingual translation. In International Conference on Learning Representations.",
+            "Biao Zhang, Philip Williams, Ivan Titov, and Rico Sennrich. 2020. Improving massively multilingual neural machine translation and zero-shot translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1628-1639, Online. Association for Computational Linguistics.",
+            "Fan Zhang, Mei Tu, Song Liu, and Jinyao Yan. 2024. A lightweight mixture-of-experts neural machine translation model with stage-wise training strategy. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2381-2392, Mexico City, Mexico. Association for Computational Linguistics.",
+            "Hao Zhao, Zihan Qiu, Huijia Wu, Zili Wang, Zhaofeng He, and Jie Fu. 2024a. HyperMoE: Towards better mixture of experts via transferring among experts. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10605-10618, Bangkok, Thailand. Association for Computational Linguistics.",
+            "Xinyu Zhao, Xuxi Chen, Yu Cheng, and Tianlong Chen. 2024b. Sparse moe with language guided routing for"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "multilingual machine translation. In The Twelfth International Conference on Learning Representations.",
+            "Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. LlamaFactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 400-410, Bangkok, Thailand. Association for Computational Linguistics.",
+            "Chulun Zhou, Yunlong Liang, Fandong Meng, Jinan Xu, Jinsong Su, and Jie Zhou. 2023. RC3: Regularized contrastive cross-lingual cross-modal pre-training. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 11747–11762, Toronto, Canada. Association for Computational Linguistics.",
+            "Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. 2022. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906.",
+            "Simiao Zuo, Xiaodong Liu, Jian Jiao, Young Jin Kim, Hany Hassan, Ruofei Zhang, Jianfeng Gao, and Tuo Zhao. 2022. Taming sparsely activated transformer with stochastic experts. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            467
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "21444",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/873b12617108b5add48059e4e4518886b5dc3f8e274a9cfbbcb7b4b1a9fe24c2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Tasks/Groups</td><td>Train</td><td>Valid</td><td>Test</td></tr><tr><td rowspan=\"5\">Multi-Domain Translation Dataset (De→En)</td><td>IT</td><td>0.22M</td><td></td><td></td></tr><tr><td>Koran</td><td>18K</td><td></td><td></td></tr><tr><td>Law</td><td>0.47M</td><td>2000</td><td>2000</td></tr><tr><td>Medical</td><td>0.25M</td><td></td><td></td></tr><tr><td>Subtitles</td><td>0.5M</td><td></td><td></td></tr><tr><td>Multilingual Translation Dataset</td><td>16</td><td>17,559,950</td><td>30*1000</td><td>30*1000</td></tr></table>",
+        "bbox": [
+            122,
+            80,
+            484,
+            199
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/0f48aa72a0571d2e986088268a9e2bec4e814d6aa3e9dc234d1fc6eec2c932be.jpg",
+        "table_caption": [
+            "Table 8: The data statistic of the multi-domain translation dataset and multilingual translation (OPUS-16) dataset. The number in Train/Valid/Test columns denotes the number of sentence pairs in each domain/language pair."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Models</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td>Qwen1.5-MoE-A2.7B (SFT, Top-8)</td><td>45.59</td><td>23.31</td><td>55.67</td><td>60.18</td><td>29.21</td><td>42.79</td></tr><tr><td>THOR-MoE (Top-p, p=0.5)</td><td>46.76</td><td>24.03</td><td>35.9</td><td>61.13</td><td>30.29</td><td>43.62</td></tr><tr><td>Qwen1.5-MoE-A2.7B (SFT, Top-8)</td><td>87.3</td><td>74.24</td><td>85.85</td><td>87.78</td><td>79.36</td><td>82.91</td></tr><tr><td>THOR-MoE (Top-p, p=0.5)</td><td>87.45</td><td>74.58</td><td>86.45</td><td>88.75</td><td>80.23</td><td>83.49</td></tr></table>",
+        "bbox": [
+            119,
+            294,
+            490,
+            351
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A Task Prediction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            458,
+            289,
+            473
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We also evaluate the performance of the Task Predictor to show whether the classifier can accurately predict suitable labels. The results are  $82.45\\%$  and  $64.89\\%$  for domain and language prediction, respectively. It suggests that our classifier can predict suitable labels and further provide effective mixed task representation for routing.",
+        "bbox": [
+            112,
+            486,
+            489,
+            598
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Comparison to Larger Model",
+        "text_level": 1,
+        "bbox": [
+            114,
+            612,
+            405,
+            627
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We implemented our proposed method based on a large-scale Qwen1.5-MoE-A2.7B (totally 14B, activated 2.7B, denoted as THOR-MoE). Besides, we also fine-tuned the vanilla Qwen1.5-MoE-A2.7B model as the baseline. The averaged COMET and BLEU results shown in Table 9 demonstrate that the THOR-MoE model at the 14B level also achieve better results than fine-tuned Qwen1.5-MoE-A2.7B model in terms of BLEU and COMET scores.",
+        "bbox": [
+            112,
+            639,
+            489,
+            797
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Analysis of Domain Similarity",
+        "text_level": 1,
+        "bbox": [
+            114,
+            813,
+            413,
+            829
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Following previous work (Wu et al., 2024), we conduct some analysis for domain similarity. Specifically, we sum the predicted soft label and then normalize it. The vertical axis and horizontal axis are unsupervised cluster label (clustering with BERT",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/67e1d8d1ac68c4db99ab4edcb46b924dd840e63195a50c23c1eca8cdead14e02.jpg",
+        "table_caption": [
+            "Table 9: BLEU (top block) / COMET (below block) scores on multi-domain translation benchmarks with decoder-only architecture. Note that top-8 means it activates 4 shared experts and 4 non-shared experts out of 60 experts."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td></tr><tr><td>IT</td><td>0.833</td><td>0.027</td><td>0.055</td><td>0.0219</td><td>0.065</td></tr><tr><td>Koran</td><td>0.012</td><td>0.817</td><td>0.003</td><td>0.019</td><td>0.149</td></tr><tr><td>Medical</td><td>0.111</td><td>0.004</td><td>0.737</td><td>0.135</td><td>0.013</td></tr><tr><td>Law</td><td>0.142</td><td>0.008</td><td>0.032</td><td>0.798</td><td>0.02</td></tr><tr><td>Subtitles</td><td>0.058</td><td>0.076</td><td>0.009</td><td>0.006</td><td>0.851</td></tr></table>",
+        "bbox": [
+            531,
+            80,
+            860,
+            170
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 10: BLEU (top block) / COMET (below block) scores on multi-domain translation benchmarks with decoder-only architecture. Note top-8 means it activates 4 shared experts and 4 non-shared experts out of 60 experts.",
+        "bbox": [
+            507,
+            179,
+            882,
+            250
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "base and  $k = 5$  ) and predicted soft labels, respectively. The similarity matrix are shown in Table 10.",
+        "bbox": [
+            507,
+            275,
+            882,
+            307
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The results show that there exist some sentences that are assigned to a cluster of another domain (i.e., hard cases). This makes sense as the mixed representation can capture such domain information, making the proposed approach works (the reason).",
+        "bbox": [
+            507,
+            309,
+            882,
+            388
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "21445",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2025/THOR-MoE_ Hierarchical Task-Guided and Context-Responsive Routing for Neural Machine Translation/f6520cf3-9239-4e76-8a25-131bf94b31d7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b17b2ce634d0b99731ad732b4061991b45fd83e12735194a4fb0d7ffa39369d
+size 513465

2025/THOR-MoE_ Hierarchical Task-Guided and Context-Responsive Routing for Neural Machine Translation/full.md

@@ -0,0 +1,408 @@
+# THOR-MoE: Hierarchical Task-Guided and Context-Responsive Routing for Neural Machine Translation
+
+Yunlong Liang, Fandong Meng*, Jie Zhou
+
+Pattern Recognition Center, WeChat AI, Tencent Inc
+
+{yunlonliang, fandongmeng, withtomzhou}@tencent.com
+
+# Abstract
+
+The sparse Mixture-of-Experts (MoE) has achieved significant progress for neural machine translation (NMT). However, there exist two limitations in current MoE solutions which may lead to sub-optimal performance: 1) they directly use the task knowledge of NMT into MoE (e.g., domain/linguistics-specific knowledge), which are generally unavailable at practical application and neglect the naturally grouped domain/linguistic properties; 2) the expert selection only depends on the localized token representation without considering the context, which fully grasps the state of each token in a global view. To address the above limitations, we propose THOR-MoE via arming the MoE with hierarchical task-guided and context-responsive routing policies. Specifically, it 1) firstly predicts the domain/language label and then extracts mixed domain/language representation to allocate task-level experts in a hierarchical manner; 2) injects the context information to enhance the token routing from the pre-selected task-level experts set, which can help each token to be accurately routed to more specialized and suitable experts. Extensive experiments on multi-domain translation and multilingual translation benchmarks with different architectures consistently demonstrate the superior performance of THOR-MoE. Additionally, the THOR-MoE operates as a plug-and-play module compatible with existing Top- $k$  (Shazeer et al., 2017a) and Top $p$  (Huang et al., 2024) routing schemes, ensuring broad applicability across diverse MoE architectures. For instance, compared with vanilla Top- $p$  (Huang et al., 2024) routing, the context-aware manner can achieve an average improvement of 0.75 BLEU with less than  $22\%$  activated parameters on multi-domain translation tasks.
+
+# 1 Introduction
+
+The rapid advancement of neural machine translation (NMT; Zhang et al. (2020); Fan et al. (2020)) has been significantly propelled by sparse Mixture-of-Experts (MoE) architectures (Shazeer et al., 2017a; Lepikhin et al., 2021a; Fedus et al., 2022), which increase model capacity while maintaining computational efficiency through conditional computation. In the literature, existing work can be roughly classified into two categories: (1) some studies mainly focus on how to effectively incorporate task information (e.g., domain/linguistics-specific knowledge) into MoE models designing task-specific modules (Pham et al., 2023; Jiang et al., 2024); (2) another line of work concentrates on how to improve the training and inference efficiency via reducing the activated experts in MoE (Jawahar et al., 2023a; Elbayad et al., 2023; Wang et al., 2024; Zeng et al., 2024).
+
+Specifically, in (1), Kudugunta et al. (2021) routes input examples to different experts based on translation language representations. Li et al. (2023c) explores multiple language-group-specific routers to incorporate language group knowledge into models. Zhao et al. (2024b) takes one step further by designing hierarchical language-guided token routing. Although intuitive and effective, they heavily rely on explicit task-specific knowledge (e.g., domain or language labels), which is often unavailable in real-world scenarios. This reliance not only restricts generalization but also overlooks the intrinsic hierarchical grouping of domain and linguistic properties inherent in multilingual or multi-domain translation tasks. In (2), Zhao et al. (2024a) distills the knowledge of unselected experts to the selected one and thus reduces the number of experts without impacting performance; Li et al. (2023a) and Huang et al. (2024) decrease the number of experts by designing a threshold based on expert probability distribution; Zhao et al.
+
+(2024b) adapts the number of experts by the difficulty of language. Generally, the global context grasp the overall situation and might know whether each token is difficult or not (Gloeckle et al., 2024). However, previous routing strategies select experts solely based on localized token-level representations, neglecting the broader contextual dependencies that govern token interactions. This myopic view limits the model's ability to allocate experts optimally in a globally coherent manner.
+
+In this work, to address the above issues, we propose an innovative hierarchical context-responsive routing (THOR-MoE) framework for neural machine translation. Specifically, the THOR-MoE first predicts which domain/language the input belongs to and then extracts the mixed task representation to allocate experts in a hierarchical manner. It is not only flexible in practice but also effectively infuses the domain knowledge into the task-level routing via encouraging experts to specialize in certain domains/languages. Furthermore, before each token routing, the THOR-MoE injects the context information to help accurately assign the experts from the pre-selected task-level experts set for each token in a global perspective. This enables the model to adaptively capture evolving contextual dependencies, such as long-range syntactic structures or discourse coherence. Crucially, the THOR-MoE operates as a plug-and-play module compatible with existing Top- $k$  (Shazeer et al., 2017a) and Top- $p$  (Huang et al., 2024) routing schemes, ensuring broad applicability across diverse MoE architectures.
+
+We validate our proposed THOR-MoE framework on the commonly-used multi-domain NMT benchmark (Aharoni and Goldberg, 2020) and multilingual NMT benchmark (Zhang et al., 2020), which contains 5 domains and 16 languages, respectively. Extensive experiments show that the hierarchical context-responsive routing is pivotal for unlocking the full potential of MoE systems.
+
+In summary, our main contributions are:
+
+- We propose two components to expand MoE: (i) a hierarchical design to enable language- and domain-specific expert routing; (ii) enabling the use of context information during expert selection, which is important for translation.  
+- Extensive experiments on both multi-domain and multilingual translation consistently show the ef
+
+fectiveness and generalization of THOR-MoE. For example, our model achieves consistent improvements of an average improvement of 0.75 BLEU with less than  $22\%$  activated parameters over vanilla Top-p routing in multi-domain translation tasks.
+
+# 2 Background
+
+# 2.1 Neural Machine Translation
+
+Given an input sentence in the source language  $X = \{x_{i}\}_{i=1}^{|X|}$ , the goal of the model is to produce its translation in the target language  $Y = \{y_{i}\}_{i=1}^{|Y|}$ . The conditional distribution of the model is:
+
+$$
+\mathcal {L} _ {\mathrm {N M T}} = - \sum_ {t = 1} ^ {| Y |} \log \left(p \left(y _ {t} \mid X, y _ {1: t - 1}\right)\right), \tag {1}
+$$
+
+where  $y_{1:t - 1}$  is the partial translation.
+
+# 2.2 Sparse Mixture-of-Experts
+
+The vanilla MoE model can be seen the variant of Transformer model via replacing the feedforward (FFN) sub-block of the transformer block with MoE layer, in which per token selects fixed number of experts (Lepikhin et al., 2021b). In each MoE layer, there are  $N$  experts, denoted as  $\mathcal{E} = \{e_1,e_2,\dots ,e_N\}$ . An input  $\mathbf{x}$  is dispatched to these experts, and the output of the MoE layer is computed as the weighted average of the outputs from the experts:
+
+$$
+\operatorname {M o E} (\mathbf {x}) = \sum_ {i = 1} ^ {N} g _ {i} (\mathbf {x}) * e _ {i} (\mathbf {x}), \tag {2}
+$$
+
+where  $g_{*}(\mathbf{x})$  is computed by a router that predicts the contribution of each expert to the final output. Given the computing efficiency, the MoE assigns each token to limited experts (e.g., 1 or 2).
+
+To derive  $g_{*}(\mathbf{x})$ , we generally compute the probability  $\mathbf{P}$  of each expert being selected for the input  $\mathbf{x}$  as follows:
+
+$$
+\mathbf {P} = \operatorname {S o f t m a x} \left(\mathbf {W} _ {\mathbf {r}} \cdot \mathbf {x} ^ {T}\right), \tag {3}
+$$
+
+where  $\mathbf{W}_{\mathbf{r}} \in N \times d$  serves as a learnable parameter and  $d$  denotes the dimension of the input  $\mathbf{x}$ . The vector  $\mathbf{P}$ , with a size of  $N$ , encapsulates the probabilities associated with the selection of each expert. And  $P_{i}$  indicates the likelihood of choosing the  $i^{th}$  expert  $e_{i}$  to calculate the input  $\mathbf{x}$ .
+
+![](images/e9a5792ab50d7a0f8f2cc6098e6ceaa16d3138cde790dd1b2357ac7a8ed36b75.jpg)  
+(1) Vanilla Token Routing;
+
+![](images/df4bed66f43b928c36629b6683b81ab72a184188338292fdadc967a04d0eaa3d.jpg)  
+(3) Context-Responsive Token Routing.
+
+![](images/8324a558aad599412debbe3de019202ac74c186c8132e29dec2106846a2765ee.jpg)  
+(2) Hierarchical Task-Guided Routing;  
+Figure 1: The overview of the proposed THOR-MoE. 1) vanilla token routing; 2) hierarchical task-guided routing; 3) context-responsive token routing. The 'Emb.' and 'Rep.' denotes the embedding and representation, respectively. The hierarchical manner denotes that the task-guided routing firstly selects different task-specific experts sets for different queries (e.g.,  $\mathcal{E}_t$  for query 1). Then the context-responsive token routing assigns experts from  $S^t$  for each token in query 1.
+
+# 2.2.1 Top-  $k$  Routing
+
+Top- $k$  routing (Shazeer et al., 2017b; Lepikhin et al., 2021b; Zuo et al., 2022) entails the selection of the top  $k$  experts, which correspond to the  $k$  highest probabilities within the vector  $\mathbf{P}$ . Then, the probabilities of the chosen experts are re-scaled through normalization, while the probabilities of the non-selected experts are set to zero, signifying their deactivation. The resultant computation for  $g_{*}(\mathbf{x})$  proceeds as below:
+
+$$
+g _ {i} (\mathbf {x}) = \left\{ \begin{array}{l l} \frac {P _ {i}}{\sum_ {j \in \operatorname {T o p K} (\mathbf {P})} P _ {j}}, & i \in \operatorname {T o p K} (\mathbf {P}) \\ 0, & i \notin \operatorname {T o p K} (\mathbf {P}) \end{array} \right. \tag {4}
+$$
+
+where  $\operatorname{TopK}(\mathbf{P})$  returns the indices of the k largest elements in  $\mathbf{P}$ .
+
+# 2.2.2 Top-  $p$  Routing
+
+Top- $p$  routing is initially proposed by (Li et al., 2023a; Huang et al., 2024), which aims to overcome the drawback of Top- $k$  activating the fixed number of experts overlooking the variability in difficulty across different inputs. We refer to readers to Huang et al. (2024) for more details.
