Add Batch 57271549-52ae-4d2f-ab66-b29b4bc051a8 data

.gitattributes

@@ -3825,3 +3825,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/Search[[:space:]]if[[:space:]]you[[:space:]]don’t[[:space:]]know![[:space:]]Knowledge-Augmented[[:space:]]Korean[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/95e2754f-d0b3-44b4-88a4-c4c1f19b9b15_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/SecureSQL_[[:space:]]Evaluating[[:space:]]Data[[:space:]]Leakage[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]as[[:space:]]Natural[[:space:]]Language[[:space:]]Interfaces[[:space:]]to[[:space:]]Databases/96d3201b-2088-4dc3-a586-903ec4bd7d6e_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Securing[[:space:]]Multi-turn[[:space:]]Conversational[[:space:]]Language[[:space:]]Models[[:space:]]From[[:space:]]Distributed[[:space:]]Backdoor[[:space:]]Attacks/ee8f5c20-3e3a-4dfb-b3f3-d20b34d55f5c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SedarEval_[[:space:]]Automated[[:space:]]Evaluation[[:space:]]using[[:space:]]Self-Adaptive[[:space:]]Rubrics/b33177e1-fdfe-4c5f-a9a9-037a138aa0cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/See[[:space:]]Detail[[:space:]]Say[[:space:]]Clear_[[:space:]]Towards[[:space:]]Brain[[:space:]]CT[[:space:]]Report[[:space:]]Generation[[:space:]]via[[:space:]]Pathological[[:space:]]Clue-driven[[:space:]]Representation[[:space:]]Learning/081b7730-6494-4c1e-8ba3-d9c947e8be2f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Seeing[[:space:]]Through[[:space:]]VisualBERT_[[:space:]]A[[:space:]]Causal[[:space:]]Adventure[[:space:]]on[[:space:]]Memetic[[:space:]]Landscapes/3b29f9b0-75b7-4106-bc6f-27cbeed85edf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Selection-p_[[:space:]]Self-Supervised[[:space:]]Task-Agnostic[[:space:]]Prompt[[:space:]]Compression[[:space:]]for[[:space:]]Faithfulness[[:space:]]and[[:space:]]Transferability/14d2ceda-5f83-49b4-88b6-f58b5179b8e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Selective[[:space:]]Annotation[[:space:]]via[[:space:]]Data[[:space:]]Allocation_[[:space:]]These[[:space:]]Data[[:space:]]Should[[:space:]]Be[[:space:]]Triaged[[:space:]]to[[:space:]]Experts[[:space:]]for[[:space:]]Annotation[[:space:]]Rather[[:space:]]Than[[:space:]]the[[:space:]]Model/bcc9d306-7f45-4b34-b97c-c05fad825ac3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Consistency[[:space:]]Boosts[[:space:]]Calibration[[:space:]]for[[:space:]]Math[[:space:]]Reasoning/610e067d-96b2-49b7-8e9d-e89fb3c34d6e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Constructed[[:space:]]Context[[:space:]]Decompilation[[:space:]]with[[:space:]]Fined-grained[[:space:]]Alignment[[:space:]]Enhancement/45f4817a-44a7-499d-b884-1bdac2c1b3e5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Distillation[[:space:]]for[[:space:]]Model[[:space:]]Stacking[[:space:]]Unlocks[[:space:]]Cross-Lingual[[:space:]]NLU[[:space:]]in[[:space:]]200+[[:space:]]Languages/7e26119d-5c03-4bda-95f0-4a298ab50b80_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Evaluation[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]based[[:space:]]on[[:space:]]Glass-box[[:space:]]Features/f272d6b4-0f30-498a-aa3d-937a38ef3871_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Evolution[[:space:]]Fine-Tuning[[:space:]]for[[:space:]]Policy[[:space:]]Optimization/f429f3a3-904d-4b42-bb55-dde0fd8a6fb5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Explore_[[:space:]]Enhancing[[:space:]]Mathematical[[:space:]]Reasoning[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Fine-grained[[:space:]]Rewards/0050fd4f-4478-48e9-b7eb-9f80dd9808cb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Recognition[[:space:]]in[[:space:]]Language[[:space:]]Models/a0a4d18a-1b19-4a7c-aa49-4b891fff8e59_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-Renewal[[:space:]]Prompt[[:space:]]Optimizing[[:space:]]with[[:space:]]Implicit[[:space:]]Reasoning/d80875f6-b94b-48a6-a34d-1d65aadf1bd7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-contradictory[[:space:]]reasoning[[:space:]]evaluation[[:space:]]and[[:space:]]detection/d9c93459-c5e5-4d9c-b755-5c75dd1c3af2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-supervised[[:space:]]Preference[[:space:]]Optimization_[[:space:]]Enhance[[:space:]]Your[[:space:]]Language[[:space:]]Model[[:space:]]with[[:space:]]Preference[[:space:]]Degree[[:space:]]Awareness/e2cc6e0c-ce9c-4749-9583-e2bc905655b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-training[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Arithmetic[[:space:]]Reasoning/435df39d-e694-4d44-b90c-da0f58e68e46_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Self-training[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Knowledge[[:space:]]Detection/bd7c9ec5-45ff-492b-81e6-65ad0110050e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semantic[[:space:]]Token[[:space:]]Reweighting[[:space:]]for[[:space:]]Interpretable[[:space:]]and[[:space:]]Controllable[[:space:]]Text[[:space:]]Embeddings[[:space:]]in[[:space:]]CLIP/c718a104-e822-4497-883a-54264c48f76e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Semi-Supervised[[:space:]]Reward[[:space:]]Modeling[[:space:]]via[[:space:]]Iterative[[:space:]]Self-Training/a3a34752-c502-4aed-863f-abf1d4fa9b59_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shall[[:space:]]We[[:space:]]Team[[:space:]]Up_[[:space:]]Exploring[[:space:]]Spontaneous[[:space:]]Cooperation[[:space:]]of[[:space:]]Competing[[:space:]]LLM[[:space:]]Agents/0ac8f0f4-3a70-41bd-8f8c-104470db2f72_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/ShieldLM_[[:space:]]Empowering[[:space:]]LLMs[[:space:]]as[[:space:]]Aligned,[[:space:]]Customizable[[:space:]]and[[:space:]]Explainable[[:space:]]Safety[[:space:]]Detectors/88b018fb-f424-4975-83b8-33bd5367cebb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Shoes-ACOSI_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Aspect-Based[[:space:]]Sentiment[[:space:]]Analysis[[:space:]]with[[:space:]]Implicit[[:space:]]Opinion[[:space:]]Extraction/02314f5e-ac0e-4e97-bb41-23a58c851beb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Should[[:space:]]Cross-Lingual[[:space:]]AMR[[:space:]]Parsing[[:space:]]go[[:space:]]Meta_[[:space:]]An[[:space:]]Empirical[[:space:]]Assessment[[:space:]]of[[:space:]]Meta-Learning[[:space:]]and[[:space:]]Joint[[:space:]]Learning[[:space:]]AMR[[:space:]]Parsing/0b8a2e9a-d6b7-46a7-a567-7bdbd1c27864_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sing[[:space:]]it,[[:space:]]Narrate[[:space:]]it_[[:space:]]Quality[[:space:]]Musical[[:space:]]Lyrics[[:space:]]Translation/192fb3fe-fb4d-4309-9e2e-942c9b36293c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Skills-in-Context_[[:space:]]Unlocking[[:space:]]Compositionality[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/ee2eaec9-bf5e-4de1-b401-debeeb57046f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SnapNTell_[[:space:]]Enhancing[[:space:]]Entity-Centric[[:space:]]Visual[[:space:]]Question[[:space:]]Answering[[:space:]]with[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]Multimodal[[:space:]]LLM/21cb7539-81fe-4b99-9023-0f34f66fc335_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SocialGaze_[[:space:]]Improving[[:space:]]the[[:space:]]Integration[[:space:]]of[[:space:]]Human[[:space:]]Social[[:space:]]Norms[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/f7ebea1a-3f6d-4e3f-856c-4d73a4214dda_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Socratic[[:space:]]Human[[:space:]]Feedback[[:space:]](SoHF)_[[:space:]]Expert[[:space:]]Steering[[:space:]]Strategies[[:space:]]for[[:space:]]LLM[[:space:]]Code[[:space:]]Generation/7ff46ffa-27ed-4652-95d5-d8f970751dd4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Soda-Eval_[[:space:]]Open-Domain[[:space:]]Dialogue[[:space:]]Evaluation[[:space:]]in[[:space:]]the[[:space:]]age[[:space:]]of[[:space:]]LLMs/f4ddbc06-e400-42db-9cdd-fe21c8954833_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Solving[[:space:]]for[[:space:]]X[[:space:]]and[[:space:]]Beyond_[[:space:]]Can[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Solve[[:space:]]Complex[[:space:]]Math[[:space:]]Problems[[:space:]]with[[:space:]]More-Than-Two[[:space:]]Unknowns_/56c27add-d150-4f00-93f9-e76e9e385bc8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Sonnet[[:space:]]or[[:space:]]Not,[[:space:]]Bot_[[:space:]]Poetry[[:space:]]Evaluation[[:space:]]for[[:space:]]Large[[:space:]]Models[[:space:]]and[[:space:]]Datasets/752d16ba-a95c-402e-8bed-482c7c05ad64_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SpeciaLex_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]In-Context[[:space:]]Specialized[[:space:]]Lexicon[[:space:]]Learning/3203a6f9-ca35-43c6-b821-567b62adb213_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/StablePT[[:space:]]_[[:space:]]Towards[[:space:]]Stable[[:space:]]Prompting[[:space:]]for[[:space:]]Few-shot[[:space:]]Learning[[:space:]]via[[:space:]]Input[[:space:]]Separation/31db5709-f610-4100-a10c-7384ac2b57bc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Stanceformer_[[:space:]]Target-Aware[[:space:]]Transformer[[:space:]]for[[:space:]]Stance[[:space:]]Detection/ab1e9460-93e1-4e53-97d1-b56aaac4b202_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Stark_[[:space:]]Social[[:space:]]Long-Term[[:space:]]Multi-Modal[[:space:]]Conversation[[:space:]]with[[:space:]]Persona[[:space:]]Commonsense[[:space:]]Knowledge/cf904d82-7b65-40c6-8422-57caf94cd8db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Step-level[[:space:]]Value[[:space:]]Preference[[:space:]]Optimization[[:space:]]for[[:space:]]Mathematical[[:space:]]Reasoning/56a3c7d3-cd95-41c5-8f51-13d2b3c7b4f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Stochastic[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Language[[:space:]]Models[[:space:]]Using[[:space:]]Masked[[:space:]]Gradients/92554702-0933-4cf6-b4ad-f25f7dc2a17d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/StraGo_[[:space:]]Harnessing[[:space:]]Strategic[[:space:]]Guidance[[:space:]]for[[:space:]]Prompt[[:space:]]Optimization/1d5a6b6e-f34d-4deb-9044-8170dc47e3b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Structured[[:space:]]Chain-of-Thought[[:space:]]Prompting[[:space:]]for[[:space:]]Few-Shot[[:space:]]Generation[[:space:]]of[[:space:]]Content-Grounded[[:space:]]QA[[:space:]]Conversations/074575c5-2d56-47b0-adff-e2aaa25fc7fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Student[[:space:]]Data[[:space:]]Paradox[[:space:]]and[[:space:]]Curious[[:space:]]Case[[:space:]]of[[:space:]]Single[[:space:]]Student-Tutor[[:space:]]Model_[[:space:]]Regressive[[:space:]]Side[[:space:]]Effects[[:space:]]of[[:space:]]Training[[:space:]]LLMs[[:space:]]for[[:space:]]Personalized[[:space:]]Learning/c1ba5906-38f5-49e4-b5a9-aaef8f0e6e39_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Style-Compress_[[:space:]]An[[:space:]]LLM-Based[[:space:]]Prompt[[:space:]]Compression[[:space:]]Framework[[:space:]]Considering[[:space:]]Task-Specific[[:space:]]Styles/db4183b0-1196-4aff-8ab9-78bea07f02cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SummaCoz_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Improving[[:space:]]the[[:space:]]Interpretability[[:space:]]of[[:space:]]Factual[[:space:]]Consistency[[:space:]]Detection[[:space:]]for[[:space:]]Summarization/64aa231a-a3f1-4e0f-a5d0-7f8404848c90_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Suri_[[:space:]]Multi-constraint[[:space:]]Instruction[[:space:]]Following[[:space:]]in[[:space:]]Long-form[[:space:]]Text[[:space:]]Generation/a66f1397-8251-46f6-af07-3c429abc5217_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Symbolic[[:space:]]Prompt[[:space:]]Program[[:space:]]Search_[[:space:]]A[[:space:]]Structure-Aware[[:space:]]Approach[[:space:]]to[[:space:]]Efficient[[:space:]]Compile-Time[[:space:]]Prompt[[:space:]]Optimization/31c7d240-86d2-4ca0-a17c-d1e94a11dfaf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SynTQA_[[:space:]]Synergistic[[:space:]]Table-based[[:space:]]Question[[:space:]]Answering[[:space:]]via[[:space:]]Mixture[[:space:]]of[[:space:]]Text-to-SQL[[:space:]]and[[:space:]]E2E[[:space:]]TQA/ba9304a6-0ffc-402b-b9c4-f430c03d805b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/SynthEval_[[:space:]]Hybrid[[:space:]]Behavioral[[:space:]]Testing[[:space:]]of[[:space:]]NLP[[:space:]]Models[[:space:]]with[[:space:]]Synthetic[[:space:]]CheckLists/7634fa12-a33c-4f1f-89da-dfebe3959e05_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Synthetic[[:space:]]Multimodal[[:space:]]Question[[:space:]]Generation/70ee8c4e-4c7f-48a2-aac9-137ba5bed752_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TAP4LLM_[[:space:]]Table[[:space:]]Provider[[:space:]]on[[:space:]]Sampling,[[:space:]]Augmenting,[[:space:]]and[[:space:]]Packing[[:space:]]Semi-structured[[:space:]]Data[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Reasoning/c6bbf5bf-966a-4d89-800c-a441ec518e9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TARA_[[:space:]]Token-level[[:space:]]Attribute[[:space:]]Relation[[:space:]]Adaptation[[:space:]]for[[:space:]]Multi-Attribute[[:space:]]Controllable[[:space:]]Text[[:space:]]Generation/7ffc13bf-9682-4964-90b2-7119db7ee815_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TOOLVERIFIER_[[:space:]]Generalization[[:space:]]to[[:space:]]New[[:space:]]Tools[[:space:]]via[[:space:]]Self-Verification/89099054-dfc3-42a9-b43a-9b2b84c2254a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TOWER_[[:space:]]Tree[[:space:]]Organized[[:space:]]Weighting[[:space:]]for[[:space:]]Evaluating[[:space:]]Complex[[:space:]]Instructions/fd7e4aad-5add-4e88-84fc-cac11bc53ad2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TRACE[[:space:]]the[[:space:]]Evidence_[[:space:]]Constructing[[:space:]]Knowledge-Grounded[[:space:]]Reasoning[[:space:]]Chains[[:space:]]for[[:space:]]Retrieval-Augmented[[:space:]]Generation/824d577f-24d7-4223-95f9-636a2ef8cf8f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TRIP[[:space:]]NEGOTIATOR_[[:space:]]A[[:space:]]Travel[[:space:]]Persona-aware[[:space:]]Reinforced[[:space:]]Dialogue[[:space:]]Generation[[:space:]]Model[[:space:]]for[[:space:]]Personalized[[:space:]]Integrative[[:space:]]Negotiation[[:space:]]in[[:space:]]Tourism/32d23a0c-4721-43b7-8311-11808513cfbf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TROPE_[[:space:]]TRaining-Free[[:space:]]Object-Part[[:space:]]Enhancement[[:space:]]for[[:space:]]Seamlessly[[:space:]]Improving[[:space:]]Fine-Grained[[:space:]]Zero-Shot[[:space:]]Image[[:space:]]Captioning/16c4526d-9c1d-4fe6-91da-652cd4aca304_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TS-Align_[[:space:]]A[[:space:]]Teacher-Student[[:space:]]Collaborative[[:space:]]Framework[[:space:]]for[[:space:]]Scalable[[:space:]]Iterative[[:space:]]Finetuning[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/50109634-d613-4f49-b559-688f98d6a47f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TWBias_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Assessing[[:space:]]Social[[:space:]]Bias[[:space:]]in[[:space:]]Traditional[[:space:]]Chinese[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]a[[:space:]]Taiwan[[:space:]]Cultural[[:space:]]Lens/8f2ab6af-954f-4361-81fb-c44a6fed3e9b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Tab2Text[[:space:]]-[[:space:]]A[[:space:]]framework[[:space:]]for[[:space:]]deep[[:space:]]learning[[:space:]]with[[:space:]]tabular[[:space:]]data/ae425a3d-5c78-4413-85cb-f5246d5599cd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Taking[[:space:]]a[[:space:]]Deep[[:space:]]Breath_[[:space:]]Enhancing[[:space:]]Language[[:space:]]Modeling[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Sentinel[[:space:]]Tokens/0bba3424-48ff-481c-acb6-7c625a961f17_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Taking[[:space:]]a[[:space:]]turn[[:space:]]for[[:space:]]the[[:space:]]better_[[:space:]]Conversation[[:space:]]redirection[[:space:]]throughout[[:space:]]the[[:space:]]course[[:space:]]of[[:space:]]mental-health[[:space:]]therapy/21b22500-890d-4883-8dc8-106830b31445_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Talking[[:space:]]the[[:space:]]Talk[[:space:]]Does[[:space:]]Not[[:space:]]Entail[[:space:]]Walking[[:space:]]the[[:space:]]Walk_[[:space:]]On[[:space:]]the[[:space:]]Limits[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Lexical[[:space:]]Entailment[[:space:]]Recognition/1a9078c6-d036-441f-a713-2cad2e47f923_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Targeted[[:space:]]Multilingual[[:space:]]Adaptation[[:space:]]for[[:space:]]Low-resource[[:space:]]Language[[:space:]]Families/3fa94ed9-dfc1-41ae-9e97-5cbca8550400_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Temperature-Centric[[:space:]]Investigation[[:space:]]of[[:space:]]Speculative[[:space:]]Decoding[[:space:]]with[[:space:]]Knowledge[[:space:]]Distillation/29fceb02-eb3a-4e93-a423-44d0c614e38c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Temporal[[:space:]]Cognitive[[:space:]]Tree_[[:space:]]A[[:space:]]Hierarchical[[:space:]]Modeling[[:space:]]Approach[[:space:]]for[[:space:]]Event[[:space:]]Temporal[[:space:]]Relation[[:space:]]Extraction/8e33950c-eb4a-4a0a-8804-5d3cef263c26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Temporal[[:space:]]Fact[[:space:]]Reasoning[[:space:]]over[[:space:]]Hyper-Relational[[:space:]]Knowledge[[:space:]]Graphs/6836f680-16b4-408e-96f4-b1314ddaab3f_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/SedarEval_ Automated Evaluation using Self-Adaptive Rubrics/b33177e1-fdfe-4c5f-a9a9-037a138aa0cf_content_list.json

@@ -0,0 +1,2305 @@
+[
+    {
+        "type": "text",
+        "text": "SedarEval: Automated Evaluation using Self-Adaptive Rubrics",
+        "text_level": 1,
+        "bbox": [
+            171,
+            90,
+            825,
+            110
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhiyuan Fan $^{1}$ , Weinong Wang $^{2*}$ , Xing Wu $^{3}$ , Debing Zhang $^{2}$",
+        "bbox": [
+            242,
+            133,
+            757,
+            152
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Hong Kong University of Science and Technology,  $^{2}$ Xiaohongshu Inc,",
+        "bbox": [
+            210,
+            167,
+            786,
+            184
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$ Institute of Information Engineering, Chinese Academy of Sciences zhiyuan.fan@connect.ust.hk, weinong.wang@hotmail.com",
+        "bbox": [
+            210,
+            184,
+            776,
+            217
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            261,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The evaluation paradigm of LLM-as-judge gains popularity due to its significant reduction in human labor and time costs. This approach utilizes one or more large language models (LLMs) to assess the quality of outputs from other LLMs. However, existing methods rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process, compromising precision and stability in assessments. Inspired by human examination scoring processes, we propose a new evaluation paradigm based on self-adaptive rubrics. Specifically, we create detailed scoring rubrics for each question, capturing the primary and secondary criteria in a structured format of scoring and deduction points that mimic a human evaluator's analytical process. Building on this paradigm, we further develop a novel benchmark called SedarEval, which covers a range of domains including long-tail knowledge, mathematics, coding, and logical reasoning. SedarEval consists of 1,000 meticulously crafted questions, each with its own self-adaptive rubric. To further streamline the evaluation, we train a specialized evaluator language model (evaluator LM) to supplant human graders. Using the same training data, our evaluator LM achieves a higher concordance rate with human grading results than other paradigms, including GPT-4, highlighting the superiority and efficiency of our approach.",
+        "bbox": [
+            144,
+            284,
+            460,
+            739
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            747,
+            258,
+            762
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The rapid advancements in large language models (LLMs) have led to their widespread use (OpenAI et al., 2024; Team et al., 2023; Anthropic, 2024; Bai et al., 2023). However, assessing these models in open-ended question-answering scenarios poses a significant challenge. Automated metric-based evaluations offer speed and convenience but often fall short due to the diversity of ground truth",
+        "bbox": [
+            112,
+            772,
+            487,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(Schluter, 2017a; Reiter, 2018; Montahaei et al., 2019; Freitag et al., 2020). In contrast, human-based evaluations provide reliable assessments but require substantial resources.",
+        "bbox": [
+            507,
+            261,
+            882,
+            324
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To bridge the gap, the LLM-as-a-judge paradigm attempts to strike a balance between automated and human evaluation. Prominent examples of this approach include MT-bench (Zheng et al., 2024) and Arena (Chiang et al., 2024), which leverage proprietary models to evaluate individual or comparative model responses. These benchmarks use pre-defined principles, such as the 3H principle (human-like, helpful, harmonious), to determine responses that align best with realistic human preferences. The widespread use of GPT-4 (OpenAI et al., 2024) as an evaluator in these studies presents challenges, including high costs for research institutions and potential data leaks.",
+        "bbox": [
+            507,
+            326,
+            884,
+            549
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Some studies (Zhu et al., 2023; Li et al., 2023a; Wang et al., 2024; Kim et al., 2024a,b) propose using open-source pretrained models (Touvron et al., 2023; Bai et al., 2023; Zeng et al., 2022) to train specialized evaluator LMs, offering a more cost-effective and secure solution. However, these methods typically use a uniform, question-agnostic rubric to guide the scoring process, overlooking the unique characteristics of each question. Each question has different emphases, with primary and secondary scoring points. A general rubric applies uniform criteria, failing to accurately reflect human preferences.",
+        "bbox": [
+            507,
+            551,
+            882,
+            759
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To adaptively align the scoring process with human judgment, we propose a novel evaluation paradigm based on self-adaptive rubrics. Unlike coarse-grained general rubrics, we provide fine-grained rubrics for each task, detailing specific scoring and penalty points with primary and secondary information. By analyzing focus points, we assign different values to each point. Additionally, we introduce penalty points to penalize models for generating rejected responses. The scoring pro",
+        "bbox": [
+            507,
+            760,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author.",
+        "bbox": [
+            136,
+            906,
+            280,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "16916",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 16916-16930",
+        "bbox": [
+            210,
+            945,
+            786,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            273,
+            959,
+            724,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/bb722d898fb4f2cae162dd772a0b5a02bb62c55d30cfeb44bb1dfbb37d5238d9.jpg",
+        "image_caption": [
+            "Figure 1: Automated evaluation pipeline using self-adaptive rubrics. This pipeline dynamically adjusts the evaluation rubric based on the input question, resulting in a scoring process that aligns more closely with human evaluators."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            87,
+            872,
+            340
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "cess considers both preferred and rejected perspectives. The inconsistent coverage of positive and penalty points ensures a more refined constraint on the scoring process. These detailed scoring trajectories simplify the evaluation process to an instruction-following task, reducing dependency on a judge model's internal knowledge and skills, leading to more accurate and stable assessments. Building on this paradigm, we construct a new benchmark called SedarEval, presenting a taxonomy comprising 5 major categories, with a dataset of 1000 queries, that fully aligns with realistic scenarios.",
+        "bbox": [
+            112,
+            417,
+            489,
+            625
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We further conduct ablation experiments on each component of the LLM-as-a-judge paradigm to investigate training a specialized LLM for scoring, revealing their respective importance. We analyze whether LLMs can correctly evaluate questions they can correctly answer and find that insufficient diversity in existing SFT data and a lack of evaluation-format data limit model performance. We also propose human-AI consistency to ensure evaluator LLMs maintain alignment with human preferences while leveraging their chain of thought capability to improve evaluation performance. Based on these findings, we develop a specialized evaluator LLM tailored to the benchmark for automated scoring. This model surpasses GPT-4 in model-level and question-level Pearson correlation, GSB, and ACC metrics, demonstrating higher consistency with human judgment. Ex",
+        "bbox": [
+            112,
+            631,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "perimental results validate the effectiveness and efficiency of our proposed paradigm.",
+        "bbox": [
+            507,
+            417,
+            880,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are summarized as follows:",
+        "bbox": [
+            527,
+            451,
+            867,
+            464
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We propose a novel evaluation paradigm using self-adaptive rubrics for each question, offering granular guidance and closely aligning the scoring process with human evaluation.",
+            "2. We develop a high-quality benchmark called SedarEval, featuring 1,000 meticulously crafted questions with detailed rubrics, and conduct manual evaluations on 20 LLMs.",
+            "3. We analyze the training of evaluator LMs, highlight existing methods' shortcomings, and use the self-adaptive rubrics paradigm to train an evaluator LM that surpasses GPT-4 in agreement with human evaluations."
+        ],
+        "bbox": [
+            522,
+            480,
+            882,
+            717
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            732,
+            665,
+            747
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Benchmark LLMs Capabilities. With the rapid advancement of LLMs (OpenAI et al., 2024; Team et al., 2023; Anthropic, 2024), it has become a substantial challenge to benchmark their broad capabilities reliably. NLU-style tasks (Hendrycks et al., 2020; Huang et al., 2024; Srivastava et al., 2022; Zhong et al., 2023), such as multi-choice QA, employ general-exam questions from various domains to assess a model's knowledge and comprehension abilities. However, their real-world usage is limited",
+        "bbox": [
+            507,
+            760,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "16917",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "due to misalignment with human preferences. Recently, reference-free benchmarks (Li et al., 2023b; Chiang et al., 2023; Zheng et al., 2024; Ye et al., 2023) have been proposed to evaluate texts' quality in a generative setting directly. Unlike previous datasets, our benchmark provides a comprehensive and stable model assessment with its diverse test cases and broad label distribution.",
+        "bbox": [
+            112,
+            84,
+            492,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Automatic NLG Evaluation. It's notably challenging to evaluate the quality of generated text in the field of natural language generation (NLG). Traditional n-gram-based metrics (Papineni et al., 2002; Lin, 2004; Snover et al., 2006) and embedding-based metrics (Li et al., 2019; Zhang et al., 2020; Risch et al., 2021) can only assess lexical or semantic similarity between the generated answers and reference answers (Schluter, 2017a; Reiter, 2018; Montahaei et al., 2019; Freitag et al., 2020). These metrics have been found to have a relatively low correlation with human preferences (Liu et al., 2023a). Recently, employing LLM as a judge (Zheng et al., 2023; Li et al., 2023b; Chan et al., 2023) is a novel evaluation paradigm that has gained widespread application. The most common approach involves using proprietary LLMs, such as GPT-4 (OpenAI et al., 2024), as judge models to rank or score outputs generated by other models. However, this method relies on closed-source models, incurs high costs, and poses risks of internal evaluation dataset leaks for companies developing LLMs. To address these issues, various works (Zhu et al., 2023; Li et al., 2023a; Wang et al., 2024; Kim et al., 2024a,b) have proposed training dedicated scoring models on open-source base models using synthetic or manually labeled data. These evaluations often use reference answers to assist in the assessment or employ general rubrics to guide the scoring process. However, these approaches overlook the differences between individual questions and the varying scoring criteria of each question, even within the same category. In contrast, we propose an evaluation paradigm based on self-adaptive rubrics that generates fine-grained, customizable rubrics for each question, guiding a more precise scoring process. It is worth noting that although Prometheus 2 also claims to use fine-grained rubrics, their rubrics remain question-agnostic.",
+        "bbox": [
+            110,
+            212,
+            490,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Quantifying Evaluation Confidence. The automatic metrics are imperfect, and we must measure their performance further. A gold standard for this is their alignment with human judgment",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "and the confidence level we can have when these metrics guide our decision-making process. However, quantifying this performance (Krishna et al., 2021; Schluter, 2017b; Stureborg et al., 2024) is difficult due to various factors (the evaluator's accuracy and stability, evaluation set size, the extent of the performance difference among competing models, etc.). (Kocmi et al., 2021; Deutsch et al., 2021; Zhang and Vogel, 2004) investigate the correlation between human judgment and traditional automatic metrics such as ROUGE and BLEU and analyze their confidence intervals. For LLM-based evaluators, commonly used metrics include Pearson, Spearman, and Kdendall-Tau to measure the alignment between the model's scores and human preferences. However, previous work has primarily focused on the correlation of rankings or overall scores at the model level without comparing the scores with human ratings at the individual question. This limits the interpretability of the scoring process and hampers its utility in guiding the development and iteration of LLMs.",
+        "bbox": [
+            507,
+            84,
+            885,
+            439
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 SedarEval Benchmark",
+        "text_level": 1,
+        "bbox": [
+            507,
+            453,
+            741,
+            469
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce SedarEval, a benchmark constructed upon the self-adaptive rubrics paradigm. We begin by delving into the intricacies of the self-adaptive rubric paradigm, followed by a detailed explanation of the benchmark's core components – questions and their corresponding rubrics – along with the methodology for model evaluation using this benchmark. To ensure the quality of SedarEval, we incorporate comprehensive human assessment into the construction process, meticulously filtering out samples that fail to meet the established quality standards.",
+        "bbox": [
+            505,
+            481,
+            885,
+            675
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Self-Adaptive Rubrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            688,
+            732,
+            703
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Previous LLM-as-a-judge approaches, which rely on general rubrics or principles for scoring, often lack specific, problem-related rubric guidance. Consequently, these methods depend heavily on the inherent capabilities of the LLM itself, leading to potential errors in evaluations due to insufficient reasoning abilities or hallucinations. Additionally, this approach introduces extraneous biases, such as position bias and order bias.",
+        "bbox": [
+            507,
+            711,
+            884,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Self-adaptive rubrics address these issues by tailoring the evaluation criteria to the specific problems at hand, incorporating the focal points of the problem and assigning different weights accordingly.",
+        "bbox": [
+            507,
+            857,
+            885,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "16918",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "By introducing penalty points, these rubrics align more closely with human judgments by deducting points for outputs that deviate from expected tendencies. To prevent human evaluators (or LLMs) from making incorrect assessments due to a lack of background information, additional context is provided for each question to assist in the scoring process. A typical self-adaptive rubric comprises three components: scoring points, penalty points, and background knowledge, as illustrated in Table 3.",
+        "bbox": [
+            112,
+            84,
+            489,
+            261
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Dataset Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            272,
+            329,
+            286
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Questions: We have defined a classification system for objective questions, with a two-tiered scoring system as shown in the diagram. Under each secondary classification, we have hired five people to create questions. Specifically, each person is required to first create their own questions to get a question pool, and then each person votes on all the questions. We only keep the questions that all five people agree on.",
+        "bbox": [
+            112,
+            293,
+            489,
+            437
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each candidate question, the annotators will select 5 LLMs to test the effectiveness of the questioned question. We only keep the questions with a larger variance in scores, which are more discriminating, and remove the questions where the answers from different models are almost the same, which are not helpful in distinguishing between different models. For example, if a question can be answered correctly by all models, or incorrectly by all models, then this question cannot show which model is better.",
+        "bbox": [
+            112,
+            438,
+            489,
+            613
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After collecting the initial questions, we hired another group of people to compare all the questions in pairs to judge the similarity of the problem-solving ideas for the two questions and delete the questions with too much similarity.",
+        "bbox": [
+            112,
+            615,
+            489,
+            695
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Rubrics: For each question, we assign it to three individuals to discuss together and generate a rubric like the one shown in Figure 1.",
+        "bbox": [
+            112,
+            696,
+            487,
+            743
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For more detailed information, please refer to Appendix D, which contains benchmark statistics and the leaderboard.",
+        "bbox": [
+            112,
+            745,
+            489,
+            791
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Evaluation Pipeline",
+        "text_level": 1,
+        "bbox": [
+            112,
+            804,
+            315,
+            819
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The entire evaluation pipeline using our benchmark is illustrated in Figure 1. Given a question, its corresponding rubrics, and the model to be evaluated, we first input the question into the model to generate a response. The response is then scored according to the predefined rubric, either by human",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "evaluators or using LLMs. Finally, all the scores are aggregated to obtain the model's total score.",
+        "bbox": [
+            507,
+            84,
+            882,
+            116
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Evaluator Language Model",
+        "text_level": 1,
+        "bbox": [
+            507,
+            129,
+            781,
+            146
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce an evaluator LM aligned with the self-adaptive rubrics paradigm to substitute human evaluators. We begin by delineating the evaluation format. Subsequently, we propose a novel data filtering strategy to align the Chain-of-Thought evaluation process with human judgments. Finally, we discuss the automation of rubric generation.",
+        "bbox": [
+            505,
+            156,
+            884,
+            285
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Evaluation Format",
+        "text_level": 1,
+        "bbox": [
+            507,
+            297,
+            705,
+            311
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The evaluation format consists of two types: direct scoring of individual model outputs and pairwise comparison of model outputs to determine the superior one. Pairwise evaluation requires significantly more comparisons as the number of candidate models increases, as shown by Equation 1. Therefore, we employ direct assessment in this paper. Notably, direct assessment scores can be compared to derive pairwise results.",
+        "bbox": [
+            505,
+            319,
+            884,
+            462
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nC (n, 2) = \\frac {n !}{2 ! (n - 2) !} - n = \\frac {n ^ {2} - 3 n}{2} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            549,
+            476,
+            882,
+            510
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use a reference-based format to organize the output. Specifically, for each question, we compile the reference answer, self-adaptive rubrics, and scoring examples to create an auto-prompt template. When evaluating answers, we incorporate the answers into this auto-prompt template as the complete input. We conduct ablation experiments on each component in zero-shot, few-shot, and instruction tuning settings.",
+        "bbox": [
+            507,
+            517,
+            885,
+            663
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Human-AI Consistency",
+        "text_level": 1,
+        "bbox": [
+            507,
+            674,
+            741,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Human annotators provide specific scores for each response without corresponding explanations, which is efficient but suboptimal for training evaluator LMs. To alleviate this issue, we use GPT-4 to generate detailed reasoning steps using Chain-of-Thought. However, scoring preferences may differ between GPT-4 and human annotators, and both may make errors. To mitigate these errors and align the scoring process with human judgment, we introduce a Human-AI Consistency strategy to improve synthetic data quality. We extract final scores from the GPT-4 scoring process and compare them with human scores, retaining only the data where GPT-4 and human results are consistent,",
+        "bbox": [
+            505,
+            696,
+            884,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "16919",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "as shown in Equation 2, where  $\\mathcal{H}$  represents human scores,  $\\mathcal{A}$  represents AI scores, and  $\\mathbb{I}$  is an indicator function.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {T} = \\{(h, a) \\mid h \\in \\mathcal {H}, a \\in \\mathcal {A}, \\mathbb {I} (h, a) = 1 \\} \\quad (2)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            132,
+            158,
+            487,
+            175
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This approach only retains instances where human and AI scores are consistent and differs from rejection sampling, which uses human scores as a reward function to select the optimal output from multiple GPT-4 results.",
+        "bbox": [
+            112,
+            185,
+            489,
+            265
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Automatic Rubric Generation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            275,
+            394,
+            290
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To reduce human annotation costs, we investigate using human-annotated datasets to train a model for automatic self-adaptive rubric generation. By providing the model with questions and corresponding reference answers, we train it to produce rubrics that delineate scoring criteria and identify deduction points.",
+        "bbox": [
+            112,
+            296,
+            487,
+            407
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Generating self-adaptive rubric format output is straightforward, but aligning rubrics with human preferences requires aligning the model with human evaluative criteria. This complexity arises because identifying scoring points, assigning specific weights, and criteria for deductions are significantly influenced by human judgment.",
+        "bbox": [
+            112,
+            409,
+            487,
+            520
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The training process for the automatic rubric generation model comprises two stages. Initially, we use human-labeled data to train a base model through Supervised Fine-Tuning (SFT), as depicted in Equation 3.",
+        "bbox": [
+            112,
+            521,
+            489,
+            602
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} (\\theta) = - \\sum_ {i = 1} ^ {N} \\log p _ {\\theta} \\left(y _ {i} \\mid x _ {i}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            613,
+            487,
+            656
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The base model generates rubrics that conform to the specified format, though they may not fully align with human preferences (quantitative metrics will be introduced in Section 5.2). In the next phase, rubrics generated by the base model are treated as rejected responses, while human-labeled rubrics serve as preferred responses to construct preference pairs. We then train the model using Direct Preference Optimization (DPO) to align it with human preferences, as shown in Equations 4 and 5.",
+        "bbox": [
+            112,
+            659,
+            489,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (y, x) = \\beta \\log \\frac {\\pi_ {\\theta} (y \\mid x)}{\\pi_ {\\mathrm {r e f}} (y \\mid x)} \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            196,
+            848,
+            487,
+            883
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} \\left(\\pi_ {\\theta}\\right) = - \\log \\sigma \\left(f \\left(y _ {w}, x\\right) - f \\left(y _ {l}, x\\right)\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            134,
+            904,
+            487,
+            922
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We also explore automating rubric generation using GPT-4 without reference answers. To ensure accuracy, GPT-4 creates both the rubric and an ideal answer for each question. If the ideal answer corresponds with the ground truth, the generated rubric is deemed acceptable. We employ a self-refinement strategy to help the model iteratively refine its outputs, aligning it with human preferences. For detailed algorithmic procedures and prompts, refer to Appendix E.1.",
+        "bbox": [
+            507,
+            84,
+            884,
+            244
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experiment",
+        "text_level": 1,
+        "bbox": [
+            509,
+            255,
+            648,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Experimental Setting",
+        "text_level": 1,
+        "bbox": [
+            507,
+            280,
+            722,
+            296
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We train the Evaluator Language Model and the Rubric Generation Model using both the opensource model LLaMA-3 (Touvron et al., 2023) and our internal model  $\\mathrm{XDG}^{1}$ . To maximize training efficiency and utilize hardware resources, we implement tensor parallelism (Shoeybi et al., 2020) with PyTorch 2.3 (Paszke et al., 2019). For the 7B/8B models, we use 128 H100 GPUs, while for the 70B models, we use 512 H100 GPUs. For the models' chat versions (i.e., instruction-tuned), we employ the same chat markup language (ChatML) as the models themselves. For the pre-trained versions, we use a unified ChatML to reduce data bias. We adopt adaptive learning rate and batch size strategies. Further training details are provided in Appendix A.",
+        "bbox": [
+            507,
+            300,
+            884,
+            557
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.2 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            567,
+            705,
+            581
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To assess the performance of the evaluator language model, we use Pearson's correlation coefficient and Spearman's rank correlation coefficient. These statistical measures assess the consistency between the outcomes of the evaluator language model and those obtained from human evaluators.",
+        "bbox": [
+            507,
+            588,
+            882,
+            683
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Each question is accompanied by a detailed rubric specifying exact scoring and deduction criteria, so we use accuracy to evaluate the model's capability in following these self-adaptive rubrics for scoring. Considering potential noise in the model scoring, we introduce a weaker threshold ACC, which considers a result correct if it falls within a specified range. The calculation formulas are presented in Equation 6.",
+        "bbox": [
+            507,
+            684,
+            882,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {A C C} _ {t} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\{ \\begin{array}{l l} 1, & \\text {i f} | y _ {\\text {p r e d} _ {i}} - y _ {\\text {t r u e} _ {i}} | \\leq \\epsilon \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            847,
+            882,
+            890
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>The name of this model has been anonymized to ensure confidentiality.",
+        "bbox": [
+            507,
+            896,
+            880,
+            922
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "16920",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To facilitate the iterative enhancement of LLMs using our benchmark, a robust metric is essential to assess whether the current model version outperforms its predecessor. Therefore, we adopt the widely used GSB (Good, Same, Bad) metric to compare model performance. Given two models, A and B, the calculation formula is presented in Equation 7. In this context, \"#good\" signifies that model A surpasses model B, \"#bad\" denotes the contrary, and \"#same\" indicates equivalent performance between the models.",
+        "bbox": [
+            112,
+            84,
+            492,
+            261
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Delta G S B = \\frac {\\# \\text {g o o d} - \\# \\text {b a d}}{\\# \\text {g o o d} + \\# \\text {s a m e} + \\# \\text {b a d}} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            159,
+            273,
+            489,
+            307
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the quality of automatically generated rubrics, we draw on the ACU (Liu et al., 2023b) and FactScore (Min et al., 2023) paradigms, using GPT-4 to calculate the match between the generated rubrics and the ground truth rubrics. The calculation formula is specified in Equation 8, where GT represents the correct rubric set containing multiple {grading points: specific score} pairs, and AT denotes the automatically generated rubric set.  $\\mathbb{I}(i\\in GT)$  is an indicator function that equals 1 if the item  $i$  from AT is present in GT, and 0 otherwise. The prompts used for this evaluation are detailed in Appendix C.2.",
+        "bbox": [
+            112,
+            312,
+            489,
+            521
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {M a t c h} (G T, A T) = \\frac {\\sum_ {i \\in A T} \\mathbb {I} (i \\in G T)}{| G T |} \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            152,
+            533,
+            487,
+            569
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.3 Selected Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            577,
+            287,
+            590
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Previous studies predominantly employ English-proficient models to generate <question, response, score> triples for training evaluator language models, often overlooking models proficient in Chinese. Additionally, several studies exclusively use GPT-3.5 or GPT-4 to construct such synthetic data. These data generation methodologies may cause discrepancies between the synthetic and real-world data distributions, introducing biases into the trained evaluator language models.",
+        "bbox": [
+            112,
+            599,
+            489,
+            759
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To alleviate this issue, we utilize a broader range of LLMs to collect responses that better reflect real-world distributions. This approach ensures greater diversity and mitigates biases introduced by relying solely on synthetic data generated from a single model. Specifically, we choose GPT-4, GPT-4-turbo, GPT-4-o, Claude Opus, DeepSeek 2.0, MiniMax 6.5, MiniMax 6, Doubao, GPT-3.5, Tongyi Qianwen 2.0, and Tongyi Qianwen 1.5-100B/70B. This selection includes models proficient in differ",
+        "bbox": [
+            112,
+            760,
+            490,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ent languages and multiple versions of the same model.",
+        "bbox": [
+            507,
+            84,
+            880,
+            112
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For open-source models, we use local deployment to infer responses. For proprietary LLMs, if API services are available, we collect model outputs by requesting the API. If only a web interface is provided, we employ people to gather the outputs.",
+        "bbox": [
+            507,
+            117,
+            880,
+            197
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6 Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            208,
+            618,
+            225
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct a comprehensive experimental analysis of the robustness of the proposed benchmark evaluations, examining the data distribution, training phases, and training paradigms of evaluator LMs. Our findings reveal limitations in current training methodologies for evaluator LMs. Building on these insights, we develop an evaluator LM aligned with the self-adaptive rubrics paradigm.",
+        "bbox": [
+            507,
+            233,
+            884,
+            378
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.1 Scaling Law for Robust Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            388,
+            830,
+            404
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A robust benchmark should effectively distinguish the capabilities of different models and maintain stability to ensure consistent rankings rather than allowing fluctuations due to the instability of individual tasks. To achieve this, the benchmark needs a sufficiently broad distribution while minimizing extraneous biases.",
+        "bbox": [
+            507,
+            409,
+            882,
+            520
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To verify the robustness of the proposed benchmark, we conduct two rounds of sampling without replacement from a pool of 1,000 questions. In each round, we select  $n$  questions, resulting in a total of  $2n$  independent questions, where  $n \\in [10,500]$ . We then compare the consistency of the model rankings obtained from these two samples.",
+        "bbox": [
+            507,
+            521,
+            884,
+            650
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/b0762fe1c2d162e955cdaf38f3906ab8a0255ca6357adcbf8544267b772456fa.jpg",
+        "image_caption": [
+            "Figure 2: Consistency of model rankings as  $n$  increases. After  $n$  reaches approximately 300, the consistency stabilizes with only minor fluctuations.",
+            "Figure 2 shows the variation in the consistency of model rankings under different question sets as  $n$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            668,
+            873,
+            814
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "16921",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "increases. When  $n$  is relatively small, the consistency is low, indicating the inconsistency caused by biases in different distributions. As  $n$  increases, the consistency improves despite the two sets remaining independent. After  $n$  reaches approximately 300, the consistency stabilizes with only minor fluctuations. This demonstrates the scaling law for robust evaluation, indicating that as the number of questions increases, the evaluation results tend to stabilize due to the broader coverage of the distribution.",
+        "bbox": [
+            112,
+            84,
+            492,
+            261
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2 Score Distribution Shift",
+        "text_level": 1,
+        "bbox": [
+            112,
+            273,
+            347,
+            288
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/3fb42723dbc22e32d626e5726793e5ebd8a6dfd4611b27479317e6f01955eda6.jpg",
+        "image_caption": [
+            "Figure 3: Data distribution comparison of different data."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            304,
+            487,
+            391
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The Prometheus approach, which relies on general rubrics not specifically tailored to the problem at hand, employs GPT-4 to generate an equal amount of synthetic data across different scores (1-5) to mitigate score bias from the evaluating language model. In contrast, our method uses self-adaptive rubrics, and our responses are genuinely collected from the model rather than artificially synthesized. Consequently, we cannot ensure that the quantity of data for each score is perfectly balanced.",
+        "bbox": [
+            112,
+            437,
+            487,
+            596
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "However, as illustrated in Figure 3, we observe that despite the score distribution shift between the training and test data, the score distribution of the model outputs, trained using the self-adaptive rubrics paradigm, closely aligns with the human-provided ground truth. This finding substantiates the robustness and efficacy of the self-adaptive rubrics paradigm in automated scoring.",
+        "bbox": [
+            112,
+            598,
+            489,
+            727
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.3 Out of Distribution Evaluation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            739,
+            401,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We establish two dimensions for evaluating the out-of-distribution capabilities of our model: model-level and question-level. For the model-level evaluation, we utilize the same set of questions, selecting a subset of models to train the evaluator language model (LM), and subsequently test on the remaining unselected models. In the question-level evaluation, a subset of questions along with all associated models are used for training, and the scoring performance is then assessed on a different",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "set of questions.",
+        "bbox": [
+            507,
+            85,
+            633,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1 presents the experimental results, showing that under the self-adaptive rubrics paradigm, the model performs well in both model-level and question-level evaluations. This indicates that our proposed method has strong generalization capabilities.",
+        "bbox": [
+            507,
+            99,
+            884,
+            195
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.4 Merged SFT or Continual SFT",
+        "text_level": 1,
+        "bbox": [
+            507,
+            206,
+            798,
+            222
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Previous research shows that a model might generate a correct answer but fail to accurately evaluate the <question, answer> pair for the same question. We argue that this issue mainly arises from the insufficient diversity of the SFT data.",
+        "bbox": [
+            507,
+            228,
+            882,
+            306
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To validate this, we conduct the following experiments:",
+        "bbox": [
+            507,
+            307,
+            882,
+            338
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Training a pretrained language model (PLM) using only traditional SFT data.",
+            "2. Training a PLM using a mix of SFT data and evaluator LM format data.",
+            "3. Performing continual SFT on an instruction-tuned model using evaluator LM format data, a widely adopted approach in other studies (Kim et al., 2024b,a)."
+        ],
+        "bbox": [
+            522,
+            348,
+            882,
+            494
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 2, we find that although the model using continual SFT performs well on evaluation tasks, its general ability significantly declines, limiting its versatility. However, starting from a PLM and using a mix of SFT data and evaluator LM format data for SFT results in excellent evaluation capability with minimal impact on general ability. This reveals the shortcomings of the previous continual SFT approach and suggests that the model's inability to evaluate the questions it can answer may simply be due to the lack of such data, highlighting the importance of diversity in SFT data.",
+        "bbox": [
+            507,
+            504,
+            882,
+            712
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We employ Human-AI Consistency to filter the evaluator LM and find that, compared to using raw Chain-of-Thought data generated by GPT-4 and data filtered by rejection sampling, the data selected using Human-AI Consistency shows significant improvements in both evaluation and general capability, demonstrating the effectiveness of this strategy.",
+        "bbox": [
+            507,
+            713,
+            882,
+            841
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.5 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            852,
+            673,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct detailed ablation experiments on the components of self-adaptive rubrics, namely, reference answers, rubrics, and in-context examples.",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "16922",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/cea7e13b2b4ba40d6727bd6fc5cbbbc2afc88163c5836a8e2e4f70c676c828a1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Type</td><td colspan=\"4\">Question-level</td><td colspan=\"4\">Model-level</td></tr><tr><td>GSB</td><td>ACC</td><td>ACC(t)</td><td>pearson</td><td>GSB</td><td>ACC</td><td>ACC(t)</td><td>pearson</td></tr><tr><td>XDG</td><td>0.952</td><td>0.590</td><td>0.794</td><td>0.738</td><td>0.952</td><td>0.590</td><td>0.794</td><td>0.380</td></tr><tr><td>GPT-3.5</td><td>0.829</td><td>0.422</td><td>0.663</td><td>0.566</td><td>0.829</td><td>0.422</td><td>0.663</td><td>0.566</td></tr><tr><td>GPT-4</td><td>0.952</td><td>0.654</td><td>0.855</td><td>0.822</td><td>0.952</td><td>0.654</td><td>0.855</td><td>0.822</td></tr></table>",
+        "bbox": [
+            181,
+            80,
+            816,
+            167
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/d3d9ce3942338c1d8d82f6d532563a0e3129aa673fed15072573631590734929.jpg",
+        "table_caption": [
+            "Table 1: Out of distribution evaluation performance in both model-level and question-level."
+        ],
+        "table_footnote": [
+            "Table 2: Experiments on training phases and training data, where v1 represents continual SFT, v1 represents SFT from PLM, v1 represents SFT incorporating evaluator LM format data, and v1 represents data filtered using the Human-AI Consistency strategy."
+        ],
+        "table_body": "<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACC(t)</td><td>pearson</td><td>general</td></tr><tr><td>baseline</td><td>0.784</td><td>0.339</td><td>0.584</td><td>0.263</td><td>730</td></tr><tr><td>XDG-v1</td><td>0.910</td><td>0.514</td><td>0.755</td><td>0.686</td><td>458</td></tr><tr><td>XDG-v2</td><td>0.895</td><td>0.551</td><td>0.802</td><td>0.761</td><td>684</td></tr><tr><td>XDG-v3</td><td>0.911</td><td>0.593</td><td>0.811</td><td>0.765</td><td>653</td></tr><tr><td>XDG-v4</td><td>0.941</td><td>0.664</td><td>0.854</td><td>0.829</td><td>685</td></tr></table>",
+        "bbox": [
+            122,
+            212,
+            475,
+            315
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 3, the consistency between the evaluator LM and human scoring significantly increases after incorporating self-adaptive rubrics. However, the improvements are not as pronounced when other components are added, indicating that the primary driver of enhanced performance is the self-adaptive rubrics themselves. This suggests that self-adaptive rubrics play a crucial role in aligning the evaluator LM with human judgment.",
+        "bbox": [
+            112,
+            419,
+            489,
+            565
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/d29fc4d7cf30a6f0dc94ba37ed00e15d6139cad03fe74aedf6520d72064022c7.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 3: Ablation study for each component, where R.A. stands for reference answer."
+        ],
+        "table_body": "<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACCT</td><td>Pearson</td></tr><tr><td>Baseline</td><td>0.963</td><td>0.636</td><td>0.802</td><td>0.733</td></tr><tr><td>+ rubric</td><td>0.957</td><td>0.706</td><td>0.871</td><td>0.843</td></tr><tr><td>+ R.A</td><td>0.952</td><td>0.717</td><td>0.877</td><td>0.848</td></tr><tr><td>+ example</td><td>0.959</td><td>0.728</td><td>0.888</td><td>0.867</td></tr></table>",
+        "bbox": [
+            117,
+            573,
+            485,
+            659
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6.6 Comparison with Alternative Paradigm",
+        "text_level": 1,
+        "bbox": [
+            112,
+            722,
+            472,
+            739
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Using the same training data, we conduct a comparative analysis between the self-adaptive rubrics paradigm and the existing general rubric paradigm, as presented in Table 4. The results demonstrate that our approach significantly outperforms existing methods. Furthermore, in addition to accurately ranking the models, our method provides fine-grained capability evaluations that closely align with human assessments. This is both crucial and practical for facilitating the iterative development of LLMs. Due to space constraints, detailed de",
+        "bbox": [
+            112,
+            744,
+            490,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "scriptions and results of other experiments are provided in Appendix E.",
+        "bbox": [
+            507,
+            216,
+            884,
+            247
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            259,
+            640,
+            274
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce a novel evaluation paradigm called self-adaptive rubrics, aligning the scoring process with human judgment and reducing bias by tailoring rubrics to specific questions. Based on this paradigm, we develop a new benchmark, INSDA. To automate scoring, we analyze existing open-source evaluator language models and identify training phase data diversity issues. We then introduce human-AI consistency to align the chain-of-thought evaluation with human judgment and propose an evaluator LM that follows the self-adaptive rubrics paradigm. Experimental results show our model achieves higher consistency with human evaluation compared to GPT-4. We hope our work inspires researchers to apply this paradigm to more tasks, aligning automated scoring with human judgment.",
+        "bbox": [
+            507,
+            284,
+            885,
+            558
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            569,
+            615,
+            583
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose an evaluation paradigm based on self-adaptive rubrics, which provides more granular process guidance to align the scoring process with human judgment. Additionally, we introduce a benchmark, INSDA, based on this framework. However, there are several limitations:",
+        "bbox": [
+            507,
+            595,
+            884,
+            690
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- For questions with multiple correct answers, it requires manually writing multiple self-adaptive rubrics. It is worth noting that, to our knowledge, no current work focuses on the multi-solution direction.",
+            "- For subjective questions, such as creative writing, poetry, and other forms of artistic expression, different groups or individuals may have varying definitions of what constitutes good work. Therefore, it is necessary for each group or individual to set their own self-adaptive rubrics rather than relying on predefined ones. This also highlights the flexibility"
+        ],
+        "bbox": [
+            531,
+            702,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "16923",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and interpretability of the self-adaptive rubrics paradigm we propose.",
+        "bbox": [
+            149,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethical Considerations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            130,
+            315,
+            145
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We propose a scoring paradigm based on self-adaptive rubrics to enhance the interpretability and controllability of the automated scoring process. This approach aims to improve the credibility of evaluation results produced by LLMs and to support the research community in advancing these models. Nevertheless, the inherent hallucinations within LLMs pose a challenge to ensuring the complete accuracy of automated evaluation outcomes. Therefore, we recommend incorporating human review of certain outputs when using LLMs as judges to increase the overall reliability and credibility of the process.",
+        "bbox": [
+            112,
+            158,
+            490,
+            366
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Additionally, when generating self-adaptive rubrics for subjective questions, different groups or individuals may have varying definitions of what constitutes a good answer, potentially leading to biases and discrepancies. We encourage dialogue and mutual understanding among groups or individuals with diverse values, promoting the use of self-adaptive rubrics that align with their respective values and preferences.",
+        "bbox": [
+            112,
+            368,
+            489,
+            512
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            541,
+            213,
+            556
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "AI Anthropic. 2024. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card.",
+            "Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.",
+            "Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. 2023. Chateval: Towards better llmbased evaluators through multi-agent debate. Preprint, arXiv:2308.07201.",
+            "Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with  $90\\%$  chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6.",
+            "Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E Gonzalez, et al. 2024. Chatbot arena: An open platform for evaluating llms by human preference. arXiv preprint arXiv:2403.04132."
+        ],
+        "bbox": [
+            112,
+            565,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Daniel Deutsch, Rotem Dror, and Dan Roth. 2021. A statistical analysis of summarization evaluation metrics using resampling methods. Transactions of the Association for Computational Linguistics, 9:1132-1146.",
+            "Markus Freitag, David Grangier, and Isaac Caswell. 2020. BLEU might be guilty but references are not innocent. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 61-71, Online. Association for Computational Linguistics.",
+            "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300.",
+            "Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. 2024. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems, 36.",
+            "Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangwoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, and Minjoon Seo. 2024a. Prometheus: Inducing fine-grained evaluation capability in language models. Preprint, arXiv:2310.08491.",
+            "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. Preprint, arXiv:2405.01535.",
+            "Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. arXiv preprint arXiv:2107.10821.",
+            "Kalpesh Krishna, Aurko Roy, and Mohit Iyyer. 2021. Hurdles to progress in long-form question answering. arXiv preprint arXiv:2103.06332.",
+            "Junlong Li, Shichao Sun, Weizhe Yuan, Run-Ze Fan, Hai Zhao, and Pengfei Liu. 2023a. Generative judge for evaluating alignment. Preprint, arXiv:2310.05470.",
+            "Siyao Li, Deren Lei, Pengda Qin, and William Yang Wang. 2019. Deep reinforcement learning with distributional semantic rewards for abstractive summarization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6038-6044, Hong Kong, China. Association for Computational Linguistics.",
+            "Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023b. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval."
+        ],
+        "bbox": [
+            507,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "16924",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "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.",
+        "bbox": [
+            112,
+            85,
+            489,
+            139
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023a. Gpteval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634.",
+        "bbox": [
+            112,
+            148,
+            489,
+            200
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Yixin Liu, Alex Fabbri, Pengfei Liu, Yilun Zhao, Linyong Nan, Ruilin Han, Simeng Han, Shafiq Joty, Chien-Sheng Wu, Caiming Xiong, and Dragomir Radev. 2023b. Revisiting the gold standard: Grounding summarization evaluation with robust human evaluation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4140-4170, Toronto, Canada. Association for Computational Linguistics.",
+        "bbox": [
+            112,
+            210,
+            490,
+            331
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076-12100, Singapore. Association for Computational Linguistics.",
+        "bbox": [
+            112,
+            338,
+            489,
+            445
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Ehsan Montahaei, Danial Alihosseini, and Mahdieh Soleymani Baghshah. 2019. Jointly measuring diversity and quality in text generation models. Preprint, arXiv:1904.03971.",
+        "bbox": [
+            112,
+            451,
+            489,
+            506
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo",
+        "bbox": [
+            112,
+            516,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reiichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tsengpreston Tuggle Nick Turley Jerry Tworek Juan Felipe Ceron Uribe Andrea Vallone Arun Vijayvergiya Chelsea VossCarroll Wainwright Justin Jay Wang Alvin Wang Ben Wang Jonathan Ward Jason Wei CJ Weinmann Akila Welihinda Peter Welinder Jiayi Weng Lilian Weng Matt Wiethoff Dave Willner Clemens Winter Samuel Wolrich Hannah Wong Lauren Workman Sherwin Wu Jeff Wu Michael Wu Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. 2024. Gpt-4 technical report. Preprint arXiv:2303.08774.",
+        "bbox": [
+            507,
+            85,
+            884,
+            778
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "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 on Association for Computational Linguistics, ACL '02, page 311-318, USA. Association for Computational Linguistics.",
+        "bbox": [
+            507,
+            788,
+            884,
+            869
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Al",
+        "bbox": [
+            507,
+            879,
+            885,
+            921
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "16925",
+        "bbox": [
+            477,
+            927,
+            524,
+            941
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. Preprint, arXiv:1912.01703.",
+            "Ehud Reiter. 2018. A structured review of the validity of BLEU. Computational Linguistics, 44(3):393-401.",
+            "Julian Risch, Timo Möller, Julian Gutsch, and Malte Pietsch. 2021. Semantic answer similarity for evaluating question answering models. In Proceedings of the 3rd Workshop on Machine Reading for Question Answering, pages 149-157, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Natalie Schluter. 2017a. The limits of automatic summarisation according to ROUGE. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 41-45, Valencia, Spain. Association for Computational Linguistics.",
+            "Natalie Schluter. 2017b. The limits of automatic summarisation according to rouge. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, pages 41-45. Association for Computational Linguistics.",
+            "Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. 2020. Megatron-lm: Training multi-billion parameter language models using model parallelism. Preprint, arXiv:1909.08053.",
+            "Matthew Snover, Bonnie Dorr, Rich Schwartz, Linnea Micciulla, and John Makhoul. 2006. A study of translation edit rate with targeted human annotation. In Proceedings of the 7th Conference of the Association for Machine Translation in the Americas: Technical Papers, pages 223-231, Cambridge, Massachusetts, USA. Association for Machine Translation in the Americas.",
+            "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.",
+            "Rickard Stureborg, Dimitris Alikaniotis, and Yoshi Suhara. 2024. Characterizing the confidence of large language model-based automatic evaluation metrics. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers), pages 76-89.",
+            "Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix,"
+        ],
+        "bbox": [
+            114,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. Preprint, arXiv:2302.13971.",
+            "Yidong Wang, Zhuohao Yu, Zhengran Zeng, Linyi Yang, Cunxiang Wang, Hao Chen, Chaoya Jiang, Rui Xie, Jindong Wang, Xing Xie, Wei Ye, Shikun Zhang, and Yue Zhang. 2024. Pandalm: An automatic evaluation benchmark for llm instruction tuning optimization. Preprint, arXiv:2306.05087.",
+            "Seonghyeon Ye, Doyoung Kim, Sungdong Kim, Hyeonbin Hwang, Seungone Kim, Yongrae Jo, James Thorne, Juho Kim, and Minjoon Seo. 2023. Flask: Fine-grained language model evaluation based on alignment skill sets. arXiv preprint arXiv:2307.10928.",
+            "Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. 2022. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414.",
+            "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. Preprint, arXiv:1904.09675.",
+            "Ying Zhang and Stephan Vogel. 2004. Measuring confidence intervals for the machine translation evaluation metrics. In Proceedings of the 10th Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages.",
+            "Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36.",
+            "Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Preprint, arXiv:2306.05685.",
+            "Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. 2023. Agieval: A human-centric benchmark for evaluating foundation models. arXiv preprint arXiv:2304.06364.",
+            "Lianghui Zhu, Xinggang Wang, and Xinlong Wang. 2023. Judgelm: Fine-tuned large language models are scalable judges. Preprint, arXiv:2310.17631."
+        ],
+        "bbox": [
+            509,
+            85,
+            882,
+            820
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "16926",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Training Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            294,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We employed a default learning rate of 2e-5, and the batch size per device was dynamically adjusted based on the total data volume and the number of machines to maintain consistent optimization steps. The Adam optimizer was utilized.",
+        "bbox": [
+            112,
+            108,
+            487,
+            187
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "For the scaling law experiments, when  $n$  was below 300, we conducted three repetitions and averaged the results to minimize error.",
+        "bbox": [
+            112,
+            189,
+            485,
+            234
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "For all invoked APIs, we used the default parameters without extensive modifications.",
+        "bbox": [
+            112,
+            237,
+            487,
+            267
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Data Annotation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            279,
+            297,
+            294
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.1 Annotator Qualifications",
+        "text_level": 1,
+        "bbox": [
+            114,
+            304,
+            359,
+            319
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "All our annotators are internal team members with at least a Master's degree. We provide additional compensation significantly higher than the standard salary, based on the amount of data annotated.",
+        "bbox": [
+            112,
+            324,
+            485,
+            388
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Prompt Templates",
+        "text_level": 1,
+        "bbox": [
+            114,
+            399,
+            312,
+            416
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C.1 AutoPrompt",
+        "text_level": 1,
+        "bbox": [
+            114,
+            423,
+            263,
+            438
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Below, I will provide a <Question>, along with the corresponding <Reference Answer> and <Scoring Rubric>. You need to evaluate the <Output Result> from the <Model Answer> of the <Model to be Assessed>. The evaluation should be divided into two parts: \"Scoring Process\" and \"Final Score.\" Please note that the scoring range is from 0 to 5 points. You must justify the score you assign based on the <Model Answer>, strictly adhering to the requirements of the <Scoring Rubric> without adding, changing, or imagining any additional criteria.",
+        "bbox": [
+            137,
+            453,
+            463,
+            677
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C.2 Prompt for Set Matching",
+        "text_level": 1,
+        "bbox": [
+            114,
+            695,
+            361,
+            711
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are a meticulous judge tasked with evaluating whether the \"Test Rubric\" provided by the user aligns with the \"Standard Rubric.\" The evaluation rules are as follows:",
+        "bbox": [
+            139,
+            724,
+            460,
+            787
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The initial total score is set to zero.",
+            "- For each item in the \"Test Rubric\":"
+        ],
+        "bbox": [
+            161,
+            796,
+            438,
+            837
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "1. If the item matches any item in the \"Standard Rubric\" exactly, one point is added to the total score.",
+        "bbox": [
+            193,
+            844,
+            460,
+            907
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "2. If the item in the \"Test Rubric\" is unrelated to any item in the \"Standard Rubric,\" the total score remains unchanged.",
+            "3. If the item in the \"Test Rubric\" is the exact opposite of any item in the \"Standard Rubric,\" one point is subtracted from the total score."
+        ],
+        "bbox": [
+            586,
+            92,
+            857,
+            221
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You need to return the entire scoring process (explaining why points were added or subtracted) along with the final score. The return format should be:",
+        "bbox": [
+            534,
+            234,
+            858,
+            296
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "{ \"Scoring Process\": \"<Here, provide the scoring process as a string>\",",
+        "bbox": [
+            571,
+            309,
+            821,
+            357
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "\"Final Score\": \"<Here, provide the final score as a mathematical expression, concluding with 'Final Score: <score>' e.g., '3/5=0.6, Final Score: 0.6>'",
+        "bbox": [
+            571,
+            357,
+            821,
+            437
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The returned format must be compatible with json.load() to be converted into a dictionary.",
+        "bbox": [
+            534,
+            450,
+            855,
+            498
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C.3 Prosecutor Prompt",
+        "text_level": 1,
+        "bbox": [
+            509,
+            527,
+            709,
+            544
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Please check if the generated answer is correct. The reference answer is:  $\\{\\mathrm{gt}\\}$ , and the generated answer is:  $\\{\\mathrm{user}\\}$ . Please respond in the following format:  $\\{\\text{\"result\": True}\\}$",
+        "bbox": [
+            534,
+            564,
+            857,
+            629
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C.4 Refinement Prompt",
+        "text_level": 1,
+        "bbox": [
+            509,
+            659,
+            714,
+            674
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Your generated answer is not the standard answer. Please reflect on this and generate a new answer.",
+        "bbox": [
+            534,
+            695,
+            855,
+            740
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The generated scoring points and the full score answer are:",
+        "bbox": [
+            534,
+            743,
+            853,
+            772
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Benchmark",
+        "text_level": 1,
+        "bbox": [
+            509,
+            814,
+            648,
+            829
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D.1 Benchmark Leaderboard",
+        "text_level": 1,
+        "bbox": [
+            509,
+            846,
+            759,
+            860
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We provide the Benchmark LeaderBoard at the following link: https://github.com/Zhiyuan-Fan/self-adaptive-rubrics",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "16927",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/a7634cb60184aae7c523b465d602181dddec21c905a1593e72dabc74c9682ab9.jpg",
+        "image_caption": [
+            "Figure 4: Benchmark multi-level classification system. To maintain conciseness, we have detailed the hierarchy only up to the third tier."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            163,
+            80,
+            831,
+            554
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D.2 Benchmark Statistics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            617,
+            332,
+            632
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We provide the Benchmark statistics at the following link: https://github.com/Zhiyuan-Fan/self-adaptive-rubrics",
+        "bbox": [
+            112,
+            644,
+            489,
+            692
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "E Additional Experiments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            715,
+            359,
+            732
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "E.1 Automatic Rubric Generation",
+        "text_level": 1,
+        "bbox": [
+            114,
+            749,
+            400,
+            764
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "In this section, we conduct experiments on automatic rubric generation and analyze the results to explore its potential. Specifically, we compare the following three experimental setups: 1) using GPT-4 to automatically generate rubrics; 2) training an automatic rubric generation model with human-labeled rubrics under the SFT paradigm; and 3) aligning the model output with human preferences using DPO.",
+        "bbox": [
+            112,
+            776,
+            489,
+            921
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "E.2 Error Propagating",
+        "text_level": 1,
+        "bbox": [
+            509,
+            617,
+            705,
+            633
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "When using rubric generation models to automatically create self-adaptive rubrics, a potential issue is that if the generated rubric is inconsistent with the human-provided rubric, errors can accumulate in the scoring pipeline, leading to a larger deviation in the final score. By incorporating a filtering strategy, the overall performance will improve.",
+        "bbox": [
+            507,
+            638,
+            884,
+            750
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "E.3 Joint Training vs. Expert Training",
+        "text_level": 1,
+        "bbox": [
+            507,
+            760,
+            830,
+            778
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We also explored whether to combine data from different categories for joint training when training the evaluator LM or rubric generation model, or to train a separate expert model for each category individually. We found that using joint training can achieve better results than expert training.",
+        "bbox": [
+            507,
+            782,
+            882,
+            879
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "16928",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/67528f74254b32d18ef6494d9cd26383e440a7319d1badaaaadad247909d067c.jpg",
+        "image_caption": [
+            "Figure 5: An illustrative example of the format used in self-adaptive rubrics."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            163,
+            86,
+            836,
+            354
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/a24f9880c21151af79994eef03b22d371fcf509650d3bf630a87fedd963e95ff.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Type</td><td>ACC</td><td>ACCT</td><td>Pearsonr</td><td>GSB</td></tr><tr><td>GPT-4</td><td>w self-adaptive rubrics</td><td>0.3241</td><td>0.7025</td><td>0.7283</td><td>0.9211</td></tr><tr><td>GPT-4</td><td>wo</td><td>0.2500</td><td>0.5035</td><td>0.4863</td><td>0.8684</td></tr><tr><td>GPT-4-turbo</td><td>w</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GPT-4-turbo</td><td>wo</td><td>0.1995</td><td>0.5473</td><td>0.5326</td><td>0.9000</td></tr><tr><td>GPT-3.5</td><td>w</td><td>0.2121</td><td>0.5569</td><td>0.3888</td><td>0.8947</td></tr><tr><td>GPT-3.5</td><td>wo</td><td>0.1677</td><td>0.3872</td><td>0.1286</td><td>0.5158</td></tr></table>",
+        "bbox": [
+            176,
+            395,
+            820,
+            521
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/82200f8acca72b714e743533c896d1fddcf41d01bcda4def9b276e10a11aa531.jpg",
+        "table_caption": [
+            "Table 4: Ablation Study"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>OOD</td><td>ID</td><td>Random</td></tr><tr><td>GPT4-turbo</td><td>0.399</td><td>0.488</td><td>0.417</td></tr><tr><td>GPT4</td><td>0.602</td><td>0.613</td><td>0.613</td></tr><tr><td>xdg-turbo</td><td>0.606</td><td>0.607</td><td>0.603</td></tr></table>",
+        "bbox": [
+            139,
+            579,
+            463,
+            649
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/e90361d0d2995eb248f48326dae2aa00ed33720deb1f0381386a05107bff4dbe.jpg",
+        "table_caption": [
+            "Table 5: Rubric Generation Results."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACCt</td><td>pearsonr</td></tr><tr><td>GPT-4</td><td>0.921</td><td>0.324</td><td>0.702</td><td>0.728</td></tr><tr><td>GPT-3.5</td><td>0.894</td><td>0.212</td><td>0.556</td><td>0.388</td></tr></table>",
+        "bbox": [
+            119,
+            703,
+            480,
+            758
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/3203498003cf76d87fd2ff4ae4b65853c707487a967f35f048b670f3889d4482.jpg",
+        "table_caption": [
+            "Table 6: General Rubrics with Ground Truth."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACCt</td><td>pearsonr</td></tr><tr><td>GPT-4-turbo</td><td>0.9</td><td>0.199</td><td>0.547</td><td>0.532</td></tr><tr><td>GPT-4</td><td>0.868</td><td>0.250</td><td>0.503</td><td>0.486</td></tr><tr><td>GPT-3.5</td><td>0.515</td><td>0.167</td><td>0.387</td><td>0.128</td></tr></table>",
+        "bbox": [
+            115,
+            812,
+            500,
+            881
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 7: General Rubrics without Groud Truth.",
+        "bbox": [
+            139,
+            891,
+            460,
+            903
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "16929",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/7ccd63d48d08b349fb2d0ead2a92dcb5085a633cb55b23c8050a36f1b8bc557f.jpg",
+        "image_caption": [
+            "Figure 6: Distribution change of the evaluator LM format data after applying the Human-AI Consistency strategy."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            156,
+            99,
+            478,
+            309
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "image",
+        "img_path": "images/ec57d49ca2e75e8c3ceb1d0a8fdc4218aa8c171a31ef1c137715a8c67ec12bd5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            551,
+            99,
+            860,
+            305
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "Algorithm 1 Self-Adaptive Rubrics Generation and Validation  \nRequire:  $Q$  ▷ Given question  \nRequire:  $GT$  ▷ Ground truth answer  \nRequire:  $n$  ▷ Maximum iterations  \n1:  $i \\gets 0$   \n2: accepted ← False  \n3: while  $i < n$  and ¬accepted do  \n4:  $R, IA \\gets \\text{GPT-4}(Q)$  ▷ Generate rubrics and ideal answer  \n5: if  $PA(IA, GT)$  then ▷ Prosecutor agent checks ideal answer  \n6: accepted ← True  \n7: else  \n8: Inform GPT-4 of incorrect IA  \n9:  $i \\gets i + 1$   \n10: end if  \n11: end while  \n12: if accepted then  \n13: return R  \n14: else  \n15: return Failure in  $n$  iterations  \n16: end if",
+        "guess_lang": "txt",
+        "bbox": [
+            114,
+            468,
+            487,
+            843
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "16930",
+        "bbox": [
+            477,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2024/SedarEval_ Automated Evaluation using Self-Adaptive Rubrics/b33177e1-fdfe-4c5f-a9a9-037a138aa0cf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:467e4da6d62d3c6e172f38b70f840cc721516aac98e78857e568ffe3938b2ad1
+size 1418433