+
+Formally, Top-  $p$  first sorts the elements in  $\mathbf{P}$  from highest to lowest, returning in a sorted index list  $I$ . Subsequently, we identify the minimal subset of experts for which the aggregate probability
+
+surpasses the threshold  $p$ . The count of selected experts, denoted as  $t$ , is then determined by:
+
+$$
+t = \underset {k \in \{1 \dots , N \}} {\arg \min } \sum_ {j <   = k} P _ {i, j} \geq p, \tag {5}
+$$
+
+where  $p$  is the threshold that controls how confident the model should be when stopping adding more experts.  $p$  is a hyper-parameter whose range is from 0 to 1. The higher the  $p$  is, the more experts will be activated.
+
+In dynamic routing mechanism, the calculation of  $g_{*}(\mathbf{x})$  is:
+
+$$
+g _ {i} (\mathbf {x}) = \left\{ \begin{array}{l l} P _ {i} & e _ {i} \in S \\ 0, & e _ {i} \notin S \end{array} \right. \tag {6}
+$$
+
+where  $S$  is the set of selected experts controlled by  $t$  in Equation 5:
+
+$$
+S = \left\{e _ {I _ {1}}, e _ {I _ {2}} \dots e _ {I _ {t}} \right\}. \tag {7}
+$$
+
+# 3 THOR-MoE
+
+In this section, we introduce the proposed hierarchical context-responsive routing framework as shown in Figure 1, including two components: hierarchical task-guided routing (§ 3.1) and context-responsive token routing (§ 3.2). Finally, we describe the training objectives (§ 3.3) in detail.
+
+# 3.1 Hierarchical Task-Guided Routing
+
+Our hierarchical task-guided routing consists of three stages, as shown in 2) of 1. The first stage is task prediction, which aims to automatically obtain the task knowledge by predicting which domain/language the input belongs to. The second stage is to inject the mixed task representation to allocates experts in a hierarchical task-guided manner (third stage).
+
+Task Predictor. During real testing for a query, it is hard to reach its domain or language category. Therefore, to automatically obtain the domain knowledge or language knowledge, we add a special token [CLS] in front of the input. Then, it represents the global sentence information and thus can effectively reflect which domain/language the input belongs to (Devlin et al., 2019). Finally, we transform the representation  $\mathbf{H}^{cls}$  into a fixed-size vector through Maxpooling and apply a fully-connected layer to predict suitable labels:
+
+$$
+\begin{array}{l} \mathbf {H} ^ {m a x} = \operatorname {M a x p o o l i n g} \left(\mathbf {H} ^ {c l s}\right), \mathbf {H} ^ {m a x} \in \mathbb {R} ^ {d}, \\ \mathcal {P} ^ {t} = \operatorname {S o f t m a x} (\mathbf {W} ^ {p} \mathbf {H} ^ {m a x}), \\ \end{array}
+$$
+
+where  $\mathbf{W}^p\in \mathbb{R}^{K\times d}$  is trainable weight and  $K$  denotes the number of domains or languages and  $d$  is the model dimension.
+
+Mixed Task Representation. Theoretically, we can directly take the predicted task label to extract corresponding task representation. However, in some cases in practice, it is hard to judge since some sentences belongs to multiple domains. In language cases, there are some code-mixed sentences, which is also confused to decide. Therefore, directly use the given label in existing work or the predicted label cannot accurately help model select the specialized experts for such cases and thus limits their potential, as shown in § 6.2. To address this issue, we use the mixed task representation as follows:
+
+$$
+\mathbf {E} _ {p} = \sum \left(\mathcal {P} ^ {t} \cdot \mathbf {E M B 1}\right), \mathbf {E} _ {p} \in \mathbb {R} ^ {d}, \tag {9}
+$$
+
+where  $\mathbf{E}_p$  denotes the mixed task representation generated by the weighted sum of the task distribution  $\mathcal{P}$  (predicted by the task predictor as Eq. 8) and the task parameter matrix EMB1. The dimension of EMB1 is num (number of tasks/language groups) * d (embedding dimension), which is randomly initialized and then optimized with the training objective.
+
+Hierarchical Task-Guided Routing. With the mixed task representation, we design a task router
+
+$g^{t}$  at the task level. In each MoE layer,  $g^{t}$  takes the mixed task representation  $\mathbf{E}_p$  as input and outputs a task-dependent expert vector. Then, based on it, we use TopK function to select task-specific candidate experts as  $S^t$  from all experts. Finally, based on the selected task-specific candidate experts  $S_{t}$ , we can conduct general token routing at the token level with Top-  $k$  or Top-  $p$  routing as described in § 2.2.1 and § 2.2.2, respectively. The final routing policy can be written as  $\sum_{i}^{|\mathcal{S}^{t}|}\sum_{j}^{|\mathcal{E}|}g_{i}^{t}\cdot g_{j}\cdot e_{i,j}(\mathbf{x})$ .
+
+However, current expert selection at the token level only depends on the current token representation without considering the complete word meaning and global context. Generally, the global context grasp the overall situation and might know which token is difficult or not. Thus, the context can schedules the token to route to suitable experts in a global view. To this end, we propose context-responsive experts selection at the token level.
+
+# 3.2 Context-Responsive Routing
+
+As shown in 3) of 1, before the experts selection at the token level, we explicitly inject the context into the token representation. Since it is hard to determine for a code-mixed input or a query that can belong to several domains, which thus limits its application in practice. Therefore, we use the averaged context representation as the context representation.  $\mathbf{H}_{ctx} = \frac{1}{|Y_{1:t - 1}|}\sum_{t = 1}^{|Y_{1:t - 1}|}\mathbf{h}_t^L$  where  $\mathbf{h}$  is the hidden state of the  $t$ -th token at the  $L$  layer. At each decoding step, the context used for each input token is the prefix of the generated next token. That is, the context is dynamically updated with decoding.
+
+To effectively incorporate the context representation into the routing process, we design a gate to dynamically control the contribution of these information:
+
+$$
+\begin{array}{l} \mathbf {x} _ {i} = g \odot \mathbf {x} _ {i} + (1 - g) \odot \mathbf {H} _ {c t x}, \\ g = \sigma ([ \mathbf {x} _ {i}; \mathbf {H} _ {c t x} ] \mathbf {W} ^ {g} + \mathbf {b} ^ {g}), \\ \end{array}
+$$
+
+where  $\mathbf{W}^g$  and  $\mathbf{b}^g$  are the trainable parameters.
+
+In this manner, we can inject the context information into each token to accurately assign specialized experts to suitable tokens and unlock the potential of the MoE.
+
+# 3.3 Training Objectives
+
+Our training objectives consist of four parts in Top $k$  routing (translation loss  $\mathcal{L}_{\mathrm{NMT}}$ , task prediction loss  $\mathcal{L}_{\mathrm{tp}}$ , load balance loss at the task level  $\mathcal{L}_{bd}$
+
+and token level  $\mathcal{L}_{bt}$  and five parts in Top-  $p$  routing  $(\mathcal{L}_{\mathrm{NMT}},\mathcal{L}_{\mathrm{tp}},\mathcal{L}_{bd},\mathcal{L}_{bt}$  and dynamic routing loss  $\mathcal{L}_d$  in Top-  $p$
+
+Task Prediction Loss. The task prediction loss is defined as:
+
+$$
+\mathcal {L} _ {t p} = \min  \left(- \log \left(\mathcal {P} ^ {t} \left[ t _ {g} \right]\right)\right), \tag {10}
+$$
+
+where  $t_g$  is the golden class label,  $\mathcal{P}^t$  is calculated as in Eq. 8.
+
+Load Balance Loss at the Task Level. To fully encourage the experts to specialize in certain tasks, we propose a task-guided routing loss. It aims to achieve the number of tasks processed by different experts to be roughly the same. Therefore, the task-guided routing loss is defined as:
+
+$$
+\mathcal {L} _ {b d} = N * \sum_ {i = 1} ^ {N} F _ {i} ^ {t} * Q _ {i} ^ {t}, \tag {11}
+$$
+
+where  $F_{i}^{t}$  represents the proportion of tasks selecting expert  $e_{i}$ , and  $Q_{i}^{t}$  denotes the proportion of the router's probability allocated to expert  $e_{i}$ . For  $K$  tasks,  $F_{i}^{t}$  and  $Q_{i}^{t}$  are defined as:  $F_{i}^{t} = \frac{1}{K}\sum_{j=1}^{K} 1\{e_{i} \in S^{t,j}\}$  and  $Q_{i}^{t} = \frac{1}{K}\sum_{j=1}^{K} P_{i}^{j}$  where  $S^{t,j}$  is the set of activated experts for task  $j$ , which is calculated by Equation 7, and  $P^{j}$  is the probability of selecting each expert for task  $j$ , calculated by Equation 3.
+
+Load Balance Loss at the Token Level. This loss is the similar to the vanilla load balance loss with a slight difference that the experts candidates is pre-selected by the hierarchical task-guided routing. Formally, the load balance loss at the token level is defined as:
+
+$$
+\mathcal {L} _ {b t} = \left| \mathcal {S} ^ {t} \right| * \sum_ {k = 1} ^ {\left| \mathcal {S} ^ {t} \right|} F _ {k} ^ {b} * Q _ {k} ^ {b}, \tag {12}
+$$
+
+where  $F_{k}^{b}$  represents the proportion of tokens selecting expert  $e_k$ , and  $Q_{k}^{b}$  denotes the proportion of the router's probability allocated to expert  $e_k$ . For a sequence comprising  $M$  tokens,  $F_{k}^{b}$  and  $Q_{k}^{b}$  are defined as:  $F_{k}^{b} = \frac{1}{M}\sum_{\ell = 1}^{M}1\{e_{k}\in S^{t,\ell}\}$  and  $Q_{k} = \frac{1}{M}\sum_{\ell = 1}^{M}P_{k}^{\ell}$  where  $S^{t,\ell}$  is the set of activated experts for token  $\ell$  from pre-selected experts set  $S^{d,\ell}$ , which is calculated by Equation 7, and  $P^{\ell}$  is the probability of selecting each experts for token  $\ell$ , calculated by Equation 3.
+
+Dynamic Routing Loss in Top- $p$ . The dynamic routing loss aims to prevent dynamic routing from using too many parameters to cheat and losing its ability to selectively choose experts. Therefore, Huang et al. (2024) introduce a constraint on  $\mathbf{P}$  and minimize the entropy of the distribution  $\mathbf{P}$ , ensuring that every token can focus on as less specific experts as possible, which is formalized as:
+
+$$
+\mathcal {L} _ {\text {t o p p}} = - \sum_ {i = 1} ^ {N} P _ {i} * \log \left(P _ {i}\right). \tag {13}
+$$
+
+Final Loss. Our approach can be flexibly applied in Top- $k$  and Top- $p$  routing strategies. Therefore, we have two final loss functions for Top- $k$  and Top- $p$ , respectively. It is a combination of the translation loss, task prediction loss, load balance loss at the task level and at the token level, and dynamic loss:
+
+$$
+\mathcal {J} _ {t o p k} = \mathcal {L} _ {\mathrm {N M T}} + \alpha \mathcal {L} _ {d p} + \beta \mathcal {L} _ {b d} + \gamma \mathcal {L} _ {b p}, \tag {14}
+$$
+
+$$
+\mathcal {J} _ {\text {t o p p}} = \mathcal {J} _ {\text {t o p k}} + \delta \mathcal {L} _ {\text {t o p p}}, \tag {15}
+$$
+
+where  $\alpha, \beta, \gamma,$  and  $\delta$  are hyper-parameters to adjust the contribution among these loss functions, respectively.
+
+# 4 Experiments
+
+# 4.1 Datasets and Metric
+
+Datasets. We use the multi-domain translation dataset proposed by Koehn and Knowles (2017). The dataset mainly covers five diverse domains: IT, Koran, Law, Medical, and Subtitles, which are available in OPUS (Aulamo and Tiedemann, 2019). Following previous work (Gu et al., 2022; Liang et al., 2024), we use the new data splitting released by Aharoni and Goldberg (2020), and perform German to English translation  $(\mathrm{De}\rightarrow \mathrm{En})$ . We use the OPUS-16 for multilingual translation following (Zhao et al., 2024b). The OPUS-16 comes from OPUS-100 (Zhang et al., 2020), which includes 16 languages (8 high resource  $(>0.9\mathrm{M})$ , 4 medium resource, and 4 low resource  $(< 0.1\mathrm{M})$ ). Please refer to Tab. 8 of § 8 for detailed data statistics.
+
+Metric. For a fair comparison, we follow previous work (Gu et al., 2022; Zhao et al., 2024b) and adopt the 4-gram case-sensitive BLEU with the SacreBLEU tool $^{2}$  (Post, 2018) and report the statistical significance test (Koehn, 2004). For
+
+<table><tr><td></td><td>Models</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td rowspan="3">Dense</td><td>SFT-3B</td><td>40.65</td><td>20.4</td><td>51.4</td><td>54.8</td><td>28.33</td><td>39.12</td></tr><tr><td>Trim-MoE (Top-1)</td><td>40.03</td><td>19.55</td><td>51.80</td><td>55.83</td><td>25.50</td><td>38.54</td></tr><tr><td>Trim-MoE (Top-2)</td><td>45.10</td><td>22.68</td><td>51.84</td><td>57.12</td><td>29.02</td><td>41.15</td></tr><tr><td rowspan="4">MoE</td><td>Trim-MoE (Top-p, p=0.5)</td><td>39.39</td><td>19.21</td><td>55.67</td><td>60.18</td><td>29.21</td><td>40.73</td></tr><tr><td>THOR-MoE (Top-1)</td><td>43.96‡</td><td>21.93‡</td><td>52.52†</td><td>56.25</td><td>28.41‡</td><td>40.61‡</td></tr><tr><td>THOR-MoE (Top-2)</td><td>46.00†</td><td>23.35</td><td>55.79‡</td><td>61.06‡</td><td>28.23</td><td>42.89‡</td></tr><tr><td>THOR-MoE (Top-p, p=0.5)</td><td>44.63‡</td><td>22.53‡</td><td>53.58</td><td>58.65</td><td>27.99</td><td>41.48†</td></tr></table>
+
+Table 1: BLEU score on multi-domain translation benchmarks with decoder-only architecture. “†” and “‡” denote that statistically significant better than the best result of the counterpart (e.g., THOR-MoE (Top-2) vs. Trim-MoE (Top-2)) with t-test  $p < 0.05$  and  $p < 0.01$  hereinafter, respectively. The best and second best results are bold and underlined, respectively.
+
+multilingual translation, the scores encompass the average ratings across all language pairs, such as English  $\rightarrow$  Any (En  $\rightarrow$  XX), and Any  $\rightarrow$  English (XX  $\rightarrow$  En) on the OPUS-16 dataset.
+
+# 4.2 Implementation Details
+
+In multi-domain translation, we use decoder-only Transformer architecture. Specifically, we use Qwen1.5-MoE-A2.7B (Team, 2024a). It has 14.3B parameters (60 non shared experts + 4 shared experts) in total and 2.7B activated parameters (4 non shared experts + 4 shared experts) during runtime. Due to limited GPU resource, we used a trimmed version, which has 3.5B parameters (8 non shared experts + 4 shared experts) in total and 2.3B activated parameters (2 non shared experts + 4 shared experts). To keep its capacity, we initialize the trimmed model with Qwen1.5-MoE-A2.7B and denote it as Trim-MoE. During training, we use Llama-Factory (Zheng et al., 2024) to instruct-tune LLMs. All LLMs are tuned on an  $8 \times$  NVIDIA A100 GPUs (40G) with 1e-5 learning rate. We set gradient accumulation to 16 and batch size to 1, which gives us  $2^{*}8^{*}16^{*}1$  batch in total. We use the DeepSpeed optimization (Rasley et al., 2020), and set ZeRO-3 optimization. Following Qin et al. (2024), we set the number of training epochs to 3. We set hyper-parameters  $\alpha$ ,  $\beta$ ,  $\gamma$ , and  $\delta$  to 1e-2, 1e-2, 1e-2, and 1e-4, which are determined by a grid search.
+
+In multilingual translation, we follow (Zhao et al., 2024b) and compare our method with the Transformer-Base model (as Dense) and its MoE variants that have 6 encoder and decoder layers, 32 experts. The input and hidden dimensions of all feed-forward networks are 512 and 2048. We set the training processes to 35K iterations with
+
+a learning rate of 5e-4, which follows the inverse square root with 4,000 warm-up steps. To keep balanced training, we use temperature-based data sampling strategy with temperature 1.5. We set hyper-parameters  $\alpha$ ,  $\beta$ ,  $\gamma$ , and  $\delta$  to 1e-2, 5e-2, 5e-2, and 1e-4, which are determined by a grid search.
+
+# 4.3 Comparison Models
+
+Our comparison models mainly include two types: Decoder-only and Encoder-Decoder based.
+
+Decoder-only. We fine-tune two dense models based on Qwen2.5-3B (Team, 2024b), which have similar parameters with the activated parameter and denoted as SFT-3B. Besides, based on the TrimMoE-A2.3B model, we also fine-tuned three MoE models with Top-1, Top-2, and Top- $p$  routing strategy.
+
+Encoder-Decoder. There are three types of comparison models: vanilla dense model, vanilla MoE model, and enhanced MoE models via language knowledge. We use the Transformer-Base model as Dense model. For vanilla MoE model, the Switch Transformer (Fedus et al., 2022) with a top-1 token-based routing (as ST-MoE) and GShard (Lepikhin et al., 2021a) with a top-2 token-based routing (as GS-MoE) are adopted. For Language-specific MoE models, the LS-MoE with fixed routing inspired by (Pires et al., 2023), assigning 2 non-overlapping experts for tokens according to their source language in the encoder and target language in the decoder; Hybrid-MoE (Kudugunta et al., 2021), with a top-2 token routing in the encoder and a top-2 target language routing in the decoder side; Residual-MoE (Elbayad et al., 2023; Rajbhandari et al., 2022; Zhang et al., 2021) that augments each MoE layer with a shared feed-forward network through a binary gate function; Lingual
+
+MoE (Zhao et al., 2024b) stands for an MoE model with linguistic-guided routing and dynamic expert allocation, where the first-level language router selects the top 8 experts and the second-level token router activates the dynamical number of experts.