2024/SedarEval_ Automated Evaluation using Self-Adaptive Rubrics/full.md

@@ -0,0 +1,443 @@
+# SedarEval: Automated Evaluation using Self-Adaptive Rubrics
+
+Zhiyuan Fan $^{1}$ , Weinong Wang $^{2*}$ , Xing Wu $^{3}$ , Debing Zhang $^{2}$
+
+$^{1}$ Hong Kong University of Science and Technology,  $^{2}$ Xiaohongshu Inc,
+
+$^{3}$ Institute of Information Engineering, Chinese Academy of Sciences zhiyuan.fan@connect.ust.hk, weinong.wang@hotmail.com
+
+# Abstract
+
+The evaluation paradigm of LLM-as-judge gains popularity due to its significant reduction in human labor and time costs. This approach utilizes one or more large language models (LLMs) to assess the quality of outputs from other LLMs. However, existing methods rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process, compromising precision and stability in assessments. Inspired by human examination scoring processes, we propose a new evaluation paradigm based on self-adaptive rubrics. Specifically, we create detailed scoring rubrics for each question, capturing the primary and secondary criteria in a structured format of scoring and deduction points that mimic a human evaluator's analytical process. Building on this paradigm, we further develop a novel benchmark called SedarEval, which covers a range of domains including long-tail knowledge, mathematics, coding, and logical reasoning. SedarEval consists of 1,000 meticulously crafted questions, each with its own self-adaptive rubric. To further streamline the evaluation, we train a specialized evaluator language model (evaluator LM) to supplant human graders. Using the same training data, our evaluator LM achieves a higher concordance rate with human grading results than other paradigms, including GPT-4, highlighting the superiority and efficiency of our approach.
+
+# 1 Introduction
+
+The rapid advancements in large language models (LLMs) have led to their widespread use (OpenAI et al., 2024; Team et al., 2023; Anthropic, 2024; Bai et al., 2023). However, assessing these models in open-ended question-answering scenarios poses a significant challenge. Automated metric-based evaluations offer speed and convenience but often fall short due to the diversity of ground truth
+
+(Schluter, 2017a; Reiter, 2018; Montahaei et al., 2019; Freitag et al., 2020). In contrast, human-based evaluations provide reliable assessments but require substantial resources.
+
+To bridge the gap, the LLM-as-a-judge paradigm attempts to strike a balance between automated and human evaluation. Prominent examples of this approach include MT-bench (Zheng et al., 2024) and Arena (Chiang et al., 2024), which leverage proprietary models to evaluate individual or comparative model responses. These benchmarks use pre-defined principles, such as the 3H principle (human-like, helpful, harmonious), to determine responses that align best with realistic human preferences. The widespread use of GPT-4 (OpenAI et al., 2024) as an evaluator in these studies presents challenges, including high costs for research institutions and potential data leaks.
+
+Some studies (Zhu et al., 2023; Li et al., 2023a; Wang et al., 2024; Kim et al., 2024a,b) propose using open-source pretrained models (Touvron et al., 2023; Bai et al., 2023; Zeng et al., 2022) to train specialized evaluator LMs, offering a more cost-effective and secure solution. However, these methods typically use a uniform, question-agnostic rubric to guide the scoring process, overlooking the unique characteristics of each question. Each question has different emphases, with primary and secondary scoring points. A general rubric applies uniform criteria, failing to accurately reflect human preferences.
+
+To adaptively align the scoring process with human judgment, we propose a novel evaluation paradigm based on self-adaptive rubrics. Unlike coarse-grained general rubrics, we provide fine-grained rubrics for each task, detailing specific scoring and penalty points with primary and secondary information. By analyzing focus points, we assign different values to each point. Additionally, we introduce penalty points to penalize models for generating rejected responses. The scoring pro
+
+![](images/bb722d898fb4f2cae162dd772a0b5a02bb62c55d30cfeb44bb1dfbb37d5238d9.jpg)  
+Figure 1: Automated evaluation pipeline using self-adaptive rubrics. This pipeline dynamically adjusts the evaluation rubric based on the input question, resulting in a scoring process that aligns more closely with human evaluators.
+
+cess considers both preferred and rejected perspectives. The inconsistent coverage of positive and penalty points ensures a more refined constraint on the scoring process. These detailed scoring trajectories simplify the evaluation process to an instruction-following task, reducing dependency on a judge model's internal knowledge and skills, leading to more accurate and stable assessments. Building on this paradigm, we construct a new benchmark called SedarEval, presenting a taxonomy comprising 5 major categories, with a dataset of 1000 queries, that fully aligns with realistic scenarios.
+
+We further conduct ablation experiments on each component of the LLM-as-a-judge paradigm to investigate training a specialized LLM for scoring, revealing their respective importance. We analyze whether LLMs can correctly evaluate questions they can correctly answer and find that insufficient diversity in existing SFT data and a lack of evaluation-format data limit model performance. We also propose human-AI consistency to ensure evaluator LLMs maintain alignment with human preferences while leveraging their chain of thought capability to improve evaluation performance. Based on these findings, we develop a specialized evaluator LLM tailored to the benchmark for automated scoring. This model surpasses GPT-4 in model-level and question-level Pearson correlation, GSB, and ACC metrics, demonstrating higher consistency with human judgment. Ex
+
+perimental results validate the effectiveness and efficiency of our proposed paradigm.
+
+Our contributions are summarized as follows:
+
+1. We propose a novel evaluation paradigm using self-adaptive rubrics for each question, offering granular guidance and closely aligning the scoring process with human evaluation.  
+2. We develop a high-quality benchmark called SedarEval, featuring 1,000 meticulously crafted questions with detailed rubrics, and conduct manual evaluations on 20 LLMs.  
+3. We analyze the training of evaluator LMs, highlight existing methods' shortcomings, and use the self-adaptive rubrics paradigm to train an evaluator LM that surpasses GPT-4 in agreement with human evaluations.
+
+# 2 Related Work
+
+Benchmark LLMs Capabilities. With the rapid advancement of LLMs (OpenAI et al., 2024; Team et al., 2023; Anthropic, 2024), it has become a substantial challenge to benchmark their broad capabilities reliably. NLU-style tasks (Hendrycks et al., 2020; Huang et al., 2024; Srivastava et al., 2022; Zhong et al., 2023), such as multi-choice QA, employ general-exam questions from various domains to assess a model's knowledge and comprehension abilities. However, their real-world usage is limited
+
+due to misalignment with human preferences. Recently, reference-free benchmarks (Li et al., 2023b; Chiang et al., 2023; Zheng et al., 2024; Ye et al., 2023) have been proposed to evaluate texts' quality in a generative setting directly. Unlike previous datasets, our benchmark provides a comprehensive and stable model assessment with its diverse test cases and broad label distribution.
+
+Automatic NLG Evaluation. It's notably challenging to evaluate the quality of generated text in the field of natural language generation (NLG). Traditional n-gram-based metrics (Papineni et al., 2002; Lin, 2004; Snover et al., 2006) and embedding-based metrics (Li et al., 2019; Zhang et al., 2020; Risch et al., 2021) can only assess lexical or semantic similarity between the generated answers and reference answers (Schluter, 2017a; Reiter, 2018; Montahaei et al., 2019; Freitag et al., 2020). These metrics have been found to have a relatively low correlation with human preferences (Liu et al., 2023a). Recently, employing LLM as a judge (Zheng et al., 2023; Li et al., 2023b; Chan et al., 2023) is a novel evaluation paradigm that has gained widespread application. The most common approach involves using proprietary LLMs, such as GPT-4 (OpenAI et al., 2024), as judge models to rank or score outputs generated by other models. However, this method relies on closed-source models, incurs high costs, and poses risks of internal evaluation dataset leaks for companies developing LLMs. To address these issues, various works (Zhu et al., 2023; Li et al., 2023a; Wang et al., 2024; Kim et al., 2024a,b) have proposed training dedicated scoring models on open-source base models using synthetic or manually labeled data. These evaluations often use reference answers to assist in the assessment or employ general rubrics to guide the scoring process. However, these approaches overlook the differences between individual questions and the varying scoring criteria of each question, even within the same category. In contrast, we propose an evaluation paradigm based on self-adaptive rubrics that generates fine-grained, customizable rubrics for each question, guiding a more precise scoring process. It is worth noting that although Prometheus 2 also claims to use fine-grained rubrics, their rubrics remain question-agnostic.
+
+Quantifying Evaluation Confidence. The automatic metrics are imperfect, and we must measure their performance further. A gold standard for this is their alignment with human judgment
+
+and the confidence level we can have when these metrics guide our decision-making process. However, quantifying this performance (Krishna et al., 2021; Schluter, 2017b; Stureborg et al., 2024) is difficult due to various factors (the evaluator's accuracy and stability, evaluation set size, the extent of the performance difference among competing models, etc.). (Kocmi et al., 2021; Deutsch et al., 2021; Zhang and Vogel, 2004) investigate the correlation between human judgment and traditional automatic metrics such as ROUGE and BLEU and analyze their confidence intervals. For LLM-based evaluators, commonly used metrics include Pearson, Spearman, and Kdendall-Tau to measure the alignment between the model's scores and human preferences. However, previous work has primarily focused on the correlation of rankings or overall scores at the model level without comparing the scores with human ratings at the individual question. This limits the interpretability of the scoring process and hampers its utility in guiding the development and iteration of LLMs.
+
+# 3 SedarEval Benchmark
+
+In this section, we introduce SedarEval, a benchmark constructed upon the self-adaptive rubrics paradigm. We begin by delving into the intricacies of the self-adaptive rubric paradigm, followed by a detailed explanation of the benchmark's core components – questions and their corresponding rubrics – along with the methodology for model evaluation using this benchmark. To ensure the quality of SedarEval, we incorporate comprehensive human assessment into the construction process, meticulously filtering out samples that fail to meet the established quality standards.
+
+# 3.1 Self-Adaptive Rubrics
+
+Previous LLM-as-a-judge approaches, which rely on general rubrics or principles for scoring, often lack specific, problem-related rubric guidance. Consequently, these methods depend heavily on the inherent capabilities of the LLM itself, leading to potential errors in evaluations due to insufficient reasoning abilities or hallucinations. Additionally, this approach introduces extraneous biases, such as position bias and order bias.
+
+Self-adaptive rubrics address these issues by tailoring the evaluation criteria to the specific problems at hand, incorporating the focal points of the problem and assigning different weights accordingly.
+
+By introducing penalty points, these rubrics align more closely with human judgments by deducting points for outputs that deviate from expected tendencies. To prevent human evaluators (or LLMs) from making incorrect assessments due to a lack of background information, additional context is provided for each question to assist in the scoring process. A typical self-adaptive rubric comprises three components: scoring points, penalty points, and background knowledge, as illustrated in Table 3.
+
+# 3.2 Dataset Construction
+
+Questions: We have defined a classification system for objective questions, with a two-tiered scoring system as shown in the diagram. Under each secondary classification, we have hired five people to create questions. Specifically, each person is required to first create their own questions to get a question pool, and then each person votes on all the questions. We only keep the questions that all five people agree on.
+
+For each candidate question, the annotators will select 5 LLMs to test the effectiveness of the questioned question. We only keep the questions with a larger variance in scores, which are more discriminating, and remove the questions where the answers from different models are almost the same, which are not helpful in distinguishing between different models. For example, if a question can be answered correctly by all models, or incorrectly by all models, then this question cannot show which model is better.
+
+After collecting the initial questions, we hired another group of people to compare all the questions in pairs to judge the similarity of the problem-solving ideas for the two questions and delete the questions with too much similarity.
+
+Rubrics: For each question, we assign it to three individuals to discuss together and generate a rubric like the one shown in Figure 1.
+
+For more detailed information, please refer to Appendix D, which contains benchmark statistics and the leaderboard.
+
+# 3.3 Evaluation Pipeline
+
+The entire evaluation pipeline using our benchmark is illustrated in Figure 1. Given a question, its corresponding rubrics, and the model to be evaluated, we first input the question into the model to generate a response. The response is then scored according to the predefined rubric, either by human
+
+evaluators or using LLMs. Finally, all the scores are aggregated to obtain the model's total score.
+
+# 4 Evaluator Language Model
+
+In this section, we introduce an evaluator LM aligned with the self-adaptive rubrics paradigm to substitute human evaluators. We begin by delineating the evaluation format. Subsequently, we propose a novel data filtering strategy to align the Chain-of-Thought evaluation process with human judgments. Finally, we discuss the automation of rubric generation.
+
+# 4.1 Evaluation Format
+
+The evaluation format consists of two types: direct scoring of individual model outputs and pairwise comparison of model outputs to determine the superior one. Pairwise evaluation requires significantly more comparisons as the number of candidate models increases, as shown by Equation 1. Therefore, we employ direct assessment in this paper. Notably, direct assessment scores can be compared to derive pairwise results.
+
+$$
+C (n, 2) = \frac {n !}{2 ! (n - 2) !} - n = \frac {n ^ {2} - 3 n}{2} \tag {1}
+$$
+
+We use a reference-based format to organize the output. Specifically, for each question, we compile the reference answer, self-adaptive rubrics, and scoring examples to create an auto-prompt template. When evaluating answers, we incorporate the answers into this auto-prompt template as the complete input. We conduct ablation experiments on each component in zero-shot, few-shot, and instruction tuning settings.
+
+# 4.2 Human-AI Consistency
+
+Human annotators provide specific scores for each response without corresponding explanations, which is efficient but suboptimal for training evaluator LMs. To alleviate this issue, we use GPT-4 to generate detailed reasoning steps using Chain-of-Thought. However, scoring preferences may differ between GPT-4 and human annotators, and both may make errors. To mitigate these errors and align the scoring process with human judgment, we introduce a Human-AI Consistency strategy to improve synthetic data quality. We extract final scores from the GPT-4 scoring process and compare them with human scores, retaining only the data where GPT-4 and human results are consistent,
+
+as shown in Equation 2, where  $\mathcal{H}$  represents human scores,  $\mathcal{A}$  represents AI scores, and  $\mathbb{I}$  is an indicator function.
+
+$$
+\mathcal {T} = \{(h, a) \mid h \in \mathcal {H}, a \in \mathcal {A}, \mathbb {I} (h, a) = 1 \} \quad (2)
+$$
+
+This approach only retains instances where human and AI scores are consistent and differs from rejection sampling, which uses human scores as a reward function to select the optimal output from multiple GPT-4 results.
+
+# 4.3 Automatic Rubric Generation
+
+To reduce human annotation costs, we investigate using human-annotated datasets to train a model for automatic self-adaptive rubric generation. By providing the model with questions and corresponding reference answers, we train it to produce rubrics that delineate scoring criteria and identify deduction points.
+
+Generating self-adaptive rubric format output is straightforward, but aligning rubrics with human preferences requires aligning the model with human evaluative criteria. This complexity arises because identifying scoring points, assigning specific weights, and criteria for deductions are significantly influenced by human judgment.
+
+The training process for the automatic rubric generation model comprises two stages. Initially, we use human-labeled data to train a base model through Supervised Fine-Tuning (SFT), as depicted in Equation 3.
+
+$$
+\mathcal {L} (\theta) = - \sum_ {i = 1} ^ {N} \log p _ {\theta} \left(y _ {i} \mid x _ {i}\right) \tag {3}
+$$
+
+The base model generates rubrics that conform to the specified format, though they may not fully align with human preferences (quantitative metrics will be introduced in Section 5.2). In the next phase, rubrics generated by the base model are treated as rejected responses, while human-labeled rubrics serve as preferred responses to construct preference pairs. We then train the model using Direct Preference Optimization (DPO) to align it with human preferences, as shown in Equations 4 and 5.
+
+$$
+f (y, x) = \beta \log \frac {\pi_ {\theta} (y \mid x)}{\pi_ {\mathrm {r e f}} (y \mid x)} \tag {4}
+$$
+
+$$
+\mathcal {L} \left(\pi_ {\theta}\right) = - \log \sigma \left(f \left(y _ {w}, x\right) - f \left(y _ {l}, x\right)\right) \tag {5}
+$$
+
+We also explore automating rubric generation using GPT-4 without reference answers. To ensure accuracy, GPT-4 creates both the rubric and an ideal answer for each question. If the ideal answer corresponds with the ground truth, the generated rubric is deemed acceptable. We employ a self-refinement strategy to help the model iteratively refine its outputs, aligning it with human preferences. For detailed algorithmic procedures and prompts, refer to Appendix E.1.
+
+# 5 Experiment
+
+# 5.1 Experimental Setting
+
+We train the Evaluator Language Model and the Rubric Generation Model using both the opensource model LLaMA-3 (Touvron et al., 2023) and our internal model  $\mathrm{XDG}^{1}$ . To maximize training efficiency and utilize hardware resources, we implement tensor parallelism (Shoeybi et al., 2020) with PyTorch 2.3 (Paszke et al., 2019). For the 7B/8B models, we use 128 H100 GPUs, while for the 70B models, we use 512 H100 GPUs. For the models' chat versions (i.e., instruction-tuned), we employ the same chat markup language (ChatML) as the models themselves. For the pre-trained versions, we use a unified ChatML to reduce data bias. We adopt adaptive learning rate and batch size strategies. Further training details are provided in Appendix A.
+
+# 5.2 Evaluation Metrics
+
+To assess the performance of the evaluator language model, we use Pearson's correlation coefficient and Spearman's rank correlation coefficient. These statistical measures assess the consistency between the outcomes of the evaluator language model and those obtained from human evaluators.
+
+Each question is accompanied by a detailed rubric specifying exact scoring and deduction criteria, so we use accuracy to evaluate the model's capability in following these self-adaptive rubrics for scoring. Considering potential noise in the model scoring, we introduce a weaker threshold ACC, which considers a result correct if it falls within a specified range. The calculation formulas are presented in Equation 6.
+
+$$
+\mathrm {A C C} _ {t} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left\{ \begin{array}{l l} 1, & \text {i f} | y _ {\text {p r e d} _ {i}} - y _ {\text {t r u e} _ {i}} | \leq \epsilon \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {6}
+$$
+
+To facilitate the iterative enhancement of LLMs using our benchmark, a robust metric is essential to assess whether the current model version outperforms its predecessor. Therefore, we adopt the widely used GSB (Good, Same, Bad) metric to compare model performance. Given two models, A and B, the calculation formula is presented in Equation 7. In this context, "#good" signifies that model A surpasses model B, "#bad" denotes the contrary, and "#same" indicates equivalent performance between the models.
+
+$$
+\Delta G S B = \frac {\# \text {g o o d} - \# \text {b a d}}{\# \text {g o o d} + \# \text {s a m e} + \# \text {b a d}} \tag {7}
+$$
+
+To evaluate the quality of automatically generated rubrics, we draw on the ACU (Liu et al., 2023b) and FactScore (Min et al., 2023) paradigms, using GPT-4 to calculate the match between the generated rubrics and the ground truth rubrics. The calculation formula is specified in Equation 8, where GT represents the correct rubric set containing multiple {grading points: specific score} pairs, and AT denotes the automatically generated rubric set.  $\mathbb{I}(i\in GT)$  is an indicator function that equals 1 if the item  $i$  from AT is present in GT, and 0 otherwise. The prompts used for this evaluation are detailed in Appendix C.2.
+
+$$
+\operatorname {M a t c h} (G T, A T) = \frac {\sum_ {i \in A T} \mathbb {I} (i \in G T)}{| G T |} \tag {8}
+$$
+
+# 5.3 Selected Models
+
+Previous studies predominantly employ English-proficient models to generate <question, response, score> triples for training evaluator language models, often overlooking models proficient in Chinese. Additionally, several studies exclusively use GPT-3.5 or GPT-4 to construct such synthetic data. These data generation methodologies may cause discrepancies between the synthetic and real-world data distributions, introducing biases into the trained evaluator language models.
+
+To alleviate this issue, we utilize a broader range of LLMs to collect responses that better reflect real-world distributions. This approach ensures greater diversity and mitigates biases introduced by relying solely on synthetic data generated from a single model. Specifically, we choose GPT-4, GPT-4-turbo, GPT-4-o, Claude Opus, DeepSeek 2.0, MiniMax 6.5, MiniMax 6, Doubao, GPT-3.5, Tongyi Qianwen 2.0, and Tongyi Qianwen 1.5-100B/70B. This selection includes models proficient in differ
+
+ent languages and multiple versions of the same model.
+
+For open-source models, we use local deployment to infer responses. For proprietary LLMs, if API services are available, we collect model outputs by requesting the API. If only a web interface is provided, we employ people to gather the outputs.
+
+# 6 Analysis
+
+In this section, we conduct a comprehensive experimental analysis of the robustness of the proposed benchmark evaluations, examining the data distribution, training phases, and training paradigms of evaluator LMs. Our findings reveal limitations in current training methodologies for evaluator LMs. Building on these insights, we develop an evaluator LM aligned with the self-adaptive rubrics paradigm.
+
+# 6.1 Scaling Law for Robust Evaluation
+
+A robust benchmark should effectively distinguish the capabilities of different models and maintain stability to ensure consistent rankings rather than allowing fluctuations due to the instability of individual tasks. To achieve this, the benchmark needs a sufficiently broad distribution while minimizing extraneous biases.
+
+To verify the robustness of the proposed benchmark, we conduct two rounds of sampling without replacement from a pool of 1,000 questions. In each round, we select  $n$  questions, resulting in a total of  $2n$  independent questions, where  $n \in [10,500]$ . We then compare the consistency of the model rankings obtained from these two samples.
+
+![](images/b0762fe1c2d162e955cdaf38f3906ab8a0255ca6357adcbf8544267b772456fa.jpg)  
+Figure 2: Consistency of model rankings as  $n$  increases. After  $n$  reaches approximately 300, the consistency stabilizes with only minor fluctuations.  
+Figure 2 shows the variation in the consistency of model rankings under different question sets as  $n$
+
+increases. When  $n$  is relatively small, the consistency is low, indicating the inconsistency caused by biases in different distributions. As  $n$  increases, the consistency improves despite the two sets remaining independent. After  $n$  reaches approximately 300, the consistency stabilizes with only minor fluctuations. This demonstrates the scaling law for robust evaluation, indicating that as the number of questions increases, the evaluation results tend to stabilize due to the broader coverage of the distribution.
+
+# 6.2 Score Distribution Shift
+
+![](images/3fb42723dbc22e32d626e5726793e5ebd8a6dfd4611b27479317e6f01955eda6.jpg)  
+Figure 3: Data distribution comparison of different data.
+
+The Prometheus approach, which relies on general rubrics not specifically tailored to the problem at hand, employs GPT-4 to generate an equal amount of synthetic data across different scores (1-5) to mitigate score bias from the evaluating language model. In contrast, our method uses self-adaptive rubrics, and our responses are genuinely collected from the model rather than artificially synthesized. Consequently, we cannot ensure that the quantity of data for each score is perfectly balanced.
+
+However, as illustrated in Figure 3, we observe that despite the score distribution shift between the training and test data, the score distribution of the model outputs, trained using the self-adaptive rubrics paradigm, closely aligns with the human-provided ground truth. This finding substantiates the robustness and efficacy of the self-adaptive rubrics paradigm in automated scoring.
+
+# 6.3 Out of Distribution Evaluation
+
+We establish two dimensions for evaluating the out-of-distribution capabilities of our model: model-level and question-level. For the model-level evaluation, we utilize the same set of questions, selecting a subset of models to train the evaluator language model (LM), and subsequently test on the remaining unselected models. In the question-level evaluation, a subset of questions along with all associated models are used for training, and the scoring performance is then assessed on a different
+
+set of questions.
+
+Table 1 presents the experimental results, showing that under the self-adaptive rubrics paradigm, the model performs well in both model-level and question-level evaluations. This indicates that our proposed method has strong generalization capabilities.
+
+# 6.4 Merged SFT or Continual SFT
+
+Previous research shows that a model might generate a correct answer but fail to accurately evaluate the <question, answer> pair for the same question. We argue that this issue mainly arises from the insufficient diversity of the SFT data.
+
+To validate this, we conduct the following experiments:
+
+1. Training a pretrained language model (PLM) using only traditional SFT data.  
+2. Training a PLM using a mix of SFT data and evaluator LM format data.  
+3. Performing continual SFT on an instruction-tuned model using evaluator LM format data, a widely adopted approach in other studies (Kim et al., 2024b,a).
+
+As shown in Table 2, we find that although the model using continual SFT performs well on evaluation tasks, its general ability significantly declines, limiting its versatility. However, starting from a PLM and using a mix of SFT data and evaluator LM format data for SFT results in excellent evaluation capability with minimal impact on general ability. This reveals the shortcomings of the previous continual SFT approach and suggests that the model's inability to evaluate the questions it can answer may simply be due to the lack of such data, highlighting the importance of diversity in SFT data.
+
+We employ Human-AI Consistency to filter the evaluator LM and find that, compared to using raw Chain-of-Thought data generated by GPT-4 and data filtered by rejection sampling, the data selected using Human-AI Consistency shows significant improvements in both evaluation and general capability, demonstrating the effectiveness of this strategy.
+
+# 6.5 Ablation Study
+
+We conduct detailed ablation experiments on the components of self-adaptive rubrics, namely, reference answers, rubrics, and in-context examples.
+
+<table><tr><td rowspan="2">Type</td><td colspan="4">Question-level</td><td colspan="4">Model-level</td></tr><tr><td>GSB</td><td>ACC</td><td>ACC(t)</td><td>pearson</td><td>GSB</td><td>ACC</td><td>ACC(t)</td><td>pearson</td></tr><tr><td>XDG</td><td>0.952</td><td>0.590</td><td>0.794</td><td>0.738</td><td>0.952</td><td>0.590</td><td>0.794</td><td>0.380</td></tr><tr><td>GPT-3.5</td><td>0.829</td><td>0.422</td><td>0.663</td><td>0.566</td><td>0.829</td><td>0.422</td><td>0.663</td><td>0.566</td></tr><tr><td>GPT-4</td><td>0.952</td><td>0.654</td><td>0.855</td><td>0.822</td><td>0.952</td><td>0.654</td><td>0.855</td><td>0.822</td></tr></table>
+
+Table 1: Out of distribution evaluation performance in both model-level and question-level.  
+
+<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACC(t)</td><td>pearson</td><td>general</td></tr><tr><td>baseline</td><td>0.784</td><td>0.339</td><td>0.584</td><td>0.263</td><td>730</td></tr><tr><td>XDG-v1</td><td>0.910</td><td>0.514</td><td>0.755</td><td>0.686</td><td>458</td></tr><tr><td>XDG-v2</td><td>0.895</td><td>0.551</td><td>0.802</td><td>0.761</td><td>684</td></tr><tr><td>XDG-v3</td><td>0.911</td><td>0.593</td><td>0.811</td><td>0.765</td><td>653</td></tr><tr><td>XDG-v4</td><td>0.941</td><td>0.664</td><td>0.854</td><td>0.829</td><td>685</td></tr></table>
+
+Table 2: Experiments on training phases and training data, where v1 represents continual SFT, v1 represents SFT from PLM, v1 represents SFT incorporating evaluator LM format data, and v1 represents data filtered using the Human-AI Consistency strategy.
+
+As shown in Table 3, the consistency between the evaluator LM and human scoring significantly increases after incorporating self-adaptive rubrics. However, the improvements are not as pronounced when other components are added, indicating that the primary driver of enhanced performance is the self-adaptive rubrics themselves. This suggests that self-adaptive rubrics play a crucial role in aligning the evaluator LM with human judgment.
+
+<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACCT</td><td>Pearson</td></tr><tr><td>Baseline</td><td>0.963</td><td>0.636</td><td>0.802</td><td>0.733</td></tr><tr><td>+ rubric</td><td>0.957</td><td>0.706</td><td>0.871</td><td>0.843</td></tr><tr><td>+ R.A</td><td>0.952</td><td>0.717</td><td>0.877</td><td>0.848</td></tr><tr><td>+ example</td><td>0.959</td><td>0.728</td><td>0.888</td><td>0.867</td></tr></table>
+
+Table 3: Ablation study for each component, where R.A. stands for reference answer.
+
+# 6.6 Comparison with Alternative Paradigm
+
+Using the same training data, we conduct a comparative analysis between the self-adaptive rubrics paradigm and the existing general rubric paradigm, as presented in Table 4. The results demonstrate that our approach significantly outperforms existing methods. Furthermore, in addition to accurately ranking the models, our method provides fine-grained capability evaluations that closely align with human assessments. This is both crucial and practical for facilitating the iterative development of LLMs. Due to space constraints, detailed de
+
+scriptions and results of other experiments are provided in Appendix E.
+
+# 7 Conclusion
+
+In this paper, we introduce a novel evaluation paradigm called self-adaptive rubrics, aligning the scoring process with human judgment and reducing bias by tailoring rubrics to specific questions. Based on this paradigm, we develop a new benchmark, INSDA. To automate scoring, we analyze existing open-source evaluator language models and identify training phase data diversity issues. We then introduce human-AI consistency to align the chain-of-thought evaluation with human judgment and propose an evaluator LM that follows the self-adaptive rubrics paradigm. Experimental results show our model achieves higher consistency with human evaluation compared to GPT-4. We hope our work inspires researchers to apply this paradigm to more tasks, aligning automated scoring with human judgment.
+
+# Limitations
+
+In this paper, we propose an evaluation paradigm based on self-adaptive rubrics, which provides more granular process guidance to align the scoring process with human judgment. Additionally, we introduce a benchmark, INSDA, based on this framework. However, there are several limitations:
+
+- For questions with multiple correct answers, it requires manually writing multiple self-adaptive rubrics. It is worth noting that, to our knowledge, no current work focuses on the multi-solution direction.  
+- For subjective questions, such as creative writing, poetry, and other forms of artistic expression, different groups or individuals may have varying definitions of what constitutes good work. Therefore, it is necessary for each group or individual to set their own self-adaptive rubrics rather than relying on predefined ones. This also highlights the flexibility
+
+and interpretability of the self-adaptive rubrics paradigm we propose.
+
+# Ethical Considerations
+
+We propose a scoring paradigm based on self-adaptive rubrics to enhance the interpretability and controllability of the automated scoring process. This approach aims to improve the credibility of evaluation results produced by LLMs and to support the research community in advancing these models. Nevertheless, the inherent hallucinations within LLMs pose a challenge to ensuring the complete accuracy of automated evaluation outcomes. Therefore, we recommend incorporating human review of certain outputs when using LLMs as judges to increase the overall reliability and credibility of the process.
+
+Additionally, when generating self-adaptive rubrics for subjective questions, different groups or individuals may have varying definitions of what constitutes a good answer, potentially leading to biases and discrepancies. We encourage dialogue and mutual understanding among groups or individuals with diverse values, promoting the use of self-adaptive rubrics that align with their respective values and preferences.
+
+# References
+
+AI Anthropic. 2024. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card.  
+Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.  
+Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. 2023. Chateval: Towards better llmbased evaluators through multi-agent debate. Preprint, arXiv:2308.07201.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with  $90\%$  chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6.  
+Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E Gonzalez, et al. 2024. Chatbot arena: An open platform for evaluating llms by human preference. arXiv preprint arXiv:2403.04132.
+
+Daniel Deutsch, Rotem Dror, and Dan Roth. 2021. A statistical analysis of summarization evaluation metrics using resampling methods. Transactions of the Association for Computational Linguistics, 9:1132-1146.  
+Markus Freitag, David Grangier, and Isaac Caswell. 2020. BLEU might be guilty but references are not innocent. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 61-71, Online. Association for Computational Linguistics.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300.  
+Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. 2024. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems, 36.  
+Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangwoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, and Minjoon Seo. 2024a. Prometheus: Inducing fine-grained evaluation capability in language models. Preprint, arXiv:2310.08491.  
+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. Preprint, arXiv:2405.01535.  
+Tom Kocmi, Christian Federmann, Roman Grundkiewicz, Marcin Junczys-Dowmunt, Hitokazu Matsushita, and Arul Menezes. 2021. To ship or not to ship: An extensive evaluation of automatic metrics for machine translation. arXiv preprint arXiv:2107.10821.  
+Kalpesh Krishna, Aurko Roy, and Mohit Iyyer. 2021. Hurdles to progress in long-form question answering. arXiv preprint arXiv:2103.06332.  
+Junlong Li, Shichao Sun, Weizhe Yuan, Run-Ze Fan, Hai Zhao, and Pengfei Liu. 2023a. Generative judge for evaluating alignment. Preprint, arXiv:2310.05470.  
+Siyao Li, Deren Lei, Pengda Qin, and William Yang Wang. 2019. Deep reinforcement learning with distributional semantic rewards for abstractive summarization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6038-6044, Hong Kong, China. Association for Computational Linguistics.  
+Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023b. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval.
+
+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.
+
+Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023a. Gpteval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634.
+
+Yixin Liu, Alex Fabbri, Pengfei Liu, Yilun Zhao, Linyong Nan, Ruilin Han, Simeng Han, Shafiq Joty, Chien-Sheng Wu, Caiming Xiong, and Dragomir Radev. 2023b. Revisiting the gold standard: Grounding summarization evaluation with robust human evaluation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4140-4170, Toronto, Canada. Association for Computational Linguistics.
+
+Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076-12100, Singapore. Association for Computational Linguistics.
+
+Ehsan Montahaei, Danial Alihosseini, and Mahdieh Soleymani Baghshah. 2019. Jointly measuring diversity and quality in text generation models. Preprint, arXiv:1904.03971.
+
+OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo
+
+Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reiichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tsengpreston Tuggle Nick Turley Jerry Tworek Juan Felipe Ceron Uribe Andrea Vallone Arun Vijayvergiya Chelsea VossCarroll Wainwright Justin Jay Wang Alvin Wang Ben Wang Jonathan Ward Jason Wei CJ Weinmann Akila Welihinda Peter Welinder Jiayi Weng Lilian Weng Matt Wiethoff Dave Willner Clemens Winter Samuel Wolrich Hannah Wong Lauren Workman Sherwin Wu Jeff Wu Michael Wu Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. 2024. Gpt-4 technical report. Preprint arXiv:2303.08774.
+
+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 on Association for Computational Linguistics, ACL '02, page 311-318, USA. Association for Computational Linguistics.
+
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Al
+
+ban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. Preprint, arXiv:1912.01703.  
+Ehud Reiter. 2018. A structured review of the validity of BLEU. Computational Linguistics, 44(3):393-401.  
+Julian Risch, Timo Möller, Julian Gutsch, and Malte Pietsch. 2021. Semantic answer similarity for evaluating question answering models. In Proceedings of the 3rd Workshop on Machine Reading for Question Answering, pages 149-157, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Natalie Schluter. 2017a. The limits of automatic summarisation according to ROUGE. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 41-45, Valencia, Spain. Association for Computational Linguistics.  
+Natalie Schluter. 2017b. The limits of automatic summarisation according to rouge. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, pages 41-45. Association for Computational Linguistics.  
+Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. 2020. Megatron-lm: Training multi-billion parameter language models using model parallelism. Preprint, arXiv:1909.08053.  
+Matthew Snover, Bonnie Dorr, Rich Schwartz, Linnea Micciulla, and John Makhoul. 2006. A study of translation edit rate with targeted human annotation. In Proceedings of the 7th Conference of the Association for Machine Translation in the Americas: Technical Papers, pages 223-231, Cambridge, Massachusetts, USA. Association for Machine Translation in the Americas.  
+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.  
+Rickard Stureborg, Dimitris Alikaniotis, and Yoshi Suhara. 2024. Characterizing the confidence of large language model-based automatic evaluation metrics. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers), pages 76-89.  
+Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix,
+
+Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. Preprint, arXiv:2302.13971.  
+Yidong Wang, Zhuohao Yu, Zhengran Zeng, Linyi Yang, Cunxiang Wang, Hao Chen, Chaoya Jiang, Rui Xie, Jindong Wang, Xing Xie, Wei Ye, Shikun Zhang, and Yue Zhang. 2024. Pandalm: An automatic evaluation benchmark for llm instruction tuning optimization. Preprint, arXiv:2306.05087.  
+Seonghyeon Ye, Doyoung Kim, Sungdong Kim, Hyeonbin Hwang, Seungone Kim, Yongrae Jo, James Thorne, Juho Kim, and Minjoon Seo. 2023. Flask: Fine-grained language model evaluation based on alignment skill sets. arXiv preprint arXiv:2307.10928.  
+Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. 2022. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. Preprint, arXiv:1904.09675.  
+Ying Zhang and Stephan Vogel. 2004. Measuring confidence intervals for the machine translation evaluation metrics. In Proceedings of the 10th Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Preprint, arXiv:2306.05685.  
+Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. 2023. Agieval: A human-centric benchmark for evaluating foundation models. arXiv preprint arXiv:2304.06364.  
+Lianghui Zhu, Xinggang Wang, and Xinlong Wang. 2023. Judgelm: Fine-tuned large language models are scalable judges. Preprint, arXiv:2310.17631.
+
+# A Training Details
+
+We employed a default learning rate of 2e-5, and the batch size per device was dynamically adjusted based on the total data volume and the number of machines to maintain consistent optimization steps. The Adam optimizer was utilized.
+
+For the scaling law experiments, when  $n$  was below 300, we conducted three repetitions and averaged the results to minimize error.
+
+For all invoked APIs, we used the default parameters without extensive modifications.
+
+# B Data Annotation
+
+# B.1 Annotator Qualifications
+
+All our annotators are internal team members with at least a Master's degree. We provide additional compensation significantly higher than the standard salary, based on the amount of data annotated.
+
+# C Prompt Templates
+
+# C.1 AutoPrompt
+
+Below, I will provide a <Question>, along with the corresponding <Reference Answer> and <Scoring Rubric>. You need to evaluate the <Output Result> from the <Model Answer> of the <Model to be Assessed>. The evaluation should be divided into two parts: "Scoring Process" and "Final Score." Please note that the scoring range is from 0 to 5 points. You must justify the score you assign based on the <Model Answer>, strictly adhering to the requirements of the <Scoring Rubric> without adding, changing, or imagining any additional criteria.
+
+# C.2 Prompt for Set Matching
+
+You are a meticulous judge tasked with evaluating whether the "Test Rubric" provided by the user aligns with the "Standard Rubric." The evaluation rules are as follows:
+
+- The initial total score is set to zero.  
+- For each item in the "Test Rubric":
+
+1. If the item matches any item in the "Standard Rubric" exactly, one point is added to the total score.
+
+2. If the item in the "Test Rubric" is unrelated to any item in the "Standard Rubric," the total score remains unchanged.  
+3. If the item in the "Test Rubric" is the exact opposite of any item in the "Standard Rubric," one point is subtracted from the total score.
+
+You need to return the entire scoring process (explaining why points were added or subtracted) along with the final score. The return format should be:
+
+{ "Scoring Process": "<Here, provide the scoring process as a string>",
+
+"Final Score": "<Here, provide the final score as a mathematical expression, concluding with 'Final Score: <score>' e.g., '3/5=0.6, Final Score: 0.6>'
+
+The returned format must be compatible with json.load() to be converted into a dictionary.
+
+# C.3 Prosecutor Prompt
+
+Please check if the generated answer is correct. The reference answer is:  $\{\mathrm{gt}\}$ , and the generated answer is:  $\{\mathrm{user}\}$ . Please respond in the following format:  $\{\text{"result": True}\}$
+
+# C.4 Refinement Prompt
+
+Your generated answer is not the standard answer. Please reflect on this and generate a new answer.
+
+The generated scoring points and the full score answer are:
+
+# D Benchmark
+
+# D.1 Benchmark Leaderboard
+
+We provide the Benchmark LeaderBoard at the following link: https://github.com/Zhiyuan-Fan/self-adaptive-rubrics
+
+![](images/a7634cb60184aae7c523b465d602181dddec21c905a1593e72dabc74c9682ab9.jpg)  
+Figure 4: Benchmark multi-level classification system. To maintain conciseness, we have detailed the hierarchy only up to the third tier.
+
+# D.2 Benchmark Statistics
+
+We provide the Benchmark statistics at the following link: https://github.com/Zhiyuan-Fan/self-adaptive-rubrics
+
+# E Additional Experiments
+
+# E.1 Automatic Rubric Generation
+
+In this section, we conduct experiments on automatic rubric generation and analyze the results to explore its potential. Specifically, we compare the following three experimental setups: 1) using GPT-4 to automatically generate rubrics; 2) training an automatic rubric generation model with human-labeled rubrics under the SFT paradigm; and 3) aligning the model output with human preferences using DPO.
+
+# E.2 Error Propagating
+
+When using rubric generation models to automatically create self-adaptive rubrics, a potential issue is that if the generated rubric is inconsistent with the human-provided rubric, errors can accumulate in the scoring pipeline, leading to a larger deviation in the final score. By incorporating a filtering strategy, the overall performance will improve.
+
+# E.3 Joint Training vs. Expert Training
+
+We also explored whether to combine data from different categories for joint training when training the evaluator LM or rubric generation model, or to train a separate expert model for each category individually. We found that using joint training can achieve better results than expert training.
+
+![](images/67528f74254b32d18ef6494d9cd26383e440a7319d1badaaaadad247909d067c.jpg)  
+Figure 5: An illustrative example of the format used in self-adaptive rubrics.
+
+<table><tr><td>Model</td><td>Type</td><td>ACC</td><td>ACCT</td><td>Pearsonr</td><td>GSB</td></tr><tr><td>GPT-4</td><td>w self-adaptive rubrics</td><td>0.3241</td><td>0.7025</td><td>0.7283</td><td>0.9211</td></tr><tr><td>GPT-4</td><td>wo</td><td>0.2500</td><td>0.5035</td><td>0.4863</td><td>0.8684</td></tr><tr><td>GPT-4-turbo</td><td>w</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>GPT-4-turbo</td><td>wo</td><td>0.1995</td><td>0.5473</td><td>0.5326</td><td>0.9000</td></tr><tr><td>GPT-3.5</td><td>w</td><td>0.2121</td><td>0.5569</td><td>0.3888</td><td>0.8947</td></tr><tr><td>GPT-3.5</td><td>wo</td><td>0.1677</td><td>0.3872</td><td>0.1286</td><td>0.5158</td></tr></table>
+
+Table 4: Ablation Study  
+
+<table><tr><td>Type</td><td>OOD</td><td>ID</td><td>Random</td></tr><tr><td>GPT4-turbo</td><td>0.399</td><td>0.488</td><td>0.417</td></tr><tr><td>GPT4</td><td>0.602</td><td>0.613</td><td>0.613</td></tr><tr><td>xdg-turbo</td><td>0.606</td><td>0.607</td><td>0.603</td></tr></table>
+
+Table 5: Rubric Generation Results.  
+
+<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACCt</td><td>pearsonr</td></tr><tr><td>GPT-4</td><td>0.921</td><td>0.324</td><td>0.702</td><td>0.728</td></tr><tr><td>GPT-3.5</td><td>0.894</td><td>0.212</td><td>0.556</td><td>0.388</td></tr></table>
+
+Table 6: General Rubrics with Ground Truth.  
+
+<table><tr><td>Type</td><td>GSB</td><td>ACC</td><td>ACCt</td><td>pearsonr</td></tr><tr><td>GPT-4-turbo</td><td>0.9</td><td>0.199</td><td>0.547</td><td>0.532</td></tr><tr><td>GPT-4</td><td>0.868</td><td>0.250</td><td>0.503</td><td>0.486</td></tr><tr><td>GPT-3.5</td><td>0.515</td><td>0.167</td><td>0.387</td><td>0.128</td></tr></table>
+
+Table 7: General Rubrics without Groud Truth.
+
+![](images/7ccd63d48d08b349fb2d0ead2a92dcb5085a633cb55b23c8050a36f1b8bc557f.jpg)  
+Figure 6: Distribution change of the evaluator LM format data after applying the Human-AI Consistency strategy.
+
+![](images/ec57d49ca2e75e8c3ceb1d0a8fdc4218aa8c171a31ef1c137715a8c67ec12bd5.jpg)
+
+```txt
+Algorithm 1 Self-Adaptive Rubrics Generation and Validation  
+Require:  $Q$  ▷ Given question  
+Require:  $GT$  ▷ Ground truth answer  
+Require:  $n$  ▷ Maximum iterations  
+1:  $i \gets 0$   
+2: accepted ← False  
+3: while  $i < n$  and ¬accepted do  
+4:  $R, IA \gets \text{GPT-4}(Q)$  ▷ Generate rubrics and ideal answer  
+5: if  $PA(IA, GT)$  then ▷ Prosecutor agent checks ideal answer  
+6: accepted ← True  
+7: else  
+8: Inform GPT-4 of incorrect IA  
+9:  $i \gets i + 1$   
+10: end if  
+11: end while  
+12: if accepted then  
+13: return R  
+14: else  
+15: return Failure in  $n$  iterations  
+16: end if
+```

2024/SedarEval_ Automated Evaluation using Self-Adaptive Rubrics/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aaf3d0ddd23b8adcfa0db934d86b98a0071825ec372db6d2113d7598e66e73cf
+size 725727

2024/See Detail Say Clear_ Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning/081b7730-6494-4c1e-8ba3-d9c947e8be2f_content_list.json

@@ -0,0 +1,1486 @@
+[
+    {
+        "type": "text",
+        "text": "See Detail Say Clear: Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning",
+        "text_level": 1,
+        "bbox": [
+            166,
+            85,
+            831,
+            126
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Chengxin Zheng $^{1}$ , Junzhong Ji $^{1}$ , Yanzhao Shi $^{1}$ , Xiaodan Zhang $^{1,*}$ , Liangqiong Qu $^{2}$",
+        "bbox": [
+            144,
+            136,
+            855,
+            156
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ College of Computer Science, Beijing University of Technology, Beijing, China  \n $^{2}$ Department of Statistics and Actuarial Science, School of Computing and Data Science, The University of Hong Kong, Hong Kong, China",
+        "bbox": [
+            137,
+            170,
+            858,
+            221
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Correspondence: zhangxiaodan@bjut.edu.cn",
+        "bbox": [
+            356,
+            223,
+            640,
+            236
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            341,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Brain CT report generation is significant to aid physicians in diagnosing cranial diseases. Recent studies concentrate on handling the consistency between visual and textual pathological features to improve the coherence of report. However, there exist some challenges: 1) Redundant visual representing: Massive irrelevant areas in 3D scans distract models from representing salient visual contexts. 2) Shifted semantic representing: Limited medical corpus causes difficulties for models to transfer the learned textual representations to generative layers. This study introduces a Pathological Clue-driven Representation Learning (PCRL) model to build cross-modal representations based on pathological clues and naturally adapt them for accurate report generation. Specifically, we construct pathological clues from perspectives of segmented regions, pathological entities, and report themes, to fully grasp visual pathological patterns and learn cross-modal feature representations. To adapt the representations for the text generation task, we bridge the gap between representation learning and report generation by using a unified large language model (LLM) with task-tailored instructions. These crafted instructions enable the LLM to be flexibly fine-tuned across tasks and smoothly transfer the semantic representation for report generation. Experiments demonstrate that our method outperforms previous methods and achieves SoTA performance. Our code is available at https://github.com/Chauncey-Jheng/PCRL-MRG.",
+        "bbox": [
+            141,
+            285,
+            460,
+            770
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            778,
+            260,
+            794
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Brain computed tomography (CT) imaging is essential for diagnosing various cranial diseases, including cerebral infarction and hemorrhage. However, it is time-consuming and error-prone for radiologists to manually interpret medical findings from these scans and write reports. Automated report",
+        "bbox": [
+            112,
+            804,
+            490,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a6944c8749122044e7cba7df9684a4584674e72bec357b874b4aa436abf20f15.jpg",
+        "image_caption": [
+            "Figure 1: Comparison of different cross-modal alignment paradigms. (a) Previous way: The origin CT images contain extraneous information unrelated to diagnosis, and separate training for two semantic tasks makes it challenging to find a shared optimal solution, resulting in inadequate cross-modal alignment. (b) Our way: The visual representation is refined to concentrate on pathological clues, and employ joint training with task-tailored instructions via unified LLM to find a transferable representation, leading to better adaption for report generation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            259,
+            882,
+            495
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "generation systems are designed to boost efficiency, reduce the workload for radiologists, and optimize resources in busy clinical scenarios.",
+        "bbox": [
+            507,
+            707,
+            882,
+            755
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "With the advancement of deep neural networks and their successful application in image captioning tasks (Vinyals et al., 2015; Xu et al., 2015), medical report generation (MRG) has gained more attention. Unlike the short sentences of traditional image captioning, MRG aims to generate lengthy and precise reports. To achieve this, various cross-modal alignment methods are required to ensure the consistency between visual and textual information (Shi et al., 2023), including attention mech-",
+        "bbox": [
+            507,
+            760,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding Author",
+        "bbox": [
+            139,
+            906,
+            289,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "16542",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 16542-16552 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            208,
+            945,
+            786,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "anisms (Jing et al., 2018; Wang et al., 2018), memory mechanisms (Chen et al., 2020, 2021), and knowledge graphs (Li et al., 2019, 2023b).",
+        "bbox": [
+            112,
+            84,
+            485,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recently, learning representations via visual-textual contrastive learning (Shi et al., 2023, 2024) or using pre-trained large language models (LLMs) (Thawakar et al., 2023) to strength representations are also proven to be effective.",
+        "bbox": [
+            112,
+            133,
+            487,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, as shown in Figure 1(a), learning cross-modal correspondences is still challenging in current methods due to the following concerns: 1) Redundant visual representing. Different from chest X-ray data, 3D brain CT scans contain extensive redundant information, e.g. background and insignificant areas. With the lack of human-crafted boxes to locate pathology regions, models struggle to capture and interpret the visual pathology patterns for generating reports. Although current advanced models use semantic prior knowledge or medical prompts (Jin et al., 2024; Bu et al., 2024) to automatically learn the salient visual areas, this may introduce noise and unstable representation and cause severe hallucinations in generated texts. 2) Shifted semantic representing. Compared to natural corpus, limited brain CT report corpus is insufficient to transfer the pathological semantic representations learned by represent learning layers (e.g., contrastive learning layer) to the language model (Huh et al., 2024), since the direct weight-sharing (Shi et al., 2023, 2024) is prone to degrade the coherence of generated diagnostic sentences. Thus, how to uniformly represent cross-modal pathological features and adapt them to report generation still remains an open question.",
+        "bbox": [
+            115,
+            214,
+            489,
+            631
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a Pathological Cluedriven Representation Learning (PCRL) model to seamlessly build cross-modal representations based on diverse pathological clues and transfer them for generating accurate brain CT reports. Specifically, we extract pathological clues from perspectives of segmented regions, pathological entities, and report themes to depict clinical scenarios. Segmented region clues are automatically generated and filtered by given pathology prompts, enabling the visual encoder to grasp visual pathological patterns. Meanwhile, entity and theme clues are respectively extracted by detailed findings and full-text reports, to handle the enriched visual-textual alignment and build cross-modal pathology representations.",
+        "bbox": [
+            112,
+            632,
+            487,
+            872
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Besides, to adapt the learned representations for the text generation task, we bridge the gap between representation learning and report generation by",
+        "bbox": [
+            112,
+            873,
+            487,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "employing a unified large language model (LLM) with task-tailored instructions, which has proven to be more effective than conventional decoders by using appropriate tokens to seamlessly connect different tasks (Yu et al., 2023). As shown in Figure 1(b), We craft a representation instruction to prompt the LLM to produce high-level pathological semantic features for cross-modal alignment, and a generation instruction to prompt LLM to generate accurate brain CT reports based on the learned representations.",
+        "bbox": [
+            507,
+            84,
+            882,
+            259
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our main contributions can be summarized as:",
+        "bbox": [
+            527,
+            261,
+            873,
+            275
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We propose a novel framework to seamlessly learn visual-textual representations from perspectives of diverse pathological clues and leverage them for enhancing the quality of generated brain CT reports.",
+            "2. We, for the first time, design a new paradigm to effectively transfer the learned pathological representations for report generation using a unified LLM prompted by task-tailored instructions.",
+            "3. We validate the model capabilities on the open-source CTRG-Brain dataset. Experimental results show that our model achieves remarkable performance in generating brain CT reports."
+        ],
+        "bbox": [
+            524,
+            286,
+            884,
+            532
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            552,
+            663,
+            568
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Medical Report Generation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            577,
+            769,
+            593
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The advancements in image captioning techniques have spurred the development of a range of radiology report generation methods (Jing et al., 2018; Chen et al., 2020, 2021; Li et al., 2019; Shi et al., 2023; Zhang et al., 2023; Li et al., 2023b; Shi et al., 2024; Shen et al., 2024). To exploit the more effective parts of medical images, Jing et al. (2018) first propose a co-attention mechanism to associate images with disease tags and improve the accuracy of generated reports; Chen et al. (2021) design a cross-modal memory matrix to learn high-level vision-language correspondence for enhancing report generation; Liu et al. (2021c) used contrastive attention to captured the difference of abnormal and normal samples. To address the semantic bias of limited medical report corpus, Jing et al. (2020) exploited the textual structure information of reports, while Liu et al. (2021a); Shen et al. (2024) used curriculum learning to alleviate the data bias of limited medical data corpus. The retrieval-based",
+        "bbox": [
+            507,
+            599,
+            882,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "16543",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/161384d732f8271cc544bc4dd7d693aa996edc8b2faa05cd3aa2e30ecce89dc8.jpg",
+        "image_caption": [
+            "Figure 2: The overall framework of our method, which mainly consists of an image encoder and a text decoder for brain CT report generation (left). The pathological clue-driven representation learning (right) is proposed to guide the encoder and decoder for more fine-grained representation by three alignment modules: (a) segmentation clue alignment (SCA), (b) entity clue alignment (ECA), and (c) theme clue alignment (TCA)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            137,
+            86,
+            860,
+            303
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "(Li et al., 2018; Liu et al., 2021b) and prior expert knowledge based methods (Zhang et al., 2020; Yang et al., 2022; You et al., 2021) also shows effectiveness to alleviate the inaccuracy of generated report caused by limited training data. To build more stable relations between multi-modal features, contrastive learning (Radford et al., 2021) acts as an unsupervised method, and is proven to be effective for learning cross-modal representations for MRG tasks (Li et al., 2023b; Shi et al., 2024).",
+        "bbox": [
+            110,
+            400,
+            487,
+            561
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, LLM-based MRG methods (Jin et al., 2024; Chen et al., 2024) harness the semantic processing capabilities of LLMs to enhance the coherence of generated medical reports. However, the current bottleneck of LLM is to interpret sparse visual patterns of pathologies, which causes severe hallucinations and degrades the trustworthiness of clinical diagnosis. To solve this issue, we propose to augment LLM with representation learning guided by pathology clues. By aligning cross-modal pathological features and transferring the learned representations to the LLM-based report decoder, our model can generate accurate and coherent brain CT reports.",
+        "bbox": [
+            112,
+            563,
+            489,
+            788
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Pre-trained Large Model",
+        "text_level": 1,
+        "bbox": [
+            112,
+            802,
+            357,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In recent years, pre-trained large models have achieved significant breakthroughs in both natural language processing (NLP) and computer vision (CV). These models leverage extensive datasets for pre-training, allowing them to perform exceptionally well in downstream tasks. The Segment",
+        "bbox": [
+            112,
+            825,
+            487,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Anything Model (SAM) Kirillov et al. (2023) is an innovative large vision model in the field of image segmentation. SAM's strengths lie in its extensive training on a large-scale dataset, enhancing its versatility and enabling real-time generation of high-quality segmentation masks across diverse tasks and image distributions. MedCLIP Wang et al. (2022b) is a notable visual-language model in the medical domain. Pre-trained on decoupling medical images and texts, MedCLIP has shown SoTA performance on zero-shot image-text retrieval. LLaMA3 is the next generation of state-of-the-art open-source LLM from Meta (2024). Trained on over 15 trillion tokens, seven times the data of its predecessor, LLaMA3 incorporates extensive multilingual and high-quality datasets. To improve the use of large models for assisting in medical report generation, we adopt SAM for region segmentation, MedCLIP for feature representation of pathological entities, and LLaMA for joint training to achieve fine-grained cross-modal alignment and improve medical report generation.",
+        "bbox": [
+            507,
+            400,
+            884,
+            755
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            509,
+            776,
+            658,
+            793
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Figure 2, our framework consists of two branches: the brain CT Report Generation (RG) on the left and the Pathological Clue-driven Representation Learning (PCRL) on the right. The interaction between these branches is facilitated through the shared use of a visual encoder and a language model.",
+        "bbox": [
+            507,
+            808,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "16544",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Brain CT Report Generation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            389,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this branch, the input comprises a set of CT scan images  $I = \\{i_1,\\dots ,i_N\\}$ , where  $N$  represents the number of CT images in each sample. The output is the corresponding brain CT findings report  $Y = \\{y_{1},\\ldots ,y_{M}\\}$ , with  $M$  representing the number of tokens. We adopt the encoder-decoder architecture for report generation. First, we use ResNet101 (He et al., 2016) to extract grid features  $G = \\{g_{1},\\dots ,g_{N}\\} \\in \\mathbf{R}^{N\\times H\\times d}(N = 24,H = 196,d = 2048)$  from  $I$ . Then, we embed these features using a visual encoder to obtain visual features  $V$ . These visual features  $V$  are passed through a multi-modal projector to obtain the visual embedding tokens  $H_{v}\\in \\mathbf{R}^{N\\times d_{w}}(d_{w} = 4096)$  which are aligned with the word embedding space of a large language model. Finally, these tokens are concatenated with the instruction token  $H_{q}$  and fed into the large language model. We train the parameters  $\\theta$  by minimizing the cross-entropy loss, which can be represented by the following formula:",
+        "bbox": [
+            112,
+            105,
+            489,
+            428
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {g} = - \\sum_ {t = 1} ^ {M} \\log P \\left(y _ {t} \\mid y _ {1: t - 1}, H _ {v}, H _ {q}; \\theta\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            136,
+            438,
+            487,
+            481
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where,  $P(y_{t}|*)$  denotes the probability conditioned on  $H_{v}, H_{q}$ , and the embeddings of previous words  $y_{1}, y_{2}, \\dots, y_{t-1}$ .",
+        "bbox": [
+            112,
+            492,
+            487,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Pathological Clue-driven Representation Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            551,
+            478,
+            583
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Learning fine-grained visual-text representations is critical for generating accurate reports (Shi et al., 2023). This branch mainly builds enriched representations by conducting multi-modal feature alignment based on pathological clues, including segmentation clue alignment (SCA), entity clues alignment (ECA), and theme clues alignment (TCA).",
+        "bbox": [
+            112,
+            588,
+            489,
+            701
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Preparation of Pathological Clues",
+        "text_level": 1,
+        "bbox": [
+            112,
+            709,
+            438,
+            725
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pathological clues are collected in three perspectives for feature alignment.",
+        "bbox": [
+            112,
+            728,
+            489,
+            760
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "**Pathology Theme Clues:** Learning the structure of professional brain CT reports is essential for MRG models to satisfy human standards and produce reliable reports. To this end, we propose to build theme clues by simply using the full-text report and whole images as global signals, which can be useful to enhance the overall quality of medical reports by TCA (illustrated in 3.2.4).",
+        "bbox": [
+            112,
+            760,
+            489,
+            888
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pathology Entity Clues: To learn detailed information about pathology entities, we regard each",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "single finding sentence in the report as an entity clue and extract the related multi-modal features in ECA 3.2.3. The construction of entities  $E = \\{e_1,\\dots,e_{N_e}\\}$  (with  $N_{e} = 24$ ) is based on expert knowledge and the frequency of words in the training corpus.",
+        "bbox": [
+            507,
+            84,
+            884,
+            180
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Segment Clues: To enhance visual representations via detailed contour of pathologies, we propose to utilize the pre-trained segmentation model SAM (Kirillov et al., 2023) to generate mask candidates and filter useful masks related to pathology entities. First, we prompt SAM by covering each brain CT image with a grid of points and integrating it into image embeddings through average sampling. In this way, the mask decoder in SAM is prompted to generate a gallery of candidate masks  $M_I = \\{m_1, \\dots, m_{N_m}\\}$ . To ensure the quality of masks, we also apply rule-based methods to filter out low-quality and duplicate masks with area size, stability scores, and IoU scores. For each sample consisting of 24 images, we generate a corresponding series of masks and combine them to form the candidate Gallery  $= \\{M_{I_1}, \\dots, M_{I_N}\\}$ .",
+        "bbox": [
+            507,
+            181,
+            884,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Then, to retrieve valuable masks related to the sample's pathological entities, we propose to use the MedCLIP (Wang et al., 2022b) for text-prompted retrieval. Based on expert knowledge, we divide the 24 images into eight-layer categories, each mapping a specific set of entities (Shi et al., 2024). Conversely, each entity also has its corresponding layers. We extract the existing entities from the sample's report and their corresponding pathological descriptions  $D = \\{d_{1},\\dots,d_{N_{d}}\\}$  ( $N_{d}\\leq N_{e}$ ), which serve as the Query. Next, we use the  $d$ -th description  $Query_{d}$  to search for the most similar matching mask in  $Gallery_{d}$ , which is a subset of  $Gallery$ , consisting of all visual masks of the layers related to the existing entities. The procedure can be represented as:",
+        "bbox": [
+            507,
+            456,
+            882,
+            712
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {R e t r i e v a l} _ {d} = \\operatorname {C L I P} \\left(\\operatorname {G a l l e r y} _ {d}, \\operatorname {Q u e r y} _ {d}\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            728,
+            882,
+            745
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $Retrieval_{d}$  denotes the retrieved masks for SCA 3.2.2.",
+        "bbox": [
+            507,
+            760,
+            882,
+            791
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Segmentation Clue Alignment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            804,
+            806,
+            820
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This module aims to learn the fine-grained visual representation based on the extracted segment masks. First, we obtain the visual features  $V_{D} = \\{v_{1},\\dots v_{N_{d}}\\}$  corresponding to the images containing the entities using a selector. Then, we input these features into a lightweight segmentation",
+        "bbox": [
+            507,
+            825,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "16545",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "head to generate the corresponding foreground entity masks  $S_{D} = \\{s_{1},\\dots,s_{N_{d}}\\}$ , which serves as discriminated pathological information. We then align these generated masks with the retrieved segment masks  $M_{D} = \\{Retrieval_{1},\\dots,Retrival_{N_{d}}\\}$  to learn detailed patterns. It is important to note that the size of the generated masks  $S_{D}$  differs from the size of the retrieved SAM masks  $M_{D}$ . To address this, we first resize the SAM masks to match the size of the masks generated by our segmentation head. We finally calculate the following loss for aligning the two types of masks:",
+        "bbox": [
+            112,
+            84,
+            489,
+            277
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {s e g} = 1 - \\frac {2 \\sum_ {i = 1} ^ {n} p _ {i} y _ {i}}{\\sum_ {i = 1} ^ {n} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {n} y _ {i} ^ {2}} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            174,
+            281,
+            487,
+            317
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $y_{i}$  is the pixel value of retrieved SAM masks and  $p_i$  is the predicted pixel value.",
+        "bbox": [
+            112,
+            321,
+            485,
+            351
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this way, the visual encoder can be effectively learned to focus on the areas of pathologies while reducing the influence of irrelevant visual features.",
+        "bbox": [
+            112,
+            353,
+            487,
+            401
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.3 Entity Clue Alignment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            409,
+            352,
+            425
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To learn the cross-modal patterns of pathology entities, we extract visual and textual pathological features based on the entity clues. First, we build a selector to obtain the visual features  $V_{D} = \\{v_{1},\\dots v_{N_{d}}\\} (N_{d} < N)$ , which corresponds to significant CT images that exist pathology entities. We then apply global average pooling (GAP) to generalize  $V_{D}$  and obtain representations of each significant CT image, denoted as  $R_v^e = \\{r_1,\\dots r_{N_d}\\}$ .",
+        "bbox": [
+            112,
+            428,
+            489,
+            571
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Different from previous work (Li et al., 2023b) use an external language model (e.g. SciBert (Beltagy et al., 2019)) to build textual representation for feature alignment, we propose to leverage a unified LLM to generate textual representation via tailored prompts. Inspired by Wang et al. (2023), we crafted a representation instruction \"Summarize the following cranial diagnosis in one word,\" which is denoted as  $H_{q_r}$ . This prompt can effectively activate the summarization ability of LLM to generate corresponding text representations for pathological entity description in  $D$ . We use the output generated by the final hidden layer of LLM as textual features for cross-modal alignment.",
+        "bbox": [
+            112,
+            573,
+            489,
+            796
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "With the carefully extracted visual and textual representations of entities, we map them into the same dimension through an embedding layer. The process can be represented by the following formula:",
+        "bbox": [
+            112,
+            797,
+            487,
+            876
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {v} ^ {e} = \\operatorname {L i n e a r} _ {v} ^ {e} (G A P \\left(V _ {D}\\right)) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            169,
+            885,
+            487,
+            901
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {w} ^ {e} = \\operatorname {L i n e a r} _ {t} ^ {e} \\left(L L M \\left(H _ {q _ {r}}, D\\right)\\right) \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            904,
+            485,
+            922
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $GAP$  denotes the global average pooling,  $Linear_v$  represents the visual mapper, and  $Linear_t$  represents the textual mapper. We employ the symmetric InfoNCE (van den Oord et al., 2018) loss for visual-textual alignment:",
+        "bbox": [
+            507,
+            84,
+            884,
+            165
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {C L _ {e}} = - \\frac {1}{2} \\left(\\alpha_ {v} \\sum_ {i = 1} ^ {N _ {d}} \\log \\frac {\\exp \\left(\\mathrm {s} _ {v} ^ {e} (i , i)\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {d}} \\exp \\left(\\mathrm {s} _ {v} ^ {e} (i , j)\\right)} \\right. \\\\ + \\alpha_ {w} \\sum_ {i = 1} ^ {N _ {d}} \\log \\frac {\\exp \\left(\\mathrm {s} _ {w} ^ {e} (i , i)\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {d}} \\exp \\left(\\mathrm {s} _ {w} ^ {e} (i , j)\\right)} \\Bigg) \\tag {6} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            509,
+            183,
+            880,
+            288
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $s_v^e (i,j) = \\frac{\\sin(R_{v_i}^e,R_{w_j}^e)}{\\tau}$  and  $s_w^e (i,j) = \\frac{\\sin(R_{w_i}^e,R_{v_j}^e)}{\\tau}$  denote the similarity between the visual representation  $R_{v}^{e}$  and the textual representation  $R_{w}^{e}$ ,  $\\tau$  is a temperature parameter.  $\\alpha_{v}$  and  $\\alpha_{w}$  are hyperparameters to balance the contrastive learning.",
+        "bbox": [
+            507,
+            294,
+            882,
+            390
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By aligning multi-modal entity clues, the model can grasp fine-grained visual-text representations to generate accurate diagnostic words.",
+        "bbox": [
+            507,
+            392,
+            880,
+            439
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.4 Theme Clue Alignment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            448,
+            752,
+            463
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Theme clue alignment aims to equip the model with comprehensive skills to generate accurate style and structure of reports, thereby enhancing clinical reliability. For global visual features, we use one-dimensional global max pooling (GMP) to represent the entire sample visually, denoted as  $R_{w}^{t}$ . For the overall report of the sample, we generate textual representation  $R_{w}^{t}$  by prompting LLM with the same representation instruction (see Section 3.2.3). This process can be represented as:",
+        "bbox": [
+            505,
+            466,
+            882,
+            626
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {v} ^ {t} = \\operatorname {L i n e a r} _ {v} ^ {t} (G M P (V)) \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            632,
+            880,
+            650
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {w} ^ {t} = \\operatorname {L i n e a r} _ {t} ^ {t} \\left(L L M \\left(H _ {q _ {s}}, Y\\right)\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            564,
+            652,
+            880,
+            670
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Similar to ECA, the loss of TCA can be formulated as:",
+        "bbox": [
+            507,
+            676,
+            880,
+            705
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {C L _ {t}} = - \\frac {1}{2} \\left(\\alpha_ {v} \\sum_ {i = 1} ^ {N _ {b}} \\log \\frac {\\exp (\\mathrm {s} _ {v} ^ {t} (i , i))}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {b}} \\exp (\\mathrm {s} _ {v} ^ {t} (i , j))}\\right) \\\\ + \\alpha_ {w} \\sum_ {i = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\mathrm {s} _ {w} ^ {t} (i , i)\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {b}} \\exp \\left(\\mathrm {s} _ {w} ^ {t} (i , j)\\right)} \\tag {9} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            509,
+            711,
+            880,
+            816
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $s_v^t (i,j) = \\sin (R_{v_i}^t,R_{w_j}^t) / \\tau$  and  $s_w^t (i,j) =$ $\\sin (R_{w_i}^t,R_{v_j}^t) / \\tau$  denote the similarity between the visual representation  $R_{v}^{e}$  and the textual representation  $R_w^e$ ,  $\\tau$  is a temperature parameter,  $N_{b}$  is the batch size. Shared with ECA,  $\\alpha_{v}$  and  $\\alpha_{w}$  are set to balance the feature alignment.",
+        "bbox": [
+            507,
+            822,
+            882,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "16546",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/c426badad0e5b81b6fa3b1318f504f5b95288b66aa4ff0a9f079464eb2279d6d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>Decoder</td><td>B1</td><td>B2</td><td>B3</td><td>B4</td><td>M</td><td>RG</td><td>C</td><td>F1</td></tr><tr><td>HRNN(Krause et al., 2017)†</td><td>LSTM</td><td>42.3</td><td>28.3</td><td>21.2</td><td>17.1</td><td>27.2</td><td>39.3</td><td>20.9</td><td>70.2</td></tr><tr><td>Up-Down(Anderson et al., 2018)†</td><td>LSTM</td><td>45.8</td><td>34.8</td><td>28.5</td><td>24.4</td><td>31.6</td><td>42.5</td><td>27.3</td><td>70.2</td></tr><tr><td>WCL(Yan et al., 2021)†</td><td>LSTM</td><td>49.5</td><td>36.5</td><td>29.4</td><td>25.1</td><td>31.3</td><td>42.8</td><td>33.3</td><td>64.5</td></tr><tr><td>R2Gen-CMN(Chen et al., 2021)†</td><td>Transformer</td><td>49.1</td><td>40.0</td><td>34.4</td><td>30.1</td><td>29.9</td><td>48.6</td><td>84.2</td><td>69.8</td></tr><tr><td>XProNet(Wang et al., 2022a)†</td><td>Transformer</td><td>50.6</td><td>41.3</td><td>34.4</td><td>29.1</td><td>31.5</td><td>51.7</td><td>83.3</td><td>70.1</td></tr><tr><td>WGAM(Yang et al., 2021)†</td><td>LSTM</td><td>49.4</td><td>36.7</td><td>29.6</td><td>25.4</td><td>32.0</td><td>42.4</td><td>31.9</td><td>68.9</td></tr><tr><td>PGCA(Shi et al., 2023)†</td><td>LSTM</td><td>50.2</td><td>37.8</td><td>30.7</td><td>26.5</td><td>32.5</td><td>43.0</td><td>34.0</td><td>69.2</td></tr><tr><td>WGAM-HI(Zhang et al., 2023)†</td><td>LSTM</td><td>50.4</td><td>37.6</td><td>30.5</td><td>26.1</td><td>31.4</td><td>43.8</td><td>33.2</td><td>67.4</td></tr><tr><td>LLaVA-med(Li et al., 2023a)†</td><td>LLaMA3-8B</td><td>50.0</td><td>39.3</td><td>32.0</td><td>26.3</td><td>31.4</td><td>46.7</td><td>38.6</td><td>64.3</td></tr><tr><td>HILT(Liu et al., 2024)†</td><td>LLaMA3-8B</td><td>50.7</td><td>39.9</td><td>33.1</td><td>27.9</td><td>30.8</td><td>46.1</td><td>43.7</td><td>68.4</td></tr><tr><td>PromptMRG(Jin et al., 2024)†</td><td>LLaMA3-8B</td><td>48.1</td><td>38.3</td><td>31.6</td><td>26.5</td><td>31.0</td><td>47.4</td><td>50.3</td><td>69.2</td></tr><tr><td>Ours (Jieba tokenizer)</td><td>LLaMA3-8B</td><td>51.5</td><td>42.0</td><td>35.7</td><td>30.9</td><td>31.4</td><td>49.0</td><td>80.0</td><td>70.6</td></tr><tr><td>Ours (LLaMA tokenizer)</td><td>LLaMA3-8B</td><td>62.0</td><td>54.7</td><td>49.4</td><td>45.3</td><td>33.1</td><td>57.7</td><td>96.4</td><td>70.6</td></tr></table>",
+        "bbox": [
+            127,
+            82,
+            878,
+            307
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The performance of our PCRL compared with previous state-of-the-art models on the Brain CT report generation dataset CTRG-Brain. The best results are highlighted in bold.  $\\dagger$  denotes the re-implementation results.",
+        "bbox": [
+            112,
+            323,
+            880,
+            354
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3 Joint Training",
+        "text_level": 1,
+        "bbox": [
+            112,
+            378,
+            273,
+            393
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In the training stage, we jointly train the RG branch and the PCRL branch to maximize the utilization of multi-granularity visual-text representations. Instead of using separate modules to learn representation and generate medical reports (Li et al., 2023b), we use a unified LLM to bridge the gap between representation learning and report generation via two task-tailored instructions, i.e., representation instruction and generation instruction. This can transfer the representations learned by the PCRL branch to optimize the RG branch effectively, thereby generating accurate reports.",
+        "bbox": [
+            112,
+            400,
+            489,
+            593
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Our final loss contains the above-mentioned losses, which can be formulated as:",
+        "bbox": [
+            112,
+            595,
+            487,
+            626
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {g} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {C L _ {e}} + \\mathcal {L} _ {C L _ {t}} \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            643,
+            487,
+            659
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            677,
+            258,
+            694
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Dataset",
+        "text_level": 1,
+        "bbox": [
+            112,
+            705,
+            220,
+            719
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We validate the performance of our model using the CTRG-Brain (Tang et al., 2024) dataset. This dataset comprises 6,000 samples, containing a total of 160,336 CT images and 6,000 Chinese medical reports. Following the mainstream division method (Shi et al., 2023; Zhang et al., 2023), we split the dataset into a training set, a validation set, and a test set in a 7:1:2 ratio. For consistent processing, we divide the brain CT image samples into 8 layers based on expert knowledge, with each layer containing 3 continuous CT images, assigning each sample with 24 CT images.",
+        "bbox": [
+            112,
+            728,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            378,
+            705,
+            392
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We chose Natural Language Generation (NLG) metrics and Clinical Evaluation (CE) metrics to evaluate our model's performance. NLG metrics include BLEU (Papineni et al., 2002),METEOR (Banerjee and Lavie), ROUGE-L (Lin, 2004), and CIDEr (Vedantam et al., 2015), denoted as B1, B2, B3,B4,M,RG,and C respectively. To measure the pathological accuracy, we use 24 keywords summarized by experienced radiologists to calculate the Clinical Evaluation (CE) metric(Shi et al., 2023; Zhang et al., 2023), i.e., F1 score, which is the harmonic mean of the precision and recall.",
+        "bbox": [
+            507,
+            405,
+            884,
+            598
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            620,
+            741,
+            636
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We reshape the size of image to  $512 \\times 512$  pixels and used a ResNet101 to extract image features, which is pre-trained on the ImageNet dataset and fine-tuned on the CQ500 dataset (Chilamkurthy et al., 2018). For our large language model, we utilize LLaMA3-8B (Meta, 2024), which is quantized to 4-bit, and use the LoRa (Hu et al., 2021) for parameter-efficient fine-tuning. The overall trainable parameter quantity of our model is  $229.9\\mathrm{M}$  with  $3.4\\mathrm{M}$  parameters in LLM (only  $0.04\\%$ ). During training, we use the AdamW optimizer with a learning rate of 1e-4. The batch size is 4, with 1050 training steps per epoch. For testing, we set the temperature coefficient of the large model to 0.6 and the top-p value to 0.9. The model is implemented using PyTorch 2.3.0, and the entire training process is conducted on a single RTX 4090 GPU.",
+        "bbox": [
+            507,
+            646,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "16547",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/34d6436b42c4da7b97f621fff444bf66b3b7236fe3af95bbc27426db04e6ca27.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">Representation</td><td colspan=\"3\">Module Loss</td><td rowspan=\"2\">B1</td><td rowspan=\"2\">B2</td><td rowspan=\"2\">B3</td><td rowspan=\"2\">B4</td><td rowspan=\"2\">M</td><td rowspan=\"2\">RG</td><td rowspan=\"2\">C</td></tr><tr><td>Visual</td><td>Textual</td><td>LSCA</td><td>LECA</td><td>LTCA</td></tr><tr><td>Baseline</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>49.6</td><td>39.9</td><td>33.4</td><td>28.6</td><td>30.7</td><td>48.4</td><td>50.8</td></tr><tr><td>(a)</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>51.3</td><td>41.0</td><td>34.5</td><td>29.7</td><td>31.2</td><td>47.7</td><td>70.8</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>51.3</td><td>41.4</td><td>35.0</td><td>30.1</td><td>31.2</td><td>48.7</td><td>71.3</td></tr><tr><td>(c)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>51.7</td><td>41.4</td><td>34.6</td><td>29.5</td><td>31.2</td><td>48.0</td><td>72.8</td></tr><tr><td>(d)</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>48.9</td><td>38.8</td><td>32.2</td><td>27.4</td><td>30.0</td><td>46.3</td><td>58.3</td></tr><tr><td>(e)</td><td>✓</td><td>✓(Bert)</td><td>✓</td><td>✓</td><td>✓</td><td>49.5</td><td>39.0</td><td>31.8</td><td>26.3</td><td>30.4</td><td>45.4</td><td>48.5</td></tr><tr><td>Ours</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>51.5</td><td>42.0</td><td>35.7</td><td>30.9</td><td>31.4</td><td>49.0</td><td>80.0</td></tr></table>",
+        "bbox": [
+            127,
+            82,
+            877,
+            227
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Ablation studies of our proposed method. The Baseline model is an encoder-decoder framework without an alignment mechanism. (a), (b), (c), (d) and (e) respectively denote the use of different representations (i.e., visual and textual in ECA and TCA) and module losses.",
+        "bbox": [
+            112,
+            236,
+            884,
+            280
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Quantitative Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            305,
+            331,
+            322
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare the proposed PCRL with some competitive brain CT report generation methods (WGAM (Yang et al., 2021), PGCA (Shi et al., 2023), WGAM-HI(Zhang et al., 2023)). Besides, we also reproduced some SOTA models in image captioning (HRNN (Krause et al., 2017), Up-Down (Anderson et al., 2018)) and chest Xray report generation (WCL (Yan et al., 2021), R2Gen-CMN (Chen et al., 2021), XProNet (Wang et al., 2022a)) for comprehensive comparisons on CTRG-Brain dataset. What's more, for fair comparisons, We also reproduce several related LLM-based methods (LLaVA-Med(Li et al., 2023a), HILT(Liu et al., 2024), PromptMRG(Jin et al., 2024)with the same LLM decoder.",
+        "bbox": [
+            112,
+            332,
+            489,
+            575
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 1, our method outperforms others across most evaluation metrics. WGAM and WGAM-HI employ weakly-supervised visual attention to extract key visual features, resulting in higher BLEU scores. With contrastive learning for cross-modal alignment, WCL and PGCA can effectively learn relations between CT images and reports, achieving better results. However, due to the lack of training data, the above methods still produce reports that fall short in fluency and readability. Here's a refined version of your sentence: The LLM-based methods (LLaVA-med, HILT, PromptMRG) demonstrate poorer performance compared to traditional transformer-based methods (R2Gen-CMN, XProNet). This suggests that prior knowledge from pretrained large models, without further cross-modal alignment, may lead to hallucinations in report generation.",
+        "bbox": [
+            112,
+            579,
+            489,
+            869
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Our PCRL achieves fine-grained cross-modal alignment by leveraging a series of pre-trained large models, generating more fluent reports and",
+        "bbox": [
+            112,
+            873,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "achieves the best performance compared to other methods. It is noteworthy that we tested our method using the Jieba and LLaMA tokenizers respectively to compute NLG scores, with the latter achieving the best performance across all NLG metrics.",
+        "bbox": [
+            507,
+            305,
+            884,
+            400
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "This discrepancy may be due to differences in tokenization methods used during training. While other models employ the traditional Jieba tool for Chinese word tokenization, our PCRL follows more advanced BPE-based subword tokenization. Nevertheless, our model also achieves competitive results in overall metrics.",
+        "bbox": [
+            507,
+            401,
+            885,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            525,
+            675,
+            539
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the effect of each component in PCRL, we have done plenty of ablation studies, as shown in Table 2. Baseline is the standard encoder-decoder (ResNet101-LLaMA3) architecture without alignment. By progressively adding visual or textual representation (visual and textual) and the module losses  $(\\mathcal{L}_{SCA},\\mathcal{L}_{ECA}$  and  $\\mathcal{L}_{TCA})$  , respectively denoting the incorporation of two modalities of representations and the utilization of three pathological clue-driven alignments (i.e., SCA, ECA, and TCA).",
+        "bbox": [
+            507,
+            545,
+            884,
+            721
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "By comparing (a) with the baseline, we observe that incorporating the segmentation alignment significantly enhances the model's performance. This demonstrates that our SCA method effectively filters out irrelevant information from CT images and extracts crucial pathology-related visual regions. In contrast, with the implementation of ECA, (b) achieves comparable or better performance across all metrics compared to (a). This indicates that fine-grained cross-modal pathological entity alignment can effectively learn strongly correlated visual-text",
+        "bbox": [
+            507,
+            722,
+            885,
+            900
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "https://github.com/fxsjy/jieba",
+        "bbox": [
+            534,
+            906,
+            721,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "16548",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/45bdebb06d6c1e987ffd7e26ccf90f172f571581eaebac5c779e9b053deb61c1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Ground Truth Sample</td><td>Ground Truth Entities</td><td>Retrieved Masks</td><td>Generated Reports</td><td></td></tr><tr><td>...</td><td>基底节区(basal ganglia):左侧基底节区见小片状低密度影 (Small patchy low-density shadows visible in the left basal ganglia area)</td><td></td><td>双侧大脑半球对称,两侧基底节区见斑点状低密度影,中线结构居中,脑沟、脑池、蛛网膜下腔稍增宽,脑回变小。(Bilateral cerebral hemispheres are symmetrical, with punctate low-density shadows in the basal ganglia on both sides. The midline structure is centered, and the sulcus, cistern, and subarachnoid space are slightly widened, while the gyrus is smaller.)</td><td rowspan=\"3\">Baseline</td></tr><tr><td rowspan=\"3\">双侧大脑半球对称,左侧基底节区见小片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑,脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A small patch of low-density shadow is seen in the left basal ganglia region, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal.)</td><td>脑室(ventricle):各脑室、脑池大小形态正常(The size and shape of all ventricles and cisterns are normal.)</td><td rowspan=\"3\"></td><td>双侧大脑半球对称,右侧丘脑见小片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑、脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A small patch of low-density shadow is seen in the right thalamus, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal. The midline structure is centered. There are no abnormalities in the cerebellum and brainstem below the tentorium.)</td></tr><tr><td>中线(midline structure):中线结构居中 (The midline structure centered)</td><td rowspan=\"2\">双侧大脑半球对称,左侧基底节区见斑片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑、脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A patch of low-density shadow is seen in the left basal ganglia region, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal. The midline structure is centered. There are no abnormalities in the cerebellum and brainstem below the tentorium.)</td></tr><tr><td>脑干(brainstem):幕下小脑,脑干无异常(There are no abnormalities in the cerebellum and brainstem below the tentorium)</td><td>Ours</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            884,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "representations, thus better supporting medical report generation. Besides, (c) combines both SCA and ECA and shows improved performance in the B1 and C metrics, which are highly correlated with keyword frequency. However, it slightly underperforms in the B2, B4 and RG metrics compared to (b), which is more related to overall text style.(d) showed SCA is essential for focusing on entity details and maintaining the thematic style.",
+        "bbox": [
+            112,
+            398,
+            487,
+            543
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Notably, to validate the contribution of joint representation learning, we replace the language model with Bert model for textual representation during alignment in (e). The result indicated that shared representation through joint learning with unified language model can effectively improve the overall quality of the task.",
+        "bbox": [
+            112,
+            545,
+            487,
+            657
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6 Qualitative Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            671,
+            319,
+            688
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We visualize the brain CT reports generated by baseline, WGAM-HI (Zhang et al., 2023) and our model in Figure 3. Given the ground truth brain CT sample and entity data, our model generates better brain CT report with the most accurate entity words (e.g. basal ganglia, ventricle, and brainstem) among the competitors, which demonstrates the effectiveness of using diverse medical clues to learn enriched multi-modal representations. It can be also seen that compared with the baseline, the report generated by our model has a better semantic structure, indicating the contribution of employing a unified LLM to transfer useful representations. Besides, we also find that the retrieved entity masks",
+        "bbox": [
+            112,
+            696,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/7719a15e36aa57ee5e55575f8b57839495bd7278f8f3dc11df92784ea5e8be81.jpg",
+        "image_caption": [
+            "Figure 3: Visualization of report generation and mask segmentation. Given the ground truth sample and corresponding entities, the retrieved entity masks are listed in the third column. Reports generated by Baseline, WGAM-HI, and our model are listed in the fourth column. Different colors denote the specific entity words and entity masks, respectively. The English translation is given for a better understanding of the original Chinese reports in CTRG-Brain.",
+            "Figure 4: Visualization of the segmentation masks gallery generated by SAM for a single sample."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            399,
+            868,
+            492
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "from segmentation masks gallery (see Figure 4) can generally match related entity words at both levels of visual and semantic. For example, the mask of \"midline structure\" and \"brainstem\" matches the empirical scan slice and fine-grained region chosen by experienced radiologists. This guarantees the model to mine accurate visual cranial patterns, therefore generating high-quality reports.",
+        "bbox": [
+            505,
+            560,
+            884,
+            689
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            701,
+            640,
+            715
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We propose a novel model to mine pathological clues for enhancing multi-modal representations and seamlessly transfer them into report generation. First, through carefully designed segmentation clue alignment, entity clue alignment, and theme clue alignment, the diverse and precise feature representation can be well-constructed. Second, we transfer the learned representation to boost the brain CT report generation via a unified LLM prompted by task-tailored instructions. Experiments demonstrate the effectiveness of our model in generating pathologically accurate reports.",
+        "bbox": [
+            505,
+            728,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "16549",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Although the segmentation clues retrieved by MedClip (Wang et al., 2022b) can generally match corresponding pathological entities and help the model neglect redundant visual information, it should be noted that a part of retrieved entity masks may not be precise. This is because MedClip is mainly pretrained by chest X-ray data with limited brain CT samples. Thus, addressing this challenge is imperative for the research community. In the future, we will work on exploring useful approaches. One potential approach is to train a unified text-prompted medical segmentation model towards 3D brain CT scans, which can not only be employed to offer fine-grained visual information for medical report generation but also for other related tasks, e.g. medical VQA.",
+        "bbox": [
+            112,
+            109,
+            492,
+            367
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            394,
+            213,
+            409
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086.",
+            "S Banerjee and A Lavie. Meteor: an automatic metric for mt evaluation with high levels of correlation with human judgments. Proceedings of ACL-WMT, pages 65-72.",
+            "Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. Scibert: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3613-3618.",
+            "Shenshen Bu, Yujie Song, Taiji Li, and Zhiming Dai. 2024. Dynamic knowledge prompt for chest x-ray report generation. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, LREC/COLING, pages 5425-5436. ELRA and ICCL.",
+            "Zhihong Chen, Yaling Shen, Yan Song, and Xiang Wan. 2021. Cross-modal memory networks for radiology report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 5904-5914.",
+            "Zhihong Chen, Yan Song, Tsung-Hui Chang, and Xi-ang Wan. 2020. Generating radiology reports via"
+        ],
+        "bbox": [
+            115,
+            417,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "memory-driven transformer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 1439-1449.",
+            "Zhixuan Chen, Luyang Luo, Yequan Bie, and Hao Chen. 2024. Dia-llama: Towards large language model-driven CT report generation. CoRR, abs/2403.16386.",
+            "Sasank Chilamkurthy, Rohit Ghosh, Swetha Tanamala, Mustafa Biviji, Norbert G. Campeau, Vasantha Kumar Venugopal, Vidur Mahajan, Pooja Rao, and Prashant Warier. 2018. Development and validation of deep learning algorithms for detection of critical findings in head ct scans. arXiv preprint arXiv:1803.05854.",
+            "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778.",
+            "Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.",
+            "Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. 2024. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987.",
+            "Haibo Jin, Haoxuan Che, Yi Lin, and Hao Chen. 2024. Promptmrg: Diagnosis-driven prompts for medical report generation. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pages 2607-2615.",
+            "Baoyu Jing, Zeya Wang, and Eric Xing. 2020. Show, describe and conclude: On exploiting the structure information of chest x-ray reports. arXiv preprint arXiv:2004.12274.",
+            "Baoyu Jing, Pengtao Xie, and Eric P. Xing. 2018. On the automatic generation of medical imaging reports. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 2577-2586.",
+            "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. 2023. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026.",
+            "Jonathan Krause, Justin Johnson, Ranjay Krishna, and Li Fei-Fei. 2017. A hierarchical approach for generating descriptive image paragraphs. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3337-3345."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "16550",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "C Li, C Wong, S Zhang, et al. 2023a. Llava-med: training a large language-and-vision assistant for biomedicine in one day. arxiv. arXiv preprint arXiv:2306.00890.",
+            "Christy Y. Li, Xiaodan Liang, Zhiting Hu, and Eric P. Xing. 2019. Knowledge-driven encode, retrieve, paraphrase for medical image report generation. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 6666-6673.",
+            "Mingjie Li, Bingqian Lin, Zicong Chen, Haokun Lin, Xiaodan Liang, and Xiaojun Chang. 2023b. Dynamic graph enhanced contrastive learning for chest x-ray report generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3334-3343.",
+            "Yuan Li, Xiaodan Liang, Zhiting Hu, and Eric P Xing. 2018. Hybrid retrieval-generation reinforced agent for medical image report generation. Advances in neural information processing systems, 31.",
+            "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.",
+            "Che Liu, Zhongwei Wan, Yuqi Wang, Hui Shen, Haozhe Wang, Kangyu Zheng, Mi Zhang, and Rossella Arcucci. 2024. Benchmarking and boosting radiology report generation for 3d high-resolution medical images. arXiv preprint arXiv:2406.07146.",
+            "Fenglin Liu, Shen Ge, and Xian Wu. 2021a. Competence-based multimodal curriculum learning for medical report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 3001-3012.",
+            "Fenglin Liu, Xian Wu, Shen Ge, Wei Fan, and Yuexian Zou. 2021b. Exploring and distilling posterior and prior knowledge for radiology report generation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 13753-13762.",
+            "Fenglin Liu, Changchang Yin, Xian Wu, Shen Ge, Yuexian Zou, Ping Zhang, and Xu Sun. 2021c. Contrastive attention for automatic chest x-ray report generation. arXiv preprint arXiv:2106.06965.",
+            "AI Meta. 2024. Introducing meta llama 3: The most capable openly available llm to date. Meta AI.",
+            "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, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "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, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763.",
+            "Qingya Shen, Yanzhao Shi, Xiaodan Zhang, Junzhong Ji, Ying Liu, and Huimin Xu. 2024. Ghcl: Gaussian heuristic curriculum learning for brain ct report generation. Multimedia Systems, 30(2):69.",
+            "Yanzhao Shi, Junzhong Ji, Xiaodan Zhang, Ying Liu, Zheng Wang, and Huimin Xu. 2024. Prior tissue knowledge-driven contrastive learning for brain CT report generation. Multim. Syst., 30(2):98.",
+            "Yanzhao Shi, Junzhong Ji, Xiaodan Zhang, Liangqiong Qu, and Ying Liu. 2023. Granularity matters: Pathological graph-driven cross-modal alignment for brain CT report generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP, pages 6617-6630.",
+            "Yuhao Tang, Haichen Yang, Liyan Zhang, and Ye Yuan. 2024. Work like a doctor: Unifying scan localizer and dynamic generator for automated computed tomography report generation. *Expert Systems with Applications*, 237:121442.",
+            "Omkar Thawakar, Abdelrahman M. Shaker, Sahal Shaji Mullappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman H. Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. 2023. Xraygpt: Chest radiographs summarization using medical vision-language models. CoRR, abs/2306.07971.",
+            "Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748.",
+            "Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566-4575.",
+            "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2015. Show and tell: A neural image caption generator. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3156-3164.",
+            "Jun Wang, Abhir Bhalerao, and Yulan He. 2022a. Cross-modal prototype driven network for radiology report generation. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 563-579. Springer.",
+            "Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving text embeddings with large language models. ArXiv, abs/2401.00368."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "16551",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, and Ronald M. Summers. 2018. Tienet: Text-image embedding network for common thorax disease classification and reporting in chest x-rays. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 9049-9058.",
+            "Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. 2022b. Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163.",
+            "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. 2015. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 2048-2057.",
+            "An Yan, Zexue He, Xing Lu, Jiang Du, Eric Y. Chang, Amilcare Gentili, Julian J. McAuley, and Chun-Nan Hsu. 2021. Weakly supervised contrastive learning for chest x-ray report generation. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, pages 4009-4015.",
+            "Shuxin Yang, Xian Wu, Shen Ge, S Kevin Zhou, and Li Xiao. 2022. Knowledge matters: Chest radiology report generation with general and specific knowledge. Medical image analysis, 80:102510.",
+            "Sisi Yang, Junzhong Ji, Xiaodan Zhang, Ying Liu, and Zheng Wang. 2021. Weakly guided hierarchical encoder-decoder network for brain ct report generation. In IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021, Houston, TX, USA, December 9-12, 2021, pages 568-573. IEEE.",
+            "Di You, Fenglin Liu, Shen Ge, Xiaoxia Xie, Jing Zhang, and Xian Wu. 2021. Aligntransformer: Hierarchical alignment of visual regions and disease tags for medical report generation. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part III, pages 72-82.",
+            "Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. 2023. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737.",
+            "Xiaodan Zhang, Sisi Yang, Yanzhao Shi, Junzhong Ji, Ying Liu, Zheng Wang, and Huimin Xu. 2023. Weakly guided attention model with hierarchical interaction for brain ct report generation. Computers in Biology and Medicine, page 107650.",
+            "Yixiao Zhang, Xiaosong Wang, Ziyue Xu, Qihang Yu, Alan L. Yuille, and Daguang Xu. 2020. When radiology report generation meets knowledge graph."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 12910-12917.",
+        "bbox": [
+            527,
+            85,
+            884,
+            175
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "16552",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2024/See Detail Say Clear_ Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning/081b7730-6494-4c1e-8ba3-d9c947e8be2f_model.json

@@ -0,0 +1,2004 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.168,
+                0.086,
+                0.833,
+                0.127
+            ],
+            "angle": 0,
+            "content": "See Detail Say Clear: Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.146,
+                0.137,
+                0.856,
+                0.157
+            ],
+            "angle": 0,
+            "content": "Chengxin Zheng\\(^{1}\\), Junzhong Ji\\(^{1}\\), Yanzhao Shi\\(^{1}\\), Xiaodan Zhang\\(^{1,*}\\), Liangqiong Qu\\(^{2}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.138,
+                0.171,
+                0.859,
+                0.222
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)College of Computer Science, Beijing University of Technology, Beijing, China  \n\\(^{2}\\)Department of Statistics and Actuarial Science, School of Computing and Data Science, The University of Hong Kong, Hong Kong, China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.357,
+                0.224,
+                0.642,
+                0.237
+            ],
+            "angle": 0,
+            "content": "Correspondence: zhangxiaodan@bjut.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.261,
+                0.342,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.286,
+                0.461,
+                0.771
+            ],
+            "angle": 0,
+            "content": "Brain CT report generation is significant to aid physicians in diagnosing cranial diseases. Recent studies concentrate on handling the consistency between visual and textual pathological features to improve the coherence of report. However, there exist some challenges: 1) Redundant visual representing: Massive irrelevant areas in 3D scans distract models from representing salient visual contexts. 2) Shifted semantic representing: Limited medical corpus causes difficulties for models to transfer the learned textual representations to generative layers. This study introduces a Pathological Clue-driven Representation Learning (PCRL) model to build cross-modal representations based on pathological clues and naturally adapt them for accurate report generation. Specifically, we construct pathological clues from perspectives of segmented regions, pathological entities, and report themes, to fully grasp visual pathological patterns and learn cross-modal feature representations. To adapt the representations for the text generation task, we bridge the gap between representation learning and report generation by using a unified large language model (LLM) with task-tailored instructions. These crafted instructions enable the LLM to be flexibly fine-tuned across tasks and smoothly transfer the semantic representation for report generation. Experiments demonstrate that our method outperforms previous methods and achieves SoTA performance. Our code is available at https://github.com/Chauncey-Jheng/PCRL-MRG."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.78,
+                0.262,
+                0.795
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.805,
+                0.491,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Brain computed tomography (CT) imaging is essential for diagnosing various cranial diseases, including cerebral infarction and hemorrhage. However, it is time-consuming and error-prone for radiologists to manually interpret medical findings from these scans and write reports. Automated report"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.516,
+                0.26,
+                0.883,
+                0.497
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.516,
+                0.885,
+                0.674
+            ],
+            "angle": 0,
+            "content": "Figure 1: Comparison of different cross-modal alignment paradigms. (a) Previous way: The origin CT images contain extraneous information unrelated to diagnosis, and separate training for two semantic tasks makes it challenging to find a shared optimal solution, resulting in inadequate cross-modal alignment. (b) Our way: The visual representation is refined to concentrate on pathological clues, and employ joint training with task-tailored instructions via unified LLM to find a transferable representation, leading to better adaption for report generation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.708,
+                0.884,
+                0.756
+            ],
+            "angle": 0,
+            "content": "generation systems are designed to boost efficiency, reduce the workload for radiologists, and optimize resources in busy clinical scenarios."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.761,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "With the advancement of deep neural networks and their successful application in image captioning tasks (Vinyals et al., 2015; Xu et al., 2015), medical report generation (MRG) has gained more attention. Unlike the short sentences of traditional image captioning, MRG aims to generate lengthy and precise reports. To achieve this, various cross-modal alignment methods are required to ensure the consistency between visual and textual information (Shi et al., 2023), including attention mech-"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.14,
+                0.907,
+                0.29,
+                0.921
+            ],
+            "angle": 0,
+            "content": "*Corresponding Author"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.477,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16542"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.21,
+                0.946,
+                0.788,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 16542-16552 November 12-16, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.486,
+                0.132
+            ],
+            "angle": 0,
+            "content": "anisms (Jing et al., 2018; Wang et al., 2018), memory mechanisms (Chen et al., 2020, 2021), and knowledge graphs (Li et al., 2019, 2023b)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.134,
+                0.488,
+                0.213
+            ],
+            "angle": 0,
+            "content": "Recently, learning representations via visual-textual contrastive learning (Shi et al., 2023, 2024) or using pre-trained large language models (LLMs) (Thawakar et al., 2023) to strength representations are also proven to be effective."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.215,
+                0.49,
+                0.632
+            ],
+            "angle": 0,
+            "content": "However, as shown in Figure 1(a), learning cross-modal correspondences is still challenging in current methods due to the following concerns: 1) Redundant visual representing. Different from chest X-ray data, 3D brain CT scans contain extensive redundant information, e.g. background and insignificant areas. With the lack of human-crafted boxes to locate pathology regions, models struggle to capture and interpret the visual pathology patterns for generating reports. Although current advanced models use semantic prior knowledge or medical prompts (Jin et al., 2024; Bu et al., 2024) to automatically learn the salient visual areas, this may introduce noise and unstable representation and cause severe hallucinations in generated texts. 2) Shifted semantic representing. Compared to natural corpus, limited brain CT report corpus is insufficient to transfer the pathological semantic representations learned by represent learning layers (e.g., contrastive learning layer) to the language model (Huh et al., 2024), since the direct weight-sharing (Shi et al., 2023, 2024) is prone to degrade the coherence of generated diagnostic sentences. Thus, how to uniformly represent cross-modal pathological features and adapt them to report generation still remains an open question."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.633,
+                0.489,
+                0.873
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose a Pathological Cluedriven Representation Learning (PCRL) model to seamlessly build cross-modal representations based on diverse pathological clues and transfer them for generating accurate brain CT reports. Specifically, we extract pathological clues from perspectives of segmented regions, pathological entities, and report themes to depict clinical scenarios. Segmented region clues are automatically generated and filtered by given pathology prompts, enabling the visual encoder to grasp visual pathological patterns. Meanwhile, entity and theme clues are respectively extracted by detailed findings and full-text reports, to handle the enriched visual-textual alignment and build cross-modal pathology representations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.874,
+                0.488,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Besides, to adapt the learned representations for the text generation task, we bridge the gap between representation learning and report generation by"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.883,
+                0.26
+            ],
+            "angle": 0,
+            "content": "employing a unified large language model (LLM) with task-tailored instructions, which has proven to be more effective than conventional decoders by using appropriate tokens to seamlessly connect different tasks (Yu et al., 2023). As shown in Figure 1(b), We craft a representation instruction to prompt the LLM to produce high-level pathological semantic features for cross-modal alignment, and a generation instruction to prompt LLM to generate accurate brain CT reports based on the learned representations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.528,
+                0.262,
+                0.875,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Our main contributions can be summarized as:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.525,
+                0.287,
+                0.883,
+                0.368
+            ],
+            "angle": 0,
+            "content": "1. We propose a novel framework to seamlessly learn visual-textual representations from perspectives of diverse pathological clues and leverage them for enhancing the quality of generated brain CT reports."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.525,
+                0.378,
+                0.885,
+                0.457
+            ],
+            "angle": 0,
+            "content": "2. We, for the first time, design a new paradigm to effectively transfer the learned pathological representations for report generation using a unified LLM prompted by task-tailored instructions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.525,
+                0.469,
+                0.884,
+                0.533
+            ],
+            "angle": 0,
+            "content": "3. We validate the model capabilities on the open-source CTRG-Brain dataset. Experimental results show that our model achieves remarkable performance in generating brain CT reports."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.525,
+                0.287,
+                0.885,
+                0.533
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.554,
+                0.665,
+                0.569
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.579,
+                0.771,
+                0.594
+            ],
+            "angle": 0,
+            "content": "2.1 Medical Report Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.6,
+                0.884,
+                0.921
+            ],
+            "angle": 0,
+            "content": "The advancements in image captioning techniques have spurred the development of a range of radiology report generation methods (Jing et al., 2018; Chen et al., 2020, 2021; Li et al., 2019; Shi et al., 2023; Zhang et al., 2023; Li et al., 2023b; Shi et al., 2024; Shen et al., 2024). To exploit the more effective parts of medical images, Jing et al. (2018) first propose a co-attention mechanism to associate images with disease tags and improve the accuracy of generated reports; Chen et al. (2021) design a cross-modal memory matrix to learn high-level vision-language correspondence for enhancing report generation; Liu et al. (2021c) used contrastive attention to captured the difference of abnormal and normal samples. To address the semantic bias of limited medical report corpus, Jing et al. (2020) exploited the textual structure information of reports, while Liu et al. (2021a); Shen et al. (2024) used curriculum learning to alleviate the data bias of limited medical data corpus. The retrieval-based"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.525,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16543"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.139,
+                0.087,
+                0.861,
+                0.304
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.318,
+                0.884,
+                0.378
+            ],
+            "angle": 0,
+            "content": "Figure 2: The overall framework of our method, which mainly consists of an image encoder and a text decoder for brain CT report generation (left). The pathological clue-driven representation learning (right) is proposed to guide the encoder and decoder for more fine-grained representation by three alignment modules: (a) segmentation clue alignment (SCA), (b) entity clue alignment (ECA), and (c) theme clue alignment (TCA)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.401,
+                0.489,
+                0.562
+            ],
+            "angle": 0,
+            "content": "(Li et al., 2018; Liu et al., 2021b) and prior expert knowledge based methods (Zhang et al., 2020; Yang et al., 2022; You et al., 2021) also shows effectiveness to alleviate the inaccuracy of generated report caused by limited training data. To build more stable relations between multi-modal features, contrastive learning (Radford et al., 2021) acts as an unsupervised method, and is proven to be effective for learning cross-modal representations for MRG tasks (Li et al., 2023b; Shi et al., 2024)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.564,
+                0.49,
+                0.789
+            ],
+            "angle": 0,
+            "content": "Recently, LLM-based MRG methods (Jin et al., 2024; Chen et al., 2024) harness the semantic processing capabilities of LLMs to enhance the coherence of generated medical reports. However, the current bottleneck of LLM is to interpret sparse visual patterns of pathologies, which causes severe hallucinations and degrades the trustworthiness of clinical diagnosis. To solve this issue, we propose to augment LLM with representation learning guided by pathology clues. By aligning cross-modal pathological features and transferring the learned representations to the LLM-based report decoder, our model can generate accurate and coherent brain CT reports."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.803,
+                0.358,
+                0.818
+            ],
+            "angle": 0,
+            "content": "2.2 Pre-trained Large Model"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.826,
+                0.489,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In recent years, pre-trained large models have achieved significant breakthroughs in both natural language processing (NLP) and computer vision (CV). These models leverage extensive datasets for pre-training, allowing them to perform exceptionally well in downstream tasks. The Segment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.401,
+                0.885,
+                0.756
+            ],
+            "angle": 0,
+            "content": "Anything Model (SAM) Kirillov et al. (2023) is an innovative large vision model in the field of image segmentation. SAM's strengths lie in its extensive training on a large-scale dataset, enhancing its versatility and enabling real-time generation of high-quality segmentation masks across diverse tasks and image distributions. MedCLIP Wang et al. (2022b) is a notable visual-language model in the medical domain. Pre-trained on decoupling medical images and texts, MedCLIP has shown SoTA performance on zero-shot image-text retrieval. LLaMA3 is the next generation of state-of-the-art open-source LLM from Meta (2024). Trained on over 15 trillion tokens, seven times the data of its predecessor, LLaMA3 incorporates extensive multilingual and high-quality datasets. To improve the use of large models for assisting in medical report generation, we adopt SAM for region segmentation, MedCLIP for feature representation of pathological entities, and LLaMA for joint training to achieve fine-grained cross-modal alignment and improve medical report generation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.777,
+                0.659,
+                0.794
+            ],
+            "angle": 0,
+            "content": "3 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.809,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "As illustrated in Figure 2, our framework consists of two branches: the brain CT Report Generation (RG) on the left and the Pathological Clue-driven Representation Learning (PCRL) on the right. The interaction between these branches is facilitated through the shared use of a visual encoder and a language model."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16544"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.085,
+                0.391,
+                0.1
+            ],
+            "angle": 0,
+            "content": "3.1 Brain CT Report Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.106,
+                0.49,
+                0.429
+            ],
+            "angle": 0,
+            "content": "In this branch, the input comprises a set of CT scan images \\( I = \\{i_1,\\dots ,i_N\\} \\), where \\( N \\) represents the number of CT images in each sample. The output is the corresponding brain CT findings report \\( Y = \\{y_{1},\\ldots ,y_{M}\\} \\), with \\( M \\) representing the number of tokens. We adopt the encoder-decoder architecture for report generation. First, we use ResNet101 (He et al., 2016) to extract grid features \\( G = \\{g_{1},\\dots ,g_{N}\\} \\in \\mathbf{R}^{N\\times H\\times d}(N = 24,H = 196,d = 2048) \\) from \\( I \\). Then, we embed these features using a visual encoder to obtain visual features \\( V \\). These visual features \\( V \\) are passed through a multi-modal projector to obtain the visual embedding tokens \\( H_{v}\\in \\mathbf{R}^{N\\times d_{w}}(d_{w} = 4096) \\) which are aligned with the word embedding space of a large language model. Finally, these tokens are concatenated with the instruction token \\( H_{q} \\) and fed into the large language model. We train the parameters \\( \\theta \\) by minimizing the cross-entropy loss, which can be represented by the following formula:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.137,
+                0.439,
+                0.488,
+                0.482
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {g} = - \\sum_ {t = 1} ^ {M} \\log P \\left(y _ {t} \\mid y _ {1: t - 1}, H _ {v}, H _ {q}; \\theta\\right) \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.493,
+                0.489,
+                0.543
+            ],
+            "angle": 0,
+            "content": "where, \\( P(y_{t}|*) \\) denotes the probability conditioned on \\( H_{v}, H_{q} \\), and the embeddings of previous words \\( y_{1}, y_{2}, \\dots, y_{t-1} \\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.552,
+                0.48,
+                0.584
+            ],
+            "angle": 0,
+            "content": "3.2 Pathological Clue-driven Representation Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.589,
+                0.49,
+                0.702
+            ],
+            "angle": 0,
+            "content": "Learning fine-grained visual-text representations is critical for generating accurate reports (Shi et al., 2023). This branch mainly builds enriched representations by conducting multi-modal feature alignment based on pathological clues, including segmentation clue alignment (SCA), entity clues alignment (ECA), and theme clues alignment (TCA)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.71,
+                0.44,
+                0.726
+            ],
+            "angle": 0,
+            "content": "3.2.1 Preparation of Pathological Clues"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.729,
+                0.49,
+                0.761
+            ],
+            "angle": 0,
+            "content": "Pathological clues are collected in three perspectives for feature alignment."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.761,
+                0.49,
+                0.889
+            ],
+            "angle": 0,
+            "content": "**Pathology Theme Clues:** Learning the structure of professional brain CT reports is essential for MRG models to satisfy human standards and produce reliable reports. To this end, we propose to build theme clues by simply using the full-text report and whole images as global signals, which can be useful to enhance the overall quality of medical reports by TCA (illustrated in 3.2.4)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.89,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Pathology Entity Clues: To learn detailed information about pathology entities, we regard each"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.181
+            ],
+            "angle": 0,
+            "content": "single finding sentence in the report as an entity clue and extract the related multi-modal features in ECA 3.2.3. The construction of entities \\(E = \\{e_1,\\dots,e_{N_e}\\}\\) (with \\(N_{e} = 24\\)) is based on expert knowledge and the frequency of words in the training corpus."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.183,
+                0.885,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Segment Clues: To enhance visual representations via detailed contour of pathologies, we propose to utilize the pre-trained segmentation model SAM (Kirillov et al., 2023) to generate mask candidates and filter useful masks related to pathology entities. First, we prompt SAM by covering each brain CT image with a grid of points and integrating it into image embeddings through average sampling. In this way, the mask decoder in SAM is prompted to generate a gallery of candidate masks \\( M_I = \\{m_1, \\dots, m_{N_m}\\} \\). To ensure the quality of masks, we also apply rule-based methods to filter out low-quality and duplicate masks with area size, stability scores, and IoU scores. For each sample consisting of 24 images, we generate a corresponding series of masks and combine them to form the candidate Gallery \\( = \\{M_{I_1}, \\dots, M_{I_N}\\} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.457,
+                0.884,
+                0.713
+            ],
+            "angle": 0,
+            "content": "Then, to retrieve valuable masks related to the sample's pathological entities, we propose to use the MedCLIP (Wang et al., 2022b) for text-prompted retrieval. Based on expert knowledge, we divide the 24 images into eight-layer categories, each mapping a specific set of entities (Shi et al., 2024). Conversely, each entity also has its corresponding layers. We extract the existing entities from the sample's report and their corresponding pathological descriptions \\( D = \\{d_{1},\\dots,d_{N_{d}}\\} \\) (\\( N_{d}\\leq N_{e} \\)), which serve as the Query. Next, we use the \\( d \\)-th description \\( Query_{d} \\) to search for the most similar matching mask in \\( Gallery_{d} \\), which is a subset of \\( Gallery \\), consisting of all visual masks of the layers related to the existing entities. The procedure can be represented as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.527,
+                0.73,
+                0.883,
+                0.746
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {R e t r i e v a l} _ {d} = \\operatorname {C L I P} \\left(\\operatorname {G a l l e r y} _ {d}, \\operatorname {Q u e r y} _ {d}\\right) \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.761,
+                0.883,
+                0.792
+            ],
+            "angle": 0,
+            "content": "where \\(Retrieval_{d}\\) denotes the retrieved masks for SCA 3.2.2."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.805,
+                0.807,
+                0.821
+            ],
+            "angle": 0,
+            "content": "3.2.2 Segmentation Clue Alignment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.826,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "This module aims to learn the fine-grained visual representation based on the extracted segment masks. First, we obtain the visual features \\( V_{D} = \\{v_{1},\\dots v_{N_{d}}\\} \\) corresponding to the images containing the entities using a selector. Then, we input these features into a lightweight segmentation"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16545"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.278
+            ],
+            "angle": 0,
+            "content": "head to generate the corresponding foreground entity masks \\( S_{D} = \\{s_{1},\\dots,s_{N_{d}}\\} \\), which serves as discriminated pathological information. We then align these generated masks with the retrieved segment masks \\( M_{D} = \\{Retrieval_{1},\\dots,Retrival_{N_{d}}\\} \\) to learn detailed patterns. It is important to note that the size of the generated masks \\( S_{D} \\) differs from the size of the retrieved SAM masks \\( M_{D} \\). To address this, we first resize the SAM masks to match the size of the masks generated by our segmentation head. We finally calculate the following loss for aligning the two types of masks:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.176,
+                0.282,
+                0.488,
+                0.318
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {s e g} = 1 - \\frac {2 \\sum_ {i = 1} ^ {n} p _ {i} y _ {i}}{\\sum_ {i = 1} ^ {n} p _ {i} ^ {2} + \\sum_ {i = 1} ^ {n} y _ {i} ^ {2}} \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.322,
+                0.487,
+                0.353
+            ],
+            "angle": 0,
+            "content": "where \\( y_{i} \\) is the pixel value of retrieved SAM masks and \\( p_i \\) is the predicted pixel value."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.354,
+                0.489,
+                0.402
+            ],
+            "angle": 0,
+            "content": "In this way, the visual encoder can be effectively learned to focus on the areas of pathologies while reducing the influence of irrelevant visual features."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.41,
+                0.353,
+                0.426
+            ],
+            "angle": 0,
+            "content": "3.2.3 Entity Clue Alignment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.429,
+                0.49,
+                0.573
+            ],
+            "angle": 0,
+            "content": "To learn the cross-modal patterns of pathology entities, we extract visual and textual pathological features based on the entity clues. First, we build a selector to obtain the visual features \\( V_{D} = \\{v_{1},\\dots v_{N_{d}}\\} (N_{d} < N) \\), which corresponds to significant CT images that exist pathology entities. We then apply global average pooling (GAP) to generalize \\( V_{D} \\) and obtain representations of each significant CT image, denoted as \\( R_v^e = \\{r_1,\\dots r_{N_d}\\} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.574,
+                0.49,
+                0.797
+            ],
+            "angle": 0,
+            "content": "Different from previous work (Li et al., 2023b) use an external language model (e.g. SciBert (Beltagy et al., 2019)) to build textual representation for feature alignment, we propose to leverage a unified LLM to generate textual representation via tailored prompts. Inspired by Wang et al. (2023), we crafted a representation instruction \"Summarize the following cranial diagnosis in one word,\" which is denoted as \\( H_{q_r} \\). This prompt can effectively activate the summarization ability of LLM to generate corresponding text representations for pathological entity description in \\( D \\). We use the output generated by the final hidden layer of LLM as textual features for cross-modal alignment."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.799,
+                0.489,
+                0.877
+            ],
+            "angle": 0,
+            "content": "With the carefully extracted visual and textual representations of entities, we map them into the same dimension through an embedding layer. The process can be represented by the following formula:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.17,
+                0.886,
+                0.488,
+                0.902
+            ],
+            "angle": 0,
+            "content": "\\[\nR _ {v} ^ {e} = \\operatorname {L i n e a r} _ {v} ^ {e} (G A P \\left(V _ {D}\\right)) \\tag {4}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.168,
+                0.906,
+                0.487,
+                0.923
+            ],
+            "angle": 0,
+            "content": "\\[\nR _ {w} ^ {e} = \\operatorname {L i n e a r} _ {t} ^ {e} \\left(L L M \\left(H _ {q _ {r}}, D\\right)\\right) \\tag {5}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.166
+            ],
+            "angle": 0,
+            "content": "where \\( GAP \\) denotes the global average pooling, \\( Linear_v \\) represents the visual mapper, and \\( Linear_t \\) represents the textual mapper. We employ the symmetric InfoNCE (van den Oord et al., 2018) loss for visual-textual alignment:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.51,
+                0.184,
+                0.882,
+                0.289
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {C L _ {e}} = - \\frac {1}{2} \\left(\\alpha_ {v} \\sum_ {i = 1} ^ {N _ {d}} \\log \\frac {\\exp \\left(\\mathrm {s} _ {v} ^ {e} (i , i)\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {d}} \\exp \\left(\\mathrm {s} _ {v} ^ {e} (i , j)\\right)} \\right. \\\\ + \\alpha_ {w} \\sum_ {i = 1} ^ {N _ {d}} \\log \\frac {\\exp \\left(\\mathrm {s} _ {w} ^ {e} (i , i)\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {d}} \\exp \\left(\\mathrm {s} _ {w} ^ {e} (i , j)\\right)} \\Bigg) \\tag {6} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.295,
+                0.884,
+                0.391
+            ],
+            "angle": 0,
+            "content": "where \\( s_v^e (i,j) = \\frac{\\sin(R_{v_i}^e,R_{w_j}^e)}{\\tau} \\) and \\( s_w^e (i,j) = \\frac{\\sin(R_{w_i}^e,R_{v_j}^e)}{\\tau} \\) denote the similarity between the visual representation \\( R_{v}^{e} \\) and the textual representation \\( R_{w}^{e} \\), \\( \\tau \\) is a temperature parameter. \\( \\alpha_{v} \\) and \\( \\alpha_{w} \\) are hyperparameters to balance the contrastive learning."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.393,
+                0.882,
+                0.44
+            ],
+            "angle": 0,
+            "content": "By aligning multi-modal entity clues, the model can grasp fine-grained visual-text representations to generate accurate diagnostic words."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.449,
+                0.753,
+                0.464
+            ],
+            "angle": 0,
+            "content": "3.2.4 Theme Clue Alignment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.467,
+                0.884,
+                0.627
+            ],
+            "angle": 0,
+            "content": "Theme clue alignment aims to equip the model with comprehensive skills to generate accurate style and structure of reports, thereby enhancing clinical reliability. For global visual features, we use one-dimensional global max pooling (GMP) to represent the entire sample visually, denoted as \\( R_{w}^{t} \\). For the overall report of the sample, we generate textual representation \\( R_{w}^{t} \\) by prompting LLM with the same representation instruction (see Section 3.2.3). This process can be represented as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.567,
+                0.633,
+                0.882,
+                0.651
+            ],
+            "angle": 0,
+            "content": "\\[\nR _ {v} ^ {t} = \\operatorname {L i n e a r} _ {v} ^ {t} (G M P (V)) \\tag {7}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.566,
+                0.653,
+                0.882,
+                0.671
+            ],
+            "angle": 0,
+            "content": "\\[\nR _ {w} ^ {t} = \\operatorname {L i n e a r} _ {t} ^ {t} \\left(L L M \\left(H _ {q _ {s}}, Y\\right)\\right) \\tag {8}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.677,
+                0.882,
+                0.706
+            ],
+            "angle": 0,
+            "content": "Similar to ECA, the loss of TCA can be formulated as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.51,
+                0.712,
+                0.882,
+                0.817
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {C L _ {t}} = - \\frac {1}{2} \\left(\\alpha_ {v} \\sum_ {i = 1} ^ {N _ {b}} \\log \\frac {\\exp (\\mathrm {s} _ {v} ^ {t} (i , i))}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {b}} \\exp (\\mathrm {s} _ {v} ^ {t} (i , j))}\\right) \\\\ + \\alpha_ {w} \\sum_ {i = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\mathrm {s} _ {w} ^ {t} (i , i)\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {N _ {b}} \\exp \\left(\\mathrm {s} _ {w} ^ {t} (i , j)\\right)} \\tag {9} \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "where \\( s_v^t (i,j) = \\sin (R_{v_i}^t,R_{w_j}^t) / \\tau \\) and \\( s_w^t (i,j) = \\) \\( \\sin (R_{w_i}^t,R_{v_j}^t) / \\tau \\) denote the similarity between the visual representation \\( R_{v}^{e} \\) and the textual representation \\( R_w^e \\), \\( \\tau \\) is a temperature parameter, \\( N_{b} \\) is the batch size. Shared with ECA, \\( \\alpha_{v} \\) and \\( \\alpha_{w} \\) are set to balance the feature alignment."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16546"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.129,
+                0.083,
+                0.88,
+                0.308
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Methods</td><td>Decoder</td><td>B1</td><td>B2</td><td>B3</td><td>B4</td><td>M</td><td>RG</td><td>C</td><td>F1</td></tr><tr><td>HRNN(Krause et al., 2017)†</td><td>LSTM</td><td>42.3</td><td>28.3</td><td>21.2</td><td>17.1</td><td>27.2</td><td>39.3</td><td>20.9</td><td>70.2</td></tr><tr><td>Up-Down(Anderson et al., 2018)†</td><td>LSTM</td><td>45.8</td><td>34.8</td><td>28.5</td><td>24.4</td><td>31.6</td><td>42.5</td><td>27.3</td><td>70.2</td></tr><tr><td>WCL(Yan et al., 2021)†</td><td>LSTM</td><td>49.5</td><td>36.5</td><td>29.4</td><td>25.1</td><td>31.3</td><td>42.8</td><td>33.3</td><td>64.5</td></tr><tr><td>R2Gen-CMN(Chen et al., 2021)†</td><td>Transformer</td><td>49.1</td><td>40.0</td><td>34.4</td><td>30.1</td><td>29.9</td><td>48.6</td><td>84.2</td><td>69.8</td></tr><tr><td>XProNet(Wang et al., 2022a)†</td><td>Transformer</td><td>50.6</td><td>41.3</td><td>34.4</td><td>29.1</td><td>31.5</td><td>51.7</td><td>83.3</td><td>70.1</td></tr><tr><td>WGAM(Yang et al., 2021)†</td><td>LSTM</td><td>49.4</td><td>36.7</td><td>29.6</td><td>25.4</td><td>32.0</td><td>42.4</td><td>31.9</td><td>68.9</td></tr><tr><td>PGCA(Shi et al., 2023)†</td><td>LSTM</td><td>50.2</td><td>37.8</td><td>30.7</td><td>26.5</td><td>32.5</td><td>43.0</td><td>34.0</td><td>69.2</td></tr><tr><td>WGAM-HI(Zhang et al., 2023)†</td><td>LSTM</td><td>50.4</td><td>37.6</td><td>30.5</td><td>26.1</td><td>31.4</td><td>43.8</td><td>33.2</td><td>67.4</td></tr><tr><td>LLaVA-med(Li et al., 2023a)†</td><td>LLaMA3-8B</td><td>50.0</td><td>39.3</td><td>32.0</td><td>26.3</td><td>31.4</td><td>46.7</td><td>38.6</td><td>64.3</td></tr><tr><td>HILT(Liu et al., 2024)†</td><td>LLaMA3-8B</td><td>50.7</td><td>39.9</td><td>33.1</td><td>27.9</td><td>30.8</td><td>46.1</td><td>43.7</td><td>68.4</td></tr><tr><td>PromptMRG(Jin et al., 2024)†</td><td>LLaMA3-8B</td><td>48.1</td><td>38.3</td><td>31.6</td><td>26.5</td><td>31.0</td><td>47.4</td><td>50.3</td><td>69.2</td></tr><tr><td>Ours (Jieba tokenizer)</td><td>LLaMA3-8B</td><td>51.5</td><td>42.0</td><td>35.7</td><td>30.9</td><td>31.4</td><td>49.0</td><td>80.0</td><td>70.6</td></tr><tr><td>Ours (LLaMA tokenizer)</td><td>LLaMA3-8B</td><td>62.0</td><td>54.7</td><td>49.4</td><td>45.3</td><td>33.1</td><td>57.7</td><td>96.4</td><td>70.6</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.324,
+                0.882,
+                0.355
+            ],
+            "angle": 0,
+            "content": "Table 1: The performance of our PCRL compared with previous state-of-the-art models on the Brain CT report generation dataset CTRG-Brain. The best results are highlighted in bold. \\(\\dagger\\) denotes the re-implementation results."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.379,
+                0.274,
+                0.394
+            ],
+            "angle": 0,
+            "content": "3.3 Joint Training"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.401,
+                0.49,
+                0.594
+            ],
+            "angle": 0,
+            "content": "In the training stage, we jointly train the RG branch and the PCRL branch to maximize the utilization of multi-granularity visual-text representations. Instead of using separate modules to learn representation and generate medical reports (Li et al., 2023b), we use a unified LLM to bridge the gap between representation learning and report generation via two task-tailored instructions, i.e., representation instruction and generation instruction. This can transfer the representations learned by the PCRL branch to optimize the RG branch effectively, thereby generating accurate reports."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.596,
+                0.489,
+                0.627
+            ],
+            "angle": 0,
+            "content": "Our final loss contains the above-mentioned losses, which can be formulated as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.178,
+                0.644,
+                0.488,
+                0.661
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {g} + \\mathcal {L} _ {\\text {s e g}} + \\mathcal {L} _ {C L _ {e}} + \\mathcal {L} _ {C L _ {t}} \\tag {10}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.678,
+                0.26,
+                0.695
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.706,
+                0.221,
+                0.72
+            ],
+            "angle": 0,
+            "content": "4.1 Dataset"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.729,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We validate the performance of our model using the CTRG-Brain (Tang et al., 2024) dataset. This dataset comprises 6,000 samples, containing a total of 160,336 CT images and 6,000 Chinese medical reports. Following the mainstream division method (Shi et al., 2023; Zhang et al., 2023), we split the dataset into a training set, a validation set, and a test set in a 7:1:2 ratio. For consistent processing, we divide the brain CT image samples into 8 layers based on expert knowledge, with each layer containing 3 continuous CT images, assigning each sample with 24 CT images."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.379,
+                0.707,
+                0.393
+            ],
+            "angle": 0,
+            "content": "4.2 Evaluation Metrics"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.406,
+                0.885,
+                0.599
+            ],
+            "angle": 0,
+            "content": "We chose Natural Language Generation (NLG) metrics and Clinical Evaluation (CE) metrics to evaluate our model's performance. NLG metrics include BLEU (Papineni et al., 2002),METEOR (Banerjee and Lavie), ROUGE-L (Lin, 2004), and CIDEr (Vedantam et al., 2015), denoted as B1, B2, B3,B4,M,RG,and C respectively. To measure the pathological accuracy, we use 24 keywords summarized by experienced radiologists to calculate the Clinical Evaluation (CE) metric(Shi et al., 2023; Zhang et al., 2023), i.e., F1 score, which is the harmonic mean of the precision and recall."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.621,
+                0.742,
+                0.637
+            ],
+            "angle": 0,
+            "content": "4.3 Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.648,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We reshape the size of image to \\(512 \\times 512\\) pixels and used a ResNet101 to extract image features, which is pre-trained on the ImageNet dataset and fine-tuned on the CQ500 dataset (Chilamkurthy et al., 2018). For our large language model, we utilize LLaMA3-8B (Meta, 2024), which is quantized to 4-bit, and use the LoRa (Hu et al., 2021) for parameter-efficient fine-tuning. The overall trainable parameter quantity of our model is \\(229.9\\mathrm{M}\\) with \\(3.4\\mathrm{M}\\) parameters in LLM (only \\(0.04\\%\\)). During training, we use the AdamW optimizer with a learning rate of 1e-4. The batch size is 4, with 1050 training steps per epoch. For testing, we set the temperature coefficient of the large model to 0.6 and the top-p value to 0.9. The model is implemented using PyTorch 2.3.0, and the entire training process is conducted on a single RTX 4090 GPU."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16547"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.128,
+                0.083,
+                0.878,
+                0.228
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">Representation</td><td colspan=\"3\">Module Loss</td><td rowspan=\"2\">B1</td><td rowspan=\"2\">B2</td><td rowspan=\"2\">B3</td><td rowspan=\"2\">B4</td><td rowspan=\"2\">M</td><td rowspan=\"2\">RG</td><td rowspan=\"2\">C</td></tr><tr><td>Visual</td><td>Textual</td><td>LSCA</td><td>LECA</td><td>LTCA</td></tr><tr><td>Baseline</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>49.6</td><td>39.9</td><td>33.4</td><td>28.6</td><td>30.7</td><td>48.4</td><td>50.8</td></tr><tr><td>(a)</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>51.3</td><td>41.0</td><td>34.5</td><td>29.7</td><td>31.2</td><td>47.7</td><td>70.8</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>51.3</td><td>41.4</td><td>35.0</td><td>30.1</td><td>31.2</td><td>48.7</td><td>71.3</td></tr><tr><td>(c)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>51.7</td><td>41.4</td><td>34.6</td><td>29.5</td><td>31.2</td><td>48.0</td><td>72.8</td></tr><tr><td>(d)</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>48.9</td><td>38.8</td><td>32.2</td><td>27.4</td><td>30.0</td><td>46.3</td><td>58.3</td></tr><tr><td>(e)</td><td>✓</td><td>✓(Bert)</td><td>✓</td><td>✓</td><td>✓</td><td>49.5</td><td>39.0</td><td>31.8</td><td>26.3</td><td>30.4</td><td>45.4</td><td>48.5</td></tr><tr><td>Ours</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>51.5</td><td>42.0</td><td>35.7</td><td>30.9</td><td>31.4</td><td>49.0</td><td>80.0</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.237,
+                0.885,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Table 2: Ablation studies of our proposed method. The Baseline model is an encoder-decoder framework without an alignment mechanism. (a), (b), (c), (d) and (e) respectively denote the use of different representations (i.e., visual and textual in ECA and TCA) and module losses."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.306,
+                0.332,
+                0.323
+            ],
+            "angle": 0,
+            "content": "4.4 Quantitative Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.334,
+                0.49,
+                0.576
+            ],
+            "angle": 0,
+            "content": "We compare the proposed PCRL with some competitive brain CT report generation methods (WGAM (Yang et al., 2021), PGCA (Shi et al., 2023), WGAM-HI(Zhang et al., 2023)). Besides, we also reproduced some SOTA models in image captioning (HRNN (Krause et al., 2017), Up-Down (Anderson et al., 2018)) and chest Xray report generation (WCL (Yan et al., 2021), R2Gen-CMN (Chen et al., 2021), XProNet (Wang et al., 2022a)) for comprehensive comparisons on CTRG-Brain dataset. What's more, for fair comparisons, We also reproduce several related LLM-based methods (LLaVA-Med(Li et al., 2023a), HILT(Liu et al., 2024), PromptMRG(Jin et al., 2024)with the same LLM decoder."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.58,
+                0.49,
+                0.87
+            ],
+            "angle": 0,
+            "content": "As shown in Table 1, our method outperforms others across most evaluation metrics. WGAM and WGAM-HI employ weakly-supervised visual attention to extract key visual features, resulting in higher BLEU scores. With contrastive learning for cross-modal alignment, WCL and PGCA can effectively learn relations between CT images and reports, achieving better results. However, due to the lack of training data, the above methods still produce reports that fall short in fluency and readability. Here's a refined version of your sentence: The LLM-based methods (LLaVA-med, HILT, PromptMRG) demonstrate poorer performance compared to traditional transformer-based methods (R2Gen-CMN, XProNet). This suggests that prior knowledge from pretrained large models, without further cross-modal alignment, may lead to hallucinations in report generation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.874,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Our PCRL achieves fine-grained cross-modal alignment by leveraging a series of pre-trained large models, generating more fluent reports and"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.306,
+                0.885,
+                0.401
+            ],
+            "angle": 0,
+            "content": "achieves the best performance compared to other methods. It is noteworthy that we tested our method using the Jieba and LLaMA tokenizers respectively to compute NLG scores, with the latter achieving the best performance across all NLG metrics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.402,
+                0.887,
+                0.514
+            ],
+            "angle": 0,
+            "content": "This discrepancy may be due to differences in tokenization methods used during training. While other models employ the traditional Jieba tool for Chinese word tokenization, our PCRL follows more advanced BPE-based subword tokenization. Nevertheless, our model also achieves competitive results in overall metrics."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.526,
+                0.676,
+                0.541
+            ],
+            "angle": 0,
+            "content": "4.5 Ablation Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.546,
+                0.885,
+                0.722
+            ],
+            "angle": 0,
+            "content": "To evaluate the effect of each component in PCRL, we have done plenty of ablation studies, as shown in Table 2. Baseline is the standard encoder-decoder (ResNet101-LLaMA3) architecture without alignment. By progressively adding visual or textual representation (visual and textual) and the module losses \\((\\mathcal{L}_{SCA},\\mathcal{L}_{ECA}\\) and \\(\\mathcal{L}_{TCA})\\) , respectively denoting the incorporation of two modalities of representations and the utilization of three pathological clue-driven alignments (i.e., SCA, ECA, and TCA)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.724,
+                0.887,
+                0.901
+            ],
+            "angle": 0,
+            "content": "By comparing (a) with the baseline, we observe that incorporating the segmentation alignment significantly enhances the model's performance. This demonstrates that our SCA method effectively filters out irrelevant information from CT images and extracts crucial pathology-related visual regions. In contrast, with the implementation of ECA, (b) achieves comparable or better performance across all metrics compared to (a). This indicates that fine-grained cross-modal pathological entity alignment can effectively learn strongly correlated visual-text"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.536,
+                0.907,
+                0.722,
+                0.922
+            ],
+            "angle": 0,
+            "content": "https://github.com/fxsjy/jieba"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16548"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.118,
+                0.082,
+                0.885,
+                0.293
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Ground Truth Sample</td><td>Ground Truth Entities</td><td>Retrieved Masks</td><td>Generated Reports</td><td></td></tr><tr><td>...</td><td>基底节区(basal ganglia):左侧基底节区见小片状低密度影 (Small patchy low-density shadows visible in the left basal ganglia area)</td><td></td><td>双侧大脑半球对称,两侧基底节区见斑点状低密度影,中线结构居中,脑沟、脑池、蛛网膜下腔稍增宽,脑回变小。(Bilateral cerebral hemispheres are symmetrical, with punctate low-density shadows in the basal ganglia on both sides. The midline structure is centered, and the sulcus, cistern, and subarachnoid space are slightly widened, while the gyrus is smaller.)</td><td rowspan=\"3\">Baseline</td></tr><tr><td rowspan=\"3\">双侧大脑半球对称,左侧基底节区见小片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑,脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A small patch of low-density shadow is seen in the left basal ganglia region, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal.)</td><td>脑室(ventricle):各脑室、脑池大小形态正常(The size and shape of all ventricles and cisterns are normal.)</td><td rowspan=\"3\"></td><td>双侧大脑半球对称,右侧丘脑见小片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑、脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A small patch of low-density shadow is seen in the right thalamus, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal. The midline structure is centered. There are no abnormalities in the cerebellum and brainstem below the tentorium.)</td></tr><tr><td>中线(midline structure):中线结构居中 (The midline structure centered)</td><td rowspan=\"2\">双侧大脑半球对称,左侧基底节区见斑片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑、脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A patch of low-density shadow is seen in the left basal ganglia region, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal. The midline structure is centered. There are no abnormalities in the cerebellum and brainstem below the tentorium.)</td></tr><tr><td>脑干(brainstem):幕下小脑,脑干无异常(There are no abnormalities in the cerebellum and brainstem below the tentorium)</td><td>Ours</td></tr></table>"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.302,
+                0.885,
+                0.375
+            ],
+            "angle": 0,
+            "content": "Figure 3: Visualization of report generation and mask segmentation. Given the ground truth sample and corresponding entities, the retrieved entity masks are listed in the third column. Reports generated by Baseline, WGAM-HI, and our model are listed in the fourth column. Different colors denote the specific entity words and entity masks, respectively. The English translation is given for a better understanding of the original Chinese reports in CTRG-Brain."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.399,
+                0.489,
+                0.544
+            ],
+            "angle": 0,
+            "content": "representations, thus better supporting medical report generation. Besides, (c) combines both SCA and ECA and shows improved performance in the B1 and C metrics, which are highly correlated with keyword frequency. However, it slightly underperforms in the B2, B4 and RG metrics compared to (b), which is more related to overall text style.(d) showed SCA is essential for focusing on entity details and maintaining the thematic style."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.546,
+                0.489,
+                0.658
+            ],
+            "angle": 0,
+            "content": "Notably, to validate the contribution of joint representation learning, we replace the language model with Bert model for textual representation during alignment in (e). The result indicated that shared representation through joint learning with unified language model can effectively improve the overall quality of the task."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.673,
+                0.32,
+                0.689
+            ],
+            "angle": 0,
+            "content": "4.6 Qualitative Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.697,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We visualize the brain CT reports generated by baseline, WGAM-HI (Zhang et al., 2023) and our model in Figure 3. Given the ground truth brain CT sample and entity data, our model generates better brain CT report with the most accurate entity words (e.g. basal ganglia, ventricle, and brainstem) among the competitors, which demonstrates the effectiveness of using diverse medical clues to learn enriched multi-modal representations. It can be also seen that compared with the baseline, the report generated by our model has a better semantic structure, indicating the contribution of employing a unified LLM to transfer useful representations. Besides, we also find that the retrieved entity masks"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.524,
+                0.4,
+                0.87,
+                0.493
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.505,
+                0.883,
+                0.536
+            ],
+            "angle": 0,
+            "content": "Figure 4: Visualization of the segmentation masks gallery generated by SAM for a single sample."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.561,
+                0.885,
+                0.69
+            ],
+            "angle": 0,
+            "content": "from segmentation masks gallery (see Figure 4) can generally match related entity words at both levels of visual and semantic. For example, the mask of \"midline structure\" and \"brainstem\" matches the empirical scan slice and fine-grained region chosen by experienced radiologists. This guarantees the model to mine accurate visual cranial patterns, therefore generating high-quality reports."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.702,
+                0.642,
+                0.717
+            ],
+            "angle": 0,
+            "content": "5 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.729,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "We propose a novel model to mine pathological clues for enhancing multi-modal representations and seamlessly transfer them into report generation. First, through carefully designed segmentation clue alignment, entity clue alignment, and theme clue alignment, the diverse and precise feature representation can be well-constructed. Second, we transfer the learned representation to boost the brain CT report generation via a unified LLM prompted by task-tailored instructions. Experiments demonstrate the effectiveness of our model in generating pathologically accurate reports."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16549"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.221,
+                0.099
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.11,
+                0.493,
+                0.368
+            ],
+            "angle": 0,
+            "content": "Although the segmentation clues retrieved by MedClip (Wang et al., 2022b) can generally match corresponding pathological entities and help the model neglect redundant visual information, it should be noted that a part of retrieved entity masks may not be precise. This is because MedClip is mainly pretrained by chest X-ray data with limited brain CT samples. Thus, addressing this challenge is imperative for the research community. In the future, we will work on exploring useful approaches. One potential approach is to train a unified text-prompted medical segmentation model towards 3D brain CT scans, which can not only be employed to offer fine-grained visual information for medical report generation but also for other related tasks, e.g. medical VQA."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.395,
+                0.214,
+                0.41
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.418,
+                0.49,
+                0.512
+            ],
+            "angle": 0,
+            "content": "Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.521,
+                0.489,
+                0.576
+            ],
+            "angle": 0,
+            "content": "S Banerjee and A Lavie. Meteor: an automatic metric for mt evaluation with high levels of correlation with human judgments. Proceedings of ACL-WMT, pages 65-72."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.585,
+                0.49,
+                0.678
+            ],
+            "angle": 0,
+            "content": "Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. Scibert: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3613-3618."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.688,
+                0.49,
+                0.768
+            ],
+            "angle": 0,
+            "content": "Shenshen Bu, Yujie Song, Taiji Li, and Zhiming Dai. 2024. Dynamic knowledge prompt for chest x-ray report generation. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, LREC/COLING, pages 5425-5436. ELRA and ICCL."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.778,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Zhihong Chen, Yaling Shen, Yan Song, and Xiang Wan. 2021. Cross-modal memory networks for radiology report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 5904-5914."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.894,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Zhihong Chen, Yan Song, Tsung-Hui Chang, and Xi-ang Wan. 2020. Generating radiology reports via"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.418,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.885,
+                0.14
+            ],
+            "angle": 0,
+            "content": "memory-driven transformer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 1439-1449."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.153,
+                0.885,
+                0.193
+            ],
+            "angle": 0,
+            "content": "Zhixuan Chen, Luyang Luo, Yequan Bie, and Hao Chen. 2024. Dia-llama: Towards large language model-driven CT report generation. CoRR, abs/2403.16386."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.206,
+                0.885,
+                0.297
+            ],
+            "angle": 0,
+            "content": "Sasank Chilamkurthy, Rohit Ghosh, Swetha Tanamala, Mustafa Biviji, Norbert G. Campeau, Vasantha Kumar Venugopal, Vidur Mahajan, Pooja Rao, and Prashant Warier. 2018. Development and validation of deep learning algorithms for detection of critical findings in head ct scans. arXiv preprint arXiv:1803.05854."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.311,
+                0.885,
+                0.378
+            ],
+            "angle": 0,
+            "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.39,
+                0.885,
+                0.456
+            ],
+            "angle": 0,
+            "content": "Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.47,
+                0.885,
+                0.511
+            ],
+            "angle": 0,
+            "content": "Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. 2024. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.524,
+                0.885,
+                0.578
+            ],
+            "angle": 0,
+            "content": "Haibo Jin, Haoxuan Che, Yi Lin, and Hao Chen. 2024. Promptmrg: Diagnosis-driven prompts for medical report generation. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pages 2607-2615."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.59,
+                0.885,
+                0.643
+            ],
+            "angle": 0,
+            "content": "Baoyu Jing, Zeya Wang, and Eric Xing. 2020. Show, describe and conclude: On exploiting the structure information of chest x-ray reports. arXiv preprint arXiv:2004.12274."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.656,
+                0.885,
+                0.736
+            ],
+            "angle": 0,
+            "content": "Baoyu Jing, Pengtao Xie, and Eric P. Xing. 2018. On the automatic generation of medical imaging reports. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 2577-2586."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.749,
+                0.885,
+                0.828
+            ],
+            "angle": 0,
+            "content": "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. 2023. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.842,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Jonathan Krause, Justin Johnson, Ranjay Krishna, and Li Fei-Fei. 2017. A hierarchical approach for generating descriptive image paragraphs. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3337-3345."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16550"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.086,
+                0.487,
+                0.138
+            ],
+            "angle": 0,
+            "content": "C Li, C Wong, S Zhang, et al. 2023a. Llava-med: training a large language-and-vision assistant for biomedicine in one day. arxiv. arXiv preprint arXiv:2306.00890."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.147,
+                0.488,
+                0.279
+            ],
+            "angle": 0,
+            "content": "Christy Y. Li, Xiaodan Liang, Zhiting Hu, and Eric P. Xing. 2019. Knowledge-driven encode, retrieve, paraphrase for medical image report generation. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 6666-6673."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.287,
+                0.487,
+                0.352
+            ],
+            "angle": 0,
+            "content": "Mingjie Li, Bingqian Lin, Zicong Chen, Haokun Lin, Xiaodan Liang, and Xiaojun Chang. 2023b. Dynamic graph enhanced contrastive learning for chest x-ray report generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3334-3343."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.361,
+                0.487,
+                0.414
+            ],
+            "angle": 0,
+            "content": "Yuan Li, Xiaodan Liang, Zhiting Hu, and Eric P Xing. 2018. Hybrid retrieval-generation reinforced agent for medical image report generation. Advances in neural information processing systems, 31."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.422,
+                0.486,
+                0.462
+            ],
+            "angle": 0,
+            "content": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.47,
+                0.487,
+                0.536
+            ],
+            "angle": 0,
+            "content": "Che Liu, Zhongwei Wan, Yuqi Wang, Hui Shen, Haozhe Wang, Kangyu Zheng, Mi Zhang, and Rossella Arcucci. 2024. Benchmarking and boosting radiology report generation for 3d high-resolution medical images. arXiv preprint arXiv:2406.07146."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.544,
+                0.487,
+                0.65
+            ],
+            "angle": 0,
+            "content": "Fenglin Liu, Shen Ge, and Xian Wu. 2021a. Competence-based multimodal curriculum learning for medical report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 3001-3012."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.658,
+                0.487,
+                0.737
+            ],
+            "angle": 0,
+            "content": "Fenglin Liu, Xian Wu, Shen Ge, Wei Fan, and Yuexian Zou. 2021b. Exploring and distilling posterior and prior knowledge for radiology report generation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 13753-13762."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.745,
+                0.487,
+                0.798
+            ],
+            "angle": 0,
+            "content": "Fenglin Liu, Changchang Yin, Xian Wu, Shen Ge, Yuexian Zou, Ping Zhang, and Xu Sun. 2021c. Contrastive attention for automatic chest x-ray report generation. arXiv preprint arXiv:2106.06965."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.807,
+                0.486,
+                0.833
+            ],
+            "angle": 0,
+            "content": "AI Meta. 2024. Introducing meta llama 3: The most capable openly available llm to date. Meta AI."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.842,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": "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, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.086,
+                0.488,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.513,
+                0.086,
+                0.882,
+                0.191
+            ],
+            "angle": 0,
+            "content": "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, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.201,
+                0.882,
+                0.253
+            ],
+            "angle": 0,
+            "content": "Qingya Shen, Yanzhao Shi, Xiaodan Zhang, Junzhong Ji, Ying Liu, and Huimin Xu. 2024. Ghcl: Gaussian heuristic curriculum learning for brain ct report generation. Multimedia Systems, 30(2):69."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.263,
+                0.882,
+                0.316
+            ],
+            "angle": 0,
+            "content": "Yanzhao Shi, Junzhong Ji, Xiaodan Zhang, Ying Liu, Zheng Wang, and Huimin Xu. 2024. Prior tissue knowledge-driven contrastive learning for brain CT report generation. Multim. Syst., 30(2):98."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.326,
+                0.882,
+                0.405
+            ],
+            "angle": 0,
+            "content": "Yanzhao Shi, Junzhong Ji, Xiaodan Zhang, Liangqiong Qu, and Ying Liu. 2023. Granularity matters: Pathological graph-driven cross-modal alignment for brain CT report generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP, pages 6617-6630."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.414,
+                0.882,
+                0.48
+            ],
+            "angle": 0,
+            "content": "Yuhao Tang, Haichen Yang, Liyan Zhang, and Ye Yuan. 2024. Work like a doctor: Unifying scan localizer and dynamic generator for automated computed tomography report generation. *Expert Systems with Applications*, 237:121442."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.49,
+                0.882,
+                0.568
+            ],
+            "angle": 0,
+            "content": "Omkar Thawakar, Abdelrahman M. Shaker, Sahal Shaji Mullappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman H. Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. 2023. Xraygpt: Chest radiographs summarization using medical vision-language models. CoRR, abs/2306.07971."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.578,
+                0.882,
+                0.618
+            ],
+            "angle": 0,
+            "content": "Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.628,
+                0.882,
+                0.694
+            ],
+            "angle": 0,
+            "content": "Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566-4575."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.704,
+                0.882,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2015. Show and tell: A neural image caption generator. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3156-3164."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.779,
+                0.882,
+                0.858
+            ],
+            "angle": 0,
+            "content": "Jun Wang, Abhir Bhalerao, and Yulan He. 2022a. Cross-modal prototype driven network for radiology report generation. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 563-579. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.868,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving text embeddings with large language models. ArXiv, abs/2401.00368."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.524,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16551"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.179
+            ],
+            "angle": 0,
+            "content": "Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, and Ronald M. Summers. 2018. Tienet: Text-image embedding network for common thorax disease classification and reporting in chest x-rays. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 9049-9058."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.187,
+                0.489,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. 2022b. Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.25,
+                0.489,
+                0.342
+            ],
+            "angle": 0,
+            "content": "Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. 2015. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 2048-2057."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.351,
+                0.489,
+                0.444
+            ],
+            "angle": 0,
+            "content": "An Yan, Zexue He, Xing Lu, Jiang Du, Eric Y. Chang, Amilcare Gentili, Julian J. McAuley, and Chun-Nan Hsu. 2021. Weakly supervised contrastive learning for chest x-ray report generation. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, pages 4009-4015."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.452,
+                0.489,
+                0.506
+            ],
+            "angle": 0,
+            "content": "Shuxin Yang, Xian Wu, Shen Ge, S Kevin Zhou, and Li Xiao. 2022. Knowledge matters: Chest radiology report generation with general and specific knowledge. Medical image analysis, 80:102510."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.515,
+                0.489,
+                0.594
+            ],
+            "angle": 0,
+            "content": "Sisi Yang, Junzhong Ji, Xiaodan Zhang, Ying Liu, and Zheng Wang. 2021. Weakly guided hierarchical encoder-decoder network for brain ct report generation. In IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021, Houston, TX, USA, December 9-12, 2021, pages 568-573. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.603,
+                0.489,
+                0.709
+            ],
+            "angle": 0,
+            "content": "Di You, Fenglin Liu, Shen Ge, Xiaoxia Xie, Jing Zhang, and Xian Wu. 2021. Aligntransformer: Hierarchical alignment of visual regions and disease tags for medical report generation. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part III, pages 72-82."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.717,
+                0.489,
+                0.796
+            ],
+            "angle": 0,
+            "content": "Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. 2023. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.805,
+                0.489,
+                0.872
+            ],
+            "angle": 0,
+            "content": "Xiaodan Zhang, Sisi Yang, Yanzhao Shi, Junzhong Ji, Ying Liu, Zheng Wang, and Huimin Xu. 2023. Weakly guided attention model with hierarchical interaction for brain ct report generation. Computers in Biology and Medicine, page 107650."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.881,
+                0.489,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Yixiao Zhang, Xiaosong Wang, Ziyue Xu, Qihang Yu, Alan L. Yuille, and Daguang Xu. 2020. When radiology report generation meets knowledge graph."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.528,
+                0.086,
+                0.885,
+                0.177
+            ],
+            "angle": 0,
+            "content": "In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 12910-12917."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "16552"
+        }
+    ]
+]

2024/See Detail Say Clear_ Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning/081b7730-6494-4c1e-8ba3-d9c947e8be2f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3835982be65f8c0a7f5679eb0bde0de158ad0b934faaf07d028ec349cf01b2f
+size 4157572

2024/See Detail Say Clear_ Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning/full.md

@@ -0,0 +1,290 @@
+# See Detail Say Clear: Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning
+
+Chengxin Zheng $^{1}$ , Junzhong Ji $^{1}$ , Yanzhao Shi $^{1}$ , Xiaodan Zhang $^{1,*}$ , Liangqiong Qu $^{2}$
+
+$^{1}$ College of Computer Science, Beijing University of Technology, Beijing, China  
+ $^{2}$ Department of Statistics and Actuarial Science, School of Computing and Data Science, The University of Hong Kong, Hong Kong, China
+
+Correspondence: zhangxiaodan@bjut.edu.cn
+
+# Abstract
+
+Brain CT report generation is significant to aid physicians in diagnosing cranial diseases. Recent studies concentrate on handling the consistency between visual and textual pathological features to improve the coherence of report. However, there exist some challenges: 1) Redundant visual representing: Massive irrelevant areas in 3D scans distract models from representing salient visual contexts. 2) Shifted semantic representing: Limited medical corpus causes difficulties for models to transfer the learned textual representations to generative layers. This study introduces a Pathological Clue-driven Representation Learning (PCRL) model to build cross-modal representations based on pathological clues and naturally adapt them for accurate report generation. Specifically, we construct pathological clues from perspectives of segmented regions, pathological entities, and report themes, to fully grasp visual pathological patterns and learn cross-modal feature representations. To adapt the representations for the text generation task, we bridge the gap between representation learning and report generation by using a unified large language model (LLM) with task-tailored instructions. These crafted instructions enable the LLM to be flexibly fine-tuned across tasks and smoothly transfer the semantic representation for report generation. Experiments demonstrate that our method outperforms previous methods and achieves SoTA performance. Our code is available at https://github.com/Chauncey-Jheng/PCRL-MRG.
+
+# 1 Introduction
+
+Brain computed tomography (CT) imaging is essential for diagnosing various cranial diseases, including cerebral infarction and hemorrhage. However, it is time-consuming and error-prone for radiologists to manually interpret medical findings from these scans and write reports. Automated report
+
+![](images/a6944c8749122044e7cba7df9684a4584674e72bec357b874b4aa436abf20f15.jpg)  
+Figure 1: Comparison of different cross-modal alignment paradigms. (a) Previous way: The origin CT images contain extraneous information unrelated to diagnosis, and separate training for two semantic tasks makes it challenging to find a shared optimal solution, resulting in inadequate cross-modal alignment. (b) Our way: The visual representation is refined to concentrate on pathological clues, and employ joint training with task-tailored instructions via unified LLM to find a transferable representation, leading to better adaption for report generation.
+
+generation systems are designed to boost efficiency, reduce the workload for radiologists, and optimize resources in busy clinical scenarios.
+
+With the advancement of deep neural networks and their successful application in image captioning tasks (Vinyals et al., 2015; Xu et al., 2015), medical report generation (MRG) has gained more attention. Unlike the short sentences of traditional image captioning, MRG aims to generate lengthy and precise reports. To achieve this, various cross-modal alignment methods are required to ensure the consistency between visual and textual information (Shi et al., 2023), including attention mech-
+
+anisms (Jing et al., 2018; Wang et al., 2018), memory mechanisms (Chen et al., 2020, 2021), and knowledge graphs (Li et al., 2019, 2023b).
+
+Recently, learning representations via visual-textual contrastive learning (Shi et al., 2023, 2024) or using pre-trained large language models (LLMs) (Thawakar et al., 2023) to strength representations are also proven to be effective.
+
+However, as shown in Figure 1(a), learning cross-modal correspondences is still challenging in current methods due to the following concerns: 1) Redundant visual representing. Different from chest X-ray data, 3D brain CT scans contain extensive redundant information, e.g. background and insignificant areas. With the lack of human-crafted boxes to locate pathology regions, models struggle to capture and interpret the visual pathology patterns for generating reports. Although current advanced models use semantic prior knowledge or medical prompts (Jin et al., 2024; Bu et al., 2024) to automatically learn the salient visual areas, this may introduce noise and unstable representation and cause severe hallucinations in generated texts. 2) Shifted semantic representing. Compared to natural corpus, limited brain CT report corpus is insufficient to transfer the pathological semantic representations learned by represent learning layers (e.g., contrastive learning layer) to the language model (Huh et al., 2024), since the direct weight-sharing (Shi et al., 2023, 2024) is prone to degrade the coherence of generated diagnostic sentences. Thus, how to uniformly represent cross-modal pathological features and adapt them to report generation still remains an open question.
+
+In this paper, we propose a Pathological Cluedriven Representation Learning (PCRL) model to seamlessly build cross-modal representations based on diverse pathological clues and transfer them for generating accurate brain CT reports. Specifically, we extract pathological clues from perspectives of segmented regions, pathological entities, and report themes to depict clinical scenarios. Segmented region clues are automatically generated and filtered by given pathology prompts, enabling the visual encoder to grasp visual pathological patterns. Meanwhile, entity and theme clues are respectively extracted by detailed findings and full-text reports, to handle the enriched visual-textual alignment and build cross-modal pathology representations.
+
+Besides, to adapt the learned representations for the text generation task, we bridge the gap between representation learning and report generation by
+
+employing a unified large language model (LLM) with task-tailored instructions, which has proven to be more effective than conventional decoders by using appropriate tokens to seamlessly connect different tasks (Yu et al., 2023). As shown in Figure 1(b), We craft a representation instruction to prompt the LLM to produce high-level pathological semantic features for cross-modal alignment, and a generation instruction to prompt LLM to generate accurate brain CT reports based on the learned representations.
+
+Our main contributions can be summarized as:
+
+1. We propose a novel framework to seamlessly learn visual-textual representations from perspectives of diverse pathological clues and leverage them for enhancing the quality of generated brain CT reports.  
+2. We, for the first time, design a new paradigm to effectively transfer the learned pathological representations for report generation using a unified LLM prompted by task-tailored instructions.  
+3. We validate the model capabilities on the open-source CTRG-Brain dataset. Experimental results show that our model achieves remarkable performance in generating brain CT reports.
+
+# 2 Related Work
+
+# 2.1 Medical Report Generation
+
+The advancements in image captioning techniques have spurred the development of a range of radiology report generation methods (Jing et al., 2018; Chen et al., 2020, 2021; Li et al., 2019; Shi et al., 2023; Zhang et al., 2023; Li et al., 2023b; Shi et al., 2024; Shen et al., 2024). To exploit the more effective parts of medical images, Jing et al. (2018) first propose a co-attention mechanism to associate images with disease tags and improve the accuracy of generated reports; Chen et al. (2021) design a cross-modal memory matrix to learn high-level vision-language correspondence for enhancing report generation; Liu et al. (2021c) used contrastive attention to captured the difference of abnormal and normal samples. To address the semantic bias of limited medical report corpus, Jing et al. (2020) exploited the textual structure information of reports, while Liu et al. (2021a); Shen et al. (2024) used curriculum learning to alleviate the data bias of limited medical data corpus. The retrieval-based
+
+![](images/161384d732f8271cc544bc4dd7d693aa996edc8b2faa05cd3aa2e30ecce89dc8.jpg)  
+Figure 2: The overall framework of our method, which mainly consists of an image encoder and a text decoder for brain CT report generation (left). The pathological clue-driven representation learning (right) is proposed to guide the encoder and decoder for more fine-grained representation by three alignment modules: (a) segmentation clue alignment (SCA), (b) entity clue alignment (ECA), and (c) theme clue alignment (TCA).
+
+(Li et al., 2018; Liu et al., 2021b) and prior expert knowledge based methods (Zhang et al., 2020; Yang et al., 2022; You et al., 2021) also shows effectiveness to alleviate the inaccuracy of generated report caused by limited training data. To build more stable relations between multi-modal features, contrastive learning (Radford et al., 2021) acts as an unsupervised method, and is proven to be effective for learning cross-modal representations for MRG tasks (Li et al., 2023b; Shi et al., 2024).
+
+Recently, LLM-based MRG methods (Jin et al., 2024; Chen et al., 2024) harness the semantic processing capabilities of LLMs to enhance the coherence of generated medical reports. However, the current bottleneck of LLM is to interpret sparse visual patterns of pathologies, which causes severe hallucinations and degrades the trustworthiness of clinical diagnosis. To solve this issue, we propose to augment LLM with representation learning guided by pathology clues. By aligning cross-modal pathological features and transferring the learned representations to the LLM-based report decoder, our model can generate accurate and coherent brain CT reports.
+
+# 2.2 Pre-trained Large Model
+
+In recent years, pre-trained large models have achieved significant breakthroughs in both natural language processing (NLP) and computer vision (CV). These models leverage extensive datasets for pre-training, allowing them to perform exceptionally well in downstream tasks. The Segment
+
+Anything Model (SAM) Kirillov et al. (2023) is an innovative large vision model in the field of image segmentation. SAM's strengths lie in its extensive training on a large-scale dataset, enhancing its versatility and enabling real-time generation of high-quality segmentation masks across diverse tasks and image distributions. MedCLIP Wang et al. (2022b) is a notable visual-language model in the medical domain. Pre-trained on decoupling medical images and texts, MedCLIP has shown SoTA performance on zero-shot image-text retrieval. LLaMA3 is the next generation of state-of-the-art open-source LLM from Meta (2024). Trained on over 15 trillion tokens, seven times the data of its predecessor, LLaMA3 incorporates extensive multilingual and high-quality datasets. To improve the use of large models for assisting in medical report generation, we adopt SAM for region segmentation, MedCLIP for feature representation of pathological entities, and LLaMA for joint training to achieve fine-grained cross-modal alignment and improve medical report generation.
+
+# 3 Methodology
+
+As illustrated in Figure 2, our framework consists of two branches: the brain CT Report Generation (RG) on the left and the Pathological Clue-driven Representation Learning (PCRL) on the right. The interaction between these branches is facilitated through the shared use of a visual encoder and a language model.
+
+# 3.1 Brain CT Report Generation
+
+In this branch, the input comprises a set of CT scan images  $I = \{i_1,\dots ,i_N\}$ , where  $N$  represents the number of CT images in each sample. The output is the corresponding brain CT findings report  $Y = \{y_{1},\ldots ,y_{M}\}$ , with  $M$  representing the number of tokens. We adopt the encoder-decoder architecture for report generation. First, we use ResNet101 (He et al., 2016) to extract grid features  $G = \{g_{1},\dots ,g_{N}\} \in \mathbf{R}^{N\times H\times d}(N = 24,H = 196,d = 2048)$  from  $I$ . Then, we embed these features using a visual encoder to obtain visual features  $V$ . These visual features  $V$  are passed through a multi-modal projector to obtain the visual embedding tokens  $H_{v}\in \mathbf{R}^{N\times d_{w}}(d_{w} = 4096)$  which are aligned with the word embedding space of a large language model. Finally, these tokens are concatenated with the instruction token  $H_{q}$  and fed into the large language model. We train the parameters  $\theta$  by minimizing the cross-entropy loss, which can be represented by the following formula:
+
+$$
+\mathcal {L} _ {g} = - \sum_ {t = 1} ^ {M} \log P \left(y _ {t} \mid y _ {1: t - 1}, H _ {v}, H _ {q}; \theta\right) \tag {1}
+$$
+
+where,  $P(y_{t}|*)$  denotes the probability conditioned on  $H_{v}, H_{q}$ , and the embeddings of previous words  $y_{1}, y_{2}, \dots, y_{t-1}$ .
+
+# 3.2 Pathological Clue-driven Representation Learning
+
+Learning fine-grained visual-text representations is critical for generating accurate reports (Shi et al., 2023). This branch mainly builds enriched representations by conducting multi-modal feature alignment based on pathological clues, including segmentation clue alignment (SCA), entity clues alignment (ECA), and theme clues alignment (TCA).
+
+# 3.2.1 Preparation of Pathological Clues
+
+Pathological clues are collected in three perspectives for feature alignment.
+
+**Pathology Theme Clues:** Learning the structure of professional brain CT reports is essential for MRG models to satisfy human standards and produce reliable reports. To this end, we propose to build theme clues by simply using the full-text report and whole images as global signals, which can be useful to enhance the overall quality of medical reports by TCA (illustrated in 3.2.4).
+
+Pathology Entity Clues: To learn detailed information about pathology entities, we regard each
+
+single finding sentence in the report as an entity clue and extract the related multi-modal features in ECA 3.2.3. The construction of entities  $E = \{e_1,\dots,e_{N_e}\}$  (with  $N_{e} = 24$ ) is based on expert knowledge and the frequency of words in the training corpus.
+
+Segment Clues: To enhance visual representations via detailed contour of pathologies, we propose to utilize the pre-trained segmentation model SAM (Kirillov et al., 2023) to generate mask candidates and filter useful masks related to pathology entities. First, we prompt SAM by covering each brain CT image with a grid of points and integrating it into image embeddings through average sampling. In this way, the mask decoder in SAM is prompted to generate a gallery of candidate masks  $M_I = \{m_1, \dots, m_{N_m}\}$ . To ensure the quality of masks, we also apply rule-based methods to filter out low-quality and duplicate masks with area size, stability scores, and IoU scores. For each sample consisting of 24 images, we generate a corresponding series of masks and combine them to form the candidate Gallery  $= \{M_{I_1}, \dots, M_{I_N}\}$ .
+
+Then, to retrieve valuable masks related to the sample's pathological entities, we propose to use the MedCLIP (Wang et al., 2022b) for text-prompted retrieval. Based on expert knowledge, we divide the 24 images into eight-layer categories, each mapping a specific set of entities (Shi et al., 2024). Conversely, each entity also has its corresponding layers. We extract the existing entities from the sample's report and their corresponding pathological descriptions  $D = \{d_{1},\dots,d_{N_{d}}\}$  ( $N_{d}\leq N_{e}$ ), which serve as the Query. Next, we use the  $d$ -th description  $Query_{d}$  to search for the most similar matching mask in  $Gallery_{d}$ , which is a subset of  $Gallery$ , consisting of all visual masks of the layers related to the existing entities. The procedure can be represented as:
+
+$$
+\operatorname {R e t r i e v a l} _ {d} = \operatorname {C L I P} \left(\operatorname {G a l l e r y} _ {d}, \operatorname {Q u e r y} _ {d}\right) \tag {2}
+$$
+
+where  $Retrieval_{d}$  denotes the retrieved masks for SCA 3.2.2.
+
+# 3.2.2 Segmentation Clue Alignment
+
+This module aims to learn the fine-grained visual representation based on the extracted segment masks. First, we obtain the visual features  $V_{D} = \{v_{1},\dots v_{N_{d}}\}$  corresponding to the images containing the entities using a selector. Then, we input these features into a lightweight segmentation
+
+head to generate the corresponding foreground entity masks  $S_{D} = \{s_{1},\dots,s_{N_{d}}\}$ , which serves as discriminated pathological information. We then align these generated masks with the retrieved segment masks  $M_{D} = \{Retrieval_{1},\dots,Retrival_{N_{d}}\}$  to learn detailed patterns. It is important to note that the size of the generated masks  $S_{D}$  differs from the size of the retrieved SAM masks  $M_{D}$ . To address this, we first resize the SAM masks to match the size of the masks generated by our segmentation head. We finally calculate the following loss for aligning the two types of masks:
+
+$$
+\mathcal {L} _ {s e g} = 1 - \frac {2 \sum_ {i = 1} ^ {n} p _ {i} y _ {i}}{\sum_ {i = 1} ^ {n} p _ {i} ^ {2} + \sum_ {i = 1} ^ {n} y _ {i} ^ {2}} \tag {3}
+$$
+
+where  $y_{i}$  is the pixel value of retrieved SAM masks and  $p_i$  is the predicted pixel value.
+
+In this way, the visual encoder can be effectively learned to focus on the areas of pathologies while reducing the influence of irrelevant visual features.
+
+# 3.2.3 Entity Clue Alignment
+
+To learn the cross-modal patterns of pathology entities, we extract visual and textual pathological features based on the entity clues. First, we build a selector to obtain the visual features  $V_{D} = \{v_{1},\dots v_{N_{d}}\} (N_{d} < N)$ , which corresponds to significant CT images that exist pathology entities. We then apply global average pooling (GAP) to generalize  $V_{D}$  and obtain representations of each significant CT image, denoted as  $R_v^e = \{r_1,\dots r_{N_d}\}$ .
+
+Different from previous work (Li et al., 2023b) use an external language model (e.g. SciBert (Beltagy et al., 2019)) to build textual representation for feature alignment, we propose to leverage a unified LLM to generate textual representation via tailored prompts. Inspired by Wang et al. (2023), we crafted a representation instruction "Summarize the following cranial diagnosis in one word," which is denoted as  $H_{q_r}$ . This prompt can effectively activate the summarization ability of LLM to generate corresponding text representations for pathological entity description in  $D$ . We use the output generated by the final hidden layer of LLM as textual features for cross-modal alignment.
+
+With the carefully extracted visual and textual representations of entities, we map them into the same dimension through an embedding layer. The process can be represented by the following formula:
+
+$$
+R _ {v} ^ {e} = \operatorname {L i n e a r} _ {v} ^ {e} (G A P \left(V _ {D}\right)) \tag {4}
+$$
+
+$$
+R _ {w} ^ {e} = \operatorname {L i n e a r} _ {t} ^ {e} \left(L L M \left(H _ {q _ {r}}, D\right)\right) \tag {5}
+$$
+
+where  $GAP$  denotes the global average pooling,  $Linear_v$  represents the visual mapper, and  $Linear_t$  represents the textual mapper. We employ the symmetric InfoNCE (van den Oord et al., 2018) loss for visual-textual alignment:
+
+$$
+\begin{array}{l} \mathcal {L} _ {C L _ {e}} = - \frac {1}{2} \left(\alpha_ {v} \sum_ {i = 1} ^ {N _ {d}} \log \frac {\exp \left(\mathrm {s} _ {v} ^ {e} (i , i)\right)}{\sum_ {j = 1 , j \neq i} ^ {N _ {d}} \exp \left(\mathrm {s} _ {v} ^ {e} (i , j)\right)} \right. \\ + \alpha_ {w} \sum_ {i = 1} ^ {N _ {d}} \log \frac {\exp \left(\mathrm {s} _ {w} ^ {e} (i , i)\right)}{\sum_ {j = 1 , j \neq i} ^ {N _ {d}} \exp \left(\mathrm {s} _ {w} ^ {e} (i , j)\right)} \Bigg) \tag {6} \\ \end{array}
+$$
+
+where  $s_v^e (i,j) = \frac{\sin(R_{v_i}^e,R_{w_j}^e)}{\tau}$  and  $s_w^e (i,j) = \frac{\sin(R_{w_i}^e,R_{v_j}^e)}{\tau}$  denote the similarity between the visual representation  $R_{v}^{e}$  and the textual representation  $R_{w}^{e}$ ,  $\tau$  is a temperature parameter.  $\alpha_{v}$  and  $\alpha_{w}$  are hyperparameters to balance the contrastive learning.
+
+By aligning multi-modal entity clues, the model can grasp fine-grained visual-text representations to generate accurate diagnostic words.
+
+# 3.2.4 Theme Clue Alignment
+
+Theme clue alignment aims to equip the model with comprehensive skills to generate accurate style and structure of reports, thereby enhancing clinical reliability. For global visual features, we use one-dimensional global max pooling (GMP) to represent the entire sample visually, denoted as  $R_{w}^{t}$ . For the overall report of the sample, we generate textual representation  $R_{w}^{t}$  by prompting LLM with the same representation instruction (see Section 3.2.3). This process can be represented as:
+
+$$
+R _ {v} ^ {t} = \operatorname {L i n e a r} _ {v} ^ {t} (G M P (V)) \tag {7}
+$$
+
+$$
+R _ {w} ^ {t} = \operatorname {L i n e a r} _ {t} ^ {t} \left(L L M \left(H _ {q _ {s}}, Y\right)\right) \tag {8}
+$$
+
+Similar to ECA, the loss of TCA can be formulated as:
+
+$$
+\begin{array}{l} \mathcal {L} _ {C L _ {t}} = - \frac {1}{2} \left(\alpha_ {v} \sum_ {i = 1} ^ {N _ {b}} \log \frac {\exp (\mathrm {s} _ {v} ^ {t} (i , i))}{\sum_ {j = 1 , j \neq i} ^ {N _ {b}} \exp (\mathrm {s} _ {v} ^ {t} (i , j))}\right) \\ + \alpha_ {w} \sum_ {i = 1} ^ {N _ {b}} \log \frac {\exp \left(\mathrm {s} _ {w} ^ {t} (i , i)\right)}{\sum_ {j = 1 , j \neq i} ^ {N _ {b}} \exp \left(\mathrm {s} _ {w} ^ {t} (i , j)\right)} \tag {9} \\ \end{array}
+$$
+
+where  $s_v^t (i,j) = \sin (R_{v_i}^t,R_{w_j}^t) / \tau$  and  $s_w^t (i,j) =$ $\sin (R_{w_i}^t,R_{v_j}^t) / \tau$  denote the similarity between the visual representation  $R_{v}^{e}$  and the textual representation  $R_w^e$ ,  $\tau$  is a temperature parameter,  $N_{b}$  is the batch size. Shared with ECA,  $\alpha_{v}$  and  $\alpha_{w}$  are set to balance the feature alignment.
+
+<table><tr><td>Methods</td><td>Decoder</td><td>B1</td><td>B2</td><td>B3</td><td>B4</td><td>M</td><td>RG</td><td>C</td><td>F1</td></tr><tr><td>HRNN(Krause et al., 2017)†</td><td>LSTM</td><td>42.3</td><td>28.3</td><td>21.2</td><td>17.1</td><td>27.2</td><td>39.3</td><td>20.9</td><td>70.2</td></tr><tr><td>Up-Down(Anderson et al., 2018)†</td><td>LSTM</td><td>45.8</td><td>34.8</td><td>28.5</td><td>24.4</td><td>31.6</td><td>42.5</td><td>27.3</td><td>70.2</td></tr><tr><td>WCL(Yan et al., 2021)†</td><td>LSTM</td><td>49.5</td><td>36.5</td><td>29.4</td><td>25.1</td><td>31.3</td><td>42.8</td><td>33.3</td><td>64.5</td></tr><tr><td>R2Gen-CMN(Chen et al., 2021)†</td><td>Transformer</td><td>49.1</td><td>40.0</td><td>34.4</td><td>30.1</td><td>29.9</td><td>48.6</td><td>84.2</td><td>69.8</td></tr><tr><td>XProNet(Wang et al., 2022a)†</td><td>Transformer</td><td>50.6</td><td>41.3</td><td>34.4</td><td>29.1</td><td>31.5</td><td>51.7</td><td>83.3</td><td>70.1</td></tr><tr><td>WGAM(Yang et al., 2021)†</td><td>LSTM</td><td>49.4</td><td>36.7</td><td>29.6</td><td>25.4</td><td>32.0</td><td>42.4</td><td>31.9</td><td>68.9</td></tr><tr><td>PGCA(Shi et al., 2023)†</td><td>LSTM</td><td>50.2</td><td>37.8</td><td>30.7</td><td>26.5</td><td>32.5</td><td>43.0</td><td>34.0</td><td>69.2</td></tr><tr><td>WGAM-HI(Zhang et al., 2023)†</td><td>LSTM</td><td>50.4</td><td>37.6</td><td>30.5</td><td>26.1</td><td>31.4</td><td>43.8</td><td>33.2</td><td>67.4</td></tr><tr><td>LLaVA-med(Li et al., 2023a)†</td><td>LLaMA3-8B</td><td>50.0</td><td>39.3</td><td>32.0</td><td>26.3</td><td>31.4</td><td>46.7</td><td>38.6</td><td>64.3</td></tr><tr><td>HILT(Liu et al., 2024)†</td><td>LLaMA3-8B</td><td>50.7</td><td>39.9</td><td>33.1</td><td>27.9</td><td>30.8</td><td>46.1</td><td>43.7</td><td>68.4</td></tr><tr><td>PromptMRG(Jin et al., 2024)†</td><td>LLaMA3-8B</td><td>48.1</td><td>38.3</td><td>31.6</td><td>26.5</td><td>31.0</td><td>47.4</td><td>50.3</td><td>69.2</td></tr><tr><td>Ours (Jieba tokenizer)</td><td>LLaMA3-8B</td><td>51.5</td><td>42.0</td><td>35.7</td><td>30.9</td><td>31.4</td><td>49.0</td><td>80.0</td><td>70.6</td></tr><tr><td>Ours (LLaMA tokenizer)</td><td>LLaMA3-8B</td><td>62.0</td><td>54.7</td><td>49.4</td><td>45.3</td><td>33.1</td><td>57.7</td><td>96.4</td><td>70.6</td></tr></table>
+
+Table 1: The performance of our PCRL compared with previous state-of-the-art models on the Brain CT report generation dataset CTRG-Brain. The best results are highlighted in bold.  $\dagger$  denotes the re-implementation results.
+
+# 3.3 Joint Training
+
+In the training stage, we jointly train the RG branch and the PCRL branch to maximize the utilization of multi-granularity visual-text representations. Instead of using separate modules to learn representation and generate medical reports (Li et al., 2023b), we use a unified LLM to bridge the gap between representation learning and report generation via two task-tailored instructions, i.e., representation instruction and generation instruction. This can transfer the representations learned by the PCRL branch to optimize the RG branch effectively, thereby generating accurate reports.
+
+Our final loss contains the above-mentioned losses, which can be formulated as:
+
+$$
+\mathcal {L} = \mathcal {L} _ {g} + \mathcal {L} _ {\text {s e g}} + \mathcal {L} _ {C L _ {e}} + \mathcal {L} _ {C L _ {t}} \tag {10}
+$$
+
+# 4 Experiments
+
+# 4.1 Dataset
+
+We validate the performance of our model using the CTRG-Brain (Tang et al., 2024) dataset. This dataset comprises 6,000 samples, containing a total of 160,336 CT images and 6,000 Chinese medical reports. Following the mainstream division method (Shi et al., 2023; Zhang et al., 2023), we split the dataset into a training set, a validation set, and a test set in a 7:1:2 ratio. For consistent processing, we divide the brain CT image samples into 8 layers based on expert knowledge, with each layer containing 3 continuous CT images, assigning each sample with 24 CT images.
+
+# 4.2 Evaluation Metrics
+
+We chose Natural Language Generation (NLG) metrics and Clinical Evaluation (CE) metrics to evaluate our model's performance. NLG metrics include BLEU (Papineni et al., 2002),METEOR (Banerjee and Lavie), ROUGE-L (Lin, 2004), and CIDEr (Vedantam et al., 2015), denoted as B1, B2, B3,B4,M,RG,and C respectively. To measure the pathological accuracy, we use 24 keywords summarized by experienced radiologists to calculate the Clinical Evaluation (CE) metric(Shi et al., 2023; Zhang et al., 2023), i.e., F1 score, which is the harmonic mean of the precision and recall.
+
+# 4.3 Implementation Details
+
+We reshape the size of image to  $512 \times 512$  pixels and used a ResNet101 to extract image features, which is pre-trained on the ImageNet dataset and fine-tuned on the CQ500 dataset (Chilamkurthy et al., 2018). For our large language model, we utilize LLaMA3-8B (Meta, 2024), which is quantized to 4-bit, and use the LoRa (Hu et al., 2021) for parameter-efficient fine-tuning. The overall trainable parameter quantity of our model is  $229.9\mathrm{M}$  with  $3.4\mathrm{M}$  parameters in LLM (only  $0.04\%$ ). During training, we use the AdamW optimizer with a learning rate of 1e-4. The batch size is 4, with 1050 training steps per epoch. For testing, we set the temperature coefficient of the large model to 0.6 and the top-p value to 0.9. The model is implemented using PyTorch 2.3.0, and the entire training process is conducted on a single RTX 4090 GPU.
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">Representation</td><td colspan="3">Module Loss</td><td rowspan="2">B1</td><td rowspan="2">B2</td><td rowspan="2">B3</td><td rowspan="2">B4</td><td rowspan="2">M</td><td rowspan="2">RG</td><td rowspan="2">C</td></tr><tr><td>Visual</td><td>Textual</td><td>LSCA</td><td>LECA</td><td>LTCA</td></tr><tr><td>Baseline</td><td>X</td><td>X</td><td>X</td><td>X</td><td>X</td><td>49.6</td><td>39.9</td><td>33.4</td><td>28.6</td><td>30.7</td><td>48.4</td><td>50.8</td></tr><tr><td>(a)</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>X</td><td>51.3</td><td>41.0</td><td>34.5</td><td>29.7</td><td>31.2</td><td>47.7</td><td>70.8</td></tr><tr><td>(b)</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>X</td><td>51.3</td><td>41.4</td><td>35.0</td><td>30.1</td><td>31.2</td><td>48.7</td><td>71.3</td></tr><tr><td>(c)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>X</td><td>51.7</td><td>41.4</td><td>34.6</td><td>29.5</td><td>31.2</td><td>48.0</td><td>72.8</td></tr><tr><td>(d)</td><td>✓</td><td>✓</td><td>X</td><td>✓</td><td>✓</td><td>48.9</td><td>38.8</td><td>32.2</td><td>27.4</td><td>30.0</td><td>46.3</td><td>58.3</td></tr><tr><td>(e)</td><td>✓</td><td>✓(Bert)</td><td>✓</td><td>✓</td><td>✓</td><td>49.5</td><td>39.0</td><td>31.8</td><td>26.3</td><td>30.4</td><td>45.4</td><td>48.5</td></tr><tr><td>Ours</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>51.5</td><td>42.0</td><td>35.7</td><td>30.9</td><td>31.4</td><td>49.0</td><td>80.0</td></tr></table>
+
+Table 2: Ablation studies of our proposed method. The Baseline model is an encoder-decoder framework without an alignment mechanism. (a), (b), (c), (d) and (e) respectively denote the use of different representations (i.e., visual and textual in ECA and TCA) and module losses.
+
+# 4.4 Quantitative Analysis
+
+We compare the proposed PCRL with some competitive brain CT report generation methods (WGAM (Yang et al., 2021), PGCA (Shi et al., 2023), WGAM-HI(Zhang et al., 2023)). Besides, we also reproduced some SOTA models in image captioning (HRNN (Krause et al., 2017), Up-Down (Anderson et al., 2018)) and chest Xray report generation (WCL (Yan et al., 2021), R2Gen-CMN (Chen et al., 2021), XProNet (Wang et al., 2022a)) for comprehensive comparisons on CTRG-Brain dataset. What's more, for fair comparisons, We also reproduce several related LLM-based methods (LLaVA-Med(Li et al., 2023a), HILT(Liu et al., 2024), PromptMRG(Jin et al., 2024)with the same LLM decoder.
+
+As shown in Table 1, our method outperforms others across most evaluation metrics. WGAM and WGAM-HI employ weakly-supervised visual attention to extract key visual features, resulting in higher BLEU scores. With contrastive learning for cross-modal alignment, WCL and PGCA can effectively learn relations between CT images and reports, achieving better results. However, due to the lack of training data, the above methods still produce reports that fall short in fluency and readability. Here's a refined version of your sentence: The LLM-based methods (LLaVA-med, HILT, PromptMRG) demonstrate poorer performance compared to traditional transformer-based methods (R2Gen-CMN, XProNet). This suggests that prior knowledge from pretrained large models, without further cross-modal alignment, may lead to hallucinations in report generation.
+
+Our PCRL achieves fine-grained cross-modal alignment by leveraging a series of pre-trained large models, generating more fluent reports and
+
+achieves the best performance compared to other methods. It is noteworthy that we tested our method using the Jieba and LLaMA tokenizers respectively to compute NLG scores, with the latter achieving the best performance across all NLG metrics.
+
+This discrepancy may be due to differences in tokenization methods used during training. While other models employ the traditional Jieba tool for Chinese word tokenization, our PCRL follows more advanced BPE-based subword tokenization. Nevertheless, our model also achieves competitive results in overall metrics.
+
+# 4.5 Ablation Study
+
+To evaluate the effect of each component in PCRL, we have done plenty of ablation studies, as shown in Table 2. Baseline is the standard encoder-decoder (ResNet101-LLaMA3) architecture without alignment. By progressively adding visual or textual representation (visual and textual) and the module losses  $(\mathcal{L}_{SCA},\mathcal{L}_{ECA}$  and  $\mathcal{L}_{TCA})$  , respectively denoting the incorporation of two modalities of representations and the utilization of three pathological clue-driven alignments (i.e., SCA, ECA, and TCA).
+
+By comparing (a) with the baseline, we observe that incorporating the segmentation alignment significantly enhances the model's performance. This demonstrates that our SCA method effectively filters out irrelevant information from CT images and extracts crucial pathology-related visual regions. In contrast, with the implementation of ECA, (b) achieves comparable or better performance across all metrics compared to (a). This indicates that fine-grained cross-modal pathological entity alignment can effectively learn strongly correlated visual-text
+
+<table><tr><td>Ground Truth Sample</td><td>Ground Truth Entities</td><td>Retrieved Masks</td><td>Generated Reports</td><td></td></tr><tr><td>...</td><td>基底节区(basal ganglia):左侧基底节区见小片状低密度影 (Small patchy low-density shadows visible in the left basal ganglia area)</td><td></td><td>双侧大脑半球对称,两侧基底节区见斑点状低密度影,中线结构居中,脑沟、脑池、蛛网膜下腔稍增宽,脑回变小。(Bilateral cerebral hemispheres are symmetrical, with punctate low-density shadows in the basal ganglia on both sides. The midline structure is centered, and the sulcus, cistern, and subarachnoid space are slightly widened, while the gyrus is smaller.)</td><td rowspan="3">Baseline</td></tr><tr><td rowspan="3">双侧大脑半球对称,左侧基底节区见小片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑,脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A small patch of low-density shadow is seen in the left basal ganglia region, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal.)</td><td>脑室(ventricle):各脑室、脑池大小形态正常(The size and shape of all ventricles and cisterns are normal.)</td><td rowspan="3"></td><td>双侧大脑半球对称,右侧丘脑见小片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑、脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A small patch of low-density shadow is seen in the right thalamus, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal. The midline structure is centered. There are no abnormalities in the cerebellum and brainstem below the tentorium.)</td></tr><tr><td>中线(midline structure):中线结构居中 (The midline structure centered)</td><td rowspan="2">双侧大脑半球对称,左侧基底节区见斑片状低密度影,边缘欠清晰,无占位效应,各脑室、脑池大小形态正常,中线结构居中,幕下小脑、脑干无异常。(The bilateral cerebral hemispheres are symmetrical. A patch of low-density shadow is seen in the left basal ganglia region, with unclear edges and no mass effect. The size and shape of all ventricles and cisterns are normal. The midline structure is centered. There are no abnormalities in the cerebellum and brainstem below the tentorium.)</td></tr><tr><td>脑干(brainstem):幕下小脑,脑干无异常(There are no abnormalities in the cerebellum and brainstem below the tentorium)</td><td>Ours</td></tr></table>
+
+representations, thus better supporting medical report generation. Besides, (c) combines both SCA and ECA and shows improved performance in the B1 and C metrics, which are highly correlated with keyword frequency. However, it slightly underperforms in the B2, B4 and RG metrics compared to (b), which is more related to overall text style.(d) showed SCA is essential for focusing on entity details and maintaining the thematic style.
+
+Notably, to validate the contribution of joint representation learning, we replace the language model with Bert model for textual representation during alignment in (e). The result indicated that shared representation through joint learning with unified language model can effectively improve the overall quality of the task.
+
+# 4.6 Qualitative Analysis
+
+We visualize the brain CT reports generated by baseline, WGAM-HI (Zhang et al., 2023) and our model in Figure 3. Given the ground truth brain CT sample and entity data, our model generates better brain CT report with the most accurate entity words (e.g. basal ganglia, ventricle, and brainstem) among the competitors, which demonstrates the effectiveness of using diverse medical clues to learn enriched multi-modal representations. It can be also seen that compared with the baseline, the report generated by our model has a better semantic structure, indicating the contribution of employing a unified LLM to transfer useful representations. Besides, we also find that the retrieved entity masks
+
+![](images/7719a15e36aa57ee5e55575f8b57839495bd7278f8f3dc11df92784ea5e8be81.jpg)  
+Figure 3: Visualization of report generation and mask segmentation. Given the ground truth sample and corresponding entities, the retrieved entity masks are listed in the third column. Reports generated by Baseline, WGAM-HI, and our model are listed in the fourth column. Different colors denote the specific entity words and entity masks, respectively. The English translation is given for a better understanding of the original Chinese reports in CTRG-Brain.  
+Figure 4: Visualization of the segmentation masks gallery generated by SAM for a single sample.
+
+from segmentation masks gallery (see Figure 4) can generally match related entity words at both levels of visual and semantic. For example, the mask of "midline structure" and "brainstem" matches the empirical scan slice and fine-grained region chosen by experienced radiologists. This guarantees the model to mine accurate visual cranial patterns, therefore generating high-quality reports.
+
+# 5 Conclusion
+
+We propose a novel model to mine pathological clues for enhancing multi-modal representations and seamlessly transfer them into report generation. First, through carefully designed segmentation clue alignment, entity clue alignment, and theme clue alignment, the diverse and precise feature representation can be well-constructed. Second, we transfer the learned representation to boost the brain CT report generation via a unified LLM prompted by task-tailored instructions. Experiments demonstrate the effectiveness of our model in generating pathologically accurate reports.
+
+# Limitations
+
+Although the segmentation clues retrieved by MedClip (Wang et al., 2022b) can generally match corresponding pathological entities and help the model neglect redundant visual information, it should be noted that a part of retrieved entity masks may not be precise. This is because MedClip is mainly pretrained by chest X-ray data with limited brain CT samples. Thus, addressing this challenge is imperative for the research community. In the future, we will work on exploring useful approaches. One potential approach is to train a unified text-prompted medical segmentation model towards 3D brain CT scans, which can not only be employed to offer fine-grained visual information for medical report generation but also for other related tasks, e.g. medical VQA.
+
+# References
+
+Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6077-6086.  
+S Banerjee and A Lavie. Meteor: an automatic metric for mt evaluation with high levels of correlation with human judgments. Proceedings of ACL-WMT, pages 65-72.  
+Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. Scibert: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3613-3618.  
+Shenshen Bu, Yujie Song, Taiji Li, and Zhiming Dai. 2024. Dynamic knowledge prompt for chest x-ray report generation. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, LREC/COLING, pages 5425-5436. ELRA and ICCL.  
+Zhihong Chen, Yaling Shen, Yan Song, and Xiang Wan. 2021. Cross-modal memory networks for radiology report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 5904-5914.  
+Zhihong Chen, Yan Song, Tsung-Hui Chang, and Xi-ang Wan. 2020. Generating radiology reports via
+
+memory-driven transformer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 1439-1449.  
+Zhixuan Chen, Luyang Luo, Yequan Bie, and Hao Chen. 2024. Dia-llama: Towards large language model-driven CT report generation. CoRR, abs/2403.16386.  
+Sasank Chilamkurthy, Rohit Ghosh, Swetha Tanamala, Mustafa Biviji, Norbert G. Campeau, Vasantha Kumar Venugopal, Vidur Mahajan, Pooja Rao, and Prashant Warier. 2018. Development and validation of deep learning algorithms for detection of critical findings in head ct scans. arXiv preprint arXiv:1803.05854.  
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.  
+Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. 2024. The platonic representation hypothesis. arXiv preprint arXiv:2405.07987.  
+Haibo Jin, Haoxuan Che, Yi Lin, and Hao Chen. 2024. Promptmrg: Diagnosis-driven prompts for medical report generation. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pages 2607-2615.  
+Baoyu Jing, Zeya Wang, and Eric Xing. 2020. Show, describe and conclude: On exploiting the structure information of chest x-ray reports. arXiv preprint arXiv:2004.12274.  
+Baoyu Jing, Pengtao Xie, and Eric P. Xing. 2018. On the automatic generation of medical imaging reports. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 2577-2586.  
+Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. 2023. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026.  
+Jonathan Krause, Justin Johnson, Ranjay Krishna, and Li Fei-Fei. 2017. A hierarchical approach for generating descriptive image paragraphs. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3337-3345.
+
+C Li, C Wong, S Zhang, et al. 2023a. Llava-med: training a large language-and-vision assistant for biomedicine in one day. arxiv. arXiv preprint arXiv:2306.00890.  
+Christy Y. Li, Xiaodan Liang, Zhiting Hu, and Eric P. Xing. 2019. Knowledge-driven encode, retrieve, paraphrase for medical image report generation. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 6666-6673.  
+Mingjie Li, Bingqian Lin, Zicong Chen, Haokun Lin, Xiaodan Liang, and Xiaojun Chang. 2023b. Dynamic graph enhanced contrastive learning for chest x-ray report generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3334-3343.  
+Yuan Li, Xiaodan Liang, Zhiting Hu, and Eric P Xing. 2018. Hybrid retrieval-generation reinforced agent for medical image report generation. Advances in neural information processing systems, 31.  
+Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.  
+Che Liu, Zhongwei Wan, Yuqi Wang, Hui Shen, Haozhe Wang, Kangyu Zheng, Mi Zhang, and Rossella Arcucci. 2024. Benchmarking and boosting radiology report generation for 3d high-resolution medical images. arXiv preprint arXiv:2406.07146.  
+Fenglin Liu, Shen Ge, and Xian Wu. 2021a. Competence-based multimodal curriculum learning for medical report generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 3001-3012.  
+Fenglin Liu, Xian Wu, Shen Ge, Wei Fan, and Yuexian Zou. 2021b. Exploring and distilling posterior and prior knowledge for radiology report generation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 13753-13762.  
+Fenglin Liu, Changchang Yin, Xian Wu, Shen Ge, Yuexian Zou, Ping Zhang, and Xu Sun. 2021c. Contrastive attention for automatic chest x-ray report generation. arXiv preprint arXiv:2106.06965.  
+AI Meta. 2024. Introducing meta llama 3: The most capable openly available llm to date. Meta AI.  
+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, July 6-12, 2002, Philadelphia, PA, USA, pages 311-318.
+
+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, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763.  
+Qingya Shen, Yanzhao Shi, Xiaodan Zhang, Junzhong Ji, Ying Liu, and Huimin Xu. 2024. Ghcl: Gaussian heuristic curriculum learning for brain ct report generation. Multimedia Systems, 30(2):69.  
+Yanzhao Shi, Junzhong Ji, Xiaodan Zhang, Ying Liu, Zheng Wang, and Huimin Xu. 2024. Prior tissue knowledge-driven contrastive learning for brain CT report generation. Multim. Syst., 30(2):98.  
+Yanzhao Shi, Junzhong Ji, Xiaodan Zhang, Liangqiong Qu, and Ying Liu. 2023. Granularity matters: Pathological graph-driven cross-modal alignment for brain CT report generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP, pages 6617-6630.  
+Yuhao Tang, Haichen Yang, Liyan Zhang, and Ye Yuan. 2024. Work like a doctor: Unifying scan localizer and dynamic generator for automated computed tomography report generation. *Expert Systems with Applications*, 237:121442.  
+Omkar Thawakar, Abdelrahman M. Shaker, Sahal Shaji Mullappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman H. Khan, Jorma Laaksonen, and Fahad Shahbaz Khan. 2023. Xraygpt: Chest radiographs summarization using medical vision-language models. CoRR, abs/2306.07971.  
+Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748.  
+Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566-4575.  
+Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2015. Show and tell: A neural image caption generator. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3156-3164.  
+Jun Wang, Abhir Bhalerao, and Yulan He. 2022a. Cross-modal prototype driven network for radiology report generation. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 563-579. Springer.  
+Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving text embeddings with large language models. ArXiv, abs/2401.00368.
+
+Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, and Ronald M. Summers. 2018. Tienet: Text-image embedding network for common thorax disease classification and reporting in chest x-rays. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 9049-9058.  
+Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. 2022b. Medclip: Contrastive learning from unpaired medical images and text. arXiv preprint arXiv:2210.10163.  
+Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron C. Courville, Ruslan Salakhutdinov, Richard S. Zemel, and Yoshua Bengio. 2015. Show, attend and tell: Neural image caption generation with visual attention. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 2048-2057.  
+An Yan, Zexue He, Xing Lu, Jiang Du, Eric Y. Chang, Amilcare Gentili, Julian J. McAuley, and Chun-Nan Hsu. 2021. Weakly supervised contrastive learning for chest x-ray report generation. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, pages 4009-4015.  
+Shuxin Yang, Xian Wu, Shen Ge, S Kevin Zhou, and Li Xiao. 2022. Knowledge matters: Chest radiology report generation with general and specific knowledge. Medical image analysis, 80:102510.  
+Sisi Yang, Junzhong Ji, Xiaodan Zhang, Ying Liu, and Zheng Wang. 2021. Weakly guided hierarchical encoder-decoder network for brain ct report generation. In IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021, Houston, TX, USA, December 9-12, 2021, pages 568-573. IEEE.  
+Di You, Fenglin Liu, Shen Ge, Xiaoxia Xie, Jing Zhang, and Xian Wu. 2021. Aligntransformer: Hierarchical alignment of visual regions and disease tags for medical report generation. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 - 24th International Conference, Strasbourg, France, September 27 - October 1, 2021, Proceedings, Part III, pages 72-82.  
+Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. 2023. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737.  
+Xiaodan Zhang, Sisi Yang, Yanzhao Shi, Junzhong Ji, Ying Liu, Zheng Wang, and Huimin Xu. 2023. Weakly guided attention model with hierarchical interaction for brain ct report generation. Computers in Biology and Medicine, page 107650.  
+Yixiao Zhang, Xiaosong Wang, Ziyue Xu, Qihang Yu, Alan L. Yuille, and Daguang Xu. 2020. When radiology report generation meets knowledge graph.
+
+In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 12910-12917.

2024/See Detail Say Clear_ Towards Brain CT Report Generation via Pathological Clue-driven Representation Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca846c2fcc1aec3fa02bb76930da2baf3a4da643aa2be480dc43906b1011fcc2
+size 606626

2024/Seeing Through VisualBERT_ A Causal Adventure on Memetic Landscapes/3b29f9b0-75b7-4106-bc6f-27cbeed85edf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f0c7a3392dd024c90236fe7ea4f0de83270197445f1a3430ea938038c61cbef
+size 4716990