+
+# 5 Main Results
+
+Table 1 shows main results on the multi-domain translation benchmark with decoder-only architecture. Table 2 presents main results on the multilingual translation with encoder-decoder architecture. For a fair comparison, all models are trained and assessed on the OPUS-16 dataset with Transformer-Base as the backbone architecture.
+
+# 5.1 Results on Multi-Domain Translation
+
+Table 1 shows that the Trim-MoE-based models generally surpass the dense one. For example, the Trim-MoE (Top-2) outperforms SFT-3B model by average 2 BLEU scores where the activated parameters is less than dense model (2.3B vs. 3B), showing the effectiveness of MoE model. Furthermore, the proposed THOR-MoE significantly and consistently surpasses the counterpart of Trim-MoE-based one. For instance, the THOR-MoE (Top-2) outperforms the Trim-MoE (Top-2) by averaged 1.74 BLEU scores. This clear advantage confirms the superiority of incorporating domain knowledge and context knowledge into the routing. What's important, the proposed approach is compatible with the Top- $k$  (Shazeer et al., 2017a) and Top $p$  (Huang et al., 2024) routing strategies, validating the generalization of THOR-MoE, which can be a plug-and-play module. Besides, we find that the scores on the same domain (e.g., Law) change largely. The reason may be that the data distribution is unbalanced. We also list the results in terms of COMET (Rei et al., 2020) score in Table 3 and we can conclude the similar findings.
+
+# 5.2 Results on Multilingual Translation
+
+Table 2 summarizes the results and we can conclude several observations:
+
+THOR-MoE vs. Dense and Vanilla MoE Baselines. The THOR-MoE substantially surpasses the Dense and vanilla MoE multilingual translation baselines with a large margin. Specifically, compared with Dense and ST-MoE, the THOR-MoE (Top-p) achieves  $\{5.39\%, 3.99\%, 4.70\% \}$  and  $2.46\%, 0.55\%, 1.51\%$  averaged improvement in terms of BLEU score for Avg1., Avg2., and All
+
+Avg., respectively. This significant margin demonstrates the effectiveness of hierarchical language-guided routing and context-responsive routing.
+
+THOR-MoE vs. Language-Guided MoE models. The THOR-MoE also consistently outperforms language-guided MoE models, including LS-MoE, hybrid-MoE, Residual-MoE, and Linguual-MoE. It shows again that the superiority of hierarchical language-guided routing and context-responsive routing. In detail, the Linguual-MoE performs the highest in baselines and it also employs a hierarchical language-group-guided and dynamical routing, which introduces the hard language id embedding and fails to consider code-mixed cases and that the token in different language can be the same (e.g., 'Internet' in German and English). In contrast, THOR-MoE applies a mixed language representation where the language ids are predicted, which comprehensively incorporates the languages knowledge. Furthermore, the dynamical routing in Linguual-MoE does not incorporate the context information. The context generally knows which token is difficult or not in a global view. The THOR-MoE fully considers the above issues and obtains better results.
+
+# 6 Analysis
+
+# 6.1 Ablation Study
+
+We conduct ablation studies to investigate how well hierarchical context-responsive routing of THORMoE works. We conclude two findings from the results in Tab. 4.
+
+(1) "w/o hierarchical task-guided routing": i.e., without using any task-related routing and decaying to vanilla manner (select experts from full set of experts), the model performance greatly degrades on both translation tasks. It shows the necessity of using task-guided routing in a hierarchical manner.  
+(2) "w/o context-responsive routing": the model performance becomes worse on both tasks when removing context. This shows that our context-responsive routing indeed can enhance the routing effectiveness and guarantee the token can be assigned to suitable and specialized experts with the indication of context, which thus benefits the model performance on both translation tasks.
+
+# 6.2 Comparison among Different Task Representations
+
+In this section, we aim to investigate the impact of task representation in different manners. Table 5
+
+<table><tr><td rowspan="2" colspan="2">Models</td><td colspan="4">En→XX</td><td colspan="4">XX→En</td><td rowspan="2">Avg.</td></tr><tr><td>high</td><td>medium</td><td>low</td><td>Avg1.</td><td>high</td><td>medium</td><td>low</td><td>Avg2.</td></tr><tr><td rowspan="5">Dense</td><td>Transformer-base</td><td>25.37</td><td>39.12</td><td>14.78</td><td>26.16</td><td>28.81</td><td>39.24</td><td>24.21</td><td>30.27</td><td>28.21</td></tr><tr><td>GS-MoE</td><td>25.55</td><td>40.71</td><td>18.97</td><td>27.70</td><td>28.77</td><td>41.70</td><td>29.25</td><td>32.12</td><td>29.91</td></tr><tr><td>ST-MoE</td><td>26.55</td><td>43.04</td><td>20.23</td><td>29.09</td><td>29.94</td><td>44.00</td><td>30.94</td><td>33.71</td><td>31.40</td></tr><tr><td>LS-MoE</td><td>20.06</td><td>37.31</td><td>11.72</td><td>22.29</td><td>23.69</td><td>42.99</td><td>32.01</td><td>30.60</td><td>26.44</td></tr><tr><td>Hybird-MoE</td><td>24.56</td><td>31.70</td><td>13.38</td><td>23.55</td><td>29.59</td><td>41.74</td><td>27.81</td><td>32.18</td><td>27.85</td></tr><tr><td rowspan="5">MoE</td><td>Residual-MoE</td><td>26.52</td><td>43.49</td><td>21.58</td><td>29.53</td><td>29.84</td><td>43.94</td><td>31.25</td><td>33.72</td><td>31.62</td></tr><tr><td>Lingual-MoE</td><td>27.11</td><td>46.24</td><td>23.34</td><td>30.95</td><td>29.87</td><td>43.49</td><td>32.02</td><td>33.81</td><td>32.38</td></tr><tr><td>THOR-MoE (Top-1)</td><td>27.21</td><td>46.37</td><td>23.88</td><td>31.17</td><td>29.68</td><td>44.01</td><td>32.58</td><td>33.99</td><td>32.57</td></tr><tr><td>THOR-MoE (Top-2)</td><td>27.88†</td><td>47.29†</td><td>24.85†</td><td>31.98†</td><td>30.38†</td><td>44.52†</td><td>33.26†</td><td>34.64†</td><td>33.31†</td></tr><tr><td>THOR-MoE (Top-p)</td><td>27.63†</td><td>46.82†</td><td>24.15†</td><td>31.55†</td><td>29.97</td><td>44.18†</td><td>32.96†</td><td>34.26</td><td>32.91†</td></tr></table>
+
+Table 2: Averaged BLEU scores on multilingual translation with encoder-decoder architectures. “†” denotes that statistically significant better than Lingual-MoE with t-test  $p < 0.05$ .  
+
+<table><tr><td>Models</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td>SFT-3B</td><td>86.39</td><td>73.17</td><td>85.43</td><td>87.19</td><td>78.52</td><td>82.14</td></tr><tr><td>THOR-MoE (Top-p)</td><td>87.31</td><td>74.09</td><td>86.09</td><td>87.88</td><td>79</td><td>82.87</td></tr></table>
+
+Table 3: COMET scores on multi-domain translation benchmarks with decoder-only architecture.  
+
+<table><tr><td></td><td>Multi-Domain</td><td>Multilingual</td></tr><tr><td>THOR-MoE (Top-p)</td><td>41.48</td><td>32.91</td></tr><tr><td>w/o hierarchical task-guided routing</td><td>40.16</td><td>31.57</td></tr><tr><td>w/o context-responsive routing</td><td>40.74</td><td>32.12</td></tr></table>
+
+shows the results where the 'Non-Mixed Representation' denotes that directly uses the automatically predict task label to extract corresponding task representation and 'Golden Representation' indicates using the golden task label to extract corresponding representation. We conclude that the task knowledge indeed has a positive impact on translation performance (vs. baseline). We also observe that using the automatically predicted task labels (actually the mixed task representation) shows better results than using ground truth and Non-Mixed manner in terms of Avg. BLEU scores. The reason may be that the mixed task representation has certain fault tolerance. We also analyze the accuracy of task prediction to further show its quality
+
+Table 4: Ablation Study with Avg. results.  
+
+<table><tr><td></td><td>Multi-Domain</td><td>Multilingual</td></tr><tr><td>baseline (w/o any task Rep.)</td><td>40.74</td><td>32.12</td></tr><tr><td>Non-Mixed Rep.</td><td>40.95</td><td>32.44</td></tr><tr><td>Golden Rep.</td><td>41.32</td><td>32.78</td></tr><tr><td>THOR-MoE (Mixed manner)</td><td>41.48</td><td>32.91</td></tr></table>
+
+Table 5: The Avg. results of different representation (Rep.) manner during hierarchical task-guided routing.  
+
+<table><tr><td></td><td>Multi-Domain</td><td>Multilingual</td></tr><tr><td>baseline (w/o any task Rep.)</td><td>40.74</td><td>32.12</td></tr><tr><td>Infuse task Rep. into token Rep.</td><td>40.95</td><td>32.44</td></tr><tr><td>THOR-MoE (Hierarchical manner)</td><td>41.48</td><td>32.91</td></tr></table>
+
+Table 6: The investigation of why hierarchical manner.
+
+![](images/7c46b75344c67dc342d848a35ef6d0fc4422e4513c6c5b4e77b20107b57dddb3.jpg)  
+Figure 2: Average activated experts number across training steps on the multi-domain translation task.
+
+in Appendix A.
+
+# 6.3 Why Hierarchical Design?
+
+In this section, we aim to investigate the manner of using mixed task representation. Table 6 shows that introducing task knowledge indeed helps tokens route well and achieve improvement (vs. baseline). The hierarchical manner greatly outperforms the direct infusion with token representation before routing, proving the effectiveness of hierarchical design, which unlock the potential of MoE with the help of the mixed task knowledge.
+
+<table><tr><td>Domains</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td>Top-p</td><td>1.87</td><td>1.95</td><td>1.82</td><td>1.92</td><td>1.77</td><td>1.87</td></tr><tr><td>Top-p w/ ctx</td><td>1.37</td><td>1.61</td><td>1.42</td><td>1.56</td><td>1.31</td><td>1.45</td></tr></table>
+
+Table 7: Average activated experts (AE) in different domain translation tasks. 'Top-  $p$  w/ctx' denotes the proposed method to infuse the context during routing.
+
+# 6.4 Analysis of Efficient Training and Inference
+
+To further explore whether our proposed method is efficient in training and inference, we calculate the average number of experts activated by the model on multi-domain translation tasks. Figure 2 and Table 7 shows the average number of experts activated per token across various translation tasks during training and inference. The result is averaged across all the layers of transformers.
+
+During training, we can see that our method converges faster to use less activated experts than the vanilla Top- $p$  (Huang et al., 2024). During inference, we can observe that across all five tasks, the number of activated experts is less than original Top- $p$  routing, averaging 1.45 activated experts (less than  $22\%$  activated parameters) with better performance. Both findings show that the context plays a key role in guiding token routing.
+
+# 7 Related Work
+
+Neural Machine Translation. The NMT have received remarkable attention in the era of LLMs (Jawahar et al., 2023b; Zhou et al., 2023). Previous work mainly focuses on the continual learning of new domains (Gu and Feng, 2020; Gu et al., 2022; Liang et al., 2024), introducing ready-made task-related (linguistics) knowledge via adapter (Zhang et al., 2020, 2021), parameter sharing (Aharoni et al., 2019), or task-specific modules (Pires et al., 2023), and multilingual representation learning (Yang et al., 2021).
+
+Mixture-of-Experts. Jacobs et al. (1991) first proposes the concept of MoE, which consists of a series of network sub-modules. The Sparately-gated MoE, as a variant, which only activates a few expert networks for each input, has been shown its superiority in diverse NLP and computer vision applications(Shazeer et al., 2017b; Zoph et al., 2022). In the context of NMT, most of previous work focuses on addressing the over/under-fitting (Elbayad et al., 2023) via regularization strategies (Elbayad et al., 2023) or modularizing MoE (Li et al., 2023c;
+
+Zhang et al., 2024), and incorporating task-related knowledge (language, domain etc) (Kudugunta et al., 2021; Zhao et al., 2024b; Pham et al., 2023; Gururangan et al., 2022; Li et al., 2023b), achieving impressive performance. Besides, there are some studies aim to improve the training/inference efficiency of routing via adaptive computation (Jawahar et al., 2023a) or reducing the number of activated experts (Li et al., 2023a; Huang et al., 2024).
+
+Different from these prior works that directly use the task-related knowledge, we propose a hierarchical context-responsive routing method where we automatically extract corresponding domain/language knowledge and design a hierarchical network to guide the task-level routing. Besides, we aim to help each token accurately select specialized and suitable experts and thus we incorporate the context to guide the token routing rather than relying on the localized token only in existing work. In this manner, the context-responsive routing can improve the training or inference efficiency, which is compatible with previous Top- $k$  and Top- $p$  routing policies.
+
+# 8 Conclusion
+
+In this paper, we propose a new hierarchical task-gudied and context-responsive routing framework for NMT. To automatically obtain the task knowledge, we propose to predict it and then use mixed task representation. Consequently, we design a hierarchical routing at the task level and the token level. Further, we propose to inject context to enhance the effectiveness of token routing. Extensive experiments on multi-domain and multilingual translation benchmarks show the superiority and generalization of our proposed approach.
+
+# Limitations
+
+While we introduce the mixed domain/language and context knowledge into routing in hierarchical manner and achieve good results, there are some limitations worth considering to study in future work: (1) In this study, the design relies on the prior (the number of tasks and language groups), which may limit its extension to broader topics (Li and Zhou, 2025); (2) This work only conduct experiments on translation tasks and does not conduct experiments other generation or discriminative tasks.
+
+# References
+
+Roee Aharoni and Yoav Goldberg. 2020. Unsupervised domain clusters in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7747-7763, Online. Association for Computational Linguistics.  
+Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Mikko Aulamo and Jörg Tiedemann. 2019. The OPUS resource repository: An open package for creating parallel corpora and machine translation services. In Proceedings of the 22nd Nordic Conference on Computational Linguistics, NoDaLiDa 2019, Turku, Finland, September 30 - October 2, 2019, pages 389-394.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Maha Elbayad, Anna Sun, and Shruti Bhosale. 2023. Fixing MoE over-fitting on low-resource languages in multilingual machine translation. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 14237-14253, Toronto, Canada. Association for Computational Linguistics.  
+Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, and Armand Joulin. 2020. Beyond english-centric multilingual machine translation. CoRR, abs/2010.11125.  
+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.  
+Fabian Gloeckle, Badr Youbi Idrissi, Baptiste Roziere, David Lopez-Paz, and Gabriel Synnaeve. 2024. Better & faster large language models via multi-token prediction. Preprint, arXiv:2404.19737.  
+Shuhao Gu and Yang Feng. 2020. Investigating catastrophic forgetting during continual training for neural machine translation. In Proceedings of the 28th International Conference on Computational Linguistics,
+
+pages 4315-4326, Barcelona, Spain (Online). International Committee on Computational Linguistics.  
+Shuhao Gu, Bojie Hu, and Yang Feng. 2022. Continual learning of neural machine translation within low-forgetting risk regions. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1707-1718, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Suchin Gururangan, Mike Lewis, Ari Holtzman, Noah A. Smith, and Luke Zettlemoyer. 2022. DEMix layers: Disentangling domains for modular language modeling. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5557-5576, Seattle, United States. Association for Computational Linguistics.  
+Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. 2024. Harder task needs more experts: Dynamic routing in MoE models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12883-12895, Bangkok, Thailand. Association for Computational Linguistics.  
+Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. 1991. Adaptive mixtures of local experts. Neural Comput., 3(1):79-87.  
+Ganesh Jawahar, Subhabrata Mukherjee, Xiaodong Liu, Young Jin Kim, Muhammad Abdul-Mageed, Laks Lakshmanan, V.S., Ahmed Hassan Awadallah, Sebastian Bubeck, and Jianfeng Gao. 2023a. AutoMoE: Heterogeneous mixture-of-experts with adaptive computation for efficient neural machine translation. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9116-9132, Toronto, Canada. Association for Computational Linguistics.  
+Ganesh Jawahar, Subhabrata Mukherjee, Xiaodong Liu, Young Jin Kim, Muhammad Abdul-Mageed, Laks V. S. Lakshmanan, Ahmed Hassan Awadallah, Sébastien Bubeck, and Jianfeng Gao. 2023b. Automoe: Heterogeneous mixture-of-experts with adaptive computation for efficient neural machine translation. In *Findings of the Association for Computational Linguistics: ACL 2023*, Toronto, Canada, July 9-14, 2023, pages 9116-9132. Association for Computational Linguistics.  
+Songtao Jiang, Tuo Zheng, Yan Zhang, Yeying Jin, Li Yuan, and Zuozhu Liu. 2024. Med-MoE: Mixture of domain-specific experts for lightweight medical vision-language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2024, pages 3843-3860, Miami, Florida, USA. Association for Computational Linguistics.  
+Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of EMNLP, pages 388-395.
+
+Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, NMT@ACL 2017, Vancouver, Canada, August 4, 2017, pages 28-39.  
+Sneha Kudugunta, Yanping Huang, Ankur Bapna, Maxim Krikun, Dmitry Lepikhin, Minh-Thang Luong, and Orhan First. 2021. Beyond distillation: Task-level mixture-of-experts for efficient inference. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3577-3599, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2021a. {GS}hard: Scaling giant models with conditional computation and automatic sharding. In International Conference on Learning Representations.  
+Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2021b. Gshard: Scaling giant models with conditional computation and automatic sharding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.  
+Jiamin Li, Qiang Su, Yitao Yang, Yimin Jiang, Cong Wang, and Hong Xu. 2023a. Adaptive gating in mixture-of-experts based language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3577-3587, Singapore. Association for Computational Linguistics.  
+Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A. Smith, and Luke Zettle-moyer. 2023b. Branch-train-merge: Embarrassingly parallel training of expert language models.  
+Shangjie Li, Xiangpeng Wei, Shaolin Zhu, Jun Xie, Baosong Yang, and Deyi Xiong. 2023c. MMNMT: Modularizing multilingual neural machine translation with flexibly assembled MoE and dense blocks. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4978-4990, Singapore. Association for Computational Linguistics.  