2024/Seeing Through VisualBERT_ A Causal Adventure on Memetic Landscapes/full.md

@@ -0,0 +1,473 @@
+# Seeing Through VisualBERT: A Causal Adventure on Memetic Landscapes
+
+Dibyanayan Bandyopadhyay  
+Indian Institute of Technology Patna  
+Patna, Bihar, India  
+dibyanayan_2321cs14@iitp.ac.in
+
+Mohammed Hasanuzzaman  
+EEECS, Queen's University Belfast  
+Belfast, BT7 1NN, UK  
+m.hasanuzzaman@qub.ac.uk
+
+Asif Ekbal  
+Indian Institute of Technology Jodhpur  
+Jodhpur, Rajasthan, India  
+asif@iitp.ac.in
+
+# Abstract
+
+Detecting offensive memes is crucial, yet standard deep neural network systems often remain opaque. Various input attribution-based methods attempt to interpret their behavior, but they face challenges with implicitly offensive memes and non-causal attributions. To address these issues, we propose a framework based on a Structural Causal Model (SCM). In this framework, VisualBERT is trained to predict the class of an input meme based on both meme input and causal concepts, allowing for transparent interpretation. Our qualitative evaluation demonstrates the framework's effectiveness in understanding model behavior, particularly in determining whether the model was right due to the right reason, and in identifying reasons behind misclassification. Additionally, quantitative analysis assesses the significance of proposed modelling choices, such as de-confounding, adversarial learning, and dynamic routing, and compares them with input attribution methods. Surprisingly, we find that input attribution methods do not guarantee causality within our framework, raising questions about their reliability in safety-critical applications<sup>1</sup>. The project page is at: https://newcodevelop.github.io/causality_adventure/
+
+# 1 Introduction
+
+Memes have evolved from spreading humor to being used for disseminating offensive content, necessitating the development of neural multimodal systems to detect such content (Kiela et al., 2021). However, these systems often lack transparency, undermining public trust in real-world applications
+
+![](images/174b624632b634349f566b76355a62ee783d76f03536dc2d2b3c09cbb2cb293e.jpg)  
+Figure 1: The Underlying notion of these memes is racism, which is implicit. None of the input attribution methods could decipher the notion of racism solely through input attribution.
+
+![](images/66e0c6bf517f4660d114a6f1b9a0c1aec28e1d7ef979c37930b31008c781a1cb.jpg)
+
+and highlighting the need for interpretability and trustworthiness.
+
+While Large Language Models (LLMs) and Vision Language Models (VLMs) could predict offensive memes and provide self-explanations, these explanations are not always faithful to model behavior (Madsen et al., 2024; Agarwal et al., 2024). Our self-consistency checks (\$Appendix B) confirm this issue for offensive meme detection, prompting us to focus on enhancing the reliability of existing multimodal classifiers. Current interpretability techniques, such as input attributions (e.g., Integrated Gradient (Sundararajan et al., 2017)), struggle with implicit content and causality. For implicitly offensive memes, attribution methods fail to capture underlying concepts like 'racism' (refer to Figure 1). Also, they only detect influential features without considering their causal impact (Chattopadhyay et al., 2019).
+
+Causality-based techniques like CausaLM (Feder et al., 2021) and Amnesic Probing (Elazar et al., 2021) offer solutions but face scalability issues and focus on global rather than local explanations like: 'Is adjective important for sentiment analysis' or 'Is part-of-speech information crucial for word prediction?' We
+
+address these limitations with a novel causal framework integrating VisualBERT (Li et al., 2019) with causal concepts. Our approach extracts implicit context from the meme as a set of causal concepts and uses dynamic routing and adversarial learning to predict meme offensiveness based on both meme content and contribution from causal concepts.
+
+Quantitative analysis (Section 6) shows that traditional interpretability techniques, which rely on correlation, do not always align with causality, emphasizing that correlation does not imply causation. Through ablation studies, we demonstrate that our framework bases its predictions on relevant causal concepts, enhancing trustworthiness. Qualitative analysis (Section 7) indicates whether the model's decisions are justifiable and explains its error cases.
+
+Our proposed framework is novel, model-agnostic, and acts as a proof-of-concept which demonstrates the potential of using causal analysis to elucidate the decision-making process of multimodal classifiers.
+
+# 2 Related Work
+
+Causal Interpretability: Causal interpretability aims to understand how counterfactuals cause model outputs to change, thus estimating the causal effect of inputs (Feder et al., 2022). A subfield, causal mediation analysis, explores the mechanisms behind these effects (Geiger et al., 2021; Vig et al., 2020; Meng et al., 2023). Generating exact counterfactuals is challenging (Abraham et al., 2022; Calderon et al., 2022), so recent work focuses on approximations (Geiger et al., 2021) or counterfactual representation (Feder et al., 2021; Elazar et al., 2021; Ravfogel et al., 2021). Our current research concentrates on counterfactual representation. Most of the existing works target single modality (e.g. text or vision) (Feder et al., 2021; Goyal et al., 2020) and answer global questions about feature importance (Elazar et al., 2021). We propose a method for answering local questions about specific concepts (e.g., 'Is the meme offensive due to the presence of Holocaust as a concept?') while addressing scalability issues of prior methods (Feder et al., 2022). Our framework incorporates concept annotations and integrates with VisualBERT for trustworthy local causal interpretability.
+
+Multimodal Interpretability. Recently, there has been a surge in multimodal models for various
+
+tasks (Ding et al., 2021; Du et al., 2022; Li et al., 2023; Liu et al., 2023; Zhu et al., 2023), yet research on generating explanations for their predictions remains limited. Researchers primarily rely on interpretability techniques like LIME (Ribeiro et al., 2016) and SHAP (Lundberg and Lee, 2017a) and various input attribution methods (Sundararajan et al., 2017; Lundberg and Lee, 2017b; Shrikumar et al., 2017). However, recently, there has been a shift towards generating natural language explanations, bridging the gap between text-only and multimodal systems. Methods like NLX-GPT (Sammani et al., 2022) and Semantify (Bandyopadhyay et al., 2024) offer solutions but fail to fully capture implicit causal meanings or the causal impact of input features (Chattopadhyay et al., 2019). This gap motivated us to develop a framework that enables causal interpretations of implicit inputs.
+
+# 3 Causal Process
+
+# 3.1 Causal Diagram
+
+Our framework is based on a Structural Causal Model (SCM) that integrates both the causal explanation process and multimodal classification objectives, drawing inspiration from Geiger et al. (2021). We assume an exogenous variable  $E_{1}$  that generates causal concepts  $c_{1},\ldots$  to  $c_{n}$ . Another exogenous variable  $E_{2}$  controls meme text  $t$  and image  $v$  representation. The collection of the concepts  $\{c_i\}_{i = 1}^n$  controls the latent representation  $L$ .  $(t,v)$  along with  $L$  controls the intermediate representation  $I$ , which further controls the output  $y$  of the model. Figure 2 represents this in details.
+
+Counterfactual Representation. Inspired by CausaLM (Feder et al., 2021), we want to intervene  $I$  such that the generated counterfactual representation  $I^{CF_i}$  is insensitive to concept  $c_{i}$  and similar to  $I$  for all the other concepts except  $c_{i}^{2}$ .
+
+To achieve this, we observe the SCM, depicted in Figure 2, where  $I = f(L,(t,v))$  and  $L$  represents the latent, a weighted sum of concept representations  $c_{i}$ , given by  $L = \sum_{i=1}^{n} w_{i} \cdot c_{i}$ . Creating a counterfactual latent  $(L^{-i})$  unaffected by a specific concept  $c_{i}$  is straightforward to achieve by setting  $w_{i} = 0$ . This explicit modelling offers simple implementation and ensures  $I^{CF_{i}}$  represents a counterfactual unaffected by  $c_{i}$ , while still influenced by other concepts.
+
+Need for De-confounding. In this formulation
+
+![](images/124a36862a0cc535d755a1c06ea74e567b235f7284d25acd0c91be17d4eb7f35.jpg)  
+Figure 2: Left: The causal process is illustrated by a SCM. Right: Causal intervention selectively intervenes on a concept  $c_{i}$  to nullify its effect on the model. This generates the intermediate counterfactual representation  $I^{CF_i}$ . To measure the causal effect of concept  $c_{i}$ , we take the Individual treatment effect (ITE) as  $|y' - y|$ . Dotted blue lines denote that meme content representations  $(t,v)$  (along with  $E_{1}$ ) generate causal concepts  $c_{i}$ . X demonstrates causal intervention, i.e. breaking the link between  $(t,v)$  and  $c_{i}$ , realized by setting  $w_{i} = 0$ .
+
+of obtaining the counterfactual representation, a challenge arises because concept  $c_{i}$  might be correlated with other concepts  $\{c_{j}\}$  (e.g. 'holocaust' is correlated to ' Nazism'), such that  $L^{-i}$  can be retrieved from  $\{c_{j}\}$ . We term these influencing concepts like  $\{c_{j}\}$  as 'control concepts' and  $c_{i}$  as the 'treated concept', following the terminology in Feder et al. (2021). To address this challenge, we propose a novel step called de-confounding (refer to Section 4.3). Here, we constrain our framework to prevent the recovery of a treated concept  $c_{i}$  from the counterfactual latent representation  $L^{-i}$ , which is essentially a linear combination of potential control concepts. The intuition is that once a concept is removed from the representation containing it, it cannot be recovered.
+
+Causal Effect of Concept  $c_{i}$ . Drawing inspiration from existing literature (Feder et al., 2021), we formulate the causal effect of concept  $c_{i}$  (for a specific input) as:
+
+$$
+\begin{array}{l} \widehat {R I T E} _ {i} = <   \phi \left(I _ {D C} ^ {C F _ {i}}\right) - \phi \left(I _ {D C}\right) > \\ = <   \phi (f ((t, v), L _ {D C} ^ {- i})) - \phi (f ((t, v), L _ {D C})) > \tag {1} \\ \end{array}
+$$
+
+Here,  $\widehat{RITE}_i$  represents the "Representation-based Individual Treatment Effect", for  $c_i$ , with  $DC$  as subscript representing the de-confounding objective. The function  $f$ , modelled as VisualBERT in this paper, takes the input meme as text and visual representation  $((t,v))$  and latent  $L_{DC}$ , outputting a representation  $I_{DC}$ . This intermediate representation is then passed through a classifier  $\phi$ . Essentially,  $\widehat{RITE}_i$  denotes the absolute change
+
+in predicted class probability due to the absence of concept  $c_{i}$  and could be used to measure its causal effect on the model.
+
+# 3.2 Concept Annotation
+
+Figure 2 illustrates the causal process, which relies on the integration of concepts alongside meme inputs to facilitate model predictions. The selection of these concepts (which form the 'concept set') is pivotal, aiming to i) encapsulate the breadth of themes present within the training dataset while ii) minimizing redundancy.
+
+Scalability. To make the annotation process efficient and scalable, we use the following approach. Starting empty, the concept set expands as new concepts are introduced. For example, if the first meme includes 'terrorism' and 'holocaust', these are added to the set. If the second meme includes 'terrorism' and 'racism', the set becomes 'terrorism', 'holocaust', 'racism'. Therefore, we only append new concepts to the set if they are absent, to minimize redundancy and counter overlap between similar concepts.
+
+Annotation Process. We enlisted three annotators, all postgraduate students aged 25-27 with expertise in multimodal machine learning and prior experience curating datasets published in reputable venues, to annotate the concept set. We ensured ethics and took active steps to ensure their wellbeing, as detailed in Appendix Section G. Let us denote the three annotated concept sets as  $\{c_1\}$ ,  $\{c_2\}$ , and  $\{c_3\}$ . We then calculate concept representation  $r_t(c_1^i)$ , where  $c_1^i$  denotes  $i$ th concept from set  $\{c_1\}$ . Similarly, let us denote meme representation as  $r_t(T^j) \odot r_v(V^j)$ , where  $T^j$ ,  $V^j$  are the text and image of the  $j$ th meme from the test set, and  $\odot$  illustrates element-wise multiplication. Here,  $r_t(T^j) \in \mathbb{R}^{1 \times 768}$  and  $r_v(V^j) \in \mathbb{R}^{1 \times 768}$  represent CLIP (Radford et al., 2021) text and vision encodings respectively for the  $j$ th meme. We then calculate the total similarity of the set  $\{c_1\}$  to the memes in the training set as  $\sum_{j=0}^{N} \sum_{i=0}^{n} r_t(c_1^i)^T \cdot (r_t(T^j) \odot r_v(V^j))$ , where  $N$  is the number of memes in the training set and  $n$  is the number of concepts in the set  $\{c_1\}$ . Similarly, this total similarity is calculated for sets  $\{c_2\}$ , and  $\{c_3\}$ . We observe that the total similarity of set  $\{c_2\}$  is the highest which leads us to choose this as the final set. The concepts in  $\{c_2\}$  are shown in Table 1.
+
+Ambiguity Resolution. Annotators were instructed to maintain precision and leverage existing annotated concepts when annotating new memes.
+
+Despite these guidelines, ambiguities often emerged, especially when new meme concepts were semantically similar to already annotated concepts. In these instances, annotators were encouraged to engage in discussions to reach a consensus if they encountered difficulties in identifying or formulating a concept. Additionally, they were advised to utilize shared resources, such as dictionaries, to identify semantically related terms.
+
+# 4 Methodology
+
+![](images/240dd0fd4e404b07e7714d418ccbeff57ae6cc15bd3a06ecc3b2244b18687d62.jpg)  
+Figure 3: Model architecture comprising of Visual-BERT, a dynamic routing layer and a gradient reversal layer.  $0/1$ : non-offensive/offensive.
+
+Our method relies on a three-step process, namely i) Dynamic Routing, ii) Adversarial learning and iii) De-confounding.
+
+Model Inputs. A meme consists of text  $T$  and image  $I$ . We extract image features using FasterRCNN (Ren et al., 2016), yielding dimensions  $v \in \mathbb{R}^{\mathbb{B} \times \mathbb{N} \times 768}$ . Text tokens pass through the model embedding layer to generate text features with dimensions  $t \in \mathbb{R}^{\mathbb{B} \times \mathbb{M} \times 768}$ . Concatenating these text and image features  $(t, v)$  results in input dimensions of  $\mathbb{R}^{\mathbb{B} \times (\mathbb{M} + \mathbb{N}) \times 768}$  for VisualBERT. Additionally, we introduce a latent representation  $L$ , with dimensions  $\mathbb{R}^{\mathbb{B} \times 1 \times 768}$ , obtained by weighted summation of concept features  $c_i$ . Formally,  $L = \sum_{i=0}^{n} (w_d^i + w_s^i)c_i$ , where  $w_d^i$  is a dynamic weight,  $w_d^i = \rho(c_i, T)$ , dependent on the  $i$ th concept  $c_i$  and the meme text  $T$ . This functional relationship is learned through dynamic routing. Similarly,  $w_s^i$  are static weights,  $w_s^i = \tau(c_i)$ , dependent only on the concept  $c_i$ .
+
+# 4.1 Dynamic Routing.
+
+Need for Dynamic Routing. The key idea behind dynamic routing is to learn dynamic weights that determine the importance of each concept for a
+
+prediction based on both the meme input and the concept itself. These weights are functions of the meme's text, and concept, serving as learnable parameters that control each concept's influence on the prediction. Without dynamic routing, static weights are used, treating all concepts equally, regardless of the specific meme input.
+
+# How is it achieved?
+
+To model the interaction between the  $M$  text features  $\{t_j\}_{j = 0}^M$  produced by VisualBERT for a given input  $T$  and the concept feature  $c_{i}$ , we learn a weight  $W$ , which modifies  $t_j$  as  $t_{ij} = W_{ij}\cdot t_j$ . Further, the interaction between  $c_{i}$  and  $t_j$  can be modelled by taking a dot product between them.  $p_{ij} = t_j^T\cdot c_i$  demonstrates this operation, where  $p_{ij}$  is a scalar. To normalize  $p_{ij}$  between 0 and 1, we use the softmax function:  $b_{ij} = \frac{\exp(p_{ij})}{\sum_{k = 0}^{n}\exp(p_{ik})}$ , where  $n$  is the number of concepts. To measure the effect of all text inputs on concept  $i$ , we calculate the weighted mean:  $s_i = \frac{1}{M}\sum_{j = 0}^{M}b_{ij}\cdot t_{ij}$  where  $M$  is the number of text tokens input to the model. Here,  $s_i$  shows the cumulative effect of all the text inputs on concept  $c_{i}$ . To model the interaction between  $c_{i}$  and  $T$ , we want the length of  $s_i$  to approximate their interaction. Specifically, longer vectors should have a unit length, whereas shorter vectors should have a length close to zero. This is achieved using the following squashing function (Sabour et al., 2017):
+
+$$
+v _ {i} = \operatorname {s q u a s h} \left(s _ {i}\right) = \frac {\left\| s _ {i} \right\| ^ {2}}{1 + \left\| s _ {i} \right\| ^ {2}} \cdot \frac {s _ {i}}{\left\| s _ {i} \right\|} \tag {2}
+$$
+
+The length of  $v_{i}$  acts as the dynamic weight between  $c_{i}$  and  $T_{i}$ , such that  $w_{d}^{i} = \rho (c_{i},T) = ||v_{i}||$ .
+
+# 4.2 Adversarial Learning
+
+Notation. Let  $m_i$  be a  $\mathbb{R}^{1 \times 768}$  dimensional vector output corresponding to the '[CLS]' token when VisualBERT processes the text  $T$  and image  $I$ . Similarly, let  $\hat{m}_i$  be the output vector when VisualBERT processes the text  $T$ , image  $I$ , and latent  $L$  (by concatenating with image representation). For classifying an input meme, we utilize a feed-forward neural network (FFN) on the representation  $\hat{m}_i$ .
+
+Need for Adversarial Learning. Typically, input text and images contain sufficient information to classify a meme into offensive or non-offensive classes, rendering the latent representation less effective compared to text and image inputs alone. To enhance the effectiveness of the latent representation to match that of text and image inputs, we
+
+<table><tr><td>0. holocaust (535)</td><td>1. nazism (510)</td><td>2. genocide (292)</td><td>3. funny (3000)</td><td>4. anti-muslim (345)</td><td>5. terrorism (276)</td></tr><tr><td>6. violence (360)</td><td>7. politics (122)</td><td>8. racism (405)</td><td>9. international-relation (88)</td><td>10. adult (172)</td><td>11. gore (178)</td></tr><tr><td>12. misogynistic (381)</td><td>13. immigration (430)</td><td>14. extremism (431)</td><td>15. immoral (151)</td><td>16. white supremacy (205)</td><td>17. indecency (831)</td></tr></table>
+
+Table 1: Annotated concepts and the number of memes that have this concept in brackets. Note that the total number of concepts may exceed the number of memes as a meme can have multiple concepts. Annotators could see the train labels while annotating and were instructed to label non-offensive memes as funny.
+
+employ adversarial learning. The objective is to make both  $m_{i}$  and  $L$  invariant to the output class, while their combined representation  $\hat{m}_{i}$  should retain discriminatory information for classifying memes. The aim is for  $L$  to be as effective as the combined text and image inputs.
+
+How is it achieved? To achieve this, we utilize a Gradient Reversal Layer (Ganin and Lempitsky, 2015) before passing  $L$  and  $m_{i}$  to two separate classifiers for the offensiveness detection task. These classifiers help learn class-invariant  $L$  and  $m_{i}$ , while the classifier utilizing  $\hat{m}_{i}$  learns class-dependent representation.
+
+# 4.3 De-confounding
+
+The premise of de-confounding and its necessity is described in Section 3.1. Before deconfounding, we assume, that there exists some  $\overline{W}$ , which can project each  $L^{-i}$  to  $c_{i}$ . More formally,  $\overline{W} \cdot L^{CF} = C$ , where  $L^{CF} = [-L^{-1}, -L^{-2}, \dots, -L^{-n}] \in \mathbb{R}^{768 \times n}$ , and  $C = [-c_{1}, -c_{2}, \dots, -c_{n}] \in \mathbb{R}^{768 \times n}$ , where  $L^{-i}, c_{i} \in \mathbb{R}^{1 \times 768}$  and there are  $n$  concepts.
+
+For De-confounding, we aim to learn a projection matrix  $P$  (a nullspace projection matrix of  $\overline{W}$ ) such that projecting each  $c_{i}$  through it yields counterfactual latent  $L^{CF'}$  incapable of reconstructing  $C$ . The following theorem illustrates this.
+
+Theorem. If  $P$  is a nullspace projection matrix of  $\overline{W}$  and  $C' = P \cdot C$  for all  $i$ , then  $\overline{W} \cdot L^{CF'} = 0$  for all  $i$ . For Proof, refer to Appendix Section §C. After de-confounding, we denote the latent  $L$  with a subscript, i.e.,  $L_{DC}$ .
+
+# 5 Experimental setups
+
+The experimental setups and dataset details are elaborated in the Appendix Section F due to space constraints. Here we introduce various metrics and baselines.
+
+# 5.1 Simulating Model Outcome
+
+Definition. Simulatability, as defined in Hase et al. (2020), refers to how well explanations from model  $M$  help an observer (e.g. another simpler model,
+
+called a simulator model) predict outputs of  $M$ . Intuitively, better simulatability would reflect faithful explanation as there is a pattern between explanation as input and model prediction as output for the simulator to learn. In our proposed framework, explanation would mean causally sorted concepts in descending order. Sorting is necessary to preserve the causal importance order between the concepts.
+
+Notation. Let us assume an ordered set  $X_{cau}^{j} = \{x_{i}^{j}\}_{i=1}^{n}$  contains  $n$  concepts sorted by their causal attribution score (measured by  $\widehat{RITE}$  score) in a decreasing manner for the  $j$ th meme in the dataset. Similarly, let us assume an ordered set  $X_{attr}^{j}$  refers to  $n$  concepts sorted by their input attribution score by an input attribution method like Integrated Gradient.
+
+Rank Correlation. A positive correlation between these sets  $X_{cau}^{j}$  and  $X_{attr}^{j}$  indicates that highly attributed concepts also influence model outcomes (causality), while a negative correlation suggests the opposite. We measure correlation by both Kendal's (KENDALL, 1938) and Spearman's rank correlation (Spearman, 1904) criteria.
+
+Accuracy Metrics. Denoting  $\hat{x}_i^j$  as the text representation of  $x_i^j$ , the rank-adjusted³ BoW representation  $\hat{X}^j$  (note that we drop cau or attr as subscript to show a generic case) of the set  $X^j$  would be  $\frac{1}{n}\sum_{i=1}^{n}\gamma^i\hat{x}_i^j$ , where  $\gamma = 0.9$  is a positive non-zero constant. We train a simulator (a support vector machine-SVM model (Cortes and Vapnik, 1995)) on: i)  $\hat{X}^j$ , ii) concatenation of  $\hat{X}^j$  and  $\hat{m}_i^j$ , denoted by  $[\hat{X}^j;\hat{m}_i^j]$ , and iii)  $\hat{m}_i^j$ , to predict the original model prediction  $\hat{y}^j$ . The jth superscript reflects the jth meme. Intuitively, the SVM simulates the original model based on the provided information (either one of cases (i), (ii), or (iii)). The performance of the simulator can be seen in Table 2. F1 w/exp denotes the simulator performance in case (i) when only the ranked concepts  $X_{cau}$  or  $X_{attr}$  were used as simulator input. Similarly, F1 w/inp denotes case (iii), where the multimodal represent
+
+tation  $(\hat{m}_i^j)$  is used for model input. Lastly,  $F1w/$  both denotes case (ii).
+
+Comprehensiveness and Sufficiency. To measure the impact of set  $X_{cau}$  or  $X_{attr}$  on simulator performance, we employ two metrics: i) Comprehensiveness and ii) Sufficiency (DeYoung et al., 2020). Comprehensiveness quantifies the reduction in simulator model confidence when  $\hat{X}^j$  replaces  $[\hat{X}^j; \hat{m}_i^j]$  as simulator input. Denoting the simulator by  $S$ , comprehensiveness is  $S([\hat{X}^j; \hat{m}_i^j])_k - S(\hat{m}_i^j)_k$  for predicted class  $k$ . A higher comprehensiveness score indicates more importance of attribution set  $X^j$  for the simulator. Sufficiency is defined as  $S([\hat{X}^j; \hat{m}_i^j])_k - S(\hat{X}^j)_k$ . It requires a higher average comprehensiveness and lower average sufficiency score for  $X^j$  to be considered more simulatable.
+
+# 5.2 Quantifying Trustworthiness
+
+To measure the trustworthiness of a model, understanding whether its predictions originate from relevant concepts within the input is crucial, akin to assessing if the model is 'right for the right reasons.' To ascertain this, we annotate offensive memes from the test set with relevant concepts from a predefined set of 18 concepts. Subsequently, we employ averaged Precision@5 (P@5), Recall@5 (R@5), and Mean Average Precision (MAP@5) to assess the relevance of the top five concepts (w.r.t. the annotation) from both the  $X_{cau}$  and the attribution set  $X_{attr}$  obtained through various input attribution methods. Technical specifics are detailed in the Appendix Section E. Better scores in all these metrics reflect that the top attributed concepts align with human judgement, thus essentially making the model more trustworthy in return.
+
+# 5.3 Baselines
+
+We employ several standard input attribution methods to calculate attribution scores, dividing them into two groups based on their underlying mechanisms:
+
+# Path-integral based Gradient Attribution:
+
+- Integrated Gradients (IG) (Sundararajan et al., 2017): IG attributes the importance of features by integrating gradients along the path from a baseline input to the actual input. It ensures that attribution is distributed across all input features in a manner that satisfies the completeness property (§Appendix A).
+
+- DeepLIFT (Shrikumar et al., 2019): This method compares the activation of each neuron to a reference, and assigns importance to the neuron based on the comparison score.  
+- DeepLIFTSHAP (Lundberg and Lee, 2017a): A variant of DeepLIFT that aligns with SHAP values, combining both methods to compute attribution based on a cooperative game-theory approach.  
+- GradientSHAP (Ancona et al., 2018): GradientSHAP samples a point between an input-baseline pair and computes the mean gradients with respect to an output class across all such pairs.
+
+# Simple Gradient Attribution:
+
+- Saliency (Simonyan et al., 2014): It Identifies key input features by computing output gradients, highlighting features most influential to the model's prediction.  
+- Input  $\times$  Gradient (Shrikumar et al., 2017): This method computes the element-wise product of the input and its gradient to measure each feature's contribution to the prediction.
+
+We use the Captum library (https://captum.ai/) to calculate attributions for each method and apply these attributions to three matrices: i) the concept matrix  $(C\in \mathbb{R}^{18\times 768})$  , ii) the textual embedding  $(t\in \mathbb{R}^{\mathbb{B}\times \mathbb{M}\times 768})$  , and iii) the image embedding  $(v\in \mathbb{R}^{\mathbb{B}\times \mathbb{N}\times 768})$
+
+Comparison with Our Methodology. Traditional input attribution methods outlined above rank features (or concepts) by assigning scores based on their importance to the model's output. In contrast, our approach adopts a causal perspective, using the RITE score to measure how much each concept causally influences the prediction.
+
+While our causal framework differs from standard attribution techniques, we explore an intriguing link between the completeness property of path-integral methods and causality, which we discuss in more detail in Appendix A.
+
+# 6 Quantitative Findings
+
+1. Classifier performance. Model performance on the test set is  $70.36\%$  as measured by the F1 score. This reflects the VisualBERT in our framework works well and even exceeds some of the benchmark models evaluated as a part of the FB Hateful Mem dataset (Kiela et al., 2021).
+
+<table><tr><td rowspan="2">Methods</td><td colspan="2">Causal Rank Correlation</td><td colspan="2">Explainability of Simulator</td><td colspan="3">Performance of Simulator</td></tr><tr><td>Kendall&#x27;s Tau</td><td>Spearman&#x27;s rho</td><td>Comprehensiveness</td><td>Sufficiency</td><td>F1 w/ both</td><td>F1 w/ inp</td><td>F1 w/ exp</td></tr><tr><td>Int. Grad.</td><td>0.017</td><td>0.025</td><td>0.030</td><td>0.089</td><td>0.66</td><td>0.62</td><td>0.53</td></tr><tr><td>Saliency</td><td>0.859</td><td>0.916</td><td>0.005</td><td>0.116</td><td>0.63</td><td>0.62</td><td>0.42</td></tr><tr><td>DeepLift</td><td>-0.001</td><td>-0.002</td><td>0.029</td><td>0.021</td><td>0.67</td><td>0.62</td><td>0.65</td></tr><tr><td>DeepLiftSHAP</td><td>-0.003</td><td>-0.005</td><td>0.035</td><td>0.017</td><td>0.67</td><td>0.62</td><td>0.68</td></tr><tr><td>GradientSHAP</td><td>0.010</td><td>0.014</td><td>0.008</td><td>0.117</td><td>0.63</td><td>0.62</td><td>0.45</td></tr><tr><td>Input x Grad</td><td>-0.001</td><td>-0.002</td><td>0.034</td><td>0.029</td><td>0.68</td><td>0.62</td><td>0.67</td></tr><tr><td>Causal</td><td>1</td><td>1</td><td>0.002</td><td>0.112</td><td>0.63</td><td>0.62</td><td>0.46</td></tr></table>
+
+Table 2: Table shows an assessment of input attribution-based methods in two directions: i) Their rank correlation with causality, ii) Explanation capability of the extracted keywords as measured by a simulator. Note that here we do not use de-confounding for input-attribute methods as we want to measure the overall effect of these concepts on the model outcome (\$list A)  
+
+<table><tr><td>Input Attr/Causal</td><td colspan="3">Full</td><td colspan="3">w/o dyn. routing</td><td colspan="3">w/o adversarial</td><td colspan="3">w/o deconfounding</td></tr><tr><td></td><td>R@5</td><td>P@5</td><td>MAP@5</td><td>R@5</td><td>P@5</td><td>MAP@5</td><td>R@5</td><td>P@5</td><td>MAP@5</td><td>R@5</td><td>P@5</td><td>MAP@5</td></tr><tr><td>Int. Grad.</td><td>0.29</td><td>0.20</td><td>0.20</td><td>0.26</td><td>0.17</td><td>0.19</td><td>0.28</td><td>0.19</td><td>0.17</td><td>0.27</td><td>0.18</td><td>0.17</td></tr><tr><td>Saliency</td><td>0.27</td><td>0.19</td><td>0.18</td><td>0.19</td><td>0.12</td><td>0.16</td><td>0.29</td><td>0.19</td><td>0.16</td><td>0.32</td><td>0.21</td><td>0.25</td></tr><tr><td>DeepLIFT</td><td>0.30</td><td>0.21</td><td>0.21</td><td>0.27</td><td>0.18</td><td>0.17</td><td>0.26</td><td>0.18</td><td>0.16</td><td>0.26</td><td>0.18</td><td>0.18</td></tr><tr><td>DeepLiftSHAP</td><td>0.31</td><td>0.21</td><td>0.22</td><td>0.31</td><td>0.22</td><td>0.22</td><td>0.30</td><td>0.20</td><td>0.20</td><td>0.29</td><td>0.19</td><td>0.20</td></tr><tr><td>GradSHAP</td><td>0.30</td><td>0.20</td><td>0.20</td><td>0.30</td><td>0.21</td><td>0.21</td><td>0.29</td><td>0.20</td><td>0.20</td><td>0.29</td><td>0.20</td><td>0.20</td></tr><tr><td>Input x Grad</td><td>0.30</td><td>0.21</td><td>0.21</td><td>0.26</td><td>0.17</td><td>0.17</td><td>0.27</td><td>0.19</td><td>0.17</td><td>0.26</td><td>0.18</td><td>0.17</td></tr><tr><td>Causal</td><td>0.26</td><td>0.18</td><td>0.17</td><td>0.21</td><td>0.14</td><td>0.16</td><td>0.20</td><td>0.14</td><td>0.13</td><td>0.30</td><td>0.20</td><td>0.19</td></tr></table>
+
+Table 3: Red border denotes scores obtained by Causal framework, i.e. sorting keywords with their RITE score. The scores from alternative input-attribute methods are also displayed.
+
+2. Simple attribution methods align better with Causality. Table 2 presents Kendall's tau and Spearman's rho (averaged across all memes in the test set) for correlation comparison. Notably, simpler methods like Saliency show a stronger correlation with the causal set than complex methods like DeepLIFT, suggesting their potential to capture causal relationships. This suggests that simpler methods may offer a clearer and more direct understanding of causality within the model's decision-making process.
+
+3. Complex attribution methods are more simulatable. Consider the  $F1$  w/ exp scores for DeepLIFT, DeepLiftSHAP, and Input x Grad as input attribution methods (§Table 2). The attribution set  $X_{attr}$  obtained by these methods achieves  $\sim 66\%$  F1 score of the simulator, indicating a high correlation between  $X_{attr}$  and the original model prediction  $\hat{y}$  (termed as highly simulatable). Also both comprehensiveness and sufficiency scores are higher for these models.
+
+4. Correlation does not imply causation. Note the negative correlation of  $X_{attr}$  with the  $X_{cau}$  set for complex attribution methods (§Table 2), highlighting the fact that Correlation (simulation) does not always imply Causation. Additionally, observe the lower  $F1w / exp$  score obtained by the Saliency
+
+![](images/4c55bfe71ecf2cedd5309dace1f905c84c1e914d95f2de1e054e763fe09bc859.jpg)  
+Figure 4: Comparison of mean  $\widehat{RITE}$  score between w/ de-confounding and w/o de-confounding strategies
+
+attribution set, despite its higher rank correlation to the Causal set. This suggests that keywords causally related to model outcome may not always be easily simulatable. This outcome is in line with the study of Bastings et al. (2022), where authors showed that simple attribution methods are more faithful compared to complex methods. Note that although the findings match, their approach is non-causal and does not overlap with ours at all.
+
+5. Dynamic routing is paramount. Within our framework (§Table 3), without any ablation of modelling components (shown by Full), the  $X_{cau}$  set achieves 0.26, 0.18, and 0.17 scores for R@5, P@5, and MAP@5, respectively. Disabling dynamic routing (w/o dyn. routing) results in static
+
+<table><tr><td>Meme Idx</td><td>Top-5 Causal</td><td>Actual</td><td>Predicted</td><td>Possible Reason</td></tr><tr><td>32</td><td>‘anti muslim’, ‘terrorism’, ‘genocide’, ‘violence’, ‘adult’</td><td>Non Offensive</td><td>Offensive</td><td>Dataset bias</td></tr><tr><td>93</td><td>‘violence’ ‘racism’, ‘genocide’, ‘anti muslim’, ‘holocaust’</td><td>Non Offensive</td><td>Offensive</td><td>Dataset bias</td></tr><tr><td>92</td><td>‘violence’, ‘genocide’, ‘terrorism’, ‘extremism’, ‘ Nazism’</td><td>Non Offensive</td><td>Offensive</td><td>Inadequate context</td></tr><tr><td>65</td><td>‘indececy’, ‘holocaust’, ‘funny’, ‘immigration’, ‘racism’</td><td>Offensive</td><td>Non Offensive</td><td>Background knowledge</td></tr><tr><td>73</td><td>‘holocaust’, ‘violence’, ‘gore’, ‘Nazism’, ‘funny’</td><td>Offensive</td><td>Non Offensive</td><td>Modality conflict</td></tr><tr><td>76</td><td>‘misogynistic’, ‘gore’, ‘holocaust’, ‘extremism’, ‘ Nazism’</td><td>Offensive</td><td>Non Offensive</td><td>Bias and modality conflict</td></tr></table>
+
+Table 4: Meme index with corresponding causal keywords and their possible reason behind error cases.
+
+weighting of concepts irrespective of meme input, leading to reduced scores for the  $X_{cau}$  set across all metrics compared to the Full framework, indicating diminished trustworthiness.
+
+6. De-confounding is necessary for establishing correlation between causal concepts and model output. Without de-confounding, input attribution-based methods like DeepLIFTSHAP and GradientSHAP do not perform as well as they do with other configurations (e.g. w/o dyn. routing). Refer to Table 3. This suggests that without de-confounding, the model struggles to find a proper correlation between causal concepts and model output. Empirically, in Figure 4, we illustrate the mean  $\widehat{RITE}$  scores of the concepts with and without de-confounding. Without deconfounding, due to widely varied mean  $\widehat{RITE}$  scores for several concepts, certain concepts dominate the top positions in the causal set, regardless of the input, indicating potential bias in the model's causal attribution and making it difficult for input attribution methods to establish a correlation between model input and output, resulting in lower scores across metrics.
+
+# 7. Necessity of de-confounding for causality.
+
+The observation that the 'causal' set performs better without de-confounding may raise questions about its necessity. However, as illustrated in Section 3, de-confounding remains a principally valid choice. Without de-confounding, as discussed in the previous paragraph, certain concepts often dominate the top positions in the causal set, regardless of the input. Conversely, with de-confounding, all concepts have similar  $\widehat{RITE}$  values and their standard deviation, ensuring equal representation across test set examples (§Figure 4). Therefore, despite achieving lower scores, the framework with de-confounding is more trustworthy due to its consistent performance, lack of bias towards certain concepts in the causal attribution set, and equal representation of concepts across inputs.
+
+# 7 Analysing the model through Causal Lens
+
+Is the model always right due to the right reason? No, it is not always the case. We gain insight into this by examining the model causally. In Figure 6, we present two examples of memes from the test set along with the model's predictions, both correctly classified as offensive.
+
+In the first example, the meme's offensiveness is linked to the Holocaust and antisemitic ideas, accurately reflected in the Top-5 causal concepts. Notably, concepts like 'Violence', 'Holocaust', and 'Nazism' from the gold standard set directly relate to the meme's context. With two out of the Top-5 causal keywords aligning with the meme's context, we conclude that the model's prediction was correct for the right reason.
+
+In the second example, despite the meme being antisemitic, the identified causal concepts are irrelevant to antisemitism, although highly offensive. However, certain keywords (such as 'Holocaust' and 'Genocide') identified by GradientSHAP are attributed. Three of these ('Racism', 'Holocaust', and 'Genocide') directly relate to the meme's context. Initially, it may seem the model was right for the right reason. However, the absence of overlap between the set of causal concepts and the gold standard concept set indicates that the model classified the meme correctly but with erroneous causal attribution. This discrepancy suggests that although accurate input attributions exist, the model may base its decision on different causal concepts. Such analyses can help a user to trust the model predictions.
+
+# 8 Error Analysis
+
+In this section, we analyze the model's performance using causal concepts. For example, in the first row of Table 4 ( meme index 32), the model identifies 'anti-muslim' among the top 5 causal keywords, even though they are irrelevant. The presence of 'Mohammed' alone leads the model to classify the
+
+![](images/53147f4a9f46b3442669ff29788b50779ca055ec3b48e2723fb0e41ef1b909f9.jpg)  
+Figure 5: Memes of Table 4.
+
+![](images/47870058272940d385bec14e27a7053cc17d4d187f73716968a278a9d479c06e.jpg)  
+Figure 6: Evaluating trustworthiness of the model
+
+meme as offensive, indicating a dataset-wide bias towards words like 'Mohammed'. Similarly for the second meme, the mere presence of the concept of 'blackness' was sufficient for classifying the meme as offensive despite having nothing related to racism in it. This also shows dataset bias where lots of racists  $(\sim 47.1\%)$  and religious memes  $(\sim 39.3\%)$  are present.
+
+Similarly, in the third meme, the word 'jew' prompts the model to associate the meme with concepts like 'violence', and 'genocide' possibly due to insufficient visual context and association of the word jew in antisemitic offensive memes.4
+
+In the fourth meme, although offensive, the model finds it humorous due to a lack of background knowledge, especially regarding the wordplay on SpongeBob, and the smiling SpongeBob face may further contribute to misclassification.
+
+The fifth meme's misclassification can be attributed to a humorous background image unrelated to the meme text, creating a modality conflict. Note that the fourth and fifth memes which are classified as non-offensive have the 'funny' keyword as a part of the Top 5 causal keywords.
+
+The sixth meme shows a sexually offensive re
+
+mark but due to conflicting visual modality (showing violence), this meme got misclassified as non offensive. The keywords generated can also be attributed to dataset bias as shown in the first two memes.
+
+By examining causal keywords alongside model inputs, this type of error analysis offers insights into why the model made mistakes.
+
+# 9 Conclusion
+
+In this paper, we introduce a multimodal causal framework aimed at transparently analyzing VisualBERT predictions. Guided by an SCM, the framework compels VisualBERT to base its decisions on interpretable human-understandable concepts. Evaluation on the Facebook hateful meme dataset reveals key insights: i) Input attribution methods may lack causal underpinning, ii) Modelling choices significantly influence relevant causal attributions, enhancing model trustworthiness. The qualitative analysis delves into whether the model is 'right for the right' reasons and uncovers causal factors behind misclassifications.
+
+The simplicity and versatility of our framework (i.e. the underlying Structural Causal Model and its translation to modelling choices) allow its application across various tasks and multimodal models. Although we show the importance of our architecture on meme offensive detection tasks as a testbed, its application may be important in medicine where the need for trustworthy systems is paramount.
+
+# Limitations
+
+While our approach demonstrates promising results, there are some limitations to consider. Firstly, the reliance on a specific dataset, such as the Facebook Hateful Mem dataset, and a specific model, like VisualBERT, may limit the generalizability of our findings to other datasets and models.
+
+Secondly, the concept annotation process introduces challenges as it relies on human annotators
+
+to define and refine the concept set. This process may introduce subjectivity and biases. To address this challenge, employing more robust annotation guidelines, inter-annotator agreement assessments, and sensitivity analyses can enhance the reliability of the concept annotation process.
+
+At the outset, the concept of the paper acts as a seed or proof of concept, further generalizability of which is to be explored through a chain of related future studies. Specifically, exploring potential applications of the framework in other domains beyond meme classification would be valuable. The framework could be applied in areas such as content moderation, sentiment analysis, and trend analysis in social media, news media, marketing, and public opinion research and medicine.
+
+# Ethical Declaration
+
+We acknowledge the potential for misuse of our annotated concepts, which could be employed to filter memes based on racial prejudices. To mitigate this risk, human moderation and intervention are crucial. The purpose of annotating concepts is to facilitate research into the analysis and understanding of offensive memes on the internet. When used appropriately, we believe it serves as a valuable resource.
+
+Further the involvement of annotators to annotate potentially triggering meme may seem problematic. On the other hand, we completely ensure annotators' well-being by making voluntary free session with institutional counsellors available at any time. Also, participation in this process was purely at their own wish and they have been warned on exposing themselves to various offensive and trigerring contents which were marked as a disclaimer. We followed four broad ethical principles during the annotation process: i) Annotators were fully briefed on the nature of the task and provided informed consent to participate. ii) Annotators had access to psychological support via our institutional counseling system. iii) Annotators were compensated fairly in line with institute regulations. iv) The privacy and confidentiality of student participants were strictly protected throughout the study.
+
+By adhering to these protocols, we ensured that the ethical concerns associated with using students to label offensive memes were adequately addressed, prioritizing their well-being and ethical treatment. Our study underwent evaluation and approval by our Institutional Review Board (IRB)
+
+before proceeding for either annotating offensive memes in the first place or using students to annotate these memes.
+
+# Acknowledgements
+
+Dibyanayan Bandyopadhyay acknowledges Prime Minister's Research Fellowship (PMRF). The authors further acknowledge anonymous reviewers for their useful feedback.
+
+# References
+
+Eldar David Abraham, Karel D'Oosterlinck, Amir Feder, Yair Ori Gat, Atticus Geiger, Christopher Potts, Roi Reichart, and Zhengxuan Wu. 2022. Gebab: Estimating the causal effects of real-world concepts on nlp model behavior. Preprint, arXiv:2205.14140.  
+Chirag Agarwal, Sree Harsha Tanneru, and Himabindu Lakkaraju. 2024. Faithfulness vs. plausibility: On the (un)reliability of explanations from large language models. Preprint, arXiv:2402.04614.  
+Marco Ancona, Enea Ceolini, Cengiz Oztireli, and Markus Gross. 2018. Towards better understanding of gradient-based attribution methods for deep neural networks. Preprint, arXiv:1711.06104.  
+Dibyanayan Bandyopadhyay, Asmit Ganguly, Baban Gain, and Asif Ekbal. 2024. Semantify: Unveiling memes with robust interpretability beyond input attribution. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 6189-6197. International Joint Conferences on Artificial Intelligence Organization. Main Track.  
+Jasmijn Bastings, Sebastian Ebert, Polina Zablotskaia, Anders Sandholm, and Katja Filippova. 2022. "will you find these shortcuts?" a protocol for evaluating the faithfulness of input salience methods for text classification. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 976-991, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Nitay Calderon, Eyal Ben-David, Amir Feder, and Roi Reichart. 2022. DoCoGen: Domain counterfactual generation for low resource domain adaptation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7727-7746, Dublin, Ireland. Association for Computational Linguistics.  
+Aditya Chattopadhyay, Piyushi Manupriya, Anirban Sarkar, and Vineeth N Balasubramanian. 2019. Neural network attributions: A causal perspective. Preprint, arXiv:1902.02302.  
+Corinna Cortes and Vladimir Vapnik. 1995. Support-vector networks. Machine learning, 20(3):273-297.
+
+Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C. Wallace. 2020. ERASER: A benchmark to evaluate rationalized NLP models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4443-4458, Online. Association for Computational Linguistics.  
+Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. 2021. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835.  
+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.  
+Yanai Elazar, Shauli Ravfogel, Alon Jacovi, and Yoav Goldberg. 2021. Amnesic probing: Behavioral explanation with amnesic counterfactuals. Preprint, arXiv:2006.00995.  
+Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, and Diyi Yang. 2022. Causal inference in natural language processing: Estimation, prediction, interpretation and beyond. Transactions of the Association for Computational Linguistics, 10:1138-1158.  
+Amir Feder, Nadav Oved, Uri Shalit, and Roi Reichart. 2021. CausaLM: Causal model explanation through counterfactual language models. Computational Linguistics, 47(2):333-386.  
+Yaroslav Ganin and Victor Lempitsky. 2015. Unsupervised domain adaptation by backpropagation. Preprint, arXiv:1409.7495.  
+Atticus Geiger, Hanson Lu, Thomas Icard, and Christopher Potts. 2021. Causal abstractions of neural networks. Preprint, arXiv:2106.02997.  
+Yash Goyal, Amir Feder, Uri Shalit, and Been Kim. 2020. Explaining classifiers with causal concept effect (cace). Preprint, arXiv:1907.07165.  
+Peter Hase, Shiyue Zhang, Harry Xie, and Mohit Bansal. 2020. Leakage-adjusted simulatability: Can models generate non-trivial explanations of their behavior in natural language? In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4351-4367, Online. Association for Computational Linguistics.  
+M. G. KENDALL. 1938. A NEW MEASURE OF RANK CORRELATION. Biometrika, 30(1-2):81-93.
+
+Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. 2021. The hateful memes challenge: Detecting hate speech in multimodal memes. Preprint, arXiv:2005.04790.  
+Diederik P. Kingma and Jimmy Ba. 2017. Adam: A method for stochastic optimization. Preprint, arXiv:1412.6980.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. Preprint, arXiv:2301.12597.  
+Lianian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. 2019. Visualbert: A simple and performant baseline for vision and language. Preprint, arXiv:1908.03557.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. Preprint, arXiv:2304.08485.  
+Scott Lundberg and Su-In Lee. 2017a. A unified approach to interpreting model predictions. *Preprint*, arXiv:1705.07874.  
+Scott M. Lundberg and Su-In Lee. 2017b. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 4768-4777, Red Hook, NY, USA. Curran Associates Inc.  
+Andreas Madsen, Sarath Chandar, and Siva Reddy. 2024. Are self-explanations from large language models faithful? Preprint, arXiv:2401.07927.  
+Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2023. Locating and editing factual associations in gpt. Preprint, arXiv:2202.05262.  
+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. Preprint, arXiv:2103.00020.  
+Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guarding protected attributes by iterative nullspace projection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237-7256, Online. Association for Computational Linguistics.  
+Shauli Ravfogel, Grusha Prasad, Tal Linzen, and Yoav Goldberg. 2021. Counterfactual interventions reveal the causal effect of relative clause representations on agreement prediction. In Proceedings of the 25th Conference on Computational Natural Language Learning, pages 194-209, Online. Association for Computational Linguistics.
+
+Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. 2016. Faster r-cnn: Towards real-time object detection with region proposal networks. Preprint, arXiv:1506.01497.  
+Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "why should i trust you?": Explaining the predictions of any classifier. Preprint, arXiv:1602.04938.  
+Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. 2017. Dynamic routing between capsules. Preprint, arXiv:1710.09829.  
+Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. 2022. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. Preprint, arXiv:2203.05081.  
+Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2019. Learning important features through propagating activation differences. Preprint, arXiv:1704.02685.  
+Avanti Shrikumar, Peyton Greenside, Anna Shcherbina, and Anshul Kundaje. 2017. Not just a black box: Learning important features through propagating activation differences. Preprint, arXiv:1605.01713.  
+Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. 2014. Deep inside convolutional networks: Visualising image classification models and saliency maps. Preprint, arXiv:1312.6034.  
+C. Spearman. 1904. The proof and measurement of association between two things. *The American Journal of Psychology*, 15(1):72-101.  
+Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. Preprint, arXiv:1703.01365.  
+Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Simas Sakenis, Jason Huang, Yaron Singer, and Stuart Shieber. 2020. Causal mediation analysis for interpreting neural nlp: The case of gender bias. Preprint, arXiv:2004.12265.  
+Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592.
+
+# A Completeness vs. Causality
+
+Completeness is a property inherent to path-integral-based methods. Consider a concept defined by  $c_{i}$ , where  $C = \{c_i\}_{i=1}^n$  and  $n = 18$  represents the number of concepts. The completeness property asserts that the attribution for  $c_{i}$ , denoted as  $\text{Attr}_i$ , when calculated separately from other concepts, is equal to  $F(c_{i}, (t,v)) - F(c_{i}', (t,v))$ . Here,  $F$  is the VisualBERT combined with the classifier,  $(t,v)$  represents other unchanged model
+
+inputs (mime content), and  $c_{i}^{\prime}$  denotes the baseline, often chosen as a zero vector.
+
+$$
+\operatorname {A t t r} _ {i} = F \left(c _ {i}, o\right) - F \left(c _ {i} ^ {\prime}, o\right) \tag {3}
+$$
+
+In this context,  $F(x) = (\phi \circ f)(x)$ . Choosing  $c_{i}$  as the zero vector makes the attribution exactly equal to the  $\widehat{RITE}_{i}$  score for concept  $c_{i}$  (see Equation 1).
+
+Using approximations (e.g., Gauss-Legendre approximation) in calculating the path integral results in  $Attr_{i} \approx \widehat{RITE}_{i}$ , implying that  $X_{cau} = X_{attr}$ . This has an interesting implication:
+
+1.  $X_{cau} = X_{attr}$  suggests that path-integral-based attribution methods exhibit causal behavior, which is a highly desirable property.
+
+However, achieving this property requires certain assumptions and modeling choices, outlined below:
+
+1. The background SCM must be chosen as a modeling choice, resulting in the same outcome as using a concept as a zero vector and setting its corresponding static and dynamic weights to zero.  
+2. We must assume that individual concept attributions are calculated separately, independent of other concepts, with separate zero-baselines for each concept. This is not the usual practice, as attributions are typically calculated with text, vision, and concepts considered as three singular inputs to the model with random baselines. Consequently, all concept embeddings  $(C = \{c_i\}_{i=1}^n)$  are treated as a single input for attribution calculation.
+
+# B LLMs/VLMs Self-consistency Checks
+
+Self-consistency checks. Self-consistency checks (Madsen et al., 2024) are methods used to verify the faithfulness of a model's explanation by ensuring that the model's behavior aligns with expected outcomes. In the context of counterfactual explanations, these checks involve modifying the input so that the model predicts the opposite label. By re-evaluating the prediction with this modified input, we can confirm whether the model produces the expected opposite outcome. If it does, the counterfactual explanation is considered faithful. This process is crucial when using an instruction-tuned language model in a
+
+conversational setting, where re-evaluation should occur in a new chat session to avoid bias from prior prompts.
+
+LLM self-consistency checks. We define two self-consistency checks for LLMs as illustrated in Figure 8 and Figure 9. In the first session, referred to as Session 1, we query the LLM with the following template: "I have a meme which can be described as caption. The meme text reads: meme text. Is this meme offensive? Only answer Yes/No." The LLM responds with either Yes or No. We then ask the LLM to explain why the meme is offensive or non-offensive. Next, we prompt the LLM to create a counterfactual version of the meme text that would yield the opposite prediction about whether the meme is offensive or non-offensive. This counterfactual text is then embedded into the template and sent back to the LLM as a counterfactual query. The LLM should respond with the opposite label from its initial prediction, as it is a counterfactual according to the model. Figures 8 and 9 show two failure cases in these counterfactual self-consistency checks for LLMs, indicating that LLM explanations are not faithful and should not be trusted to reveal the model's inner workings, despite their plausible appearance.
+
+VLM self-consistency checks. Similar to language models, Visual Language Models (VLMs) can be assessed for self-consistency. Figure 7 presents two instances of self-consistency failures in VLMs using counterfactual explanations. Initially, the VLM is prompted to label a meme as offensive or non-offensive. Subsequently, the same VLM is asked to generate a counterfactual version of the meme text, which is then in-painted back into the meme to create a counterfactual sample. This counterfactual sample is used as a query in the following session. As a counterfactual, the original label should be reversed. However, as demonstrated in Figure 7, this label reversal does not occur in either case, indicating a failure in self-consistency checks for the VLM.
+
+# C Proof
+
+The latent representation  $L^{\prime}$  after projecting  $c_{i}$  with  $P$  is given by
+
+$$
+L ^ {\prime} = \sum_ {i = 0} ^ {n} w _ {i} P \cdot c _ {i} = P \cdot \sum_ {i = 0} ^ {n} w _ {i} c _ {i} = P \cdot L \tag {4}
+$$
+
+From Equation 4,  $\overline{W} \cdot L' = \overline{W} \cdot (P \cdot L) = 0$ , as  $\overline{W} \cdot (P \cdot x) = 0$  for all  $x$ .
+
+Furthermore,  $w_{i}\neq 0$  , and  $w_{i}\overline{W}\cdot (P\cdot c_{i}) = 0$  Therefore,  $\overline{W}\cdot L^{CF^{\prime}} = \overline{W}\cdot L^{\prime} - w_{i}\overline{W}\cdot (P\cdot c_{i}) = 0$  for all  $i$
+
+Following this theorem and its proof, we construct a nullspace projection  $P$  of  $\overline{W}$ .  $\overline{W}$  is learned by training a neural network to reconstruct  $C$  from  $L^{CF}$  for all  $i$ . To de-confound the system, we project the concept representations via  $P$ , updating  $C$  as  $C \gets P \cdot C$ . This approach is inspired by Iterative Nullspace Projection (INLP)(Ravfogel et al., 2020). After this operation, we denote the latent  $L$  with a subscript, i.e.,  $L_{DC}$ .
+
+# D FAQ
+
+1. Question: What is the need of the framework?
+
+Ans: The proposed framework incorporates a VisualBERT and forces it to base its prediction on causal concepts. This makes the VisualBERT transparent because now we can base its decision on the external causal concepts by looking at their RITE scores. As proposed in the Introduction, input memes contain a lot of implicit contexts, which cannot be attributed by Input attribution methods as they are missing from the meme input itself.
+
+2. Question: What is the need for Adversarial Learning?
+
+Adversarial Learning ensures the model utilizes the concept alongside the meme inputs. Normally, the meme inputs contain enough information to classify the input meme and simply adding concept representation on top of it does not make the multimodal information any richer. Adversarial learning ensures we learn a non-trivial interaction between multimodal representation and concept representation such that both play an equal role in the final classification task.
+
+3. Question: With De-confounding, the possibility of top-5 causal concepts being relevant decreases than without it. Then what is its use?
+
+De-confounding is principally valid because from the counterfactual representation of concept  $c_{i}$ , we should not be able to recover  $c_{i}$  (i.e. its representation). Otherwise, it is not even a counterfactual. Our main goal is to estimate an unbiased value of  $\widehat{RITE}$  score, which is obtained after deconfounding. So, although w/o de-confounding scores are higher, they may not reflect true causality as without de-confounding, true counterfactual representations are not generated. Also, an empirical argument is validated in the Section 6 and Figure 4.
+
+![](images/47f270fa2c450baaeed7ced0a1190a92f0f0c175590f1bd118b676985debafb6.jpg)
+
+![](images/8702cdadec729ea5623047e4c3107effd50a8a8fee30db7789d3e702530faa24.jpg)
+
+![](images/9912f52930eb0c5e0f080cfb05e4836e8023eacf7c40bbe38f3d273229bf9c6e.jpg)  
+Figure 7: Counterfactual self-consistency checks for VLM, LLaVA-1.6-34B (https://github.com/haotian-liu/LLaVA).
+
+![](images/914515ffe5046aacd0d6af40219244096eabdbf417c7dbe7627b7c62c0a8d0fb.jpg)
+
+4. Question: Why do you call your framework causal when the only thing 'causal' is how you select the concepts based on their RITE score?
+
+We assumed that a meme is offensive due to the meme content itself and some contribution from a fixed collection of concepts. That is illustrated as a structural causal model (SCM) in Figure 3.1. Note that every modelling choice is based on the SCM itself. Essentially the framework encompasses an SCM which implements itself as an addition to VisualBERT in the form of i) Adversarial Learning, ii) Dynamic Routing and iii) De-confounding. This forces the SCM to base its decision on meme content as well as on interpretable concepts. Essentially the concepts are called "causal" because they are being modelled as causal in the SCM and the
+
+framework implements the SCM on top of VisualBERT.
+
+5. Question: What is the generality of the proposed framework?
+
+The framework is overall generic. The framework which implements the SCM is general for every task where implicit domain knowledge is necessary. One such task is meme offensiveness detection. Although the scope of this framework is general, we use it on top of meme offensive detection tasks because of its i) generality of using nontrivial implicit context which is unavailable in input space, and ii) The task has social good applications, and therefore is important to solve. We leave it in the hands of future researchers to use this framework as a proof of concept to describe its
+
+# Session 1
+
+![](images/3c628b97f69b669da3798272680b044eb4b6c16b3f91eb55ff44878d5ab12bd4.jpg)  
+Figure 8: An instance where GPT-3.5 offers an counterfactual self-explanation and employs self-consistency verification to assess its faithfulness.
+
+effect in critical domains, such as medicine, where the trustworthiness of machine learning (ML) systems is of utmost importance.
+
+So in summary, the proposed framework is generic whose applicability is shown in one interesting task. The applicability of it to other domains remains as an avenue for further research.
+
+6. Question: In the proposed SCM (Equation 1), when a concept is intervened or masked, its information in the text and image remains, potentially affecting the reliability of the causal effects. How do you address this concern?
+
+We address this issue in two ways: a) Gradient Reversal (GR) Layer: It ensures that the text and image alone, without the explicit presence of the concept, cannot classify the meme as offensive or non-offensive. It forces the model to rely on the explicit presence of the concept for classification.
+
+b) De-confounding Strategy: It nullifies the effects of any concept on other interrelated concepts and provides a clearer analysis of each concept's individual impact on predictions.
+
+These mechanisms ensure the reliability of the causal effects derived from equation (1).
+
+7. Question: The performance of  $X_{cau}$  in "explainability of simulator" and "performance of simulator" is not good compared to  $X_{attr}$ . Is that problematic?
+
+It's important to note that evaluating the causal framework based on simulator scores can be misleading. We are using the causal framework to judge how causally important the input attribution based methods are, not the other way around. Better simulator performance of  $X_{attr}$  indicates that it has better recognizable patterns, suggesting it can explain the model's decision well. Lower scores
+
+# Session 1
+
+![](images/beb93a6c5509fd3c892c47cf69e3246cee76a9815f8a46507461ff75f8822774.jpg)  
+Figure 9: Another instance where GPT-3.5 offers an counterfactual self-explanation and employs self-consistency verification to assess its faithfulness.
+
+only on simulatability metrics indicate that  $X_{cau}$  does not exhibit patterns like  $X_{attr}$ . On the flip side, despite high explainability, input attribution methods do not necessarily capture the underlying causality as indicated by lower rank correlation among various highly simulatable input attribution methods (e.g. DeepLIFT, DeepLIFTSHAP). As noted in this paper, "Correlation does not imply causation."
+
+8. Question: Why most of the examples in the paper are related to racism? Is it because offense related to racism are easy to see?
+
+This is not the case. Most of the examples are racist or religious because they incorporate most of the memes in the dataset, comprising of  $47.1\%$  and  $39.3\%$  of memes respectively.
+
+# E Evaluation Metrics
+
+$\mathbf{P}@\mathbf{5}$ : This is defined as among the top 5 causal keywords, how many are also relevant. This can be mathematically formulated as  $\frac{n(A \cap B)}{n(B)}$ , where  $A$  is the set of relevant gold standard keywords, whereas  $B$  is the set of top 5 causal keywords, and  $n(A)$  shows the number of elements in  $A$ .  $n(B) = 5$ , as we are considering the Top 5 keywords. Consider the first example in Figure 6.  $A = \{\text{holocaust, Nazism, genocide, extremism}\}$  and  $B = \{\text{violence, Holocaust, Nazism, anti-muslim, terrorism}\}$ .  $A \cap B = \{\text{holocaust, Nazism}\}$ , which entails  $n(A \cap B) = 2$  and  $n(B) = 5$ . So the  $\mathbf{P}@\mathbf{5}$  score is  $2/5 = 0.4$ . We report the averaged  $\mathbf{P}@\mathbf{5}$  scores across all test set examples in Tables.
+
+R@5: This is defined as among the gold standard relevant keywords, how many are chosen from the top5 causal set. Following the previousnota
+
+tion, this can be defined as  $\frac{n(A\cap B)}{n(A)}$  . For the previous example,  $n(A\cap B) = 2$  and  $n(A) = 4$  , which entails R@5 is  $2 / 4 = 0.5$
+
+MAP@5: This metric is defined as a mean of  $\mathrm{P}@\mathrm{k}$  where  $\mathbf{k}$  ranges from 1 to 5.  $\mathrm{P}@\mathrm{k}$  values are calculated following the previous paragraph. Mathematically,  $\mathrm{MAP}@5 = (\mathrm{P}@1 + \mathrm{P}@2 + \mathrm{P}@3 + \mathrm{P}@4 + \mathrm{P}@5)/5$ .
+
+# F Experimental Setup
+
+Our proposed model was implemented using PyTorch, a Python-based deep-learning library. All experiments (excluding the baselines) were performed on a single Nvidia P100 16GB GPU on CuDA driver version 12.4. The baselines were run on a different CuDA version compared to the baselines (12.1 vs 12.4). We employed the Adam optimizer (Kingma and Ba, 2017) with a learning rate set to  $5e - 5$  for optimization. The experiments were conducted using the Facebook Hateful Mem dataset (Kiela et al., 2021). The whole set of 8500 memes is broken down into 7840, 660 examples as the train-test split using a random seed 42. The total number of parameters of the model is 350M and to train a single epoch, 30 mins of continuous GPU execution is needed with a fixed batch size of 32.
+
+To mitigate the impact of non-deterministic GPU operations, we used a fixed random seed of 42 across all the experiments.
+
+Note that many attribution methods, such as saliency, rely on gradient backpropagation (particularly the chain rule) to compute attribution scores. Small precision differences caused by changes in GPU hardware can lead to significant variations in these scores.
+
+# G Addressing Ethical Concerns Related to Annotation of Offensive Memes
+
+The involvement of annotators in labeling potentially triggering memes may seem problematic. However, we have implemented several measures to ensure the well-being of the annotators:
+
+Voluntary Participation and Informed Consent: Participation in the annotation process was entirely voluntary. Annotators were fully informed about the nature of the task and the potential exposure to offensive and triggering content through a detailed disclaimer. They provided their informed consent before participating.
+
+Psychological Support: We ensured that annotators had access to free, voluntary sessions with institutional counselors at any time. This support system was established to address any psychological distress that might arise from the task.
+
+Fair Compensation: Annotators were fairly compensated for their work, receiving rates according to institutional policies. This compensation was provided in addition to the fellowships that students receive, in accordance with institute regulations.
+
+Confidentiality: We maintained strict confidentiality and privacy for all student participants throughout the study.
+
+Ethical Review and Approval: Our study, including the use of students to annotate offensive memes, underwent evaluation and received approval from our Institutional Review Board (IRB). The IRB reviewed our study design, the nature of the content, and the involvement of student participants to ensure compliance with ethical standards.
+
+Use of Publicly Available Dataset: The memes included in our study were selected from the well-known, publicly available FB Hateful Meme Dataset. This dataset represents a wide range of themes found on the internet, including some offensive and extreme content. Our goal was to analyze how different types of content are perceived and categorized by labeling them with keywords, necessitating the inclusion of a broad spectrum of examples.

2024/Seeing Through VisualBERT_ A Causal Adventure on Memetic Landscapes/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40e15e766e3dd918150811c13e3714c61e5b1b026e6b725c5d0bf40924b15178
+size 769184

2024/Selection-p_ Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability/14d2ceda-5f83-49b4-88b6-f58b5179b8e9_content_list.json

@@ -0,0 +1,2000 @@
+[
+    {
+        "type": "text",
+        "text": "Selection- $p$ : Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability",
+        "text_level": 1,
+        "bbox": [
+            144,
+            89,
+            852,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Tsz Ting Chung $^{1}$  Leyang Cui $^{2}$  Lemao Liu $^{2}$  Xinting Huang $^{2}$  Shuming Shi $^{2}$  Dit-Yan Yeung $^{1}$",
+        "bbox": [
+            213,
+            145,
+            786,
+            178
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>The Hong Kong University of Science and Technology <sup>2</sup>Tencent AILab {ttchungc, nealcly.nlp, lemaoliu, shuming}@gmail.com timxthuang@tencent.com dyyeung@cse.ust.hk",
+        "bbox": [
+            196,
+            179,
+            801,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models (LLMs) have demonstrated impressive capabilities in a wide range of natural language processing tasks when leveraging in-context learning. To mitigate the additional computational and financial costs associated with in-context learning, several prompt compression methods have been proposed to compress the in-context learning prompts. Despite their success, these methods face challenges with transferability due to model-specific compression, or rely on external training data, such as GPT-4. In this paper, we investigate the ability of LLMs to develop a unified compression method that discretizes uninformative tokens, utilizing a self-supervised pre-training technique. By introducing a small number of parameters during the continual pretraining, the proposed Selection- $p$  produces a probability for each input token, indicating whether to preserve or discard it. Experiments show Selection- $p$  achieves state-of-the-art performance across numerous classification tasks, achieving compression rates of up to 10 times while experiencing only a marginal  $0.8\\%$  decrease in performance. Moreover, it exhibits superior transferability to different models compared to prior work. Additionally, we further analyze how Selection- $p$  helps maintain performance on in-context learning with long contexts.",
+        "bbox": [
+            144,
+            291,
+            460,
+            715
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            733,
+            258,
+            747
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In-context learning has shown remarkable success in various natural language processing tasks (Brown et al., 2020), such as classification task (Min et al., 2022), and mathematical reasoning task (Wei et al., 2023), enabling Large Language Models (LLMs) to tackle complex and diverse tasks using only few-shot samples. However, in-context learning also significantly extends the length of prompts, resulting in increased computational and financial costs. Recently, a line of work has been",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/1cce267e8b12d68049f13240838dcd5b6ff21bcaf32913baea6a9a43854736a3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Transferable</td><td>Single Run</td><td>-External</td></tr><tr><td>AutoCompressor</td><td>X</td><td>✓</td><td>✓</td></tr><tr><td>LLMLingua</td><td>✓</td><td>X</td><td>✓</td></tr><tr><td>LLMLingua-2</td><td>✓</td><td>✓</td><td>X</td></tr><tr><td>Selection-p</td><td>✓</td><td>✓</td><td>✓</td></tr></table>",
+        "bbox": [
+            507,
+            258,
+            887,
+            335
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Comparison between the proposed Selection- $p$  and existing content compression approaches. Selection- $p$  exhibits great transferability (Transferable), does not require multi-round iterative decoding (Single Run), and does not rely on costly external resources for training ( $\\neg$ External).",
+        "bbox": [
+            507,
+            344,
+            885,
+            431
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "focusing on prompt compression, which aims to compress the original prompts while minimizing information loss. They are categorized into discrete compression (Li et al., 2023; Jiang et al., 2023; Pan et al., 2024) and continuous compression (Mu et al., 2023; Chevalier et al., 2023; Ge et al., 2024); the former compresses the context into discrete tokens, while the latter compresses it into a short sequence of continuous vectors.",
+        "bbox": [
+            507,
+            463,
+            882,
+            607
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Observing redundant and repetitive content in a given input, discrete compression methods aim to eliminate less informative context without significantly compromising the model's performance. For example, LLM Lingua (Jiang et al., 2023) proposes to perform iterative token truncation based on the content perplexity, requiring multi-round decoding (Jiang et al., 2023). LLM Lingua-2 (Pan et al., 2024), distilled from GPT-4 (OpenAI, 2023), addresses the potential misalignment between entropy and the compression objective, as well as the distribution gap between the perplexity of the compression model and the target model. High costs are still involved in the training data construction. Meanwhile, optimizing the distribution for specific LLMs (GPT-4) may, on the other hand, hinder the transferability of the compressed content to other LLMs. More details are discussed in Section 4.4.",
+        "bbox": [
+            507,
+            611,
+            884,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Continuous compression (Bulatov et al., 2022;",
+        "bbox": [
+            527,
+            904,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Wingate et al., 2022) teaches pre-trained LMs the ability to compress text into a short sequence of continuous vectors. AutoCompressors (Chevalier et al., 2023) uses an unsupervised learning objective, which motivates the model to cache crucial information within the summary vectors. Despite their success, these methods have poor generalization as they can only compress to the length specified during training. Additionally, since the continuous vector cannot be transferred between models, a separate compressor must be trained for each model.",
+        "bbox": [
+            112,
+            84,
+            489,
+            275
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "This raises an interesting research question:",
+        "bbox": [
+            132,
+            278,
+            457,
+            292
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "\"Can LLMs learn to identify less informative tokens within a given context without external annotated signals?\"",
+        "bbox": [
+            112,
+            294,
+            487,
+            341
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To answer this question, we propose a pretraining strategy with a self-supervised signal, which enables the model to autonomously learn to predict the next token based on compressed context. With a small additional number of parameters, a forward pass on the proposed selection-  $p$  creates the probability vector  $p$  corresponding to each input token, indicating whether to preserve or discard the token. During inference, we can apply the detokenized compressed tokens to any downstream LLMs with only single-turn decoding and without reliance on any costly external resources as depicted in Table 1.",
+        "bbox": [
+            112,
+            343,
+            489,
+            549
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The main contributions of this work are fourfold:",
+        "bbox": [
+            132,
+            551,
+            487,
+            565
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We present Selection- $p$  which achieves only  $0.8\\%$  drops in performance under 10x compression rate across nine traditional classification tasks, surpassing the performance of the existing compression models. Under this setting, a speedup of 5.3x can be achieved during inference with in-context learning.",
+            "- Selection- $p$  demonstrated great transferability, which surpasses the performance of prior work in performing hard compression for both open-source and close-source models.",
+            "- We further analyze how Selection- $p$  helps in in-context learning in long-context settings, presenting a potential solution to address the performance declination of long-context models in ICL.",
+            "- We connect in-domain prior works and make comparisons with these state-of-the-art compression models, providing a complete picture."
+        ],
+        "bbox": [
+            134,
+            573,
+            487,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            665,
+            98
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Hard Compression",
+        "text_level": 1,
+        "bbox": [
+            509,
+            114,
+            705,
+            130
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Some studies focus on token pruning (Goyal et al., 2020; Kim and Cho, 2021; Rao et al., 2021; Kim et al., 2022; Modarressi et al., 2022) and token merging (Bolya et al., 2023) but they are designed primarily for smaller models like BERT. More recently, Selective Context (Li et al., 2023) is the first to propose to prune less important tokens based on information entropy. Subsequently, LLMLingua (Jiang et al., 2023) refined the approach by integrating the selection of demonstrations and the allocation of compression budgets for various segments of the input prompt. No training is required in these models but their efficacy in downstream tasks with compression applied to in-context demonstrations remains limited. Pan et al. (2024) extended the idea and addressed the potential misalignment between entropy and the compression objective, leveraging full bidirectional context by training on their proposed GPT-4 distilled compression dataset. Our simple yet effective approaches outperform previous studies.",
+        "bbox": [
+            507,
+            137,
+            885,
+            475
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Soft Compression",
+        "text_level": 1,
+        "bbox": [
+            509,
+            494,
+            694,
+            508
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Gist (Mu et al., 2023) is first proposed to compress prompt with soft tokens. Subsequently, Autocompressor (Chevalier et al., 2023) and ICAE (Ge et al., 2024) extend the idea to handle long contexts with different pretraining approaches. ICAE further conducts instruction tuning to enhance model performance. The downstream performance of these models heavily relies on the tuned compression model, with a fixed compression rate. Additionally, retraining is necessary for different versions of LLMs. Compared to these approaches, our work offers greater flexibility and transferability, while simultaneously surpassing the performance of existing compression models.",
+        "bbox": [
+            507,
+            518,
+            885,
+            743
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            509,
+            762,
+            658,
+            778
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Under the intuition that redundant texts often exist and their removal does not hinder human understanding of the text, we assume that LLMs behave in a similar manner. To efficiently identify less informative tokens within a given context, we propose a simple pre-training objective encouraging the model to predict the same token both before and after discarding less informative tokens.",
+        "bbox": [
+            507,
+            791,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/1153cb586b1ec2012e25950e87672fbf4d302723f611451a86c1d38d66563ab3.jpg",
+        "image_caption": [
+            "Figure 1: Illustration with the training process. Areas in orange are learnable parameters. For the input context  $[x_1, x_2, \\ldots, x_{n-1}]$ , inference without parameters update is performed first to create the attention mask  $\\bar{p}$ . These subsequently form the model input for LoRA training and updating the parameters of the additional linear layer."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            80,
+            487,
+            382
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            112,
+            507,
+            257,
+            524
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Language Model. Given a context  $[x_1, x_2, \\ldots, x_{n-1}]$ , the objective of the language model is to predict the next token  $x_n$ , formed as  $P(x_n | x_1, x_2, \\ldots, x_{n-1})$ . In case of training with the causal language modeling (CLM) loss, we will have,",
+        "bbox": [
+            112,
+            530,
+            489,
+            626
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L M}} = - \\sum \\log P (x _ {i} \\mid x _ {1}, x _ {2}, \\dots , x _ {i - 1}; \\theta)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            639,
+            472,
+            663
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Tokens Selection Models. LLMLingua (Jiang et al., 2023) and its variant (Li et al., 2023; Pan et al., 2024) select tokens according to the computed distribution of the targeted LLM. Specifically, suppose  $x_{i}$  is a token in the prompt, if the probability of a token  $x_{i}$  is less than a threshold, then the token is selected to be compressed.",
+        "bbox": [
+            112,
+            675,
+            489,
+            788
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Selection-  $p$",
+        "text_level": 1,
+        "bbox": [
+            112,
+            802,
+            247,
+            818
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following the tokens selection models (Li et al., 2023; Jiang et al., 2023; Pan et al., 2024), Selection $p$  includes the two steps for testing, i.e. the selection (compression) step and the inference step. In the selection step, unlike LLMLingua and its variant, we instead define a selection model to select",
+        "bbox": [
+            112,
+            824,
+            490,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "less informative tokens within the context in a discriminative way. In the inference step, the tokens selected via the selection model are passed to our targeted inference model.",
+        "bbox": [
+            507,
+            84,
+            884,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Selection. Assume  $p_i \\in [0,1]$  denote a measure of the informativeness of token  $x_i$ . Theoretically, we can customize a deep neural model to instantiate  $p_i$ . In practice, we directly take a pre-trained language model and adopt the last layer of hidden representation  $\\mathbf{h}^L$  from a pre-trained LM to the new linear projection layer. Formally, suppose  $\\mathbf{h}^L = \\{h_1, h_2, \\dots, h_n\\}$  denotes the sequence of hidden states for all tokens  $x_i$  at the last layer of a pre-trained language model. The selection model  $\\mathbf{p} = \\{p_1, p_2, \\dots, p_n\\}$  is defined as follows:",
+        "bbox": [
+            507,
+            159,
+            884,
+            336
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {p} = \\sigma \\left(\\mathbf {W h} ^ {L} + b\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            623,
+            350,
+            882,
+            368
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\sigma$  is the sigmoid function,  $\\mathbf{W}$  and  $b$  are parameters of the projection matrix and bias vector, respectively.",
+        "bbox": [
+            507,
+            382,
+            882,
+            430
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "By using the selection model  $\\mathbf{p}$ , it is straightforward to compress the context for inference: we directly prune the corresponding tokens according to our desired compression rate in retaining the top  $k\\%$  of tokens in the context. To ensure the efficiency in performing compression, a single forward pass on our model creates the token preservation probability for all input tokens for simplicity.",
+        "bbox": [
+            507,
+            432,
+            882,
+            561
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Training. Our training criterion for the selection model aims to preserve the language modeling ability of the LLM while also learning to discard tokens effectively. To this end, we employ the self-supervised approach to optimize the selection model and therefore we do not need external resources to train the selection model compared with Pan et al. (2024).",
+        "bbox": [
+            507,
+            571,
+            882,
+            699
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To keep the training process consistent with the inference process, we first discretize the selection model  $\\mathbf{p}$ . Let  $\\bar{p}_i$  denote the discretized binary model of  $p_i$ . In other words,  $\\bar{p}_i$  is 1 if it ranks the top  $k\\%$  of tokens with the highest  $p$  values and 0 otherwise. Then we use the discretized model as a mask to define the CLM loss function as follows:",
+        "bbox": [
+            507,
+            700,
+            882,
+            812
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathcal {L}} _ {\\mathrm {C L M}} = - \\sum \\log P \\left(x _ {i} \\mid \\bar {p} _ {1} x _ {1}, \\dots , \\bar {p} _ {i - 1} x _ {i - 1}; \\theta\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            510,
+            821,
+            880,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\bar{p}_i x_i$  denotes whether the token  $x_i$  is masked or not depending on the value of  $\\bar{p}_i$ , where the above language model  $P$  is set as the same model as that used in the selection model in Eq. 1.",
+        "bbox": [
+            507,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Training Details for Transferability. One of our goals is to achieve a transferable compression method. Therefore, the parameters that achieve the best loss on the language models in Eq. 1 and Eq. 2 in training may not be transferred to the targeted language model in inference since the language models in training and inference can be different. As a result, to ensure the better transferability of the optimized selection model, we freeze the pre-trained language model in Eq. 1 and employ LoRA (Hu et al., 2021) to train a partial parameter in the language model in Eq. 2. In summary, the CLM-based training loss is illustrated in Figure 1.",
+        "bbox": [
+            112,
+            84,
+            489,
+            294
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            306,
+            252,
+            322
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Setting",
+        "text_level": 1,
+        "bbox": [
+            112,
+            332,
+            216,
+            348
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We finetune a LLaMA-2-7b model (Touvron et al., 2023) on 100M tokens from RedPajama (Togeth erAI, 2023) on split segments of 1,024 tokens via LoRA (Hu et al., 2022). To provide a comprehensive analysis of the model capabilities, our evaluation is conducted on traditional classification tasks as well as the long-context classification task.",
+        "bbox": [
+            112,
+            353,
+            489,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each task, we randomly sample from the training set to construct the demonstration set for In-Context Learning (ICL), which also serves as our compression target for token selection. During inference, the compression process only needs to be computed once for all subsequent inferences on the testing instances.",
+        "bbox": [
+            112,
+            466,
+            489,
+            580
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Traditional Classification Tasks. Following Chevalier et al. (2023), we evaluate and compare different compression models on nine classification tasks, including six tasks from SuperGlue (Wang et al., 2019). The predictions by LLMs are determined by iterating through all possible answer options for the instance and selecting the option with the minimum negative log-likelihood. The in-context demonstrations have been carefully selected to approximate a size of 750 tokens, and the complete demonstration is employed. This is referred to as the \"full-shot\". Since the average token length for a single demonstration varies for different tasks (e.g., a single demonstration in RTE averages about 75 tokens, resulting in a 10-shot setup under the full-shot setting), the exact number of shots differs depending on the task.",
+        "bbox": [
+            112,
+            588,
+            489,
+            863
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To ensure a fair comparison among different compression models, a compression rate of 0.1 is used. For each task, four sets of demonstrations are selected, and the average result across these four trials is presented in Table 2. Additionally, the average accuracy of the nine tasks is presented for a clearer performance comparison.",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To achieve a compression rate of 0.1, a simpler approach is to directly retain one-tenth of the full-shot demonstration (e.g., using 1-shot instead of 10-shot for the RTE task) instead of performing selection at the token level. We also include this method as a baseline to further demonstrate the effectiveness of our approach.",
+        "bbox": [
+            507,
+            198,
+            884,
+            310
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Long Context Classification Tasks. Recent research by Li et al. (2024) shows the failure of in-context learning tasks when applied to long-context scenarios. To investigate whether compression models can serve as a viable solution in long-context settings, we compare Selection- $p$  with long-context models, including LLaMA-2-7B-LongLora (Chen et al., 2023) and Long-LLaMA-code-7B (Tworkowski et al., 2023) on the BANKING77 dataset (Casanueva et al., 2020). The dataset contains 77 classes where traversing all the instances with unique labels requires approximately two thousand tokens. Evaluation is conducted at 2K, 4K, and 7K token levels, and we adopt a compression rate of  $10\\mathrm{x}$  for Selection- $p$  and LLMingua-2 among all levels. Since the long in-context demonstration is used, chunking is performed for every 2,048 tokens in Selection- $p$ . The compressed results are concatenated together with a space token between each pair of chunks. We again follow the evaluation setting by Chevalier et al. (2023) on the models' prediction, with the result presented in Table 3.",
+        "bbox": [
+            507,
+            318,
+            884,
+            688
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            699,
+            628,
+            712
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We compare the Selection-  $p$  with the following state-of-the-art compression models.",
+        "bbox": [
+            507,
+            719,
+            880,
+            751
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "- LLMLingua (Jiang et al., 2023) employs an iterative compression algorithm to filter less informative tokens based on the token-level perplexity. To further boost the performance, Jiang et al. (2023) also conducts a budget controller to allocate varying budgets across different demonstrations and questions. We find that there is also a significant discrepancy observed between the prescribed compression rate and the actual compression rate through",
+        "bbox": [
+            531,
+            760,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>In our preliminary experiments, we tried the Gumbel-softmax trick to optimize the loss in Eq. 2 but we did not observe gains over the direct optimization implemented in our paper.",
+        "bbox": [
+            112,
+            870,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "LLMLingua API calls. For a fair comparison, we exclusively utilize the token-level prompt compression algorithm from LLMLingua in Table 2. We additionally compare LLMLingua with Selection-  $p$  equipped with Budget Controller in Section 5.5.",
+        "bbox": [
+            147,
+            84,
+            487,
+            179
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- LLMLingua-2 (Pan et al., 2024) is derived from data distillation obtained by instructing GPT-4 to perform compression. Similar to ours, the model is trained as a binary classifier on each token, determining whether each token should be preserved.",
+            "- AutoCompressor (Chevalier et al., 2023) is constructed based on the RMT architecture (Bulatov et al., 2022). It compresses text into summary vectors that can be reused in subsequent segments. It is the only soft compression model adopted for comparison, considering that we followed the experiment setting for assessing the in-context learning ability of LLMs."
+        ],
+        "bbox": [
+            136,
+            193,
+            485,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Evaluation Result",
+        "text_level": 1,
+        "bbox": [
+            112,
+            470,
+            302,
+            485
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Traditional Classification Tasks. None of the compression models can achieve superior performance compared to the full-shot demonstration setting, which is in line with our expectations given the information loss during compression. However, certain tasks show a notable improvement when compared to both the zero-shot and full-shot approaches, e.g., all the hard compression models surpass zero-shot and full-shot by approximately  $20\\%$  in the WSC task. Among all compression models, Selection- $p$  demonstrates the highest performance in conducting ICL, with an average accuracy of  $67.4\\%$  across all 10 tasks as presented in Table 2. Examples of in-context demonstration before and after compression are shown in Appendix A. To demonstrate the effectiveness of our model, we have included one-tenth of the original demonstration set as the baseline. Our model significantly outperforms the baseline with a comparable number of tokens, therefore highlighting the effect of performing compression at the token level.",
+        "bbox": [
+            112,
+            492,
+            489,
+            831
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Long Context Classification Tasks. Our model outperforms LLaMA-2-7B-LongLora, Long-LLaMA-code-7B and LLMLingua-2 at all token size levels, and achieves similar results to Li et al. (2024)'s findings on the long-context models.",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In addition, a growing trend with variations is observed with increasing compressed demonstrations, indicating that our model can successfully learn from additional information after compression. Examples of in-context demonstration before and after compression are presented in Appendix A.",
+        "bbox": [
+            507,
+            84,
+            884,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.4 Transferability",
+        "text_level": 1,
+        "bbox": [
+            507,
+            191,
+            672,
+            206
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Compression is first performed on the demonstration set for ICL with Selection- $p$ . Subsequently, the compressed tokens are passed to a separate downstream model (i.e. LLaMA-2-13B or the black-box models) as the compressed demonstration prompt for evaluation.",
+        "bbox": [
+            507,
+            212,
+            882,
+            306
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To LLaMA-2-13B. We follow the same setting of evaluation across different classification tasks in Section 4.3. To assess the transferability of the compression models, we compress demonstrations with Selection- $p$  and input the compressed demonstration tokens into LLaMA-2-13B. In comparing different compression models, since retraining is required for soft compression methods, no results can be obtained for AutoCompressor (Chevalier et al., 2023). In the case of LMLingua, token-level perplexity is calculated with LLaMA-2-13B instead of LLaMaA-2-7B in this experiment.",
+        "bbox": [
+            507,
+            317,
+            882,
+            508
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our approach outperforms all other compression models as shown in Table 4. In addition, a small deviation is observed between the  $10\\mathrm{x}$  compression rate and the full shot setting, demonstrating the great transferability of our models. Notably, with Selection-  $p$  ,the tasks that outperform the full-shot setting in LLaMA-2-7b also exhibit similar patterns in LLaMA-2-13B.",
+        "bbox": [
+            507,
+            511,
+            880,
+            637
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To Black-box Models. Taking cost into consideration, we select ChatGPT (OpenAI, 2023) and Gemini (Team, 2023) for evaluation to examine its transferability to LLMs. Traditional classification tasks often have a simple nature and the potential issue of data contamination, leading to high accuracy and causing an insignificant evaluation. Therefore, we use BANKING77 (Casanueva et al., 2020) for evaluation. Following a similar setup as described in Section 4.3, we adopt a token size of 750 for examination. However, in case the compression rate is too high, ChatGPT and Gemini are more likely to deviate from the instructions and provide task-irrelevant responses. Therefore, we adopt a compression rate of  $3\\mathrm{x}$  and use the EM metric for this experiment given their black-box nature. Note that there may be variation in the result of Gemini",
+        "bbox": [
+            507,
+            646,
+            882,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/efbf2d1744cac689f6a044add781c277accb87d9b881b5896faee3991dbb48d2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Subj</td><td>RTE</td><td>WSC</td><td>BoolQ</td><td>MultiRC</td><td>SST-2</td><td>WIC</td><td>COPA</td><td>AG News</td><td>AVG</td></tr><tr><td>Zero-shot</td><td>49.3</td><td>58.8</td><td>43.4</td><td>67.4</td><td>52.5</td><td>67.7</td><td>50.8</td><td>52.5</td><td>63.3</td><td>56.2</td></tr><tr><td>“One-tenth”-shot</td><td>48.6</td><td>65.3</td><td>52.6</td><td>68.3</td><td>49.2</td><td>84.0</td><td>53.6</td><td>83.9</td><td>55.5</td><td>62.3</td></tr><tr><td>Full-shot</td><td>80.7</td><td>70.7</td><td>41.8</td><td>62.8</td><td>46.8</td><td>92.5</td><td>56.4</td><td>85.6</td><td>76.3</td><td>68.2</td></tr><tr><td>AutoCompressor</td><td>57.9</td><td>56.1</td><td>39.4</td><td>66.5</td><td>51.8</td><td>92.8</td><td>53.0</td><td>84.4</td><td>80.9</td><td>64.7</td></tr><tr><td>LLMLingua</td><td>56.7</td><td>60.9</td><td>63.0</td><td>69.4</td><td>50.3</td><td>69.9</td><td>51.6</td><td>76.4</td><td>61.5</td><td>62.2</td></tr><tr><td>LLMLingua-2</td><td>57.3</td><td>67.8</td><td>63.7</td><td>69.8</td><td>52.8</td><td>66.2</td><td>50.3</td><td>71.4</td><td>61.9</td><td>62.3</td></tr><tr><td>Selection-p</td><td>68.5</td><td>68.5</td><td>61.1</td><td>69.7</td><td>54.4</td><td>90.7</td><td>50.3</td><td>76.9</td><td>66.8</td><td>67.4</td></tr></table>",
+        "bbox": [
+            154,
+            80,
+            843,
+            212
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/a04a3f2bf762e1eb58daf1a266810dbd8c3e68cd7a7dc6400463487e0ebb4b18.jpg",
+        "table_caption": [
+            "Table 2: Evaluation result on traditional classification tasks. Four sets of random demonstrations were selected, with the average result being presented. The average result across different classification tasks is presented under AVG."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>2K</td><td>4K</td><td>7K</td></tr><tr><td>LLaMA-2-7B-LongLora</td><td>0.0</td><td>0.0</td><td>0.0</td></tr><tr><td>Long-LLaMA-code-7B</td><td>0.0</td><td>0.0</td><td>0.0</td></tr><tr><td>LLMlingua-2</td><td>41.2</td><td>31.6</td><td>35.2</td></tr><tr><td>Selection-p</td><td>46.9</td><td>50.9</td><td>51.6</td></tr></table>",
+        "bbox": [
+            152,
+            288,
+            450,
+            370
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Evaluation result on BANKING77 with increasing in-content demonstrations tokens length.",
+        "bbox": [
+            112,
+            382,
+            487,
+            411
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "since the discrete compressed tokens sometimes trigger the SAFETY error. The prompt used is presented in Appendix B.",
+        "bbox": [
+            112,
+            438,
+            485,
+            485
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Though the performance of our model still deviates from the full-shot setting, it achieved the best performance compared to the existing works as presented in Table 4, demonstrating fair transferability even in LLMs like ChatGPT. Surprisingly, though LLM Lingua-2 is distilled from GPT-4, it exhibits poor generalization compared to other compression models.",
+        "bbox": [
+            112,
+            487,
+            487,
+            613
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            630,
+            221,
+            646
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Flexibility",
+        "text_level": 1,
+        "bbox": [
+            112,
+            657,
+            238,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Performance with Different Number of Initial Tokens. The result in long context classification tasks in Section 4.3 shows the effectiveness of chunk-wise compression in long context. We further analyze if compression models work well in normal few-shot settings in classification tasks. In this experiment, the in-context demonstrations are selected with an approximate size of 250 tokens. The comparison to the result with the token size of 750 in Section 4.3 is presented in Table 6.",
+        "bbox": [
+            112,
+            678,
+            487,
+            839
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Selection-  $p$  shows the best performance under the constraint of 250 tokens when compared to other compression models. Additionally, it also follows the full-shot (i.e. 750 tokens level) trend, the average performance across all classification",
+        "bbox": [
+            112,
+            841,
+            487,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "tasks increases along with the number of provided demonstrations. On the contrary, other compression models didn't achieve an improvement in accuracy with more demonstrations. For instance, there is a drop of  $2\\%$  recorded with an additional 500 tokens of information for AutoCompressor.",
+        "bbox": [
+            507,
+            292,
+            884,
+            388
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Latency Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            400,
+            690,
+            416
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We analyze end-to-end latency on A100-80G GPU with the WSC task, illustrated in Table 7. Our method can achieve  $5.3\\mathrm{x}$  speed up on  $10\\mathrm{x}$  compressed in-context demonstration. Compared to the inference time, negligible time is required for compression on the ICL task setting, demonstrating high efficiency in adopting our models for compression. We also compared LLMLingua with the disabled content Budget Controller. It requires iterative decoding on the segmented context while Selection- $p$  only requires a single inference on all tokens and demonstrates a good performance.",
+        "bbox": [
+            507,
+            420,
+            882,
+            615
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.3 Correlation with Attention and Perplexity",
+        "text_level": 1,
+        "bbox": [
+            507,
+            626,
+            880,
+            642
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "With the  $p$ -value ranging between 0 and 1 for each token, we further study whether any correlations exist among  $p$ , the mean attention value during the forward pass, and the tokens level perplexity (i.e. a core component in LLMLingua (Jiang et al., 2023)). Since the value of  $p$  is derived from the last hidden state of the model, we only consider the last layer mean attention of our tuned model. We employed Spearman's Rank Correlation Coefficient (Spearman, 1904) to compute the correlation between the three variables. It is calculated for different traditional classification tasks and the averaged value across tasks. The result presented in Figure 2 indicates only a weak correlation observed between the  $p$  value and the other two variables while the correlation between the last layer mean attention and perplexity is more significant. Among all tasks,",
+        "bbox": [
+            507,
+            646,
+            882,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/fbee90128c89e235f29e5621695770001274f636805e382599faeff62fefda7c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"10\">LLaMA-2-13B</td></tr><tr><td></td><td>Subj</td><td>RTE</td><td>WSC</td><td>BoolQ</td><td>MultiRC</td><td>SST-2</td><td>WIC</td><td>COPA</td><td>AG News</td><td>AVG</td></tr><tr><td>Full-shot</td><td>91.6</td><td>74.8</td><td>46.9</td><td>67.7</td><td>45.7</td><td>94.7</td><td>54.6</td><td>77.6</td><td>79.2</td><td>70.3</td></tr><tr><td>LLMingua</td><td>53.6</td><td>61.0</td><td>63.2</td><td>70.6</td><td>51.5</td><td>64.3</td><td>50.0</td><td>78.1</td><td>58.0</td><td>61.4</td></tr><tr><td>LLMingua-2</td><td>48.3</td><td>68.7</td><td>41.3</td><td>75.8</td><td>52.2</td><td>51.3</td><td>49.0</td><td>48.2</td><td>71.1</td><td>56.2</td></tr><tr><td>Selection-p</td><td>69.3</td><td>69.5</td><td>65.4</td><td>74.7</td><td>50.7</td><td>81.1</td><td>50.5</td><td>87.5</td><td>63.2</td><td>68.0</td></tr></table>",
+        "bbox": [
+            159,
+            80,
+            840,
+            186
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/325ae2643eaab6fd436ab6226353aaba7714df60f404cc1e22fadc70eb7ad4a1.jpg",
+        "table_caption": [
+            "Table 4: Analysis of transferability to open-source model LLaMA-2-13B. The experiment is performed on 750 tokens in-context demonstrations with a 10x compression rate."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>ChatGPT\nGPT-3.5-Turbo</td><td>Gemini\nGemini-1.0-Pro</td></tr><tr><td>Full-shot</td><td>74.2</td><td>73.3</td></tr><tr><td>LLMlingua</td><td>58.6</td><td>40.2</td></tr><tr><td>LLMlingua-2</td><td>55.7</td><td>51.9</td></tr><tr><td>Selection-p</td><td>62.9</td><td>58.9</td></tr></table>",
+        "bbox": [
+            139,
+            244,
+            467,
+            340
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Analysis of transferability to blackbox models (i.e. ChatGPT and Gemini). The experiment is performed on 750 tokens in-context demonstrations with a 3x compression rate.",
+        "bbox": [
+            112,
+            350,
+            489,
+            407
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "WIC demonstrates a prominently high value compared to others, this may explain the small variation in accuracy across different compression models and different experimental settings.",
+        "bbox": [
+            112,
+            434,
+            489,
+            500
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Tokens Level Part-of-Speech Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            512,
+            455,
+            529
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To further interpret the rationale behind our compression models, we analyze what kinds of words are likely preserved by Selection-  $p$ . Under the discreteness of our compression result, we locate the corresponding words from the compressed tokens and obtain the Part-of-Speech (PoS) tags with an NLTK tagger. For each type of PoS tag, we compute the token preservation percentage with",
+        "bbox": [
+            112,
+            533,
+            489,
+            664
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\frac {\\left| \\text {c o m p r e s s e d} _ {t} \\text {k o t e n} _ {\\text {t a g} _ {i}} \\right|}{\\left| \\text {t o t a l} _ {t} \\text {t o k e n} _ {\\text {t a g} _ {i}} \\right|}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            213,
+            676,
+            386,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "for each PoS tag  $\\mathrm{tag}_{i}$ . The experiment is conducted between the compressed result and the original demonstrations among the nine traditional classification tasks with four demonstration sets per task. We analyze tags with a frequency of appearance greater than  $1\\%$ .",
+        "bbox": [
+            112,
+            727,
+            489,
+            822
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From the result presented in Figure 3, PRP and punctuations (i.e. indicating the start of the next sentence or phrase) are more likely preserved. The potential reason for the high preservation ratio on PRP (personal pronoun) likely corresponds to the pronoun resolution task of WSC. Under the task",
+        "bbox": [
+            112,
+            824,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/71879c9b24ec7215b4977e0372906f4bf7db657088b7556e080f0810d16064b0.jpg",
+        "image_caption": [
+            "Figure 2: Spearman's Rank Correlation Coefficient (spearmanr) between  $p$  (p) value, mean attention (a) and token-level perplexity (ppl) across different traditional classification tasks."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            236,
+            843,
+            514
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "setting, it can be useful hints for the answer derivation.",
+        "bbox": [
+            507,
+            598,
+            882,
+            629
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The high preservation ratio of punctuation may indicate a large redundancy in a sentence, and truncating sentence separation tokens is undesirable. Additionally, as highlighted by Wang et al. (2023), formatting information (i.e. structure of the demonstrations) matters a lot in in-context learning. However, formatting tokens (i.e. “:”) are unlikely to be preserved with Selection- $p$  compared to the original distribution in our case. In general, we also observe a higher degree of preservation of noun phrases compared to verbs.",
+        "bbox": [
+            507,
+            631,
+            884,
+            808
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.5 On Fair Comparison with LLMLingua",
+        "text_level": 1,
+        "bbox": [
+            507,
+            819,
+            862,
+            835
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As described in Section 4.2, LLMLingua conducts demonstration selection prior to compression at the token level, while other methods compress directly on the token level. Since the demonstration selection process can also be incorporated into other",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/3fdd07e0580d58de74c7ae5b2d690ce5399f1a57fcd57f5434cd95bdd313c2a0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Subj</td><td>RTE</td><td>WSC</td><td>BoolQ</td><td>MultiRC</td><td>SST-2</td><td>WIC</td><td>COPA</td><td>AG News</td><td>AVG</td></tr><tr><td>Zero-shot</td><td>49.3</td><td>58.8</td><td>43.4</td><td>67.4</td><td>52.5</td><td>67.7</td><td>50.8</td><td>52.5</td><td>63.3</td><td>56.2</td></tr><tr><td>Full-shot</td><td>81.3</td><td>69.9</td><td>51.2</td><td>62.7</td><td>46.8</td><td>89.3</td><td>51.8</td><td>85.1</td><td>67.9</td><td>67.3</td></tr><tr><td>ΔFull-shot</td><td>-0.6</td><td>+0.8</td><td>-9.4</td><td>+0.1</td><td>±0</td><td>+3.2</td><td>+4.6</td><td>+0.5</td><td>+8.4</td><td>+0.9</td></tr><tr><td>AutoCompressor</td><td>56.2</td><td>61.5</td><td>44.2</td><td>68.3</td><td>52.7</td><td>93.0</td><td>51.5</td><td>83.6</td><td>76.1</td><td>65.2</td></tr><tr><td>ΔAutoCompressor</td><td>+1.7</td><td>-5.4</td><td>-4.8</td><td>-1.8</td><td>-0.9</td><td>-0.2</td><td>+1.5</td><td>+0.8</td><td>+4.8</td><td>-0.5</td></tr><tr><td>LLMLingua*</td><td>55.6</td><td>61.4</td><td>61.3</td><td>68.2</td><td>53.1</td><td>81.1</td><td>50.2</td><td>75.8</td><td>70.9</td><td>64.2</td></tr><tr><td>ΔLLMLingua</td><td>+1.1</td><td>-0.5</td><td>+1.7</td><td>+1.2</td><td>-2.8</td><td>-11.2</td><td>+1.4</td><td>+0.6</td><td>-9.4</td><td>-2.0</td></tr><tr><td>LLMLingua-2</td><td>52.4</td><td>65.3</td><td>63.9</td><td>66.1</td><td>50.9</td><td>82.9</td><td>50.8</td><td>77.8</td><td>56.3</td><td>62.9</td></tr><tr><td>ΔLLMLingua-2</td><td>+4.9</td><td>+2.5</td><td>-0.2</td><td>+3.7</td><td>+1.9</td><td>-16.7</td><td>-0.5</td><td>-6.4</td><td>+5.6</td><td>-0.6</td></tr><tr><td>Selection-p</td><td>65.7</td><td>65.5</td><td>58.7</td><td>67.5</td><td>54.3</td><td>81.3</td><td>50.4</td><td>77.9</td><td>68.3</td><td>65.5</td></tr><tr><td>ΔSelection-p</td><td>+2.8</td><td>+3.0</td><td>+2.4</td><td>+2.3</td><td>+0.1</td><td>+9.4</td><td>-0.1</td><td>-1.0</td><td>-1.5</td><td>+1.9</td></tr></table>",
+        "bbox": [
+            147,
+            80,
+            845,
+            275
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a61bc460fe771ecbc98abe9805443c6b5f79c1338916ed5810906d538a6550d8.jpg",
+        "table_caption": [
+            "Table 6: Performance with different number of demonstrations from about 250 tokens to about 750 tokens.  $\\Delta$  refers to the performance enhancement that can be achieved by increasing the demonstration tokens size to 750."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>1x</td><td>2x</td><td>5x</td><td>10x</td></tr><tr><td>End-to-End without compression</td><td>298.6</td><td></td><td></td><td></td></tr><tr><td>End-to-End with Selection-p</td><td></td><td>167.0 (1.8x)</td><td>81.6 (3.7x)</td><td>55.6 (5.3x)</td></tr><tr><td>LLMLingua per demonstrations set</td><td>-</td><td>0.82</td><td>0.82</td><td>0.81</td></tr><tr><td>Selection-p per demonstrations set</td><td>-</td><td>0.68</td><td>0.67</td><td>0.67</td></tr></table>",
+        "bbox": [
+            226,
+            326,
+            771,
+            411
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "compression models, we only utilize the modified version of LLMLingua in the previous experiments to ensure a fair comparison.",
+        "bbox": [
+            112,
+            502,
+            487,
+            551
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we further analyze the performance by comparing our proposed method, equipped with the Budget Controller, with the whole LLMLingua to provide a comprehensive analysis. We follow the setting described in Section 4.3 and select the WSC task for our experiment. To illustrate, in the original demonstration set consisting of 16 demonstrations, the LLMLingua API retains only four demonstrations. This leads to two options with Selection-  $p$ : (1) continuously applying the  $10\\mathrm{x}$  compression directly to the filtered set of four demonstrations and resulting in a final compression rate of  $38\\mathrm{x}$ , and (2) adjusting the compression rate of Selection-  $p$  to achieve a final compression rate of  $10\\mathrm{x}$ .",
+        "bbox": [
+            112,
+            558,
+            489,
+            800
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The result presented in Table 8 demonstrates the significant impact of the Budget Controller. Similar trends in performance for both Selection-  $p$  and LLMLingua are observed (i.e., a decrease in performance on the WSC task). Notably, the performance of Selection-  $p$  surpasses LLMLingua after equipping with the Budget Controller.",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/617409adb1273099cf7049734e4762f30e18f88ec24bf8c4105ff9930f4cbfde.jpg",
+        "table_caption": [
+            "Table 7: Latency(s) comparison on WSC in 750 tokens level with about 16 demonstrations. We present the averaged complete end-to-end inference with and without Selection-  $p$  among four sets of demonstrations. Comparison is conducted with LLMLingua which also builds upon the LLaMA-2-7B backbone, with the averaged compression time of the in-context demonstrations being presented."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>WSC</td><td>rate</td></tr><tr><td>Selection-p</td><td>61.1</td><td>10x</td></tr><tr><td>LLMLingua</td><td>63.0</td><td>10x</td></tr><tr><td>Selection-p (+ Budget Controller)</td><td>47.6</td><td>10x</td></tr><tr><td>Selection-p (+ Budget Controller)</td><td>57.0</td><td>38x</td></tr><tr><td>LLMLingua (whole)</td><td>44.7</td><td>10x</td></tr></table>",
+        "bbox": [
+            557,
+            499,
+            836,
+            595
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Comparison with LLMLingua on Budget Controller. Adopting different strategies in equipping Selection-  $p$  with Budget Controller, leads to the two different compression rates (rate) of  ${10}\\mathrm{x}$  and  ${38}\\mathrm{x}$  .",
+        "bbox": [
+            507,
+            605,
+            882,
+            663
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, there is a disparity between the instructed compression rate and the actual compression rate in LLMLingua. The target size for compressed tokens is 75, while LLMLingua typically achieves an average compressed token size of around 192.1, which is more than 1.5 times higher than the desired rate across all classification tasks.",
+        "bbox": [
+            507,
+            688,
+            882,
+            800
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            814,
+            640,
+            829
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We introduce a simple yet effective self-supervised approach in context compression and conduct evaluation across 10 classification tasks in both few-shot and long-context settings. Our approach also demonstrated great transferability to both the open-",
+        "bbox": [
+            507,
+            841,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/a77d617378c73b3204182d5fe86570211ad68f6deafbea4efd61db42f87bae3d.jpg",
+        "image_caption": [
+            "Figure 3: Analysis of the token preservation percentage with respect to different types of Part-of-Speech tags under  $10\\mathrm{x}$  compression rate."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            159,
+            85,
+            836,
+            200
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "source (i.e. LLaMA-2-13B) and black-box models (i.e. ChatGPT and Gemini), with performance surpassing the existing state-of-the-art compression models. Analysis is also conducted among different compression rates and demonstration lengths. With a 10x compression rate, our model only shows a 0.8-point drop in performance across different traditional classification tasks with a 5.3x speedup. Through experiments in long-context settings, our work also presents the possibility of addressing the in-context learning issue of the recent long-context models. Both efficiency enhancement as well as performance preservation are shown in our model.",
+        "bbox": [
+            110,
+            269,
+            489,
+            479
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement This research has been made possible by the Hong Kong PhD Fellowship provided to Tsz Ting Chung and the Research Impact Fund project R6003-21 provided by the Research Grants Council of Hong Kong to Dit-Yan Yeung.",
+        "bbox": [
+            112,
+            488,
+            489,
+            570
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            579,
+            220,
+            594
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Under the consideration of cost, we did not perform further analysis on other LLMs apart from ChatGPT and Gemini. In addition, our model which builds up LLaMA-2-7B does not achieve better latency than models like LLMLingua-2 and AutoCompressor. Under the ICL setting, minimal time is required for compression, leading to insignificance in end-to-end inference time. While AutoCompressor offers better latency, its soft compression nature limits its applicability to other LLMs. Overall, our experiments across various tasks and settings demonstrate better performance and transferability, with the benefits outweighing the latency issue.",
+        "bbox": [
+            112,
+            605,
+            489,
+            829
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            857,
+            213,
+            872
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. 2023. Token merging: Your vit but faster. In The",
+        "bbox": [
+            114,
+            879,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Eleventh International Conference on Learning Representations.",
+        "bbox": [
+            526,
+            269,
+            882,
+            297
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.",
+        "bbox": [
+            510,
+            305,
+            884,
+            501
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Aydar Bulatov, Yuri Kuratov, and Mikhail Burtsev. 2022. Recurrent memory transformer. In Advances in Neural Information Processing Systems.",
+        "bbox": [
+            509,
+            508,
+            882,
+            550
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Inigo Casanueva, Tadas Temčinas, Daniela Gerz, Matthew Henderson, and Ivan Vulić. 2020. Efficient intent detection with dual sentence encoders. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 38-45, Online. Association for Computational Linguistics.",
+        "bbox": [
+            510,
+            557,
+            882,
+            638
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. 2023. Longlora: Efficient fine-tuning of long-context large language models. ArXiv preprint, abs/2309.12307.",
+        "bbox": [
+            510,
+            644,
+            882,
+            697
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. 2023. Adapting language models to compress contexts. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3829-3846, Singapore. Association for Computational Linguistics.",
+        "bbox": [
+            510,
+            705,
+            882,
+            785
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Tao Ge, Hu Jing, Lei Wang, Xun Wang, Si-Qing Chen, and Furu Wei. 2024. In-context autoencoder for context compression in a large language model. In The Twelfth International Conference on Learning Representations.",
+        "bbox": [
+            510,
+            793,
+            882,
+            858
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Saurabh Goyal, Anamitra Roy Choudhury, Saurabh Raje, Venkatesan T. Chakaravarthy, Yogish Sabharwal, and Ashish Verma. 2020. Power-bert: Accelerating BERT inference via progressive word-vector",
+        "bbox": [
+            509,
+            866,
+            882,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "elimination. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 3690-3699. PMLR.",
+            "Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.",
+            "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.",
+            "Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2023. LLMLingua: Compressing prompts for accelerated inference of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13358-13376, Singapore. Association for Computational Linguistics.",
+            "Gyuwan Kim and Kyunghyun Cho. 2021. Length-adaptive transformer: Train once with length drop, use anytime with search. 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 6501-6511, Online. Association for Computational Linguistics.",
+            "Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. 2022. Learned token pruning for transformers.",
+            "Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. 2024. Long-context llms struggle with long in-context learning.",
+            "Yucheng Li, Bo Dong, Frank Guerin, and Chenghua Lin. 2023. Compressing context to enhance inference efficiency of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6342-6353, Singapore. Association for Computational Linguistics.",
+            "Sewon Min, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Noisy channel language model prompting for few-shot text classification. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5316-5330, Dublin, Ireland. Association for Computational Linguistics.",
+            "Ali Modarressi, Hosein Moebbi, and Mohammad Taher Pilehvar. 2022. AdapLeR: Speeding up inference by adaptive length reduction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1-15, Dublin, Ireland. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jesse Mu, Xiang Li, and Noah Goodman. 2023. Learning to compress prompts with gist tokens. In Advances in Neural Information Processing Systems, volume 36, pages 19327-19352. Curran Associates, Inc.",
+            "OpenAI. 2023. GPT-4 technical report. ArXiv preprint, abs/2303.08774.",
+            "Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Ruhle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. 2024. Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression.",
+            "Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. 2021. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 13937-13949.",
+            "C. Spearman. 1904. The proof and measurement of association between two things. American Journal of Psychology, 15:88-103.",
+            "Gemini Team. 2023. Gemini: A family of highly capable multimodal models.",
+            "TogetherAI. 2023. Redpajama: An open dataset for training large language models.",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models.",
+            "Szymon Tworkowski, Konrad Staniszewski, Mikołaj Pacek, Yuhuai Wu, Henryk Michalewski, and Piotr Miłos. 2023. Focused transformer: Contrastive training for context scaling.",
+            "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 3261-3275.",
+            "Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. 2023. Label words are anchors: An information flow perspective for understanding in-context learning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9840-9855, Singapore. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models.",
+            "David Wingate, Mohammad Shoeybi, and Taylor Sorensen. 2022. Prompt compression and contrastive conditioning for controllability and toxicity reduction in language models. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5621-5634, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            241
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Examples Illustration of Compression Results",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            480,
+            116
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 shows examples of two traditional classification tasks (i.e. Subj and WSC) with Selection- $p$  under  $10\\mathrm{x}$  compression, displaying both the compressed results and the original demonstration sets before compression. Additionally, Figure 5 illustrates an example for BANKING77 under  $10\\mathrm{x}$  compression.",
+        "bbox": [
+            112,
+            128,
+            489,
+            240
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Prompt of Evaluation on BANKING77",
+        "text_level": 1,
+        "bbox": [
+            112,
+            256,
+            485,
+            273
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In our prompt to ChatGPT, we first list out all 77 labels and then provide a list of demonstrations. The template is detailed below,",
+        "bbox": [
+            112,
+            285,
+            489,
+            332
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Answer in activate_my_card or age_limit or apple_pay_or_google_pay or atm_support or automatic_top_up or balance_not UPDATED_after_bank_transfer or ... or wrong_amount_of_cash_received or wrong_exchange_rate_for_cash_withdrawal.",
+        "bbox": [
+            112,
+            351,
+            489,
+            447
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Context: <context>",
+        "bbox": [
+            112,
+            464,
+            263,
+            478
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Answer: <answer>",
+        "bbox": [
+            112,
+            481,
+            262,
+            494
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Training Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            527,
+            294,
+            544
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We use LLaMA-2-7B for compression (i.e. token selection). It takes roughly 50 hours on a single A100 GPU to train on 100M tokens from RedPajama.",
+        "bbox": [
+            112,
+            556,
+            489,
+            620
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Detailed Latency Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            636,
+            378,
+            653
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "For the token-based compression models, the time needed for compression per demonstration set on the WSC task is presented in Table 9.",
+        "bbox": [
+            112,
+            664,
+            487,
+            712
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/f25308a7fbffe2be27be2a85589d6566f881d96d952e5d735923c0b81e8521d1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>LLMLingua-2</td><td>Selection-p</td><td>LLMLingua</td></tr><tr><td>0.15</td><td>0.67</td><td>0.81</td></tr></table>",
+        "bbox": [
+            161,
+            726,
+            438,
+            766
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Under the ICL setting, the same demonstration set is used consistently, and compression on the demonstration set only needs to be computed once for all subsequent inferences. This contributed to the short compression time in the complete end-to-end process.",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Considering that the time taken for compression on the demonstration set is negligible when compared to the time required for inference (approximately at a ratio of 0.01), the end-to-end inference time for the token-selection-based compression models is roughly the same. The end-to-end inference time result for the WSC task on LLaMA2-7B is shown in Table 10. The table is presented in order of the inference time for clarity.",
+        "bbox": [
+            507,
+            84,
+            884,
+            229
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/87f30e9a005ce26c2498c17e3eff99dd83c35a8220b058b724295295c9191aa6.jpg",
+        "table_caption": [
+            "Table 9: Time needed for compression per demonstration set on the WSC task."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>AutoCompressor</td><td>LLMLingua-2</td><td>Selection-p</td><td>LLMLingua</td></tr><tr><td>22.73</td><td>55.6</td><td>55.6</td><td>55.6</td></tr></table>",
+        "bbox": [
+            531,
+            239,
+            860,
+            281
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 10: End-to-end inference time on the WSC task.",
+        "bbox": [
+            509,
+            291,
+            878,
+            305
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[Subj] Original Demonstrations:",
+        "text_level": 1,
+        "bbox": [
+            142,
+            105,
+            400,
+            121
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "input: each of the principals has a radically different way of dealing with it. type: objective",
+        "bbox": [
+            142,
+            123,
+            702,
+            154
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "input: well-intentioned though it may be, its soap-opera morality tales have the antiseptic, preprogrammed feel of an after-school special. type: subjective",
+        "bbox": [
+            141,
+            170,
+            858,
+            219
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "input: an astonishing feat for a major star let alone a 27 year old from pickum , south carolina who only two years ago was sleeping in a cardboard box in the back alleys of detroit with her mother , connie , and her uncle clutch , while playing guitar on the streets for spare change .type: objective",
+        "bbox": [
+            141,
+            234,
+            855,
+            299
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "input: may is a young strange girl who had a very disturbed childhood and does not still know the meaning of true friendship or love. type: objective",
+        "bbox": [
+            141,
+            315,
+            855,
+            363
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[Subj] Compressed demonstrations (10x):",
+        "text_level": 1,
+        "bbox": [
+            142,
+            379,
+            470,
+            395
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "input:-int it may- mor exc lord treasure planet- the ste moments- it thr en a r with he and I crowd gru a-erm own - into impossibly ris gar \"ir wh sh de prere prede it mag m ro philosophvag rede thes v de spcer twz-passer aston fe two with cl for dist and still",
+        "bbox": [
+            141,
+            395,
+            855,
+            444
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[WSC] Original Demonstrations:",
+        "text_level": 1,
+        "bbox": [
+            142,
+            476,
+            403,
+            491
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: In the sentence \"James asked Robert for a favor but he was refused.\", does the pronoun 'he' refer to Robert?  \nAnswer: no",
+        "bbox": [
+            141,
+            493,
+            855,
+            538
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: In the sentence \"What about the time you cut up tulip bulbs in the hamburgers because you thought they were onions?\", does the pronoun 'they' refer to tulip bulbs? Answer: yes",
+        "bbox": [
+            141,
+            556,
+            855,
+            604
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: In the sentence \"When Mr. Bond, the veterinarian, came to look at the black horse that lay groaning on the grass, he felt him all over, and shook his head; one of his legs was broken.\", does the pronoun 'his' refer to the black horse?  \nAnswer: no",
+        "bbox": [
+            142,
+            620,
+            855,
+            684
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: In the sentence \"Sam took French classes from Adam, because he was eager to speak it fluently,\" does the pronoun 'he' refer to Adam?",
+        "bbox": [
+            141,
+            701,
+            855,
+            733
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Answer: no",
+        "bbox": [
+            142,
+            734,
+            231,
+            747
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[WSC] Compressed demonstrations (10x):",
+        "text_level": 1,
+        "bbox": [
+            142,
+            765,
+            475,
+            781
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "\" pron:: \" tul pron ' tul: he pron ' told P which P. He have pron ' P a the would only Gru une pron pronI put cfr It pron refriger man pronJohn wheng. He very im Wainws d Fol he pron veteran gro his pron the repa pron w t win gro \" he pron '",
+        "bbox": [
+            141,
+            782,
+            855,
+            829
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: Illustration of the compression result by Selection- $p$  for Subj and WSC tasks under  $10\\mathrm{x}$  compression rate. Compression is performed with 19 demonstrations for Subj while it is performed with 16 demonstrations for WSC with total sum of about 750 tokens respectively.",
+        "bbox": [
+            114,
+            859,
+            884,
+            903
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "[BANKING77] Original Demonstrations:",
+        "text_level": 1,
+        "bbox": [
+            141,
+            281,
+            468,
+            297
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Context: Why did using an ATM cause me to be charged an additional fee?",
+        "bbox": [
+            141,
+            298,
+            697,
+            313
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answer: cash Withdrawal charge",
+        "bbox": [
+            141,
+            315,
+            394,
+            330
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Context: I asked for a refund but its not here yet.",
+        "bbox": [
+            141,
+            346,
+            497,
+            362
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answer: Refund_not Showing_up",
+        "bbox": [
+            141,
+            363,
+            396,
+            379
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Context: is there a reason i need to verify top up",
+        "bbox": [
+            141,
+            395,
+            499,
+            411
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answer: verify_top_up",
+        "bbox": [
+            141,
+            412,
+            315,
+            426
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "···",
+        "bbox": [
+            141,
+            435,
+            159,
+            441
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Context: There is a payment on my card that I do not recognize. I've never seen the name on the transaction before.",
+        "bbox": [
+            141,
+            443,
+            853,
+            473
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answer: card-payment_not_recognised",
+        "bbox": [
+            141,
+            476,
+            436,
+            491
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Context: I happened to forget my passcode",
+        "bbox": [
+            141,
+            508,
+            460,
+            523
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answer: passcode_forgotten",
+        "bbox": [
+            141,
+            524,
+            354,
+            539
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Context: I made a cash withdrawal and it is still listed as a pending transaction.",
+        "bbox": [
+            141,
+            556,
+            724,
+            571
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Answer: pending_cash_withdrawal",
+        "bbox": [
+            141,
+            573,
+            403,
+            587
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "[BANKING77] Compressed demonstrations (10x):",
+        "text_level": 1,
+        "bbox": [
+            141,
+            604,
+            539,
+            619
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "c_with asked ref Refnoting_up_recogn_c wrong_rece_transaction_chargtw unblock activ activ_not_fe收费标准 : ex sho Please revert_ the cardWr: I: pending_: the card.not_recogn: passf:_c",
+        "bbox": [
+            139,
+            621,
+            858,
+            668
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Figure 5: Illustration of the compression result by Selection- $p$  for BANKING77 under 10x compression rate. Compression is performed with 27 demonstrations with total sum of about 750 tokens.",
+        "bbox": [
+            112,
+            697,
+            884,
+            726
+        ],
+        "page_idx": 13
+    }
+]

2024/Selection-p_ Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability/14d2ceda-5f83-49b4-88b6-f58b5179b8e9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69ab15c4e9e38ba25eece21902b740b55158f757bf0a7d8d2f9720780ffb3f8a
+size 537994

2024/Selection-p_ Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability/full.md

@@ -0,0 +1,386 @@
+# Selection- $p$ : Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability
+
+Tsz Ting Chung $^{1}$  Leyang Cui $^{2}$  Lemao Liu $^{2}$  Xinting Huang $^{2}$  Shuming Shi $^{2}$  Dit-Yan Yeung $^{1}$
+
+<sup>1</sup>The Hong Kong University of Science and Technology <sup>2</sup>Tencent AILab {ttchungc, nealcly.nlp, lemaoliu, shuming}@gmail.com timxthuang@tencent.com dyyeung@cse.ust.hk
+
+# Abstract
+
+Large Language Models (LLMs) have demonstrated impressive capabilities in a wide range of natural language processing tasks when leveraging in-context learning. To mitigate the additional computational and financial costs associated with in-context learning, several prompt compression methods have been proposed to compress the in-context learning prompts. Despite their success, these methods face challenges with transferability due to model-specific compression, or rely on external training data, such as GPT-4. In this paper, we investigate the ability of LLMs to develop a unified compression method that discretizes uninformative tokens, utilizing a self-supervised pre-training technique. By introducing a small number of parameters during the continual pretraining, the proposed Selection- $p$  produces a probability for each input token, indicating whether to preserve or discard it. Experiments show Selection- $p$  achieves state-of-the-art performance across numerous classification tasks, achieving compression rates of up to 10 times while experiencing only a marginal  $0.8\%$  decrease in performance. Moreover, it exhibits superior transferability to different models compared to prior work. Additionally, we further analyze how Selection- $p$  helps maintain performance on in-context learning with long contexts.
+
+# 1 Introduction
+
+In-context learning has shown remarkable success in various natural language processing tasks (Brown et al., 2020), such as classification task (Min et al., 2022), and mathematical reasoning task (Wei et al., 2023), enabling Large Language Models (LLMs) to tackle complex and diverse tasks using only few-shot samples. However, in-context learning also significantly extends the length of prompts, resulting in increased computational and financial costs. Recently, a line of work has been
+
+<table><tr><td></td><td>Transferable</td><td>Single Run</td><td>-External</td></tr><tr><td>AutoCompressor</td><td>X</td><td>✓</td><td>✓</td></tr><tr><td>LLMLingua</td><td>✓</td><td>X</td><td>✓</td></tr><tr><td>LLMLingua-2</td><td>✓</td><td>✓</td><td>X</td></tr><tr><td>Selection-p</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+Table 1: Comparison between the proposed Selection- $p$  and existing content compression approaches. Selection- $p$  exhibits great transferability (Transferable), does not require multi-round iterative decoding (Single Run), and does not rely on costly external resources for training ( $\neg$ External).
+
+focusing on prompt compression, which aims to compress the original prompts while minimizing information loss. They are categorized into discrete compression (Li et al., 2023; Jiang et al., 2023; Pan et al., 2024) and continuous compression (Mu et al., 2023; Chevalier et al., 2023; Ge et al., 2024); the former compresses the context into discrete tokens, while the latter compresses it into a short sequence of continuous vectors.
+
+Observing redundant and repetitive content in a given input, discrete compression methods aim to eliminate less informative context without significantly compromising the model's performance. For example, LLM Lingua (Jiang et al., 2023) proposes to perform iterative token truncation based on the content perplexity, requiring multi-round decoding (Jiang et al., 2023). LLM Lingua-2 (Pan et al., 2024), distilled from GPT-4 (OpenAI, 2023), addresses the potential misalignment between entropy and the compression objective, as well as the distribution gap between the perplexity of the compression model and the target model. High costs are still involved in the training data construction. Meanwhile, optimizing the distribution for specific LLMs (GPT-4) may, on the other hand, hinder the transferability of the compressed content to other LLMs. More details are discussed in Section 4.4.
+
+Continuous compression (Bulatov et al., 2022;
+
+Wingate et al., 2022) teaches pre-trained LMs the ability to compress text into a short sequence of continuous vectors. AutoCompressors (Chevalier et al., 2023) uses an unsupervised learning objective, which motivates the model to cache crucial information within the summary vectors. Despite their success, these methods have poor generalization as they can only compress to the length specified during training. Additionally, since the continuous vector cannot be transferred between models, a separate compressor must be trained for each model.
+
+This raises an interesting research question:
+
+"Can LLMs learn to identify less informative tokens within a given context without external annotated signals?"
+
+To answer this question, we propose a pretraining strategy with a self-supervised signal, which enables the model to autonomously learn to predict the next token based on compressed context. With a small additional number of parameters, a forward pass on the proposed selection-  $p$  creates the probability vector  $p$  corresponding to each input token, indicating whether to preserve or discard the token. During inference, we can apply the detokenized compressed tokens to any downstream LLMs with only single-turn decoding and without reliance on any costly external resources as depicted in Table 1.
+
+The main contributions of this work are fourfold:
+
+- We present Selection- $p$  which achieves only  $0.8\%$  drops in performance under 10x compression rate across nine traditional classification tasks, surpassing the performance of the existing compression models. Under this setting, a speedup of 5.3x can be achieved during inference with in-context learning.  
+- Selection- $p$  demonstrated great transferability, which surpasses the performance of prior work in performing hard compression for both open-source and close-source models.  
+- We further analyze how Selection- $p$  helps in in-context learning in long-context settings, presenting a potential solution to address the performance declination of long-context models in ICL.  
+- We connect in-domain prior works and make comparisons with these state-of-the-art compression models, providing a complete picture.
+
+# 2 Related Work
+
+# 2.1 Hard Compression
+
+Some studies focus on token pruning (Goyal et al., 2020; Kim and Cho, 2021; Rao et al., 2021; Kim et al., 2022; Modarressi et al., 2022) and token merging (Bolya et al., 2023) but they are designed primarily for smaller models like BERT. More recently, Selective Context (Li et al., 2023) is the first to propose to prune less important tokens based on information entropy. Subsequently, LLMLingua (Jiang et al., 2023) refined the approach by integrating the selection of demonstrations and the allocation of compression budgets for various segments of the input prompt. No training is required in these models but their efficacy in downstream tasks with compression applied to in-context demonstrations remains limited. Pan et al. (2024) extended the idea and addressed the potential misalignment between entropy and the compression objective, leveraging full bidirectional context by training on their proposed GPT-4 distilled compression dataset. Our simple yet effective approaches outperform previous studies.
+
+# 2.2 Soft Compression
+
+Gist (Mu et al., 2023) is first proposed to compress prompt with soft tokens. Subsequently, Autocompressor (Chevalier et al., 2023) and ICAE (Ge et al., 2024) extend the idea to handle long contexts with different pretraining approaches. ICAE further conducts instruction tuning to enhance model performance. The downstream performance of these models heavily relies on the tuned compression model, with a fixed compression rate. Additionally, retraining is necessary for different versions of LLMs. Compared to these approaches, our work offers greater flexibility and transferability, while simultaneously surpassing the performance of existing compression models.
+
+# 3 Methodology
+
+Under the intuition that redundant texts often exist and their removal does not hinder human understanding of the text, we assume that LLMs behave in a similar manner. To efficiently identify less informative tokens within a given context, we propose a simple pre-training objective encouraging the model to predict the same token both before and after discarding less informative tokens.
+
+![](images/1153cb586b1ec2012e25950e87672fbf4d302723f611451a86c1d38d66563ab3.jpg)  
+Figure 1: Illustration with the training process. Areas in orange are learnable parameters. For the input context  $[x_1, x_2, \ldots, x_{n-1}]$ , inference without parameters update is performed first to create the attention mask  $\bar{p}$ . These subsequently form the model input for LoRA training and updating the parameters of the additional linear layer.
+
+# 3.1 Preliminary
+
+Language Model. Given a context  $[x_1, x_2, \ldots, x_{n-1}]$ , the objective of the language model is to predict the next token  $x_n$ , formed as  $P(x_n | x_1, x_2, \ldots, x_{n-1})$ . In case of training with the causal language modeling (CLM) loss, we will have,
+
+$$
+\mathcal {L} _ {\mathrm {C L M}} = - \sum \log P (x _ {i} \mid x _ {1}, x _ {2}, \dots , x _ {i - 1}; \theta)
+$$
+
+Tokens Selection Models. LLMLingua (Jiang et al., 2023) and its variant (Li et al., 2023; Pan et al., 2024) select tokens according to the computed distribution of the targeted LLM. Specifically, suppose  $x_{i}$  is a token in the prompt, if the probability of a token  $x_{i}$  is less than a threshold, then the token is selected to be compressed.
+
+# 3.2 Selection-  $p$
+
+Following the tokens selection models (Li et al., 2023; Jiang et al., 2023; Pan et al., 2024), Selection $p$  includes the two steps for testing, i.e. the selection (compression) step and the inference step. In the selection step, unlike LLMLingua and its variant, we instead define a selection model to select
+
+less informative tokens within the context in a discriminative way. In the inference step, the tokens selected via the selection model are passed to our targeted inference model.
+
+Selection. Assume  $p_i \in [0,1]$  denote a measure of the informativeness of token  $x_i$ . Theoretically, we can customize a deep neural model to instantiate  $p_i$ . In practice, we directly take a pre-trained language model and adopt the last layer of hidden representation  $\mathbf{h}^L$  from a pre-trained LM to the new linear projection layer. Formally, suppose  $\mathbf{h}^L = \{h_1, h_2, \dots, h_n\}$  denotes the sequence of hidden states for all tokens  $x_i$  at the last layer of a pre-trained language model. The selection model  $\mathbf{p} = \{p_1, p_2, \dots, p_n\}$  is defined as follows:
+
+$$
+\mathbf {p} = \sigma \left(\mathbf {W h} ^ {L} + b\right) \tag {1}
+$$
+
+where  $\sigma$  is the sigmoid function,  $\mathbf{W}$  and  $b$  are parameters of the projection matrix and bias vector, respectively.
+
+By using the selection model  $\mathbf{p}$ , it is straightforward to compress the context for inference: we directly prune the corresponding tokens according to our desired compression rate in retaining the top  $k\%$  of tokens in the context. To ensure the efficiency in performing compression, a single forward pass on our model creates the token preservation probability for all input tokens for simplicity.
+
+Training. Our training criterion for the selection model aims to preserve the language modeling ability of the LLM while also learning to discard tokens effectively. To this end, we employ the self-supervised approach to optimize the selection model and therefore we do not need external resources to train the selection model compared with Pan et al. (2024).
+
+To keep the training process consistent with the inference process, we first discretize the selection model  $\mathbf{p}$ . Let  $\bar{p}_i$  denote the discretized binary model of  $p_i$ . In other words,  $\bar{p}_i$  is 1 if it ranks the top  $k\%$  of tokens with the highest  $p$  values and 0 otherwise. Then we use the discretized model as a mask to define the CLM loss function as follows:
+
+$$
+\hat {\mathcal {L}} _ {\mathrm {C L M}} = - \sum \log P \left(x _ {i} \mid \bar {p} _ {1} x _ {1}, \dots , \bar {p} _ {i - 1} x _ {i - 1}; \theta\right) \tag {2}
+$$
+
+where  $\bar{p}_i x_i$  denotes whether the token  $x_i$  is masked or not depending on the value of  $\bar{p}_i$ , where the above language model  $P$  is set as the same model as that used in the selection model in Eq. 1.
+
+Training Details for Transferability. One of our goals is to achieve a transferable compression method. Therefore, the parameters that achieve the best loss on the language models in Eq. 1 and Eq. 2 in training may not be transferred to the targeted language model in inference since the language models in training and inference can be different. As a result, to ensure the better transferability of the optimized selection model, we freeze the pre-trained language model in Eq. 1 and employ LoRA (Hu et al., 2021) to train a partial parameter in the language model in Eq. 2. In summary, the CLM-based training loss is illustrated in Figure 1.
+
+# 4 Experiment
+
+# 4.1 Setting
+
+We finetune a LLaMA-2-7b model (Touvron et al., 2023) on 100M tokens from RedPajama (Togeth erAI, 2023) on split segments of 1,024 tokens via LoRA (Hu et al., 2022). To provide a comprehensive analysis of the model capabilities, our evaluation is conducted on traditional classification tasks as well as the long-context classification task.
+
+For each task, we randomly sample from the training set to construct the demonstration set for In-Context Learning (ICL), which also serves as our compression target for token selection. During inference, the compression process only needs to be computed once for all subsequent inferences on the testing instances.
+
+Traditional Classification Tasks. Following Chevalier et al. (2023), we evaluate and compare different compression models on nine classification tasks, including six tasks from SuperGlue (Wang et al., 2019). The predictions by LLMs are determined by iterating through all possible answer options for the instance and selecting the option with the minimum negative log-likelihood. The in-context demonstrations have been carefully selected to approximate a size of 750 tokens, and the complete demonstration is employed. This is referred to as the "full-shot". Since the average token length for a single demonstration varies for different tasks (e.g., a single demonstration in RTE averages about 75 tokens, resulting in a 10-shot setup under the full-shot setting), the exact number of shots differs depending on the task.
+
+To ensure a fair comparison among different compression models, a compression rate of 0.1 is used. For each task, four sets of demonstrations are selected, and the average result across these four trials is presented in Table 2. Additionally, the average accuracy of the nine tasks is presented for a clearer performance comparison.
+
+To achieve a compression rate of 0.1, a simpler approach is to directly retain one-tenth of the full-shot demonstration (e.g., using 1-shot instead of 10-shot for the RTE task) instead of performing selection at the token level. We also include this method as a baseline to further demonstrate the effectiveness of our approach.
+
+Long Context Classification Tasks. Recent research by Li et al. (2024) shows the failure of in-context learning tasks when applied to long-context scenarios. To investigate whether compression models can serve as a viable solution in long-context settings, we compare Selection- $p$  with long-context models, including LLaMA-2-7B-LongLora (Chen et al., 2023) and Long-LLaMA-code-7B (Tworkowski et al., 2023) on the BANKING77 dataset (Casanueva et al., 2020). The dataset contains 77 classes where traversing all the instances with unique labels requires approximately two thousand tokens. Evaluation is conducted at 2K, 4K, and 7K token levels, and we adopt a compression rate of  $10\mathrm{x}$  for Selection- $p$  and LLMingua-2 among all levels. Since the long in-context demonstration is used, chunking is performed for every 2,048 tokens in Selection- $p$ . The compressed results are concatenated together with a space token between each pair of chunks. We again follow the evaluation setting by Chevalier et al. (2023) on the models' prediction, with the result presented in Table 3.
+
+# 4.2 Baselines
+
+We compare the Selection-  $p$  with the following state-of-the-art compression models.
+
+- LLMLingua (Jiang et al., 2023) employs an iterative compression algorithm to filter less informative tokens based on the token-level perplexity. To further boost the performance, Jiang et al. (2023) also conducts a budget controller to allocate varying budgets across different demonstrations and questions. We find that there is also a significant discrepancy observed between the prescribed compression rate and the actual compression rate through
+
+LLMLingua API calls. For a fair comparison, we exclusively utilize the token-level prompt compression algorithm from LLMLingua in Table 2. We additionally compare LLMLingua with Selection-  $p$  equipped with Budget Controller in Section 5.5.
+
+- LLMLingua-2 (Pan et al., 2024) is derived from data distillation obtained by instructing GPT-4 to perform compression. Similar to ours, the model is trained as a binary classifier on each token, determining whether each token should be preserved.  
+- AutoCompressor (Chevalier et al., 2023) is constructed based on the RMT architecture (Bulatov et al., 2022). It compresses text into summary vectors that can be reused in subsequent segments. It is the only soft compression model adopted for comparison, considering that we followed the experiment setting for assessing the in-context learning ability of LLMs.
+
+# 4.3 Evaluation Result
+
+Traditional Classification Tasks. None of the compression models can achieve superior performance compared to the full-shot demonstration setting, which is in line with our expectations given the information loss during compression. However, certain tasks show a notable improvement when compared to both the zero-shot and full-shot approaches, e.g., all the hard compression models surpass zero-shot and full-shot by approximately  $20\%$  in the WSC task. Among all compression models, Selection- $p$  demonstrates the highest performance in conducting ICL, with an average accuracy of  $67.4\%$  across all 10 tasks as presented in Table 2. Examples of in-context demonstration before and after compression are shown in Appendix A. To demonstrate the effectiveness of our model, we have included one-tenth of the original demonstration set as the baseline. Our model significantly outperforms the baseline with a comparable number of tokens, therefore highlighting the effect of performing compression at the token level.
+
+Long Context Classification Tasks. Our model outperforms LLaMA-2-7B-LongLora, Long-LLaMA-code-7B and LLMLingua-2 at all token size levels, and achieves similar results to Li et al. (2024)'s findings on the long-context models.
+
+In addition, a growing trend with variations is observed with increasing compressed demonstrations, indicating that our model can successfully learn from additional information after compression. Examples of in-context demonstration before and after compression are presented in Appendix A.
+
+# 4.4 Transferability
+
+Compression is first performed on the demonstration set for ICL with Selection- $p$ . Subsequently, the compressed tokens are passed to a separate downstream model (i.e. LLaMA-2-13B or the black-box models) as the compressed demonstration prompt for evaluation.
+
+To LLaMA-2-13B. We follow the same setting of evaluation across different classification tasks in Section 4.3. To assess the transferability of the compression models, we compress demonstrations with Selection- $p$  and input the compressed demonstration tokens into LLaMA-2-13B. In comparing different compression models, since retraining is required for soft compression methods, no results can be obtained for AutoCompressor (Chevalier et al., 2023). In the case of LMLingua, token-level perplexity is calculated with LLaMA-2-13B instead of LLaMaA-2-7B in this experiment.
+
+Our approach outperforms all other compression models as shown in Table 4. In addition, a small deviation is observed between the  $10\mathrm{x}$  compression rate and the full shot setting, demonstrating the great transferability of our models. Notably, with Selection-  $p$  ,the tasks that outperform the full-shot setting in LLaMA-2-7b also exhibit similar patterns in LLaMA-2-13B.
+
+To Black-box Models. Taking cost into consideration, we select ChatGPT (OpenAI, 2023) and Gemini (Team, 2023) for evaluation to examine its transferability to LLMs. Traditional classification tasks often have a simple nature and the potential issue of data contamination, leading to high accuracy and causing an insignificant evaluation. Therefore, we use BANKING77 (Casanueva et al., 2020) for evaluation. Following a similar setup as described in Section 4.3, we adopt a token size of 750 for examination. However, in case the compression rate is too high, ChatGPT and Gemini are more likely to deviate from the instructions and provide task-irrelevant responses. Therefore, we adopt a compression rate of  $3\mathrm{x}$  and use the EM metric for this experiment given their black-box nature. Note that there may be variation in the result of Gemini
+
+<table><tr><td></td><td>Subj</td><td>RTE</td><td>WSC</td><td>BoolQ</td><td>MultiRC</td><td>SST-2</td><td>WIC</td><td>COPA</td><td>AG News</td><td>AVG</td></tr><tr><td>Zero-shot</td><td>49.3</td><td>58.8</td><td>43.4</td><td>67.4</td><td>52.5</td><td>67.7</td><td>50.8</td><td>52.5</td><td>63.3</td><td>56.2</td></tr><tr><td>“One-tenth”-shot</td><td>48.6</td><td>65.3</td><td>52.6</td><td>68.3</td><td>49.2</td><td>84.0</td><td>53.6</td><td>83.9</td><td>55.5</td><td>62.3</td></tr><tr><td>Full-shot</td><td>80.7</td><td>70.7</td><td>41.8</td><td>62.8</td><td>46.8</td><td>92.5</td><td>56.4</td><td>85.6</td><td>76.3</td><td>68.2</td></tr><tr><td>AutoCompressor</td><td>57.9</td><td>56.1</td><td>39.4</td><td>66.5</td><td>51.8</td><td>92.8</td><td>53.0</td><td>84.4</td><td>80.9</td><td>64.7</td></tr><tr><td>LLMLingua</td><td>56.7</td><td>60.9</td><td>63.0</td><td>69.4</td><td>50.3</td><td>69.9</td><td>51.6</td><td>76.4</td><td>61.5</td><td>62.2</td></tr><tr><td>LLMLingua-2</td><td>57.3</td><td>67.8</td><td>63.7</td><td>69.8</td><td>52.8</td><td>66.2</td><td>50.3</td><td>71.4</td><td>61.9</td><td>62.3</td></tr><tr><td>Selection-p</td><td>68.5</td><td>68.5</td><td>61.1</td><td>69.7</td><td>54.4</td><td>90.7</td><td>50.3</td><td>76.9</td><td>66.8</td><td>67.4</td></tr></table>
+
+Table 2: Evaluation result on traditional classification tasks. Four sets of random demonstrations were selected, with the average result being presented. The average result across different classification tasks is presented under AVG.  
+
+<table><tr><td></td><td>2K</td><td>4K</td><td>7K</td></tr><tr><td>LLaMA-2-7B-LongLora</td><td>0.0</td><td>0.0</td><td>0.0</td></tr><tr><td>Long-LLaMA-code-7B</td><td>0.0</td><td>0.0</td><td>0.0</td></tr><tr><td>LLMlingua-2</td><td>41.2</td><td>31.6</td><td>35.2</td></tr><tr><td>Selection-p</td><td>46.9</td><td>50.9</td><td>51.6</td></tr></table>
+
+Table 3: Evaluation result on BANKING77 with increasing in-content demonstrations tokens length.
+
+since the discrete compressed tokens sometimes trigger the SAFETY error. The prompt used is presented in Appendix B.
+
+Though the performance of our model still deviates from the full-shot setting, it achieved the best performance compared to the existing works as presented in Table 4, demonstrating fair transferability even in LLMs like ChatGPT. Surprisingly, though LLM Lingua-2 is distilled from GPT-4, it exhibits poor generalization compared to other compression models.
+
+# 5 Analysis
+
+# 5.1 Flexibility
+
+Performance with Different Number of Initial Tokens. The result in long context classification tasks in Section 4.3 shows the effectiveness of chunk-wise compression in long context. We further analyze if compression models work well in normal few-shot settings in classification tasks. In this experiment, the in-context demonstrations are selected with an approximate size of 250 tokens. The comparison to the result with the token size of 750 in Section 4.3 is presented in Table 6.
+
+Selection-  $p$  shows the best performance under the constraint of 250 tokens when compared to other compression models. Additionally, it also follows the full-shot (i.e. 750 tokens level) trend, the average performance across all classification
+
+tasks increases along with the number of provided demonstrations. On the contrary, other compression models didn't achieve an improvement in accuracy with more demonstrations. For instance, there is a drop of  $2\%$  recorded with an additional 500 tokens of information for AutoCompressor.
+
+# 5.2 Latency Analysis
+
+We analyze end-to-end latency on A100-80G GPU with the WSC task, illustrated in Table 7. Our method can achieve  $5.3\mathrm{x}$  speed up on  $10\mathrm{x}$  compressed in-context demonstration. Compared to the inference time, negligible time is required for compression on the ICL task setting, demonstrating high efficiency in adopting our models for compression. We also compared LLMLingua with the disabled content Budget Controller. It requires iterative decoding on the segmented context while Selection- $p$  only requires a single inference on all tokens and demonstrates a good performance.
+
+# 5.3 Correlation with Attention and Perplexity
+
+With the  $p$ -value ranging between 0 and 1 for each token, we further study whether any correlations exist among  $p$ , the mean attention value during the forward pass, and the tokens level perplexity (i.e. a core component in LLMLingua (Jiang et al., 2023)). Since the value of  $p$  is derived from the last hidden state of the model, we only consider the last layer mean attention of our tuned model. We employed Spearman's Rank Correlation Coefficient (Spearman, 1904) to compute the correlation between the three variables. It is calculated for different traditional classification tasks and the averaged value across tasks. The result presented in Figure 2 indicates only a weak correlation observed between the  $p$  value and the other two variables while the correlation between the last layer mean attention and perplexity is more significant. Among all tasks,
+
+<table><tr><td></td><td colspan="10">LLaMA-2-13B</td></tr><tr><td></td><td>Subj</td><td>RTE</td><td>WSC</td><td>BoolQ</td><td>MultiRC</td><td>SST-2</td><td>WIC</td><td>COPA</td><td>AG News</td><td>AVG</td></tr><tr><td>Full-shot</td><td>91.6</td><td>74.8</td><td>46.9</td><td>67.7</td><td>45.7</td><td>94.7</td><td>54.6</td><td>77.6</td><td>79.2</td><td>70.3</td></tr><tr><td>LLMingua</td><td>53.6</td><td>61.0</td><td>63.2</td><td>70.6</td><td>51.5</td><td>64.3</td><td>50.0</td><td>78.1</td><td>58.0</td><td>61.4</td></tr><tr><td>LLMingua-2</td><td>48.3</td><td>68.7</td><td>41.3</td><td>75.8</td><td>52.2</td><td>51.3</td><td>49.0</td><td>48.2</td><td>71.1</td><td>56.2</td></tr><tr><td>Selection-p</td><td>69.3</td><td>69.5</td><td>65.4</td><td>74.7</td><td>50.7</td><td>81.1</td><td>50.5</td><td>87.5</td><td>63.2</td><td>68.0</td></tr></table>
+
+Table 4: Analysis of transferability to open-source model LLaMA-2-13B. The experiment is performed on 750 tokens in-context demonstrations with a 10x compression rate.  
+
+<table><tr><td></td><td>ChatGPT
+GPT-3.5-Turbo</td><td>Gemini
+Gemini-1.0-Pro</td></tr><tr><td>Full-shot</td><td>74.2</td><td>73.3</td></tr><tr><td>LLMlingua</td><td>58.6</td><td>40.2</td></tr><tr><td>LLMlingua-2</td><td>55.7</td><td>51.9</td></tr><tr><td>Selection-p</td><td>62.9</td><td>58.9</td></tr></table>
+
+Table 5: Analysis of transferability to blackbox models (i.e. ChatGPT and Gemini). The experiment is performed on 750 tokens in-context demonstrations with a 3x compression rate.
+
+WIC demonstrates a prominently high value compared to others, this may explain the small variation in accuracy across different compression models and different experimental settings.
+
+# 5.4 Tokens Level Part-of-Speech Analysis
+
+To further interpret the rationale behind our compression models, we analyze what kinds of words are likely preserved by Selection-  $p$ . Under the discreteness of our compression result, we locate the corresponding words from the compressed tokens and obtain the Part-of-Speech (PoS) tags with an NLTK tagger. For each type of PoS tag, we compute the token preservation percentage with
+
+$$
+\frac {\left| \text {c o m p r e s s e d} _ {t} \text {k o t e n} _ {\text {t a g} _ {i}} \right|}{\left| \text {t o t a l} _ {t} \text {t o k e n} _ {\text {t a g} _ {i}} \right|}
+$$
+
+for each PoS tag  $\mathrm{tag}_{i}$ . The experiment is conducted between the compressed result and the original demonstrations among the nine traditional classification tasks with four demonstration sets per task. We analyze tags with a frequency of appearance greater than  $1\%$ .
+
+From the result presented in Figure 3, PRP and punctuations (i.e. indicating the start of the next sentence or phrase) are more likely preserved. The potential reason for the high preservation ratio on PRP (personal pronoun) likely corresponds to the pronoun resolution task of WSC. Under the task
+
+![](images/71879c9b24ec7215b4977e0372906f4bf7db657088b7556e080f0810d16064b0.jpg)  
+Figure 2: Spearman's Rank Correlation Coefficient (spearmanr) between  $p$  (p) value, mean attention (a) and token-level perplexity (ppl) across different traditional classification tasks.
+
+setting, it can be useful hints for the answer derivation.
+
+The high preservation ratio of punctuation may indicate a large redundancy in a sentence, and truncating sentence separation tokens is undesirable. Additionally, as highlighted by Wang et al. (2023), formatting information (i.e. structure of the demonstrations) matters a lot in in-context learning. However, formatting tokens (i.e. “:”) are unlikely to be preserved with Selection- $p$  compared to the original distribution in our case. In general, we also observe a higher degree of preservation of noun phrases compared to verbs.
+
+# 5.5 On Fair Comparison with LLMLingua
+
+As described in Section 4.2, LLMLingua conducts demonstration selection prior to compression at the token level, while other methods compress directly on the token level. Since the demonstration selection process can also be incorporated into other
+
+<table><tr><td></td><td>Subj</td><td>RTE</td><td>WSC</td><td>BoolQ</td><td>MultiRC</td><td>SST-2</td><td>WIC</td><td>COPA</td><td>AG News</td><td>AVG</td></tr><tr><td>Zero-shot</td><td>49.3</td><td>58.8</td><td>43.4</td><td>67.4</td><td>52.5</td><td>67.7</td><td>50.8</td><td>52.5</td><td>63.3</td><td>56.2</td></tr><tr><td>Full-shot</td><td>81.3</td><td>69.9</td><td>51.2</td><td>62.7</td><td>46.8</td><td>89.3</td><td>51.8</td><td>85.1</td><td>67.9</td><td>67.3</td></tr><tr><td>ΔFull-shot</td><td>-0.6</td><td>+0.8</td><td>-9.4</td><td>+0.1</td><td>±0</td><td>+3.2</td><td>+4.6</td><td>+0.5</td><td>+8.4</td><td>+0.9</td></tr><tr><td>AutoCompressor</td><td>56.2</td><td>61.5</td><td>44.2</td><td>68.3</td><td>52.7</td><td>93.0</td><td>51.5</td><td>83.6</td><td>76.1</td><td>65.2</td></tr><tr><td>ΔAutoCompressor</td><td>+1.7</td><td>-5.4</td><td>-4.8</td><td>-1.8</td><td>-0.9</td><td>-0.2</td><td>+1.5</td><td>+0.8</td><td>+4.8</td><td>-0.5</td></tr><tr><td>LLMLingua*</td><td>55.6</td><td>61.4</td><td>61.3</td><td>68.2</td><td>53.1</td><td>81.1</td><td>50.2</td><td>75.8</td><td>70.9</td><td>64.2</td></tr><tr><td>ΔLLMLingua</td><td>+1.1</td><td>-0.5</td><td>+1.7</td><td>+1.2</td><td>-2.8</td><td>-11.2</td><td>+1.4</td><td>+0.6</td><td>-9.4</td><td>-2.0</td></tr><tr><td>LLMLingua-2</td><td>52.4</td><td>65.3</td><td>63.9</td><td>66.1</td><td>50.9</td><td>82.9</td><td>50.8</td><td>77.8</td><td>56.3</td><td>62.9</td></tr><tr><td>ΔLLMLingua-2</td><td>+4.9</td><td>+2.5</td><td>-0.2</td><td>+3.7</td><td>+1.9</td><td>-16.7</td><td>-0.5</td><td>-6.4</td><td>+5.6</td><td>-0.6</td></tr><tr><td>Selection-p</td><td>65.7</td><td>65.5</td><td>58.7</td><td>67.5</td><td>54.3</td><td>81.3</td><td>50.4</td><td>77.9</td><td>68.3</td><td>65.5</td></tr><tr><td>ΔSelection-p</td><td>+2.8</td><td>+3.0</td><td>+2.4</td><td>+2.3</td><td>+0.1</td><td>+9.4</td><td>-0.1</td><td>-1.0</td><td>-1.5</td><td>+1.9</td></tr></table>
+
+Table 6: Performance with different number of demonstrations from about 250 tokens to about 750 tokens.  $\Delta$  refers to the performance enhancement that can be achieved by increasing the demonstration tokens size to 750.  
+
+<table><tr><td></td><td>1x</td><td>2x</td><td>5x</td><td>10x</td></tr><tr><td>End-to-End without compression</td><td>298.6</td><td></td><td></td><td></td></tr><tr><td>End-to-End with Selection-p</td><td></td><td>167.0 (1.8x)</td><td>81.6 (3.7x)</td><td>55.6 (5.3x)</td></tr><tr><td>LLMLingua per demonstrations set</td><td>-</td><td>0.82</td><td>0.82</td><td>0.81</td></tr><tr><td>Selection-p per demonstrations set</td><td>-</td><td>0.68</td><td>0.67</td><td>0.67</td></tr></table>
+
+compression models, we only utilize the modified version of LLMLingua in the previous experiments to ensure a fair comparison.
+
+In this section, we further analyze the performance by comparing our proposed method, equipped with the Budget Controller, with the whole LLMLingua to provide a comprehensive analysis. We follow the setting described in Section 4.3 and select the WSC task for our experiment. To illustrate, in the original demonstration set consisting of 16 demonstrations, the LLMLingua API retains only four demonstrations. This leads to two options with Selection-  $p$ : (1) continuously applying the  $10\mathrm{x}$  compression directly to the filtered set of four demonstrations and resulting in a final compression rate of  $38\mathrm{x}$ , and (2) adjusting the compression rate of Selection-  $p$  to achieve a final compression rate of  $10\mathrm{x}$ .
+
+The result presented in Table 8 demonstrates the significant impact of the Budget Controller. Similar trends in performance for both Selection-  $p$  and LLMLingua are observed (i.e., a decrease in performance on the WSC task). Notably, the performance of Selection-  $p$  surpasses LLMLingua after equipping with the Budget Controller.
+
+Table 7: Latency(s) comparison on WSC in 750 tokens level with about 16 demonstrations. We present the averaged complete end-to-end inference with and without Selection-  $p$  among four sets of demonstrations. Comparison is conducted with LLMLingua which also builds upon the LLaMA-2-7B backbone, with the averaged compression time of the in-context demonstrations being presented.  
+
+<table><tr><td></td><td>WSC</td><td>rate</td></tr><tr><td>Selection-p</td><td>61.1</td><td>10x</td></tr><tr><td>LLMLingua</td><td>63.0</td><td>10x</td></tr><tr><td>Selection-p (+ Budget Controller)</td><td>47.6</td><td>10x</td></tr><tr><td>Selection-p (+ Budget Controller)</td><td>57.0</td><td>38x</td></tr><tr><td>LLMLingua (whole)</td><td>44.7</td><td>10x</td></tr></table>
+
+Table 8: Comparison with LLMLingua on Budget Controller. Adopting different strategies in equipping Selection-  $p$  with Budget Controller, leads to the two different compression rates (rate) of  ${10}\mathrm{x}$  and  ${38}\mathrm{x}$  .
+
+Furthermore, there is a disparity between the instructed compression rate and the actual compression rate in LLMLingua. The target size for compressed tokens is 75, while LLMLingua typically achieves an average compressed token size of around 192.1, which is more than 1.5 times higher than the desired rate across all classification tasks.
+
+# 6 Conclusion
+
+We introduce a simple yet effective self-supervised approach in context compression and conduct evaluation across 10 classification tasks in both few-shot and long-context settings. Our approach also demonstrated great transferability to both the open-
+
+![](images/a77d617378c73b3204182d5fe86570211ad68f6deafbea4efd61db42f87bae3d.jpg)  
+Figure 3: Analysis of the token preservation percentage with respect to different types of Part-of-Speech tags under  $10\mathrm{x}$  compression rate.
+
+source (i.e. LLaMA-2-13B) and black-box models (i.e. ChatGPT and Gemini), with performance surpassing the existing state-of-the-art compression models. Analysis is also conducted among different compression rates and demonstration lengths. With a 10x compression rate, our model only shows a 0.8-point drop in performance across different traditional classification tasks with a 5.3x speedup. Through experiments in long-context settings, our work also presents the possibility of addressing the in-context learning issue of the recent long-context models. Both efficiency enhancement as well as performance preservation are shown in our model.
+
+Acknowledgement This research has been made possible by the Hong Kong PhD Fellowship provided to Tsz Ting Chung and the Research Impact Fund project R6003-21 provided by the Research Grants Council of Hong Kong to Dit-Yan Yeung.
+
+# Limitations
+
+Under the consideration of cost, we did not perform further analysis on other LLMs apart from ChatGPT and Gemini. In addition, our model which builds up LLaMA-2-7B does not achieve better latency than models like LLMLingua-2 and AutoCompressor. Under the ICL setting, minimal time is required for compression, leading to insignificance in end-to-end inference time. While AutoCompressor offers better latency, its soft compression nature limits its applicability to other LLMs. Overall, our experiments across various tasks and settings demonstrate better performance and transferability, with the benefits outweighing the latency issue.
+
+# References
+
+Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. 2023. Token merging: Your vit but faster. In The
+
+Eleventh International Conference on Learning Representations.
+
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
+
+Aydar Bulatov, Yuri Kuratov, and Mikhail Burtsev. 2022. Recurrent memory transformer. In Advances in Neural Information Processing Systems.
+
+Inigo Casanueva, Tadas Temčinas, Daniela Gerz, Matthew Henderson, and Ivan Vulić. 2020. Efficient intent detection with dual sentence encoders. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 38-45, Online. Association for Computational Linguistics.
+
+Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. 2023. Longlora: Efficient fine-tuning of long-context large language models. ArXiv preprint, abs/2309.12307.
+
+Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. 2023. Adapting language models to compress contexts. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3829-3846, Singapore. Association for Computational Linguistics.
+
+Tao Ge, Hu Jing, Lei Wang, Xun Wang, Si-Qing Chen, and Furu Wei. 2024. In-context autoencoder for context compression in a large language model. In The Twelfth International Conference on Learning Representations.
+
+Saurabh Goyal, Anamitra Roy Choudhury, Saurabh Raje, Venkatesan T. Chakaravarthy, Yogish Sabharwal, and Ashish Verma. 2020. Power-bert: Accelerating BERT inference via progressive word-vector
+
+elimination. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 3690-3699. PMLR.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.  
+Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.  
+Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2023. LLMLingua: Compressing prompts for accelerated inference of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13358-13376, Singapore. Association for Computational Linguistics.  
+Gyuwan Kim and Kyunghyun Cho. 2021. Length-adaptive transformer: Train once with length drop, use anytime with search. 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 6501-6511, Online. Association for Computational Linguistics.  
+Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. 2022. Learned token pruning for transformers.  
+Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. 2024. Long-context llms struggle with long in-context learning.  
+Yucheng Li, Bo Dong, Frank Guerin, and Chenghua Lin. 2023. Compressing context to enhance inference efficiency of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6342-6353, Singapore. Association for Computational Linguistics.  
+Sewon Min, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Noisy channel language model prompting for few-shot text classification. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5316-5330, Dublin, Ireland. Association for Computational Linguistics.  
+Ali Modarressi, Hosein Moebbi, and Mohammad Taher Pilehvar. 2022. AdapLeR: Speeding up inference by adaptive length reduction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1-15, Dublin, Ireland. Association for Computational Linguistics.
+
+Jesse Mu, Xiang Li, and Noah Goodman. 2023. Learning to compress prompts with gist tokens. In Advances in Neural Information Processing Systems, volume 36, pages 19327-19352. Curran Associates, Inc.  
+OpenAI. 2023. GPT-4 technical report. ArXiv preprint, abs/2303.08774.  
+Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor Ruhle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. 2024. Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression.  
+Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. 2021. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 13937-13949.  
+C. Spearman. 1904. The proof and measurement of association between two things. American Journal of Psychology, 15:88-103.  
+Gemini Team. 2023. Gemini: A family of highly capable multimodal models.  
+TogetherAI. 2023. Redpajama: An open dataset for training large language models.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models.  
+Szymon Tworkowski, Konrad Staniszewski, Mikołaj Pacek, Yuhuai Wu, Henryk Michalewski, and Piotr Miłos. 2023. Focused transformer: Contrastive training for context scaling.  
+Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 3261-3275.  
+Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. 2023. Label words are anchors: An information flow perspective for understanding in-context learning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9840-9855, Singapore. Association for Computational Linguistics.
+
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models.  
+David Wingate, Mohammad Shoeybi, and Taylor Sorensen. 2022. Prompt compression and contrastive conditioning for controllability and toxicity reduction in language models. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5621-5634, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+# A Examples Illustration of Compression Results
+
+Figure 4 shows examples of two traditional classification tasks (i.e. Subj and WSC) with Selection- $p$  under  $10\mathrm{x}$  compression, displaying both the compressed results and the original demonstration sets before compression. Additionally, Figure 5 illustrates an example for BANKING77 under  $10\mathrm{x}$  compression.
+
+# B Prompt of Evaluation on BANKING77
+
+In our prompt to ChatGPT, we first list out all 77 labels and then provide a list of demonstrations. The template is detailed below,
+
+Answer in activate_my_card or age_limit or apple_pay_or_google_pay or atm_support or automatic_top_up or balance_not UPDATED_after_bank_transfer or ... or wrong_amount_of_cash_received or wrong_exchange_rate_for_cash_withdrawal.
+
+Context: <context>
+
+Answer: <answer>
+
+# C Training Details
+
+We use LLaMA-2-7B for compression (i.e. token selection). It takes roughly 50 hours on a single A100 GPU to train on 100M tokens from RedPajama.
+
+# D Detailed Latency Analysis
+
+For the token-based compression models, the time needed for compression per demonstration set on the WSC task is presented in Table 9.
+
+<table><tr><td>LLMLingua-2</td><td>Selection-p</td><td>LLMLingua</td></tr><tr><td>0.15</td><td>0.67</td><td>0.81</td></tr></table>
+
+Under the ICL setting, the same demonstration set is used consistently, and compression on the demonstration set only needs to be computed once for all subsequent inferences. This contributed to the short compression time in the complete end-to-end process.
+
+Considering that the time taken for compression on the demonstration set is negligible when compared to the time required for inference (approximately at a ratio of 0.01), the end-to-end inference time for the token-selection-based compression models is roughly the same. The end-to-end inference time result for the WSC task on LLaMA2-7B is shown in Table 10. The table is presented in order of the inference time for clarity.
+
+Table 9: Time needed for compression per demonstration set on the WSC task.  
+
+<table><tr><td>AutoCompressor</td><td>LLMLingua-2</td><td>Selection-p</td><td>LLMLingua</td></tr><tr><td>22.73</td><td>55.6</td><td>55.6</td><td>55.6</td></tr></table>
+
+Table 10: End-to-end inference time on the WSC task.
+
+# [Subj] Original Demonstrations:
+
+input: each of the principals has a radically different way of dealing with it. type: objective
+
+input: well-intentioned though it may be, its soap-opera morality tales have the antiseptic, preprogrammed feel of an after-school special. type: subjective
+
+input: an astonishing feat for a major star let alone a 27 year old from pickum , south carolina who only two years ago was sleeping in a cardboard box in the back alleys of detroit with her mother , connie , and her uncle clutch , while playing guitar on the streets for spare change .type: objective
+
+input: may is a young strange girl who had a very disturbed childhood and does not still know the meaning of true friendship or love. type: objective
+
+# [Subj] Compressed demonstrations (10x):
+
+input:-int it may- mor exc lord treasure planet- the ste moments- it thr en a r with he and I crowd gru a-erm own - into impossibly ris gar "ir wh sh de prere prede it mag m ro philosophvag rede thes v de spcer twz-passer aston fe two with cl for dist and still
+
+# [WSC] Original Demonstrations:
+
+Question: In the sentence "James asked Robert for a favor but he was refused.", does the pronoun 'he' refer to Robert?  
+Answer: no
+
+Question: In the sentence "What about the time you cut up tulip bulbs in the hamburgers because you thought they were onions?", does the pronoun 'they' refer to tulip bulbs? Answer: yes
+
+Question: In the sentence "When Mr. Bond, the veterinarian, came to look at the black horse that lay groaning on the grass, he felt him all over, and shook his head; one of his legs was broken.", does the pronoun 'his' refer to the black horse?  
+Answer: no
+
+Question: In the sentence "Sam took French classes from Adam, because he was eager to speak it fluently," does the pronoun 'he' refer to Adam?
+
+Answer: no
+
+# [WSC] Compressed demonstrations (10x):
+
+" pron:: " tul pron ' tul: he pron ' told P which P. He have pron ' P a the would only Gru une pron pronI put cfr It pron refriger man pronJohn wheng. He very im Wainws d Fol he pron veteran gro his pron the repa pron w t win gro " he pron '
+
+Figure 4: Illustration of the compression result by Selection- $p$  for Subj and WSC tasks under  $10\mathrm{x}$  compression rate. Compression is performed with 19 demonstrations for Subj while it is performed with 16 demonstrations for WSC with total sum of about 750 tokens respectively.
+
+# [BANKING77] Original Demonstrations:
+
+Context: Why did using an ATM cause me to be charged an additional fee?
+
+Answer: cash Withdrawal charge
+
+Context: I asked for a refund but its not here yet.
+
+Answer: Refund_not Showing_up
+
+Context: is there a reason i need to verify top up
+
+Answer: verify_top_up
+
+···
+
+Context: There is a payment on my card that I do not recognize. I've never seen the name on the transaction before.
+
+Answer: card-payment_not_recognised
+
+Context: I happened to forget my passcode
+
+Answer: passcode_forgotten
+
+Context: I made a cash withdrawal and it is still listed as a pending transaction.
+
+Answer: pending_cash_withdrawal
+
+# [BANKING77] Compressed demonstrations (10x):
+
+c_with asked ref Refnoting_up_recogn_c wrong_rece_transaction_chargtw unblock activ activ_not_fe收费标准 : ex sho Please revert_ the cardWr: I: pending_: the card.not_recogn: passf:_c
+
+Figure 5: Illustration of the compression result by Selection- $p$  for BANKING77 under 10x compression rate. Compression is performed with 27 demonstrations with total sum of about 750 tokens.

2024/Selection-p_ Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1925fa8a83bf72e64826a8a441eaa0aa3c6167b494492cefd60b76733b96f82b
+size 379446

2024/Selective Annotation via Data Allocation_ These Data Should Be Triaged to Experts for Annotation Rather Than the Model/bcc9d306-7f45-4b34-b97c-c05fad825ac3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:458af176634060dbb2f5d103ad0992d7420a17fff7251f4f001b66b9a9581ac7
+size 2073643

2024/Selective Annotation via Data Allocation_ These Data Should Be Triaged to Experts for Annotation Rather Than the Model/full.md