+Ziyue Li and Tianyi Zhou. 2025. Your mixture-of-experts LLM is secretly an embedding model for free. In The Thirteenth International Conference on Learning Representations.  
+Yunlong Liang, Fandong Meng, Jiaan Wang, Jinan Xu, Yufeng Chen, and Jie Zhou. 2024. Continual learning with semi-supervised contrastive distillation for incremental neural machine translation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10914-10928, Bangkok, Thailand. Association for Computational Linguistics.
+
+Hai Pham, Young Jin Kim, Subhabrata Mukherjee, David P. Woodruff, Barnabas Poczos, and Hany Hassan. 2023. Task-based MoE for multitask multilingual machine translation. In Proceedings of the 3rd Workshop on Multi-lingual Representation Learning (MRL), pages 164–172, Singapore. Association for Computational Linguistics.  
+Telmo Pires, Robin Schmidt, Yi-Hsiu Liao, and Stephan Peitz. 2023. Learning language-specific layers for multilingual machine translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14767-14783, Toronto, Canada. Association for Computational Linguistics.  
+Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of WMT, pages 186-191.  
+Yiwei Qin, Xuefeng Li, Haoyang Zou, Yixiu Liu, Shijie Xia, Zhen Huang, Yixin Ye, Weizhe Yuan, Hector Liu, Yuanzhi Li, et al. 2024. O1 replication journey: A strategic progress report-part 1. arXiv preprint arXiv:2410.18982.  
+Samyam Rajbhandari, Conglong Li, Zhewei Yao, Minjia Zhang, Reza Yazdani Aminabadi, Ammar Ahmad Awan, Jeff Rasley, and Yuxiong He. 2022. Deepspeed-moe: Advancing mixture-of-experts inference and training to power next-generation ai scale. Preprint, arXiv:2201.05596.  
+Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 3505-3506.  
+Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702, Online. Association for Computational Linguistics.  
+Noam Shazeer, *Azalia Mirhoseini, *Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017a. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations.  
+Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. 2017b. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.  
+Qwen Team. 2024a. Qwen1.5-moe: Matching 7b model performance with 1/3 activated parameters".  
+Qwen Team. 2024b. Qwen2.5: A party of foundation models.
+
+An Wang, Xingwu Sun, Ruobing Xie, Shuaipeng Li, Ji- aqi Zhu, Zhen Yang, Pinxue Zhao, J. N. Han, Zhanhui Kang, Di Wang, Naoaki Okazaki, and Cheng zhong Xu. 2024. Hmoe: Heterogeneous mixture of experts for language modeling. Preprint, arXiv:2408.10681.  
+Junhong Wu, Yuchen Liu, and Chengqing Zong. 2024. F-MALLOC: Feed-forward memory allocation for continual learning in neural machine translation. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7180-7192, Mexico City, Mexico. Association for Computational Linguistics.  
+Yilin Yang, Akiko Eriguchi, Alexandre Muzio, Prasad Tadepalli, Stefan Lee, and Hany Hassan. 2021. Improving multilingual translation by representation and gradient regularization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7266-7279, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Zihao Zeng, Yibo Miao, Hongcheng Gao, Hao Zhang, and Zhijie Deng. 2024. AdaMoE: Token-adaptive routing with null experts for mixture-of-experts language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6223-6235, Miami, Florida, USA. Association for Computational Linguistics.  
+Biao Zhang, Ankur Bapna, Rico Sennrich, and Orhan First. 2021. Share or not? learning to schedule language-specific capacity for multilingual translation. In International Conference on Learning Representations.  
+Biao Zhang, Philip Williams, Ivan Titov, and Rico Sennrich. 2020. Improving massively multilingual neural machine translation and zero-shot translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1628-1639, Online. Association for Computational Linguistics.  
+Fan Zhang, Mei Tu, Song Liu, and Jinyao Yan. 2024. A lightweight mixture-of-experts neural machine translation model with stage-wise training strategy. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2381-2392, Mexico City, Mexico. Association for Computational Linguistics.  
+Hao Zhao, Zihan Qiu, Huijia Wu, Zili Wang, Zhaofeng He, and Jie Fu. 2024a. HyperMoE: Towards better mixture of experts via transferring among experts. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10605-10618, Bangkok, Thailand. Association for Computational Linguistics.  
+Xinyu Zhao, Xuxi Chen, Yu Cheng, and Tianlong Chen. 2024b. Sparse moe with language guided routing for
+
+multilingual machine translation. In The Twelfth International Conference on Learning Representations.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. LlamaFactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 400-410, Bangkok, Thailand. Association for Computational Linguistics.  
+Chulun Zhou, Yunlong Liang, Fandong Meng, Jinan Xu, Jinsong Su, and Jie Zhou. 2023. RC3: Regularized contrastive cross-lingual cross-modal pre-training. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 11747–11762, Toronto, Canada. Association for Computational Linguistics.  
+Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. 2022. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906.  
+Simiao Zuo, Xiaodong Liu, Jian Jiao, Young Jin Kim, Hany Hassan, Ruofei Zhang, Jianfeng Gao, and Tuo Zhao. 2022. Taming sparsely activated transformer with stochastic experts. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.
+
+<table><tr><td></td><td>Tasks/Groups</td><td>Train</td><td>Valid</td><td>Test</td></tr><tr><td rowspan="5">Multi-Domain Translation Dataset (De→En)</td><td>IT</td><td>0.22M</td><td></td><td></td></tr><tr><td>Koran</td><td>18K</td><td></td><td></td></tr><tr><td>Law</td><td>0.47M</td><td>2000</td><td>2000</td></tr><tr><td>Medical</td><td>0.25M</td><td></td><td></td></tr><tr><td>Subtitles</td><td>0.5M</td><td></td><td></td></tr><tr><td>Multilingual Translation Dataset</td><td>16</td><td>17,559,950</td><td>30*1000</td><td>30*1000</td></tr></table>
+
+Table 8: The data statistic of the multi-domain translation dataset and multilingual translation (OPUS-16) dataset. The number in Train/Valid/Test columns denotes the number of sentence pairs in each domain/language pair.  
+
+<table><tr><td>Models</td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td><td>Avg.</td></tr><tr><td>Qwen1.5-MoE-A2.7B (SFT, Top-8)</td><td>45.59</td><td>23.31</td><td>55.67</td><td>60.18</td><td>29.21</td><td>42.79</td></tr><tr><td>THOR-MoE (Top-p, p=0.5)</td><td>46.76</td><td>24.03</td><td>35.9</td><td>61.13</td><td>30.29</td><td>43.62</td></tr><tr><td>Qwen1.5-MoE-A2.7B (SFT, Top-8)</td><td>87.3</td><td>74.24</td><td>85.85</td><td>87.78</td><td>79.36</td><td>82.91</td></tr><tr><td>THOR-MoE (Top-p, p=0.5)</td><td>87.45</td><td>74.58</td><td>86.45</td><td>88.75</td><td>80.23</td><td>83.49</td></tr></table>
+
+# A Task Prediction
+
+We also evaluate the performance of the Task Predictor to show whether the classifier can accurately predict suitable labels. The results are  $82.45\%$  and  $64.89\%$  for domain and language prediction, respectively. It suggests that our classifier can predict suitable labels and further provide effective mixed task representation for routing.
+
+# B Comparison to Larger Model
+
+We implemented our proposed method based on a large-scale Qwen1.5-MoE-A2.7B (totally 14B, activated 2.7B, denoted as THOR-MoE). Besides, we also fine-tuned the vanilla Qwen1.5-MoE-A2.7B model as the baseline. The averaged COMET and BLEU results shown in Table 9 demonstrate that the THOR-MoE model at the 14B level also achieve better results than fine-tuned Qwen1.5-MoE-A2.7B model in terms of BLEU and COMET scores.
+
+# C Analysis of Domain Similarity
+
+Following previous work (Wu et al., 2024), we conduct some analysis for domain similarity. Specifically, we sum the predicted soft label and then normalize it. The vertical axis and horizontal axis are unsupervised cluster label (clustering with BERT
+
+Table 9: BLEU (top block) / COMET (below block) scores on multi-domain translation benchmarks with decoder-only architecture. Note that top-8 means it activates 4 shared experts and 4 non-shared experts out of 60 experts.  
+
+<table><tr><td></td><td>IT</td><td>Koran</td><td>Medical</td><td>Law</td><td>Subtitles</td></tr><tr><td>IT</td><td>0.833</td><td>0.027</td><td>0.055</td><td>0.0219</td><td>0.065</td></tr><tr><td>Koran</td><td>0.012</td><td>0.817</td><td>0.003</td><td>0.019</td><td>0.149</td></tr><tr><td>Medical</td><td>0.111</td><td>0.004</td><td>0.737</td><td>0.135</td><td>0.013</td></tr><tr><td>Law</td><td>0.142</td><td>0.008</td><td>0.032</td><td>0.798</td><td>0.02</td></tr><tr><td>Subtitles</td><td>0.058</td><td>0.076</td><td>0.009</td><td>0.006</td><td>0.851</td></tr></table>
+
+Table 10: BLEU (top block) / COMET (below block) scores on multi-domain translation benchmarks with decoder-only architecture. Note top-8 means it activates 4 shared experts and 4 non-shared experts out of 60 experts.
+
+base and  $k = 5$  ) and predicted soft labels, respectively. The similarity matrix are shown in Table 10.
+
+The results show that there exist some sentences that are assigned to a cluster of another domain (i.e., hard cases). This makes sense as the mixed representation can capture such domain information, making the proposed approach works (the reason).

2025/THOR-MoE_ Hierarchical Task-Guided and Context-Responsive Routing for Neural Machine Translation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b030181c0d458ee1a249eb7a65010c0376abf91be09eb9fcd6339cf8d8670fd6
+size 488140

2025/TRACT_ Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge/142e9116-0830-41df-88fb-0f4abce5fead_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c735cf6337165b3d3f1d66b43de26d83a2516ab231cbc713cde1c46bbd2be781
+size 798380

2025/TRACT_ Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge/full.md

@@ -0,0 +1,549 @@
+# TRACT: Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge
+
+Cheng-Han Chiang
+
+NTU GICE
+
+Taiwan
+
+dcml0714@gmail.com
+
+Hung-yi Lee
+
+NTUGICE
+
+Taiwan
+
+hungyilee@ntu.edu.tw
+
+Michal Lukasik
+
+Google Research
+
+New York
+
+mlukasik@google.com
+
+# Abstract
+
+The LLM-as-a-judge paradigm uses large language models (LLMs) for automated text evaluation, which assigns a score to the text based on some scoring rubrics. Existing methods for LLM-as-a-judge use cross-entropy (CE) loss for fine-tuning, which neglects the numeric nature of score prediction. Recent work addresses numerical prediction limitations of LLM fine-tuning through regression-aware fine-tuning, which, however, does not consider chain-of-thought (CoT) reasoning for score prediction. In this paper, we introduce TRACT (Two-stage Regression-Aware fine-tuning with CoT), a method combining CoT reasoning with regression-aware training. The training objective of TRACT combines the CE loss for learning the CoT reasoning and the regression-aware loss for the score prediction. TRACT consists of two stages: first, a seed LLM is fine-tuned to generate CoTs; next, we retrain the seed LLM using the CoTs generated by the LLM trained in stage 1. Experiments across four LLM-as-a-judge datasets and two LLMs show that TRACT significantly outperforms existing methods. Extensive ablation studies validate the importance of each component in TRACT. $^{1}$
+
+# 1 Introduction
+
+Large Language Models (LLMs) have been applied to evaluate written based on the fine-grained evaluation rubrics specified in the input by outputting a score indicating the quality (Chiang and Lee, 2023a). The fine-grained evaluation rubrics define the criteria for scoring answers, and each instance to be evaluated can have a set of customized fine-grained evaluation rubrics. This LLM-as-a-judge paradigm has become the standard way to evaluate LLMs (Zheng et al., 2023; Dubois et al., 2024; Abdin et al., 2024; Lambert et al., 2025), and has applications across diverse scenarios (Yuan
+
+![](images/304fddb0ccec2057b1be7dfdb086b7272545d30fbc7750ea16118903a9fa88fd.jpg)  
+(a) CoT-RAFT fine-tuning objective (§3.2).
+
+![](images/7a012221cea9c567fef42ac3266871cfcdbadb9458c98b3b305e137829eaae16.jpg)  
+(b) TRACT algorithm (§3.3).  
+Figure 1: TRACT method overview. (a) Illustration of the CoT-RAFT fine-tuning objective (Eq. 4), used in both stages of TRACT. (b) Two fine-tuning stages of TRACT (also see Algorithm 1). Stage 1: model  $p_{\mathrm{s}}$  is trained over the ground truth scores and the annotation CoTs (generated by the annotation model  $p_{\mathrm{a}}$ ). Stage 2: CoT supervision is sampled from  $p_{\mathrm{s}}$  (frozen at this stage) and used to fine-tune the final model  $p_{\mathrm{tract}}$ .
+
+et al., 2024; Chiang et al., 2024). To induce the fine-grained assessment ability of LLMs, LLM-as-a-judge is typically trained with cross-entropy (CE) loss to predict chain-of-thought (CoT) reasoning about the evaluation, followed by a score (Kim et al., 2024a,b; Li et al., 2024).
+
+Although fine-tuning LLM-as-a-judge with CE loss is a well-established practice (Kim et al., 2024a,b), it has intuitive limitations when applied to predicting numerical targets. For example, given a ground truth score '1', placing high probability on the token '5' is penalized the same as placing high probability on the token '2', despite the fact that they induce very different errors in terms of numerical metrics such as squared error. To miti-
+
+gate this, Lukasik et al. (2025) proposed regression-aware fine-tuning (RAFT; see §2.3), a method using squared error loss in fine-tuning LLMs.
+
+RAFT has shown promising results on regression tasks; however, it does not incorporate CoT reasoning (Wei et al., 2022; Kojima et al., 2022). Chiang and Lee (2023b) show that not using CoT can harm the performance of LLM-as-a-judge. This makes us ask: Can CoT reasoning be effectively integrated into regression-aware fine-tuning?
+
+In this paper, we propose Two-stage Regression-Aware fine-tuning with CoT reasoning (TRACT), a method that combines the benefits of CoT reasoning with RAFT for improved numerical prediction abilities.. In stage 1, TRACT fine-tunes the seed LLM over both the CoT supervision and the ground truth scores in the original training data. TRACT's fine-tuning objective (which we refer to as CoT-RAFT) is the sum of CE loss over the CoT, and the RAFT loss over score prediction (see Figure 1a). In stage 2, TRACT fine-tunes the seed LLM from the CoTs generated by the model trained in stage 1 using the same CoT-RAFT objective. This two-stage pipeline is for reducing the distribution between the seed LLM and the CoT supervision used for fine-tuning in stage 2 (see Figure 1b). We summarize TRACT in Algorithm 1 and contrast its formulation against baselines in Table 1.
+
+We conduct experiments on five datasets and two LLMs and show that TRACT consistently outperforms baselines by a large margin, including Prometheus-2-7B (Kim et al., 2024b), which is the best model of equal size from prior works. Our contributions are as follows:
+
+(i) We propose TRACT, a fine-tuning method that combines CoT reasoning with regression-aware loss to improve the numerical scoring ability of LLMs (§3).  
+(ii) We show that TRACT yields significant performance improvements over existing baselines for LLM-as-a-judge. We validate the importance of each component in TRACT with extensive ablations (§5, §6).  
+(iii) We show that TRACT, while trained on pointwise LLM-as-a-judge datasets, can work well on RewardBench (Lambert et al., 2024), a pairwise comparison dataset ( $\S 7$ ).  
+(iv) We release our models. Our best model surpasses the prior state-of-the-art model of comparable size by 0.05 in Pearson's correlation when the inference-time compute is limited.
+
+# 2 Background
+
+We first introduce the notation and next review previous works on regression-aware numerical output prediction with autoregressive models.
+
+# 2.1 Notation
+
+Let  $\mathcal{V}$  denote a finite vocabulary of tokens,  $\mathcal{X} \subseteq \mathcal{V}^*$  be a set of inputs comprising strings of tokens,  $S \subseteq \mathcal{V}^*$  the set of CoT reasonings, and  $\mathcal{Y} \subset \mathbb{R}$  be a set of numerical targets. We assume that each  $y \in \mathcal{V}$  has a unique string representation  $\operatorname{str}(y) \in \mathcal{V}^*$ . Let  $\mathbb{P}$  denote a ground-truth distribution over  $\mathcal{X} \times S \times \mathcal{Y}$ , with the decomposition  $\mathbb{P}(x, s, y) = \mathbb{P}(x) \cdot \mathbb{P}(s \mid x) \cdot \mathbb{P}(y \mid [x, s])$ , where  $[x, s]$  denotes a string concatenation of  $x$  and  $s$ . Denote the training data by  $D_{\mathrm{train}} \in (\mathcal{X} \times \mathcal{Y})^N$ , and each element in  $D_{\mathrm{train}}$  is composed of  $(x, y^*)$ . Denote the training data augmented with CoT by  $D_{\mathrm{train}}^{\mathrm{CoT}} \in (\mathcal{X} \times S \times \mathcal{Y})^N$ , and each element in  $D_{\mathrm{train}}^{\mathrm{CoT}}$  is composed of  $(x, s^*, y^*)$ , where  $s^*$  can come from the human or LLM annotation.
+
+In this paper, we focus on LLM-as-a-judge application; in this case, the input  $x \in \mathcal{X}$  consists of the instructions for LLM as a judge, the scoring rubrics and a sample to be evaluated. The CoT reasoning  $s$  contains an analysis of the sample in the inputs, and the target space is  $\mathcal{Y} = \{1,2,3,4,5\}$ . An example input-output pair is shown in Table 4 in the Appendix.
+
+We denote an LLM by the probability distribution that it models given the input, that is,  $p(\cdot \mid x)$ . We use subscripts to denote specific LLM variants:  $p_0$  represents the seed LLM before fine-tuning, and  $p_{\mathrm{a}}$  represents the LLM annotator used to generate the CoTs in  $D_{\mathrm{train}}^{\mathrm{CoT}}$ .