@@ -0,0 +1,439 @@
+# Selective Annotation via Data Allocation: These Data Should Be Triaged to Experts for Annotation Rather Than the Model
+
+Chen Huang\*, Yang Deng\*, Wenqiang Lei\*, Jiancheng Lv\*, Ido Dagan
+
+$\spadesuit$  Sichuan University Singapore Management University Bar-Ilan University
+
+Engineering Research Center of Machine Learning and Industry Intelligence, Ministry of Education, China
+
+huangc.scu@gmail.com ydeng@smu.edu.sg
+
+{wenqianglei, lvjiancheng} @scu.edu.cn dagan@cs.biu.ac.il
+
+# Abstract
+
+To obtain high-quality annotations under limited budget, semi-automatic annotation methods are commonly used, where a portion of the data is annotated by experts and a model is then trained to complete the annotations for the remaining data. However, these methods mainly focus on selecting informative data for expert annotations to improve the model predictive ability (i.e., triage-to-human data), while the rest of the data is indiscriminately assigned to model annotation (i.e., triage-to-model data). This may lead to inefficiencies in budget allocation for annotations, as easy data that the model could accurately annotate may be unnecessarily assigned to the expert, and hard data may be misclassified by the model. As a result, the overall annotation quality may be compromised. To address this issue, we propose a selective annotation framework called SANT. It effectively takes advantage of both the triage-to-human and triage-to-model data through the proposed error-aware triage and bi-weighting mechanisms. As such, informative or hard data is assigned to the expert for annotation, while easy data is handled by the model. Experimental results show that SANT consistently outperforms other baselines, leading to higher-quality annotation through its proper allocation of data to both expert and model workers. We provide pioneering work on data annotation within budget constraints, establishing a landmark for future triage-based annotation studies.
+
+# 1 Introduction
+
+Creating a high-quality and fully annotated dataset by human experts is criticized as being expensive. It is impossible to complete the annotation if we are operating on a limited budget (Chen et al., 2021), especially when dealing with large-scale data corpora on the web (Hedderich et al., 2021; Feng and Lapata, 2008). In this paper, we study the data
+
+annotation problem within the constraints of limited budgets, where the goal is to achieve high-quality annotations despite having insufficient budgets to hire experts for annotating the entire dataset. To tackle this challenge, a typical solution is to conduct semi-automatic annotation (Chen et al., 2020b; Yao et al., 2023) (cf. Fig. 1 (middle)). It typically allocates the budget to hire experts for annotating a portion of the data, and then utilizes the annotated data to train a prediction model for completing the annotations for the remaining data. The choice of the model can vary, ranging from sophisticated models with large parameters (Bryant et al., 2017; Schulz et al., 2019; Hedderich et al., 2021) to lightweight models (Chen et al., 2020a; Desmond et al., 2021a) depending on the available hardware budgets. However, such methods often overlook the data allocation problem, i.e., the process of deciding which data should be annotated by human experts and which should be left for the prediction models to annotate. This may lead to inefficiencies in budget allocation for annotations, as certain data that the model could accurately annotate is unnecessarily assigned to experts. Consequently, the overall annotation quality within the same budget may be compromised.
+
+In the light of this challenge, we take the first step to re-formulate the semi-automatic annotation under limited budgets as a data triage problem, which aims to optimize the annotation quality under limited budgets by determining the allocation of data to human annotator (namely triage-to-human data) and data to be assigned for model prediction (namely triage-to-model data). Existing semi-automatic annotation methods (Desmond et al., 2021a; Su et al., 2022; Li et al., 2023) mainly focus on selecting unlabeled data for expert annotations to improve the model predictive ability (i.e., triage-to-human data) through random strategies or active learning (AL), while the rest of data is indiscriminately assigned to model annotation (i.e.,
+
+![](images/7e3422a2f2cf777e33e98ef63926606d233eca3c750ab77c05edddca5b060b6c.jpg)  
+Figure 1: Semi-automatic annotation focuses on triage-to-human data, overlooking triage-to-model data.
+
+triage-to-model data). However, as shown in Fig. 1 (bottom), even if we constantly improve the predictive ability via triage-to-human data, the model still inevitably makes errors in its annotations. Not to mention that the data allocation strategies of existing methods may not effectively improve the model predictive ability (Tang and Huang, 2019; Mindermann et al., 2022). To address this issue, determining what data should be assigned for model annotation (triage-to-model data) is crucial. A model could achieve high-quality annotations if assigned with the data that is easy for model to predict, while experts annotate the hard ones. Therefore, when solving the data triage challenge, it's important to take both triage-to-human and triage-to-model data into account. In this way, the budget utilization could be improved through efficiently allocating both human and model annotations.
+
+To this end, we propose a novel Selective ANnoTation framework (SANT), which optimizes the utilization of limited annotation budgets by balancing the importance of triage-to-human and triage-to-model data. Firstly, an AL-based mechanism is deployed to assign higher weights to data that help improve the model predictive ability. Secondly, an Error-Aware Triage (EAT) mechanism is proposed to assess the "hardness" weight of data, with the aim of redirecting the focus toward triage-to-model data. In specific, EAT dynamically estimates the probability of model error using triage-to-human data from previous rounds. Finally, a bi-weighting mechanism is adopted to aggregate different advantages of prioritizing triage-to-human and triage-to-model data. Such mechanism adaptively adjusts the importance of data so that informative or hard-to-predict data is assigned to humans for annotation. As such, the annotation budget are efficiently utilized to secure the model predictive ability and to leverage the strengths of the expert on hard data annotation simultaneously.
+
+Our experiment results show that SANT consistently improves annotation quality across various annotation budgets, surpassing the best semiautomatic annotation method by an average of  $+0.50\%$ ,  $+4.86\%$ , and  $+4.54\%$  on the three annotation tasks, respectively. Recently, as large language models (LLMs) possess strong zero-shot and few-shot capabilities for data annotation (He et al., 2023; Gilardi et al., 2023; Ding et al., 2023), we further adopt ChatGPT as a strong automatic annotation baseline. Notably, SANT exhibits a better annotation quality over ChatGPT, even with the aid of few-shot in-context learning and Chain-of-Thought (CoT) prompting (Brown et al., 2020; Wei et al., 2022). These results highlight that proper allocation of data to both human and model workers results in higher-quality annotation. Further analysis highlights the importance of the triage-to-model data, which are more beneficial to obtain higher-quality annotations compared to the triage-to-human ones. In conclusion, we pioneer work on data annotation with limited budgets by formulating a data triage problem. We set a landmark for future triage-based annotation methods. Our contributions are as follows:
+
+- We present the first work to study the roles of both triage-to-human and triage-to-model data. We verify the importance of triage-to-model data, which is largely overlooked by existing studies.  
+- We re-formulate semi-automatic annotation as data triage and propose a selective annotation framework, achieved by optimizing two novel mechanisms (i.e., EAT and bi-weighting mechanisms) for utilizing limited budgets.  
+- We show that proper allocation of data to both the expert and model workers results in higher-quality annotation. SANT effectively balances the advantages of two factors: improving model predictive ability (i.e., triage-to-human data) and triaging data based on the model's predictive ability (triage-to-model data).
+
+# 2 Related Work
+
+Data Annotation. Human annotations are known to be expensive (Chen et al., 2021; Hedderich et al., 2021; Feng and Lapata, 2008). To this end, many studies have explored human-machine cooperative data annotation (Lu et al., 2024; Li et al., 2023; Ding et al., 2024; Huang et al., 2024a,b). In this case, a portion of the data is designated for human annotation (triage-to-human data), while
+
+the remaining data is allocated for model annotation (triage-to-model data). To further enhance the model's predictive capabilities, existing methods are often formulated as a "model training and prediction" paradigm, where the model iteratively learns from the annotated data and then makes predictions. (Dalvi et al., 2016; Li et al., 2015; Bryant et al., 2017; Chen et al., 2020b; Wang et al., 2023; Tan et al., 2023). Considering the hardware budgets and time efficiency requirements, light-weighted models (Desmond et al., 2021a; Chen et al., 2020a), such as label propagation (Zhou et al., 2003) and MLP are preferred. Recently, uncertainty-based methods have been used to select a data subset for human annotations (Zhang et al., 2021; Desmond et al., 2021a; Su et al., 2022; Ein-Dor et al., 2020), improving the overall annotation quality by refining the model using more informative data. However, current methods concentrate solely on selecting triage-to-human data, while neglecting the importance of triage-to-model data. This fails to optimize data allocation effectively. In contrast, SANT considers both types of data.
+
+Active learning. Current data allocation practices in the field of data annotation draw inspiration from the active learning (Ren et al., 2020), where data with high uncertainty are prioritized for human annotation. This aligns with active learning's preference for informative or representative data (Desmond et al., 2021b; Loquercio et al., 2020). To our knowledge, no existing research has explored the application of other active learning techniques for data allocation in annotation. Our work, SANT, extends this approach by introducing a novel perspective. It prioritizes triage-to-model data, essentially introducing the anti-current learning (Braun et al., 2017) into the task of data allocation (i.e., we assign hard-to-predict data for human annotation). SANT seamlessly integrates active learning algorithms with the anti-current learning approach through our proposed bi-weighting mechanism. Our experiments demonstrate the effectiveness of this novel framework in enhancing the quality of data annotation.
+
+# 3 SANT: Selective Annotation
+
+We introduce the proposed Selective Annotation (SANT) framework. In Section 3.1, we provide the problem definition of the data annotation process, while Section 3.2 introduces the proposed framework SANT and two novel mechanisms. In
+
+Appendix F, we introduce the learning details of our framework. Following previous works on semiautomatic annotation (Chen et al., 2020a; Desmond et al., 2021a; Hedderich et al., 2021) and human-model interaction (Hwa, 2000; Kristjansson et al., 2004; Huang et al., 2023), we assume the cost required for human annotation is constant for each data point, and estimate the annotation budgets using the number of data points.
+
+# 3.1 Problem Definition
+
+We reformulate the data annotation under limited budgets as a data triage problem, where a data triage module  $d$  assigns data to either the expert or the model  $f$  on-the-fly. The triage module outputs a triage signal of either 0 or 1 to delegate the annotation duty. If the signal is 0, the model is responsible for data annotation. Conversely, if the signal is 1, the human is assigned, and the model is updated using previous human-annotated data.
+
+The annotation process terminates when one of two conditions is met: either the limited budgets are exhausted or all the data is annotated. In the first case, if the budgets are depleted before completing the entire dataset, the model annotator must continue with the remaining annotations. In the second case, it is possible for all data to be annotated before the budgets are exhausted if the amount of data assigned to the human during process is less than the budgets. To maximize the utilization of the human budget, a portion of the model-annotated data, selected by the customized data triage strategy, is re-allocated to the human annotator until their budget is depleted.
+
+# 3.2 Proposed Framework
+
+Semi-automatic annotation methods use active learning (AL) in the data triage module, which primarily prioritize triage-to-human data but overlook the triage-to-model data. As a result, they fail to efficiently and directly optimize budget utilization. To this end, SANT introduces two innovative mechanisms, namely EAT and bi-weighting mechanisms, which consider both human and model-based triage data for budget optimization.
+
+# 3.2.1 AL-based mechanism
+
+Semi-automatic annotation and SANT are flexible to any off-the-shelf AL methods<sup>1</sup>, such as,
+
+leveraging the uncertainty of the model prediction (Desmond et al., 2021b) or using a differentiable neural network (Wang et al., 2022). Formally speaking, given data  $x_{t}$  and the model  $f_{t}$  at round  $t$ ,  $x_{t}$  is assigned to the expert if the AL-based mechanism  $d_{t}^{AL}$  gives  $x_{t}$  a large weight. If the AL-based mechanism is implemented using a learnable neural network, both AL and  $f_{t}$  are jointly optimized/updated, otherwise only  $f_{t}$  is updated. Note that the AL-based mechanism solely focuses on triage-to-human data, which limits its ability to optimize budget utilization directly and efficiently.
+
+# 3.2.2 Error-aware triage mechanism (EAT)
+
+Unlike existing AL-based semi-automatic annotation, we propose EAT to prioritize triage-to-model data by assigning hard-to-predict data to the expert and reserving easy-to-predict data to the model. Specifically, EAT assigns data  $x_{t}$  to the expert if the EAT-based mechanism  $d_{t}^{EAT}$  gives  $x_{t}$  a high weight, which is determined by the probability of  $f_{t}$  misclassifying  $x_{t}$ . The model  $f_{t}$  and EAT are then jointly optimized/updated. If  $d_{t}^{EAT}$  does not assign a high weight to  $x_{t}$ , then the model  $f_{t}$  acts as the annotator without requiring human involvement. Notably, EAT can be parameterized using any neural network, depending on the available hardware budgets. Here, EAT's input and learning objectives are presented below.
+
+Input of EAT. Instead of learning from training data independently (Mozannar and Sontag, 2020; Wilder et al., 2021; Raghu et al., 2019; Okati et al., 2021), to enhance EAT's input features, we propose capturing the relationship between input data and its neighboring data from the same batch. This includes the use of weighted neighborhood entropy, which measures the model's predictive uncertainty on the local area of input data  $x_{t}$ . To calculate this entropy, we retrieve the top-  $k$  neighbors  $N_{t}(x_{t})$  and calculate the cosine similarity between  $x_{t}$  and each neighbor  $x_{i} \in N_{t}$ . We store these similarity values in vector  $C_{t}(x_{t}) \in R^{k}$ . Then, we calculate the entropy of the model prediction  $f_{t}(x_{i})$  for each  $x_{i} \in N_{t}(x_{t})$  and store them in vector  $E_{t}(x_{t}) \in R^{k}$ . Finally, we obtain the weighted neighborhood entropy of  $x_{t}$  by using the element-wise multiplication operator  $\odot$  on vectors  $E_{t}(x_{t})$  and  $C_{t}(x_{t})$ . The input of EAT is then the concatenation of the embedding of unlabeled data  $x_{t}$ , the model output  $f_{t}(x_{t})$ , and the weighted neighborhood entropy.
+
+Learning objective of EAT. EAT is designed to allocate hard-to-predict data, which are beyond the
+
+model's predictive ability, to the expert. To achieve this, EAT estimates the likelihood of the model making errors on unlabeled data  $x_{t}$ . Specifically, EAT is required to minimize the following loss.
+
+$$
+L _ {d} = \sum_ {\left\{x _ {t}, y _ {t} \right\}} \ell_ {d} \left(\mathbb {1} \left[ y _ {t} \neq \max  \left(f _ {t} \left(x _ {t}\right)\right) \right], d _ {t} ^ {E A T} \left(x _ {t}\right)\right). \tag {1}
+$$
+
+Here,  $\mathbb{1}[\cdot ]$  is an indicator function.  $y_{t}\in R$  and  $f_{t}(x_{t})\in R^{\mathbf{C}}$  are the ground truth from the human and predicted outputs of the model, where  $\mathbf{C}$  is the number of classes. Also, the operator  $\max (\cdot)$  outputs the index of the maximal value of the input vector. In our experiments,  $\ell_d$  comes in the form of the negative log-likelihood or binary cross-entropy loss, depending on the annotation tasks. Moreover, We further follow recent works on few-shot learning (Wang et al., 2018; Li et al., 2020) and utilize the max-margin loss to improve SANT's ability to learn and generalize from limited labeled data on the fly. Apart from the aforementioned loss  $L_{d}$ , we add the following max-margin loss  $L_{m}$  (See Appendix C for ablation experiments on  $L_{m}$ ).
+
+$$
+L _ {m} = \max  (0, \alpha + \bar {L} _ {f} ^ {m} - \bar {L} _ {f} ^ {h}),
+$$
+
+$$
+\overline {{L _ {f} ^ {m}}} = \frac {\sum_ {\{x _ {t} , y _ {t} \}} \mathbb {1} \left[ d _ {t} ^ {E A T} \left(x _ {t}\right) <   0 . 5 \right] \ell_ {f} \left(x _ {t} , y _ {t}\right)}{\sum_ {\{x _ {t} , y _ {t} \}} \mathbb {1} \left[ d _ {t} ^ {E A T} \left(x _ {t}\right) <   0 . 5 \right]}, \tag {2}
+$$
+
+$$
+\overline {{L _ {f} ^ {h}}} = \frac {\sum_ {\{x _ {t} , y _ {t} \}} \mathbb {1} [ d _ {t} ^ {E A T} (x _ {t}) \geq 0 . 5 ] \ell_ {f} (x _ {t} , y _ {t})}{\sum_ {\{x _ {t} , y _ {t} \}} \mathbb {1} [ d _ {t} ^ {E A T} (x _ {t}) \geq 0 . 5 ]}.
+$$
+
+Here,  $\alpha > 0$  is a margin hyper-parameter. Given the fixed model, EAT carefully allocate the expert and the model annotations so that it ensures that the average loss of the model on the triage-to-human data  $\overline{L_f^h}$  is much higher than the triage-to-model data  $\overline{L_f^m}$ . Finally, the learning objective of EAT is given by  $L_{EAT} = L_d + L_m$ .
+
+# 3.2.3 Bi-weighting mechanism
+
+The AL-based mechanism focuses on triage-to-human data to improve the model's predictive ability. However, it may still struggle with hard data beyond the model capacity. In contrast, our proposed EAT-based mechanism prioritizes triage-to-model data and reserves hard data for human annotation, which allows the model to handle relatively easy data. However, models trained on hard data may have weaker predictive ability. To find the optimal trade-off, the bi-weighting mechanism aims to ensure that the model is trained properly and can annotate relatively easy data effectively.
+
+The bi-weighting mechanism aggregates the advantages of prioritizing both triage-to-human and
+
+triage-to-model data by introducing a simple yet efficient method. It scores  $x_{t}$  using  $d_{t}^{bi}(x_{t})$ , which is the product of  $d_{t}^{AL}(x_{t}) \in [0,1]$  and  $d_{t}^{EAT}(x_{t}) \in [0,1]$ , to adjust the importance of data. This ensures that informative or hard data are assigned to humans for annotation, and that the model is trained properly. Considering the sample efficiency of AL is more advantageous when the model is trained based on a few labeled data, we further control the importance of the AL score  $d_{t}^{AL}(x_{t})$  by  $\tau \geq 0$ , which is the percentage of annotated data in the total data. Given a function  $\beta(\tau) = e^{\tau - T_0}$  parameterized by a hyper-parameter  $T_0 \in [0,1]$  and the total amount of dataset  $|X|$ , the weight of  $x_{t} \in X$  is given by the following equation.
+
+$$
+d _ {t} ^ {b i} \left(x _ {t}\right) = \left(d _ {t} ^ {A L} \left(x _ {t}\right)\right) ^ {\eta (t)} * d _ {t} ^ {E A T} \left(x _ {t}\right) \tag {3}
+$$
+
+$$
+\eta (t) = \beta (t / | X |).
+$$
+
+As such, the bi-weighting mechanism highlights the importance of  $d_t^{AL}(x_t)$  in the early stages of the annotation process (i.e., when  $\tau < T_0$ ) and decreases it as the annotation process progresses. This allows for efficient use of the AL score. Notably, the bi-weighting mechanism is efficient because it allows for the fusion of AL and EAT in a post-hoc manner.
+
+Optimization of SANT. We optimize the model annotator  $f_{t}$ , EAT  $d_{t}^{EAT}$ , and AL method  $d_{t}^{AL}$  independently. The loss function for AL, denoted as  $L_{AL}$  (if applicable), is included in the overall loss function  $\mathcal{L}$  of SANT, which is the sum of  $L_{f}$ ,  $L_{EAT}$ , and  $L_{AL}$ . Note that all human-annotated data from past human-model interactions are utilized for optimization. For optimization details, please refer to Appendix F.
+
+# 4 Evaluation
+
+In this section, we conduct extensive experiments to assess the effectiveness of SANT. Given limited annotation budgets, we evaluate if SANT is more desirable to obtain high-quality annotations, compared to existing annotation methods (cf. Section 4.2). Furthermore, we comprehensively analyze the advantages of SANT and uncover the characteristics of triage-to-human and triage-to-model data in data annotation task (cf. Section 4.3).
+
+# 4.1 Experiment Setup
+
+Tasks & Datasets. To assess SANT, we perform three annotation tasks, namely sentiment annotation, knowledge graph completion, and tagging.
+
+These tasks have been highlighted as crucial in various web applications (Yao et al., 2021; Marcos-Pablos and Garcia-Penalvo, 2020; Tovstogan et al., 2020; Lee et al., 2022). Each task corresponds to a different type of classification problem: binary classification, multi-class classification, and multi-label classification, respectively. The difficulty level of these tasks ranges from easy to hard. To simulate the annotation process, we utilize three commonly used datasets: IMDB for sentiment annotation, WN18RR for knowledge graph completion, and CiteULike for scientific paper tagging. More details about the datasets and tasks can be found in Appendix D.
+
+Evaluation Metrics. We have selected appropriate metrics to evaluate the quality of annotations. For sentiment annotation and knowledge graph completion tasks, we use accuracy as the metric. Similarly, we use Hit Ratio  $(HR@10)$  for the multi-label tagging task, as suggested in Chen et al. (2020a). Our evaluation pays special attention to the annotation quality of model-annotated data, but also discusses the annotation quality of overall data (including expert-annotated and model-annotated data).
+
+Comparative methods. For a fair comparison, we use the same model annotator for both SANT and semi-automatic annotation, with the only difference being the addition of a data triage module. We consider both classic and advanced active learning methods, which focus on triage-to-human data, and Random strategy that allocates data randomly.
+
+- MaxEntropy uses entropy-based strategy as data triage to assign data with high uncertainty to the expert (Xiao et al., 2023; Li et al., 2023).  
+- Calibrated MaxEntropy aims to measure uncertainty more precisely. We equip MaxEntropy with the model calibration, implemented by the temperature scaling (Guo et al., 2017) with the scalar default parameter (i.e., 1.5).  
+- Ent-gn uses an advanced AL method (Wang et al., 2022) as the data triage module. It selects data that leads to a lower upper-bound of test loss, where the entropy is used to compute the gradient norm in an unsupervised way.  
+- Exp-gn uses labels from past interactions to compute an expected empirical loss (Wang et al., 2022) instead of using entropy in Ent-gn.  
+- LLM-based automatic annotation involves using ChatGPT as a strong baseline for automatic annotation. We test various prompts, including zero-shot, few-shot, zero-shot CoT, and few-shot
+
+![](images/3b03a96b0771737f3ec33ad1e20060d80af3c271ae042879a3e5d0deee02b05f.jpg)
+
+![](images/8aaf6fb7fed9aae3727415ee355697e1932f9e51d41b0ac2ee7f674584dd213e.jpg)
+
+![](images/8934d4017dc4f7f75f7eebf4e729bbc4fb46131c373e7bff91eef66a5ad6e8b1.jpg)  
+Figure 2: Annotation quality of SANT and ChatGPT-based automatic annotation. The X-axis means the proportion of annotation budgets. As the annotation tasks become increasingly difficult (from task a to task c), human experts are indispensable to achieving high-quality annotation, despite the efficiency of adopting LLMs as annotators.
+
+CoT. Refer to Appendix G for more details.
+
+To analyze SANT's advantages and uncover its characteristics, we consider the following ablation baselines used in Section 4.3.
+
+- SANT w/o AL prioritizes the triage-to-model data only. It assigns easy-to-predict data to the model and hard-to-predict data to the expert.  
+- SANT w/o EAT prioritizes the triage-to-human data only, which is essentially semi-automatic annotation with the AL enhancement. The off-the-shelf Exp-gn is used in this case.
+
+Implementation details. Considering the human-model annotation efficiency, light-weighted annotation models are always preferred (Desmond et al., 2021a; Chen et al., 2020a; Hedderich et al., 2021). Regarding the semi-automatic annotation implementation, we use customized model annotators for different annotation tasks. Specifically, we use FastText for sentiment annotation (Hedderich et al., 2021), the distributional model for the knowledge graph completion task (Roller et al., 2014; Kober et al., 2021), and the item tagging model for tagging (Chen et al., 2020a). Regarding the implementation of SANT, MLP is used as the backbone of EAT following Chen et al. (2020a). The hyperparameter  $T_{0}$  in the bi-weighting mechanism is set to be 0.2 for all experiments without tuning. To simulate human annotation behavior, we mask out the ground truth in the datasets. the ground truth is revealed when the expert is selected by the data triage module. Moreover, to simulate the limited annotation budgets setting, we tune the annotation budgets from  $10\%$  to  $90\%$  with a step size being  $10\%$ , where  $10\%$  budgets mean the human can only annotate  $10\%$  data. More implementation details are shown in Appendix E, and engineering-related suggestions are in Appendix H.
+
+# 4.2 Annotation Quality Evaluation
+
+This section presents a comparison of the annotation quality of SANT with LLM-based automatic methods and semi-automatic methods. The results are presented in Figure 2 and Table 1.
+
+Despite the efficiency of adopting LLMs as annotators, human experts are indispensable to achieving high-quality annotation. As illustrated in Figure 2, while ChatGPT's performance on simple annotation tasks (i.e., sentiment annotation) is comparable to SANT, it falls behind SANT significantly when it comes to complex annotation tasks (i.e., knowledge graph completion and tagging). In particular, ChatGPT struggles with the tagging task due to the large number of candidate tags (i.e., 500). It often generates non-existent tags. Compared to SANT, which uses a light-weighted model architecture and a certain amount of annotation budgets, we suggest that experts play an indispensable role in achieving high-quality annotations, especially for challenging annotation tasks. Therefore, significant effort should be invested in researching methods that make full use of limited budgets to achieve high-quality annotations, which is our motivation.
+
+SANT secures higher-quality annotations compared to semi-automatic annotation. Table 1 shows the annotation quality of SANT and baselines on the model-annotated data, excluding the influence of expert annotation. Compared to random-based method, almost all triage-based methods could achieve performance improvements. Such an exception motivates the necessity of clear division of duty between the human and model. This is beneficial for maximizing the annotation quality and optimizing the utilization of the limited budgets. Compared to AL-based semi-automatic annotation, SANT secures higher quality. Moreover, SANT consistently outperforms Exp-gn, which shares the same model annotator and AL-based mechanism, across different annotation budgets As the annotation budgets increase, the performance gain of
+
+<table><tr><td>Task</td><td>Triage Method</td><td>Triage Types</td><td>10%</td><td>20%</td><td>30%</td><td>40%</td><td>50%</td><td>60%</td><td>70%</td><td>80%</td><td>90%</td></tr><tr><td rowspan="6">Sentiment Annotation (binary classification)</td><td>Random</td><td>Random</td><td>82.90</td><td>82.38</td><td>82.24</td><td>82.74</td><td>82.15</td><td>82.23</td><td>82.27</td><td>82.49</td><td>82.44</td></tr><tr><td>Calibrated MaxEntropy</td><td>Classic AL</td><td>84.28</td><td>85.42</td><td>87.45</td><td>89.31</td><td>92.58</td><td>96.17</td><td>96.17</td><td>96.10</td><td>96.16</td></tr><tr><td>MaxEntropy</td><td>Classic AL</td><td>83.15</td><td>85.09</td><td>87.03</td><td>89.61</td><td>93.21</td><td>95.94</td><td>95.93</td><td>95.91</td><td>95.93</td></tr><tr><td>Ent-gn</td><td>Advanced AL</td><td>84.37</td><td>85.59</td><td>87.53</td><td>90.76</td><td>93.00</td><td>96.32</td><td>96.33</td><td>96.32</td><td>96.47</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>84.38</td><td>85.57</td><td>87.47</td><td>90.76</td><td>92.98</td><td>96.29</td><td>96.33</td><td>96.40</td><td>96.43</td></tr><tr><td>SANT (w/ Exp-gn)</td><td>Advanced AL + EAT</td><td>84.38</td><td>85.57</td><td>87.93</td><td>90.79</td><td>93.77</td><td>96.28</td><td>97.00</td><td>97.67</td><td>97.77</td></tr><tr><td rowspan="6">Knowledge Graph Completion (multi-class classification)</td><td>Random</td><td>Random</td><td>55.97</td><td>55.78</td><td>55.92</td><td>55.81</td><td>55.87</td><td>57.08</td><td>57.03</td><td>57.06</td><td>56.82</td></tr><tr><td>Calibrated MaxEntropy</td><td>Classic AL</td><td>56.66</td><td>58.91</td><td>60.75</td><td>64.17</td><td>64.48</td><td>64.72</td><td>64.59</td><td>64.59</td><td>64.29</td></tr><tr><td>MaxEntropy</td><td>Classic AL</td><td>56.88</td><td>58.90</td><td>61.27</td><td>62.95</td><td>63.08</td><td>63.92</td><td>64.13</td><td>63.18</td><td>62.89</td></tr><tr><td>Ent-gn</td><td>Advanced AL</td><td>54.48</td><td>56.08</td><td>59.34</td><td>60.15</td><td>65.19</td><td>68.71</td><td>75.49</td><td>78.09</td><td>79.41</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>55.14</td><td>56.81</td><td>60.28</td><td>61.51</td><td>65.40</td><td>68.07</td><td>76.90</td><td>81.76</td><td>81.41</td></tr><tr><td>SANT (w/ Exp-gn)</td><td>Advanced AL + EAT</td><td>55.14</td><td>56.81</td><td>62.83</td><td>66.32</td><td>71.78</td><td>78.67</td><td>85.03</td><td>87.31</td><td>87.16</td></tr><tr><td rowspan="6">Tagging (multi-label classification)</td><td>Random</td><td>Random</td><td>48.78</td><td>49.45</td><td>50.28</td><td>51.75</td><td>49.58</td><td>50.48</td><td>49.86</td><td>50.11</td><td>55.06</td></tr><tr><td>Calibrated MaxEntropy</td><td>Classic AL</td><td>48.78</td><td>48.13</td><td>48.06</td><td>48.20</td><td>48.27</td><td>48.99</td><td>49.55</td><td>49.28</td><td>49.37</td></tr><tr><td>MaxEntropy</td><td>Classic AL</td><td>48.78</td><td>48.20</td><td>48.40</td><td>48.40</td><td>48.47</td><td>48.68</td><td>49.24</td><td>49.27</td><td>49.35</td></tr><tr><td>Ent-gn</td><td>Advanced AL</td><td>48.78</td><td>49.88</td><td>51.34</td><td>52.94</td><td>49.52</td><td>52.02</td><td>54.80</td><td>56.43</td><td>58.43</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>48.78</td><td>49.90</td><td>50.94</td><td>52.89</td><td>49.50</td><td>50.24</td><td>54.48</td><td>55.47</td><td>57.84</td></tr><tr><td>SANT (w/ Exp-gn)</td><td>Advanced AL + EAT</td><td>48.78</td><td>49.90</td><td>52.24</td><td>54.35</td><td>55.25</td><td>56.60</td><td>61.13</td><td>68.24</td><td>68.53</td></tr></table>
+
+Table 1: The annotation quality on model-annotated data (\%). The best performance is marked in bold and second best is underlined. Overall, SANT enjoys higher-quality annotations. We analysis the cons of SANT on very limited data in Section 4.3. More experiments are provided in Appendix, Table 5
+
+SANT becomes more significant due to better training of the EAT and AL components. These experiments highlight that solely improving model performance through triage-to-human data (AL) may not be sufficient for optimizing annotation quality under limited budgets. Determining which data should be assigned to the model is crucial for improving annotation quality.
+
+# 4.3 Analysis and Characteristics of SANT
+
+We aim to find the key predictor for the annotation prediction of SANT: triage-to-human data or triage-to-model data. The former improves the predictive ability of the model via prioritizing informative data selected by AL, while the latter carefully selects easy-to-predict data for model annotation and prioritizes hard ones for expert annotation via EAT. We consider two ablation baselines that separate the impact of AL and EAT. The results are shown in Figure 3 and Table 2.
+
+Prioritizing triage-to-human data for active learning can enhance model predictions, but this advantage is not always the case. We first study the effect of triage-to-human data by measuring the predictive ability. However, although conducting experiments under the same annotation budgets, models are trained on different human-annotated data and tested on different unlabeled data due to different data triage preferences of AL and EAT. This hinders the measurement of predictive ability. To address this, controlled experiments are conducted to assess the predictive ability of each model on the same extra test data. The results
+
+are illustrated in Figure 3. AL performs better in predicting sentiment annotations but loses its advantages in completing knowledge graph tasks. On the other hand, EAT, which implies the anticurriculum learning (Braun et al., 2017), or hard example mining (Shrivastava et al., 2016), achieves surprisingly good results in knowledge graph completion. Although prioritizing hard instances has been shown effective in training models (Mindermann et al., 2022; Shrivastava et al., 2016), our experiment does not find statistically significant evidence of its effectiveness. Therefore, determining which instances should be triaged to humans for training a good model remains a challenge. Our results suggest that future studies should consider incorporating hard instances.
+
+Importance of triage-to-human and triage-to-model data varies at different annotation stage, which is effectively balanced by SANT. We further take both the triage-to-human and triage-to-model data into account simultaneously. Table 2 shows the annotation quality of SANT and its ablations on the model-annotated data². When working with very few budgets (10%-20%), using AL to utilize triage-to-human data is the optimal choice as SANT w/o EAT achieves higher quality on two-thirds of the datasets compared to SANT w/o AL. On the contrary, when the budgets increase, EAT, which estimates model misclassification, tends to be a more stable and accurate method for human-model collaborative annotation compared to AL.
+
+![](images/a4014ec477817f63e82e33dc833b2611fe3d836638604883b8527cb87a884d7b.jpg)  
+Figure 3: Model predictive ability evaluation on the same extra test dataset. While prioritizing triage-to-human data by AL has some advantages over triage-to-model data in promoting model predictive ability, it is not always the case (e.g., SANT w/o EAT loses its advantage in knowledge graph completion task).
+
+![](images/862153f1135dcf0695480e5c3438d7cd189d503df1f80c07286d6eeebeb30c7b.jpg)
+
+![](images/9c731eb67c2899dd0361cbc0098df33bddeff52b2c87d48916e574141261f8ee.jpg)
+
+<table><tr><td>Task</td><td>Annotation Method</td><td>Data Triage</td><td>10%</td><td>20%</td><td>30%</td><td>40%</td><td>50%</td><td>60%</td><td>70%</td><td>80%</td><td>90%</td></tr><tr><td rowspan="3">Sentiment Annotation</td><td>SANT w/o EAT</td><td>Advanced AL</td><td>84.38</td><td>85.57</td><td>87.47</td><td>90.76</td><td>92.98</td><td>96.29</td><td>96.33</td><td>96.40</td><td>96.43</td></tr><tr><td>SANT w/o AL</td><td>EAT</td><td>83.90</td><td>85.16</td><td>87.89</td><td>90.78</td><td>93.40</td><td>96.13</td><td>96.96</td><td>97.47</td><td>97.71</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>84.38</td><td>85.57</td><td>87.93</td><td>90.79</td><td>93.77</td><td>96.28</td><td>97.00</td><td>97.67</td><td>97.77</td></tr><tr><td rowspan="3">Knowledge Graph Completion</td><td>SANT w/o EAT</td><td>Advanced AL</td><td>55.14</td><td>56.81</td><td>60.28</td><td>61.51</td><td>65.40</td><td>68.07</td><td>76.90</td><td>81.76</td><td>81.41</td></tr><tr><td>SANT w/o AL</td><td>EAT</td><td>57.06</td><td>59.75</td><td>62.75</td><td>66.08</td><td>71.61</td><td>78.57</td><td>84.99</td><td>87.28</td><td>87.12</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>55.14</td><td>56.81</td><td>62.83</td><td>66.32</td><td>71.78</td><td>78.67</td><td>85.03</td><td>87.31</td><td>87.16</td></tr><tr><td rowspan="3">Tagging</td><td>SANT w/o EAT</td><td>Advanced AL</td><td>48.78</td><td>49.90</td><td>50.94</td><td>52.89</td><td>49.50</td><td>50.24</td><td>54.48</td><td>55.47</td><td>57.84</td></tr><tr><td>SANT w/o AL</td><td>EAT</td><td>48.78</td><td>49.81</td><td>51.14</td><td>52.97</td><td>55.23</td><td>56.53</td><td>61.08</td><td>68.20</td><td>68.50</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>48.78</td><td>49.90</td><td>52.24</td><td>54.35</td><td>55.25</td><td>56.60</td><td>61.13</td><td>68.24</td><td>68.53</td></tr></table>
+
+Table 2: The annotation quality of SANT and its ablation counterparts on the model-annotated data (\%). Importance of triage-to-human (i.e., AL) and triage-to-model data (i.e., EAT) varies at different annotation stage. SANT effectively balances the advantages of the triage-to-human and triage-to-model data by integrating EAT and AL.
+
+This shifts the focus to triage-to-model data. In conjunction with EAT, SANT  $w/o AL$  achieves the better results compared to SANT  $w/o EAT$ , particularly in the middle range of annotation budgets (40%-70%). Therefore, the importance of triage-to-human and triage-to-model data varies at different annotation stage. To bring together two types of data, SANT utilizes the bi-weighting mechanism. We found that the annotation quality of SANT is optimal in most cases, indicating that SANT effectively allocates triage-to-human data to train the model and enhance its predictive capability, while also allocating triage-to-model data to improve the model annotation quality by providing easy-to-predict data for the model.
+
+The key factor influencing the quality of annotations in SANT lies in the careful selection of data to be assigned for model annotation. According to Figure 3, prioritizing triage-to-human data using AL or EAT does not lead to a significant improvement in the model's predictive ability. While determining which instances should be triaged to human is crucial for model training, our experiment does not find statistically significant evidence of its advantages. It is possible that the potential values of data have not been fully utilized (Tang and Huang, 2019; Mindermann et al., 2022). Therefore, triage-to-human data is expected
+
+to have a limited contribution to improving annotation quality in SANT. Additionally, our analysis of Table 2 shows that SANT w/o AL has a more significant advantage in model annotation data quality compared to SANT w/o EAT. Therefore, we argue that carefully selecting easy-to-predict data for model prediction is crucial in significantly improving overall annotation accuracy, even though training the model on manual labeled hard-to-predict data may not result in a stronger predictive ability. This finding highlights that deciding which data to use for model prediction is more important than training itself, particularly as budgets increase. In some cases, the improvement in annotation quality can compensate for the disadvantages of an under-trained model with lower predictive ability. To sum up, prioritizing triage-to-model data is a pivotal factor in determining the success of SANT.
+
+# 5 Conclusion
+
+We present the first work to study the roles of both triage-to-human and triage-to-model data. We formulate semi-automatic data annotation as a data triage problem via SANT, achieved by optimizing two novel mechanisms (i.e., EAT and bi-weighting mechanisms) in utilizing limited annotation budgets. Empirical experiments show that SANT sig-
+
+nificantly outperforms the existing semi-automatic annotation methods. Although the effectiveness of LLM-based annotator, we find that experts are still indispensable to achieving high-quality annotations, especially on challenging data annotation tasks. Coordinately delegating the human and the model as a team can be a feasible solution. We also observe that relying on triage-to-human data to create a better model may not be the most effective approach for high-quality annotations. It may be more beneficial to identify triage-to-model data.
+
+# Limitations
+
+Scale of Model Annotator. One limitation of our work is that we use lightweight models for annotation due to limited computational resources. While lightweight models are commonly employed in data annotation (Chen et al., 2020a; Desmond et al., 2021a), we do not investigate the effectiveness of larger language models as the model annotator in this study. However, SANT, based on lightweight models, still outperforms the LLM-based automatic annotator.
+
+Model annotation cost. While the use of models for data annotation may incur computational costs, our current research focuses solely on human annotation costs, neglecting model annotation costs. This simplification, while not entirely reflective of real-world scenarios, is justified by the significantly higher cost of human annotation compared to model annotation (Li et al., 2023; Gilardi et al., 2023). Additionally, the light-weighted models employed in this study further minimize the model annotation cost, rendering it negligible. To enhance the practicality of our work, future research will incorporate the cost of model annotation.
+
+Human annotation cost simulation. Human annotation cost on each piece of data may not necessarily be constant. The cognitive abilities of different humans and the difficulty level of different data can both impact the human annotation cost. Our paper, like the previous research, assumes that the human annotation cost on each data is constant. In future research, constructing a user simulator to simulate human cognitive abilities and evaluating the difficulty of the data could be the next research direction to improve our work.
+
+Performance under extremely limited budgets. Our method may not exhibit a substantial advantage over the best baseline under extremely limited budgets. However, it consistently performs
+
+at least on par, and often slightly better in this scenario. This, however, does not diminish the substantial effectiveness advantage demonstrated by our method in other scenarios. We have conducted a thorough analysis of the reasons behind our method's limited performance improvement under extremely limited budgets. This analysis will guide our future work in developing targeted improvements to further enhance the effectiveness of our method.
+
+Data annotation can be an intricate and complex engineering problem. While our work focuses on the data allocation problem and introduces the SANT framework for human-machine cooperative annotation, real-world deployment of SANT necessitates further systematic engineering considerations. These aspects, often overlooked in both our paper and other research on data allocation or data annotation, are addressed in Appendix H, where deployment suggestions are provided.
+
+# Acknowledgements
+
+This work was supported in part by the National Natural Science Foundation of China (No. 62272330); in part by the Fundamental Research Funds for the Central Universities (No. YJ202219); in part by the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant (No. MSS24C004).
+
+# References
+
+Jonathan F Bard. 2013. Practical bilevel optimization: algorithms and applications, volume 30. Springer Science & Business Media.  
+Stefan Braun, Daniel Neil, and Shih-Chii Liu. 2017. A curriculum learning method for improved noise robustness in automatic speech recognition. In 2017 25th European Signal Processing Conference (EU-SIPCO), pages 548-552. IEEE.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Christopher Bryant, Mariano Felice, and Ted Briscoe. 2017. Automatic annotation and evaluation of error types for grammatical error correction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 793-805, Vancouver, Canada. Association for Computational Linguistics.
+
+Jhonny Cerezo, Felipe Bravo-Marquez, and Alexandre Henri Bergel. 2021. Tools impact on the quality of annotations for chat untangling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: Student Research Workshop, pages 215-220, Online. Association for Computational Linguistics.  
+Nicolo Cesa-Bianchi, Claudio Gentile, and Francesco Orabona. 2009. Robust bounds for classification via selective sampling. In Proceedings of the 26th annual international conference on machine learning, pages 121-128.  
+Nicolo Cesa-Bianchi, Claudio Gentile, Luca Zaniboni, and Manfred Warmuth. 2006. Worst-case analysis of selective sampling for linear classification. Journal of Machine Learning Research, 7(7).  
+Jia Chen, Yike Wu, Shiwan Zhao, and Qin Jin. 2021. Language resource efficient learning for captioning. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 1887-1895, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Xu Chen, Changying Du, Xiuqiang He, and Jun Wang. 2020a. JIT2r: A joint framework for item tagging and tag-based recommendation. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, pages 1681-1684.  
+Yilu Chen, Xiaojun Zeng, Xing Chen, and Wenzhong Guo. 2020b. A survey on automatic image annotation. Applied Intelligence, 50(10):3412-3428.  
+Bhavana Dalvi, Sumithra Bhakthavatsalam, Chris Clark, Peter Clark, Oren Etzioni, Anthony Fader, and Dirk Groeneveld. 2016. Ike-an interactive tool for knowledge extraction. In Proceedings of the 5th Workshop on Automated Knowledge Base Construction, pages 12-17.  
+Stefan Depeweg, Jose-Miguel Hernandez-Lobato, Finale Doshi-Velez, and Steffen Udluft. 2018. Decomposition of uncertainty in bayesian deep learning for efficient and risk-sensitive learning. In International Conference on Machine Learning, pages 1184-1193. PMLR.  
+Michael Desmond, Michelle Brachman, Evelyn Duesterwald, Casey Dugan, Narendra Nath Joshi, Qian Pan, and Carolina Spina. 2022. Ai assisted data labeling with interactive auto label. In AAAI Conference on Artificial Intelligence.  
+Michael Desmond, Evelyn Duesterwald, Kristina Brimijoin, Michelle Brachman, and Qian Pan. 2021a. Semi-automated data labeling. In Proceedings of the NeurIPS 2020 Competition and Demonstration Track, volume 133 of Proceedings of Machine Learning Research, pages 156-169. PMLR.
+
+Michael Desmond, Evelyn Duesterwald, Kristina Brimijoin, Michelle Brachman, and Qian Pan. 2021b. Semi-automated data labeling. In NeurIPS 2020 Competition and Demonstration Track, pages 156-169. PMLR.  
+Bosheng Ding, Chengwei Qin, Linlin Liu, Yew Ken Chia, Boyang Li, Shafiq Joty, and Lidong Bing. 2023. Is GPT-3 a good data annotator? In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11173–11195, Toronto, Canada. Association for Computational Linguistics.  
+Bosheng Ding, Chengwei Qin, Ruochen Zhao, Tianze Luo, Xinze Li, Guizhen Chen, Wenhan Xia, Junjie Hu, Anh Tuan Luu, and Shafiq Joty. 2024. Data augmentation using llms: Data perspectives, learning paradigms and challenges. arXiv preprint arXiv:2403.02990.  
+Liat Ein-Dor, Alon Halfon, Ariel Gera, Eyal Shnarch, Lena Dankin, Leshem Choshen, Marina Danilevsky, Ranit Aharonov, Yoav Katz, and Noam Slonim. 2020. Active Learning for BERT: An Empirical Study. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7949-7962, Online. Association for Computational Linguistics.  
+Yansong Feng and Mirella Lapata. 2008. Automatic image annotation using auxiliary text information. In Proceedings of ACL-08: HLT, pages 272-280, Columbus, Ohio. Association for Computational Linguistics.  
+Fabrizio Gilardi, Meysam Alizadeh, and Mael Kubli. 2023. Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30):e2305016120.  
+Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR.  
+Xingwei He, Zhenghao Lin, Yeyun Gong, A-Long Jin, Hang Zhang, Chen Lin, Jian Jiao, Siu Ming Yiu, Nan Duan, and Weizhu Chen. 2023. Annollm: Making large language models to be better crowdsourced annotators.  
+Michael A Hedderich, Lukas Lange, and Dietrich Klakow. 2021. Area: distant supervision for low-resource named entity recognition. the Practical Machine Learning For Developing Countries Workshop at ICLR.  
+Yan Hu, Iqra Ameer, Xu Zuo, Xueqing Peng, Yujia Zhou, Zehan Li, Yiming Li, Jianfu Li, Xiaogian Jiang, and Hua Xu. 2023. Zero-shot clinical entity recognition using chatgpt. arXiv preprint arXiv:2303.16416.  
+Chen Huang, Yiping Jin, Ilija Ilievski, Wenqiang Lei, and Jiancheng Lv. 2024a. ARAIDA: Analogical reasoning-augmented interactive data annotation. In
+
+Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10660-10675, Bangkok, Thailand. Association for Computational Linguistics.  
+Chen Huang, Peixin Qin, Wenqiang Lei, and Jiancheng Lv. 2023. Reduce human labor on evaluating conversational information retrieval system: A human-machine collaboration approach. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10876-10891, Singapore. Association for Computational Linguistics.  
+Chen Huang, Xinwei Yang, Yang Deng, Wenqiang Lei, JianCheng Lv, and Tat-Seng Chua. 2024b. Co-matching: Towards human-machine collaborative legal case matching.  
+Rebecca Hwa. 2000. Sample selection for statistical grammar induction. In 2000 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora, pages 45-52.  
+Eric Jang, Shixiang Gu, and Ben Poole. 2016. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144.  
+Heinrich Jiang, Been Kim, Melody Guan, and Maya Gupta. 2018. To trust or not to trust a classifier. NeurIPS, 31.  
+Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431. Association for Computational Linguistics.  
+Jan-Christoph Klie, Michael Bugert, Beto Boullosa, Richard Eckart de Castilho, and Iryna Gurevych. 2018. The INCEpTION platform: Machine-assisted and knowledge-oriented interactive annotation. In COLING: System Demonstrations, pages 5-9, Santa Fe, New Mexico. ACL.  
+Thomas Kober, Julie Weeds, Lorenzo Bertolini, and David J. Weir. 2021. Data augmentation for hypernymy detection. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1034-1048.  
+Oyku Deniz Kose and Yanning Shen. 2022. Fairness-aware selective sampling on attributed graphs. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5682-5686.  
+Trausti Kristjansson, Aron Culotta, Paul Viola, and Andrew McCallum. 2004. Interactive information extraction with constrained conditional random fields. In AAAI, volume 4, pages 412-418.
+
+Stefan Larson, Adrian Cheung, Anish Mahendran, Kevin Leach, and Jonathan K Kummerfeld. 2020. Inconsistencies in crowdsourced slot-filling annotations: A typology and identification methods. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5035-5046.  
+Yoonjoo Lee, John Joon Young Chung, Tae Soo Kim, Jean Y Song, and Juho Kim. 2022. Promptiverse: Scalable generation of scaffolding prompts through human-ai hybrid knowledge graph annotation. In CHI Conference on Human Factors in Computing Systems, pages 1-18.  
+Aoxue Li, Weiran Huang, Xu Lan, Jiashi Feng, Zhenguo Li, and Liwei Wang. 2020. Boosting few-shot learning with adaptive margin loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12576-12584.  
+Minzhi Li, Taiwei Shi, Caleb Ziems, Min-Yen Kan, Nancy F. Chen, Zhengyuan Liu, and Diyi Yang. 2023. Coannotating: Uncertainty-guided work allocation between human and large language models for data annotation.  
+Yunyao Li, Elmer Kim, Marc A Touchette, Ramiya Venkatachalam, and Hao Wang. 2015. Vinery: A visual ide for information extraction. Proceedings of the VLDB Endowment, 8(12):1948-1951.  
+Aiwei Liu, Xuming Hu, Lijie Wen, and Philip S Yu. 2023. A comprehensive evaluation of chatgpt's zero-shot text-to-sql capability. arXiv preprint arXiv:2303.13547.  
+Christina Lohr, Johannes Kiesel, Stephanie Luther, Johannes Hellrich, Tobias Kolditz, Benno Stein, and Udo Hahn. 2019. Continuous quality control and advanced text segment annotation with WAT-SL 2.0. In Proceedings of the 13th Linguistic Annotation Workshop, pages 215-219, Florence, Italy. ACL.  
+Antonio Loquercio, Mattia Segu, and Davide Scaramuzza. 2020. A general framework for uncertainty estimation in deep learning. IEEE Robotics and Automation Letters, 5(2):3153-3160.  
+Bo-Ru Lu, Nikita Hahuong, Chia-Hsuan Lee, Zeqiu Wu, Hao Cheng, Paul Koester, Jean Utke, Tao Yu, Noah A. Smith, and Mari Ostendorf. 2024. Does collaborative human-lm dialogue generation help information extraction from human dialogues?  
+David Madras, Toni Pitassi, and Richard Zemel. 2018. Predict responsibly: improving fairness and accuracy by learning to defer. NeurIPS, 31.  
+Marian Marchal, Merel Scholman, and Vera Demberg. 2021. Semi-automatic discourse annotation in a low-resource language: Developing a connective lexicon for nigerian pidgin. In Proceedings of the 2nd Workshop on Computational Approaches to Discourse, pages 84-94.
+
+Samuel Marcos-Pablos and Francisco J Garcia-Penalvo. 2020. Information retrieval methodology for aiding scientific database search. Soft Computing, 24(8):5551-5560.  
+Sören Mindermann, Jan M Brauner, Muhammed T Razzak, Mrinank Sharma, Andreas Kirsch, Winnie Xu, Benedikt Höltgen, Aidan N Gomez, Adrien Morisot, Sebastian Farquhar, et al. 2022. Prioritized training on points that are learnable, worth learning, and not yet learnt. In International Conference on Machine Learning, pages 15630-15649. PMLR.  
+Hussein Mozannar and David Sontag. 2020. Consistent estimators for learning to defer to an expert. In International Conference on Machine Learning, pages 7076-7087. PMLR.  
+Chenri Ni, Nontawat Charoenphakdee, Junya Honda, and Masashi Sugiyama. 2019. On the calibration of multiclass classification with rejection. NeurIPS, 32.  
+Nastaran Okati, Abir De, and Manuel Rodriguez. 2021. Differentiable learning under triage. NeurIPS, 34.  
+Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for word representation. In EMNLP, pages 1532-1543.  
+Maithra Raghu, Katy Blumer, Greg Corrado, Jon Kleinberg, Ziad Obermeyer, and Sendhil Mullainathan. 2019. The algorithmic automation problem: Prediction, triage, and human effort. arXiv preprint arXiv:1903.12220.  
+Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Xiaojiang Chen, and X. Wang. 2020. A survey of deep active learning. ArXiv, abs/2009.00236.  
+Stephen Roller, Katrin Erk, and Gemma Boleda. 2014. Inclusive yet selective: Supervised distributional hypernymy detection. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1025-1036.  
+Claudia Schulz, Christian M Meyer, Jan Kiesewetter, Michael Sailer, Elisabeth Bauer, Martin R Fischer, Frank Fischer, and Iryna Gurevych. 2019. Analysis of automatic annotation suggestions for hard discourse-level tasks in expert domains. In 57th Annual Meeting of the Association for Computational Linguistics (Acl 2019), pages 2761-2772. Association for Computational Linguistics.  
+Agam Shah and Sudheer Chava. 2023. Zero is not hero yet: Benchmarking zero-shot performance of llms for financial tasks. arXiv preprint arXiv:2305.16633.  
+Amirreza Shirani, Franck Dernoncourt, Paul Asente, Nedim Lipka, Seokhwan Kim, Jose Echevarria, and Thamar Solorio. 2019. Learning emphasis selection for written text in visual media from crowd-sourced label distributions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1167-1172, Florence, Italy. Association for Computational Linguistics.
+
+Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. 2016. Training region-based object detectors with online hard example mining. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 761-769.  
+Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. 2022. Selective annotation makes language models better few-shot learners.  
+Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023. Evaluation of chatgpt as a question answering system for answering complex questions. arXiv preprint arXiv:2303.07992.  
+Ying-Peng Tang and Sheng-Jun Huang. 2019. Self-paced active learning: Query the right thing at the right time. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 5117-5124.  
+Philip Tovstogan, Xavier Serra, and Dmitry Bogdanov. 2020. Web interface for exploration of latent and tag spaces in music auto-tagging. In Proceedings of the 37th International Conference on Machine Learning. Vienna, Austria. ICML.  
+Hao Wang, Binyi Chen, and Wu-Jun Li. 2013. Collaborative topic regression with social regularization for tag recommendation. In Twenty-Third International Joint Conference on Artificial Intelligence.  
+Jiaan Wang, Yunlong Liang, Fandong Meng, Zengkui Sun, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, and Jie Zhou. 2023. Is chatgpt a good nlg evaluator? a preliminary study. arXiv preprint arXiv:2303.04048.  
+Tianyang Wang, Xingjian Li, Pengkun Yang, Guosheng Hu, Xiangrui Zeng, Siyu Huang, Cheng-Zhong Xu, and Min Xu. 2022. Boosting active learning via improving test performance. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8566-8574.  
+Xin Wang and Siu Ming Yiu. 2021. Classification with rejection: scaling generative classifiers with supervised deep infomax. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 2980-2986.  
+Yong Wang, Xiao-Ming Wu, Qimai Li, Jiatao Gu, Wangmeng Xiang, Lei Zhang, and Victor OK Li. 2018. Large margin few-shot learning. arXiv preprint arXiv:1807.02872.  
+Zhuang Wang, Koby Crammer, and Slobodan Vucetic. 2012. Breaking the curse of kernelization: Budgeted stochastic gradient descent for large-scalesvm training. The Journal of Machine Learning Research, 13(1):3103-3131.
+
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837.
+
+Bryan Wilder, Eric Horvitz, and Ece Kamar. 2021. Learning to complement humans. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 1526-1533.
+
+Ruixuan Xiao, Yiwen Dong, Junbo Zhao, Runze Wu, Minmin Lin, Gang Chen, and Haobo Wang. 2023. FreeAL: Towards human-free active learning in the era of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14520-14535, Singapore. Association for Computational Linguistics.
+
+Bingsheng Yao, Ishan Jindal, Lucian Popa, Yannis Katsis, Sayan Ghosh, Lihong He, Yuxuan Lu, Shashank Srivastava, James Hendler, and Dakuo Wang. 2023. Beyond labels: Empowering human with natural language explanations through a novel active-learning architecture.
+
+Kaichun Yao, Chuan Qin, Hengshu Zhu, Chao Ma, Jingshuai Zhang, Yi Du, and Hui Xiong. 2021. An interactive neural network approach to keyphrase extraction in talent recruitment. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 2383-2393.
+
+Yangjun Zhang, Pengjie Ren, and Maarten de Rijke. 2021. A human-machine collaborative framework for evaluating malevolence in dialogues. 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 5612-5623.
+
+Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Schölkopf. 2003. Learning with local and global consistency. In Advances in Neural Information Processing Systems, volume 16. MIT Press.
+
+# A More Related Work
+
+Our research is also tied to data annotation and active learning. Furthermore, we address the data triage problem by utilizing the idea of learning to defer. To elaborate on these approaches, we provide a literature review and highlight the differences from existing methods.
+
+Data Annotation. Computer-assisted annotation (or interactive annotation) becomes an alternative option. Such methods enable a model to facilitate human annotation by providing annotation suggestions on the fly (Klie et al., 2018; Lohr et al., 2019; Marchal et al., 2021). However, these
+
+methods still require manual validation on each prediction, which is inefficient and time-consuming. Thus, it is not feasible for the setting with limited annotation budgets.
+
+Active Learning. It targets to select informative or representative data that can maximize the performance of a model (Ren et al., 2020). One way to achieve this is through the uncertainty-based method (Desmond et al., 2021b), which selects data with high prediction uncertainty for human annotation. Density-based methods are also commonly used, but recent studies show that these two methods are highly correlated (Loquercio et al., 2020). Additionally, research on selective sampling tries to combine active learning and online learning (Cesa-Bianchi et al., 2006, 2009), where the model is allowed to adaptively query the label of an observed data on the fly, according to its potential to improve the model's performance. Basically, selective sampling has shown its effectiveness in online large-scale kernel (Wang et al., 2012) and graph-based online learning (Kose and Shen, 2022). Semi-automatic annotation utilizes the online active learning (or selective sampling) methods to triage data for expert annotations and efficiently train a good model. However, our SANT goes beyond active learning by introducing the EAT mechanism, which identifies data that the model struggles to predict. SANT then employs the bi-weighting mechanism to combine EAT and active learning, enabling it to leverage both triage-to-human data (active learning) and triage-to-model data (EAT). This integration allows SANT to take full advantage of both methods, resulting in more efficient and accurate annotations.
+
+Learning to Defer. The collaboration between humans and machines can be described as "learning to defer", where tasks are jointly completed by either the machine or deferred to humans for decision-making (Raghu et al., 2019; Okati et al., 2021; Madras et al., 2018). To achieve the overall goal, a triage strategy is employed to dynamically assign tasks to different collaborators, optimizing global utilization. Depending on the customized triage strategies, existing approaches assign data with high prediction uncertainty to humans (Ni et al., 2019), while out-of-distribution and noisy data to humans for decision-making (Wang and Yiu, 2021). Additionally, Jiang et al. (2018) assign data with unreliable model predictions to humans to mitigate prediction bias. Existing semi-automatic data annotation methods can also be incorporated
+
+into the learning to defer framework, where the informative data are deferred to humans by AL methods. Recent work tries to assign data based on the uncertainty of LLM predictions (Li et al., 2023), where the parameters of LLM are frozen during the whole annotation process, as is the uncertainty measurement method. However, this approach hinders the model's ability to improve its annotation results by learning from the data, relying heavily on LLM's data annotation capability for general tasks. As highlighted by our work and other researchers (Xiao et al., 2023), the annotation outcomes of LLM in general tasks are not satisfying. To update the annotation model on the fly, we notice that applying "learning to defer" to data annotation tasks is particularly challenging. This is because existing works on learning to defer primarily focus on fully supervised settings where labels are available, whereas in the annotation task, labels are only accessible when a human is selected for annotation. This motivates us to develop a model that can learn and generalize from limited labeled data on the fly (cf. Section 3.2.3). Furthermore, we propose two tailored mechanisms, i.e., EAT and biweighting mechanisms, and experimentally reveal some insights for the data annotation task.
+
+# B More Evaluation Results of SANT on Overall Dataset
+
+Tables 3 and 4 demonstrate the annotation quality of each method on the overall data. These findings align with those in Tables 1 and 2, indicating that proper allocation of data to both human and model workers results in higher-quality annotation. Moreover, SANT effectively balances the advantages of triage-to-human data and triage-to-model data by allocating triage-to-model data to improve model annotation quality through easy-to-predict instances.
+
+Moreover, we report the more results of model annotation quality on two datasets, including SST-5 and FreeBase in Table 5. Due to financial limits, we compare our method with the advanced AL method (Exp-gn) and the random-based method. Based on the table results, it is clear that our method continues to be effective. Moreover, when compared to LLM-based annotators, our method can generate higher-quality annotations with some human assistance. This highlights the importance of human experts in achieving high annotation quality, even though LLMs may be efficient annotators. In
+
+the revision, we will conduct more experiments to provide stronger validation.
+
+# C Ablation study on the proposed max-margin loss  $L_{m}$
+
+As shown in Section 3.2.2, EAT uses the max-margin loss  $L_{m}$  to take advantage of the partially labeled data. The loss is parameterized by a margin hyper-parameter  $\tau$  to force the average loss on the assign-to-human data to be much higher than the assign-to-model data. In this section, we offer a detailed discussion of the loss itself and the sensitivity analysis. Specifically, the ablation study on  $L_{m}$  and the hyper-parameter sensitivity study on  $\tau$  are given in Figure 4. Here, we take the sentiment annotation task as an example, and require SANT w/o AL to reject  $50\%$  of the data to the model in each batch, making sure our ablation studies are performed on the training data of the same size. We focus on the model performances on model-annotated data, marked as  $MA_{Accept}$ . Here,  $\tau$  is tuned in  $\{0,0.1,0.3,0.5,0.7,1.0\}$ . According to the results, we believe that the max-margin loss  $L_{m}$  helps train a better EAT. Moreover, the  $L_{m}$  is relatively less sensitive to its hyper-parameter  $\tau$ . Thus, we simply choose  $\tau = 0.3$ .
+
+![](images/e16e612f763484fb2410cc0870a48f0740bf291e29bef0557dc39019a410e21c.jpg)  
+Figure 4: Ablation study on  $L_{m}$  and hyper-parameter tuning analysis on  $\tau$ . We simply choose  $\tau = 0.3$  in our experiments.
+
+# D Details on Datasets
+
+To comprehensively evaluate SANT, we use IMDB $^3$ , WN18RR $^4$ , and CiteULike $^5$  to simulate sentiment annotation, knowledge graph completion, and tagging tasks, respectively. Those
+
+<table><tr><td>Task</td><td>Annotation Method</td><td>Data Triage</td><td>10%</td><td>20%</td><td>30%</td><td>40%</td><td>50%</td><td>60%</td><td>70%</td><td>80%</td><td>90%</td></tr><tr><td rowspan="6">Sentiment Annotation (binary classification)</td><td>Random</td><td>Random</td><td>84.61</td><td>85.90</td><td>87.57</td><td>89.64</td><td>91.08</td><td>92.89</td><td>94.68</td><td>96.50</td><td>98.24</td></tr><tr><td>Calibrated MaxEntropy</td><td>Classic AL</td><td>85.85</td><td>88.34</td><td>91.22</td><td>93.59</td><td>96.29</td><td>98.47</td><td>98.85</td><td>99.22</td><td>99.62</td></tr><tr><td>MaxEntropy</td><td>Classic AL</td><td>84.84</td><td>88.07</td><td>90.92</td><td>93.77</td><td>96.61</td><td>98.38</td><td>98.78</td><td>99.18</td><td>99.59</td></tr><tr><td>Ent-gn</td><td>Advanced AL</td><td>85.93</td><td>88.47</td><td>91.27</td><td>94.46</td><td>96.50</td><td>98.53</td><td>98.90</td><td>99.26</td><td>99.65</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>85.94</td><td>88.46</td><td>91.23</td><td>94.46</td><td>96.49</td><td>98.52</td><td>98.90</td><td>99.28</td><td>99.64</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>85.94</td><td>88.46</td><td>91.55</td><td>94.47</td><td>96.89</td><td>98.51</td><td>99.10</td><td>99.53</td><td>99.78</td></tr><tr><td rowspan="6">Knowledge Graph Completion (multi-class classification)</td><td>Random</td><td>Random</td><td>60.37</td><td>64.62</td><td>69.14</td><td>73.49</td><td>77.94</td><td>82.83</td><td>87.11</td><td>91.41</td><td>95.68</td></tr><tr><td>Calibrated MaxEntropy</td><td>Classic AL</td><td>60.99</td><td>67.13</td><td>72.53</td><td>78.50</td><td>82.24</td><td>85.89</td><td>89.38</td><td>92.92</td><td>96.43</td></tr><tr><td>MaxEntropy</td><td>Classic AL</td><td>61.19</td><td>67.12</td><td>72.89</td><td>77.77</td><td>81.54</td><td>85.57</td><td>89.24</td><td>92.64</td><td>96.29</td></tr><tr><td>Ent-gn</td><td>Advanced AL</td><td>59.03</td><td>64.86</td><td>71.54</td><td>76.09</td><td>82.60</td><td>87.48</td><td>92.65</td><td>95.62</td><td>97.94</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>59.63</td><td>65.45</td><td>72.20</td><td>76.91</td><td>82.70</td><td>87.23</td><td>93.07</td><td>96.35</td><td>98.14</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>59.63</td><td>65.45</td><td>73.98</td><td>79.79</td><td>85.89</td><td>91.47</td><td>95.51</td><td>97.46</td><td>98.72</td></tr><tr><td rowspan="6">Tagging (multi-label classification)</td><td>Random</td><td>Random</td><td>53.90</td><td>59.56</td><td>65.20</td><td>71.05</td><td>74.79</td><td>80.19</td><td>84.96</td><td>90.02</td><td>95.51</td></tr><tr><td>Calibrated MaxEntropy</td><td>Classic AL</td><td>53.90</td><td>58.50</td><td>63.64</td><td>68.92</td><td>74.14</td><td>79.60</td><td>84.87</td><td>89.86</td><td>94.94</td></tr><tr><td>MaxEntropy</td><td>Classic AL</td><td>53.90</td><td>58.56</td><td>63.88</td><td>69.04</td><td>74.24</td><td>79.47</td><td>84.77</td><td>89.85</td><td>94.94</td></tr><tr><td>Ent-gn</td><td>Advanced AL</td><td>53.90</td><td>59.90</td><td>65.94</td><td>71.76</td><td>74.76</td><td>80.81</td><td>86.44</td><td>91.29</td><td>95.84</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>53.90</td><td>59.92</td><td>65.66</td><td>71.73</td><td>74.75</td><td>80.10</td><td>86.34</td><td>91.09</td><td>95.78</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>53.90</td><td>59.92</td><td>66.57</td><td>72.61</td><td>77.63</td><td>82.64</td><td>88.34</td><td>93.65</td><td>96.85</td></tr></table>
+
+Table 3: The annotation quality of SANT and baselines on overall data  $(\%)$ . We consider three annotation tasks with different difficulties. The best performance is marked in bold and second best is underlined.  
+
+<table><tr><td>Task</td><td>Annotation Method</td><td>Data Triage</td><td>10%</td><td>20%</td><td>30%</td><td>40%</td><td>50%</td><td>60%</td><td>70%</td><td>80%</td><td>90%</td></tr><tr><td rowspan="3">Sentiment Annotation</td><td>SANT w/o EAT</td><td>Advanced AL</td><td>85.94</td><td>88.46</td><td>91.23</td><td>94.46</td><td>96.49</td><td>98.52</td><td>98.90</td><td>99.28</td><td>99.64</td></tr><tr><td>SANT w/o AL</td><td>EAT</td><td>85.51</td><td>88.13</td><td>91.52</td><td>94.47</td><td>96.70</td><td>98.45</td><td>99.09</td><td>99.49</td><td>99.77</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>85.94</td><td>88.46</td><td>91.55</td><td>94.47</td><td>96.89</td><td>98.51</td><td>99.10</td><td>99.53</td><td>99.78</td></tr><tr><td rowspan="3">Knowledge Graph Completion</td><td>SANT w/o EAT</td><td>Advanced AL</td><td>59.63</td><td>65.45</td><td>72.20</td><td>76.91</td><td>82.70</td><td>87.23</td><td>93.07</td><td>96.35</td><td>98.14</td></tr><tr><td>SANT w/o AL</td><td>EAT</td><td>61.35</td><td>67.80</td><td>73.93</td><td>79.65</td><td>85.81</td><td>91.43</td><td>95.50</td><td>97.46</td><td>98.71</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>59.63</td><td>65.45</td><td>73.98</td><td>79.79</td><td>85.89</td><td>91.47</td><td>95.51</td><td>97.46</td><td>98.72</td></tr><tr><td rowspan="3">Tagging</td><td>SANT w/o EAT</td><td>Advanced AL</td><td>53.90</td><td>59.92</td><td>65.66</td><td>71.73</td><td>74.75</td><td>80.10</td><td>86.34</td><td>91.09</td><td>95.78</td></tr><tr><td>SANT w/o AL</td><td>EAT</td><td>53.90</td><td>59.85</td><td>65.80</td><td>71.78</td><td>77.62</td><td>82.61</td><td>88.32</td><td>93.64</td><td>96.85</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>53.90</td><td>59.92</td><td>66.57</td><td>72.61</td><td>77.63</td><td>82.64</td><td>88.34</td><td>93.65</td><td>96.85</td></tr></table>
+
+Table 4: The annotation quality of SANT and its ablation counterparts on the overall data  $(\%)$
+
+three datasets are all English datasets and they formulate three different classification problems, including binary-class classification, multi-class classification, and multi-label classification problems. For the sentiment annotation, the annotator is required to label the binary sentiment based on the input movie review. As for the knowledge graph completion, the annotator is required to label the semantic relation of an input word pair. As for the tagging task, the annotator is required to label the tags of a scientific paper.
+
+Specifically, IMDB is a common-used data with 50K movie reviews and binary sentiment classes. WN18RR we use is a subset of WN18 with no inverse relations. WN18RR contains 93003 triples, 40943 entities, and 11 relation types. Lastly, CiteULike (Chen et al., 2020a; Wang et al., 2013) is a commonly used tagging dataset, collected from CiteULike and Google Scholar. It includes 7288 authors' 160272 citations on 8212 papers, together with 46391 different tags. Each paper in this dataset has its abstract, title, and multiple tags. The con
+
+catenation of them are used as the input in our experiments. Following (Chen et al., 2020a), we also use 500 most frequent tags in our tagging experiment.
+
+Considering the human-model interaction efficiency, light-weighted annotation models or systems are always preferred (Desmond et al., 2021a; Chen et al., 2020a). In our experiments, we use the pre-trained GloVe $^6$  (Pennington et al., 2014) for all tasks to encode the word representation without fine-tuning. For all the dataset, we first filter out words that are not contained in the pre-trained Glove embedding. For word pairs in WN18RR, we only retain those pairs that both words have Glove embeddings. Following previous works (Roller et al., 2014; Kober et al., 2021), the word pair embedding is simply obtained by concatenating the embedding of two words. As for the sentence embedding used in CiteULike and IMDB, we follow
+
+<table><tr><td>Task</td><td>Annotation Method</td><td>Data Triage</td><td>10%</td><td>20%</td><td>30%</td><td>40%</td><td>50%</td><td>60%</td><td>70%</td><td>80%</td><td>90%</td></tr><tr><td rowspan="5">Sentiment Annotation (multi-class annotation) SST-5</td><td>Zero-shot ChatGPT</td><td>None</td><td colspan="9">41.30</td></tr><tr><td>Few-shot ChatGPT</td><td>None</td><td colspan="9">53.85</td></tr><tr><td>Random</td><td>Random</td><td>30.02</td><td>33.59</td><td>36.64</td><td>37.56</td><td>38.23</td><td>38.79</td><td>39.33</td><td>39.54</td><td>39.95</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>30.13</td><td>34.04</td><td>38.43</td><td>40.09</td><td>45.74</td><td>49.63</td><td>54.08</td><td>60.31</td><td>64.95</td></tr><tr><td>SANT</td><td>Advanced AL + EAT</td><td>30.13</td><td>34.04</td><td>40.77</td><td>43.60</td><td>49.20</td><td>54.87</td><td>59.81</td><td>65.26</td><td>70.51</td></tr><tr><td rowspan="5">Knowledge graph completion (multi-class annotation) FreeBase</td><td>Zero-shot ChatGPT</td><td>None</td><td colspan="9">38.85</td></tr><tr><td>Few-shot ChatGPT</td><td>None</td><td colspan="9">51.60</td></tr><tr><td>Random</td><td>Random</td><td>22.46</td><td>35.19</td><td>42.77</td><td>48.54</td><td>52.50</td><td>55.31</td><td>57.75</td><td>60.26</td><td>61.88</td></tr><tr><td>Exp-gn</td><td>Advanced AL</td><td>22.86</td><td>36.56</td><td>46.84</td><td>50.56</td><td>64.18</td><td>69.99</td><td>73.98</td><td>79.22</td><td>83.48</td></tr><tr><td>SANT (Ours)</td><td>Advanced AL + EAT</td><td>22.86</td><td>36.56</td><td>45.03</td><td>52.17</td><td>58.79</td><td>74.53</td><td>77.04</td><td>83.77</td><td>88.13</td></tr></table>
+
+Table 5: More results on the annotation quality of SANT and baselines on model-annotated data. Two more datasets are considered here, including the SST-5 and FreeBase. Due to financial constraint, we compare our method with the ChatGPT, advanced AL method (Exp-gn), and random-based method.
+
+the FastText (Joulin et al., 2017) and obtain the sentence embedding by averaging the word embedding. We also remove the data if its  $90\%$  words (or more) have no Glove embeddings.
+
+# E Implementation details
+
+# E.1 Implementation of SANT
+
+SANT is flexible to integrate fancy models. It shares the same model annotator with semi-automatic annotation. However, considering the efficiency requirement of the human-model interaction, light-weighted annotation models or systems are always preferred (Desmond et al., 2021a; Chen et al., 2020a). Referring to Chen et al. (2020a), MLP is used as the EAT. For Implementation, our EAT outputs a two-dimensional predictive distribution as the triage signal rather than a real number between 0 and 1. By this means, we avoid setting a threshold to map the number to binary values of 0 and 1. We also use the class weight to  $L_{d}$  to handle the potential class imbalance problem and penalize mistakes in the minority class. Specifically, the weight is calculated based on the cumulative class distribution of human annotations. As for the model annotator in SANT, we use the corresponding model used in the semi-automatic annotation method. All our codes are implemented in Pytorch with one Nvidia GeForce RTX 3090. The three-layer FC with ReLu activation and dropout is utilized. Moreover, we set  $k = 3$  for retrieving top-k local information to integrate the input correlations. To ease the optimization, We also relax  $\mathbb{1}[y_t \neq \max(f_t(x_t))]$  to  $\mathbb{1}[HR@10 > 0]$  in  $L_{d}$  during the optimization of multi-label tagging annotator. Regarding the hyper-parameter  $T_{0}$  in the bi-weighting mechanism, we set it to be 0.2 for all experiments without tuning.
+
+Finally, to simulate the limited annotation budgets setting, we tune the annotation budgets from  $10\%$  to  $90\%$  with a step size of  $10\%$ , where  $10\%$  budgets mean the human can only annotate  $10\%$  data. We also require the human to annotate the first 1k data to initialize the model annotator for each annotation method. Note that when it comes to the Tagging task, the first 1k data and the  $10\%$  annotation budgets are almost equal. Therefore, in the case of  $10\%$  annotation budgets, all the budgets (i.e., the first 1k data) are used for human annotations. The corresponding results of the different triage-based methods in Table 1 and Table 2 are the same.
+
+# E.2 Implementation of Baselines
+
+Following Wang et al. (2022), our AL-based methods update the task model on the fly and output the model at the end of interactive data selection/annotation. As for MaxEntropy and Calibrated MaxEntropy, we measure uncertainty in sentiment annotation and knowledge graph completion, entropy is used, following a recent annotation method (Desmond et al., 2021b; Li et al., 2023). In the case of the tagging task, which is a multi-label classification task, total entropy (Depeweg et al., 2018) is used. Next, a dynamic mechanism (Wilder et al., 2021) is used to assign the input data, rather than adjusting the threshold. Specifically, an data  $x_{t}$  is assigned to human if  $(1 - d_{t}^{AL}(x_{t}))\max (f_{t}(x_{t})) < d_{t}^{AL}(x_{t})$  holds. Here,  $\max (\cdot)$  outputs the maximal value of the input vector. By this means, the human receives data with sufficiently high AL score, or sufficiently uncertain model output. As for Exp-gn and Ent-gn, the outputs of the active learning method (Wang et al., 2022) can not be easily normalized. In this case, half the data in a batch with the highest AL
+
+scores<sup>7</sup> are assigned to the human and the rest to the model.
+
+# E.3 Implementation of Human Simulation
+
+We consider human-machine cooperative data annotation under limited budgets, specifically a budget that restricts the number of human annotators that can be employed to annotate the entire dataset. In this case, our experiments focus solely on estimating the cost of human annotation, as it significantly outweighs the cost of model annotation, as evidenced by previous studies (Li et al., 2023; Gilardi et al., 2023). In particular, following previous works on data annotation (Chen et al., 2020a; Desmond et al., 2021a; Hedderich et al., 2021) and human-model interaction (Hwa, 2000; Kristjansson et al., 2004), we assume a consistent cost for human annotation per data point. This allows us to estimate the overall annotation budget using the number of human-annotated data points. To simulate human annotation behavior, we mask out the ground truth in the datasets. The ground truth is revealed when the human is selected by the data triage module. Moreover, to simulate the constraints of limited annotation budgets, we adjust the budget allocation from  $10\%$  to  $90\%$  of the dataset, with increments of  $10\%$ . This means that under a  $10\%$  budget, human annotators can only label  $10\%$  of the total dataset. This range of budget constraints allows us to assess the performance of SANT under varying budget limitations.
+
+# F Optimization Details of SANT
+
+In SANT, whenever the human reveals the ground truth of  $x_{t}$ , the model  $f_{t}$  is updated with  $L_{f}$  serving as the learning loss for the model (e.g., NLLLoss). However, our research focus is not on building a better model annotator. Therefore, we use different task-specified off-the-shelf model annotators in our experiments.
+
+To simplify the optimization process, we optimize the model annotator  $f_{t}$ , EAT  $d_{t}^{EAT}$ , and AL method  $d_{t}^{AL}$  independently. The loss function for AL, denoted as  $L_{AL}$  (if applicable), is included in the overall loss function  $\mathcal{L}$  of SANT, which is the sum of  $L_{f}$ ,  $L_{EAT}$ , and  $L_{AL}$ . It is important to note that all human-annotated data from past human-model interactions are utilized for optimization.
+
+During optimization, three things are taken into consideration. First, the output of the data triage module is relaxed to continuous values over  $[0,1]$  to avoid the integer optimization problem. Second, the top-k retrieval and data triage modules are implemented using the Gumbel-softmax reparameterization trick (Jang et al., 2016) to ease the non-differentiable problem. Finally, the overall loss function is a bi-level optimization problem (Bard, 2013), where optimizing the data allocation (i.e.,  $d_{t}^{AL}$  and  $d_{t}^{EAT}$ ) is nested within another problem of optimizing the model annotator  $f_{t}$ . Thus, we update  $f_{t}$  and data allocation independently and iteratively in a coordinate-descent style. Specifically, let  $\theta_f^k$ ,  $\theta_{dEAT}^k$ , and  $\theta_{dAL}^k$  be the network parameters for  $f$ ,  $d^{EAT}$ , and  $d^{AL}$  at optimization iteration  $k$ . The update procedure is given in Eq.(4). In the  $k^{th}$  optimization iteration, we first optimize the parameter of model  $f$  with the parameter of  $d^{EAT}$  and  $d^{AL}$  fixed. In the  $(k + 1)^{th}$  epoch, we exchange and optimize  $d^{EAT}$  and  $d^{AL}$  with  $f$  fixed, and so on.
+
+$$
+\begin{array}{l} \theta_ {f} ^ {k + 1} = \theta_ {f} ^ {k} - \bigtriangledown_ {f} \mathcal {L} \left(f ^ {k}, (d ^ {E A T}) ^ {k}, (d ^ {A L}) ^ {k}\right), \\ \theta_ {d ^ {E A T}} ^ {k + 1} = \theta_ {d ^ {E A T}} ^ {k} - \bigtriangledown_ {d ^ {E A T}} \mathcal {L} \left(f ^ {k + 1}, (d ^ {E A T}) ^ {k}, (d ^ {A L}) ^ {k}\right), \\ \theta_ {d ^ {A L}} ^ {k + 1} = \theta_ {d ^ {A L}} ^ {k} - \bigtriangledown_ {d ^ {A L}} \mathcal {L} \left(f ^ {k + 1}, \left(d ^ {E A T}\right) ^ {k}, \left(d ^ {A L}\right) ^ {k}\right). \tag {4} \\ \end{array}
+$$
+
+In our approach, EAT is optimized towards measuring the predictive ability of  $f_{t + 1}$  instead of  $f_{t}$ . It is important to note that if the AL method does not contain a learnable parameter, such as entropy-based AL, it can be omitted from the optimization process.
+
+# G Automatic annotation via ChatGPT
+
+Since ChatGPT has demonstrated strong zero-shot, few-shot, and CoT capabilities in various domains and tasks (Liu et al., 2023; Hu et al., 2023; Shah and Chava, 2023), we use ChatGPT for automatic data annotation experiments. Specifically, we test various prompting methods for ChatGPT, including zero-shot, few-shot, zero-shot CoT, and few-shot CoT prompts. The prompts for three annotation tasks are provided in Table 7. Due to the cost of using ChatGPT, we sample a subset for each task and then deliver it to ChatGPT for data annotation. Specifically, for the sentiment annotation task, we randomly sample 1000 data for ChatGPT annotation. As for the knowledge graph completion task, the WN18RR data contains multiple relation types.
+
+<table><tr><td rowspan="2">Res.</td><td colspan="3">KGC</td><td colspan="3">SA</td><td colspan="3">Tagging</td></tr><tr><td>SANT (re-anno.)</td><td>SANT (re-train+re-anno.)</td><td>SANT</td><td>SANT (re-anno.)</td><td>SANT (re-train+re-anno.)</td><td>SANT</td><td>SANT (re-anno.)</td><td>SANT (re-train+re-anno.)</td><td>SANT</td></tr><tr><td>10%</td><td>61.93</td><td>-</td><td>59.63</td><td>86.05</td><td>-</td><td>85.94</td><td>53.90</td><td>-</td><td>53.90</td></tr><tr><td>20%</td><td>68.15</td><td>-</td><td>65.45</td><td>88.73</td><td>-</td><td>88.46</td><td>60.15</td><td>-</td><td>59.92</td></tr><tr><td>30%</td><td>74.67</td><td>-</td><td>73.98</td><td>92.09</td><td>-</td><td>91.55</td><td>66.32</td><td>-</td><td>66.57</td></tr><tr><td>40%</td><td>79.72</td><td>-</td><td>79.79</td><td>95.01</td><td>-</td><td>94.47</td><td>71.91</td><td>-</td><td>72.61</td></tr><tr><td>50%</td><td>86.04</td><td>-</td><td>85.89</td><td>97.32</td><td>-</td><td>96.89</td><td>78.10</td><td>-</td><td>77.63</td></tr><tr><td>60%</td><td>91.49</td><td>93.93</td><td>91.47</td><td>98.64</td><td>99.09</td><td>98.51</td><td>82.63</td><td>87.20</td><td>82.64</td></tr><tr><td>70%</td><td>95.60</td><td>95.61</td><td>95.51</td><td>99.15</td><td>99.24</td><td>99.10</td><td>88.33</td><td>89.31</td><td>88.34</td></tr><tr><td>80%</td><td>97.46</td><td>97.48</td><td>97.46</td><td>99.55</td><td>99.53</td><td>99.53</td><td>93.64</td><td>93.66</td><td>93.65</td></tr><tr><td>90%</td><td>98.71</td><td>98.72</td><td>98.72</td><td>99.80</td><td>99.76</td><td>99.78</td><td>96.85</td><td>96.91</td><td>96.85</td></tr></table>
+
+Table 6: The overall annotation quality of SANT after the model annotator (re-train and) re-annotating the data that are previously model-annotated. Here, '-' means no human budgets left when data run out.
+
+As a result, we choose 100 data at random from each class for ChatGPT annotation. Last, ChatGPT fails at the tagging task, because the dataset CiteU-Like contains 500 different tags in our experiments. Despite our several attempts, ChatGPT is unable to select one or more tags from the 500 available ones based on the semantics of the input words. Therefore, we cut down the candidates to 200.
+
+In practice, as ChatGPT has been banned in some countries, we use Monica (refer to https://monica.im/) for annotations, which is a Chrome extension powered by ChatGPT API. As free users of Monica, we have a daily usage limit. Thus, we require Monica to annotate multiple instances at a time (precisely, 20 at a time).
+
+# H Practical suggestions for the interactive annotation system implementation
+
+To implement SANT into an annotation system in the practical scenario, some tricks might be helpful to further reduce human annotation labor and improve the annotation quality of both the human and model.
+
+# H.1 Improving user experience
+
+Considering generating an annotated dataset is usually time-consuming and expensive for the human, building an annotation tool with a user-friendly UI into the system plays an important role. A previous study on the tool design shows that if offering a better user experience during the annotation (Cerezo et al., 2021), the annotation quality improves.
+
+# H.2 Providing annotation suggestions to human
+
+When the human is selected to annotate data, the annotation system could attach the model predictions as suggestions to the human (Desmond et al.,
+
+2022; Klie et al., 2018), avoiding the need to annotate data from scratch and thus reducing human cognitive labor.
+
+# H.3 On error-prone human annotations
+
+Human annotations could be error-prone and inconsistent. It would usually hamper the overall annotation quality. Some works from other research domains are devoted to examining the error-prone human annotations, such as the crowd-sourced data annotation (Shirani et al., 2019; Larson et al., 2020). Considering it is not the research focus of this paper, we suggest the reader to those works before implementing a real human-model interactive system.
+
+# H.4 Post hoc correction for model annotation
+
+The system might also involve more strategies to improve model annotations in a post hoc way. The system could use the model annotator from the final round, and re-announce data, that was originally model-annotated at the early annotation rounds. If the final model in SANT is properly trained, the quality of the annotations might be further enhanced. If there are some budgets left for the human to annotate, the final model could be re-trained before the re-annotating, together with the newly human-annotated data.
+
+Taking SANT for example, Table 6 demonstrates the improvement in annotation quality after the model annotator re-annotated the data that were previously model-annotated. The re-annotation does help to improve the quality, and such improvement could be further enhanced by the re-training procedure. However, model re-training and model re-annotating require extra time cost. In the real-world scenario, one should decide whether to use the two extra modules (i.e., the model re-training and
+
+model re-annotating) based on the product form of his or her annotation system.
+
+<table><tr><td>Task</td><td>Zero-shot</td><td>Zero-shot CoT</td></tr><tr><td>SA</td><td>##### Read the sentences and identify their sentiment. There are two options: negative or positive ### 1. [INPUT_WORD1] 2. [INPUT_WORD2]</td><td>#### Read the sentences and identify their sentiment. You should think step by step and output the sentiment of each sentence. There are two optional sentiments: negative or positive ### 1. [INPUT_WORD1] 2. [INPUT_WORD2]</td></tr><tr><td>KGC</td><td>##### Read the word pairs and identify their semantic relation. There are 11 options: entailment, hyponym, hypernym , member, synonym, antonym, ... ### 1. [INPUT_WORD1], [INPUT_WORD2] 2. [INPUT_WORD3], [INPUT_WORD4]</td><td>#### Read the word pairs and identify their semantic relation. You should think step by step and output the semantic relation of each sentence. You have the following options: entailment, hyponym, hypernym, member, synonym, antonym, ... ### 1. [INPUT_WORD1], [INPUT_WORD2] 2. [INPUT_WORD3], [INPUT_WORD4]</td></tr><tr><td>Tagging</td><td>##### Read the following sentences, choose one or more tags from the TagSet that correspond to each sentence&#x27;s semantics. TagSet{...}  ### 1. [INPUT_WORD1] 2. [INPUT_WORD2]</td><td>#### Read the following sentences, choose one or more tags from the TagSet that correspond to each sentence&#x27;s semantics. You should think step by step and output the tag list of each sentence. TagSet{...}  ### 1. [INPUT_WORD1] 2. [INPUT_WORD2]</td></tr><tr><td>Task</td><td>Few-shot</td><td>Few-shot CoT</td></tr><tr><td>SA</td><td>##### Read the sentences and identify their sentiment. There are two options: negative or positive Here are two examples that you can refer to. Example 1: Input sentence: What a script, what a story, what a mess! Sentiment: negative Example 2: Input sentence: This is a great movie. Too bad it is not available on home video. Sentiment: positive  ### 1. [INPUT_WORD1] 2. [INPUT_WORD2]</td><td>## Read the sentences and identify their sentiment. You should think step by step and output the sentiment of each sentence. There are two optional sentiments: negative or positive Example 1: Input sentence: What a script, what a story, what a mess! Think step by step: He/she thinks the script and story are messy. Therefore, the sentiment is negative. Sentiment: negative Example 2: Input sentence: This is a great movie. Too bad it is not available on home video. Think step by step: He/she thinks the movie is great and he/she would like to watch it on home video. He/she really love this movie. Therefore, the sentiment is positive Sentiment: positive  ### 1. [INPUT_WORD1] 2. [INPUT_WORD2]</td></tr><tr><td>KGC</td><td>##### Read the word pairs and identify their semantic relation. You have the following options: entailment, hyponym, hypernym, ... Here are some examples that you can refer to. Example 1: Input word pair: 3D, film semantic relation: hyponym Example 2: Input word pair: ability, unfitness semantic relation: antonym Example 3: Input word pair: abominably,atrociously semantic relation: synonym Example 4: Input word pair: acclaim, sanction semantic relation: entailment  ### 1. [INPUT_WORD1], [INPUT_WORD2] 2. [INPUT_WORD3], [INPUT_WORD4] ...</td><td>## Read the word pairs and identify their semantic relation. You should think step by step and output the semantic relation of each sentence. You have the following options: hyponym, antonym, synonym, hypernym, entailment, member,... Here are some examples that you can refer to. Example 1: Input word pair: 3D, film Think step by step: 3D films are a type of film, so the relationship between the two words is hyponym semantic relation: hyponym Example 2: Input word pair: ability, unfitness Think step by step: In most cases, the term &quot;unfitness&quot; refers to anything or someone who is not appropriate for the task at hand, meaning that their ability is inadequate. Therefore, the relationship between the two words is antonym semantic relation: antonym Example 3: Input word pair: abominably,atrociously Think step by step: Abhorrently and atrociously both refer to something that is extremely unpleasant or unfair. Therefore, the relationship between the two words is synonym semantic relation: synonym Example 4: Input word pair: acclaim, sanction Think step by step: Acknowledge is used to describe enthusiastic public praise. The verb sanction can be used to signify to grant permission or approval for (an activity). So the relationship between the two words is entailment semantic relation: entailment  ### 1. [INPUT_WORD1], [INPUT_WORD2] 2. [INPUT_WORD3], [INPUT_WORD4]</td></tr><tr><td>Tagging</td><td>##### Read the following sentences, choose one or more tags from the TagSet that correspond to each sentence&#x27;s semantics. Here are one example that you can refer to. Example 1: Input sentence: the metabolic world of escherichia coli ... Tag list: ...  ### 1. [INPUT Sentence1] 2. [INPUT Sentence2]</td><td>## Read the following sentences, choose one or more tags from the TagSet that correspond to each sentence&#x27;s semantics. You should think step by step and output the tag list of each sentence. Here are one example that you can refer to. Example 1: Input sentence: the metabolic world of escherichia coli is not small to elucidate ... Think step by step: ... Tag list: ...  ### 1. [INPUT Sentence1] 2. [INPUT Sentence2]</td></tr></table>
+
+Table 7: ChatGPT prompts for three annotation tasks. We consider four kinds prompt, including zero-shot, few-shot, zero-shot CoT, and few-shot CoT prompts.

2024/Selective Annotation via Data Allocation_ These Data Should Be Triaged to Experts for Annotation Rather Than the Model/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22447281cbc74af74f36433b7fd053ef55ec46f7ab8a9a855454311a24957cdf
+size 1078018

2024/Self-Consistency Boosts Calibration for Math Reasoning/610e067d-96b2-49b7-8e9d-e89fb3c34d6e_content_list.json

@@ -0,0 +1,1117 @@
+[
+    {
+        "type": "text",
+        "text": "Self-Consistency Boosts Calibration for Math Reasoning",
+        "text_level": 1,
+        "bbox": [
+            203,
+            84,
+            791,
+            104
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Ante Wang $^{1,2}$ , Linfeng Song $^{3*}$ , Ye Tian $^{3}$ , Baolin Peng $^{3}$ , Lifeng Jin $^{3}$ , Haitao Mi $^{3}$ , Jinsong Su $^{1,2*}$  and Dong Yu $^{3}$",
+        "bbox": [
+            161,
+            112,
+            840,
+            146
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ School of Informatics, Xiamen University, China",
+        "bbox": [
+            297,
+            147,
+            704,
+            164
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural Heritage",
+        "bbox": [
+            112,
+            164,
+            889,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "of Fujian and Taiwan (Xiamen University), Ministry of Culture and Tourism, China",
+        "bbox": [
+            161,
+            181,
+            838,
+            197
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$ Tencent AI Lab, Bellevue, WA",
+        "bbox": [
+            369,
+            198,
+            631,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "wangante@stu.xmu.edu.cn, lfsong@global.tencent.com, jssu@xmu.edu.cn",
+        "bbox": [
+            159,
+            214,
+            838,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Calibration, which establishes the correlation between accuracy and model confidence, is important for LLM development. We design three off-the-shelf calibration methods based on self-consistency (Wang et al., 2022) for math reasoning tasks. Evaluation on two popular benchmarks (GSM8K and MathQA) using strong open-source LLMs (Mistral and LLaMA2) shows that our methods better bridge model confidence and accuracy than existing methods based on  $p(\\text{True})$  (Kadavath et al., 2022) or logit (Guo et al., 2017).",
+        "bbox": [
+            141,
+            281,
+            460,
+            451
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            466,
+            258,
+            481
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Mathematical reasoning tasks (Cobbe et al., 2021; Hendrycks et al., 2021; Amini et al., 2019) involve mapping a question into a series of equations, which are then solved to obtain the final answer. Math reasoning has long been recognized challenging. Existing solutions propose to map input questions into equations via semantic parsing (Matsuzaki et al., 2017; Hopkins et al., 2017) or AST decoding (Li et al., 2019; Qin et al., 2021; Wu et al., 2021). Yet, the performance can degrade dramatically even with slight changes to the questions (Patel et al., 2021; Li et al., 2022).",
+        "bbox": [
+            112,
+            493,
+            489,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, large language models (LLM, Achiam et al. 2023; Touvron et al. 2023; Jiang et al. 2024) have shown great potential for solving many math reasoning tasks, even though they are not specifically trained on these tasks. For instance, with chain-of-thought prompting (Wei et al., 2022) and self-consistency (Wang et al., 2022), open-source LLMs, such as Mixtral  $8 \\times 7\\mathrm{B}$  (Jiang et al., 2024), can reach an accuracy of around  $80\\%$  on the GSM8K benchmark (Cobbe et al., 2021). On the other hand, conventional pretrained models (e.g., T5 (Raffel et al., 2020)) that are specifically finetuned on the GSM8K training set can only report",
+        "bbox": [
+            112,
+            686,
+            489,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c21cca1ce5f7f370f4252825d74038935451cef2b3e1f95e1400d3e24645e1b0.jpg",
+        "image_caption": [
+            "Figure 1: Comparison of several calibration methods on Mistral-7B, where  $SC w / \\mathcal{F}_{CN}$  is one of our methods based on self-consistency, which will be introduced in §3."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            249,
+            857,
+            444
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "accuracies around  $10\\%$  to  $20\\%$  (Shridhar et al., 2023; Magister et al., 2023).",
+        "bbox": [
+            507,
+            543,
+            882,
+            575
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, LLMs lack adequate calibration out of the box – the probabilities of model predictions are often poorly aligned with the actual accuracy (Xiong et al., 2023; Chen et al., 2023). Calibration is important for LLM development, as a well-calibrated LLM can precisely tell how likely its responses are correct or not. With such information, LLM developers can take multiple options to handle low-confidence responses, such as letting the LLM refuse to answer or keep resampling until a confident response is produced.",
+        "bbox": [
+            505,
+            577,
+            884,
+            755
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose calibration methods based on self-consistency (Wang et al., 2022) for math reasoning tasks. Self-consistency performs clustering over multiple LLM samples before picking one from the largest cluster as the response to each input query. Here we consider several ways to estimate model confidence using the clustering results: cluster size that estimates how many samples agree with the selected one, cluster number that measures to what extent samples disagree with",
+        "bbox": [
+            507,
+            758,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding Authors",
+        "bbox": [
+            141,
+            904,
+            297,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6023",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6023-6029",
+        "bbox": [
+            216,
+            945,
+            779,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            273,
+            958,
+            724,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "each other, and pairwise comparison that captures relative differences between pairs of clusters.",
+        "bbox": [
+            112,
+            84,
+            485,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments using strong open-source LLMs: Mistral (Jiang et al., 2023, 2024) and LLaMA2 (Touvron et al., 2023) series models with / without being aligned with instructions. Results on GSM8K (Cobbe et al., 2021) and MathQA (Amini et al., 2019) show that all our methods better calibrate these models than exiting popular methods, such as  $p(\\text{True})$  (Kadavath et al., 2022) and logit (Guo et al., 2017) over the whole reasoning path or target answer span only.",
+        "bbox": [
+            112,
+            117,
+            489,
+            277
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Preview: Self-Consistency with CoT Prompting",
+        "text_level": 1,
+        "bbox": [
+            112,
+            288,
+            460,
+            322
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "For math reasoning, there are usually multiple trajectories to reach the final solution. To replicate this process, Wang et al. (2022) initially sample various reasoning paths  $r_1, \\dots, r_N$  from the LLM given input  $x$  with Chain-of-Thought (CoT) prompting. Then, the answers  $a_1, \\dots, a_N$  are extracted from the paths, and the most consistent answer (the one win by majority vote among the answers) is selected as the final answer  $\\mathbf{a}$ :",
+        "bbox": [
+            112,
+            331,
+            489,
+            475
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {a} = \\max  _ {\\hat {a}} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left(a _ {i} = \\hat {a}\\right), \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            205,
+            483,
+            487,
+            526
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nr _ {i}, a _ {i} \\sim \\mathrm {L L M} _ {\\theta} (x),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            258,
+            531,
+            391,
+            546
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $r_i, a_i$  denote the  $i$ -th sampled reasoning path and its corresponding answer, respectively.",
+        "bbox": [
+            112,
+            558,
+            485,
+            590
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Calibration using Self-Consistency",
+        "text_level": 1,
+        "bbox": [
+            112,
+            601,
+            445,
+            618
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "After performing self-consistency on input  $x$  using  $\\mathrm{LLM}_{\\theta}$ , we obtain a set of clusters  $\\mathcal{C} = \\{c_1, \\dots, c_{|\\mathcal{C}|}\\}$  with each cluster  $c_i$  comprising  $n_i$  sampled responses with the same answers. We design the following strategies, tailored to the characteristics of these clusters, to estimate the confidence of  $\\mathrm{LLM}_{\\theta}$ .",
+        "bbox": [
+            112,
+            627,
+            487,
+            722
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Cluster Number We initially consider the Cluster Number  $|\\mathcal{C}|$ . This is motivated by the finding of previous work (Wang et al., 2022; Xiong et al., 2023): LLMs tend to generate consistent answers when they are confident about their predictions, and thus the cluster number (number of distinct answers) tends to be small. We further divide the cluster number by the sample size  $N$  to normalize",
+        "bbox": [
+            112,
+            732,
+            487,
+            860
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the score into the range of [0, 1], before reversing it by “ $1 - x$ ”:",
+        "bbox": [
+            507,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {C N} (x, \\theta) = 1 - \\frac {| \\mathcal {C} |}{N}. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            608,
+            128,
+            882,
+            159
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Cluster Size In a similar vein, we adopt the Cluster Size: the number of samples (e.g.,  $n_i$ ) within a specific cluster (e.g.,  $c_i$ ). Again, we compute its proportion relative to the total sample size to normalize the score into the range [0, 1]:",
+        "bbox": [
+            507,
+            170,
+            882,
+            250
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {C S} (x, \\theta) = \\frac {n _ {i}}{N}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            628,
+            262,
+            882,
+            288
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In contrast to the cluster number, the cluster size is more universally applicable across diverse prompts, as the cluster number can easily become ineffective when the output space of an LLM is restricted, such as when options for a question are provided.",
+        "bbox": [
+            507,
+            300,
+            882,
+            381
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Pairwise Comparison The Cluster Number and Cluster Size primarily consider the number of distinct answers and the number of sampled paths within a single cluster, respectively. They both overlook the information by comparing different clusters. For example, they may fail to consider the situation when the sizes of the top-ranked clusters are close. Consequently, we introduce the Pairwise Comparison method, which computes the winning rate of the chosen cluster  $(c_i)$  against each of the remaining clusters:",
+        "bbox": [
+            507,
+            390,
+            882,
+            567
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {F} _ {P C} (x, \\theta) = \\prod_ {j \\neq i} ^ {| \\mathcal {C} |} \\frac {n _ {i}}{n _ {i} + n _ {j}}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            594,
+            580,
+            882,
+            627
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $\\frac{n_i}{n_i + n_j}$  represents the winning rate of selected cluster  $c_i$  against another cluster  $c_j$  by comparing the respective cluster sizes.",
+        "bbox": [
+            507,
+            639,
+            880,
+            687
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            700,
+            655,
+            715
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "4.1 Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            727,
+            601,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Datasets We conduct experiments on two popular math reasoning benchmarks of different type of questions, GSM8K (Cobbe et al., 2021) and MathQA (Amini et al., 2019). Particularly, GSM8K comprises 1,319 linguistically diverse grade school math word problems for testing. On the other hand, MathQA offers 2,985 multiple-choice math word problems for evaluation.",
+        "bbox": [
+            507,
+            747,
+            882,
+            876
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics We adopt Brier Score and Expected Calibration Error (ECE) as evaluating",
+        "bbox": [
+            507,
+            887,
+            882,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Here we follow common practice to adopt demonstrations with rationales for pretrained only models (e.g., Mistral-7B) and use \"Let's think step by step\" (Kojima et al., 2022) for instruction-tuned models (e.g., Mistral-7B-Inst).",
+        "bbox": [
+            112,
+            868,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6024",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/a5af16ddb9cc04ea9da3f0d8ca4038ea58dd919380229d2cb443d54ce49c75ed.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\"></td><td colspan=\"2\">Mistral-7B</td><td colspan=\"2\">Mistral-7B-Inst</td><td colspan=\"2\">Mixtral-8×7B</td><td colspan=\"2\">Mixtral-8×7B-Inst</td></tr><tr><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td></tr><tr><td rowspan=\"8\">GSM8K</td><td>logit w/ Path</td><td>0.394</td><td>0.399</td><td>0.414</td><td>0.414</td><td>0.178</td><td>0.265</td><td>0.233</td><td>0.252</td></tr><tr><td>logit w/ Answer</td><td>0.505</td><td>0.488</td><td>0.467</td><td>0.458</td><td>0.307</td><td>0.312</td><td>0.236</td><td>0.238</td></tr><tr><td>p(True)</td><td>0.127</td><td>0.267</td><td>0.406</td><td>0.407</td><td>0.070</td><td>0.201</td><td>0.195</td><td>0.198</td></tr><tr><td>Self-Consistency</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ FCN</td><td>0.092</td><td>0.186</td><td>0.125</td><td>0.182</td><td>0.136</td><td>0.157</td><td>0.075</td><td>0.092</td></tr><tr><td>w/ FCS</td><td>0.148</td><td>0.185</td><td>0.163</td><td>0.180</td><td>0.173</td><td>0.156</td><td>0.085</td><td>0.086</td></tr><tr><td>w/ FPC</td><td>0.248</td><td>0.229</td><td>0.253</td><td>0.226</td><td>0.238</td><td>0.194</td><td>0.110</td><td>0.096</td></tr><tr><td>w/ ALL</td><td>0.101</td><td>0.172</td><td>0.123</td><td>0.162</td><td>0.135</td><td>0.141</td><td>0.078</td><td>0.083</td></tr><tr><td rowspan=\"8\">MathQA</td><td>logit w/ Path</td><td>0.500</td><td>0.499</td><td>0.539</td><td>0.510</td><td>0.333</td><td>0.380</td><td>0.364</td><td>0.373</td></tr><tr><td>logit w/ Answer</td><td>0.356</td><td>0.362</td><td>0.291</td><td>0.319</td><td>0.266</td><td>0.290</td><td>0.220</td><td>0.281</td></tr><tr><td>p(True)</td><td>0.350</td><td>0.309</td><td>0.271</td><td>0.317</td><td>0.228</td><td>0.253</td><td>0.273</td><td>0.272</td></tr><tr><td>Self-Consistency</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ FCN</td><td>0.331</td><td>0.336</td><td>0.374</td><td>0.359</td><td>0.143</td><td>0.236</td><td>0.128</td><td>0.215</td></tr><tr><td>w/ FCS</td><td>0.091</td><td>0.225</td><td>0.114</td><td>0.227</td><td>0.080</td><td>0.190</td><td>0.035</td><td>0.171</td></tr><tr><td>w/ FPC</td><td>0.052</td><td>0.220</td><td>0.065</td><td>0.219</td><td>0.143</td><td>0.203</td><td>0.054</td><td>0.174</td></tr><tr><td>w/ ALL</td><td>0.144</td><td>0.238</td><td>0.172</td><td>0.242</td><td>0.089</td><td>0.190</td><td>0.072</td><td>0.177</td></tr></table>",
+        "bbox": [
+            147,
+            80,
+            850,
+            342
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Main test results on GSM8K and MathQA when using Mistral family models. Specifically,  $*$ -Inst indicates instruction-tuned models.",
+        "bbox": [
+            112,
+            351,
+            880,
+            380
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "metrics following common practice (Geng et al., 2023).",
+        "bbox": [
+            112,
+            406,
+            487,
+            436
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given instances  $(x_{1},y_{1}),\\dots,(x_{\\bar{N}},y_{\\bar{N}})$  and their corresponding LLM predictions  $\\hat{y}_1,\\dots,\\hat{y}_{\\bar{N}}$  , ECE is computed by first binning the predictions into  $M = 10$  intervals based on their LLM confidence levels (e.g.,  $p(\\hat{y}_i)$  ). For each bin (e.g.  $B_{m}$  ), it then calculates the accuracy (acc  $(B_{m})$  ) and the average confidence  $(\\mathrm{conf}(B_m))$  ..",
+        "bbox": [
+            112,
+            437,
+            485,
+            550
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {a c c} \\left(B _ {m}\\right) = \\frac {1}{\\left| B _ {m} \\right|} \\sum_ {i \\in B _ {m}} \\mathbb {1} \\left(y _ {i} = \\hat {y} _ {i}\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            173,
+            555,
+            433,
+            593
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {c o n f} \\left(B _ {m}\\right) = \\frac {1}{\\left| B _ {m} \\right|} \\sum_ {i \\in B _ {m}} p \\left(\\hat {y} _ {i}\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            164,
+            595,
+            394,
+            633
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $|B_m|$  is the number of samples in bin  $B_m$ . Finally, the difference between accuracy and confidence is averaged across all bins to obtain the ECE score:",
+        "bbox": [
+            112,
+            639,
+            487,
+            702
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {E C E} = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{\\bar {N}} \\left| \\operatorname {a c c} \\left(B _ {m}\\right) - \\operatorname {c o n f} \\left(B _ {m}\\right) \\right| \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            132,
+            707,
+            487,
+            749
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "As another popular metric, Brier score is similar to ECE but conducted at the instance level:",
+        "bbox": [
+            112,
+            755,
+            487,
+            785
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {B r i e r} = \\frac {1}{\\bar {N}} \\sum_ {i = 1} ^ {\\bar {N}} \\left(p \\left(\\hat {y} _ {i}\\right) - \\mathbb {1} \\left(y _ {i} = \\hat {y} _ {i}\\right)\\right) ^ {2}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            157,
+            790,
+            487,
+            833
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Both metrics range from 0 to 1 with lower values indicating better calibration. We take Brier score as the main metric, as it is more robust to unbalanced distribution across bins (e.g. instances concentrate to one or two bins).",
+        "bbox": [
+            112,
+            839,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Settings We conduct experiments on LLaMA2 and Mistral-family models and investigate both pretrained or instruction-tuned variations. We use nucleus sampling to obtain  $N = 16$  samples by default for each instance and use temperatures of 0.6 / 1.0 for all pretrained / instruction-tuned models.",
+        "bbox": [
+            507,
+            406,
+            882,
+            501
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Baselines We take the three representative baselines below for comparison:",
+        "bbox": [
+            507,
+            512,
+            882,
+            544
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- logit w/ Path: It averages the probabilities of the tokens from the whole path to estimate the confidence of each prediction.",
+            "- logit w/ Answer: It is similar to logit w/ Path but only consider the tokens from the predicted answer span.",
+            "-  $p(\\text{True})$ : It asks the LLM itself to classify its prediction as True or False. Then, it takes the predicted probability of True as its confidence. We follow Kadavath et al. (2022) to construct 8-shot demonstrations for prompting pretrained models but directly use instruction for instruction-tuned models."
+        ],
+        "bbox": [
+            517,
+            556,
+            884,
+            787
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.2 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            801,
+            717,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Main Results Table 1 presents the main results obtained from both benchmarks using Mistral-family models.  $p(\\text{True})$  performs best among the baselines, echoing the findings of Kadavath et al. (2022). However, due to its reliance on prompt design and in-context examples to aid the LLM",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6025",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/a1e06ad14dcc56e9e946cbc40d777400b77ac7d6596c476f5f54e1e3485ea829.jpg",
+        "image_caption": [
+            "Figure 2: Calibration results on GSM8K when using Mixtral-  $8\\times 7\\mathrm{B}$  -Inst with different  $N$"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            80,
+            440,
+            262
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "to classify its predictions, it can be challenging to construct effective demonstrations or instructions.",
+        "bbox": [
+            112,
+            330,
+            487,
+            361
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In general, self-consistency-based methods surpass baselines in most cases regarding Brier and ECE, validating the efficacy of employing self-consistency features for estimating model confidence. We also note that baselines can occasionally yield impressive ECE scores ( $p(\\text{True})$  on GSM8K with Mixtral-  $8 \\times 7\\mathrm{B}$ ). However, we observe that this is attributed to the concentration of most samples in just a few bins (e.g., Figure 1), leading to unreliable measurements. Nevertheless, our approaches still exhibit strong performance in terms of ECE scores across various settings.",
+        "bbox": [
+            112,
+            363,
+            489,
+            554
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Among the self-consistency-based methods,  $\\mathcal{F}_{CN}$  yields better ECE results on GSM8K, while  $\\mathcal{F}_{CS}$  achieves the highest Brier score. Conversely, for MathQA,  $\\mathcal{F}_{CN}$  performs significantly worse than the other two. This is because MathQA is a multi-choice task, and thus the cluster number of LLM answers is strictly limited by the provided choices. In conclusion,  $\\mathcal{F}_{CS}$  demonstrates greater generality across diverse settings, but  $\\mathcal{F}_{CN}$  and  $\\mathcal{F}_{PC}$  do offer improved estimation in certain cases.",
+        "bbox": [
+            112,
+            556,
+            489,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, we also explore averaging the proposed three metrics as a unified one, denoted as \"ALL\". Results show that it does not typically yield the best results. However, it serves as a robust choice, demonstrating competitive performance.",
+        "bbox": [
+            112,
+            717,
+            489,
+            797
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Influence of Sample Size  $N$  Previous research (Wang et al., 2022) has demonstrated that the sample size  $N$  can significantly affect the accuracy of self-consistency. When  $N$  increases, the model performance initially continues to improve before stabilizing once  $N$  reaches a sufficient level. Therefore, we take Mixtral-8×7B-Inst as a case study to",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/1cd8a97db0979f26313a691b83b5d230c757a396671c64c7a6ab36eefa059a52.jpg",
+        "image_caption": [
+            "Figure 3: Performance and calibration results on GSM8K using different models below sorted by their performance: ① LLaMA2-7B-Chat, ② LLaMA2-13B-Chat, ③ Mistral-7B-Inst, ④ LLaMA2-70B-Chat, ⑤ Mixtral-8×7B-Inst."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            82,
+            870,
+            261
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "examine the impact of  $N$  on calibration.",
+        "bbox": [
+            507,
+            356,
+            810,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As illustrated in Figure 2, the Brier scores for all our methods initially decline and then remain constant as  $N$  grows. For  $\\mathcal{F}_{CS}$  and  $\\mathcal{F}_{PC}$ ,  $N = 8$  is adequate for accurate estimation. In contrast,  $\\mathcal{F}_{CN}$  requires a larger  $N$ , indicating that the cluster number is more susceptible to the randomness of sampling.",
+        "bbox": [
+            507,
+            372,
+            884,
+            483
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Correlation between Performance and Calibration We finally explore the associations between model performance (Accuracy) and calibration. Figure 3 showcases the results on instruction-tuned LLaMA2 and Mistral series models, arranged in ascending order based on their performance. We generally observe a positively correlated trend between calibration (lower the better) and performance (higher the better) among the studied models. This observation indicates that more powerful models also exhibit enhanced calibration, echoing the findings of Kadavath et al. (2022). This phenomenon can be attributed to the fact that when a tested LLM is stronger, it is capable of generating more reasonable and consistent responses, leading to improved calibration.",
+        "bbox": [
+            507,
+            494,
+            885,
+            751
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            764,
+            640,
+            778
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we extend the widely-used inference strategy, self-consistency, to the field of calibration. Specifically, we develop three off-the-shelf calibration methods based on self-consistency for math reasoning tasks. Compared to conventional methods (p(True) and logit), our approaches yield significantly improved ECE and Brier scores on popular GSM8K and MathQA datasets. Future",
+        "bbox": [
+            507,
+            790,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6026",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "research directions include designing more effective calibration methods, leveraging richer features and employing more strategies (e.g., temperature scaling (Guo et al., 2017)) to enhance calibration performance. Our ultimate goal is to construct reliable and honest LLMs with the help of accurate confidence estimation.",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            209,
+            220,
+            224
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our methods are founded on the principle of self-consistency, which relies on sampling multiple times for prediction. This approach, however, needs additional cost for inference, which may not be efficient and eco-friendly. Besides, our current work is limited to mathematical problems and does not explore other types of tasks, such as question-answering. Although it is crucial to extend our methods to encompass other tasks, this is non-trivial due to the inherent difficulty in dividing certain tasks' model predictions into distinct clusters.",
+        "bbox": [
+            112,
+            236,
+            489,
+            429
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            442,
+            265,
+            457
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We focus on ethical AI research and strive to achieve a balance between technological advancements and our ethical responsibilities. This work studies calibration, which aims to enhance the reliability of LLMs. Besides, we conduct experiments only on publicly available datasets, upholding privacy and anonymity rules.",
+        "bbox": [
+            112,
+            468,
+            489,
+            582
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            609,
+            213,
+            624
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.",
+            "Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. 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 2357-2367.",
+            "Yangyi Chen, Lifan Yuan, Ganqu Cui, Zhiyuan Liu, and Heng Ji. 2023. A close look into the calibration of pre-trained language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1343-1367."
+        ],
+        "bbox": [
+            115,
+            632,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.",
+            "Jiahui Geng, Fengyu Cai, Yuxia Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. 2023. A survey of language model confidence estimation and calibration. arXiv preprint arXiv:2311.08298.",
+            "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR.",
+            "Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).",
+            "Mark Hopkins, Cristian Petrescu-Prahova, Roie Levin, Ronan Le Bras, Alvaro Herrasti, and Vidur Joshi. 2017. Beyond sentential semantic parsing: Tackling the math sat with a cascade of tree transducers. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 795-804.",
+            "Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.",
+            "Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. 2024. Mixtral of experts. arXiv preprint arXiv:2401.04088.",
+            "Mandar Joshi, Eunsol Choi, Daniel S 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.",
+            "Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221.",
+            "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.",
+            "Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6027",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "The Eleventh International Conference on Learning Representations.",
+            "Jierui Li, Lei Wang, Jipeng Zhang, Yan Wang, Bing Tian Dai, and Dongxiang Zhang. 2019. Modeling intra-relation in math word problems with different functional multi-head attentions. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 6162-6167.",
+            "Zhongli Li, Wenxuan Zhang, Chao Yan, Qingyu Zhou, Chao Li, Hongzhi Liu, and Yunbo Cao. 2022. Seeking patterns, not just memorizing procedures: Contrastive learning for solving math word problems. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 2486-2496.",
+            "Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. 2023. Teaching small language models to reason. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1773-1781.",
+            "Takuya Matsuzaki, Takumi Ito, Hidenao Iwane, Hirokazu Anai, and Noriko H Arai. 2017. Semantic parsing of pre-university math problems. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2131-2141.",
+            "Arkil Patel, Satwik Bhattachamishra, and Navin Goyal. 2021. Are nlp models really able to solve simple math word problems? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2080-2094.",
+            "Jinghui Qin, Xiaodan Liang, Yining Hong, Jianheng Tang, and Liang Lin. 2021. Neural-symbolic solver for math word problems with auxiliary tasks. 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 5870-5881.",
+            "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.",
+            "Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2023. Distilling reasoning capabilities into smaller language models. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 7059-7073.",
+            "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.",
+            "Qinzhuo Wu, Qi Zhang, Zhongyu Wei, and Xuan-Jing Huang. 2021. Math word problem solving with explicit numerical values. 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 5859-5869.",
+            "Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi. 2023. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. In The Twelfth International Conference on Learning Representations."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            418
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A Experiments on TriviaQA",
+        "text_level": 1,
+        "bbox": [
+            510,
+            430,
+            771,
+            447
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Although this work mainly focuses on math problems, we also verify the effectiveness of our approach on TriviaQA (Joshi et al., 2017), a popular question-answering benchmark. Different from math problems, the outputs of LLMs for this task sometimes lack clear answers for clustering. For instance, Mistral-7B-Inst provides a lengthy response, \"The musical Phantom of the Opera premiered in the US on 10th December 1993...\", instead of an answer span \"Phantom of the Opera\", to the question \"Which Lloyd Webber musical premiered in the US on 10th December 1993?\".",
+        "bbox": [
+            507,
+            456,
+            884,
+            646
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To address this issue, we follow Kuhn et al. to employ natural language inference (NLI) to identify responses with the same answers<sup>2</sup>.",
+        "bbox": [
+            507,
+            649,
+            884,
+            696
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results are shown in Table 2. Our approach significantly surpasses the baselines, particularly when utilizing instruction-tuned models, further demonstrating the effectiveness of our method.",
+        "bbox": [
+            507,
+            697,
+            884,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ We adopt a publicly accessible model at https://huggingface.co/MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli.",
+        "bbox": [
+            507,
+            879,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6028",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/59d0cca0592af98845d2a0fd4964192044d951909b01097df92834b16602fd1a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"2\">Mistral-7B</td><td colspan=\"2\">Mistral-7B-Inst</td><td colspan=\"2\">Mixtral-8×7B</td><td colspan=\"2\">Mixtral-8×7B-Inst</td></tr><tr><td></td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td></tr><tr><td>logit</td><td>0.119</td><td>0.156</td><td>0.385</td><td>0.377</td><td>0.069</td><td>0.116</td><td>0.256</td><td>0.298</td></tr><tr><td>p(True)</td><td>0.077</td><td>0.162</td><td>0.326</td><td>0.318</td><td>0.099</td><td>0.131</td><td>0.314</td><td>0.316</td></tr><tr><td colspan=\"9\">Self-Consistency</td></tr><tr><td>w/ FCN</td><td>0.044</td><td>0.133</td><td>0.054</td><td>0.173</td><td>0.077</td><td>0.119</td><td>0.068</td><td>0.115</td></tr><tr><td>w/ FCS</td><td>0.043</td><td>0.132</td><td>0.072</td><td>0.186</td><td>0.039</td><td>0.107</td><td>0.039</td><td>0.116</td></tr><tr><td>w/ FPC</td><td>0.071</td><td>0.137</td><td>0.109</td><td>0.186</td><td>0.102</td><td>0.129</td><td>0.121</td><td>0.126</td></tr></table>",
+        "bbox": [
+            184,
+            423,
+            811,
+            548
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Test results on TriviaQA.",
+        "bbox": [
+            381,
+            558,
+            613,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "6029",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    }
+]

2024/Self-Consistency Boosts Calibration for Math Reasoning/610e067d-96b2-49b7-8e9d-e89fb3c34d6e_model.json

@@ -0,0 +1,1424 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.205,
+                0.085,
+                0.793,
+                0.105
+            ],
+            "angle": 0,
+            "content": "Self-Consistency Boosts Calibration for Math Reasoning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.162,
+                0.113,
+                0.842,
+                0.147
+            ],
+            "angle": 0,
+            "content": "Ante Wang\\(^{1,2}\\), Linfeng Song\\(^{3*}\\), Ye Tian\\(^{3}\\), Baolin Peng\\(^{3}\\), Lifeng Jin\\(^{3}\\), Haitao Mi\\(^{3}\\), Jinsong Su\\(^{1,2*}\\) and Dong Yu\\(^{3}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.298,
+                0.148,
+                0.705,
+                0.165
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)School of Informatics, Xiamen University, China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.165,
+                0.89,
+                0.181
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural Heritage"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.163,
+                0.182,
+                0.84,
+                0.198
+            ],
+            "angle": 0,
+            "content": "of Fujian and Taiwan (Xiamen University), Ministry of Culture and Tourism, China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.371,
+                0.199,
+                0.632,
+                0.214
+            ],
+            "angle": 0,
+            "content": "\\(^{3}\\)Tencent AI Lab, Bellevue, WA"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.16,
+                0.215,
+                0.84,
+                0.23
+            ],
+            "angle": 0,
+            "content": "wangante@stu.xmu.edu.cn, lfsong@global.tencent.com, jssu@xmu.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.282,
+                0.461,
+                0.452
+            ],
+            "angle": 0,
+            "content": "Calibration, which establishes the correlation between accuracy and model confidence, is important for LLM development. We design three off-the-shelf calibration methods based on self-consistency (Wang et al., 2022) for math reasoning tasks. Evaluation on two popular benchmarks (GSM8K and MathQA) using strong open-source LLMs (Mistral and LLaMA2) shows that our methods better bridge model confidence and accuracy than existing methods based on \\( p(\\text{True}) \\) (Kadavath et al., 2022) or logit (Guo et al., 2017)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.467,
+                0.26,
+                0.482
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.494,
+                0.49,
+                0.686
+            ],
+            "angle": 0,
+            "content": "Mathematical reasoning tasks (Cobbe et al., 2021; Hendrycks et al., 2021; Amini et al., 2019) involve mapping a question into a series of equations, which are then solved to obtain the final answer. Math reasoning has long been recognized challenging. Existing solutions propose to map input questions into equations via semantic parsing (Matsuzaki et al., 2017; Hopkins et al., 2017) or AST decoding (Li et al., 2019; Qin et al., 2021; Wu et al., 2021). Yet, the performance can degrade dramatically even with slight changes to the questions (Patel et al., 2021; Li et al., 2022)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.687,
+                0.49,
+                0.897
+            ],
+            "angle": 0,
+            "content": "Recently, large language models (LLM, Achiam et al. 2023; Touvron et al. 2023; Jiang et al. 2024) have shown great potential for solving many math reasoning tasks, even though they are not specifically trained on these tasks. For instance, with chain-of-thought prompting (Wei et al., 2022) and self-consistency (Wang et al., 2022), open-source LLMs, such as Mixtral \\(8 \\times 7\\mathrm{B}\\) (Jiang et al., 2024), can reach an accuracy of around \\(80\\%\\) on the GSM8K benchmark (Cobbe et al., 2021). On the other hand, conventional pretrained models (e.g., T5 (Raffel et al., 2020)) that are specifically finetuned on the GSM8K training set can only report"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.25,
+                0.858,
+                0.445
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.456,
+                0.883,
+                0.514
+            ],
+            "angle": 0,
+            "content": "Figure 1: Comparison of several calibration methods on Mistral-7B, where \\(SC w / \\mathcal{F}_{CN}\\) is one of our methods based on self-consistency, which will be introduced in §3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.544,
+                0.884,
+                0.576
+            ],
+            "angle": 0,
+            "content": "accuracies around \\(10\\%\\) to \\(20\\%\\) (Shridhar et al., 2023; Magister et al., 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.579,
+                0.885,
+                0.756
+            ],
+            "angle": 0,
+            "content": "However, LLMs lack adequate calibration out of the box – the probabilities of model predictions are often poorly aligned with the actual accuracy (Xiong et al., 2023; Chen et al., 2023). Calibration is important for LLM development, as a well-calibrated LLM can precisely tell how likely its responses are correct or not. With such information, LLM developers can take multiple options to handle low-confidence responses, such as letting the LLM refuse to answer or keep resampling until a confident response is produced."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.759,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In this work, we propose calibration methods based on self-consistency (Wang et al., 2022) for math reasoning tasks. Self-consistency performs clustering over multiple LLM samples before picking one from the largest cluster as the response to each input query. Here we consider several ways to estimate model confidence using the clustering results: cluster size that estimates how many samples agree with the selected one, cluster number that measures to what extent samples disagree with"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.142,
+                0.905,
+                0.298,
+                0.919
+            ],
+            "angle": 0,
+            "content": "*Corresponding Authors"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6023"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6023-6029"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.275,
+                0.959,
+                0.725,
+                0.972
+            ],
+            "angle": 0,
+            "content": "November 12-16, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.486,
+                0.117
+            ],
+            "angle": 0,
+            "content": "each other, and pairwise comparison that captures relative differences between pairs of clusters."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.118,
+                0.49,
+                0.278
+            ],
+            "angle": 0,
+            "content": "We conduct experiments using strong open-source LLMs: Mistral (Jiang et al., 2023, 2024) and LLaMA2 (Touvron et al., 2023) series models with / without being aligned with instructions. Results on GSM8K (Cobbe et al., 2021) and MathQA (Amini et al., 2019) show that all our methods better calibrate these models than exiting popular methods, such as \\( p(\\text{True}) \\) (Kadavath et al., 2022) and logit (Guo et al., 2017) over the whole reasoning path or target answer span only."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.289,
+                0.461,
+                0.323
+            ],
+            "angle": 0,
+            "content": "2 Preview: Self-Consistency with CoT Prompting"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.332,
+                0.49,
+                0.476
+            ],
+            "angle": 0,
+            "content": "For math reasoning, there are usually multiple trajectories to reach the final solution. To replicate this process, Wang et al. (2022) initially sample various reasoning paths \\( r_1, \\dots, r_N \\) from the LLM given input \\( x \\) with Chain-of-Thought (CoT) prompting. Then, the answers \\( a_1, \\dots, a_N \\) are extracted from the paths, and the most consistent answer (the one win by majority vote among the answers) is selected as the final answer \\( \\mathbf{a} \\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.206,
+                0.485,
+                0.488,
+                0.527
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathbf {a} = \\max  _ {\\hat {a}} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left(a _ {i} = \\hat {a}\\right), \\tag {1}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.26,
+                0.532,
+                0.393,
+                0.548
+            ],
+            "angle": 0,
+            "content": "\\[\nr _ {i}, a _ {i} \\sim \\mathrm {L L M} _ {\\theta} (x),\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.559,
+                0.487,
+                0.591
+            ],
+            "angle": 0,
+            "content": "where \\(r_i, a_i\\) denote the \\(i\\)-th sampled reasoning path and its corresponding answer, respectively."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.602,
+                0.447,
+                0.619
+            ],
+            "angle": 0,
+            "content": "3 Calibration using Self-Consistency"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.628,
+                0.489,
+                0.724
+            ],
+            "angle": 0,
+            "content": "After performing self-consistency on input \\(x\\) using \\(\\mathrm{LLM}_{\\theta}\\), we obtain a set of clusters \\(\\mathcal{C} = \\{c_1, \\dots, c_{|\\mathcal{C}|}\\}\\) with each cluster \\(c_i\\) comprising \\(n_i\\) sampled responses with the same answers. We design the following strategies, tailored to the characteristics of these clusters, to estimate the confidence of \\(\\mathrm{LLM}_{\\theta}\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.733,
+                0.489,
+                0.862
+            ],
+            "angle": 0,
+            "content": "Cluster Number We initially consider the Cluster Number \\(|\\mathcal{C}|\\). This is motivated by the finding of previous work (Wang et al., 2022; Xiong et al., 2023): LLMs tend to generate consistent answers when they are confident about their predictions, and thus the cluster number (number of distinct answers) tends to be small. We further divide the cluster number by the sample size \\(N\\) to normalize"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.882,
+                0.117
+            ],
+            "angle": 0,
+            "content": "the score into the range of [0, 1], before reversing it by “\\(1 - x\\)”:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.61,
+                0.129,
+                0.883,
+                0.16
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {F} _ {C N} (x, \\theta) = 1 - \\frac {| \\mathcal {C} |}{N}. \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.171,
+                0.884,
+                0.252
+            ],
+            "angle": 0,
+            "content": "Cluster Size In a similar vein, we adopt the Cluster Size: the number of samples (e.g., \\( n_i \\)) within a specific cluster (e.g., \\( c_i \\)). Again, we compute its proportion relative to the total sample size to normalize the score into the range [0, 1]:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.629,
+                0.263,
+                0.883,
+                0.29
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {F} _ {C S} (x, \\theta) = \\frac {n _ {i}}{N}. \\tag {3}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.301,
+                0.884,
+                0.382
+            ],
+            "angle": 0,
+            "content": "In contrast to the cluster number, the cluster size is more universally applicable across diverse prompts, as the cluster number can easily become ineffective when the output space of an LLM is restricted, such as when options for a question are provided."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.391,
+                0.884,
+                0.568
+            ],
+            "angle": 0,
+            "content": "Pairwise Comparison The Cluster Number and Cluster Size primarily consider the number of distinct answers and the number of sampled paths within a single cluster, respectively. They both overlook the information by comparing different clusters. For example, they may fail to consider the situation when the sizes of the top-ranked clusters are close. Consequently, we introduce the Pairwise Comparison method, which computes the winning rate of the chosen cluster \\((c_i)\\) against each of the remaining clusters:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.595,
+                0.581,
+                0.883,
+                0.628
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {F} _ {P C} (x, \\theta) = \\prod_ {j \\neq i} ^ {| \\mathcal {C} |} \\frac {n _ {i}}{n _ {i} + n _ {j}}, \\tag {4}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.64,
+                0.882,
+                0.688
+            ],
+            "angle": 0,
+            "content": "where \\( \\frac{n_i}{n_i + n_j} \\) represents the winning rate of selected cluster \\( c_i \\) against another cluster \\( c_j \\) by comparing the respective cluster sizes."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.701,
+                0.656,
+                0.717
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.728,
+                0.602,
+                0.743
+            ],
+            "angle": 0,
+            "content": "4.1 Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.749,
+                0.884,
+                0.877
+            ],
+            "angle": 0,
+            "content": "Datasets We conduct experiments on two popular math reasoning benchmarks of different type of questions, GSM8K (Cobbe et al., 2021) and MathQA (Amini et al., 2019). Particularly, GSM8K comprises 1,319 linguistically diverse grade school math word problems for testing. On the other hand, MathQA offers 2,985 multiple-choice math word problems for evaluation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.884,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Evaluation Metrics We adopt Brier Score and Expected Calibration Error (ECE) as evaluating"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.869,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>Here we follow common practice to adopt demonstrations with rationales for pretrained only models (e.g., Mistral-7B) and use \"Let's think step by step\" (Kojima et al., 2022) for instruction-tuned models (e.g., Mistral-7B-Inst)."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6024"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.148,
+                0.082,
+                0.851,
+                0.343
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\" colspan=\"2\"></td><td colspan=\"2\">Mistral-7B</td><td colspan=\"2\">Mistral-7B-Inst</td><td colspan=\"2\">Mixtral-8×7B</td><td colspan=\"2\">Mixtral-8×7B-Inst</td></tr><tr><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td></tr><tr><td rowspan=\"8\">GSM8K</td><td>logit w/ Path</td><td>0.394</td><td>0.399</td><td>0.414</td><td>0.414</td><td>0.178</td><td>0.265</td><td>0.233</td><td>0.252</td></tr><tr><td>logit w/ Answer</td><td>0.505</td><td>0.488</td><td>0.467</td><td>0.458</td><td>0.307</td><td>0.312</td><td>0.236</td><td>0.238</td></tr><tr><td>p(True)</td><td>0.127</td><td>0.267</td><td>0.406</td><td>0.407</td><td>0.070</td><td>0.201</td><td>0.195</td><td>0.198</td></tr><tr><td>Self-Consistency</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ FCN</td><td>0.092</td><td>0.186</td><td>0.125</td><td>0.182</td><td>0.136</td><td>0.157</td><td>0.075</td><td>0.092</td></tr><tr><td>w/ FCS</td><td>0.148</td><td>0.185</td><td>0.163</td><td>0.180</td><td>0.173</td><td>0.156</td><td>0.085</td><td>0.086</td></tr><tr><td>w/ FPC</td><td>0.248</td><td>0.229</td><td>0.253</td><td>0.226</td><td>0.238</td><td>0.194</td><td>0.110</td><td>0.096</td></tr><tr><td>w/ ALL</td><td>0.101</td><td>0.172</td><td>0.123</td><td>0.162</td><td>0.135</td><td>0.141</td><td>0.078</td><td>0.083</td></tr><tr><td rowspan=\"8\">MathQA</td><td>logit w/ Path</td><td>0.500</td><td>0.499</td><td>0.539</td><td>0.510</td><td>0.333</td><td>0.380</td><td>0.364</td><td>0.373</td></tr><tr><td>logit w/ Answer</td><td>0.356</td><td>0.362</td><td>0.291</td><td>0.319</td><td>0.266</td><td>0.290</td><td>0.220</td><td>0.281</td></tr><tr><td>p(True)</td><td>0.350</td><td>0.309</td><td>0.271</td><td>0.317</td><td>0.228</td><td>0.253</td><td>0.273</td><td>0.272</td></tr><tr><td>Self-Consistency</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ FCN</td><td>0.331</td><td>0.336</td><td>0.374</td><td>0.359</td><td>0.143</td><td>0.236</td><td>0.128</td><td>0.215</td></tr><tr><td>w/ FCS</td><td>0.091</td><td>0.225</td><td>0.114</td><td>0.227</td><td>0.080</td><td>0.190</td><td>0.035</td><td>0.171</td></tr><tr><td>w/ FPC</td><td>0.052</td><td>0.220</td><td>0.065</td><td>0.219</td><td>0.143</td><td>0.203</td><td>0.054</td><td>0.174</td></tr><tr><td>w/ ALL</td><td>0.144</td><td>0.238</td><td>0.172</td><td>0.242</td><td>0.089</td><td>0.190</td><td>0.072</td><td>0.177</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.352,
+                0.882,
+                0.381
+            ],
+            "angle": 0,
+            "content": "Table 1: Main test results on GSM8K and MathQA when using Mistral family models. Specifically, \\(*\\)-Inst indicates instruction-tuned models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.407,
+                0.489,
+                0.437
+            ],
+            "angle": 0,
+            "content": "metrics following common practice (Geng et al., 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.438,
+                0.487,
+                0.551
+            ],
+            "angle": 0,
+            "content": "Given instances \\((x_{1},y_{1}),\\dots,(x_{\\bar{N}},y_{\\bar{N}})\\) and their corresponding LLM predictions \\(\\hat{y}_1,\\dots,\\hat{y}_{\\bar{N}}\\) , ECE is computed by first binning the predictions into \\(M = 10\\) intervals based on their LLM confidence levels (e.g., \\(p(\\hat{y}_i)\\) ). For each bin (e.g. \\(B_{m}\\) ), it then calculates the accuracy (acc \\((B_{m})\\) ) and the average confidence \\((\\mathrm{conf}(B_m))\\) .."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.174,
+                0.556,
+                0.435,
+                0.594
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {a c c} \\left(B _ {m}\\right) = \\frac {1}{\\left| B _ {m} \\right|} \\sum_ {i \\in B _ {m}} \\mathbb {1} \\left(y _ {i} = \\hat {y} _ {i}\\right),\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.166,
+                0.596,
+                0.395,
+                0.634
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {c o n f} \\left(B _ {m}\\right) = \\frac {1}{\\left| B _ {m} \\right|} \\sum_ {i \\in B _ {m}} p \\left(\\hat {y} _ {i}\\right),\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.64,
+                0.489,
+                0.703
+            ],
+            "angle": 0,
+            "content": "where \\( |B_m| \\) is the number of samples in bin \\( B_m \\). Finally, the difference between accuracy and confidence is averaged across all bins to obtain the ECE score:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.134,
+                0.708,
+                0.488,
+                0.75
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathrm {E C E} = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{\\bar {N}} \\left| \\operatorname {a c c} \\left(B _ {m}\\right) - \\operatorname {c o n f} \\left(B _ {m}\\right) \\right| \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.756,
+                0.488,
+                0.786
+            ],
+            "angle": 0,
+            "content": "As another popular metric, Brier score is similar to ECE but conducted at the instance level:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.158,
+                0.791,
+                0.488,
+                0.834
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {B r i e r} = \\frac {1}{\\bar {N}} \\sum_ {i = 1} ^ {\\bar {N}} \\left(p \\left(\\hat {y} _ {i}\\right) - \\mathbb {1} \\left(y _ {i} = \\hat {y} _ {i}\\right)\\right) ^ {2}. \\tag {7}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.84,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Both metrics range from 0 to 1 with lower values indicating better calibration. We take Brier score as the main metric, as it is more robust to unbalanced distribution across bins (e.g. instances concentrate to one or two bins)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.407,
+                0.884,
+                0.502
+            ],
+            "angle": 0,
+            "content": "Settings We conduct experiments on LLaMA2 and Mistral-family models and investigate both pretrained or instruction-tuned variations. We use nucleus sampling to obtain \\(N = 16\\) samples by default for each instance and use temperatures of 0.6 / 1.0 for all pretrained / instruction-tuned models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.513,
+                0.884,
+                0.545
+            ],
+            "angle": 0,
+            "content": "Baselines We take the three representative baselines below for comparison:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.518,
+                0.557,
+                0.882,
+                0.605
+            ],
+            "angle": 0,
+            "content": "- logit w/ Path: It averages the probabilities of the tokens from the whole path to estimate the confidence of each prediction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.518,
+                0.617,
+                0.882,
+                0.665
+            ],
+            "angle": 0,
+            "content": "- logit w/ Answer: It is similar to logit w/ Path but only consider the tokens from the predicted answer span."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.518,
+                0.677,
+                0.885,
+                0.788
+            ],
+            "angle": 0,
+            "content": "- \\( p(\\text{True}) \\): It asks the LLM itself to classify its prediction as True or False. Then, it takes the predicted probability of True as its confidence. We follow Kadavath et al. (2022) to construct 8-shot demonstrations for prompting pretrained models but directly use instruction for instruction-tuned models."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.518,
+                0.557,
+                0.885,
+                0.788
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.802,
+                0.719,
+                0.818
+            ],
+            "angle": 0,
+            "content": "4.2 Results and Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Main Results Table 1 presents the main results obtained from both benchmarks using Mistral-family models. \\( p(\\text{True}) \\) performs best among the baselines, echoing the findings of Kadavath et al. (2022). However, due to its reliance on prompt design and in-context examples to aid the LLM"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6025"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.119,
+                0.081,
+                0.442,
+                0.263
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.276,
+                0.489,
+                0.307
+            ],
+            "angle": 0,
+            "content": "Figure 2: Calibration results on GSM8K when using Mixtral- \\(8\\times 7\\mathrm{B}\\) -Inst with different \\(N\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.331,
+                0.489,
+                0.362
+            ],
+            "angle": 0,
+            "content": "to classify its predictions, it can be challenging to construct effective demonstrations or instructions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.364,
+                0.49,
+                0.555
+            ],
+            "angle": 0,
+            "content": "In general, self-consistency-based methods surpass baselines in most cases regarding Brier and ECE, validating the efficacy of employing self-consistency features for estimating model confidence. We also note that baselines can occasionally yield impressive ECE scores (\\( p(\\text{True}) \\) on GSM8K with Mixtral- \\( 8 \\times 7\\mathrm{B} \\)). However, we observe that this is attributed to the concentration of most samples in just a few bins (e.g., Figure 1), leading to unreliable measurements. Nevertheless, our approaches still exhibit strong performance in terms of ECE scores across various settings."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.557,
+                0.49,
+                0.717
+            ],
+            "angle": 0,
+            "content": "Among the self-consistency-based methods, \\(\\mathcal{F}_{CN}\\) yields better ECE results on GSM8K, while \\(\\mathcal{F}_{CS}\\) achieves the highest Brier score. Conversely, for MathQA, \\(\\mathcal{F}_{CN}\\) performs significantly worse than the other two. This is because MathQA is a multi-choice task, and thus the cluster number of LLM answers is strictly limited by the provided choices. In conclusion, \\(\\mathcal{F}_{CS}\\) demonstrates greater generality across diverse settings, but \\(\\mathcal{F}_{CN}\\) and \\(\\mathcal{F}_{PC}\\) do offer improved estimation in certain cases."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.718,
+                0.49,
+                0.799
+            ],
+            "angle": 0,
+            "content": "Finally, we also explore averaging the proposed three metrics as a unified one, denoted as \"ALL\". Results show that it does not typically yield the best results. However, it serves as a robust choice, demonstrating competitive performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Influence of Sample Size \\(N\\) Previous research (Wang et al., 2022) has demonstrated that the sample size \\(N\\) can significantly affect the accuracy of self-consistency. When \\(N\\) increases, the model performance initially continues to improve before stabilizing once \\(N\\) reaches a sufficient level. Therefore, we take Mixtral-8×7B-Inst as a case study to"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.527,
+                0.083,
+                0.872,
+                0.262
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.272,
+                0.886,
+                0.343
+            ],
+            "angle": 0,
+            "content": "Figure 3: Performance and calibration results on GSM8K using different models below sorted by their performance: ① LLaMA2-7B-Chat, ② LLaMA2-13B-Chat, ③ Mistral-7B-Inst, ④ LLaMA2-70B-Chat, ⑤ Mixtral-8×7B-Inst."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.357,
+                0.811,
+                0.372
+            ],
+            "angle": 0,
+            "content": "examine the impact of \\(N\\) on calibration."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.373,
+                0.885,
+                0.485
+            ],
+            "angle": 0,
+            "content": "As illustrated in Figure 2, the Brier scores for all our methods initially decline and then remain constant as \\(N\\) grows. For \\(\\mathcal{F}_{CS}\\) and \\(\\mathcal{F}_{PC}\\), \\(N = 8\\) is adequate for accurate estimation. In contrast, \\(\\mathcal{F}_{CN}\\) requires a larger \\(N\\), indicating that the cluster number is more susceptible to the randomness of sampling."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.495,
+                0.886,
+                0.752
+            ],
+            "angle": 0,
+            "content": "Correlation between Performance and Calibration We finally explore the associations between model performance (Accuracy) and calibration. Figure 3 showcases the results on instruction-tuned LLaMA2 and Mistral series models, arranged in ascending order based on their performance. We generally observe a positively correlated trend between calibration (lower the better) and performance (higher the better) among the studied models. This observation indicates that more powerful models also exhibit enhanced calibration, echoing the findings of Kadavath et al. (2022). This phenomenon can be attributed to the fact that when a tested LLM is stronger, it is capable of generating more reasonable and consistent responses, leading to improved calibration."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.765,
+                0.642,
+                0.78
+            ],
+            "angle": 0,
+            "content": "5 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.791,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In this paper, we extend the widely-used inference strategy, self-consistency, to the field of calibration. Specifically, we develop three off-the-shelf calibration methods based on self-consistency for math reasoning tasks. Compared to conventional methods (p(True) and logit), our approaches yield significantly improved ECE and Brier scores on popular GSM8K and MathQA datasets. Future"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6026"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.198
+            ],
+            "angle": 0,
+            "content": "research directions include designing more effective calibration methods, leveraging richer features and employing more strategies (e.g., temperature scaling (Guo et al., 2017)) to enhance calibration performance. Our ultimate goal is to construct reliable and honest LLMs with the help of accurate confidence estimation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.21,
+                0.221,
+                0.225
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.237,
+                0.49,
+                0.43
+            ],
+            "angle": 0,
+            "content": "Our methods are founded on the principle of self-consistency, which relies on sampling multiple times for prediction. This approach, however, needs additional cost for inference, which may not be efficient and eco-friendly. Besides, our current work is limited to mathematical problems and does not explore other types of tasks, such as question-answering. Although it is crucial to extend our methods to encompass other tasks, this is non-trivial due to the inherent difficulty in dividing certain tasks' model predictions into distinct clusters."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.443,
+                0.266,
+                0.458
+            ],
+            "angle": 0,
+            "content": "Ethics Statement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.469,
+                0.49,
+                0.583
+            ],
+            "angle": 0,
+            "content": "We focus on ethical AI research and strive to achieve a balance between technological advancements and our ethical responsibilities. This work studies calibration, which aims to enhance the reliability of LLMs. Besides, we conduct experiments only on publicly available datasets, upholding privacy and anonymity rules."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.61,
+                0.214,
+                0.625
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.633,
+                0.49,
+                0.7
+            ],
+            "angle": 0,
+            "content": "Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.71,
+                0.49,
+                0.828
+            ],
+            "angle": 0,
+            "content": "Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. 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 2357-2367."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.84,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Yangyi Chen, Lifan Yuan, Ganqu Cui, Zhiyuan Liu, and Heng Ji. 2023. A close look into the calibration of pre-trained language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1343-1367."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.633,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.883,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.162,
+                0.885,
+                0.215
+            ],
+            "angle": 0,
+            "content": "Jiahui Geng, Fengyu Cai, Yuxia Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. 2023. A survey of language model confidence estimation and calibration. arXiv preprint arXiv:2311.08298."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.224,
+                0.885,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.286,
+                0.885,
+                0.365
+            ],
+            "angle": 0,
+            "content": "Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.375,
+                0.885,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Mark Hopkins, Cristian Petrescu-Prahova, Roie Levin, Ronan Le Bras, Alvaro Herrasti, and Vidur Joshi. 2017. Beyond sentential semantic parsing: Tackling the math sat with a cascade of tree transducers. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 795-804."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.476,
+                0.885,
+                0.543
+            ],
+            "angle": 0,
+            "content": "Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.551,
+                0.885,
+                0.618
+            ],
+            "angle": 0,
+            "content": "Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. 2024. Mixtral of experts. arXiv preprint arXiv:2401.04088."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.627,
+                0.885,
+                0.706
+            ],
+            "angle": 0,
+            "content": "Mandar Joshi, Eunsol Choi, Daniel S 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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.715,
+                0.883,
+                0.793
+            ],
+            "angle": 0,
+            "content": "Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.803,
+                0.885,
+                0.869
+            ],
+            "angle": 0,
+            "content": "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.878,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6027"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.133,
+                0.086,
+                0.489,
+                0.113
+            ],
+            "angle": 0,
+            "content": "The Eleventh International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.121,
+                0.49,
+                0.201
+            ],
+            "angle": 0,
+            "content": "Jierui Li, Lei Wang, Jipeng Zhang, Yan Wang, Bing Tian Dai, and Dongxiang Zhang. 2019. Modeling intra-relation in math word problems with different functional multi-head attentions. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 6162-6167."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.21,
+                0.49,
+                0.29
+            ],
+            "angle": 0,
+            "content": "Zhongli Li, Wenxuan Zhang, Chao Yan, Qingyu Zhou, Chao Li, Hongzhi Liu, and Yunbo Cao. 2022. Seeking patterns, not just memorizing procedures: Contrastive learning for solving math word problems. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 2486-2496."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.298,
+                0.49,
+                0.378
+            ],
+            "angle": 0,
+            "content": "Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. 2023. Teaching small language models to reason. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1773-1781."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.386,
+                0.49,
+                0.466
+            ],
+            "angle": 0,
+            "content": "Takuya Matsuzaki, Takumi Ito, Hidenao Iwane, Hirokazu Anai, and Noriko H Arai. 2017. Semantic parsing of pre-university math problems. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2131-2141."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.474,
+                0.489,
+                0.554
+            ],
+            "angle": 0,
+            "content": "Arkil Patel, Satwik Bhattachamishra, and Navin Goyal. 2021. Are nlp models really able to solve simple math word problems? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2080-2094."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.562,
+                0.49,
+                0.667
+            ],
+            "angle": 0,
+            "content": "Jinghui Qin, Xiaodan Liang, Yining Hong, Jianheng Tang, and Liang Lin. 2021. Neural-symbolic solver for math word problems with auxiliary tasks. 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 5870-5881."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.676,
+                0.49,
+                0.755
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.764,
+                0.49,
+                0.83
+            ],
+            "angle": 0,
+            "content": "Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2023. Distilling reasoning capabilities into smaller language models. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 7059-7073."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.84,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.175,
+                0.885,
+                0.241
+            ],
+            "angle": 0,
+            "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.25,
+                0.885,
+                0.343
+            ],
+            "angle": 0,
+            "content": "Qinzhuo Wu, Qi Zhang, Zhongyu Wei, and Xuan-Jing Huang. 2021. Math word problem solving with explicit numerical values. 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 5859-5869."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.352,
+                0.885,
+                0.419
+            ],
+            "angle": 0,
+            "content": "Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi. 2023. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. In The Twelfth International Conference on Learning Representations."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.419
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.431,
+                0.772,
+                0.448
+            ],
+            "angle": 0,
+            "content": "A Experiments on TriviaQA"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.457,
+                0.885,
+                0.648
+            ],
+            "angle": 0,
+            "content": "Although this work mainly focuses on math problems, we also verify the effectiveness of our approach on TriviaQA (Joshi et al., 2017), a popular question-answering benchmark. Different from math problems, the outputs of LLMs for this task sometimes lack clear answers for clustering. For instance, Mistral-7B-Inst provides a lengthy response, \"The musical Phantom of the Opera premiered in the US on 10th December 1993...\", instead of an answer span \"Phantom of the Opera\", to the question \"Which Lloyd Webber musical premiered in the US on 10th December 1993?\"."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.65,
+                0.885,
+                0.697
+            ],
+            "angle": 0,
+            "content": "To address this issue, we follow Kuhn et al. to employ natural language inference (NLI) to identify responses with the same answers<sup>2</sup>."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.698,
+                0.885,
+                0.762
+            ],
+            "angle": 0,
+            "content": "Results are shown in Table 2. Our approach significantly surpasses the baselines, particularly when utilizing instruction-tuned models, further demonstrating the effectiveness of our method."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.881,
+                0.883,
+                0.918
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)We adopt a publicly accessible model at https://huggingface.co/MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6028"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.186,
+                0.424,
+                0.812,
+                0.549
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td colspan=\"2\">Mistral-7B</td><td colspan=\"2\">Mistral-7B-Inst</td><td colspan=\"2\">Mixtral-8×7B</td><td colspan=\"2\">Mixtral-8×7B-Inst</td></tr><tr><td></td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td></tr><tr><td>logit</td><td>0.119</td><td>0.156</td><td>0.385</td><td>0.377</td><td>0.069</td><td>0.116</td><td>0.256</td><td>0.298</td></tr><tr><td>p(True)</td><td>0.077</td><td>0.162</td><td>0.326</td><td>0.318</td><td>0.099</td><td>0.131</td><td>0.314</td><td>0.316</td></tr><tr><td colspan=\"9\">Self-Consistency</td></tr><tr><td>w/ FCN</td><td>0.044</td><td>0.133</td><td>0.054</td><td>0.173</td><td>0.077</td><td>0.119</td><td>0.068</td><td>0.115</td></tr><tr><td>w/ FCS</td><td>0.043</td><td>0.132</td><td>0.072</td><td>0.186</td><td>0.039</td><td>0.107</td><td>0.039</td><td>0.116</td></tr><tr><td>w/ FPC</td><td>0.071</td><td>0.137</td><td>0.109</td><td>0.186</td><td>0.102</td><td>0.129</td><td>0.121</td><td>0.126</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.383,
+                0.559,
+                0.614,
+                0.573
+            ],
+            "angle": 0,
+            "content": "Table 2: Test results on TriviaQA."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6029"
+        }
+    ]
+]

2024/Self-Consistency Boosts Calibration for Math Reasoning/610e067d-96b2-49b7-8e9d-e89fb3c34d6e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efe49d49efc8fd6a51a2012f027db23ec5f53ee040e89563fc1068e177cb78dc
+size 288418

2024/Self-Consistency Boosts Calibration for Math Reasoning/full.md

@@ -0,0 +1,210 @@
+# Self-Consistency Boosts Calibration for Math Reasoning
+
+Ante Wang $^{1,2}$ , Linfeng Song $^{3*}$ , Ye Tian $^{3}$ , Baolin Peng $^{3}$ , Lifeng Jin $^{3}$ , Haitao Mi $^{3}$ , Jinsong Su $^{1,2*}$  and Dong Yu $^{3}$
+
+$^{1}$ School of Informatics, Xiamen University, China
+
+$^{2}$ Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural Heritage
+
+of Fujian and Taiwan (Xiamen University), Ministry of Culture and Tourism, China
+
+$^{3}$ Tencent AI Lab, Bellevue, WA
+
+wangante@stu.xmu.edu.cn, lfsong@global.tencent.com, jssu@xmu.edu.cn
+
+# Abstract
+
+Calibration, which establishes the correlation between accuracy and model confidence, is important for LLM development. We design three off-the-shelf calibration methods based on self-consistency (Wang et al., 2022) for math reasoning tasks. Evaluation on two popular benchmarks (GSM8K and MathQA) using strong open-source LLMs (Mistral and LLaMA2) shows that our methods better bridge model confidence and accuracy than existing methods based on  $p(\text{True})$  (Kadavath et al., 2022) or logit (Guo et al., 2017).
+
+# 1 Introduction
+
+Mathematical reasoning tasks (Cobbe et al., 2021; Hendrycks et al., 2021; Amini et al., 2019) involve mapping a question into a series of equations, which are then solved to obtain the final answer. Math reasoning has long been recognized challenging. Existing solutions propose to map input questions into equations via semantic parsing (Matsuzaki et al., 2017; Hopkins et al., 2017) or AST decoding (Li et al., 2019; Qin et al., 2021; Wu et al., 2021). Yet, the performance can degrade dramatically even with slight changes to the questions (Patel et al., 2021; Li et al., 2022).
+
+Recently, large language models (LLM, Achiam et al. 2023; Touvron et al. 2023; Jiang et al. 2024) have shown great potential for solving many math reasoning tasks, even though they are not specifically trained on these tasks. For instance, with chain-of-thought prompting (Wei et al., 2022) and self-consistency (Wang et al., 2022), open-source LLMs, such as Mixtral  $8 \times 7\mathrm{B}$  (Jiang et al., 2024), can reach an accuracy of around  $80\%$  on the GSM8K benchmark (Cobbe et al., 2021). On the other hand, conventional pretrained models (e.g., T5 (Raffel et al., 2020)) that are specifically finetuned on the GSM8K training set can only report
+
+![](images/c21cca1ce5f7f370f4252825d74038935451cef2b3e1f95e1400d3e24645e1b0.jpg)  
+Figure 1: Comparison of several calibration methods on Mistral-7B, where  $SC w / \mathcal{F}_{CN}$  is one of our methods based on self-consistency, which will be introduced in §3.
+
+accuracies around  $10\%$  to  $20\%$  (Shridhar et al., 2023; Magister et al., 2023).
+
+However, LLMs lack adequate calibration out of the box – the probabilities of model predictions are often poorly aligned with the actual accuracy (Xiong et al., 2023; Chen et al., 2023). Calibration is important for LLM development, as a well-calibrated LLM can precisely tell how likely its responses are correct or not. With such information, LLM developers can take multiple options to handle low-confidence responses, such as letting the LLM refuse to answer or keep resampling until a confident response is produced.
+
+In this work, we propose calibration methods based on self-consistency (Wang et al., 2022) for math reasoning tasks. Self-consistency performs clustering over multiple LLM samples before picking one from the largest cluster as the response to each input query. Here we consider several ways to estimate model confidence using the clustering results: cluster size that estimates how many samples agree with the selected one, cluster number that measures to what extent samples disagree with
+
+each other, and pairwise comparison that captures relative differences between pairs of clusters.
+
+We conduct experiments using strong open-source LLMs: Mistral (Jiang et al., 2023, 2024) and LLaMA2 (Touvron et al., 2023) series models with / without being aligned with instructions. Results on GSM8K (Cobbe et al., 2021) and MathQA (Amini et al., 2019) show that all our methods better calibrate these models than exiting popular methods, such as  $p(\text{True})$  (Kadavath et al., 2022) and logit (Guo et al., 2017) over the whole reasoning path or target answer span only.
+
+# 2 Preview: Self-Consistency with CoT Prompting
+
+For math reasoning, there are usually multiple trajectories to reach the final solution. To replicate this process, Wang et al. (2022) initially sample various reasoning paths  $r_1, \dots, r_N$  from the LLM given input  $x$  with Chain-of-Thought (CoT) prompting. Then, the answers  $a_1, \dots, a_N$  are extracted from the paths, and the most consistent answer (the one win by majority vote among the answers) is selected as the final answer  $\mathbf{a}$ :
+
+$$
+\mathbf {a} = \max  _ {\hat {a}} \sum_ {i = 1} ^ {N} \mathbb {1} \left(a _ {i} = \hat {a}\right), \tag {1}
+$$
+
+$$
+r _ {i}, a _ {i} \sim \mathrm {L L M} _ {\theta} (x),
+$$
+
+where  $r_i, a_i$  denote the  $i$ -th sampled reasoning path and its corresponding answer, respectively.
+
+# 3 Calibration using Self-Consistency
+
+After performing self-consistency on input  $x$  using  $\mathrm{LLM}_{\theta}$ , we obtain a set of clusters  $\mathcal{C} = \{c_1, \dots, c_{|\mathcal{C}|}\}$  with each cluster  $c_i$  comprising  $n_i$  sampled responses with the same answers. We design the following strategies, tailored to the characteristics of these clusters, to estimate the confidence of  $\mathrm{LLM}_{\theta}$ .
+
+Cluster Number We initially consider the Cluster Number  $|\mathcal{C}|$ . This is motivated by the finding of previous work (Wang et al., 2022; Xiong et al., 2023): LLMs tend to generate consistent answers when they are confident about their predictions, and thus the cluster number (number of distinct answers) tends to be small. We further divide the cluster number by the sample size  $N$  to normalize
+
+the score into the range of [0, 1], before reversing it by “ $1 - x$ ”:
+
+$$
+\mathcal {F} _ {C N} (x, \theta) = 1 - \frac {| \mathcal {C} |}{N}. \tag {2}
+$$
+
+Cluster Size In a similar vein, we adopt the Cluster Size: the number of samples (e.g.,  $n_i$ ) within a specific cluster (e.g.,  $c_i$ ). Again, we compute its proportion relative to the total sample size to normalize the score into the range [0, 1]:
+
+$$
+\mathcal {F} _ {C S} (x, \theta) = \frac {n _ {i}}{N}. \tag {3}
+$$
+
+In contrast to the cluster number, the cluster size is more universally applicable across diverse prompts, as the cluster number can easily become ineffective when the output space of an LLM is restricted, such as when options for a question are provided.
+
+Pairwise Comparison The Cluster Number and Cluster Size primarily consider the number of distinct answers and the number of sampled paths within a single cluster, respectively. They both overlook the information by comparing different clusters. For example, they may fail to consider the situation when the sizes of the top-ranked clusters are close. Consequently, we introduce the Pairwise Comparison method, which computes the winning rate of the chosen cluster  $(c_i)$  against each of the remaining clusters:
+
+$$
+\mathcal {F} _ {P C} (x, \theta) = \prod_ {j \neq i} ^ {| \mathcal {C} |} \frac {n _ {i}}{n _ {i} + n _ {j}}, \tag {4}
+$$
+
+where  $\frac{n_i}{n_i + n_j}$  represents the winning rate of selected cluster  $c_i$  against another cluster  $c_j$  by comparing the respective cluster sizes.
+
+# 4 Experiments
+
+# 4.1 Setup
+
+Datasets We conduct experiments on two popular math reasoning benchmarks of different type of questions, GSM8K (Cobbe et al., 2021) and MathQA (Amini et al., 2019). Particularly, GSM8K comprises 1,319 linguistically diverse grade school math word problems for testing. On the other hand, MathQA offers 2,985 multiple-choice math word problems for evaluation.
+
+Evaluation Metrics We adopt Brier Score and Expected Calibration Error (ECE) as evaluating
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="2">Mistral-7B</td><td colspan="2">Mistral-7B-Inst</td><td colspan="2">Mixtral-8×7B</td><td colspan="2">Mixtral-8×7B-Inst</td></tr><tr><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td></tr><tr><td rowspan="8">GSM8K</td><td>logit w/ Path</td><td>0.394</td><td>0.399</td><td>0.414</td><td>0.414</td><td>0.178</td><td>0.265</td><td>0.233</td><td>0.252</td></tr><tr><td>logit w/ Answer</td><td>0.505</td><td>0.488</td><td>0.467</td><td>0.458</td><td>0.307</td><td>0.312</td><td>0.236</td><td>0.238</td></tr><tr><td>p(True)</td><td>0.127</td><td>0.267</td><td>0.406</td><td>0.407</td><td>0.070</td><td>0.201</td><td>0.195</td><td>0.198</td></tr><tr><td>Self-Consistency</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ FCN</td><td>0.092</td><td>0.186</td><td>0.125</td><td>0.182</td><td>0.136</td><td>0.157</td><td>0.075</td><td>0.092</td></tr><tr><td>w/ FCS</td><td>0.148</td><td>0.185</td><td>0.163</td><td>0.180</td><td>0.173</td><td>0.156</td><td>0.085</td><td>0.086</td></tr><tr><td>w/ FPC</td><td>0.248</td><td>0.229</td><td>0.253</td><td>0.226</td><td>0.238</td><td>0.194</td><td>0.110</td><td>0.096</td></tr><tr><td>w/ ALL</td><td>0.101</td><td>0.172</td><td>0.123</td><td>0.162</td><td>0.135</td><td>0.141</td><td>0.078</td><td>0.083</td></tr><tr><td rowspan="8">MathQA</td><td>logit w/ Path</td><td>0.500</td><td>0.499</td><td>0.539</td><td>0.510</td><td>0.333</td><td>0.380</td><td>0.364</td><td>0.373</td></tr><tr><td>logit w/ Answer</td><td>0.356</td><td>0.362</td><td>0.291</td><td>0.319</td><td>0.266</td><td>0.290</td><td>0.220</td><td>0.281</td></tr><tr><td>p(True)</td><td>0.350</td><td>0.309</td><td>0.271</td><td>0.317</td><td>0.228</td><td>0.253</td><td>0.273</td><td>0.272</td></tr><tr><td>Self-Consistency</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>w/ FCN</td><td>0.331</td><td>0.336</td><td>0.374</td><td>0.359</td><td>0.143</td><td>0.236</td><td>0.128</td><td>0.215</td></tr><tr><td>w/ FCS</td><td>0.091</td><td>0.225</td><td>0.114</td><td>0.227</td><td>0.080</td><td>0.190</td><td>0.035</td><td>0.171</td></tr><tr><td>w/ FPC</td><td>0.052</td><td>0.220</td><td>0.065</td><td>0.219</td><td>0.143</td><td>0.203</td><td>0.054</td><td>0.174</td></tr><tr><td>w/ ALL</td><td>0.144</td><td>0.238</td><td>0.172</td><td>0.242</td><td>0.089</td><td>0.190</td><td>0.072</td><td>0.177</td></tr></table>
+
+Table 1: Main test results on GSM8K and MathQA when using Mistral family models. Specifically,  $*$ -Inst indicates instruction-tuned models.
+
+metrics following common practice (Geng et al., 2023).
+
+Given instances  $(x_{1},y_{1}),\dots,(x_{\bar{N}},y_{\bar{N}})$  and their corresponding LLM predictions  $\hat{y}_1,\dots,\hat{y}_{\bar{N}}$  , ECE is computed by first binning the predictions into  $M = 10$  intervals based on their LLM confidence levels (e.g.,  $p(\hat{y}_i)$  ). For each bin (e.g.  $B_{m}$  ), it then calculates the accuracy (acc  $(B_{m})$  ) and the average confidence  $(\mathrm{conf}(B_m))$  ..
+
+$$
+\operatorname {a c c} \left(B _ {m}\right) = \frac {1}{\left| B _ {m} \right|} \sum_ {i \in B _ {m}} \mathbb {1} \left(y _ {i} = \hat {y} _ {i}\right),
+$$
+
+$$
+\operatorname {c o n f} \left(B _ {m}\right) = \frac {1}{\left| B _ {m} \right|} \sum_ {i \in B _ {m}} p \left(\hat {y} _ {i}\right),
+$$
+
+where  $|B_m|$  is the number of samples in bin  $B_m$ . Finally, the difference between accuracy and confidence is averaged across all bins to obtain the ECE score:
+
+$$
+\mathrm {E C E} = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{\bar {N}} \left| \operatorname {a c c} \left(B _ {m}\right) - \operatorname {c o n f} \left(B _ {m}\right) \right| \tag {6}
+$$
+
+As another popular metric, Brier score is similar to ECE but conducted at the instance level:
+
+$$
+\operatorname {B r i e r} = \frac {1}{\bar {N}} \sum_ {i = 1} ^ {\bar {N}} \left(p \left(\hat {y} _ {i}\right) - \mathbb {1} \left(y _ {i} = \hat {y} _ {i}\right)\right) ^ {2}. \tag {7}
+$$
+
+Both metrics range from 0 to 1 with lower values indicating better calibration. We take Brier score as the main metric, as it is more robust to unbalanced distribution across bins (e.g. instances concentrate to one or two bins).
+
+Settings We conduct experiments on LLaMA2 and Mistral-family models and investigate both pretrained or instruction-tuned variations. We use nucleus sampling to obtain  $N = 16$  samples by default for each instance and use temperatures of 0.6 / 1.0 for all pretrained / instruction-tuned models.
+
+Baselines We take the three representative baselines below for comparison:
+
+- logit w/ Path: It averages the probabilities of the tokens from the whole path to estimate the confidence of each prediction.  
+- logit w/ Answer: It is similar to logit w/ Path but only consider the tokens from the predicted answer span.  
+-  $p(\text{True})$ : It asks the LLM itself to classify its prediction as True or False. Then, it takes the predicted probability of True as its confidence. We follow Kadavath et al. (2022) to construct 8-shot demonstrations for prompting pretrained models but directly use instruction for instruction-tuned models.
+
+# 4.2 Results and Analysis
+
+Main Results Table 1 presents the main results obtained from both benchmarks using Mistral-family models.  $p(\text{True})$  performs best among the baselines, echoing the findings of Kadavath et al. (2022). However, due to its reliance on prompt design and in-context examples to aid the LLM
+
+![](images/a1e06ad14dcc56e9e946cbc40d777400b77ac7d6596c476f5f54e1e3485ea829.jpg)  
+Figure 2: Calibration results on GSM8K when using Mixtral-  $8\times 7\mathrm{B}$  -Inst with different  $N$
+
+to classify its predictions, it can be challenging to construct effective demonstrations or instructions.
+
+In general, self-consistency-based methods surpass baselines in most cases regarding Brier and ECE, validating the efficacy of employing self-consistency features for estimating model confidence. We also note that baselines can occasionally yield impressive ECE scores ( $p(\text{True})$  on GSM8K with Mixtral-  $8 \times 7\mathrm{B}$ ). However, we observe that this is attributed to the concentration of most samples in just a few bins (e.g., Figure 1), leading to unreliable measurements. Nevertheless, our approaches still exhibit strong performance in terms of ECE scores across various settings.
+
+Among the self-consistency-based methods,  $\mathcal{F}_{CN}$  yields better ECE results on GSM8K, while  $\mathcal{F}_{CS}$  achieves the highest Brier score. Conversely, for MathQA,  $\mathcal{F}_{CN}$  performs significantly worse than the other two. This is because MathQA is a multi-choice task, and thus the cluster number of LLM answers is strictly limited by the provided choices. In conclusion,  $\mathcal{F}_{CS}$  demonstrates greater generality across diverse settings, but  $\mathcal{F}_{CN}$  and  $\mathcal{F}_{PC}$  do offer improved estimation in certain cases.
+
+Finally, we also explore averaging the proposed three metrics as a unified one, denoted as "ALL". Results show that it does not typically yield the best results. However, it serves as a robust choice, demonstrating competitive performance.
+
+Influence of Sample Size  $N$  Previous research (Wang et al., 2022) has demonstrated that the sample size  $N$  can significantly affect the accuracy of self-consistency. When  $N$  increases, the model performance initially continues to improve before stabilizing once  $N$  reaches a sufficient level. Therefore, we take Mixtral-8×7B-Inst as a case study to
+
+![](images/1cd8a97db0979f26313a691b83b5d230c757a396671c64c7a6ab36eefa059a52.jpg)  
+Figure 3: Performance and calibration results on GSM8K using different models below sorted by their performance: ① LLaMA2-7B-Chat, ② LLaMA2-13B-Chat, ③ Mistral-7B-Inst, ④ LLaMA2-70B-Chat, ⑤ Mixtral-8×7B-Inst.
+
+examine the impact of  $N$  on calibration.
+
+As illustrated in Figure 2, the Brier scores for all our methods initially decline and then remain constant as  $N$  grows. For  $\mathcal{F}_{CS}$  and  $\mathcal{F}_{PC}$ ,  $N = 8$  is adequate for accurate estimation. In contrast,  $\mathcal{F}_{CN}$  requires a larger  $N$ , indicating that the cluster number is more susceptible to the randomness of sampling.
+
+Correlation between Performance and Calibration We finally explore the associations between model performance (Accuracy) and calibration. Figure 3 showcases the results on instruction-tuned LLaMA2 and Mistral series models, arranged in ascending order based on their performance. We generally observe a positively correlated trend between calibration (lower the better) and performance (higher the better) among the studied models. This observation indicates that more powerful models also exhibit enhanced calibration, echoing the findings of Kadavath et al. (2022). This phenomenon can be attributed to the fact that when a tested LLM is stronger, it is capable of generating more reasonable and consistent responses, leading to improved calibration.
+
+# 5 Conclusion
+
+In this paper, we extend the widely-used inference strategy, self-consistency, to the field of calibration. Specifically, we develop three off-the-shelf calibration methods based on self-consistency for math reasoning tasks. Compared to conventional methods (p(True) and logit), our approaches yield significantly improved ECE and Brier scores on popular GSM8K and MathQA datasets. Future
+
+research directions include designing more effective calibration methods, leveraging richer features and employing more strategies (e.g., temperature scaling (Guo et al., 2017)) to enhance calibration performance. Our ultimate goal is to construct reliable and honest LLMs with the help of accurate confidence estimation.
+
+# Limitations
+
+Our methods are founded on the principle of self-consistency, which relies on sampling multiple times for prediction. This approach, however, needs additional cost for inference, which may not be efficient and eco-friendly. Besides, our current work is limited to mathematical problems and does not explore other types of tasks, such as question-answering. Although it is crucial to extend our methods to encompass other tasks, this is non-trivial due to the inherent difficulty in dividing certain tasks' model predictions into distinct clusters.
+
+# Ethics Statement
+
+We focus on ethical AI research and strive to achieve a balance between technological advancements and our ethical responsibilities. This work studies calibration, which aims to enhance the reliability of LLMs. Besides, we conduct experiments only on publicly available datasets, upholding privacy and anonymity rules.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. 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 2357-2367.  
+Yangyi Chen, Lifan Yuan, Ganqu Cui, Zhiyuan Liu, and Heng Ji. 2023. A close look into the calibration of pre-trained language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1343-1367.
+
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.  
+Jiahui Geng, Fengyu Cai, Yuxia Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. 2023. A survey of language model confidence estimation and calibration. arXiv preprint arXiv:2311.08298.  
+Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR.  
+Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).  
+Mark Hopkins, Cristian Petrescu-Prahova, Roie Levin, Ronan Le Bras, Alvaro Herrasti, and Vidur Joshi. 2017. Beyond sentential semantic parsing: Tackling the math sat with a cascade of tree transducers. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 795-804.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. 2024. Mixtral of experts. arXiv preprint arXiv:2401.04088.  
+Mandar Joshi, Eunsol Choi, Daniel S 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.  
+Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.  
+Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In
+
+The Eleventh International Conference on Learning Representations.  
+Jierui Li, Lei Wang, Jipeng Zhang, Yan Wang, Bing Tian Dai, and Dongxiang Zhang. 2019. Modeling intra-relation in math word problems with different functional multi-head attentions. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 6162-6167.  
+Zhongli Li, Wenxuan Zhang, Chao Yan, Qingyu Zhou, Chao Li, Hongzhi Liu, and Yunbo Cao. 2022. Seeking patterns, not just memorizing procedures: Contrastive learning for solving math word problems. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 2486-2496.  
+Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. 2023. Teaching small language models to reason. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1773-1781.  
+Takuya Matsuzaki, Takumi Ito, Hidenao Iwane, Hirokazu Anai, and Noriko H Arai. 2017. Semantic parsing of pre-university math problems. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2131-2141.  
+Arkil Patel, Satwik Bhattachamishra, and Navin Goyal. 2021. Are nlp models really able to solve simple math word problems? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2080-2094.  
+Jinghui Qin, Xiaodan Liang, Yining Hong, Jianheng Tang, and Liang Lin. 2021. Neural-symbolic solver for math word problems with auxiliary tasks. 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 5870-5881.  
+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.  
+Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2023. Distilling reasoning capabilities into smaller language models. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 7059-7073.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.
+
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Qinzhuo Wu, Qi Zhang, Zhongyu Wei, and Xuan-Jing Huang. 2021. Math word problem solving with explicit numerical values. 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 5859-5869.  
+Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi. 2023. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. In The Twelfth International Conference on Learning Representations.
+
+# A Experiments on TriviaQA
+
+Although this work mainly focuses on math problems, we also verify the effectiveness of our approach on TriviaQA (Joshi et al., 2017), a popular question-answering benchmark. Different from math problems, the outputs of LLMs for this task sometimes lack clear answers for clustering. For instance, Mistral-7B-Inst provides a lengthy response, "The musical Phantom of the Opera premiered in the US on 10th December 1993...", instead of an answer span "Phantom of the Opera", to the question "Which Lloyd Webber musical premiered in the US on 10th December 1993?".
+
+To address this issue, we follow Kuhn et al. to employ natural language inference (NLI) to identify responses with the same answers<sup>2</sup>.
+
+Results are shown in Table 2. Our approach significantly surpasses the baselines, particularly when utilizing instruction-tuned models, further demonstrating the effectiveness of our method.
+
+<table><tr><td></td><td colspan="2">Mistral-7B</td><td colspan="2">Mistral-7B-Inst</td><td colspan="2">Mixtral-8×7B</td><td colspan="2">Mixtral-8×7B-Inst</td></tr><tr><td></td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td><td>ECE ↓</td><td>Brier ↓</td></tr><tr><td>logit</td><td>0.119</td><td>0.156</td><td>0.385</td><td>0.377</td><td>0.069</td><td>0.116</td><td>0.256</td><td>0.298</td></tr><tr><td>p(True)</td><td>0.077</td><td>0.162</td><td>0.326</td><td>0.318</td><td>0.099</td><td>0.131</td><td>0.314</td><td>0.316</td></tr><tr><td colspan="9">Self-Consistency</td></tr><tr><td>w/ FCN</td><td>0.044</td><td>0.133</td><td>0.054</td><td>0.173</td><td>0.077</td><td>0.119</td><td>0.068</td><td>0.115</td></tr><tr><td>w/ FCS</td><td>0.043</td><td>0.132</td><td>0.072</td><td>0.186</td><td>0.039</td><td>0.107</td><td>0.039</td><td>0.116</td></tr><tr><td>w/ FPC</td><td>0.071</td><td>0.137</td><td>0.109</td><td>0.186</td><td>0.102</td><td>0.129</td><td>0.121</td><td>0.126</td></tr></table>
+
+Table 2: Test results on TriviaQA.

2024/Self-Consistency Boosts Calibration for Math Reasoning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78a4d8c262060d5275cc6199e3347fd4394a40021b20ef34f38900701daf6918
+size 249996

2024/Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement/45f4817a-44a7-499d-b884-1bdac2c1b3e5_content_list.json

@@ -0,0 +1,1618 @@
+[
+    {
+        "type": "text",
+        "text": "Self-Constructed Context Decomposition with Fined-grained Alignment Enhancement",
+        "text_level": 1,
+        "bbox": [
+            127,
+            89,
+            870,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yunlong Feng, Dechuan Teng, Yang Xu, Honglin Mu, Xiao Xu, Libo Qin, Qingfu Zhu*, Wanxiang Che",
+        "bbox": [
+            270,
+            152,
+            732,
+            186
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Harbin Institute of Technology, China",
+        "bbox": [
+            344,
+            186,
+            653,
+            202
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{ylfeng,dcteng,yxu,hlmu,xxu,lbqin,qfzhu,car}  $@$  ir.hit.edu.cn",
+        "bbox": [
+            255,
+            203,
+            742,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Decompilation transforms compiled code back into a high-level programming language for analysis when source code is unavailable. Previous work has primarily focused on enhancing decomposition performance by increasing the scale of model parameters or training data for pre-training. Based on the characteristics of the decomposition task, we propose two methods: (1) Without fine-tuning, the Self-Constructed Context Decomposition  $(\\mathrm{sc}^2\\mathrm{dec})$  method recombines the LLM's decomposition results to construct pairs for in-context learning, helping the model improve decomposition performance. (2) Fine-grained Alignment Enhancement (FAE), which meticulously aligns assembly code with source code at the statement level by leveraging debugging information, is employed during the fine-tuning phase to achieve further improvements in decomposition. By integrating these two methods, we achieved a Re-Executability performance improvement of approximately  $3.90\\%$  on the Decompile-Eval benchmark, establishing a new state-of-the-art performance of  $52.41\\%$ . The code, data, and models are available at https://github.com/AlongWY/sccdec.",
+        "bbox": [
+            141,
+            286,
+            460,
+            656
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            665,
+            260,
+            682
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Decompilation is the process of converting compiled machine code or bytecode back into a high-level programming language. This process is typically used to analyze how the software works when the source code is not accessible (Brumley et al., 2013; Katz et al., 2018; Hosseini and Dolan-Gavitt, 2022; Xu et al., 2023; Armengol-Estape et al., 2023; Jiang et al., 2023; Wong et al., 2023). Many tools have been developed for decompilation, like Ghidra (Ghidra, 2024) and IDA Pro (Hex-Rays, 2024). However, these tools often struggle to generate human-readable code. The main challenge in decompilation is that it's hard to fully reconstruct",
+        "bbox": [
+            112,
+            690,
+            490,
+            901
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/345d99328aa3b056e3a95f9fb636e4ed6b7204e8a9c7f04701ac4c4bd7362e29.jpg",
+        "image_caption": [
+            "Figure 1: Pipeline of the decompilation. The input for decompilation tasks is typically assembly, with the source code invisible. Additionally, the source code obtained through decompilation usually does not exactly match the original code."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            564,
+            256,
+            826,
+            450
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the source code, especially details such as variable names (Lacomis et al., 2019) and primary structures (Wei et al., 2007), which are frequently lost during the compilation process.",
+        "bbox": [
+            507,
+            563,
+            884,
+            627
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in large language models (LLMs) have prompted researchers to view programming languages as distinct linguistic systems, utilizing pre-trained code LLMs to accomplish various coding tasks (Lippincott, 2020; Rozière et al., 2023; Guo et al., 2024; OpenAI, 2023). These models have demonstrated significant performance improvements over traditional techniques (Zeng et al., 2022; Xu et al., 2022), making it feasible to apply LLMs to decompilation challenges. For instance, Transformer-based models such as Slade (Armengol-Estapé et al., 2023) and Nova (Jiang et al., 2023) have shown potential in using language models to translate binary code back into more readable and structured source code. Recently, Tan et al. (2024) developed and released the first open-source LLM focused on decompilation, alongside the construction of the first",
+        "bbox": [
+            507,
+            631,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding Author",
+        "bbox": [
+            136,
+            906,
+            284,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6603",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6603-6614",
+        "bbox": [
+            216,
+            945,
+            779,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            273,
+            958,
+            724,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "decompilation benchmark to evaluate the decompilation capabilities of models. These efforts primarily treat the decompilation task as a translation task, training a robust decompilation model on extensive synthetic data (Hosseini and Dolan-Gavitt, 2022; Armengol-Estapé et al., 2023; Jiang et al., 2023; Tan et al., 2024).",
+        "bbox": [
+            112,
+            84,
+            487,
+            195
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "By analyzing the characteristics of decomposition tasks, we have found two features:",
+        "bbox": [
+            112,
+            197,
+            489,
+            228
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- The decompiled code generated by the model is usually compileable.",
+            "- The compiler can produce external debugging information for the debugger."
+        ],
+        "bbox": [
+            136,
+            241,
+            485,
+            305
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Based on the characteristics of the decomposition task, we propose two approaches: (1) Self-Constructed Context Decomposition  $(\\mathrm{sc}^2\\mathrm{dec})$ , a tuning-free approach, leverages the decomposition results from the model and specific compiler to derive ground-truth (assembly code, source code) pair closely related to the test sample. By incorporating the obtained pair as a demonstration example within the context, we can fully leverage the model's in-context learning capabilities to enhance performance. (2) Fine-grained Alignment Enhancement (FAE) introduces a new step-by-step decomposition training objective in addition to the end-to-end decomposition training objective that relies on large-scale pre-training for implicit alignment. By leveraging debugging information designed for debuggers, we can align assembly code with high-level code at a finer statement level. This allows the model to more faithfully preserve the original functionality after decomposition.",
+        "bbox": [
+            112,
+            316,
+            489,
+            638
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Combining these two methods, we have achieved an approximately  $3.90\\%$  improvement in re-executability performance based on lvm4decompile model (Tan et al., 2024), reaching a new state-of-the-art performance of  $52.41\\%$  re-executability.",
+        "bbox": [
+            112,
+            639,
+            489,
+            718
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, our contributions are as follows:",
+        "bbox": [
+            132,
+            720,
+            487,
+            734
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce Self-Constructed Context Decomposition, which leverages the compileability of decomposition results to construct better contexts.",
+            "- We introduce Fine-grained Alignment Enhancement, which fine-tunes the model using fine-grained alignment data between assembly code and source code derived from debug information. We also propose how to synthesize its training data automatically."
+        ],
+        "bbox": [
+            136,
+            749,
+            487,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- The experiments on the Decompile-Eval benchmark show that our proposed model achieves new state-of-the-art results.",
+        "bbox": [
+            531,
+            84,
+            880,
+            131
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Works",
+        "text_level": 1,
+        "bbox": [
+            509,
+            146,
+            672,
+            161
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Decompilation has a long history, centered on converting binary files into human-readable source code. Traditional tools such as Hex-Rays IDA Pro (Hex-Rays, 2024) and Ghidra (Ghidra, 2024) predominantly rely on analyzing a program's control flow graph (CFG) to perform this task (Brumley et al., 2013). These tools operate by scrutinizing instructions within assembly code to construct CFGs and identify common programming structures like if-else statements and loops (Wei et al., 2007). However, they rely on intricate and error-prone expert-built rule systems that frequently fail with optimized binary code, a common practice in commercial software development. Moreover, traditional tools often generate outputs that closely resemble assembly code representations, translating variables into registers and utilizing low-level operations such as goto statements. This output is not only challenging to comprehend but also occasionally cannot be recompiled (Liu and Wang, 2020).",
+        "bbox": [
+            507,
+            173,
+            884,
+            508
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Inspired by neural machine translation, researchers have begun to conceptualize decompilation as a translation task, wherein machine-level instructions are converted into readable source code using neural networks. Initial endeavors employing recurrent neural networks (RNNs) demonstrated limited success (Katz et al., 2018). However, recent advancements in natural language processing (NLP) have facilitated the application of pre-trained language models (LMs) in decompilation. Notable examples include BTC (Hosseini and Dolan-Gavitt, 2022), Slade (Armengol-Estapé et al., 2023), and Nova (Jiang et al., 2023), which mark significant strides in the field. Recently, Tan et al. (2024) created and released the first open-source large language models specifically for decompilation and established an evaluation benchmark that considers the re-compilability and re-executability of decompiled code.",
+        "bbox": [
+            507,
+            511,
+            882,
+            816
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            509,
+            829,
+            611,
+            845
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the Self-Constructed Context Decomposition method and Fine-grained Alignment Enhancement fine-tuning. Figures 2 and 3 illustrate the methods.",
+        "bbox": [
+            507,
+            857,
+            880,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6604",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/4de304b57cea5f28244a505f652396ec27592f53b88121741ae9476006cb6dbc.jpg",
+        "image_caption": [
+            "Figure 2: The pipeline for Self-Constructed Context Decomposition operates as follows: when the LLM decomposes and generates compilable code, we compile this code to construct the context, and then use it to decompile the initial assembly code again."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            487,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Self-Constructed Context Decomposition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            461,
+            478,
+            476
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the Self-Constructed Context Decomposition (sc²dec) method designed to enhance the performance of decomposition models. The method recombines the decomposed code generated by the model to construct the context. The sc²dec method involves the following key steps, as the Figure 2 shows:",
+        "bbox": [
+            112,
+            482,
+            487,
+            594
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Decompile: Initially, the model decomposes the assembly code back to the target programming language. We call it the initial decomposition result.",
+            "2. Compile and Disassembly: Then we recompile the initial decomposition result to obtain the corresponding assembly code if the generated code is compileable.",
+            "3. Constructed Context: The self-constructed context is then formed by concatenating the generated C code with its corresponding assembly code.",
+            "4. Decompile Again: Decompile the assembly code with the self-constructed context again."
+        ],
+        "bbox": [
+            127,
+            605,
+            487,
+            829
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this context, the assembly code precedes the C code, resembling an example in the context. The final input to the model is a concatenation of the self-constructed context (with assembly code first and C code second) and the original assembly code. This",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/8175bcafcd20bbbaf885b8b49ea7224062f0c5761bcc4fa9a6763924b75ed0e1.jpg",
+        "image_caption": [
+            "Figure 3: An example of step-by-step decomposition is presented, where the training objective requires the model to generate C code progressively after each assembly block. Fine-tuning the model with this objective aids in learning the fine-grained correspondences between assembly and C code."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            531,
+            83,
+            862,
+            420
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "method leverages the additional context provided by the generated C code and its corresponding assembly code to improve the overall decompilation performance.",
+        "bbox": [
+            507,
+            539,
+            882,
+            602
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Fine-grained Alignment Enhancement",
+        "text_level": 1,
+        "bbox": [
+            507,
+            615,
+            855,
+            630
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To enhance the model's ability to perceive the correspondence between assembly code and its corresponding source code, we propose a Fine-grained Alignment Enhancement (FAE) method, which constructs fine-grained alignment data between assembly code and source code derived from debug information. The method primarily includes two training objectives: the End-to-end Decomposition Objective and the Step-by-Step Decomposition Objective.",
+        "bbox": [
+            505,
+            636,
+            884,
+            796
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "- End-to-end Decomposition Objective: In the end-to-end decomposition task, the model takes assembly code as input and directly generates the corresponding decomposed C code. The goal of this task is to enable the model to translate from assembly language to C language without any intermediate steps. This ob",
+        "bbox": [
+            531,
+            808,
+            884,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6605",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "jective aims to prevent catastrophic forgetting in the model. During training, we increase the context length and ensure that the same sample is not truncated, which enhances the model's ability to align assembly and C code.",
+        "bbox": [
+            147,
+            84,
+            487,
+            164
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "- Step-by-step Decomposition Objective: As the Figure 3 shows, in the step-by-step decomposition process, an intermediate form containing both the assembly code and the corresponding C code is first generated, followed by the generation of the complete C code. This step-by-step approach allows the model to better understand the correspondence between assembly code and C code. Training the model with the step-by-step decomposition objective can enhance the model's ability to finely align assembly code with C code. It is important to note that this form is used only during training and not during inference.",
+        "bbox": [
+            136,
+            166,
+            489,
+            390
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "During the training process, we employed two different training objectives to enhance the model's decompile performance: an end-to-end decompile objective and a step-by-step decompile objective. They all take assembly code as input, but their output formats are different.",
+        "bbox": [
+            112,
+            401,
+            489,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Data Processing",
+        "text_level": 1,
+        "bbox": [
+            112,
+            506,
+            302,
+            521
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "By following the steps below, we can synthesize some fine-grained alignment training data. The synthesized data, as shown in Figure 3, is in a format of interleaved assembly and code output. We expect the model to learn finer-grained alignment between assembly and code from such data.",
+        "bbox": [
+            112,
+            524,
+            489,
+            621
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "1. Compilation: We employed the \"gcc\" compiler to compile the selected function code into binary shared libraries at various optimization levels (including O0, O1, O2, and O3). The debug option was enabled during compilation to ensure the inclusion of DWARF debugging information in the compiled output. This debugging information is crucial for subsequent disassembly and parsing steps.",
+        "bbox": [
+            129,
+            631,
+            489,
+            791
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2. Disassembly: Subsequently, we used the \"objdump\" tool to disassemble the compiled binary shared libraries. The \"-S\" option was utilized to include source code in the disassembly output. Additionally, we used the \"-source/comment=;,\" option toptides each line of source code with a \";,\" symbol, facilitating easier parsing in later stages.",
+        "bbox": [
+            127,
+            793,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "3. Data Parsing: The disassembly process yielded a mixed output containing both assembly code and source code, organized in an alternating sequence of source code followed by assembly code. We parsed the output to extract the corresponding assembly and source code for each function, ensuring accurate alignment between the two.",
+            "4. Data Reorganization: Finally, we reorganized the parsed content into a format where the assembly code precedes the corresponding source code in an alternating sequence."
+        ],
+        "bbox": [
+            522,
+            84,
+            884,
+            277
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Through these steps, we constructed a high-quality training dataset that ensures each sample contains comprehensive source code information along with the corresponding assembly code. This dataset will be used to train our model, aiming to enhance its performance in tasks such as code understanding and compilation optimization.",
+        "bbox": [
+            507,
+            288,
+            884,
+            401
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            413,
+            655,
+            430
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            438,
+            640,
+            451
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we will introduce the benchmark used in our evaluation process, Decompile-Eval (Tan et al., 2024), which is specifically designed to assess the decompilation capabilities of large language models.",
+        "bbox": [
+            507,
+            458,
+            882,
+            539
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The Decompile-Eval benchmark is adapted from the HumanEval benchmark, which includes 164 problems initially designed for code generation tasks. These problems were translated into the C programming language, and the corresponding assembly code was generated at four optimization levels (O0, O1, O2, and O3). The correctness of the decomposition results was tested using the test cases from HumanEval. The primary metrics of the Decompile-Eval benchmark are as follows:",
+        "bbox": [
+            507,
+            539,
+            882,
+            700
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Re-compilability: This metric evaluates whether the decompiled code produced by our model can be successfully recompiled into executable binary without errors. A high recompiliability rate indicates that the decompiled code is syntactically correct and adheres to the constraints of the target language (in this case, C).",
+            "- Re-executability: This metric assesses the functional correctness of the decompiled code. Specifically, it measures whether the recom-piled binaries produce the expected outputs when executed. The correctness of the output"
+        ],
+        "bbox": [
+            531,
+            712,
+            885,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6606",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "is determined using the testing methodology provided by the HumanEval dataset, ensuring a comprehensive evaluation of the logical accuracy of the decompiled code.",
+        "bbox": [
+            149,
+            84,
+            489,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In our experiment, we pay more attention to the Re-executability, as it more accurately reflects the overall decompilation capability of the model.",
+        "bbox": [
+            112,
+            166,
+            487,
+            214
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Training Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            229,
+            287,
+            243
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Implementation We utilize Lora (Hu et al., 2022) to fine-tune the llm4decompile-6.7b obtained on Hugging Face (Wolf et al., 2019), with rank set to 32, alpha set to 64 and target set to all projection layers  $^{1}$ . The optimizer is AdamW (Loshchilov and Hutter, 2019), with a learning rate of 5e-5. The maximum sequence length is set to 16384, and the learning rate scheduler type is cosine, with a warm-up period of 20 steps. The training process was conducted for one epoch. LlamaFactory (Zheng et al., 2024) and FlashAttention 2 (Dao, 2024) were used for the fine-tuning of the model. All experiments were done with an A100-SXM4-80GB. We use greedy decoding for all experiments.",
+        "bbox": [
+            112,
+            250,
+            489,
+            476
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training data We selected 10,000 samples from the train_real_compilable subset of Exebench (Armengol-Estapé et al., 2022) to synthesis the training data. The selected functions exclusively utilize the standard C library and do not include additional data structures. The training data were synthesized with gcc 11.4 provided by Ubuntu 22.04.",
+        "bbox": [
+            112,
+            489,
+            489,
+            602
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            617,
+            233,
+            631
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We present several baselines in our experiments and demonstrate the effectiveness of our proposed methods.",
+        "bbox": [
+            112,
+            640,
+            487,
+            686
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3.1 Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            701,
+            231,
+            714
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- GPT 4: One of the most powerful OpenAI models, known for its advanced language understanding and generation capabilities (OpenAI, 2023).",
+            "- deepseek-chat: A powerful conversational AI model from DeepSeek, excelling in generating coherent and contextually relevant dialogues (DeepSeek-AI, 2024).",
+            "- llm4decompile-6.7b: The 6.7B open-access decomposition LLM pre-trained on 15 billion"
+        ],
+        "bbox": [
+            136,
+            721,
+            487,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "bool func0(int num){ if (num <= 1) { return false; } for (int i = 2; i * i <= num; i++) { if (num % i == 0) { return false; } } return true; }",
+        "guess_lang": "txt",
+        "bbox": [
+            517,
+            84,
+            855,
+            225
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: The example for 1-shot learning. In this example, we have tried to cover common control logic such as if statements, loops, and early returns as much as possible. The example will be compiled with the same optimization level as the target assembly code.",
+        "bbox": [
+            507,
+            239,
+            882,
+            312
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "tokens of C source code and the corresponding assembly code (Tan et al., 2024) $^2$ .",
+        "bbox": [
+            544,
+            337,
+            880,
+            368
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- llm4decompile-6.7b+FAE: The new model was obtained by applying the Fine-grained Alignment Enhancement method for further fine-tuning based on the llm4decompile-6.7b $^2$ .",
+        "bbox": [
+            531,
+            369,
+            882,
+            447
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3.2 Methods",
+        "text_level": 1,
+        "bbox": [
+            509,
+            462,
+            638,
+            476
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- vanilla: Using specific prompts for different models to directly request decomposition.",
+        "bbox": [
+            531,
+            480,
+            880,
+            512
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [
+            "Prompt For GPT-4 and Deepseek Chat"
+        ],
+        "code_body": "What is the C source code of  $\\leftrightarrow$  the assembly code below: ..asm <asm>",
+        "guess_lang": "txt",
+        "bbox": [
+            571,
+            552,
+            835,
+            626
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [
+            "Prompt For 1lm4decompile"
+        ],
+        "code_body": "This is the assembly code: <asm>",
+        "guess_lang": "txt",
+        "bbox": [
+            571,
+            684,
+            831,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "What is the source code?",
+        "guess_lang": "txt",
+        "bbox": [
+            573,
+            733,
+            815,
+            747
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "- 1-shot: We write a fixed example as the context (Brown et al., 2020). The example will be compiled with the same optimization level as the target assembly code. The example shown in Figure 4. In this example, we have tried to cover common control logic such as if statements, loops, and early returns as much as possible.",
+        "bbox": [
+            531,
+            769,
+            884,
+            898
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ q_Proj, k_Proj, v_Proj, o_Proj, gate_Proj, up_Proj, down_Proj",
+        "bbox": [
+            112,
+            894,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "2We use the latest version, llm4decompile-6.7b-v1.5.",
+        "bbox": [
+            529,
+            906,
+            853,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6607",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/db4f41bd954fee6f85a4d2fe724dcbccdbcb189ea311055320729ca84949537d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td colspan=\"11\">GPT-4 (OpenAI, 2023)</td></tr><tr><td>vanilla</td><td>76.83</td><td>67.68</td><td>67.07</td><td>57.93</td><td>67.38</td><td>20.73</td><td>13.41</td><td>13.41</td><td>10.98</td><td>14.63</td></tr><tr><td>+sc2dec</td><td>76.22</td><td>67.03</td><td>65.24</td><td>56.10</td><td>66.15</td><td>32.93</td><td>20.12</td><td>19.51</td><td>13.41</td><td>21.49</td></tr><tr><td>1-shot</td><td>76.22</td><td>78.05</td><td>73.78</td><td>70.73</td><td>74.70</td><td>31.71</td><td>20.12</td><td>19.51</td><td>21.34</td><td>23.17</td></tr><tr><td>+sc2dec</td><td>74.39</td><td>76.22</td><td>71.95</td><td>68.90</td><td>72.87</td><td>40.24</td><td>25.61</td><td>23.78</td><td>23.17</td><td>28.20</td></tr><tr><td colspan=\"11\">Deepseek Chat (DeepSeek-AI, 2024)</td></tr><tr><td>vanilla</td><td>62.20</td><td>51.83</td><td>47.56</td><td>45.73</td><td>51.83</td><td>10.37</td><td>5.49</td><td>4.88</td><td>5.49</td><td>6.55</td></tr><tr><td>+sc2dec</td><td>57.93</td><td>50.00</td><td>46.95</td><td>45.12</td><td>50.00</td><td>13.41</td><td>6.71</td><td>6.71</td><td>7.32</td><td>8.54</td></tr><tr><td>1-shot</td><td>67.68</td><td>70.73</td><td>64.63</td><td>56.71</td><td>64.94</td><td>13.41</td><td>10.37</td><td>7.93</td><td>7.32</td><td>9.76</td></tr><tr><td>+sc2dec</td><td>65.85</td><td>70.12</td><td>62.20</td><td>53.66</td><td>62.96</td><td>17.68</td><td>11.59</td><td>10.37</td><td>9.15</td><td>12.20</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b (Tan et al., 2024)</td></tr><tr><td>vanilla</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>92.68</td><td>92.80</td><td>90.12</td><td>93.41</td><td>92.26</td><td>71.34</td><td>47.56</td><td>45.98</td><td>41.71</td><td>51.65</td></tr><tr><td>1-shot</td><td>93.05</td><td>93.29</td><td>90.98</td><td>93.90</td><td>92.80</td><td>70.37</td><td>44.88</td><td>41.46</td><td>37.68</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>93.05</td><td>92.68</td><td>90.37</td><td>93.29</td><td>92.35</td><td>71.59</td><td>47.20</td><td>46.34</td><td>41.46</td><td>51.65</td></tr><tr><td>retrieval</td><td>94.02</td><td>88.05</td><td>84.51</td><td>85.12</td><td>87.93</td><td>65.73</td><td>32.68</td><td>38.66</td><td>36.22</td><td>43.32</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b + FAE</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            884,
+            501
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: The main results of different methods across four optimization levels (O0, O1, O2, O3), as well as their average scores (AVG). The results in bold represent the optimal performance, while those underlined indicate the second-best performance. More results can be found in the Table 5.",
+        "bbox": [
+            112,
+            510,
+            882,
+            555
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- retrieval: We use the retrieval method (BM25 (Lù, 2024)) to get the context. The context is generated by searching for the most similar assembly code in the training set.",
+            "-  $\\mathbf{sc}^2\\mathbf{dec}$ : The self-constructed context decomposition we described in the Section 3.1.",
+            "- 1-shot+sc²dec: We apply the sc²dec on the 1-shot method. Note that the 1-shot example can only be applied to generate the initial decomposition result in Figure 2."
+        ],
+        "bbox": [
+            136,
+            579,
+            487,
+            739
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Results",
+        "text_level": 1,
+        "bbox": [
+            114,
+            763,
+            218,
+            778
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we present and analyze the performance of different models. Our experiments aim to evaluate the impact of various techniques, including one-shot, self-constructed context decomposition, and fine-tuning, on the model's performance. The main results are shown in Table 1.",
+        "bbox": [
+            112,
+            784,
+            489,
+            879
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Performance of the Base Model Through a comparative analysis of the experimental results in Ta",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ble 1, it is evident that different models exhibit significant differences in re-executability across various optimization levels (O0, O1, O2, O3). The base model, llm4decompile-6.7b, performs best at the O0 optimization level with a score of  $70.24\\%$ , but its performance gradually declines at other optimization levels, with an average score of  $48.60\\%$ . Among models not fine-tuned with decompilation tasks, GPT-4 achieved the best performance, with an average re-executability score of  $14.64\\%$ . Generally, their results on re-compilability metrics are significantly higher than their performance on re-executability metrics.",
+        "bbox": [
+            507,
+            579,
+            884,
+            788
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Context Sensitivity By comparing vanilla, 1-shot, and retrieval methods based on the llm4decompile-6.7b, it is observed that the model's performance can even decline after providing a sample. In contrast, the dynamic samples generated through our self-constructed context method yield an improvement in executability of over  $2\\%$",
+        "bbox": [
+            507,
+            808,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6608",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/527b3b2f11358683d0d7922e49bab92e4ee46534fe28d49c02c545b902b4c657.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>sc2dec (O0)</td><td>92.68</td><td>92.20</td><td>91.71</td><td>92.32</td><td>92.23</td><td>70.24</td><td>48.90</td><td>47.68</td><td>42.93</td><td>51.83</td></tr><tr><td>sc2dec (O1)</td><td>92.68</td><td>91.59</td><td>93.54</td><td>92.93</td><td>92.68</td><td>68.41</td><td>48.54</td><td>47.44</td><td>43.05</td><td>51.86</td></tr><tr><td>sc2dec (O2)</td><td>92.68</td><td>92.20</td><td>91.71</td><td>92.32</td><td>92.23</td><td>67.80</td><td>48.90</td><td>47.68</td><td>42.93</td><td>51.68</td></tr><tr><td>sc2dec (O3)</td><td>92.68</td><td>91.59</td><td>92.93</td><td>92.32</td><td>92.38</td><td>67.80</td><td>47.07</td><td>47.80</td><td>43.41</td><td>51.52</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            884,
+            227
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: The results of our method across optimization levels. For cases such as \"sc²dec (O0)\", we assume that the optimization level of asm code is entirely unknown and select a specific optimization level to compile the code to construct the context.",
+        "bbox": [
+            112,
+            235,
+            882,
+            280
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "int func0(int a){ for (int i = 0; i * i * i <= abs(a); i++) if (i * i * i == abs(a)) return 1; return 0; } Source code   \nint func0(int n){ int i; if  $(n <   0)$  n  $=$  -n; for  $(\\mathrm{i} = 1;\\mathrm{i}\\ast \\mathrm{i}\\ast \\mathrm{i} <   = \\mathrm{n};\\mathrm{i} + + )$  if  $(i*\\bar{i}*i==n)$  return 1; return 0; } Decompile   \nint func0(int n){ int i; if  $(n <   0)$  n  $=$  -n; if  $(n = = 0)$  return 1; for  $(\\mathrm{i} = 1;\\mathrm{i}\\ast \\mathrm{i}\\ast \\mathrm{i} <   = \\mathrm{n};\\mathrm{i} + + )$  if  $(i*\\bar{i}*i==n)$  return 1; return 0; } SC2Dec",
+        "guess_lang": "c",
+        "bbox": [
+            124,
+            302,
+            480,
+            620
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 5: A case study for  $\\mathrm{sc^2dec}$ , which is based on the llm4decompile-6.7b with FAE. By applying the  $\\mathrm{sc^2dec}$  method, the model detects and fixes the mismatch between the code and assembly.",
+        "bbox": [
+            112,
+            633,
+            489,
+            692
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "before and after fine-tuning. This indicates that inappropriate samples may even have a detrimental effect in the decompilation domain, while our method achieves a nearly stable performance enhancement without the need for retrieval.",
+        "bbox": [
+            112,
+            717,
+            489,
+            797
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "1-shot is a Strong Baseline As shown in Table 1, utilizing the 1-shot method significantly enhances the model's performance. Specifically, for GPT-4, the model's re-executability performance increased from  $14.64\\%$  to  $23.17\\%$ , an approximate improvement of  $7\\%$ . In the open-source model Deepseek Chat, performance improved from  $6.55\\%$  to  $9.76\\%$ ,",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "reflecting a  $3.21\\%$  enhancement. However, in the case of the llm4decompile model, which had already been trained on decompilation tasks, there was no performance improvement. Moreover, in the llm4decompile model fine-tuned with Fine-grained Alignment Enhancement, performance actually decreased from  $50.88\\%$  to  $50.61\\%$ , a reduction of approximately  $0.27\\%$ .",
+        "bbox": [
+            507,
+            304,
+            884,
+            434
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sc2dec Brings Further Performance Improvement The Table 1 shows that applying sc2dec independently results in performance improvements of  $6.86\\%$  and  $1.99\\%$  on GPT-4 and Deepseek Chat, respectively. These improvements are lower than those achieved by the 1-shot method, likely due to the model's low Re-Compilability metric. In other words, a substantial portion of the code obtained through decompilation is un-compilable, which impedes the application of the sc2dec method. Furthermore, when combined with the 1-shot method, sc2dec can further enhance performance, achieving an additional  $5\\%$  improvement for GPT-4, bringing it to  $28.20\\%$ , and a  $2.44\\%$  improvement for Deepseek Chat, bringing it to  $12.20\\%$ . In the case of llm4decompile-6.7b, due to the models' relatively high Re-Compilability metric, the sc2dec method significantly outperformed the 1-shot method, increasing performance from  $48.60\\%$  to  $51.65\\%$  and from  $50.88\\%$  to  $52.41\\%$ .",
+        "bbox": [
+            507,
+            444,
+            884,
+            765
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Best Performance with Combined Methods As the Table 1 shows, by combining our Step-by-Step Decompile tasks and context-based decomposition methods, we ultimately achieved approximately a  $3.90\\%$  performance improvement, surpassing the performance of these methods when applied independently. This demonstrates that fine-tuning and self-construct context methods are orthogonal. Most notably, the combined method of fined",
+        "bbox": [
+            507,
+            776,
+            885,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "6609",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/4b5b10e67ed9721d865d3c3fbbba995924642533246e0500de8ebc93b1762263.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>w/o tune</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td colspan=\"11\">FAE (End-to-End &amp; Step-by-Step Decomposition Objective)</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr><tr><td colspan=\"11\">End-to-End Decomposition Objective</td></tr><tr><td>vanilla</td><td>93.9</td><td>90.61</td><td>90.61</td><td>91.34</td><td>91.62</td><td>69.51</td><td>43.29</td><td>42.2</td><td>40.73</td><td>48.93</td></tr><tr><td>+sc2dec</td><td>92.8</td><td>90.24</td><td>90.00</td><td>90.12</td><td>90.79</td><td>71.34</td><td>46.10</td><td>43.78</td><td>42.44</td><td>50.91</td></tr><tr><td>1-shot</td><td>93.9</td><td>90.73</td><td>90.37</td><td>91.22</td><td>91.55</td><td>69.51</td><td>44.02</td><td>42.68</td><td>40.61</td><td>49.21</td></tr><tr><td>+sc2dec</td><td>92.8</td><td>90.24</td><td>89.76</td><td>90.00</td><td>90.7</td><td>71.22</td><td>46.46</td><td>44.27</td><td>42.56</td><td>51.13</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            884,
+            323
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3: The results of ablation study. All results are based on the llm4decompile-6.7b model.",
+        "bbox": [
+            179,
+            331,
+            815,
+            346
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "grained alignment enhancement and self-construct context decomposition achieves the highest scores across all optimization levels expect \"O0\", with scores of  $48.54\\%$ ,  $47.56\\%$ , and  $43.29\\%$  at the O1, O2, and O3 levels, respectively, and an overall average score of  $52.41\\%$ .",
+        "bbox": [
+            112,
+            370,
+            489,
+            467
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            479,
+            228,
+            495
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we will further analyze the experimental results to study the effectiveness and sensitivity of our method:",
+        "bbox": [
+            112,
+            500,
+            489,
+            548
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Mismatched Optimization Levels Leads to Performance Degradation As the Table 2 shows, the performance can degrade by  $1\\%$  to  $2\\%$  overall when constructing contexts that mismatch the optimization levels of the target assembly code. However, specific optimization levels may exhibit different behaviors. For instance, using a fixed O3 optimization level to construct contexts while decompiling assembly optimized with O2 does not result in significant performance degradation and even better. In practical scenarios, using O3 to construct contexts for binaries with unknown optimization levels might be a reasonable choice.",
+        "bbox": [
+            112,
+            558,
+            489,
+            766
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Case study of  $\\mathbf{sc}^2\\mathbf{dec}$  As the Figure 5 shows, the compiler has expanded the abs function, reducing the overhead of function calls. The model initially generated a piece of compilable but incorrect code during direct decompilation, omitting the case where the variable  $n$  equals 0. We compiled and disassembled this code to construct a new sample. By applying the  $\\mathrm{sc^2dec}$  method with the context, the model successfully identified the missing part",
+        "bbox": [
+            112,
+            776,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "in the short segment of assembly code and generated the correct code, thus rectifying the error from the previous inference.",
+        "bbox": [
+            507,
+            370,
+            884,
+            419
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Contribution of FAE The results in Table 1 show that the application of Fine-grained Alignment Enhancement (FAE) for further fine-tuning the model significantly enhanced its re-executability performance across all optimization levels except \"O0\". Notably, at optimization levels O2 and O3, the scores improved by  $5.12\\%$  and  $4.27\\%$ , reaching  $45.73\\%$  and  $42.32\\%$ , respectively. The average performance increased by  $2.28\\%$ , achieving  $50.88\\%$ . This demonstrates the effectiveness of the Fine-grained Alignment Enhancement method.",
+        "bbox": [
+            507,
+            429,
+            885,
+            606
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Ablation Study of FAE The results of ablation experiments in Table 3 have demonstrated the effectiveness of our proposed Step-by-Step Decompile training objective. Removing the training objective results in approximately a  $2\\%$  decline in Re-Executability performance, regardless of whether the  $\\mathrm{sc}^2$  dec method is used, and notably, a  $3.78\\%$  decline at the \"O2\" optimization level. However, the performance of the 1-shot remains unaffected, supporting our hypothesis that the mismatch between the training data form and the inference form leads to performance degradation. Furthermore, retaining only the End-to-end Decompilation training objective still yields little performance improvement, about  $0.33\\%$ . This improvement can be attributed to the longer context provided during training, as the samples are not truncated during training, allowing the model to fully observe the assembly code and its corresponding C code.",
+        "bbox": [
+            507,
+            615,
+            885,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "6610",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            247,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce two methods based on the characteristics of the decompilation task: Self-Constructed Context Decomposition, which leverages the compileability of decomposition results to construct better contexts, and Fine-grained Alignment Enhancement, which fine-tunes the model using fine-grained alignment data between assembly code and source code derived from debug information. By integrating these methods, we achieved a  $3.90\\%$  improvement in decompilation performance, reaching a new state-of-the-art level of  $52.41\\%$ .",
+        "bbox": [
+            112,
+            108,
+            490,
+            285
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledge",
+        "text_level": 1,
+        "bbox": [
+            114,
+            296,
+            235,
+            313
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We gratefully acknowledge the support of the National Natural Science Foundation of China (NSFC) via grant 62236004, 62206078, 62441603 and 62476073.",
+        "bbox": [
+            112,
+            319,
+            489,
+            384
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            395,
+            218,
+            410
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The  $\\mathrm{sc^2dec}$  depends on the decomposed code generated by the model. If the decomposed code cannot be recompiled,  $\\mathrm{sc^2dec}$  will not be able to benefit from it, necessitating that the decomposed code has high re-compilability performance. For Fine-grained Alignment Enhancement, we created a Step-by-Step Decompile training objective, training with only 10,000 samples across four optimization levels (O0-O3), without validating its performance on larger datasets or bigger models. Moreover, fine-tuning led to a further decline in performance in the 1-shot scenario.",
+        "bbox": [
+            112,
+            419,
+            489,
+            612
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Potential risks",
+        "text_level": 1,
+        "bbox": [
+            114,
+            624,
+            243,
+            639
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Decompilation is a technique that converts compiled binary code back into human-readable source code. While decompilation can be legal and useful in certain contexts, it also entails several potential risks. The primary potential risks are as follows:",
+        "bbox": [
+            112,
+            648,
+            489,
+            728
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Intellectual Property Infringement: Decompilation may violate the copyright and licensing agreements of software. Unauthorized decomposition can lead to copyright infringement, thereby instigating legal disputes.",
+            "- Security Risks: Decomposed code may expose the internal structure and implementation details of the software, providing attack vectors for hackers. Malicious actors can exploit this information to identify and exploit vulnerabilities in the software."
+        ],
+        "bbox": [
+            136,
+            734,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "- Ethical Concerns: Decompiling and analyzing another's code can be considered unethical, especially when done without authorization.",
+        "bbox": [
+            531,
+            84,
+            882,
+            133
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            173,
+            608,
+            187
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Jordi Armengol-Estapé, Jackson Woodruff, Alexander Brauckmann, José Wesley de Souza Magalhães, and Michael F. P. O’Boyle. 2022. Exebench: An ml-scale dataset of executable c functions. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, MAPS 2022, page 50–59, New York, NY, USA. Association for Computing Machinery.",
+            "Jordi Armengol-Estape, Jackson Woodruff, Chris Cummins, and Michael F. P. O'Boyle. 2023. Slade: A portable small language model decoder for optimized assembler. CoRR, abs/2305.12520.",
+            "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. Preprint, arXiv:2005.14165.",
+            "David Brumley, JongHyup Lee, Edward J. Schwartz, and Maverick Woo. 2013. Native x86 decomposition using semantics-preserving structural analysis and iterative control-flow structuring. In Proceedings of the 22th USENIX Security Symposium, Washington, DC, USA, August 14-16, 2013, pages 353-368. USENIX Association.",
+            "Tri Dao. 2024. FlashAttention-2: Faster attention with better parallelism and work partitioning. In International Conference on Learning Representations (ICLR).",
+            "DeepSeek-AI. 2024. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. Preprint, arXiv:2405.04434.",
+            "Ghidra. 2024. Ghidra software reverse engineering framework.",
+            "Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. 2024. Deepseek-coder: When the large language model meets programming – the rise of code intelligence. Preprint, arXiv:2401.14196.",
+            "Hex-Rays. 2024. Ida pro: a cross-platform multiprocessor disassembler and debugger."
+        ],
+        "bbox": [
+            509,
+            197,
+            884,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "6611",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Iman Hosseini and Brendan Dolan-Gavitt. 2022. Beyond the C: retargetable decompilation using neural machine translation. CoRR, abs/2212.08950.",
+            "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.",
+            "Nan Jiang, Chengxiao Wang, Kevin Liu, Xiangzhe Xu, Lin Tan, and Xiangyu Zhang. 2023. Nova+: Generative language models for binaries. Preprint, arXiv:2311.13721.",
+            "Deborah S. Katz, Jason Ruchti, and Eric M. Schulte. 2018. Using recurrent neural networks for decompilation. In 25th International Conference on Software Analysis, Evolution and Reengineering, SANER 2018, Campobasso, Italy, March 20-23, 2018, pages 346-356. IEEE Computer Society.",
+            "Jeremy Lacomis, Pengcheng Yin, Edward J. Schwartz, Miltiadis Allamanis, Claire Le Goues, Graham Neubig, and Bogdan Vasilescu. 2019. DIRE: A neural approach to decompiled identifier naming. In 34th IEEE/ACM International Conference on Automated Software Engineering, ASE 2019, San Diego, CA, USA, November 11-15, 2019, pages 628-639. IEEE.",
+            "Thomas Lippincott. 2020. Starcoder: A general neural ensemble technique to support traditional scholarship, illustrated with a study of the post-atlantic slave trade. In 15th Annual International Conference of the Alliance of Digital Humanities Organizations, DH 2020, Ottawa, Canada, July 20-25, 2020, Conference Abstracts.",
+            "Zhibo Liu and Shuai Wang. 2020. How far we have come: testing decomposition correctness of C decompliers. In ISSTA '20: 29th ACM SIGSOFT International Symposium on Software Testing and Analysis, Virtual Event, USA, July 18-22, 2020, pages 475-487. ACM.",
+            "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.",
+            "Xing Han Lu. 2024. Bm25s: Orders of magnitude faster lexical search via eager sparse scoring. Preprint, arXiv:2407.03618.",
+            "OpenAI. 2022. ChatGPT.",
+            "OpenAI. 2023. GPT-4 Technical Report. CoRR, abs/2303.08774.",
+            "Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton-Ferrer, Aaron Grattaflori, Wenhan Xiong, Alexandre Défossez, Jade Copet,"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. 2023. Code llama: Open foundation models for code. CoRR, abs/2308.12950.",
+            "Hanzhuo Tan, Qi Luo, Jing Li, and Yuqun Zhang. 2024. Llm4decompile: Decompiling binary code with large language models. Preprint, arXiv:2403.05286.",
+            "Tao Wei, Jian Mao, Wei Zou, and Yu Chen. 2007. A new algorithm for identifying loops in decomposition. In Static Analysis, 14th International Symposium, SAS 2007, Kongens Lyngby, Denmark, August 22-24, 2007, Proceedings, volume 4634 of Lecture Notes in Computer Science, pages 170-183. Springer.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. CoRR, abs/1910.03771.",
+            "Wai Kin Wong, Huaijin Wang, Zongjie Li, Zhibo Liu, Shuai Wang, Qiyi Tang, Sen Nie, and Shi Wu. 2023. Refining decomposed C code with large language models. CoRR, abs/2310.06530.",
+            "Frank F. Xu, Uri Alon, Graham Neubig, and Vincent Josua Hellendoorn. 2022. A systematic evaluation of large language models of code. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, MAPS 2022, page 1-10, New York, NY, USA. Association for Computing Machinery.",
+            "Xiangzhe Xu, Zhuo Zhang, Shiwei Feng, Yapeng Ye, Zian Su, Nan Jiang, Siyuan Cheng, Lin Tan, and Xiangyu Zhang. 2023. Lmpa: Improving decomposition by synergy of large language model and program analysis. CoRR, abs/2306.02546.",
+            "Zhengran Zeng, Hanzhuo Tan, Haotian Zhang, Jing Li, Yuqun Zhang, and Lingming Zhang. 2022. An extensive study on pre-trained models for program understanding and generation. In ISSTA '22: 31st ACM SIGSOFT International Symposium on Software Testing and Analysis, Virtual Event, South Korea, July 18 - 22, 2022, pages 39-51. ACM.",
+            "Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. arXiv preprint arXiv:2403.13372."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            769
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "6612",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Training examples",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            314,
+            99
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [
+            "Figure 6: The training example of end-to-end decompile objective."
+        ],
+        "code_body": "User: # This is the assembly: <func0>:\nmov %edi,%eax\nand $0x1,%eax\nret\n# What is the source code?\nAssistant: bool func0(int num) {\nreturn num % 2 != 0;\n}",
+        "guess_lang": "txt",
+        "bbox": [
+            124,
+            124,
+            436,
+            258
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [
+            "Figure 7: The training example of step-by-step decompile objective. Note that the workflow is only for training."
+        ],
+        "code_body": "User: Add inline comments that contain the corresponding C code for the assembly code first, then write the equivalent C function. ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ``  \n`` `` `````````````````````````````````````````````````````````````````````````````````````````````````````  \nAssistant: Here is the assembly code with inline comments containing the corresponding C code:  \n``` ` ` ` ` ` ` ` \n```\n<func0>:\n  mov %edi,%eax\n  and $0x1,%eax\n  ret\n```\nAssistant: Here is the assembly code with inline comments containing the corresponding C code:  \n``` ` ` ` ` ` \n```\n<func0>:\n  ; bool func0(int num) {\n    mov %edi,%eax\n    and $0x1,%eax\n  ; return num % 2 != 0;\n  ret\n}\n```\nAnd here is the equivalent C function:  \n``` c\nbool func0(int num) {\n  return num % 2 != 0;\n}",
+        "guess_lang": "txt",
+        "bbox": [
+            127,
+            318,
+            813,
+            659
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Extra Results",
+        "text_level": 1,
+        "bbox": [
+            114,
+            725,
+            270,
+            739
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The Model is More Sensitive to Compiler Series than to Versions: As the Table 4 shows, using a compiler different from the one used during training to build the context can lead to significant performance degradation. The model performs better when the context is built using the GCC compiler series, whereas the performance is slightly worse when using the clang series. This may be due to the fact that the context used during both the continued pre-training and fine-tuning stages was constructed with the GCC compiler series. When different versions of GCC are used for construction, the performance of  $\\mathrm{sc^2dec}$  stabilizes around  $52\\%$  for the fine-tuned model except \"x86-64 gcc 14.1\". In contrast, when the context is built using the clang series, the performance of  $\\mathrm{sc^2dec}$  is slightly lower and more consistent, stabilizing around  $51\\%$  for the fine-tuned model.",
+        "bbox": [
+            112,
+            750,
+            884,
+            895
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "6613",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/46c9a3409371a01f5163b2ae94267d8afabc99fdb82aa140bf2cdb8700af3a7b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Compiler</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>x86-64 gcc 10.5</td><td>92.07</td><td>90.98</td><td>91.71</td><td>92.32</td><td>91.77</td><td>70.85</td><td>47.32</td><td>47.80</td><td>43.05</td><td>52.26</td></tr><tr><td>x86-64 gcc 11.4</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>48.66</td><td>47.68</td><td>42.68</td><td>52.47</td></tr><tr><td>x86-64 gcc 12.3</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>47.68</td><td>46.95</td><td>42.56</td><td>52.01</td></tr><tr><td>x86-64 gcc 13.3</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>47.93</td><td>46.59</td><td>42.68</td><td>52.01</td></tr><tr><td>x86-64 gcc 14.1</td><td>92.07</td><td>91.83</td><td>92.44</td><td>92.56</td><td>92.23</td><td>70.85</td><td>47.07</td><td>45.73</td><td>41.95</td><td>51.40</td></tr><tr><td>x86-64 clang 14.0.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.05</td><td>45.73</td><td>42.44</td><td>51.31</td></tr><tr><td>x86-64 clang 15.0.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.29</td><td>45.73</td><td>42.32</td><td>51.34</td></tr><tr><td>x86-64 clang 16.0.0</td><td>92.68</td><td>92.2</td><td>93.66</td><td>93.17</td><td>92.93</td><td>69.02</td><td>48.05</td><td>46.46</td><td>42.32</td><td>51.46</td></tr><tr><td>x86-64 clang 17.0.1</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.05</td><td>46.71</td><td>42.32</td><td>51.52</td></tr><tr><td>x86-64 clang 18.1.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.17</td><td>46.46</td><td>42.32</td><td>51.49</td></tr></table>",
+        "bbox": [
+            154,
+            80,
+            845,
+            272
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/279de1401224e693e4acabb1b71d9d10f130e70be8505c561e24a51021dd0aea.jpg",
+        "table_caption": [
+            "Table 4: The re-executability of differnet models across different compilers. We selected the popular three versions of both the clang and gcc series in Compiler Explorer. The compiler of Compiler Explorer may differ from the compilers provided by Ubuntu in some default compilation options, which might result in slight differences in the generated assembly code."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td colspan=\"11\">GPT-4 (OpenAI, 2023)</td></tr><tr><td>vanilla</td><td>76.83</td><td>67.68</td><td>67.07</td><td>57.93</td><td>67.38</td><td>20.73</td><td>13.41</td><td>13.41</td><td>10.98</td><td>14.63</td></tr><tr><td>+sc2dec</td><td>76.22</td><td>67.03</td><td>65.24</td><td>56.10</td><td>66.15</td><td>32.93</td><td>20.12</td><td>19.51</td><td>13.41</td><td>21.49</td></tr><tr><td>1-shot</td><td>76.22</td><td>78.05</td><td>73.78</td><td>70.73</td><td>74.70</td><td>31.71</td><td>20.12</td><td>19.51</td><td>21.34</td><td>23.17</td></tr><tr><td>+sc2dec</td><td>74.39</td><td>76.22</td><td>71.95</td><td>68.90</td><td>72.87</td><td>40.24</td><td>25.61</td><td>23.78</td><td>23.17</td><td>28.20</td></tr><tr><td colspan=\"11\">GPT-3.5 Turbo (OpenAI, 2022)</td></tr><tr><td>vanilla</td><td>20.73</td><td>25.00</td><td>29.27</td><td>21.95</td><td>24.24</td><td>4.27</td><td>3.66</td><td>3.05</td><td>3.05</td><td>3.51</td></tr><tr><td>+sc2dec</td><td>18.29</td><td>22.56</td><td>28.66</td><td>20.12</td><td>22.41</td><td>5.49</td><td>3.66</td><td>3.66</td><td>4.27</td><td>4.27</td></tr><tr><td>1-shot</td><td>49.39</td><td>46.34</td><td>42.07</td><td>40.85</td><td>44.66</td><td>10.37</td><td>7.32</td><td>6.10</td><td>5.49</td><td>7.31</td></tr><tr><td>+sc2dec</td><td>42.07</td><td>42.68</td><td>36.59</td><td>34.15</td><td>38.87</td><td>12.20</td><td>9.15</td><td>6.71</td><td>7.93</td><td>8.99</td></tr><tr><td colspan=\"11\">Deepseek Chat (DeepSeek-AI, 2024)</td></tr><tr><td>vanilla</td><td>62.20</td><td>51.83</td><td>47.56</td><td>45.73</td><td>51.83</td><td>10.37</td><td>5.49</td><td>4.88</td><td>5.49</td><td>6.55</td></tr><tr><td>+sc2dec</td><td>57.93</td><td>50.00</td><td>46.95</td><td>45.12</td><td>50.00</td><td>13.41</td><td>6.71</td><td>6.71</td><td>7.32</td><td>8.54</td></tr><tr><td>1-shot</td><td>67.68</td><td>70.73</td><td>64.63</td><td>56.71</td><td>64.94</td><td>13.41</td><td>10.37</td><td>7.93</td><td>7.32</td><td>9.76</td></tr><tr><td>+sc2dec</td><td>65.85</td><td>70.12</td><td>62.20</td><td>53.66</td><td>62.96</td><td>17.68</td><td>11.59</td><td>10.37</td><td>9.15</td><td>12.20</td></tr><tr><td colspan=\"11\">Deepseek Coder (Guo et al., 2024)</td></tr><tr><td>vanilla</td><td>66.46</td><td>57.93</td><td>47.56</td><td>51.83</td><td>55.95</td><td>9.15</td><td>6.10</td><td>6.71</td><td>6.98</td><td>7.01</td></tr><tr><td>+sc2dec</td><td>62.80</td><td>54.27</td><td>43.29</td><td>50.00</td><td>52.59</td><td>12.20</td><td>7.93</td><td>9.15</td><td>8.54</td><td>9.45</td></tr><tr><td>1-shot</td><td>57.93</td><td>53.66</td><td>52.44</td><td>48.17</td><td>53.05</td><td>11.59</td><td>10.37</td><td>8.54</td><td>7.93</td><td>9.60</td></tr><tr><td>+sc2dec</td><td>56.10</td><td>51.83</td><td>51.22</td><td>46.34</td><td>51.37</td><td>15.24</td><td>11.58</td><td>10.37</td><td>8.54</td><td>11.43</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b (Tan et al., 2024)</td></tr><tr><td>vanilla</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>92.68</td><td>92.80</td><td>90.12</td><td>93.41</td><td>92.26</td><td>71.34</td><td>47.56</td><td>45.98</td><td>41.71</td><td>51.65</td></tr><tr><td>1-shot</td><td>93.05</td><td>93.29</td><td>90.98</td><td>93.90</td><td>92.80</td><td>70.37</td><td>44.88</td><td>41.46</td><td>37.68</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>93.05</td><td>92.68</td><td>90.37</td><td>93.29</td><td>92.35</td><td>71.59</td><td>47.20</td><td>46.34</td><td>41.46</td><td>51.65</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b + FAE</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr></table>",
+        "bbox": [
+            152,
+            348,
+            845,
+            864
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 5: The extra results of different methods across four optimization levels (O0, O1, O2, O3), as well as their average scores (AVG). The results in bold represent the optimal performance, while those underlined indicate the second-best performance.",
+        "bbox": [
+            112,
+            872,
+            882,
+            916
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "6614",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    }
+]

2024/Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement/45f4817a-44a7-499d-b884-1bdac2c1b3e5_model.json