+
+In the following sections, we describe different methods for training and inference of predictor  $\hat{y}$ .
+
+# 2.2 Standard Fine-tuning and Decoding with No CoT
+
+Fine-tuning without CoT seeks to adapt a seed LLM  $p_0(\cdot |x)$  to the target distribution  $\mathbb{P}$  by minimizing a suitable loss function  $\ell \colon \mathcal{V}\times \Delta_{\mathcal{V}^*}\to \mathbb{R}$ , where  $\Delta_{\mathcal{V}^*}$  denotes the set of distributions over a set  $V^{*}$ . A standard choice of  $\ell$  is the cross-entropy loss:  $\ell_{\mathrm{CE}}(y^{*},p) = -\log p(\mathsf{str}(y^{*})|x)$ , where  $\mathsf{str}(\cdot)$  converts a numerical target  $y^{*}$  to its string representation. Inference is typically conducted by standard (autoregressive) decoding, approximately seeking the mode of the distribution  $p(\cdot |x)$ , i.e.,  $\hat{y}_{\mathrm{mode}}(x)\coloneqq \arg \max_{y\in \mathcal{V}}p(y|x)$ .
+
+<table><tr><td>Category</td><td>Approach</td><td>Predictor function帽子(x)</td><td>Fine-tuning loss</td><td>Sec</td></tr><tr><td rowspan="4">Baselines</td><td>Standard fine-tuning and decoding w/o CoT</td><td>arg maxy∈y p(y|x)</td><td>- logp(y*|x)</td><td>§2.2</td></tr><tr><td>RAIL zero-shot (Lukasik et al., 2024)</td><td>∑y′∈y y′·p(y&#x27;|x)</td><td>N/A</td><td>§2.3</td></tr><tr><td>RAFT (Lukasik et al., 2025)</td><td>∑y′∈y y′·p(y|x)</td><td>(ˆy(x) - y*)2</td><td>§2.3</td></tr><tr><td>Standard fine-tuning and decoding w/ CoT</td><td>arg maxy∈y p(str(y) | [x,hat]);帽子ˆs~p(ˆx)</td><td>- logp([s*,y*] | x)</td><td>§3.1</td></tr><tr><td>Ours</td><td>TRACT</td><td>∑y∈y p(str(y) | [x,hat])·y;帽子ˆs~p(ˆx)</td><td>Algorithm 1</td><td>§3</td></tr></table>
+
+Table 1: Approaches to autoregressive regression with decoder-based LLMs. Here,  $p(\cdot \mid x)$  denotes a distribution over possible outputs given an input string  $x$ , and  $\hat{y}(x) \in \mathbb{R}$  a predictor function.  $\mathcal{V}$  denotes the targets space (unless otherwise stated,  $\{1,2,3,4,5\}$ ).
+
+# 2.3 Regression-Aware Inference and Fine-tuning
+
+Recently, Lukasik et al. (2024) noted that standard autoregressive decoding in LLMs implicitly minimizes the 0-1 error metric,  $m(y, \hat{y}) = \mathbb{1}(y \neq \hat{y})$ , making it suboptimal for regression metrics. Motivated by the minimum Bayes risk decoding framework (Kumar and Byrne, 2004), Lukasik et al. (2024) propose Regression-Aware Inference for Language models (RAIL). RAIL aims to determine the Bayes-optimal prediction by minimizing the expected loss:  $\hat{y}_{\mathrm{RAIL}}(x) \doteq \arg \min_{v \in \mathbb{R}} \mathbb{E}_{y \sim p(\cdot | x)} [m(\mathsf{num}(y), v)]$ , where  $m$  is the metric of interests and  $\mathsf{num}(\cdot)$  converts a string output to its numerical equivalent. For squared error loss,  $m(y, \hat{y}) = (y - \hat{y})^2$ , the optimal decision rule simplifies to the expected value of the numerical predictions. In the case of finite numerical targets  $\mathcal{V}$ , this can be exactly calculated by scoring and averaging a set of candidate targets:
+
+$$
+\hat {y} _ {\text {R A I L}} (x) \doteq \sum_ {y \in \mathcal {Y}} p (\operatorname {s t r} (y) \mid x) \cdot y. \tag {1}
+$$
+
+While RAIL is based on the Bayes-optimal decision rule, directly applying this rule to a model finetuned with CE loss can still result in high regression error. This is due to the misalignment between the CE fine-tuning loss and the regression evaluation metrics (Lukasik et al., 2025). To address this, Lukasik et al. (2025) proposed incorporating the RAIL optimal decision rule directly into the finetuning process using squared loss, leading to the Regression-Aware Fine-Tuning (RAFT) loss:
+
+$$
+\ell_ {\mathrm {R A F T}} (y ^ {*}, p) = \left(y ^ {*} - \hat {y} _ {\mathrm {R A I L}} (x)\right) ^ {2}. \tag {2}
+$$
+
+# 3 Method: TRACT
+
+While achieving notable success on natural language regression tasks such as STS-B (Cer et al.,
+
+2017), RAIL and RAFT do not use CoT reasoning. In this section, we start by discussing learning with CoT, and next introduce our method that incorporates CoT into RAIL and RAFT.
+
+# 3.1 Learning and Inference with CoT
+
+LLM-as-a-judge typically decodes a CoT sequence prior to outputting the final score (Chiang and Lee, 2023b). LLM-as-a-judge is fine-tuned over a dataset  $D_{\mathrm{train}}^{\mathrm{CoT}} \in (\mathcal{X} \times \mathcal{S} \times \mathcal{Y})^N$  of  $N$  (input, CoT, target score) tuples drawn from  $\mathbb{P}$ . The CoT annotations  $s^*$  often come from a powerful LLM (e.g. GPT-4); we denote the distribution of CoT annotations provided in the dataset as  $p_{\mathrm{a}}(\cdot | x)$ . The empirical loss for fine-tuning is  $\hat{L}_{\mathrm{CoT}}(p) = \frac{1}{N} \sum_{(x, y^*, s^*) \in D_{\mathrm{train}}^{\mathrm{CoT}}} \ell(y^*, s^*, p)$ .
+
+A typical choice is again the CE loss, i.e.,  $\ell_{\mathrm{CoT}}(y^{*},s^{*},p) = -\log p([s^{*},y^{*}]|x)$ . Inference is typically done by generatively predicting a CoT reasoning and the score, which samples from the distribution  $p$ , thus approximating the mode of the distribution:  $\hat{y}_{\mathrm{CoT~mode}}(x)\doteq \arg \max_{y\in \mathcal{Y}}p(\mathsf{str}(y)\mid [x,\hat{s}])$ ;  $\hat{s}\sim p(\cdot \mid x)$ .
+
+# 3.2 Regression-Aware Fine-Tuning with CoT
+
+CoT reasoning elicits a reasoning  $s$  before outputting the score  $y$ . We propose and define the CR (CoT-RAIL) predictor as first sampling a CoT  $\hat{s}$  conditioning on the input  $x$ , and then applying the RAIL predictor when conditioning on  $x$  and  $\hat{s}$ ,
+
+$$
+\hat {y} _ {\mathrm {C R}} (x) \doteq \sum_ {y \in \mathcal {Y}} p (\mathbf {s t r} (y) \mid [ x, \hat {s} ]) \cdot y; \hat {s} \sim p (\cdot \mid x) \tag {3}
+$$
+
+A practical question is how to determine where the CoT reasoning ends and where the score starts so that the RAIL predictor can be used at the appropriate position. In practice, we assume each CoT ends with the string  $d =$  'So the overall score is'. We include the string  $d$  at the end of each CoT in
+
+the training set to ensure that the fine-tuned model always ends its CoT with  $d$ .
+
+In principle, we can sample a CoT and obtain the score prediction using Equation 3, repeat the CoT sampling multiple times to estimate the score, and average the scores. Unless stated otherwise, we sample  $\hat{s}$  once, as shown in Equation 3.
+
+For training, analogously to the original RAFT algorithm, we use the CoT-RAIL predictor in the squared error loss. To guide the model in generating CoTs, we apply CE loss to the CoTs. We combine the two losses using a weighting coefficient  $\lambda$ , forming the CoT-RAFT objective:
+
+$$
+\begin{array}{l} \ell_ {\mathrm {C o T - R A F T}} ^ {\lambda} \left(y ^ {*}, p _ {\mathrm {t}}, p\right) = \\ \lambda \left(\sum_ {y \in \mathcal {Y}} p (\operatorname {s t r} (y) \mid [ x, \hat {s} ]) \cdot y - y ^ {*}\right) ^ {2} \tag {4} \\ - \log p ([ \hat {s}, y ^ {*} ] \mid x); \hat {s} \sim p _ {\mathrm {t}} (\cdot \mid x), \\ \end{array}
+$$
+
+where  $p_{\mathrm{t}}$  denotes the (target) model used to generate the CoTs for training, which can be  $p_{\mathrm{a}}$  (GPT-4), but we explore other options next.
+
+# 3.3 Self-Generated Chain-of-Thoughts (CoTs)
+
+When training with CoT-RAFT (Eq. 4), the score predictor is conditioned on the CoT generated by a target model, denoted as  $\hat{s} \sim p_{\mathrm{t}}(\cdot \mid x)$ . However, at inference time, CoT-RAIL (Eq. 3) relies on self-generated CoTs from the fine-tuned model  $p$ , i.e.,  $\hat{s} \sim p(\cdot \mid x)$ . In our case, the CoTs used for training are from GPT-4 ( $p_{\mathrm{a}}$ ) and may be very different from the CoTs generated by the fine-tuned model  $p$ . This creates a mismatch between the CoTs used during training and inference, which can harm the performance, as empirically shown in §5.
+
+To mitigate the distribution shift in annotation and model CoTs, we propose to fine-tune the model using CoTs more closely aligned with those sampled from the trained model  $p(\cdot \mid x)$ . Let  $p_{\mathrm{s}}$  denote the model initialized from the seed LLM  $p_0$  and fine-tuned on the annotation CoTs sampled from  $p_{\mathrm{a}}$ . For each input  $x$  in the training data, we prompt  $p_{\mathrm{s}}$  to generate a response containing a CoT and a score prediction. We then discard the score prediction and pair the generated CoT with the corresponding original ground truth score, creating a new training dataset, denoted as  $D_{\mathrm{self}}$ . Formally,  $D_{\mathrm{self}} = \{(x,\mathrm{CoT}_{\mathrm{s}}(x),y^{*})|(x,y^{*})\in D_{\mathrm{train}},\mathrm{CoT}_{\mathrm{s}}(x)\sim p_{\mathrm{s}}(\cdot |x)\}$ , where  $D_{\mathrm{train}}$  is the original training dataset,  $y^{*}$  is the ground truth
+
+score, and  $\mathrm{CoT_s}(x)$  is the CoT generated by  $p_{\mathrm{s}}$  for input  $x$ .
+
+We use  $D_{\mathrm{self}}$  to train from the seed LLM  $p_0$  to obtain the final TRACT model,  $p_{\mathrm{tract}}$ . The complete algorithm for training the TRACT model is shown in Algorithm 1 and Figure 1. In summary, we first train the seed LLM on the training data with the CoT annotation from  $p_{\mathrm{a}}$  to obtain  $p_{\mathrm{s}}$ , use  $p_{\mathrm{s}}$  to generate CoTs to form  $D_{\mathrm{self}}$ , and use  $D_{\mathrm{self}}$  to retrain  $p_0$  again, leading to the resulting model  $p_{\mathrm{tract}}$ .
+
+# 4 Experiment Setup
+
+Models We fine-tune TRACT from two LLMs, Mistral-7B-Instruct-v0.2 (Jiang et al., 2023) and Llama-3.1-8B-Instruct (Grattafori et al., 2024).
+
+Training Dataset We follow Kim et al. (2024b) and use Feedback Collection (Kim et al., 2024a) as the training set. The training set contains roughly 100K samples. The whole dataset, including the samples to be evaluated, the evaluation responses, and ground truth scores, are generated by GPT-4. We follow the procedure in §3.3 to construct  $D_{\mathrm{self}}$ . We analyze the quality of the self-generated CoTs in Appendix B.1.
+
+Test Datasets We use four datasets for point-wise LLM-as-a-judge, following Kim et al. (2024b).
+
+(1) Feedback Bench (Kim et al., 2024a): This is the official test set of Feedback Collection. Feedback Bench contains 1K responses to evaluate. The instructions and scoring rubrics in Feedback Bench do not overlap with the training set.  
+(2) FLASK (Ye et al., 2024): This is a fine-grained evaluation benchmark with 200 test prompts and 2000 responses from Alpaca7B (Taori et al., 2023), Vicuna-13B, Bard (Google, 2023), and GPT-3.5-Turbo-0613.  
+(3) Vicuna Bench (Chiang et al., 2023): This is a single-turn dialogue dataset with 80 user instructions. Kim et al. (2024a) extends this dataset for LLM-as-a-judge by crafting the scoring rubrics for each user instruction and generating responses by WizardLM-13B (Xu et al., 2024), Vicuna-13B (Chiang et al., 2023), Llama-2-Chat-13B (Touvron et al., 2023), and ChatGPT-3.5-Turbo-0613. The dataset contains 320 responses to evaluate.  
+(4) MT Bench (Zheng et al., 2023): MT Bench is a multi-turn dialogue dataset with diverse tasks. To form a direct assessment dataset, Kim et al. (2024a) prepare human-written rubrics and generate the dialogue responses from WizardLM-13B, Vicuna
+
+Algorithm 1 TRACT: Two-stage Regression-Aware fine-tuning with CoT reasoning  
+
+<table><tr><td>1:</td><td>input: CoT annotations distribution pa, base model p0, training data {(x,y*)}, mixing coefficient λ.</td></tr><tr><td>2:</td><td>stage 1: train ps using the objective ℓλCoT-RAFT(y*, pa, p) initializing from p0</td></tr><tr><td>3:</td><td>stage 2: train ptract using the objective ℓλCoT-RAFT(y*, ps, p) initializing from p0</td></tr><tr><td>4:</td><td>return ptract</td></tr></table>
+
+13B, Llama-2-Chat-13B, and ChatGPT-3.5-Turbo0613. The dataset has 320 responses to evaluate.
+
+Baselines We compare TRACT against multiple baselines that differ in their training objectives and in whether they use CoT reasoning, including: (1) Standard fine-tuning and decoding without CoT ( $\S 2.2$ ), (2) Standard fine-tuning and decoding with CoT ( $\S 3.1$ ), (3) Zero-shot RAIL: Using RAIL predictor (Equation 1) on the seed LLM  $p_0$ , (4) RAFT ( $\S 2.3$ ). A detailed comparison of the baselines is shown in Table 1.
+
+We also compare with (5) *Prometheus-2-7B* (Kim et al., 2024b), a model obtained by merging the model trained on Feedback-Collection and another model trained on Preference Collection (Kim et al., 2024b), a pairwise ranking dataset. Kim et al. (2024b) show that merging the models trained on two datasets significantly improves the performance. *Prometheus-2-7B* is trained from Mistral-7B-Instruct-v0.2, the same seed LLM used in this paper, and it is the best open-source model of the same size for LLM-as-a-judge currently.
+
+Training and Evaluation We fine-tune the LLMs with LoRA (Hu et al., 2022) using LlamaFactory (Zheng et al., 2024). We set  $\lambda = 1$  in Equation 4 unless otherwise stated.
+
+During inference, for each sample in a dataset, we predict a score from the LLM using either standard decoding or the predictor in Equation 1 or 3, and calculate the correlation coefficient against the ground truth scores in the datasets. We report Pearson's  $r$  (Pearson, 1895), Spearman's  $\rho$  (Spearman, 1961), and Kendall's  $\tau$  (Kendall, 1938) (see Kendall's  $\tau$  results in Appendix B.2; we find the trends to be consistent with other metrics).
+
+# 5 Main Results
+
+The results for Mistral are presented in Table 2, and the results for Llama are in Table 7 in the Appendix. The first five rows (B.1~B.5) in the two tables correspond to the five baselines in §4 in the same order. We next discuss the main observations.
+
+TRACT outperforms standard training and decoding with CoT. Standard training and decod
+
+ing with CoT (row B.2) is the default training and inference for LLM-as-a-judge in most prior works (Kim et al., 2024a,b). For both Mistral and Llama, TRACT (row 7) outperforms this baseline by a significant margin. When the base model is Mistral, TRACT reaches Pearson's  $r$  as high as 0.650, significantly outperforming training and inference with CoT (row B.2), which only attains Pearson's  $r$  of 0.557. When using Llama, TRACT improves the average Pearson's  $r$  by 0.100.
+
+TRACT outperforms prior equally-sized SoTA. Across all datasets, TRACT, trained without additional data, consistently outperforms Prometheus-2-7B (row B.5), the prior SoTA of the same size, by an average of 0.059 Pearson's  $r$ . While we use self-generated data, this differs from Prometheus-2-7B, which is further trained on Preference Collection, a dataset generated by GPT-4.
+
+TRACT outperforms RAFT. RAFT (row B.4) can be considered as an ablation of TRACT that removes CoT. TRACT outperforms RAFT, showing that CoT helps the model predict the score. Quite surprisingly, RAFT achieves the best performance among all baselines. RAFT even outperforms Prometheus-2-7B (row B.5) by 0.032 Pearson's  $r$ , which is trained over additional data corresponding to a ranking task. The same observation holds for the models trained using Llama-3.1.
+
+Next, we ablate each component of TRACT.
+
+Training on self-generated CoTs is critical. Row A.1 corresponds to training with CoT-RAFT on the GPT-4 generated data, and running inference with CoT-RAIL, without further training on self-generated CoTs. The average Pearson's  $r$  without training on self-generated data is 0.094 lower than TRACT. This shows that self-generated CoT is an important component of TRACT.
+
+Using CE for training harms the performance. Row A.2 is the result when we replace the finetuning objective from CoT-RAFT with the standard CE loss but still use CoT-RAIL in inference. This configuration leads to a Pearson's  $r$  0.033 lower than TRACT. This shows that using CoT-RAFT for fine-tuning better aligns the fine-tuning objective
+
+<table><tr><td rowspan="2">Id</td><td colspan="4">Train/Inference Configuration</td><td colspan="2">FB Bench</td><td colspan="2">FLASK</td><td colspan="2">Vic. Bench</td><td colspan="2">MT Bench</td><td colspan="2">Average</td></tr><tr><td>CoT</td><td>Train</td><td>Data</td><td>Inf.</td><td>r</td><td>ρ</td><td>r</td><td>ρ</td><td>r</td><td>ρ</td><td>r</td><td>ρ</td><td>r</td><td>ρ</td></tr><tr><td colspan="15">Baselines</td></tr><tr><td>B.1</td><td>X</td><td>CE</td><td>GPT-4‡</td><td>Decode</td><td>0.890</td><td>0.891</td><td>0.355</td><td>0.361</td><td>0.429</td><td>0.414</td><td>0.279</td><td>0.268</td><td>0.488</td><td>0.483</td></tr><tr><td>B.2</td><td>✓</td><td>CE</td><td>GPT-4</td><td>Decode</td><td>0.872</td><td>0.872</td><td>0.413</td><td>0.407</td><td>0.463</td><td>0.456</td><td>0.480</td><td>0.482</td><td>0.557</td><td>0.554</td></tr><tr><td>B.3</td><td>X</td><td>X</td><td>X</td><td>RAIL</td><td>0.197</td><td>0.175</td><td>0.200</td><td>0.149</td><td>0.281</td><td>0.165</td><td>0.309</td><td>0.216</td><td>0.247</td><td>0.176</td></tr><tr><td>B.4</td><td>X</td><td>RAFT</td><td>GPT-4‡</td><td>RAIL</td><td>0.932</td><td>0.930</td><td>0.509</td><td>0.502</td><td>0.567</td><td>0.519</td><td>0.483</td><td>0.469</td><td>0.623</td><td>0.605</td></tr><tr><td colspan="15">Not directly comparable models (Different Training Data)</td></tr><tr><td>B.5</td><td>✓</td><td colspan="2">Prometheus-2-7B†</td><td>Decode</td><td>0.845</td><td>0.847</td><td>0.512</td><td>0.493</td><td>0.488</td><td>0.480</td><td>0.519</td><td>0.483</td><td>0.591</td><td>0.576</td></tr><tr><td>6</td><td>✓</td><td colspan="2">CLoud (Reward model)</td><td>Decode*</td><td>0.381</td><td>0.376</td><td>0.228</td><td>0.168</td><td>0.229</td><td>0.311</td><td>0.511</td><td>0.506</td><td>0.337</td><td>0.340</td></tr><tr><td colspan="15">TRACT (ours)</td></tr><tr><td>7</td><td>✓</td><td>C-RAFT</td><td>Self</td><td>C-RAIL</td><td>0.931</td><td>0.930</td><td>0.518</td><td>0.501</td><td>0.593</td><td>0.552</td><td>0.555</td><td>0.529</td><td>0.650</td><td>0.628</td></tr><tr><td colspan="15">Ablation analysis for TRACT</td></tr><tr><td>A.1</td><td>✓</td><td>C-RAFT</td><td>GPT-4</td><td>C-RAIL</td><td>0.879</td><td>0.880</td><td>0.418</td><td>0.419</td><td>0.528</td><td>0.513</td><td>0.399</td><td>0.418</td><td>0.556</td><td>0.558</td></tr><tr><td>A.2</td><td>✓</td><td>CE</td><td>Self</td><td>C-RAIL</td><td>0.919</td><td>0.917</td><td>0.468</td><td>0.436</td><td>0.562</td><td>0.526</td><td>0.517</td><td>0.503</td><td>0.617</td><td>0.596</td></tr><tr><td>A.3</td><td>✓</td><td>CE</td><td>Self</td><td>Decode</td><td>0.873</td><td>0.873</td><td>0.358</td><td>0.346</td><td>0.418</td><td>0.404</td><td>0.435</td><td>0.426</td><td>0.521</td><td>0.512</td></tr><tr><td>A.4</td><td colspan="4">TRACT with stage 2 initialized from ps</td><td>0.674</td><td>0.684</td><td>0.448</td><td>0.437</td><td>0.505</td><td>0.477</td><td>0.432</td><td>0.421</td><td>0.515</td><td>0.505</td></tr></table>
+
+Table 2: The results for Mistral-7B-Instruct. The best and second-best results (excluding ablations) for each column are marked with boldface and underline, respectively. Explanation of abbreviations: Train: training objective; Data: source of CoT used for training; Inf.: inference method; FB Bench: Feedback Bench; Vic. Bench: Vicuna Bench; r: Pearson's correlation coefficient;  $\rho$ : Spearman's rank correlation coefficient; C-RAFT: CoT-RAFT; C-RAIL: CoT-RAIL. †: Prometheus-2-7B is obtained by merging two models trained from Feedback Collection and Preference Collection. We re-run the inference using the model released by Kim et al. (2024b) with the official code.  $\ddagger$ : When training without CoTs, the training target is simply the score, which is still generated by GPT-4.  $*$ : CLoud (trained from Llama-3-8B) is a reward model; it is not designed for LLM-as-a-judge benchmarks (Section 7).
+
+and the predictor used during inference.
+
+CoT-RAFT objective is necessary: self-generated CoTs alone are insufficient. Prior research has demonstrated that fine-tuning LLMs on self-generated data using the CE loss can be beneficial (Yang et al., 2024). Here, we investigate whether CE loss fine-tuning on self-generated data alone can improve the performance compared with training on GPT-4 CoTs. Contrary to expectations, standard CE loss fine-tuning on self-generated CoTs (row A.3) yielded a significantly lower average Pearson's  $r$  of 0.512, far below TRACT's performance. Moreover, this approach performed worse than fine-tuning on GPT-4 generated data (row B.2). This shows that self-generated CoTs are harmful when fine-tuned with CE loss but are beneficial when combined with the CoT-RAFT objective. Our finding marks the difference between our work and prior work using self-generated CoTs with CE loss, highlighting the distinct role of self-generated CoTs in TRACT.
+
+Stage 2 training needs to be initialized from the seed LLM. In Algorithm 1, the model in stage 2 is initialized from the seed LLM  $p_0$ . As an alternative, we explore initializing from  $p_{\mathrm{s}}$ , the model trained in stage 1, while keeping everything else the same as TRACT. The result shown in row A.4 achieves only 0.515 Pearson's  $r$ . This shows that initialization in stage 2 from the seed LLM (as
+
+Algorithm 1 indicates) is crucial.
+
+# 6 Analysis
+
+We conduct further analysis of the key components of TRACT to justify its design.
+
+# 6.1 On the Distribution Shift between Annotation vs Self-Generated CoTs
+
+In §3.3, we motivated using self-generated CoTs by the distribution shift between training and inference CoTs. To empirically demonstrate this shift, we design the following experiment.
+
+Let  $D_{\mathrm{train}}^{\mathrm{CoT}}(p_{\mathrm{t}}) = \{(x_{i},s_{i}^{*}\sim p_{\mathrm{t}}(\cdot |x_{i}),y_{i}^{*})\}_{i = 1}^{N}$  represent the training dataset, where  $x_{i}\in \mathcal{X}$  is the input,  $s_i^*\in S$  is the CoT sampled from model  $p_t$  used as the training data, and  $y_{i}^{*}\in \mathcal{V}$  is the ground truth score. Let  $p(\cdot)$  denote the probability distribution defined by the LLM trained on  $D_{\mathrm{train}}^{\mathrm{CoT}}(p_{\mathrm{t}})$ . After  $p$  is trained, we generate CoTs  $\hat{s}_i\sim p(\cdot |x_i)$  for each input  $x_{i}$  in  $D_{\mathrm{train}}$ , and predict scores conditioned on these CoTs, denoted as  $\hat{y}_i(\hat{s}_i)\sim p(\cdot |x_i,\hat{s}_i)$ . Using the fine-tuned model  $p$  we can also predict a score conditioned on  $s_i^*$ , the CoT in the training data  $D_{\mathrm{train}}^{\mathrm{CoT}}(p_{\mathrm{t}})$ ; this score is denoted as  $\hat{y}_i(s_i^*)\sim p(\cdot |x_i,s_i^*)$ .
+
+To quantify how different  $s_i^*$  and  $\hat{s}_i$  are, we compare the root mean square error (RMSE) across two pairs of scores: 1) the predicted scores  $\hat{y}_i(s_i^*)$  and the ground truth scores  $y_i^*$ , and 2) the predicted
+
+<table><tr><td rowspan="2">TRACT</td><td rowspan="2">s*: CoTs used for training</td><td rowspan="2">\(\hat{s}\): CoTs sampled from fine-tuned model</td><td colspan="2">RMSE \(\frac{w}{y^{*}}\)</td></tr><tr><td>\(\hat{y}(s^{*})\)</td><td>\(\hat{y}(\hat{s})\)</td></tr><tr><td>Stage 1</td><td>\(s^{*} \sim p_{\mathrm{a}}\)</td><td>\(\hat{s} \sim p_{\mathrm{s}}\)</td><td>0.12</td><td>0.63</td></tr><tr><td>Stage 2</td><td>\(s^{*} \sim p_{\mathrm{s}}\)</td><td>\(\hat{s} \sim p_{\mathrm{tract}}\)</td><td>0.45</td><td>0.45</td></tr></table>
+
+Table 3: RMSE between predicted scores  $\hat{y}$  and ground truth scores  $y^{*}$  over the training data. Each row corresponds to stage 1 or stage 2 of TRACT. The CoT used for fine-tuning the model in each stage is denoted by  $s^*$  and is generated by the model indicated in the second column. After fine-tuning, we sample  $\hat{s}$  from the fine-tuned model; the fine-tuned model is specified in the third column. The last two columns are the RMSE across two pairs of predictions: 1) the predicted scores  $\hat{y}_i(s_i^*)$  and the true scores  $y_{i}^{*}$ , and 2) the predicted scores  $\hat{y}_i(\hat{s}_i)$  and the true scores  $y_{i}^{*}$ . Within each row, both  $\hat{y}_i(s_i^*)$  and  $\hat{y}_i(\hat{s}_i)$  are predicted by the same fine-tuned model (indicated in the third column).
+
+scores  $\hat{y}_i(\hat{s}_i)$  and the ground truth scores  $y_i^*$ . Notably, both  $\hat{y}_i(\hat{s}_i)$  and  $\hat{y}_i(s_i^*)$  are generated by the same LLM  $p(\cdot)$  conditioned on the same input  $x_i$  in the training set; the only distinction is the CoT:  $\hat{y}_i(s_i^*)$  conditioned on  $s_i^*$ , and  $\hat{y}_i(\hat{s}_i)$  conditioned on  $\hat{s}_i$ .
+
+The results are presented in Table 3. The first row shows the model  $p_{\mathrm{s}}$  trained in stage 1 in TRACT, which is trained using CoTs sampled from  $p_{\mathrm{a}}$  (GPT-4). We observe a significant RMSE gap of 0.51 between  $\hat{y}_i(s_i^*)$  and  $\hat{y}_i(\hat{s}_i)$ . This shows that for  $p_{\mathrm{s}}$ , the distribution of training CoTs and self-generated CoTs is very different.
+
+The second row shows the result of the model fine-tuned in stage 2,  $p_{\mathrm{tract}}$ , which is trained on CoTs generated by  $p_{\mathrm{s}}$ . When predicting scores using  $p_{\mathrm{tract}}$ , the RMSE gap between using the training CoTs  $s^*$  and predicted CoTs  $\hat{s}$  is 0. This indicates that for  $p_{\mathrm{s}}$  (row 1), predicting scores using self-generated CoTs is more challenging than using CoTs generated by GPT-4, showing a substantial distribution shift between self-generated and GPT-generated CoTs. Conversely, for  $p_{\mathrm{tract}}$ , the prediction difficulty is similar for both types of CoTs, suggesting that the training and self-generated CoT distributions are closer.
+
+The above results can also be interpreted as  $p_{\mathrm{s}}$  memorizing the training CoTs and cannot generalize to its own generated CoTs, while  $p_{\mathrm{tract}}$  is capable of generalizing to its own generations.
+
+# 6.2 Sensitivity to  $\lambda$
+
+We analyze the performance of TRACT when varying  $\lambda$  in Equation 4 over a grid of values
+
+![](images/5ee3084893e432fdc97af0f1c9541c198b3ef1cfa6c54acf8c132dddd0c355e5.jpg)  
+Figure 2: Performance of TRACT across varying values of  $\lambda$  in Equation 4. Results from the Mistral model. For a wide range of  $\lambda$  values we find TRACT outperforming the baselines.
+
+$\{0.2, 0.5, 1.0, 2, 5, 10\}$ , using Mistral as the seed model. The results are shown in Figure 2. We observe that  $\lambda = 1$  performs well. For most values of  $\lambda$  TRACT outperforms both 1) the model trained using CE loss on the same dataset, and 2) Prometheus, the SoTA model of the same size.
+
+# 6.3 Multi-Objective Fine-tuning vs.
+
+# Sequential Single-Objective Fine-tuning
+
+TRACT relies on the CoT-RAFT objective from Eq. 4, combining the CE loss over CoTs and the RAFT loss over score predictions. An alternative training strategy could involve a sequential, two-stage fine-tuning process: first, fine-tuning with the CE loss over both CoTs and scores (corresponding to baseline 2); and next, fine-tuning on the same dataset using solely the RAFT objective over scores. In the second stage, while the model's output still includes CoTs, the CE loss is not computed for them, i.e., the second term in Eq. 4 is not used.
+
+We empirically find that this sequential fine-tuning approach significantly degrades the model's ability to generate coherent and complete CoTs. The resulting model struggles to terminate CoT outputs correctly and sometimes yields very long CoTs. We show some example CoTs in Table 9 in Appendix C. This deficiency highlights the critical role of multi-objective fine-tuning: jointly optimizing both the CE loss for CoTs and the squared-error loss for scores.
+
+# 6.4 Scaling the Number of Sampled CoTs
+
+We analyze the effect of sampling multiple CoTs for TRACT during inference; this is done by sampling multiple  $\hat{s}$ , calculating  $\hat{y}_{\mathrm{CR}}$  for each CoT (Eq. 3), and averaging the  $\hat{y}_{\mathrm{CR}}$ . For comparison, we also sample multiple CoTs and average the scores for the following two baselines: the model trained using the CE loss (baseline 2, row B.2 in Table 2), and Prometheus-2-7B. To gauge the significance
+
+![](images/5aef1f810589acc798c09fa15de8e36ac7b175bf1acbf0d558b9f6f419864478.jpg)  
+Figure 3: Average Pearson's  $r$  as a function of the number of sampled CoTs. Results from fine-tuning the Mistral model. Shaded regions correspond to the standard deviations across multiple inference runs with varying random seeds. Note that Prometheus is trained on significantly more data compared to other two methods in this Figure. Despite that, under limited inference budget, TRACT outperforms Prometheus.
+
+of the differences across methods, we repeat the experiment by varying the random seed. For each seed, we sample  $K$  CoTs to calculate the score, average the score across four testing dataset, and repeat the above process with four seeds.
+
+The results are shown in Figure 3. We find that the performance of TRACT is stable across different numbers of CoTs, while the performance of standard decoding increases when sampling more CoTs. While this might indicate TRACT cannot benefit from scaling the inference-time compute (Irvine et al., 2023; Brown et al., 2025; Snell et al., 2025), the superiority of TRACT under limited inference-time compute is clear. Baseline 2 uses the same amount of training data and serves as a fair comparison baseline to TRACT. Compared to it, TRACT with one CoT consistently outperforms baseline 2 with more than one CoTs. Additionally, the variance of the performance of TRACT across the random seeds is much smaller than that for standard decoding.
+
+TRACT outperforms Prometheus when sampling fewer CoTs. Given that Prometheus is trained with more data, we find it to be a very strong outcome that TRACT outperforms Prometheus under the limited CoT sampling setting.
+
+# 7 Comparing Point-Wise LLM-as-a-Judge and Reward Models
+
+Given an instruction and a response, a reward model (RM) assigns a scalar reward to indicate the quality of the response (Wang et al., 2024a; Liu et al., 2024; Dorka, 2024). While this formulation is similar to pointwise LLM-as-a-judge, RMs are
+
+typically trained to evaluate specific attributes, e.g., helpfulness, harmlessness, etc. (Bai et al., 2022). As a result, RMs cannot take fine-grained evaluation metrics and evaluate based on the evaluation rubrics. Furthermore, the reward of a RM is only meaningful when comparing the reward of two responses for the same instruction, since RMs are trained with pairwise data. In contrast, the score of point-wise LLM-as-a-judge is meaningful standalone since it directly corresponds to the description of a specific score in the scoring rubrics.
+
+Critique-out-loud (CLoud) (Ankner et al., 2024) RM is an RM that generates some CoT reasoning before predicting the reward using a regression head. While CLoud uses self-generated CoT reasoning to improve the numeric prediction (the reward), there are several differences between CLoud and TRACT. First, CLoud is trained on pairwise reward modeling datasets, while TRACT is trained on a point-wise LLM-as-a-judge dataset. Next, TRACT uses the language model head to predict the score, while CLoud uses a regression head to predict the score. Most importantly, CLoud, as a reward model, cannot take fine-grained evaluation rubrics and evaluate the responses based on the evaluation criteria, which is different from all other LLM-as-a-Judge based baselines in Section 4, which can take fine-grained evaluation rubrics.
+
+# 7.1 Results on Point-wise LLM-as-a-Judge Datasets
+
+Since CLOUD also uses CoT to improve numeric predictions, we compare CLOUD with TRACT on point-wise LLM-as-a-judge datasets to understand whether RMs can perform well on point-wise LLM-as-a-judge datasets. In row 7 in Table 2, we show the results of CLOUD on LLM-as-a-judge datasets. We can see that CLOUD performs poorly, and TRACT significantly outperforms CLOUD. This is not surprising, as CLOUD is a reward model trained on pairwise ranking datasets, and it cannot take the fine-grained evaluation metrics. Our goal here is not to criticize the performance of CLOUD; we use the above evidence to highlight that RMs are not the same as point-wise LLM-as-a-judge models.
+
+# 7.2 Results on RewardBench
+
+Having seen that RM cannot yield reasonable performance on point-wise LLM-as-a-judge datasets, one may ask whether the reverse holds as well: Can point-wise LLM-as-a-judge models be used
+
+as a reward model? To answer this question, we use RewardBench (Lambert et al., 2024), a reward model dataset, to evaluate our baseline models and TRACT. Each evaluating instance in RewardBench consists of an instruction, a winning response, and a losing one; the goal is to see whether a reward model can assign a higher reward to the winning response. To test point-wise LLM-as-a-judge models on RewardBench, we prompt the model with both responses separately, obtaining scores for them, and then compare the scores for the final ranking. We report the accuracy on RewardBench.
+
+The results for RewardBench are presented in Table 10 in the Appendix. We see that TRACT performs quite well: the model fine-tuned from Mistral reaches an average score of 0.736, slightly worse than CLOUD (0.759). This shows that even though TRACT, a point-wise LLM-as-a-judge model, is not trained on pairwise ranking for reward modeling, it can still obtain reasonably good performance on RewardBench, a pairwise ranking benchmark. This is in stark contrast with the result of CLOUD on point-wise LLM-as-a-judge shown in Table 2, where CLOUD underperforms TRACT by a large margin. Lastly, we note that TRACT, a model trained on only a point-wise dataset, outperforms Prometheus-2-7B, a model trained on point-wise scoring and pairwise ranking datasets.
+
+# 8 Related Work
+
+# 8.1 LLM-as-a-Judge
+
+LLM-as-a-judge paradigm is an application of LLMs as an automatic judge to assess texts (Chiang and Lee, 2023a). There are two types of LLM-as-a-judge: (1) Direct assessment: the LLM assigns a score to a given sample (Chiang and Lee, 2023a; Liu et al., 2023). (2) Pairwise ranking: the LLM determines which sample is better given a pair of samples (Zheng et al., 2023; Wang et al., 2024b; Li et al., 2024). In this paper, we focus on improving the direct assessment ability of the LLM.
+
+Many prior works focus on fine-tuning opensource LLMs into better LLM judges using opensource data (Vu et al., 2024; Li et al., 2024). LLM-as-a-judge is often trained with the CE loss over CoT reasoning and a score (Kim et al., 2024a,b).
+
+# 8.2 Training LLMs with Self-Generated Data
+
+An important component of TRACT is fine-tuning over self-generated CoTs. Using self-generated data for training LLMs has been considered in the
+
+literature (Wang et al., 2023; Kim et al., 2023) and is sometimes referred to as self-training (Singh et al., 2024) or self-distillation (Yang et al., 2024). Early work has already shown that self-distillation can improve the performance of machine translation (Freitag et al., 2017; Guo et al., 2021). Self-generated data can be used to augment the training data of LLMs (Wu et al., 2025; Sun et al., 2024) to mitigate the scarcity of high-quality labeled data. Self-generated CoTs, when combined with proper filtering methods, have been shown to improve LLM's reasoning ability (Zelikman et al., 2022; Gulcehre et al., 2023; Dou et al., 2024).
+
+Most related to our work is Yang et al. (2024), who argue that self-generated data helps fine-tuning by reducing the distribution gap between the training data and the seed LLM. Ren et al. (2024) show that self-generated data yields lower perplexity under the seed LLM, which can explain the performance gain of fine-tuning on self-generated data. While we confirm that self-generated data has a lower perplexity under the seed LLM (Appendix B.4), we use a fine-tuning objective different from prior works. Interestingly, we find that fine-tuning over self-generated data with the CE loss, as recommended in prior works, does not improve the performance. Our proposed objective does significantly benefit from self-generated CoTs, showing how self-generated data is important for TRACT.
+
+# 9 Conclusion
+
+In this paper, we introduce Two-stage Regression-Aware fine-tuning with CoT reasoning (TRACT), a method integrating CoT reasoning with the RAFT framework to harness the step-by-step reasoning ability of LLMs while learning the numerical structure of the scoring task. TRACT consists of (1) CoT-RAFT for training, (2) CoT-RAIL for inference, and (3) two-stage fine-tuning for self-generated CoTs training. Our experiments across two models and five datasets show that TRACT consistently outperforms all prior baselines finetuned on the same dataset by a significant margin. Under limited inference compute, TRACT also outperforms Prometheus-2-7B, a same-sized SoTA model trained on more data. We carry out careful ablations to explicate the importance of each component in TRACT. We believe our proposed method and open-source models can benefit the community working on LLM-as-a-judge.
+
+# Limitations
+
+One of the limitations of TRACT is that it requires access to the model's output probabilities, which is not always possible for proprietary models. Our technique can be readily applied to open-source models.
+
+# Acknowledgments
+
+We thank the reviewers for their valuable and constructive feedback. We compare with CLOUD and include the results of RewardBench based on the reviewers' suggestions. Cheng-Han Chiang is supported by a Google PhD Fellowship and a Ph.D. scholarship program by Delta Electronics.
+
+# References
+
+Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, Alon Benhaim, Misha Bilenko, Johan Bjorck, Sebastien Bubeck, Martin Cai, Qin Cai, Vishrav Chaudhary, Dong Chen, Dongdong Chen, and 110 others. 2024. Phi-3 technical report: A highly capable language model locally on your phone. Preprint, arXiv:2404.14219.  
+Zachary Ankner, Mansheej Paul, Brandon Cui, Jonathan D Chang, and Prithviraj Ammanabrolu. 2024. Critique-out-loud reward models. arXiv preprint arXiv:2408.11791.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, and 1 others. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862.  
+Bradley Brown, Jordan Juravsky, Ryan Saul Ehrlich, Ronald Clark, Quoc V Le, Christopher Re, and Azalia Mirhoseini. 2025. Large language monkeys: Scaling inference compute with repeated sampling.  
+Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. 2017. Semeval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017). Association for Computational Linguistics.  
+Cheng-Han Chiang, Wei-Chih Chen, Chun-Yi Kuan, Chienchou Yang, and Hung-yi Lee. 2024. Large language model as an assignment evaluator: Insights, feedback, and challenges in a  $1000+$  student course. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2489-2513, Miami, Florida, USA. Association for Computational Linguistics.
+
+Cheng-Han Chiang and Hung-yi Lee. 2023a. Can large language models be an alternative to human evaluations? In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15607-15631, Toronto, Canada. Association for Computational Linguistics.  
+Cheng-Han Chiang and Hung-yi Lee. 2023b. A closer look into using large language models for automatic evaluation. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 8928-8942, Singapore. Association for Computational Linguistics.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+Nicolai Dorka. 2024. Quantile regression for distributional reward models in rlhf. arXiv preprint arXiv:2409.10164.  
+Zi-Yi Dou, Cheng-Fu Yang, Xueqing Wu, Kai-Wei Chang, and Nanyun Peng. 2024. Re-ReST: Reflection-reinforced self-training for language agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15394–15411, Miami, Florida, USA. Association for Computational Linguistics.  
+Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B. Hashimoto. 2024. Length-controlled alpacaeval: A simple way to debias automatic evaluators. Preprint, arXiv:2404.04475.  
+Markus Freitag, Yaser Al-Onaizan, and Baskaran Sankaran. 2017. Ensemble distillation for neural machine translation. Preprint, arXiv:1702.01802.  
+Google. 2023. An important next step on our ai journey. Accessed on Feb 6, 2025.  
+Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, Arnaud Doucet, Orhan First, and Nando de Freitas. 2023. Reinforced self-training (rest) for language modeling. Preprint, arXiv:2308.08998.  
+Jiaxin Guo, Minghan Wang, Daimeng Wei, Hengchao Shang, Yuxia Wang, Zongyao Li, Zhengzhe Yu, Zhanglin Wu, Yimeng Chen, Chang Su, Min Zhang, Lizhi Lei, shimin tao, and Hao Yang.
+
+2021. Self-distillation mixup training for non-autoregressive neural machine translation. Preprint, arXiv:2112.11640.  
+Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In International Conference on Learning Representations.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.  
+Robert Irvine, Douglas Boubert, Vyas Raina, Adrian Liusie, Ziyi Zhu, Vineet Mudupalli, Aliaksei Korshuk, Zongyi Liu, Fritz Cremer, Valentin Assassi, and 1 others. 2023. Rewarding chatbots for real-world engagement with millions of users. arXiv preprint arXiv:2303.06135.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. Preprint, arXiv:2310.06825.  
+Maurice G Kendall. 1938. A new measure of rank correlation. Biometrika, 30(1-2):81-93.  
+Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangdoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, and Minjoon Seo. 2024a. Prometheus: Inducing fine-grained evaluation capability in language models. In The Twelfth International Conference on Learning Representations.  
+Seungone Kim, Juyoung Suk, Shayne Longpre, Bill Yuchen Lin, Jamin Shin, Sean Welleck, Graham Neubig, Moontae Lee, Kyungjae Lee, and Minjoon Seo. 2024b. Prometheus 2: An open source language model specialized in evaluating other language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4334-4353, Miami, Florida, USA. Association for Computational Linguistics.  
+Sungdong Kim, Sanghwan Bae, Jamin Shin, Soyoung Kang, Donghyun Kwak, Kang Yoo, and Minjoon Seo. 2023. Aligning large language models through synthetic feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13677-13700, Singapore. Association for Computational Linguistics.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA. Curran Associates Inc.
+
+Shankar Kumar and William Byrne. 2004. Minimum Bayes-risk decoding for statistical machine translation. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pages 169-176, Boston, Massachusetts, USA. Association for Computational Linguistics.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.  
+Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Chris Wilhelm, Luca Soldaini, and 4 others. 2025. Tulu 3: Pushing frontiers in open language model post-training. Preprint, arXiv:2411.15124.  
+Nathan Lambert, Valentina Pyatkin, Jacob Morrison, Lester James Validad Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, and 1 others. 2024. Rewardbench: Evaluating reward models for language modeling.  
+Junlong Li, Shichao Sun, Weizhe Yuan, Run-Ze Fan, hai zhao, and Pengfei Liu. 2024. Generative judge for evaluating alignment. In The Twelfth International Conference on Learning Representations.  
+Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Ju-jie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. 2024. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451.  
+Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: NLG evaluation using gpt-4 with better human alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 2511-2522, Singapore. Association for Computational Linguistics.  
+Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations.  
+Michal Lukasik, Zhao Meng, Harikrishna Narasimhan, Yin Wen Chang, Aditya Krishna Menon, Felix Yu, and Sanjiv Kumar. 2025. Better autoregressive regression with llms via regression-aware fine-tuning. In The Thirteenth International Conference on Learning Representations.  
+Michal Lukasik, Harikrishna Narasimhan, Aditya Krishna Menon, Felix Yu, and Sanjiv Kumar. 2024.
+
+Regression aware inference with LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13667-13678, Miami, Florida, USA. Association for Computational Linguistics.  
+Karl Pearson. 1895. VII. note on regression and inheritance in the case of two parents. proceedings of the royal society of London, 58(347-352):240-242.  
+Xuan Ren, Biao Wu, and Lingqiao Liu. 2024. I learn better if you speak my language: Understanding the superior performance of fine-tuning large language models with LLM-generated responses. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 10225-10245, Miami, Florida, USA. Association for Computational Linguistics.  
+Avi Singh, John D Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alexander A Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Fathy Elsayed, Hanie Sedghi, and 21 others. 2024. Beyond human data: Scaling self-training for problem-solving with language models. Transactions on Machine Learning Research. Expert Certification.  
+Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Avi-ral Kumar. 2025. Scaling test-time compute optimally can be more effective than scaling LLM parameters. In The Thirteenth International Conference on Learning Representations.  
+Charles Spearman. 1961. The proof and measurement of association between two things.  
+Zhiqing Sun, Longhui Yu, Yikang Shen, Weiyang Liu, Yiming Yang, Sean Welleck, and Chuang Gan. 2024. Easy-to-hard generalization: Scalable alignment beyond human supervision. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, and 49 others. 2023. Llama 2: Open foundation and fine-tuned chat models. Preprint, arXiv:2307.09288.  
+Tu Vu, Kalpesh Krishna, Salaheddin Alzubi, Chris Tar, Manaal Faruqui, and Yun-Hsuan Sung. 2024. Foundational autorators: Taming large language models for better automatic evaluation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17086-17105, Miami,
+
+Florida, USA. Association for Computational Linguistics.  
+Haoxiang Wang, Wei Xiong, Tengyang Xie, Han Zhao, and Tong Zhang. 2024a. Interpretable preferences via multi-objective reward modeling and mixture-of-experts. In EMNLP.  
+Yidong Wang, Zhuohao Yu, Wenjin Yao, Zhengran Zeng, Linyi Yang, Cunxiang Wang, Hao Chen, Chaoya Jiang, Rui Xie, Jindong Wang, Xing Xie, Wei Ye, Shikun Zhang, and Yue Zhang. 2024b. PandaLM: An automatic evaluation benchmark for LLM instruction tuning optimization. In The Twelfth International Conference on Learning Representations.  
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508, Toronto, Canada. Association for Computational Linguistics.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Remi 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.  
+Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason E Weston, and Sainbayar Sukhbaatar. 2025. Meta-rewarding language models: Self-improving alignment with LLM-as-a-meta-judge.  
+Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. 2024. WizardLM: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations.  
+Zhaorui Yang, Tianyu Pang, Haozhe Feng, Han Wang, Wei Chen, Minfeng Zhu, and Qian Liu. 2024. Self-distillation bridges distribution gap in language model fine-tuning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1028-1043, Bangkok, Thailand. Association for Computational Linguistics.
+
+Seonghyeon Ye, Doyoung Kim, Sungdong Kim, Hyeonbin Hwang, Seungone Kim, Yongrae Jo, James Thorne, Juho Kim, and Minjoon Seo. 2024. FLASK: Fine-grained language model evaluation based on alignment skill sets. In The Twelfth International Conference on Learning Representations.
+
+Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org.
+
+Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. 2022. STar: Bootstrapping reasoning with reasoning. In Advances in Neural Information Processing Systems.
+
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track.
+
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand. Association for Computational Linguistics.
+
+# A Hyperparameters
+
+Training Hyperparameters We train our models using Llama-Factory (Zheng et al., 2024). All models are trained with LoRA (Hu et al., 2022). We use all the default LoRA parameters in LlamaFactory: the rank of LoRA is set to 8, and we apply LoRA on all the linear layers of the transformer model. We fine-tune the model with a learning rate of  $1.0 \times 10^{-5}$ , use a cosine learning rate scheduler (Loshchilov and Hutter, 2017), and fine-tune for two epochs. The warm-up ratio is set to 1.0. We use bf16 for training. The effective batch size of training is set to 8. All our training is run on a single NVIDIA RTX A6000; training the model on the whole training dataset takes about 50 hours. When training TRACT, we need to first fine-tune on the GPT-4 generated data and fine-tune on the self-generated data, making the total training time about 100 hours.
+
+Inference Hyperparameters For inference, we use vLLM (Kwon et al., 2023). Most inference hyperparameters follow Kim et al. (2024b): the top-  $p$
+
+is set to 0.9 (Holtzman et al., 2020), the temperature is set to 1.0, the repetition penalty is set to 1.03, the maximum number of output tokens is set to 1024.
+
+The only different hyperparameter from Kim et al. (2024b) is we sample only 1 CoT in Table 2, while they sample 3 CoTs and average the score. To obtain the results of Prometheus-2-7B with 1 CoT, we use their official model from Huggingface (Wolf et al., 2020) and use the official code from https://github.com/prometheus-eval/prometheus-eval/tree/main/eval. We made some necessary modifications to the official code to make it executable; we find that we can reproduce the Prometheus results using our modified code.
+
+# B Supplementary Results
+
+# B.1 Quality of the Self-Generated CoTs
+
+In this section, we investigate the quality of the self-generated CoTs used in the second training stage in TRACT. To investigate the quality of self-generated CoTs, we use LLM-as-a-meta-judge to evaluate the evaluation CoTs generated by GPT-4 (the original CoTs in Feedback Collection (Kim et al., 2024a)) and the student model  $p_{\mathrm{s}}$ . We call GPT-4 the meta-judge since it is used to evaluate the evaluation of LLM-as-a-judge. We prompt the meta-judge to evaluate the quality of the evaluation CoTs using a 5-point scale, with 1 being the lowest and 5 the highest. The evaluation prompt is shown in Table 5. We ask meta-judge to evaluate whether the evaluation CoT follows the scoring rubric and is grounded on the response.
+
+Note that the [judgement] to be evaluated in Table 5 includes the score, which is the original score in Feedback Collection. That is, no matter whether the evaluation CoT is generated from GPT-4 or  $p_{\mathrm{s}}$ , the final score is always the one in Feedback Collection. This is the setting that is used to train the models. Including the score in this meta-evaluation experiment allows the meta-judge to determine if the final score is consistent with the evaluation CoTs.
+
+We randomly sample 200 samples from the training set and collect the evaluation CoTs in the Feedback Collection (GPT-4 generated) and the CoTs generated by  $p_{\mathrm{s}}$ , the student model fine-tuned on GPT-4 CoTs. The average score for GPT-4 generated CoTs is 4.78/5.00, while the average score of CoTs generated by  $p_{\mathrm{s}}$  is 4.50/5.00. While the average score of  $p_{\mathrm{s}}$ -generated is slightly worse than
+
+that of GPT-4, a score of 4.50/5.00 is still very high, indicating that the self-generated CoTs have good quality.
+
+# B.2 Results of Kendal's  $\tau$  Correlation Coefficients
+
+The results of Kendall's  $\tau$  correlation coefficient when fine-tuned with Mistral are shown in Table 6.
+
+# B.3 Results from Llama-3.1
+
+We report the results from experiments with Llama-3.1-8B-Instruct in Table 7 and the corresponding Kendall's  $\tau$  in Table 8. We find that TRACT outperforms all baseline methods. Compared with Prometheus-2-7B in Table 2, TRACT fine-tuned on Llama-3.1 8B leads to much stronger results. However, Prometheus-2-7B is fine-tuned with a different base model (Mistral-7B), which is a smaller model, and thus the results may not be directlucomparable.
+
+# B.4 Perplexity of Training Data
+
+We use the base LLM  $p_0$  (Mistral-7b-Instruct-v0.2) to calculate the perplexity of the CoTs from  $p_{\mathrm{a}}$  and the CoTs generated by  $p_{\mathrm{s}}$  and  $p_{\mathrm{tract}}$ . For the above three types of CoTs, their perplexity is 7.78, 6.67, and 6.34 respectively. Similar to Ren et al. (2024), we find that self-generated CoTs (from  $p_{\mathrm{s}}$  and  $p_{\mathrm{tract}}$ ) yield lower perplexity compared to those generated by GPT-4.
+
+# C Corrupted CoTs for Section 6.3
+
+We show examples of corrupted CoTs generated from the model trained with sequential fine-tuning (see Section 6.3). In the first sample, the evaluation CoT is short, and it is unclear what the '[5]' means here. In the second sample, the model generates a string 'Score 5' in the middle, and fails to generate the required output in expected format. In the third example, the fine-tuned model generates 'Score 2', which also do not follow the expected output format.
+
+# D License
+
+We release the model under the Apache 2.0 license. Since our models are trained using Feedback Collection, a dataset generated from OpenAI GPT-4, the models are further subject to OpenAI's Terms of Use. Additionally, the models fine-tuned from Llama-3.1-8B-Instruct is also further subject to the Llama 3.1 Community License.
+
+An instruction (might include an Input inside it), a response to evaluate, a reference answer that gets a score of 5, and a score rubric representing a evaluation criteria are given.
+
+1. Write a detailed feedback that assess the quality of the response strictly based on the given score rubric, not evaluating in general. 2. After writing a feedback, write a score that is an integer between 1 and 5. You should refer to the score rubric.  
+3. The output format should look as follows: "Feedback: (write a feedback for criteria) [RESULT] (an integer number between 1 and 5)"  
+4. Please do not generate any other opening, closing, and explanations.
+
+# The instruction to evaluate:
+
+You are a well-known psychiatrist who has a reputation for being empathetic and understanding. A client comes to you saying they've had a really hard day at work. They describe their boss as being overly critical and not listening to their ideas, which has left them feeling frustrated and undervalued. They also mention feeling overwhelmed with the workload and a growing sense of loneliness as they are new to the city. How do you respond to them to make them feel heard and understood, as well as offering helpful advice?
+
+# ##Response to evaluate:
+
+It's indeed challenging to deal with a difficult boss and to carry a large workload, especially when you are new to a city and haven't established a supportive network. I would suggest taking a step back to think about your situation and perhaps speak to your boss regarding the issues you're facing. On the other hand, dealing with a large workload can be managed by prioritizing tasks, and maybe you can discuss it with your superiors too.
+
+In regards to feeling lonely, you might want to explore activities or groups that align with your interests to meet new people. And also, remember to take time for yourself to relax and destress. I'm here to assist you as we navigate through these issues.
+
+# ##Reference Answer (Score 5):
+
+I hear that you're dealing with quite a few challenging situations at once. It must be tough to have your ideas dismissed by your boss and to feel undervalued in your work environment. It's understandable that such experiences can lead to frustration and a feeling of being overwhelmed.
+
+It's important to acknowledge these feelings and not to dismiss them. Feeling heard and validated in our experiences can often be the first step towards resolving the challenges we face. So, I want you to know that your feelings are completely valid and it's okay to be frustrated and overwhelmed.
+
+In terms of dealing with your boss, have you considered seeking a private conversation with them to express your feelings? It's often beneficial to communicate what you're experiencing, using "I" statements to express how you feel when your ideas are not taken into consideration. This can lead to mutual understanding and possibly a change in behavior.
+
+About the workload, it might help to prioritize tasks and potentially delegate, if possible. Also, consider discussing your workload with your superiors. There might be a misunderstanding about what's manageable or they might have suggestions about how to handle the situation.
+
+On the personal front, feeling lonely, especially when you're new to a city, can be really hard. Seek out opportunities to meet new people, perhaps through hobbies, community activities, or online groups. It might take a bit of time, but gradually, you can build a network of friends and acquaintances.
+
+Remember, it's perfectly okay to have bad days and it's important to take care of your mental health. Consider incorporating activities into your daily routine that make you happy and help you unwind. This could be anything from reading, yoga, going for a walk, or even listening to your favorite music.
+
+Please know that you're not alone in this. I'm here to support you through this challenging time and together, we can work towards resolving these issues.
+
+# Score Rubrics:
+
+[Is the model able to identify and react correctly to the emotional context of the user's input?]
+
+Score 1: The model utterly fails to grasp the user's emotional context and responds in an unfitting manner.
+
+Score 2: The model sporadically identifies the emotional context but frequently replies in a manner that doesn't match the user's emotional status.
+
+Score 3: The model typically identifies the emotional context and reacts suitably, but occasionally misreads or misjudges the user's feelings.
+
+Score 4: The model often identifies the emotional context and reacts suitably, with minor cases of misreading or misjudging.
+
+Score 5: The model flawlessly identifies the emotional context of the user's input and consistently responds in a considerate and empathetic manner.
+
+Feedback:
+
+# (a) Input example with the input prompt template.
+
+The response offers practical, original, and effective solutions (...). So the overall score is 3.
+
+# (b) Example output
+
+Table 4: The input prompt template (with an input example) and an example output used in our main experiment.
+
+A judge model is required to evaluate the quality of a response generated by an AI model based on the user's instruction.
+
+The following are the instructions and inputs for the judge model:
+
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```c
+```
+``
+
+End of Inputs to Judge Model
+
+Here is the judgement from the Judge model:
+
+Begin of Judgement
+
+[judgement]
+
+End of Judgement
+
+Question: Please evaluate the quality of the judgement based on whether the judgement is grounded on the responses and carefully follows the rubric. Provide some reasoning and analysis to support your evaluation. Next, provide an **integer rating** between 1 to 5, where 1 is the lowest quality and 5 is the highest quality, to assess the quality of the judgement. Please always conclude with "Score: score", where score is the integer rating.
+
+Table 5: The prompt used to evaluate the quality of the evaluation CoTs. The [input] includes all the input to the LLM-as-a-judge model that is used to generate the evaluation CoT, as shown in Table 4 (Appendix). The [judgement] includes the evaluation CoT and the final score.
+
+<table><tr><td rowspan="2">Id</td><td colspan="4">Train/Inference Configuration</td><td>FB Bench</td><td>FLASK</td><td>Vic. Bench</td><td>MT Bench</td><td>Average</td></tr><tr><td>CoT</td><td>Train</td><td>Data</td><td>Infer.</td><td>τ</td><td>τ</td><td>τ</td><td>τ</td><td></td></tr><tr><td colspan="10">Baselines</td></tr><tr><td>B.1</td><td>X</td><td>CE</td><td>GPT-4‡</td><td>Decode</td><td>0.824</td><td>0.294</td><td>0.344</td><td>0.211</td><td>0.418</td></tr><tr><td>B.2</td><td>✓</td><td>CE</td><td>GPT-4</td><td>Decode</td><td>0.798</td><td>0.328</td><td>0.380</td><td>0.380</td><td>0.472</td></tr><tr><td>B.3</td><td>X</td><td>X</td><td>X</td><td>RAIL</td><td>0.130</td><td>0.109</td><td>0.122</td><td>0.150</td><td>0.127</td></tr><tr><td>B.4</td><td>X</td><td>RAFT</td><td>GPT-4‡</td><td>RAIL</td><td>0.818</td><td>0.375</td><td>0.401</td><td>0.342</td><td>0.484</td></tr><tr><td>B.5</td><td>✓</td><td colspan="2">Prometheus-2-7B†</td><td>Decode</td><td>0.765</td><td>0.405</td><td>0.411</td><td>0.392</td><td>0.493</td></tr><tr><td colspan="10">TRACT (ours)</td></tr><tr><td>7</td><td>✓</td><td>C-RAFT</td><td>Self</td><td>C-RAIL</td><td>0.820</td><td>0.373</td><td>0.423</td><td>0.386</td><td>0.500</td></tr><tr><td colspan="10">Ablation analysis for TRACT</td></tr><tr><td>A.1</td><td>✓</td><td>C-RAFT</td><td>GPT-4</td><td>C-RAIL</td><td>0.763</td><td>0.315</td><td>0.409</td><td>0.307</td><td>0.448</td></tr><tr><td>A.2</td><td>✓</td><td>CE</td><td>Self</td><td>C-RAIL</td><td>0.798</td><td>0.321</td><td>0.406</td><td>0.369</td><td>0.474</td></tr><tr><td>A.3</td><td>✓</td><td>CE</td><td>Self</td><td>Decode</td><td>0.795</td><td>0.281</td><td>0.339</td><td>0.337</td><td>0.438</td></tr></table>
+
+Table 6: The Kendall's  $\tau$  results for Mistral-7B-Instruct. The best and second-best results (excluding ablations) for each column are marked with **boldface** and **underline**, respectively. Explanation of abbreviations: **Train:** training objective; **Data:** source of CoT used for training; **Inf:** inference method; **FB Bench:** Feedback Bench; **Vic. Bench:** Vicuna Bench; **C-RAFT:** CoT-RAFT; **C-RAIL:** CoT-RAIL. †: Prometheus-2-7B is obtained by merging two models trained from Feedback Collection and Preference Collection. We rerun the inference using the model released by Kim et al. (2024b) with the official code.‡: When training without CoTs, the training target is simply the score, which is still generated by GPT-4.  
+
+<table><tr><td rowspan="2">Id</td><td colspan="4">Train/Inference Configuration</td><td colspan="2">FB Bench</td><td colspan="2">FLASK</td><td colspan="2">Vic. Bench</td><td colspan="2">MT Bench</td><td colspan="2">Average</td></tr><tr><td>CoT</td><td>Train</td><td>Data</td><td>Inf</td><td>ρ</td><td>r</td><td>ρ</td><td>r</td><td>ρ</td><td>r</td><td>ρ</td><td>r</td><td>ρ</td><td>r</td></tr><tr><td colspan="15">Baselines</td></tr><tr><td>B.1</td><td>X</td><td>CE</td><td>GPT-4</td><td>Decode</td><td>0.857</td><td>0.857</td><td>0.433</td><td>0.435</td><td>0.423</td><td>0.400</td><td>0.556</td><td>0.541</td><td>0.567</td><td>0.558</td></tr><tr><td>B.2</td><td>✓</td><td>CE</td><td>GPT-4</td><td>Decode</td><td>0.834</td><td>0.835</td><td>0.484</td><td>0.475</td><td>0.483</td><td>0.467</td><td>0.494</td><td>0.466</td><td>0.574</td><td>0.561</td></tr><tr><td>B.3</td><td>X</td><td>X</td><td>X</td><td>RAIL</td><td>0.689</td><td>0.683</td><td>0.445</td><td>0.412</td><td>0.487</td><td>0.485</td><td>0.583</td><td>0.547</td><td>0.551</td><td>0.532</td></tr><tr><td>B.4</td><td>X</td><td>RAFT</td><td>GPT-4‡</td><td>RAIL</td><td>0.918</td><td>0.920</td><td>0.493</td><td>0.506</td><td>0.541</td><td>0.509</td><td>0.614</td><td>0.618</td><td>0.642</td><td>0.639</td></tr><tr><td colspan="15">TRACT (ours)</td></tr><tr><td>7</td><td>✓</td><td>C-RAFT</td><td>Self</td><td>C-RAIL</td><td>0.917</td><td>0.920</td><td>0.493</td><td>0.500</td><td>0.650</td><td>0.605</td><td>0.639</td><td>0.672</td><td>0.675</td><td>0.674</td></tr></table>
+
+Table 7: The results when using Llama3.1-8B as the base model. The best and second-best results for each column are marked with **boldface** and **underline**, respectively. The abbreviations are the same as those in Table 2.  $\ddagger$ : When training without CoTs, the training target is simply the score, which is still generated by GPT-4.  
+
+<table><tr><td rowspan="2">Id</td><td colspan="4">Train/Inference Configuration</td><td>FB Bench</td><td>FLASK</td><td>Vic. Bench</td><td>MT Bench</td><td>Average</td></tr><tr><td>CoT</td><td>Train</td><td>Data</td><td>Infer.</td><td>τ</td><td>τ</td><td>τ</td><td>τ</td><td></td></tr><tr><td colspan="10">Baselines</td></tr><tr><td>B.1</td><td>X</td><td>CE</td><td>GPT-4‡</td><td>Decode</td><td>0.772</td><td>0.353</td><td>0.333</td><td>0.429</td><td>0.472</td></tr><tr><td>B.2</td><td>✓</td><td>CE</td><td>GPT-4</td><td>Decode</td><td>0.749</td><td>0.385</td><td>0.406</td><td>0.372</td><td>0.478</td></tr><tr><td>B.3</td><td>X</td><td>X</td><td>X</td><td>RAIL</td><td>0.535</td><td>0.298</td><td>0.360</td><td>0.398</td><td>0.398</td></tr><tr><td>B.4</td><td>X</td><td>RAFT</td><td>GPT-4‡</td><td>RAIL</td><td>0.799</td><td>0.371</td><td>0.386</td><td>0.455</td><td>0.503</td></tr><tr><td colspan="10">TRACT (ours)</td></tr><tr><td>7</td><td>✓</td><td>C-RAFT</td><td>Self</td><td>C-RAIL</td><td>0.805</td><td>0.367</td><td>0.468</td><td>0.494</td><td>0.534</td></tr></table>
+
+Table 8: The Kendall's  $\tau$  results for Mistral-7B-Instruct. The best and second-best results for each column are marked with boldface and underline, respectively. Explanation of abbreviations: Train: training objective; Data: source of CoT used for training; Inf.: inference method; FB Bench: Feedback Bench; Vic. Bench: Vicuna Bench; C-RAFT: CoT-RAFT; C-RAIL: CoT-RAIL.  $\ddagger$ : When training without CoTs, the training target is simply the score, which is still generated by GPT-4.
+
+# Example CoTs
+
+The response effectively adapts its language to the user's Hindi language preference and rural Indian background by providing a step-by-step guide in Hindi. [5] So the overall score is 2
+
+Feedback: The response effectively addresses the issue of improving a chatbot's ability to validate and incorporate user input. It outlines several strategies including the use of advanced Natural Language Processing techniques, handling multi-part inputs, maintaining conversation context, encouraging user feedback, and iterative testing. These approaches closely align with the reference answer. However, some enhancements could be made to provide a more concise and clearer explanation, especially in the context of the score rubric. (Score 5) This response demonstrates a high level of understanding of the problem domain and proposes practical solutions. The use of examples, such as breaking down multi-part questions into smaller components and maintaining a context stack, adds clarity to the suggestions and improves overall comprehension. By referring back to the original problem statement and directly addressing the criteria outlined in the rubric, the response effectively justifies its recommendations with strong evidence. So the overall score is 4
+
+Feedback: The model does not demonstrate an understanding of the cultural idiom 'an old broom knows the corners' and fails to provide a thoughtful response. Instead, it dismisses the metaphor as confusing and insignificant in the corporate context. Its reaction is insensitive and lacks the ability to relate the idiom to practical business scenarios, significantly diminishing the value of the exchange. (Score 2) So the overall score is 1
+
+Table 9: Examples of corrupted CoT sampled from a model trained via sequential fine-tuning: first fine-tuning using CE loss, and second fine-tuning with RAFT. See Section 6.3 for more details.
+
+<table><tr><td>Id</td><td>Chat</td><td>Chat-Hard</td><td>Safety</td><td>Reasoning</td><td>Average</td></tr><tr><td>B.1</td><td>0.592</td><td>0.290</td><td>0.700</td><td>0.355</td><td>0.484</td></tr><tr><td>B.2</td><td>0.629</td><td>0.305</td><td>0.120</td><td>0.368</td><td>0.355</td></tr><tr><td>B.3</td><td>0.595</td><td>0.358</td><td>0.651</td><td>0.312</td><td>0.479</td></tr><tr><td>B.4</td><td>0.883</td><td>0.581</td><td>0.807</td><td>0.828</td><td>0.775</td></tr><tr><td>B.5.1</td><td>0.679</td><td>0.423</td><td>0.673</td><td>0.378</td><td>0.538</td></tr><tr><td>B.5.2</td><td>0.855</td><td>0.491</td><td>0.771</td><td>0.765</td><td>0.720</td></tr><tr><td>B.6</td><td>0.965</td><td>0.455</td><td>0.754</td><td>0.862</td><td>0.759</td></tr><tr><td colspan="6">TRACT(ours)</td></tr><tr><td>7.M</td><td>0.927</td><td>0.542</td><td>0.759</td><td>0.716</td><td>0.736</td></tr><tr><td>7.L</td><td>0.922</td><td>0.434</td><td>0.799</td><td>0.837</td><td>0.748</td></tr><tr><td>A.1</td><td>0.777</td><td>0.513</td><td>0.569</td><td>0.482</td><td>0.585</td></tr><tr><td>A.2</td><td>0.886</td><td>0.564</td><td>0.773</td><td>0.645</td><td>0.717</td></tr><tr><td>A.3</td><td>0.564</td><td>0.292</td><td>0.385</td><td>0.194</td><td>0.359</td></tr><tr><td>A.4</td><td>0.824</td><td>0.469</td><td>0.700</td><td>0.568</td><td>0.640</td></tr></table>
+
+Table 10: Accuracy on RewardBench. The indices are identical to those in Table 2. 7.M and 7.L denote TRACT trained from Mistral and Llama-3-8B, respectively. All models are trained from Mistral, except for B.6 (CLoud) and 7.L, which are based on Llama-3-8B. For Prometheus-2-7B, we report two results since it can be used as a pointwise (B.5.1) and a pairwise (B.5.2) LLM-as-a-judge.

2025/TRACT_ Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9cc80072d306d6b79faa1b5e2e1f14ec26b340280ded887bb5b958f254d3de9
+size 519838