@@ -0,0 +1,2127 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.129,
+                0.09,
+                0.872,
+                0.13
+            ],
+            "angle": 0,
+            "content": "Self-Constructed Context Decomposition with Fined-grained Alignment Enhancement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.272,
+                0.153,
+                0.734,
+                0.187
+            ],
+            "angle": 0,
+            "content": "Yunlong Feng, Dechuan Teng, Yang Xu, Honglin Mu, Xiao Xu, Libo Qin, Qingfu Zhu*, Wanxiang Che"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.346,
+                0.187,
+                0.654,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Harbin Institute of Technology, China"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.257,
+                0.204,
+                0.744,
+                0.22
+            ],
+            "angle": 0,
+            "content": "{ylfeng,dcteng,yxu,hlmu,xxu,lbqin,qfzhu,car} \\(@\\) ir.hit.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.261,
+                0.341,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.287,
+                0.461,
+                0.657
+            ],
+            "angle": 0,
+            "content": "Decompilation transforms compiled code back into a high-level programming language for analysis when source code is unavailable. Previous work has primarily focused on enhancing decomposition performance by increasing the scale of model parameters or training data for pre-training. Based on the characteristics of the decomposition task, we propose two methods: (1) Without fine-tuning, the Self-Constructed Context Decomposition \\((\\mathrm{sc}^2\\mathrm{dec})\\) method recombines the LLM's decomposition results to construct pairs for in-context learning, helping the model improve decomposition performance. (2) Fine-grained Alignment Enhancement (FAE), which meticulously aligns assembly code with source code at the statement level by leveraging debugging information, is employed during the fine-tuning phase to achieve further improvements in decomposition. By integrating these two methods, we achieved a Re-Executability performance improvement of approximately \\(3.90\\%\\) on the Decompile-Eval benchmark, establishing a new state-of-the-art performance of \\(52.41\\%\\). The code, data, and models are available at https://github.com/AlongWY/sccdec."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.667,
+                0.262,
+                0.683
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.692,
+                0.491,
+                0.902
+            ],
+            "angle": 0,
+            "content": "Decompilation is the process of converting compiled machine code or bytecode back into a high-level programming language. This process is typically used to analyze how the software works when the source code is not accessible (Brumley et al., 2013; Katz et al., 2018; Hosseini and Dolan-Gavitt, 2022; Xu et al., 2023; Armengol-Estape et al., 2023; Jiang et al., 2023; Wong et al., 2023). Many tools have been developed for decompilation, like Ghidra (Ghidra, 2024) and IDA Pro (Hex-Rays, 2024). However, these tools often struggle to generate human-readable code. The main challenge in decompilation is that it's hard to fully reconstruct"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.565,
+                0.258,
+                0.828,
+                0.451
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.46,
+                0.885,
+                0.532
+            ],
+            "angle": 0,
+            "content": "Figure 1: Pipeline of the decompilation. The input for decompilation tasks is typically assembly, with the source code invisible. Additionally, the source code obtained through decompilation usually does not exactly match the original code."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.564,
+                0.885,
+                0.629
+            ],
+            "angle": 0,
+            "content": "the source code, especially details such as variable names (Lacomis et al., 2019) and primary structures (Wei et al., 2007), which are frequently lost during the compilation process."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.632,
+                0.886,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Recent advances in large language models (LLMs) have prompted researchers to view programming languages as distinct linguistic systems, utilizing pre-trained code LLMs to accomplish various coding tasks (Lippincott, 2020; Rozière et al., 2023; Guo et al., 2024; OpenAI, 2023). These models have demonstrated significant performance improvements over traditional techniques (Zeng et al., 2022; Xu et al., 2022), making it feasible to apply LLMs to decompilation challenges. For instance, Transformer-based models such as Slade (Armengol-Estapé et al., 2023) and Nova (Jiang et al., 2023) have shown potential in using language models to translate binary code back into more readable and structured source code. Recently, Tan et al. (2024) developed and released the first open-source LLM focused on decompilation, alongside the construction of the first"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.907,
+                0.285,
+                0.922
+            ],
+            "angle": 0,
+            "content": "*Corresponding Author"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6603"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.218,
+                0.946,
+                0.78,
+                0.958
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 6603-6614"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.275,
+                0.959,
+                0.725,
+                0.972
+            ],
+            "angle": 0,
+            "content": "November 12-16, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.196
+            ],
+            "angle": 0,
+            "content": "decompilation benchmark to evaluate the decompilation capabilities of models. These efforts primarily treat the decompilation task as a translation task, training a robust decompilation model on extensive synthetic data (Hosseini and Dolan-Gavitt, 2022; Armengol-Estapé et al., 2023; Jiang et al., 2023; Tan et al., 2024)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.198,
+                0.49,
+                0.229
+            ],
+            "angle": 0,
+            "content": "By analyzing the characteristics of decomposition tasks, we have found two features:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.242,
+                0.486,
+                0.273
+            ],
+            "angle": 0,
+            "content": "- The decompiled code generated by the model is usually compileable."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.274,
+                0.485,
+                0.306
+            ],
+            "angle": 0,
+            "content": "- The compiler can produce external debugging information for the debugger."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.242,
+                0.486,
+                0.306
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.317,
+                0.49,
+                0.639
+            ],
+            "angle": 0,
+            "content": "Based on the characteristics of the decomposition task, we propose two approaches: (1) Self-Constructed Context Decomposition \\((\\mathrm{sc}^2\\mathrm{dec})\\), a tuning-free approach, leverages the decomposition results from the model and specific compiler to derive ground-truth (assembly code, source code) pair closely related to the test sample. By incorporating the obtained pair as a demonstration example within the context, we can fully leverage the model's in-context learning capabilities to enhance performance. (2) Fine-grained Alignment Enhancement (FAE) introduces a new step-by-step decomposition training objective in addition to the end-to-end decomposition training objective that relies on large-scale pre-training for implicit alignment. By leveraging debugging information designed for debuggers, we can align assembly code with high-level code at a finer statement level. This allows the model to more faithfully preserve the original functionality after decomposition."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.64,
+                0.49,
+                0.719
+            ],
+            "angle": 0,
+            "content": "Combining these two methods, we have achieved an approximately \\(3.90\\%\\) improvement in re-executability performance based on lvm4decompile model (Tan et al., 2024), reaching a new state-of-the-art performance of \\(52.41\\%\\) re-executability."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.721,
+                0.488,
+                0.736
+            ],
+            "angle": 0,
+            "content": "To summarize, our contributions are as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.75,
+                0.488,
+                0.812
+            ],
+            "angle": 0,
+            "content": "- We introduce Self-Constructed Context Decomposition, which leverages the compileability of decomposition results to construct better contexts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.826,
+                0.488,
+                0.921
+            ],
+            "angle": 0,
+            "content": "- We introduce Fine-grained Alignment Enhancement, which fine-tunes the model using fine-grained alignment data between assembly code and source code derived from debug information. We also propose how to synthesize its training data automatically."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.75,
+                0.488,
+                0.921
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.085,
+                0.882,
+                0.133
+            ],
+            "angle": 0,
+            "content": "- The experiments on the Decompile-Eval benchmark show that our proposed model achieves new state-of-the-art results."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.147,
+                0.673,
+                0.162
+            ],
+            "angle": 0,
+            "content": "2 Related Works"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.174,
+                0.885,
+                0.51
+            ],
+            "angle": 0,
+            "content": "Decompilation has a long history, centered on converting binary files into human-readable source code. Traditional tools such as Hex-Rays IDA Pro (Hex-Rays, 2024) and Ghidra (Ghidra, 2024) predominantly rely on analyzing a program's control flow graph (CFG) to perform this task (Brumley et al., 2013). These tools operate by scrutinizing instructions within assembly code to construct CFGs and identify common programming structures like if-else statements and loops (Wei et al., 2007). However, they rely on intricate and error-prone expert-built rule systems that frequently fail with optimized binary code, a common practice in commercial software development. Moreover, traditional tools often generate outputs that closely resemble assembly code representations, translating variables into registers and utilizing low-level operations such as goto statements. This output is not only challenging to comprehend but also occasionally cannot be recompiled (Liu and Wang, 2020)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.512,
+                0.884,
+                0.818
+            ],
+            "angle": 0,
+            "content": "Inspired by neural machine translation, researchers have begun to conceptualize decompilation as a translation task, wherein machine-level instructions are converted into readable source code using neural networks. Initial endeavors employing recurrent neural networks (RNNs) demonstrated limited success (Katz et al., 2018). However, recent advancements in natural language processing (NLP) have facilitated the application of pre-trained language models (LMs) in decompilation. Notable examples include BTC (Hosseini and Dolan-Gavitt, 2022), Slade (Armengol-Estapé et al., 2023), and Nova (Jiang et al., 2023), which mark significant strides in the field. Recently, Tan et al. (2024) created and released the first open-source large language models specifically for decompilation and established an evaluation benchmark that considers the re-compilability and re-executability of decompiled code."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.831,
+                0.613,
+                0.846
+            ],
+            "angle": 0,
+            "content": "3 Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.858,
+                0.882,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce the Self-Constructed Context Decomposition method and Fine-grained Alignment Enhancement fine-tuning. Figures 2 and 3 illustrate the methods."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6604"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.116,
+                0.082,
+                0.489,
+                0.358
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.366,
+                0.49,
+                0.438
+            ],
+            "angle": 0,
+            "content": "Figure 2: The pipeline for Self-Constructed Context Decomposition operates as follows: when the LLM decomposes and generates compilable code, we compile this code to construct the context, and then use it to decompile the initial assembly code again."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.462,
+                0.48,
+                0.478
+            ],
+            "angle": 0,
+            "content": "3.1 Self-Constructed Context Decomposition"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.483,
+                0.489,
+                0.595
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce the Self-Constructed Context Decomposition (sc²dec) method designed to enhance the performance of decomposition models. The method recombines the decomposed code generated by the model to construct the context. The sc²dec method involves the following key steps, as the Figure 2 shows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.606,
+                0.489,
+                0.669
+            ],
+            "angle": 0,
+            "content": "1. Decompile: Initially, the model decomposes the assembly code back to the target programming language. We call it the initial decomposition result."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.67,
+                0.489,
+                0.734
+            ],
+            "angle": 0,
+            "content": "2. Compile and Disassembly: Then we recompile the initial decomposition result to obtain the corresponding assembly code if the generated code is compileable."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.735,
+                0.489,
+                0.798
+            ],
+            "angle": 0,
+            "content": "3. Constructed Context: The self-constructed context is then formed by concatenating the generated C code with its corresponding assembly code."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.799,
+                0.488,
+                0.831
+            ],
+            "angle": 0,
+            "content": "4. Decompile Again: Decompile the assembly code with the self-constructed context again."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.129,
+                0.606,
+                0.489,
+                0.831
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.842,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "In this context, the assembly code precedes the C code, resembling an example in the context. The final input to the model is a concatenation of the self-constructed context (with assembly code first and C code second) and the original assembly code. This"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.532,
+                0.084,
+                0.863,
+                0.421
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.43,
+                0.885,
+                0.515
+            ],
+            "angle": 0,
+            "content": "Figure 3: An example of step-by-step decomposition is presented, where the training objective requires the model to generate C code progressively after each assembly block. Fine-tuning the model with this objective aids in learning the fine-grained correspondences between assembly and C code."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.541,
+                0.884,
+                0.604
+            ],
+            "angle": 0,
+            "content": "method leverages the additional context provided by the generated C code and its corresponding assembly code to improve the overall decompilation performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.616,
+                0.857,
+                0.631
+            ],
+            "angle": 0,
+            "content": "3.2 Fine-grained Alignment Enhancement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.637,
+                0.885,
+                0.797
+            ],
+            "angle": 0,
+            "content": "To enhance the model's ability to perceive the correspondence between assembly code and its corresponding source code, we propose a Fine-grained Alignment Enhancement (FAE) method, which constructs fine-grained alignment data between assembly code and source code derived from debug information. The method primarily includes two training objectives: the End-to-end Decomposition Objective and the Step-by-Step Decomposition Objective."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.809,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "- End-to-end Decomposition Objective: In the end-to-end decomposition task, the model takes assembly code as input and directly generates the corresponding decomposed C code. The goal of this task is to enable the model to translate from assembly language to C language without any intermediate steps. This ob"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6605"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.149,
+                0.085,
+                0.488,
+                0.165
+            ],
+            "angle": 0,
+            "content": "jective aims to prevent catastrophic forgetting in the model. During training, we increase the context length and ensure that the same sample is not truncated, which enhances the model's ability to align assembly and C code."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.167,
+                0.49,
+                0.391
+            ],
+            "angle": 0,
+            "content": "- Step-by-step Decomposition Objective: As the Figure 3 shows, in the step-by-step decomposition process, an intermediate form containing both the assembly code and the corresponding C code is first generated, followed by the generation of the complete C code. This step-by-step approach allows the model to better understand the correspondence between assembly code and C code. Training the model with the step-by-step decomposition objective can enhance the model's ability to finely align assembly code with C code. It is important to note that this form is used only during training and not during inference."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.402,
+                0.49,
+                0.497
+            ],
+            "angle": 0,
+            "content": "During the training process, we employed two different training objectives to enhance the model's decompile performance: an end-to-end decompile objective and a step-by-step decompile objective. They all take assembly code as input, but their output formats are different."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.507,
+                0.303,
+                0.523
+            ],
+            "angle": 0,
+            "content": "3.2.1 Data Processing"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.525,
+                0.49,
+                0.622
+            ],
+            "angle": 0,
+            "content": "By following the steps below, we can synthesize some fine-grained alignment training data. The synthesized data, as shown in Figure 3, is in a format of interleaved assembly and code output. We expect the model to learn finer-grained alignment between assembly and code from such data."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.131,
+                0.632,
+                0.49,
+                0.793
+            ],
+            "angle": 0,
+            "content": "1. Compilation: We employed the \"gcc\" compiler to compile the selected function code into binary shared libraries at various optimization levels (including O0, O1, O2, and O3). The debug option was enabled during compilation to ensure the inclusion of DWARF debugging information in the compiled output. This debugging information is crucial for subsequent disassembly and parsing steps."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.794,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "2. Disassembly: Subsequently, we used the \"objdump\" tool to disassemble the compiled binary shared libraries. The \"-S\" option was utilized to include source code in the disassembly output. Additionally, we used the \"-source/comment=;,\" option toptides each line of source code with a \";,\" symbol, facilitating easier parsing in later stages."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.524,
+                0.085,
+                0.885,
+                0.213
+            ],
+            "angle": 0,
+            "content": "3. Data Parsing: The disassembly process yielded a mixed output containing both assembly code and source code, organized in an alternating sequence of source code followed by assembly code. We parsed the output to extract the corresponding assembly and source code for each function, ensuring accurate alignment between the two."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.524,
+                0.214,
+                0.885,
+                0.278
+            ],
+            "angle": 0,
+            "content": "4. Data Reorganization: Finally, we reorganized the parsed content into a format where the assembly code precedes the corresponding source code in an alternating sequence."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.524,
+                0.085,
+                0.885,
+                0.278
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.29,
+                0.885,
+                0.403
+            ],
+            "angle": 0,
+            "content": "Through these steps, we constructed a high-quality training dataset that ensures each sample contains comprehensive source code information along with the corresponding assembly code. This dataset will be used to train our model, aiming to enhance its performance in tasks such as code understanding and compilation optimization."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.414,
+                0.656,
+                0.431
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.439,
+                0.642,
+                0.453
+            ],
+            "angle": 0,
+            "content": "4.1 Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.46,
+                0.883,
+                0.54
+            ],
+            "angle": 0,
+            "content": "In this section, we will introduce the benchmark used in our evaluation process, Decompile-Eval (Tan et al., 2024), which is specifically designed to assess the decompilation capabilities of large language models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.54,
+                0.884,
+                0.701
+            ],
+            "angle": 0,
+            "content": "The Decompile-Eval benchmark is adapted from the HumanEval benchmark, which includes 164 problems initially designed for code generation tasks. These problems were translated into the C programming language, and the corresponding assembly code was generated at four optimization levels (O0, O1, O2, and O3). The correctness of the decomposition results was tested using the test cases from HumanEval. The primary metrics of the Decompile-Eval benchmark are as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.713,
+                0.885,
+                0.84
+            ],
+            "angle": 0,
+            "content": "- Re-compilability: This metric evaluates whether the decompiled code produced by our model can be successfully recompiled into executable binary without errors. A high recompiliability rate indicates that the decompiled code is syntactically correct and adheres to the constraints of the target language (in this case, C)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.842,
+                0.887,
+                0.922
+            ],
+            "angle": 0,
+            "content": "- Re-executability: This metric assesses the functional correctness of the decompiled code. Specifically, it measures whether the recom-piled binaries produce the expected outputs when executed. The correctness of the output"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.713,
+                0.887,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6606"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.085,
+                0.49,
+                0.15
+            ],
+            "angle": 0,
+            "content": "is determined using the testing methodology provided by the HumanEval dataset, ensuring a comprehensive evaluation of the logical accuracy of the decompiled code."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.167,
+                0.489,
+                0.215
+            ],
+            "angle": 0,
+            "content": "In our experiment, we pay more attention to the Re-executability, as it more accurately reflects the overall decompilation capability of the model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.23,
+                0.289,
+                0.244
+            ],
+            "angle": 0,
+            "content": "4.2 Training Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.252,
+                0.49,
+                0.478
+            ],
+            "angle": 0,
+            "content": "Implementation We utilize Lora (Hu et al., 2022) to fine-tune the llm4decompile-6.7b obtained on Hugging Face (Wolf et al., 2019), with rank set to 32, alpha set to 64 and target set to all projection layers \\(^{1}\\). The optimizer is AdamW (Loshchilov and Hutter, 2019), with a learning rate of 5e-5. The maximum sequence length is set to 16384, and the learning rate scheduler type is cosine, with a warm-up period of 20 steps. The training process was conducted for one epoch. LlamaFactory (Zheng et al., 2024) and FlashAttention 2 (Dao, 2024) were used for the fine-tuning of the model. All experiments were done with an A100-SXM4-80GB. We use greedy decoding for all experiments."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.49,
+                0.49,
+                0.603
+            ],
+            "angle": 0,
+            "content": "Training data We selected 10,000 samples from the train_real_compilable subset of Exebench (Armengol-Estapé et al., 2022) to synthesis the training data. The selected functions exclusively utilize the standard C library and do not include additional data structures. The training data were synthesized with gcc 11.4 provided by Ubuntu 22.04."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.618,
+                0.235,
+                0.632
+            ],
+            "angle": 0,
+            "content": "4.3 Baselines"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.641,
+                0.489,
+                0.687
+            ],
+            "angle": 0,
+            "content": "We present several baselines in our experiments and demonstrate the effectiveness of our proposed methods."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.702,
+                0.233,
+                0.715
+            ],
+            "angle": 0,
+            "content": "4.3.1 Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.722,
+                0.489,
+                0.784
+            ],
+            "angle": 0,
+            "content": "- GPT 4: One of the most powerful OpenAI models, known for its advanced language understanding and generation capabilities (OpenAI, 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.787,
+                0.489,
+                0.849
+            ],
+            "angle": 0,
+            "content": "- deepseek-chat: A powerful conversational AI model from DeepSeek, excelling in generating coherent and contextually relevant dialogues (DeepSeek-AI, 2024)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.851,
+                0.488,
+                0.882
+            ],
+            "angle": 0,
+            "content": "- llm4decompile-6.7b: The 6.7B open-access decomposition LLM pre-trained on 15 billion"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.722,
+                0.489,
+                0.882
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.895,
+                0.489,
+                0.922
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)q_Proj, k_Proj, v_Proj, o_Proj, gate_Proj, up_Proj, down_Proj"
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.518,
+                0.085,
+                0.856,
+                0.226
+            ],
+            "angle": 0,
+            "content": "bool func0(int num){ if (num <= 1) { return false; } for (int i = 2; i * i <= num; i++) { if (num % i == 0) { return false; } } return true; }"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.24,
+                0.883,
+                0.313
+            ],
+            "angle": 0,
+            "content": "Figure 4: The example for 1-shot learning. In this example, we have tried to cover common control logic such as if statements, loops, and early returns as much as possible. The example will be compiled with the same optimization level as the target assembly code."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.545,
+                0.338,
+                0.882,
+                0.369
+            ],
+            "angle": 0,
+            "content": "tokens of C source code and the corresponding assembly code (Tan et al., 2024)\\(^2\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.37,
+                0.884,
+                0.448
+            ],
+            "angle": 0,
+            "content": "- llm4decompile-6.7b+FAE: The new model was obtained by applying the Fine-grained Alignment Enhancement method for further fine-tuning based on the llm4decompile-6.7b\\(^2\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.463,
+                0.64,
+                0.477
+            ],
+            "angle": 0,
+            "content": "4.3.2 Methods"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.481,
+                0.882,
+                0.513
+            ],
+            "angle": 0,
+            "content": "- vanilla: Using specific prompts for different models to directly request decomposition."
+        },
+        {
+            "type": "code_caption",
+            "bbox": [
+                0.574,
+                0.525,
+                0.856,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Prompt For GPT-4 and Deepseek Chat"
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.572,
+                0.553,
+                0.836,
+                0.627
+            ],
+            "angle": 0,
+            "content": "What is the C source code of \\(\\leftrightarrow\\) the assembly code below: ..asm <asm>"
+        },
+        {
+            "type": "code_caption",
+            "bbox": [
+                0.574,
+                0.657,
+                0.777,
+                0.672
+            ],
+            "angle": 0,
+            "content": "Prompt For 1lm4decompile"
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.573,
+                0.686,
+                0.832,
+                0.716
+            ],
+            "angle": 0,
+            "content": "This is the assembly code: <asm>"
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.574,
+                0.734,
+                0.816,
+                0.749
+            ],
+            "angle": 0,
+            "content": "What is the source code?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.77,
+                0.885,
+                0.899
+            ],
+            "angle": 0,
+            "content": "- 1-shot: We write a fixed example as the context (Brown et al., 2020). The example will be compiled with the same optimization level as the target assembly code. The example shown in Figure 4. In this example, we have tried to cover common control logic such as if statements, loops, and early returns as much as possible."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.531,
+                0.907,
+                0.854,
+                0.921
+            ],
+            "angle": 0,
+            "content": "2We use the latest version, llm4decompile-6.7b-v1.5."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6607"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.115,
+                0.082,
+                0.885,
+                0.502
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td colspan=\"11\">GPT-4 (OpenAI, 2023)</td></tr><tr><td>vanilla</td><td>76.83</td><td>67.68</td><td>67.07</td><td>57.93</td><td>67.38</td><td>20.73</td><td>13.41</td><td>13.41</td><td>10.98</td><td>14.63</td></tr><tr><td>+sc2dec</td><td>76.22</td><td>67.03</td><td>65.24</td><td>56.10</td><td>66.15</td><td>32.93</td><td>20.12</td><td>19.51</td><td>13.41</td><td>21.49</td></tr><tr><td>1-shot</td><td>76.22</td><td>78.05</td><td>73.78</td><td>70.73</td><td>74.70</td><td>31.71</td><td>20.12</td><td>19.51</td><td>21.34</td><td>23.17</td></tr><tr><td>+sc2dec</td><td>74.39</td><td>76.22</td><td>71.95</td><td>68.90</td><td>72.87</td><td>40.24</td><td>25.61</td><td>23.78</td><td>23.17</td><td>28.20</td></tr><tr><td colspan=\"11\">Deepseek Chat (DeepSeek-AI, 2024)</td></tr><tr><td>vanilla</td><td>62.20</td><td>51.83</td><td>47.56</td><td>45.73</td><td>51.83</td><td>10.37</td><td>5.49</td><td>4.88</td><td>5.49</td><td>6.55</td></tr><tr><td>+sc2dec</td><td>57.93</td><td>50.00</td><td>46.95</td><td>45.12</td><td>50.00</td><td>13.41</td><td>6.71</td><td>6.71</td><td>7.32</td><td>8.54</td></tr><tr><td>1-shot</td><td>67.68</td><td>70.73</td><td>64.63</td><td>56.71</td><td>64.94</td><td>13.41</td><td>10.37</td><td>7.93</td><td>7.32</td><td>9.76</td></tr><tr><td>+sc2dec</td><td>65.85</td><td>70.12</td><td>62.20</td><td>53.66</td><td>62.96</td><td>17.68</td><td>11.59</td><td>10.37</td><td>9.15</td><td>12.20</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b (Tan et al., 2024)</td></tr><tr><td>vanilla</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>92.68</td><td>92.80</td><td>90.12</td><td>93.41</td><td>92.26</td><td>71.34</td><td>47.56</td><td>45.98</td><td>41.71</td><td>51.65</td></tr><tr><td>1-shot</td><td>93.05</td><td>93.29</td><td>90.98</td><td>93.90</td><td>92.80</td><td>70.37</td><td>44.88</td><td>41.46</td><td>37.68</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>93.05</td><td>92.68</td><td>90.37</td><td>93.29</td><td>92.35</td><td>71.59</td><td>47.20</td><td>46.34</td><td>41.46</td><td>51.65</td></tr><tr><td>retrieval</td><td>94.02</td><td>88.05</td><td>84.51</td><td>85.12</td><td>87.93</td><td>65.73</td><td>32.68</td><td>38.66</td><td>36.22</td><td>43.32</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b + FAE</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.511,
+                0.884,
+                0.556
+            ],
+            "angle": 0,
+            "content": "Table 1: The main results of different methods across four optimization levels (O0, O1, O2, O3), as well as their average scores (AVG). The results in bold represent the optimal performance, while those underlined indicate the second-best performance. More results can be found in the Table 5."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.58,
+                0.487,
+                0.643
+            ],
+            "angle": 0,
+            "content": "- retrieval: We use the retrieval method (BM25 (Lù, 2024)) to get the context. The context is generated by searching for the most similar assembly code in the training set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.644,
+                0.488,
+                0.675
+            ],
+            "angle": 0,
+            "content": "- \\(\\mathbf{sc}^2\\mathbf{dec}\\): The self-constructed context decomposition we described in the Section 3.1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.676,
+                0.488,
+                0.74
+            ],
+            "angle": 0,
+            "content": "- 1-shot+sc²dec: We apply the sc²dec on the 1-shot method. Note that the 1-shot example can only be applied to generate the initial decomposition result in Figure 2."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.58,
+                0.488,
+                0.74
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.764,
+                0.22,
+                0.779
+            ],
+            "angle": 0,
+            "content": "4.4 Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.785,
+                0.49,
+                0.88
+            ],
+            "angle": 0,
+            "content": "In this section, we present and analyze the performance of different models. Our experiments aim to evaluate the impact of various techniques, including one-shot, self-constructed context decomposition, and fine-tuning, on the model's performance. The main results are shown in Table 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.89,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Performance of the Base Model Through a comparative analysis of the experimental results in Ta"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.58,
+                0.885,
+                0.789
+            ],
+            "angle": 0,
+            "content": "ble 1, it is evident that different models exhibit significant differences in re-executability across various optimization levels (O0, O1, O2, O3). The base model, llm4decompile-6.7b, performs best at the O0 optimization level with a score of \\(70.24\\%\\), but its performance gradually declines at other optimization levels, with an average score of \\(48.60\\%\\). Among models not fine-tuned with decompilation tasks, GPT-4 achieved the best performance, with an average re-executability score of \\(14.64\\%\\). Generally, their results on re-compilability metrics are significantly higher than their performance on re-executability metrics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.809,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Context Sensitivity By comparing vanilla, 1-shot, and retrieval methods based on the llm4decompile-6.7b, it is observed that the model's performance can even decline after providing a sample. In contrast, the dynamic samples generated through our self-constructed context method yield an improvement in executability of over \\(2\\%\\)"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6608"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.082,
+                0.885,
+                0.228
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>sc2dec (O0)</td><td>92.68</td><td>92.20</td><td>91.71</td><td>92.32</td><td>92.23</td><td>70.24</td><td>48.90</td><td>47.68</td><td>42.93</td><td>51.83</td></tr><tr><td>sc2dec (O1)</td><td>92.68</td><td>91.59</td><td>93.54</td><td>92.93</td><td>92.68</td><td>68.41</td><td>48.54</td><td>47.44</td><td>43.05</td><td>51.86</td></tr><tr><td>sc2dec (O2)</td><td>92.68</td><td>92.20</td><td>91.71</td><td>92.32</td><td>92.23</td><td>67.80</td><td>48.90</td><td>47.68</td><td>42.93</td><td>51.68</td></tr><tr><td>sc2dec (O3)</td><td>92.68</td><td>91.59</td><td>92.93</td><td>92.32</td><td>92.38</td><td>67.80</td><td>47.07</td><td>47.80</td><td>43.41</td><td>51.52</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.236,
+                0.884,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Table 2: The results of our method across optimization levels. For cases such as \"sc²dec (O0)\", we assume that the optimization level of asm code is entirely unknown and select a specific optimization level to compile the code to construct the context."
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.125,
+                0.303,
+                0.482,
+                0.621
+            ],
+            "angle": 0,
+            "content": "int func0(int a){ for (int i = 0; i * i * i <= abs(a); i++) if (i * i * i == abs(a)) return 1; return 0; } Source code   \nint func0(int n){ int i; if \\((n <   0)\\) n \\(=\\) -n; for \\((\\mathrm{i} = 1;\\mathrm{i}\\ast \\mathrm{i}\\ast \\mathrm{i} <   = \\mathrm{n};\\mathrm{i} + + )\\) if \\((i*\\bar{i}*i==n)\\) return 1; return 0; } Decompile   \nint func0(int n){ int i; if \\((n <   0)\\) n \\(=\\) -n; if \\((n = = 0)\\) return 1; for \\((\\mathrm{i} = 1;\\mathrm{i}\\ast \\mathrm{i}\\ast \\mathrm{i} <   = \\mathrm{n};\\mathrm{i} + + )\\) if \\((i*\\bar{i}*i==n)\\) return 1; return 0; } SC2Dec"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.634,
+                0.49,
+                0.693
+            ],
+            "angle": 0,
+            "content": "Figure 5: A case study for \\(\\mathrm{sc^2dec}\\), which is based on the llm4decompile-6.7b with FAE. By applying the \\(\\mathrm{sc^2dec}\\) method, the model detects and fixes the mismatch between the code and assembly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.718,
+                0.49,
+                0.799
+            ],
+            "angle": 0,
+            "content": "before and after fine-tuning. This indicates that inappropriate samples may even have a detrimental effect in the decompilation domain, while our method achieves a nearly stable performance enhancement without the need for retrieval."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.809,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "1-shot is a Strong Baseline As shown in Table 1, utilizing the 1-shot method significantly enhances the model's performance. Specifically, for GPT-4, the model's re-executability performance increased from \\(14.64\\%\\) to \\(23.17\\%\\), an approximate improvement of \\(7\\%\\). In the open-source model Deepseek Chat, performance improved from \\(6.55\\%\\) to \\(9.76\\%\\),"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.305,
+                0.885,
+                0.435
+            ],
+            "angle": 0,
+            "content": "reflecting a \\(3.21\\%\\) enhancement. However, in the case of the llm4decompile model, which had already been trained on decompilation tasks, there was no performance improvement. Moreover, in the llm4decompile model fine-tuned with Fine-grained Alignment Enhancement, performance actually decreased from \\(50.88\\%\\) to \\(50.61\\%\\), a reduction of approximately \\(0.27\\%\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.445,
+                0.885,
+                0.766
+            ],
+            "angle": 0,
+            "content": "Sc2dec Brings Further Performance Improvement The Table 1 shows that applying sc2dec independently results in performance improvements of \\(6.86\\%\\) and \\(1.99\\%\\) on GPT-4 and Deepseek Chat, respectively. These improvements are lower than those achieved by the 1-shot method, likely due to the model's low Re-Compilability metric. In other words, a substantial portion of the code obtained through decompilation is un-compilable, which impedes the application of the sc2dec method. Furthermore, when combined with the 1-shot method, sc2dec can further enhance performance, achieving an additional \\(5\\%\\) improvement for GPT-4, bringing it to \\(28.20\\%\\), and a \\(2.44\\%\\) improvement for Deepseek Chat, bringing it to \\(12.20\\%\\). In the case of llm4decompile-6.7b, due to the models' relatively high Re-Compilability metric, the sc2dec method significantly outperformed the 1-shot method, increasing performance from \\(48.60\\%\\) to \\(51.65\\%\\) and from \\(50.88\\%\\) to \\(52.41\\%\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.777,
+                0.887,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Best Performance with Combined Methods As the Table 1 shows, by combining our Step-by-Step Decompile tasks and context-based decomposition methods, we ultimately achieved approximately a \\(3.90\\%\\) performance improvement, surpassing the performance of these methods when applied independently. This demonstrates that fine-tuning and self-construct context methods are orthogonal. Most notably, the combined method of fined"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6609"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.082,
+                0.885,
+                0.324
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>w/o tune</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td colspan=\"11\">FAE (End-to-End &amp; Step-by-Step Decomposition Objective)</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr><tr><td colspan=\"11\">End-to-End Decomposition Objective</td></tr><tr><td>vanilla</td><td>93.9</td><td>90.61</td><td>90.61</td><td>91.34</td><td>91.62</td><td>69.51</td><td>43.29</td><td>42.2</td><td>40.73</td><td>48.93</td></tr><tr><td>+sc2dec</td><td>92.8</td><td>90.24</td><td>90.00</td><td>90.12</td><td>90.79</td><td>71.34</td><td>46.10</td><td>43.78</td><td>42.44</td><td>50.91</td></tr><tr><td>1-shot</td><td>93.9</td><td>90.73</td><td>90.37</td><td>91.22</td><td>91.55</td><td>69.51</td><td>44.02</td><td>42.68</td><td>40.61</td><td>49.21</td></tr><tr><td>+sc2dec</td><td>92.8</td><td>90.24</td><td>89.76</td><td>90.00</td><td>90.7</td><td>71.22</td><td>46.46</td><td>44.27</td><td>42.56</td><td>51.13</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.18,
+                0.332,
+                0.816,
+                0.347
+            ],
+            "angle": 0,
+            "content": "Table 3: The results of ablation study. All results are based on the llm4decompile-6.7b model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.372,
+                0.49,
+                0.468
+            ],
+            "angle": 0,
+            "content": "grained alignment enhancement and self-construct context decomposition achieves the highest scores across all optimization levels expect \"O0\", with scores of \\(48.54\\%\\), \\(47.56\\%\\), and \\(43.29\\%\\) at the O1, O2, and O3 levels, respectively, and an overall average score of \\(52.41\\%\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.48,
+                0.229,
+                0.496
+            ],
+            "angle": 0,
+            "content": "4.5 Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.501,
+                0.49,
+                0.549
+            ],
+            "angle": 0,
+            "content": "In this section, we will further analyze the experimental results to study the effectiveness and sensitivity of our method:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.559,
+                0.49,
+                0.768
+            ],
+            "angle": 0,
+            "content": "Mismatched Optimization Levels Leads to Performance Degradation As the Table 2 shows, the performance can degrade by \\(1\\%\\) to \\(2\\%\\) overall when constructing contexts that mismatch the optimization levels of the target assembly code. However, specific optimization levels may exhibit different behaviors. For instance, using a fixed O3 optimization level to construct contexts while decompiling assembly optimized with O2 does not result in significant performance degradation and even better. In practical scenarios, using O3 to construct contexts for binaries with unknown optimization levels might be a reasonable choice."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.777,
+                0.49,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Case study of \\(\\mathbf{sc}^2\\mathbf{dec}\\) As the Figure 5 shows, the compiler has expanded the abs function, reducing the overhead of function calls. The model initially generated a piece of compilable but incorrect code during direct decompilation, omitting the case where the variable \\(n\\) equals 0. We compiled and disassembled this code to construct a new sample. By applying the \\(\\mathrm{sc^2dec}\\) method with the context, the model successfully identified the missing part"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.372,
+                0.885,
+                0.42
+            ],
+            "angle": 0,
+            "content": "in the short segment of assembly code and generated the correct code, thus rectifying the error from the previous inference."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.43,
+                0.886,
+                0.607
+            ],
+            "angle": 0,
+            "content": "Contribution of FAE The results in Table 1 show that the application of Fine-grained Alignment Enhancement (FAE) for further fine-tuning the model significantly enhanced its re-executability performance across all optimization levels except \"O0\". Notably, at optimization levels O2 and O3, the scores improved by \\(5.12\\%\\) and \\(4.27\\%\\), reaching \\(45.73\\%\\) and \\(42.32\\%\\), respectively. The average performance increased by \\(2.28\\%\\), achieving \\(50.88\\%\\). This demonstrates the effectiveness of the Fine-grained Alignment Enhancement method."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.616,
+                0.886,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Ablation Study of FAE The results of ablation experiments in Table 3 have demonstrated the effectiveness of our proposed Step-by-Step Decompile training objective. Removing the training objective results in approximately a \\(2\\%\\) decline in Re-Executability performance, regardless of whether the \\(\\mathrm{sc}^2\\) dec method is used, and notably, a \\(3.78\\%\\) decline at the \"O2\" optimization level. However, the performance of the 1-shot remains unaffected, supporting our hypothesis that the mismatch between the training data form and the inference form leads to performance degradation. Furthermore, retaining only the End-to-end Decompilation training objective still yields little performance improvement, about \\(0.33\\%\\). This improvement can be attributed to the longer context provided during training, as the samples are not truncated during training, allowing the model to fully observe the assembly code and its corresponding C code."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6610"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.248,
+                0.099
+            ],
+            "angle": 0,
+            "content": "5 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.109,
+                0.491,
+                0.286
+            ],
+            "angle": 0,
+            "content": "In this paper, we introduce two methods based on the characteristics of the decompilation task: Self-Constructed Context Decomposition, which leverages the compileability of decomposition results to construct better contexts, and Fine-grained Alignment Enhancement, which fine-tunes the model using fine-grained alignment data between assembly code and source code derived from debug information. By integrating these methods, we achieved a \\(3.90\\%\\) improvement in decompilation performance, reaching a new state-of-the-art level of \\(52.41\\%\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.297,
+                0.236,
+                0.314
+            ],
+            "angle": 0,
+            "content": "Acknowledge"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.321,
+                0.49,
+                0.385
+            ],
+            "angle": 0,
+            "content": "We gratefully acknowledge the support of the National Natural Science Foundation of China (NSFC) via grant 62236004, 62206078, 62441603 and 62476073."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.396,
+                0.22,
+                0.411
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.42,
+                0.49,
+                0.613
+            ],
+            "angle": 0,
+            "content": "The \\(\\mathrm{sc^2dec}\\) depends on the decomposed code generated by the model. If the decomposed code cannot be recompiled, \\(\\mathrm{sc^2dec}\\) will not be able to benefit from it, necessitating that the decomposed code has high re-compilability performance. For Fine-grained Alignment Enhancement, we created a Step-by-Step Decompile training objective, training with only 10,000 samples across four optimization levels (O0-O3), without validating its performance on larger datasets or bigger models. Moreover, fine-tuning led to a further decline in performance in the 1-shot scenario."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.625,
+                0.244,
+                0.64
+            ],
+            "angle": 0,
+            "content": "Potential risks"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.649,
+                0.49,
+                0.73
+            ],
+            "angle": 0,
+            "content": "Decompilation is a technique that converts compiled binary code back into human-readable source code. While decompilation can be legal and useful in certain contexts, it also entails several potential risks. The primary potential risks are as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.736,
+                0.488,
+                0.816
+            ],
+            "angle": 0,
+            "content": "- Intellectual Property Infringement: Decompilation may violate the copyright and licensing agreements of software. Unauthorized decomposition can lead to copyright infringement, thereby instigating legal disputes."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.826,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "- Security Risks: Decomposed code may expose the internal structure and implementation details of the software, providing attack vectors for hackers. Malicious actors can exploit this information to identify and exploit vulnerabilities in the software."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.137,
+                0.736,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.085,
+                0.883,
+                0.134
+            ],
+            "angle": 0,
+            "content": "- Ethical Concerns: Decompiling and analyzing another's code can be considered unethical, especially when done without authorization."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.511,
+                0.174,
+                0.61,
+                0.189
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.198,
+                0.884,
+                0.304
+            ],
+            "angle": 0,
+            "content": "Jordi Armengol-Estapé, Jackson Woodruff, Alexander Brauckmann, José Wesley de Souza Magalhães, and Michael F. P. O’Boyle. 2022. Exebench: An ml-scale dataset of executable c functions. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, MAPS 2022, page 50–59, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.314,
+                0.884,
+                0.368
+            ],
+            "angle": 0,
+            "content": "Jordi Armengol-Estape, Jackson Woodruff, Chris Cummins, and Michael F. P. O'Boyle. 2023. Slade: A portable small language model decoder for optimized assembler. CoRR, abs/2305.12520."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.378,
+                0.885,
+                0.537
+            ],
+            "angle": 0,
+            "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. Preprint, arXiv:2005.14165."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.547,
+                0.884,
+                0.639
+            ],
+            "angle": 0,
+            "content": "David Brumley, JongHyup Lee, Edward J. Schwartz, and Maverick Woo. 2013. Native x86 decomposition using semantics-preserving structural analysis and iterative control-flow structuring. In Proceedings of the 22th USENIX Security Symposium, Washington, DC, USA, August 14-16, 2013, pages 353-368. USENIX Association."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.65,
+                0.883,
+                0.703
+            ],
+            "angle": 0,
+            "content": "Tri Dao. 2024. FlashAttention-2: Faster attention with better parallelism and work partitioning. In International Conference on Learning Representations (ICLR)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.511,
+                0.714,
+                0.884,
+                0.755
+            ],
+            "angle": 0,
+            "content": "DeepSeek-AI. 2024. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. Preprint, arXiv:2405.04434."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.765,
+                0.882,
+                0.792
+            ],
+            "angle": 0,
+            "content": "Ghidra. 2024. Ghidra software reverse engineering framework."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.804,
+                0.884,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. 2024. Deepseek-coder: When the large language model meets programming – the rise of code intelligence. Preprint, arXiv:2401.14196."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.894,
+                0.884,
+                0.922
+            ],
+            "angle": 0,
+            "content": "Hex-Rays. 2024. Ida pro: a cross-platform multiprocessor disassembler and debugger."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.51,
+                0.198,
+                0.885,
+                0.922
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.52,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6611"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.125
+            ],
+            "angle": 0,
+            "content": "Iman Hosseini and Brendan Dolan-Gavitt. 2022. Beyond the C: retargetable decompilation using neural machine translation. CoRR, abs/2212.08950."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.136,
+                0.487,
+                0.202
+            ],
+            "angle": 0,
+            "content": "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."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.212,
+                0.487,
+                0.263
+            ],
+            "angle": 0,
+            "content": "Nan Jiang, Chengxiao Wang, Kevin Liu, Xiangzhe Xu, Lin Tan, and Xiangyu Zhang. 2023. Nova+: Generative language models for binaries. Preprint, arXiv:2311.13721."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.274,
+                0.487,
+                0.353
+            ],
+            "angle": 0,
+            "content": "Deborah S. Katz, Jason Ruchti, and Eric M. Schulte. 2018. Using recurrent neural networks for decompilation. In 25th International Conference on Software Analysis, Evolution and Reengineering, SANER 2018, Campobasso, Italy, March 20-23, 2018, pages 346-356. IEEE Computer Society."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.363,
+                0.487,
+                0.455
+            ],
+            "angle": 0,
+            "content": "Jeremy Lacomis, Pengcheng Yin, Edward J. Schwartz, Miltiadis Allamanis, Claire Le Goues, Graham Neubig, and Bogdan Vasilescu. 2019. DIRE: A neural approach to decompiled identifier naming. In 34th IEEE/ACM International Conference on Automated Software Engineering, ASE 2019, San Diego, CA, USA, November 11-15, 2019, pages 628-639. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.465,
+                0.487,
+                0.555
+            ],
+            "angle": 0,
+            "content": "Thomas Lippincott. 2020. Starcoder: A general neural ensemble technique to support traditional scholarship, illustrated with a study of the post-atlantic slave trade. In 15th Annual International Conference of the Alliance of Digital Humanities Organizations, DH 2020, Ottawa, Canada, July 20-25, 2020, Conference Abstracts."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.567,
+                0.487,
+                0.644
+            ],
+            "angle": 0,
+            "content": "Zhibo Liu and Shuai Wang. 2020. How far we have come: testing decomposition correctness of C decompliers. In ISSTA '20: 29th ACM SIGSOFT International Symposium on Software Testing and Analysis, Virtual Event, USA, July 18-22, 2020, pages 475-487. ACM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.655,
+                0.487,
+                0.72
+            ],
+            "angle": 0,
+            "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.731,
+                0.487,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Xing Han Lu. 2024. Bm25s: Orders of magnitude faster lexical search via eager sparse scoring. Preprint, arXiv:2407.03618."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.781,
+                0.292,
+                0.794
+            ],
+            "angle": 0,
+            "content": "OpenAI. 2022. ChatGPT."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.805,
+                0.487,
+                0.83
+            ],
+            "angle": 0,
+            "content": "OpenAI. 2023. GPT-4 Technical Report. CoRR, abs/2303.08774."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.842,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton-Ferrer, Aaron Grattaflori, Wenhan Xiong, Alexandre Défossez, Jade Copet,"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.92
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.138
+            ],
+            "angle": 0,
+            "content": "Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. 2023. Code llama: Open foundation models for code. CoRR, abs/2308.12950."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.149,
+                0.882,
+                0.188
+            ],
+            "angle": 0,
+            "content": "Hanzhuo Tan, Qi Luo, Jing Li, and Yuqun Zhang. 2024. Llm4decompile: Decompiling binary code with large language models. Preprint, arXiv:2403.05286."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.198,
+                0.882,
+                0.277
+            ],
+            "angle": 0,
+            "content": "Tao Wei, Jian Mao, Wei Zou, and Yu Chen. 2007. A new algorithm for identifying loops in decomposition. In Static Analysis, 14th International Symposium, SAS 2007, Kongens Lyngby, Denmark, August 22-24, 2007, Proceedings, volume 4634 of Lecture Notes in Computer Science, pages 170-183. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.287,
+                0.882,
+                0.364
+            ],
+            "angle": 0,
+            "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. CoRR, abs/1910.03771."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.375,
+                0.882,
+                0.428
+            ],
+            "angle": 0,
+            "content": "Wai Kin Wong, Huaijin Wang, Zongjie Li, Zhibo Liu, Shuai Wang, Qiyi Tang, Sen Nie, and Shi Wu. 2023. Refining decomposed C code with large language models. CoRR, abs/2310.06530."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.438,
+                0.882,
+                0.53
+            ],
+            "angle": 0,
+            "content": "Frank F. Xu, Uri Alon, Graham Neubig, and Vincent Josua Hellendoorn. 2022. A systematic evaluation of large language models of code. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, MAPS 2022, page 1-10, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.54,
+                0.882,
+                0.606
+            ],
+            "angle": 0,
+            "content": "Xiangzhe Xu, Zhuo Zhang, Shiwei Feng, Yapeng Ye, Zian Su, Nan Jiang, Siyuan Cheng, Lin Tan, and Xiangyu Zhang. 2023. Lmpa: Improving decomposition by synergy of large language model and program analysis. CoRR, abs/2306.02546."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.616,
+                0.882,
+                0.708
+            ],
+            "angle": 0,
+            "content": "Zhengran Zeng, Hanzhuo Tan, Haotian Zhang, Jing Li, Yuqun Zhang, and Lingming Zhang. 2022. An extensive study on pre-trained models for program understanding and generation. In ISSTA '22: 31st ACM SIGSOFT International Symposium on Software Testing and Analysis, Virtual Event, South Korea, July 18 - 22, 2022, pages 39-51. ACM."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.718,
+                0.882,
+                0.77
+            ],
+            "angle": 0,
+            "content": "Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of \\(100+\\) language models. arXiv preprint arXiv:2403.13372."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.77
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6612"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.315,
+                0.101
+            ],
+            "angle": 0,
+            "content": "A Training examples"
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.126,
+                0.125,
+                0.437,
+                0.259
+            ],
+            "angle": 0,
+            "content": "User: # This is the assembly: <func0>:\nmov %edi,%eax\nand $0x1,%eax\nret\n# What is the source code?\nAssistant: bool func0(int num) {\nreturn num % 2 != 0;\n}"
+        },
+        {
+            "type": "code_caption",
+            "bbox": [
+                0.271,
+                0.273,
+                0.724,
+                0.288
+            ],
+            "angle": 0,
+            "content": "Figure 6: The training example of end-to-end decompile objective."
+        },
+        {
+            "type": "code",
+            "bbox": [
+                0.129,
+                0.319,
+                0.814,
+                0.66
+            ],
+            "angle": 0,
+            "content": "User: Add inline comments that contain the corresponding C code for the assembly code first, then write the equivalent C function. ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ``  \n`` `` `````````````````````````````````````````````````````````````````````````````````````````````````````  \nAssistant: Here is the assembly code with inline comments containing the corresponding C code:  \n``` ` ` ` ` ` ` ` \n```\n<func0>:\n  mov %edi,%eax\n  and $0x1,%eax\n  ret\n```\nAssistant: Here is the assembly code with inline comments containing the corresponding C code:  \n``` ` ` ` ` ` \n```\n<func0>:\n  ; bool func0(int num) {\n    mov %edi,%eax\n    and $0x1,%eax\n  ; return num % 2 != 0;\n  ret\n}\n```\nAnd here is the equivalent C function:  \n``` c\nbool func0(int num) {\n  return num % 2 != 0;\n}"
+        },
+        {
+            "type": "code_caption",
+            "bbox": [
+                0.119,
+                0.683,
+                0.877,
+                0.699
+            ],
+            "angle": 0,
+            "content": "Figure 7: The training example of step-by-step decompile objective. Note that the workflow is only for training."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.726,
+                0.271,
+                0.74
+            ],
+            "angle": 0,
+            "content": "B Extra Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.751,
+                0.885,
+                0.896
+            ],
+            "angle": 0,
+            "content": "The Model is More Sensitive to Compiler Series than to Versions: As the Table 4 shows, using a compiler different from the one used during training to build the context can lead to significant performance degradation. The model performs better when the context is built using the GCC compiler series, whereas the performance is slightly worse when using the clang series. This may be due to the fact that the context used during both the continued pre-training and fine-tuning stages was constructed with the GCC compiler series. When different versions of GCC are used for construction, the performance of \\(\\mathrm{sc^2dec}\\) stabilizes around \\(52\\%\\) for the fine-tuned model except \"x86-64 gcc 14.1\". In contrast, when the context is built using the clang series, the performance of \\(\\mathrm{sc^2dec}\\) is slightly lower and more consistent, stabilizing around \\(51\\%\\) for the fine-tuned model."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6613"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.155,
+                0.082,
+                0.846,
+                0.273
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Compiler</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>x86-64 gcc 10.5</td><td>92.07</td><td>90.98</td><td>91.71</td><td>92.32</td><td>91.77</td><td>70.85</td><td>47.32</td><td>47.80</td><td>43.05</td><td>52.26</td></tr><tr><td>x86-64 gcc 11.4</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>48.66</td><td>47.68</td><td>42.68</td><td>52.47</td></tr><tr><td>x86-64 gcc 12.3</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>47.68</td><td>46.95</td><td>42.56</td><td>52.01</td></tr><tr><td>x86-64 gcc 13.3</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>47.93</td><td>46.59</td><td>42.68</td><td>52.01</td></tr><tr><td>x86-64 gcc 14.1</td><td>92.07</td><td>91.83</td><td>92.44</td><td>92.56</td><td>92.23</td><td>70.85</td><td>47.07</td><td>45.73</td><td>41.95</td><td>51.40</td></tr><tr><td>x86-64 clang 14.0.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.05</td><td>45.73</td><td>42.44</td><td>51.31</td></tr><tr><td>x86-64 clang 15.0.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.29</td><td>45.73</td><td>42.32</td><td>51.34</td></tr><tr><td>x86-64 clang 16.0.0</td><td>92.68</td><td>92.2</td><td>93.66</td><td>93.17</td><td>92.93</td><td>69.02</td><td>48.05</td><td>46.46</td><td>42.32</td><td>51.46</td></tr><tr><td>x86-64 clang 17.0.1</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.05</td><td>46.71</td><td>42.32</td><td>51.52</td></tr><tr><td>x86-64 clang 18.1.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.17</td><td>46.46</td><td>42.32</td><td>51.49</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.283,
+                0.884,
+                0.341
+            ],
+            "angle": 0,
+            "content": "Table 4: The re-executability of differnet models across different compilers. We selected the popular three versions of both the clang and gcc series in Compiler Explorer. The compiler of Compiler Explorer may differ from the compilers provided by Ubuntu in some default compilation options, which might result in slight differences in the generated assembly code."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.153,
+                0.349,
+                0.847,
+                0.865
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"5\">Re-Compilability</td><td colspan=\"5\">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td colspan=\"11\">GPT-4 (OpenAI, 2023)</td></tr><tr><td>vanilla</td><td>76.83</td><td>67.68</td><td>67.07</td><td>57.93</td><td>67.38</td><td>20.73</td><td>13.41</td><td>13.41</td><td>10.98</td><td>14.63</td></tr><tr><td>+sc2dec</td><td>76.22</td><td>67.03</td><td>65.24</td><td>56.10</td><td>66.15</td><td>32.93</td><td>20.12</td><td>19.51</td><td>13.41</td><td>21.49</td></tr><tr><td>1-shot</td><td>76.22</td><td>78.05</td><td>73.78</td><td>70.73</td><td>74.70</td><td>31.71</td><td>20.12</td><td>19.51</td><td>21.34</td><td>23.17</td></tr><tr><td>+sc2dec</td><td>74.39</td><td>76.22</td><td>71.95</td><td>68.90</td><td>72.87</td><td>40.24</td><td>25.61</td><td>23.78</td><td>23.17</td><td>28.20</td></tr><tr><td colspan=\"11\">GPT-3.5 Turbo (OpenAI, 2022)</td></tr><tr><td>vanilla</td><td>20.73</td><td>25.00</td><td>29.27</td><td>21.95</td><td>24.24</td><td>4.27</td><td>3.66</td><td>3.05</td><td>3.05</td><td>3.51</td></tr><tr><td>+sc2dec</td><td>18.29</td><td>22.56</td><td>28.66</td><td>20.12</td><td>22.41</td><td>5.49</td><td>3.66</td><td>3.66</td><td>4.27</td><td>4.27</td></tr><tr><td>1-shot</td><td>49.39</td><td>46.34</td><td>42.07</td><td>40.85</td><td>44.66</td><td>10.37</td><td>7.32</td><td>6.10</td><td>5.49</td><td>7.31</td></tr><tr><td>+sc2dec</td><td>42.07</td><td>42.68</td><td>36.59</td><td>34.15</td><td>38.87</td><td>12.20</td><td>9.15</td><td>6.71</td><td>7.93</td><td>8.99</td></tr><tr><td colspan=\"11\">Deepseek Chat (DeepSeek-AI, 2024)</td></tr><tr><td>vanilla</td><td>62.20</td><td>51.83</td><td>47.56</td><td>45.73</td><td>51.83</td><td>10.37</td><td>5.49</td><td>4.88</td><td>5.49</td><td>6.55</td></tr><tr><td>+sc2dec</td><td>57.93</td><td>50.00</td><td>46.95</td><td>45.12</td><td>50.00</td><td>13.41</td><td>6.71</td><td>6.71</td><td>7.32</td><td>8.54</td></tr><tr><td>1-shot</td><td>67.68</td><td>70.73</td><td>64.63</td><td>56.71</td><td>64.94</td><td>13.41</td><td>10.37</td><td>7.93</td><td>7.32</td><td>9.76</td></tr><tr><td>+sc2dec</td><td>65.85</td><td>70.12</td><td>62.20</td><td>53.66</td><td>62.96</td><td>17.68</td><td>11.59</td><td>10.37</td><td>9.15</td><td>12.20</td></tr><tr><td colspan=\"11\">Deepseek Coder (Guo et al., 2024)</td></tr><tr><td>vanilla</td><td>66.46</td><td>57.93</td><td>47.56</td><td>51.83</td><td>55.95</td><td>9.15</td><td>6.10</td><td>6.71</td><td>6.98</td><td>7.01</td></tr><tr><td>+sc2dec</td><td>62.80</td><td>54.27</td><td>43.29</td><td>50.00</td><td>52.59</td><td>12.20</td><td>7.93</td><td>9.15</td><td>8.54</td><td>9.45</td></tr><tr><td>1-shot</td><td>57.93</td><td>53.66</td><td>52.44</td><td>48.17</td><td>53.05</td><td>11.59</td><td>10.37</td><td>8.54</td><td>7.93</td><td>9.60</td></tr><tr><td>+sc2dec</td><td>56.10</td><td>51.83</td><td>51.22</td><td>46.34</td><td>51.37</td><td>15.24</td><td>11.58</td><td>10.37</td><td>8.54</td><td>11.43</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b (Tan et al., 2024)</td></tr><tr><td>vanilla</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>92.68</td><td>92.80</td><td>90.12</td><td>93.41</td><td>92.26</td><td>71.34</td><td>47.56</td><td>45.98</td><td>41.71</td><td>51.65</td></tr><tr><td>1-shot</td><td>93.05</td><td>93.29</td><td>90.98</td><td>93.90</td><td>92.80</td><td>70.37</td><td>44.88</td><td>41.46</td><td>37.68</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>93.05</td><td>92.68</td><td>90.37</td><td>93.29</td><td>92.35</td><td>71.59</td><td>47.20</td><td>46.34</td><td>41.46</td><td>51.65</td></tr><tr><td colspan=\"11\">llm4decompile-6.7b + FAE</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.873,
+                0.884,
+                0.917
+            ],
+            "angle": 0,
+            "content": "Table 5: The extra results of different methods across four optimization levels (O0, O1, O2, O3), as well as their average scores (AVG). The results in bold represent the optimal performance, while those underlined indicate the second-best performance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "6614"
+        }
+    ]
+]

2024/Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement/45f4817a-44a7-499d-b884-1bdac2c1b3e5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18e27cb45ce2064a65eefa067ae96c8c4a9e1da612840566a40b90e742483159
+size 706962

2024/Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement/full.md

@@ -0,0 +1,327 @@
+# Self-Constructed Context Decomposition with Fined-grained Alignment Enhancement
+
+Yunlong Feng, Dechuan Teng, Yang Xu, Honglin Mu, Xiao Xu, Libo Qin, Qingfu Zhu*, Wanxiang Che
+
+Harbin Institute of Technology, China
+
+{ylfeng,dcteng,yxu,hlmu,xxu,lbqin,qfzhu,car}  $@$  ir.hit.edu.cn
+
+# Abstract
+
+Decompilation transforms compiled code back into a high-level programming language for analysis when source code is unavailable. Previous work has primarily focused on enhancing decomposition performance by increasing the scale of model parameters or training data for pre-training. Based on the characteristics of the decomposition task, we propose two methods: (1) Without fine-tuning, the Self-Constructed Context Decomposition  $(\mathrm{sc}^2\mathrm{dec})$  method recombines the LLM's decomposition results to construct pairs for in-context learning, helping the model improve decomposition performance. (2) Fine-grained Alignment Enhancement (FAE), which meticulously aligns assembly code with source code at the statement level by leveraging debugging information, is employed during the fine-tuning phase to achieve further improvements in decomposition. By integrating these two methods, we achieved a Re-Executability performance improvement of approximately  $3.90\%$  on the Decompile-Eval benchmark, establishing a new state-of-the-art performance of  $52.41\%$ . The code, data, and models are available at https://github.com/AlongWY/sccdec.
+
+# 1 Introduction
+
+Decompilation is the process of converting compiled machine code or bytecode back into a high-level programming language. This process is typically used to analyze how the software works when the source code is not accessible (Brumley et al., 2013; Katz et al., 2018; Hosseini and Dolan-Gavitt, 2022; Xu et al., 2023; Armengol-Estape et al., 2023; Jiang et al., 2023; Wong et al., 2023). Many tools have been developed for decompilation, like Ghidra (Ghidra, 2024) and IDA Pro (Hex-Rays, 2024). However, these tools often struggle to generate human-readable code. The main challenge in decompilation is that it's hard to fully reconstruct
+
+![](images/345d99328aa3b056e3a95f9fb636e4ed6b7204e8a9c7f04701ac4c4bd7362e29.jpg)  
+Figure 1: Pipeline of the decompilation. The input for decompilation tasks is typically assembly, with the source code invisible. Additionally, the source code obtained through decompilation usually does not exactly match the original code.
+
+the source code, especially details such as variable names (Lacomis et al., 2019) and primary structures (Wei et al., 2007), which are frequently lost during the compilation process.
+
+Recent advances in large language models (LLMs) have prompted researchers to view programming languages as distinct linguistic systems, utilizing pre-trained code LLMs to accomplish various coding tasks (Lippincott, 2020; Rozière et al., 2023; Guo et al., 2024; OpenAI, 2023). These models have demonstrated significant performance improvements over traditional techniques (Zeng et al., 2022; Xu et al., 2022), making it feasible to apply LLMs to decompilation challenges. For instance, Transformer-based models such as Slade (Armengol-Estapé et al., 2023) and Nova (Jiang et al., 2023) have shown potential in using language models to translate binary code back into more readable and structured source code. Recently, Tan et al. (2024) developed and released the first open-source LLM focused on decompilation, alongside the construction of the first
+
+decompilation benchmark to evaluate the decompilation capabilities of models. These efforts primarily treat the decompilation task as a translation task, training a robust decompilation model on extensive synthetic data (Hosseini and Dolan-Gavitt, 2022; Armengol-Estapé et al., 2023; Jiang et al., 2023; Tan et al., 2024).
+
+By analyzing the characteristics of decomposition tasks, we have found two features:
+
+- The decompiled code generated by the model is usually compileable.  
+- The compiler can produce external debugging information for the debugger.
+
+Based on the characteristics of the decomposition task, we propose two approaches: (1) Self-Constructed Context Decomposition  $(\mathrm{sc}^2\mathrm{dec})$ , a tuning-free approach, leverages the decomposition results from the model and specific compiler to derive ground-truth (assembly code, source code) pair closely related to the test sample. By incorporating the obtained pair as a demonstration example within the context, we can fully leverage the model's in-context learning capabilities to enhance performance. (2) Fine-grained Alignment Enhancement (FAE) introduces a new step-by-step decomposition training objective in addition to the end-to-end decomposition training objective that relies on large-scale pre-training for implicit alignment. By leveraging debugging information designed for debuggers, we can align assembly code with high-level code at a finer statement level. This allows the model to more faithfully preserve the original functionality after decomposition.
+
+Combining these two methods, we have achieved an approximately  $3.90\%$  improvement in re-executability performance based on lvm4decompile model (Tan et al., 2024), reaching a new state-of-the-art performance of  $52.41\%$  re-executability.
+
+To summarize, our contributions are as follows:
+
+- We introduce Self-Constructed Context Decomposition, which leverages the compileability of decomposition results to construct better contexts.  
+- We introduce Fine-grained Alignment Enhancement, which fine-tunes the model using fine-grained alignment data between assembly code and source code derived from debug information. We also propose how to synthesize its training data automatically.
+
+- The experiments on the Decompile-Eval benchmark show that our proposed model achieves new state-of-the-art results.
+
+# 2 Related Works
+
+Decompilation has a long history, centered on converting binary files into human-readable source code. Traditional tools such as Hex-Rays IDA Pro (Hex-Rays, 2024) and Ghidra (Ghidra, 2024) predominantly rely on analyzing a program's control flow graph (CFG) to perform this task (Brumley et al., 2013). These tools operate by scrutinizing instructions within assembly code to construct CFGs and identify common programming structures like if-else statements and loops (Wei et al., 2007). However, they rely on intricate and error-prone expert-built rule systems that frequently fail with optimized binary code, a common practice in commercial software development. Moreover, traditional tools often generate outputs that closely resemble assembly code representations, translating variables into registers and utilizing low-level operations such as goto statements. This output is not only challenging to comprehend but also occasionally cannot be recompiled (Liu and Wang, 2020).
+
+Inspired by neural machine translation, researchers have begun to conceptualize decompilation as a translation task, wherein machine-level instructions are converted into readable source code using neural networks. Initial endeavors employing recurrent neural networks (RNNs) demonstrated limited success (Katz et al., 2018). However, recent advancements in natural language processing (NLP) have facilitated the application of pre-trained language models (LMs) in decompilation. Notable examples include BTC (Hosseini and Dolan-Gavitt, 2022), Slade (Armengol-Estapé et al., 2023), and Nova (Jiang et al., 2023), which mark significant strides in the field. Recently, Tan et al. (2024) created and released the first open-source large language models specifically for decompilation and established an evaluation benchmark that considers the re-compilability and re-executability of decompiled code.
+
+# 3 Method
+
+In this section, we introduce the Self-Constructed Context Decomposition method and Fine-grained Alignment Enhancement fine-tuning. Figures 2 and 3 illustrate the methods.
+
+![](images/4de304b57cea5f28244a505f652396ec27592f53b88121741ae9476006cb6dbc.jpg)  
+Figure 2: The pipeline for Self-Constructed Context Decomposition operates as follows: when the LLM decomposes and generates compilable code, we compile this code to construct the context, and then use it to decompile the initial assembly code again.
+
+# 3.1 Self-Constructed Context Decomposition
+
+In this section, we introduce the Self-Constructed Context Decomposition (sc²dec) method designed to enhance the performance of decomposition models. The method recombines the decomposed code generated by the model to construct the context. The sc²dec method involves the following key steps, as the Figure 2 shows:
+
+1. Decompile: Initially, the model decomposes the assembly code back to the target programming language. We call it the initial decomposition result.  
+2. Compile and Disassembly: Then we recompile the initial decomposition result to obtain the corresponding assembly code if the generated code is compileable.  
+3. Constructed Context: The self-constructed context is then formed by concatenating the generated C code with its corresponding assembly code.  
+4. Decompile Again: Decompile the assembly code with the self-constructed context again.
+
+In this context, the assembly code precedes the C code, resembling an example in the context. The final input to the model is a concatenation of the self-constructed context (with assembly code first and C code second) and the original assembly code. This
+
+![](images/8175bcafcd20bbbaf885b8b49ea7224062f0c5761bcc4fa9a6763924b75ed0e1.jpg)  
+Figure 3: An example of step-by-step decomposition is presented, where the training objective requires the model to generate C code progressively after each assembly block. Fine-tuning the model with this objective aids in learning the fine-grained correspondences between assembly and C code.
+
+method leverages the additional context provided by the generated C code and its corresponding assembly code to improve the overall decompilation performance.
+
+# 3.2 Fine-grained Alignment Enhancement
+
+To enhance the model's ability to perceive the correspondence between assembly code and its corresponding source code, we propose a Fine-grained Alignment Enhancement (FAE) method, which constructs fine-grained alignment data between assembly code and source code derived from debug information. The method primarily includes two training objectives: the End-to-end Decomposition Objective and the Step-by-Step Decomposition Objective.
+
+- End-to-end Decomposition Objective: In the end-to-end decomposition task, the model takes assembly code as input and directly generates the corresponding decomposed C code. The goal of this task is to enable the model to translate from assembly language to C language without any intermediate steps. This ob
+
+jective aims to prevent catastrophic forgetting in the model. During training, we increase the context length and ensure that the same sample is not truncated, which enhances the model's ability to align assembly and C code.
+
+- Step-by-step Decomposition Objective: As the Figure 3 shows, in the step-by-step decomposition process, an intermediate form containing both the assembly code and the corresponding C code is first generated, followed by the generation of the complete C code. This step-by-step approach allows the model to better understand the correspondence between assembly code and C code. Training the model with the step-by-step decomposition objective can enhance the model's ability to finely align assembly code with C code. It is important to note that this form is used only during training and not during inference.
+
+During the training process, we employed two different training objectives to enhance the model's decompile performance: an end-to-end decompile objective and a step-by-step decompile objective. They all take assembly code as input, but their output formats are different.
+
+# 3.2.1 Data Processing
+
+By following the steps below, we can synthesize some fine-grained alignment training data. The synthesized data, as shown in Figure 3, is in a format of interleaved assembly and code output. We expect the model to learn finer-grained alignment between assembly and code from such data.
+
+1. Compilation: We employed the "gcc" compiler to compile the selected function code into binary shared libraries at various optimization levels (including O0, O1, O2, and O3). The debug option was enabled during compilation to ensure the inclusion of DWARF debugging information in the compiled output. This debugging information is crucial for subsequent disassembly and parsing steps.
+
+2. Disassembly: Subsequently, we used the "objdump" tool to disassemble the compiled binary shared libraries. The "-S" option was utilized to include source code in the disassembly output. Additionally, we used the "-source/comment=;," option toptides each line of source code with a ";," symbol, facilitating easier parsing in later stages.
+
+3. Data Parsing: The disassembly process yielded a mixed output containing both assembly code and source code, organized in an alternating sequence of source code followed by assembly code. We parsed the output to extract the corresponding assembly and source code for each function, ensuring accurate alignment between the two.  
+4. Data Reorganization: Finally, we reorganized the parsed content into a format where the assembly code precedes the corresponding source code in an alternating sequence.
+
+Through these steps, we constructed a high-quality training dataset that ensures each sample contains comprehensive source code information along with the corresponding assembly code. This dataset will be used to train our model, aiming to enhance its performance in tasks such as code understanding and compilation optimization.
+
+# 4 Experiments
+
+# 4.1 Evaluation
+
+In this section, we will introduce the benchmark used in our evaluation process, Decompile-Eval (Tan et al., 2024), which is specifically designed to assess the decompilation capabilities of large language models.
+
+The Decompile-Eval benchmark is adapted from the HumanEval benchmark, which includes 164 problems initially designed for code generation tasks. These problems were translated into the C programming language, and the corresponding assembly code was generated at four optimization levels (O0, O1, O2, and O3). The correctness of the decomposition results was tested using the test cases from HumanEval. The primary metrics of the Decompile-Eval benchmark are as follows:
+
+- Re-compilability: This metric evaluates whether the decompiled code produced by our model can be successfully recompiled into executable binary without errors. A high recompiliability rate indicates that the decompiled code is syntactically correct and adheres to the constraints of the target language (in this case, C).  
+- Re-executability: This metric assesses the functional correctness of the decompiled code. Specifically, it measures whether the recom-piled binaries produce the expected outputs when executed. The correctness of the output
+
+is determined using the testing methodology provided by the HumanEval dataset, ensuring a comprehensive evaluation of the logical accuracy of the decompiled code.
+
+In our experiment, we pay more attention to the Re-executability, as it more accurately reflects the overall decompilation capability of the model.
+
+# 4.2 Training Details
+
+Implementation We utilize Lora (Hu et al., 2022) to fine-tune the llm4decompile-6.7b obtained on Hugging Face (Wolf et al., 2019), with rank set to 32, alpha set to 64 and target set to all projection layers  $^{1}$ . The optimizer is AdamW (Loshchilov and Hutter, 2019), with a learning rate of 5e-5. The maximum sequence length is set to 16384, and the learning rate scheduler type is cosine, with a warm-up period of 20 steps. The training process was conducted for one epoch. LlamaFactory (Zheng et al., 2024) and FlashAttention 2 (Dao, 2024) were used for the fine-tuning of the model. All experiments were done with an A100-SXM4-80GB. We use greedy decoding for all experiments.
+
+Training data We selected 10,000 samples from the train_real_compilable subset of Exebench (Armengol-Estapé et al., 2022) to synthesis the training data. The selected functions exclusively utilize the standard C library and do not include additional data structures. The training data were synthesized with gcc 11.4 provided by Ubuntu 22.04.
+
+# 4.3 Baselines
+
+We present several baselines in our experiments and demonstrate the effectiveness of our proposed methods.
+
+# 4.3.1 Models
+
+- GPT 4: One of the most powerful OpenAI models, known for its advanced language understanding and generation capabilities (OpenAI, 2023).  
+- deepseek-chat: A powerful conversational AI model from DeepSeek, excelling in generating coherent and contextually relevant dialogues (DeepSeek-AI, 2024).  
+- llm4decompile-6.7b: The 6.7B open-access decomposition LLM pre-trained on 15 billion
+
+```txt
+bool func0(int num){ if (num <= 1) { return false; } for (int i = 2; i * i <= num; i++) { if (num % i == 0) { return false; } } return true; }
+```
+
+Figure 4: The example for 1-shot learning. In this example, we have tried to cover common control logic such as if statements, loops, and early returns as much as possible. The example will be compiled with the same optimization level as the target assembly code.
+
+tokens of C source code and the corresponding assembly code (Tan et al., 2024) $^2$ .
+
+- llm4decompile-6.7b+FAE: The new model was obtained by applying the Fine-grained Alignment Enhancement method for further fine-tuning based on the llm4decompile-6.7b $^2$ .
+
+# 4.3.2 Methods
+
+- vanilla: Using specific prompts for different models to directly request decomposition.
+
+Prompt For GPT-4 and Deepseek Chat  
+```txt
+What is the C source code of  $\leftrightarrow$  the assembly code below: ..asm <asm>
+```
+
+Prompt For 1lm4decompile  
+```txt
+This is the assembly code: <asm>
+```
+
+```txt
+What is the source code?
+```
+
+- 1-shot: We write a fixed example as the context (Brown et al., 2020). The example will be compiled with the same optimization level as the target assembly code. The example shown in Figure 4. In this example, we have tried to cover common control logic such as if statements, loops, and early returns as much as possible.
+
+<table><tr><td rowspan="2">Method</td><td colspan="5">Re-Compilability</td><td colspan="5">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td colspan="11">GPT-4 (OpenAI, 2023)</td></tr><tr><td>vanilla</td><td>76.83</td><td>67.68</td><td>67.07</td><td>57.93</td><td>67.38</td><td>20.73</td><td>13.41</td><td>13.41</td><td>10.98</td><td>14.63</td></tr><tr><td>+sc2dec</td><td>76.22</td><td>67.03</td><td>65.24</td><td>56.10</td><td>66.15</td><td>32.93</td><td>20.12</td><td>19.51</td><td>13.41</td><td>21.49</td></tr><tr><td>1-shot</td><td>76.22</td><td>78.05</td><td>73.78</td><td>70.73</td><td>74.70</td><td>31.71</td><td>20.12</td><td>19.51</td><td>21.34</td><td>23.17</td></tr><tr><td>+sc2dec</td><td>74.39</td><td>76.22</td><td>71.95</td><td>68.90</td><td>72.87</td><td>40.24</td><td>25.61</td><td>23.78</td><td>23.17</td><td>28.20</td></tr><tr><td colspan="11">Deepseek Chat (DeepSeek-AI, 2024)</td></tr><tr><td>vanilla</td><td>62.20</td><td>51.83</td><td>47.56</td><td>45.73</td><td>51.83</td><td>10.37</td><td>5.49</td><td>4.88</td><td>5.49</td><td>6.55</td></tr><tr><td>+sc2dec</td><td>57.93</td><td>50.00</td><td>46.95</td><td>45.12</td><td>50.00</td><td>13.41</td><td>6.71</td><td>6.71</td><td>7.32</td><td>8.54</td></tr><tr><td>1-shot</td><td>67.68</td><td>70.73</td><td>64.63</td><td>56.71</td><td>64.94</td><td>13.41</td><td>10.37</td><td>7.93</td><td>7.32</td><td>9.76</td></tr><tr><td>+sc2dec</td><td>65.85</td><td>70.12</td><td>62.20</td><td>53.66</td><td>62.96</td><td>17.68</td><td>11.59</td><td>10.37</td><td>9.15</td><td>12.20</td></tr><tr><td colspan="11">llm4decompile-6.7b (Tan et al., 2024)</td></tr><tr><td>vanilla</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>92.68</td><td>92.80</td><td>90.12</td><td>93.41</td><td>92.26</td><td>71.34</td><td>47.56</td><td>45.98</td><td>41.71</td><td>51.65</td></tr><tr><td>1-shot</td><td>93.05</td><td>93.29</td><td>90.98</td><td>93.90</td><td>92.80</td><td>70.37</td><td>44.88</td><td>41.46</td><td>37.68</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>93.05</td><td>92.68</td><td>90.37</td><td>93.29</td><td>92.35</td><td>71.59</td><td>47.20</td><td>46.34</td><td>41.46</td><td>51.65</td></tr><tr><td>retrieval</td><td>94.02</td><td>88.05</td><td>84.51</td><td>85.12</td><td>87.93</td><td>65.73</td><td>32.68</td><td>38.66</td><td>36.22</td><td>43.32</td></tr><tr><td colspan="11">llm4decompile-6.7b + FAE</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr></table>
+
+Table 1: The main results of different methods across four optimization levels (O0, O1, O2, O3), as well as their average scores (AVG). The results in bold represent the optimal performance, while those underlined indicate the second-best performance. More results can be found in the Table 5.
+
+- retrieval: We use the retrieval method (BM25 (Lù, 2024)) to get the context. The context is generated by searching for the most similar assembly code in the training set.  
+-  $\mathbf{sc}^2\mathbf{dec}$ : The self-constructed context decomposition we described in the Section 3.1.  
+- 1-shot+sc²dec: We apply the sc²dec on the 1-shot method. Note that the 1-shot example can only be applied to generate the initial decomposition result in Figure 2.
+
+# 4.4 Results
+
+In this section, we present and analyze the performance of different models. Our experiments aim to evaluate the impact of various techniques, including one-shot, self-constructed context decomposition, and fine-tuning, on the model's performance. The main results are shown in Table 1.
+
+Performance of the Base Model Through a comparative analysis of the experimental results in Ta
+
+ble 1, it is evident that different models exhibit significant differences in re-executability across various optimization levels (O0, O1, O2, O3). The base model, llm4decompile-6.7b, performs best at the O0 optimization level with a score of  $70.24\%$ , but its performance gradually declines at other optimization levels, with an average score of  $48.60\%$ . Among models not fine-tuned with decompilation tasks, GPT-4 achieved the best performance, with an average re-executability score of  $14.64\%$ . Generally, their results on re-compilability metrics are significantly higher than their performance on re-executability metrics.
+
+Context Sensitivity By comparing vanilla, 1-shot, and retrieval methods based on the llm4decompile-6.7b, it is observed that the model's performance can even decline after providing a sample. In contrast, the dynamic samples generated through our self-constructed context method yield an improvement in executability of over  $2\%$
+
+<table><tr><td rowspan="2">Method</td><td colspan="5">Re-Compilability</td><td colspan="5">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>sc2dec (O0)</td><td>92.68</td><td>92.20</td><td>91.71</td><td>92.32</td><td>92.23</td><td>70.24</td><td>48.90</td><td>47.68</td><td>42.93</td><td>51.83</td></tr><tr><td>sc2dec (O1)</td><td>92.68</td><td>91.59</td><td>93.54</td><td>92.93</td><td>92.68</td><td>68.41</td><td>48.54</td><td>47.44</td><td>43.05</td><td>51.86</td></tr><tr><td>sc2dec (O2)</td><td>92.68</td><td>92.20</td><td>91.71</td><td>92.32</td><td>92.23</td><td>67.80</td><td>48.90</td><td>47.68</td><td>42.93</td><td>51.68</td></tr><tr><td>sc2dec (O3)</td><td>92.68</td><td>91.59</td><td>92.93</td><td>92.32</td><td>92.38</td><td>67.80</td><td>47.07</td><td>47.80</td><td>43.41</td><td>51.52</td></tr></table>
+
+Table 2: The results of our method across optimization levels. For cases such as "sc²dec (O0)", we assume that the optimization level of asm code is entirely unknown and select a specific optimization level to compile the code to construct the context.
+
+```c
+int func0(int a){ for (int i = 0; i * i * i <= abs(a); i++) if (i * i * i == abs(a)) return 1; return 0; } Source code   
+int func0(int n){ int i; if  $(n <   0)$  n  $=$  -n; for  $(\mathrm{i} = 1;\mathrm{i}\ast \mathrm{i}\ast \mathrm{i} <   = \mathrm{n};\mathrm{i} + + )$  if  $(i*\bar{i}*i==n)$  return 1; return 0; } Decompile   
+int func0(int n){ int i; if  $(n <   0)$  n  $=$  -n; if  $(n = = 0)$  return 1; for  $(\mathrm{i} = 1;\mathrm{i}\ast \mathrm{i}\ast \mathrm{i} <   = \mathrm{n};\mathrm{i} + + )$  if  $(i*\bar{i}*i==n)$  return 1; return 0; } SC2Dec
+```
+
+Figure 5: A case study for  $\mathrm{sc^2dec}$ , which is based on the llm4decompile-6.7b with FAE. By applying the  $\mathrm{sc^2dec}$  method, the model detects and fixes the mismatch between the code and assembly.
+
+before and after fine-tuning. This indicates that inappropriate samples may even have a detrimental effect in the decompilation domain, while our method achieves a nearly stable performance enhancement without the need for retrieval.
+
+1-shot is a Strong Baseline As shown in Table 1, utilizing the 1-shot method significantly enhances the model's performance. Specifically, for GPT-4, the model's re-executability performance increased from  $14.64\%$  to  $23.17\%$ , an approximate improvement of  $7\%$ . In the open-source model Deepseek Chat, performance improved from  $6.55\%$  to  $9.76\%$ ,
+
+reflecting a  $3.21\%$  enhancement. However, in the case of the llm4decompile model, which had already been trained on decompilation tasks, there was no performance improvement. Moreover, in the llm4decompile model fine-tuned with Fine-grained Alignment Enhancement, performance actually decreased from  $50.88\%$  to  $50.61\%$ , a reduction of approximately  $0.27\%$ .
+
+Sc2dec Brings Further Performance Improvement The Table 1 shows that applying sc2dec independently results in performance improvements of  $6.86\%$  and  $1.99\%$  on GPT-4 and Deepseek Chat, respectively. These improvements are lower than those achieved by the 1-shot method, likely due to the model's low Re-Compilability metric. In other words, a substantial portion of the code obtained through decompilation is un-compilable, which impedes the application of the sc2dec method. Furthermore, when combined with the 1-shot method, sc2dec can further enhance performance, achieving an additional  $5\%$  improvement for GPT-4, bringing it to  $28.20\%$ , and a  $2.44\%$  improvement for Deepseek Chat, bringing it to  $12.20\%$ . In the case of llm4decompile-6.7b, due to the models' relatively high Re-Compilability metric, the sc2dec method significantly outperformed the 1-shot method, increasing performance from  $48.60\%$  to  $51.65\%$  and from  $50.88\%$  to  $52.41\%$ .
+
+Best Performance with Combined Methods As the Table 1 shows, by combining our Step-by-Step Decompile tasks and context-based decomposition methods, we ultimately achieved approximately a  $3.90\%$  performance improvement, surpassing the performance of these methods when applied independently. This demonstrates that fine-tuning and self-construct context methods are orthogonal. Most notably, the combined method of fined
+
+<table><tr><td rowspan="2">Model</td><td colspan="5">Re-Compilability</td><td colspan="5">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>w/o tune</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td colspan="11">FAE (End-to-End &amp; Step-by-Step Decomposition Objective)</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr><tr><td colspan="11">End-to-End Decomposition Objective</td></tr><tr><td>vanilla</td><td>93.9</td><td>90.61</td><td>90.61</td><td>91.34</td><td>91.62</td><td>69.51</td><td>43.29</td><td>42.2</td><td>40.73</td><td>48.93</td></tr><tr><td>+sc2dec</td><td>92.8</td><td>90.24</td><td>90.00</td><td>90.12</td><td>90.79</td><td>71.34</td><td>46.10</td><td>43.78</td><td>42.44</td><td>50.91</td></tr><tr><td>1-shot</td><td>93.9</td><td>90.73</td><td>90.37</td><td>91.22</td><td>91.55</td><td>69.51</td><td>44.02</td><td>42.68</td><td>40.61</td><td>49.21</td></tr><tr><td>+sc2dec</td><td>92.8</td><td>90.24</td><td>89.76</td><td>90.00</td><td>90.7</td><td>71.22</td><td>46.46</td><td>44.27</td><td>42.56</td><td>51.13</td></tr></table>
+
+Table 3: The results of ablation study. All results are based on the llm4decompile-6.7b model.
+
+grained alignment enhancement and self-construct context decomposition achieves the highest scores across all optimization levels expect "O0", with scores of  $48.54\%$ ,  $47.56\%$ , and  $43.29\%$  at the O1, O2, and O3 levels, respectively, and an overall average score of  $52.41\%$ .
+
+# 4.5 Analysis
+
+In this section, we will further analyze the experimental results to study the effectiveness and sensitivity of our method:
+
+Mismatched Optimization Levels Leads to Performance Degradation As the Table 2 shows, the performance can degrade by  $1\%$  to  $2\%$  overall when constructing contexts that mismatch the optimization levels of the target assembly code. However, specific optimization levels may exhibit different behaviors. For instance, using a fixed O3 optimization level to construct contexts while decompiling assembly optimized with O2 does not result in significant performance degradation and even better. In practical scenarios, using O3 to construct contexts for binaries with unknown optimization levels might be a reasonable choice.
+
+Case study of  $\mathbf{sc}^2\mathbf{dec}$  As the Figure 5 shows, the compiler has expanded the abs function, reducing the overhead of function calls. The model initially generated a piece of compilable but incorrect code during direct decompilation, omitting the case where the variable  $n$  equals 0. We compiled and disassembled this code to construct a new sample. By applying the  $\mathrm{sc^2dec}$  method with the context, the model successfully identified the missing part
+
+in the short segment of assembly code and generated the correct code, thus rectifying the error from the previous inference.
+
+Contribution of FAE The results in Table 1 show that the application of Fine-grained Alignment Enhancement (FAE) for further fine-tuning the model significantly enhanced its re-executability performance across all optimization levels except "O0". Notably, at optimization levels O2 and O3, the scores improved by  $5.12\%$  and  $4.27\%$ , reaching  $45.73\%$  and  $42.32\%$ , respectively. The average performance increased by  $2.28\%$ , achieving  $50.88\%$ . This demonstrates the effectiveness of the Fine-grained Alignment Enhancement method.
+
+Ablation Study of FAE The results of ablation experiments in Table 3 have demonstrated the effectiveness of our proposed Step-by-Step Decompile training objective. Removing the training objective results in approximately a  $2\%$  decline in Re-Executability performance, regardless of whether the  $\mathrm{sc}^2$  dec method is used, and notably, a  $3.78\%$  decline at the "O2" optimization level. However, the performance of the 1-shot remains unaffected, supporting our hypothesis that the mismatch between the training data form and the inference form leads to performance degradation. Furthermore, retaining only the End-to-end Decompilation training objective still yields little performance improvement, about  $0.33\%$ . This improvement can be attributed to the longer context provided during training, as the samples are not truncated during training, allowing the model to fully observe the assembly code and its corresponding C code.
+
+# 5 Conclusion
+
+In this paper, we introduce two methods based on the characteristics of the decompilation task: Self-Constructed Context Decomposition, which leverages the compileability of decomposition results to construct better contexts, and Fine-grained Alignment Enhancement, which fine-tunes the model using fine-grained alignment data between assembly code and source code derived from debug information. By integrating these methods, we achieved a  $3.90\%$  improvement in decompilation performance, reaching a new state-of-the-art level of  $52.41\%$ .
+
+# Acknowledge
+
+We gratefully acknowledge the support of the National Natural Science Foundation of China (NSFC) via grant 62236004, 62206078, 62441603 and 62476073.
+
+# Limitations
+
+The  $\mathrm{sc^2dec}$  depends on the decomposed code generated by the model. If the decomposed code cannot be recompiled,  $\mathrm{sc^2dec}$  will not be able to benefit from it, necessitating that the decomposed code has high re-compilability performance. For Fine-grained Alignment Enhancement, we created a Step-by-Step Decompile training objective, training with only 10,000 samples across four optimization levels (O0-O3), without validating its performance on larger datasets or bigger models. Moreover, fine-tuning led to a further decline in performance in the 1-shot scenario.
+
+# Potential risks
+
+Decompilation is a technique that converts compiled binary code back into human-readable source code. While decompilation can be legal and useful in certain contexts, it also entails several potential risks. The primary potential risks are as follows:
+
+- Intellectual Property Infringement: Decompilation may violate the copyright and licensing agreements of software. Unauthorized decomposition can lead to copyright infringement, thereby instigating legal disputes.  
+- Security Risks: Decomposed code may expose the internal structure and implementation details of the software, providing attack vectors for hackers. Malicious actors can exploit this information to identify and exploit vulnerabilities in the software.
+
+- Ethical Concerns: Decompiling and analyzing another's code can be considered unethical, especially when done without authorization.
+
+# References
+
+Jordi Armengol-Estapé, Jackson Woodruff, Alexander Brauckmann, José Wesley de Souza Magalhães, and Michael F. P. O’Boyle. 2022. Exebench: An ml-scale dataset of executable c functions. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, MAPS 2022, page 50–59, New York, NY, USA. Association for Computing Machinery.  
+Jordi Armengol-Estape, Jackson Woodruff, Chris Cummins, and Michael F. P. O'Boyle. 2023. Slade: A portable small language model decoder for optimized assembler. CoRR, abs/2305.12520.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. Preprint, arXiv:2005.14165.  
+David Brumley, JongHyup Lee, Edward J. Schwartz, and Maverick Woo. 2013. Native x86 decomposition using semantics-preserving structural analysis and iterative control-flow structuring. In Proceedings of the 22th USENIX Security Symposium, Washington, DC, USA, August 14-16, 2013, pages 353-368. USENIX Association.  
+Tri Dao. 2024. FlashAttention-2: Faster attention with better parallelism and work partitioning. In International Conference on Learning Representations (ICLR).  
+DeepSeek-AI. 2024. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. Preprint, arXiv:2405.04434.  
+Ghidra. 2024. Ghidra software reverse engineering framework.  
+Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. 2024. Deepseek-coder: When the large language model meets programming – the rise of code intelligence. Preprint, arXiv:2401.14196.  
+Hex-Rays. 2024. Ida pro: a cross-platform multiprocessor disassembler and debugger.
+
+Iman Hosseini and Brendan Dolan-Gavitt. 2022. Beyond the C: retargetable decompilation using neural machine translation. CoRR, abs/2212.08950.  
+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.  
+Nan Jiang, Chengxiao Wang, Kevin Liu, Xiangzhe Xu, Lin Tan, and Xiangyu Zhang. 2023. Nova+: Generative language models for binaries. Preprint, arXiv:2311.13721.  
+Deborah S. Katz, Jason Ruchti, and Eric M. Schulte. 2018. Using recurrent neural networks for decompilation. In 25th International Conference on Software Analysis, Evolution and Reengineering, SANER 2018, Campobasso, Italy, March 20-23, 2018, pages 346-356. IEEE Computer Society.  
+Jeremy Lacomis, Pengcheng Yin, Edward J. Schwartz, Miltiadis Allamanis, Claire Le Goues, Graham Neubig, and Bogdan Vasilescu. 2019. DIRE: A neural approach to decompiled identifier naming. In 34th IEEE/ACM International Conference on Automated Software Engineering, ASE 2019, San Diego, CA, USA, November 11-15, 2019, pages 628-639. IEEE.  
+Thomas Lippincott. 2020. Starcoder: A general neural ensemble technique to support traditional scholarship, illustrated with a study of the post-atlantic slave trade. In 15th Annual International Conference of the Alliance of Digital Humanities Organizations, DH 2020, Ottawa, Canada, July 20-25, 2020, Conference Abstracts.  
+Zhibo Liu and Shuai Wang. 2020. How far we have come: testing decomposition correctness of C decompliers. In ISSTA '20: 29th ACM SIGSOFT International Symposium on Software Testing and Analysis, Virtual Event, USA, July 18-22, 2020, pages 475-487. ACM.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.  
+Xing Han Lu. 2024. Bm25s: Orders of magnitude faster lexical search via eager sparse scoring. Preprint, arXiv:2407.03618.  
+OpenAI. 2022. ChatGPT.  
+OpenAI. 2023. GPT-4 Technical Report. CoRR, abs/2303.08774.  
+Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton-Ferrer, Aaron Grattaflori, Wenhan Xiong, Alexandre Défossez, Jade Copet,
+
+Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. 2023. Code llama: Open foundation models for code. CoRR, abs/2308.12950.  
+Hanzhuo Tan, Qi Luo, Jing Li, and Yuqun Zhang. 2024. Llm4decompile: Decompiling binary code with large language models. Preprint, arXiv:2403.05286.  
+Tao Wei, Jian Mao, Wei Zou, and Yu Chen. 2007. A new algorithm for identifying loops in decomposition. In Static Analysis, 14th International Symposium, SAS 2007, Kongens Lyngby, Denmark, August 22-24, 2007, Proceedings, volume 4634 of Lecture Notes in Computer Science, pages 170-183. Springer.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. CoRR, abs/1910.03771.  
+Wai Kin Wong, Huaijin Wang, Zongjie Li, Zhibo Liu, Shuai Wang, Qiyi Tang, Sen Nie, and Shi Wu. 2023. Refining decomposed C code with large language models. CoRR, abs/2310.06530.  
+Frank F. Xu, Uri Alon, Graham Neubig, and Vincent Josua Hellendoorn. 2022. A systematic evaluation of large language models of code. In Proceedings of the 6th ACM SIGPLAN International Symposium on Machine Programming, MAPS 2022, page 1-10, New York, NY, USA. Association for Computing Machinery.  
+Xiangzhe Xu, Zhuo Zhang, Shiwei Feng, Yapeng Ye, Zian Su, Nan Jiang, Siyuan Cheng, Lin Tan, and Xiangyu Zhang. 2023. Lmpa: Improving decomposition by synergy of large language model and program analysis. CoRR, abs/2306.02546.  
+Zhengran Zeng, Hanzhuo Tan, Haotian Zhang, Jing Li, Yuqun Zhang, and Lingming Zhang. 2022. An extensive study on pre-trained models for program understanding and generation. In ISSTA '22: 31st ACM SIGSOFT International Symposium on Software Testing and Analysis, Virtual Event, South Korea, July 18 - 22, 2022, pages 39-51. ACM.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. arXiv preprint arXiv:2403.13372.
+
+# A Training examples
+
+Figure 6: The training example of end-to-end decompile objective.  
+```txt
+User: # This is the assembly: <func0>:
+mov %edi,%eax
+and $0x1,%eax
+ret
+# What is the source code?
+Assistant: bool func0(int num) {
+return num % 2 != 0;
+}
+```
+
+Figure 7: The training example of step-by-step decompile objective. Note that the workflow is only for training.  
+```txt
+User: Add inline comments that contain the corresponding C code for the assembly code first, then write the equivalent C function. ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ` ``  
+`` `` `````````````````````````````````````````````````````````````````````````````````````````````````````  
+Assistant: Here is the assembly code with inline comments containing the corresponding C code:  
+``` ` ` ` ` ` ` ` 
+```
+<func0>:
+  mov %edi,%eax
+  and $0x1,%eax
+  ret
+```
+Assistant: Here is the assembly code with inline comments containing the corresponding C code:  
+``` ` ` ` ` ` 
+```
+<func0>:
+  ; bool func0(int num) {
+    mov %edi,%eax
+    and $0x1,%eax
+  ; return num % 2 != 0;
+  ret
+}
+```
+And here is the equivalent C function:  
+``` c
+bool func0(int num) {
+  return num % 2 != 0;
+}
+```
+
+# B Extra Results
+
+The Model is More Sensitive to Compiler Series than to Versions: As the Table 4 shows, using a compiler different from the one used during training to build the context can lead to significant performance degradation. The model performs better when the context is built using the GCC compiler series, whereas the performance is slightly worse when using the clang series. This may be due to the fact that the context used during both the continued pre-training and fine-tuning stages was constructed with the GCC compiler series. When different versions of GCC are used for construction, the performance of  $\mathrm{sc^2dec}$  stabilizes around  $52\%$  for the fine-tuned model except "x86-64 gcc 14.1". In contrast, when the context is built using the clang series, the performance of  $\mathrm{sc^2dec}$  is slightly lower and more consistent, stabilizing around  $51\%$  for the fine-tuned model.
+
+<table><tr><td rowspan="2">Compiler</td><td colspan="5">Re-Compilability</td><td colspan="5">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td>x86-64 gcc 10.5</td><td>92.07</td><td>90.98</td><td>91.71</td><td>92.32</td><td>91.77</td><td>70.85</td><td>47.32</td><td>47.80</td><td>43.05</td><td>52.26</td></tr><tr><td>x86-64 gcc 11.4</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>48.66</td><td>47.68</td><td>42.68</td><td>52.47</td></tr><tr><td>x86-64 gcc 12.3</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>47.68</td><td>46.95</td><td>42.56</td><td>52.01</td></tr><tr><td>x86-64 gcc 13.3</td><td>92.07</td><td>91.59</td><td>92.32</td><td>92.32</td><td>92.07</td><td>70.85</td><td>47.93</td><td>46.59</td><td>42.68</td><td>52.01</td></tr><tr><td>x86-64 gcc 14.1</td><td>92.07</td><td>91.83</td><td>92.44</td><td>92.56</td><td>92.23</td><td>70.85</td><td>47.07</td><td>45.73</td><td>41.95</td><td>51.40</td></tr><tr><td>x86-64 clang 14.0.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.05</td><td>45.73</td><td>42.44</td><td>51.31</td></tr><tr><td>x86-64 clang 15.0.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.29</td><td>45.73</td><td>42.32</td><td>51.34</td></tr><tr><td>x86-64 clang 16.0.0</td><td>92.68</td><td>92.2</td><td>93.66</td><td>93.17</td><td>92.93</td><td>69.02</td><td>48.05</td><td>46.46</td><td>42.32</td><td>51.46</td></tr><tr><td>x86-64 clang 17.0.1</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.05</td><td>46.71</td><td>42.32</td><td>51.52</td></tr><tr><td>x86-64 clang 18.1.0</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>69.02</td><td>48.17</td><td>46.46</td><td>42.32</td><td>51.49</td></tr></table>
+
+Table 4: The re-executability of differnet models across different compilers. We selected the popular three versions of both the clang and gcc series in Compiler Explorer. The compiler of Compiler Explorer may differ from the compilers provided by Ubuntu in some default compilation options, which might result in slight differences in the generated assembly code.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="5">Re-Compilability</td><td colspan="5">Re-Executability</td></tr><tr><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td><td>O0</td><td>O1</td><td>O2</td><td>O3</td><td>AVG</td></tr><tr><td colspan="11">GPT-4 (OpenAI, 2023)</td></tr><tr><td>vanilla</td><td>76.83</td><td>67.68</td><td>67.07</td><td>57.93</td><td>67.38</td><td>20.73</td><td>13.41</td><td>13.41</td><td>10.98</td><td>14.63</td></tr><tr><td>+sc2dec</td><td>76.22</td><td>67.03</td><td>65.24</td><td>56.10</td><td>66.15</td><td>32.93</td><td>20.12</td><td>19.51</td><td>13.41</td><td>21.49</td></tr><tr><td>1-shot</td><td>76.22</td><td>78.05</td><td>73.78</td><td>70.73</td><td>74.70</td><td>31.71</td><td>20.12</td><td>19.51</td><td>21.34</td><td>23.17</td></tr><tr><td>+sc2dec</td><td>74.39</td><td>76.22</td><td>71.95</td><td>68.90</td><td>72.87</td><td>40.24</td><td>25.61</td><td>23.78</td><td>23.17</td><td>28.20</td></tr><tr><td colspan="11">GPT-3.5 Turbo (OpenAI, 2022)</td></tr><tr><td>vanilla</td><td>20.73</td><td>25.00</td><td>29.27</td><td>21.95</td><td>24.24</td><td>4.27</td><td>3.66</td><td>3.05</td><td>3.05</td><td>3.51</td></tr><tr><td>+sc2dec</td><td>18.29</td><td>22.56</td><td>28.66</td><td>20.12</td><td>22.41</td><td>5.49</td><td>3.66</td><td>3.66</td><td>4.27</td><td>4.27</td></tr><tr><td>1-shot</td><td>49.39</td><td>46.34</td><td>42.07</td><td>40.85</td><td>44.66</td><td>10.37</td><td>7.32</td><td>6.10</td><td>5.49</td><td>7.31</td></tr><tr><td>+sc2dec</td><td>42.07</td><td>42.68</td><td>36.59</td><td>34.15</td><td>38.87</td><td>12.20</td><td>9.15</td><td>6.71</td><td>7.93</td><td>8.99</td></tr><tr><td colspan="11">Deepseek Chat (DeepSeek-AI, 2024)</td></tr><tr><td>vanilla</td><td>62.20</td><td>51.83</td><td>47.56</td><td>45.73</td><td>51.83</td><td>10.37</td><td>5.49</td><td>4.88</td><td>5.49</td><td>6.55</td></tr><tr><td>+sc2dec</td><td>57.93</td><td>50.00</td><td>46.95</td><td>45.12</td><td>50.00</td><td>13.41</td><td>6.71</td><td>6.71</td><td>7.32</td><td>8.54</td></tr><tr><td>1-shot</td><td>67.68</td><td>70.73</td><td>64.63</td><td>56.71</td><td>64.94</td><td>13.41</td><td>10.37</td><td>7.93</td><td>7.32</td><td>9.76</td></tr><tr><td>+sc2dec</td><td>65.85</td><td>70.12</td><td>62.20</td><td>53.66</td><td>62.96</td><td>17.68</td><td>11.59</td><td>10.37</td><td>9.15</td><td>12.20</td></tr><tr><td colspan="11">Deepseek Coder (Guo et al., 2024)</td></tr><tr><td>vanilla</td><td>66.46</td><td>57.93</td><td>47.56</td><td>51.83</td><td>55.95</td><td>9.15</td><td>6.10</td><td>6.71</td><td>6.98</td><td>7.01</td></tr><tr><td>+sc2dec</td><td>62.80</td><td>54.27</td><td>43.29</td><td>50.00</td><td>52.59</td><td>12.20</td><td>7.93</td><td>9.15</td><td>8.54</td><td>9.45</td></tr><tr><td>1-shot</td><td>57.93</td><td>53.66</td><td>52.44</td><td>48.17</td><td>53.05</td><td>11.59</td><td>10.37</td><td>8.54</td><td>7.93</td><td>9.60</td></tr><tr><td>+sc2dec</td><td>56.10</td><td>51.83</td><td>51.22</td><td>46.34</td><td>51.37</td><td>15.24</td><td>11.58</td><td>10.37</td><td>8.54</td><td>11.43</td></tr><tr><td colspan="11">llm4decompile-6.7b (Tan et al., 2024)</td></tr><tr><td>vanilla</td><td>92.80</td><td>93.05</td><td>90.73</td><td>94.02</td><td>92.65</td><td>70.24</td><td>45.49</td><td>40.61</td><td>38.05</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>92.68</td><td>92.80</td><td>90.12</td><td>93.41</td><td>92.26</td><td>71.34</td><td>47.56</td><td>45.98</td><td>41.71</td><td>51.65</td></tr><tr><td>1-shot</td><td>93.05</td><td>93.29</td><td>90.98</td><td>93.90</td><td>92.80</td><td>70.37</td><td>44.88</td><td>41.46</td><td>37.68</td><td>48.60</td></tr><tr><td>+sc2dec</td><td>93.05</td><td>92.68</td><td>90.37</td><td>93.29</td><td>92.35</td><td>71.59</td><td>47.20</td><td>46.34</td><td>41.46</td><td>51.65</td></tr><tr><td colspan="11">llm4decompile-6.7b + FAE</td></tr><tr><td>vanilla</td><td>92.68</td><td>92.44</td><td>93.66</td><td>93.17</td><td>92.99</td><td>67.80</td><td>47.68</td><td>45.73</td><td>42.32</td><td>50.88</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.59</td><td>91.71</td><td>92.32</td><td>91.92</td><td>70.24</td><td>48.54</td><td>47.56</td><td>43.29</td><td>52.41</td></tr><tr><td>1-shot</td><td>92.68</td><td>92.32</td><td>93.66</td><td>92.80</td><td>92.87</td><td>67.68</td><td>47.20</td><td>45.61</td><td>41.95</td><td>50.61</td></tr><tr><td>+sc2dec</td><td>92.07</td><td>91.71</td><td>91.46</td><td>91.95</td><td>91.80</td><td>70.24</td><td>48.05</td><td>47.20</td><td>43.17</td><td>52.16</td></tr></table>
+
+Table 5: The extra results of different methods across four optimization levels (O0, O1, O2, O3), as well as their average scores (AVG). The results in bold represent the optimal performance, while those underlined indicate the second-best performance.

2024/Self-Constructed Context Decompilation with Fined-grained Alignment Enhancement/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f5418f7ea394f7fed48fd012e45d836ff391dff7b243e714c79e4f8a6a55f82
+size 871787

2024/Self-Distillation for Model Stacking Unlocks Cross-Lingual NLU in 200+ Languages/7e26119d-5c03-4bda-95f0-4a298ab50b80_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bf0b43acc827a60c831191cb2fa6cd51e18624fd9c8a05b5cb731a7fb2232ca
+size 681335

2024/Self-Distillation for Model Stacking Unlocks Cross-Lingual NLU in 200+ Languages/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc619ede67e572ff67f8a630dde7040018fbfb4679d983a85a3a9b0f76a3dc2c
+size 2026732

2024/Self-Evaluation of Large Language Model based on Glass-box Features/f272d6b4-0f30-498a-aa3d-937a38ef3871_content_list.json

@@ -0,0 +1,1387 @@
+[
+    {
+        "type": "text",
+        "text": "Self-Evaluation of Large Language Model based on Glass-box Features",
+        "text_level": 1,
+        "bbox": [
+            129,
+            79,
+            867,
+            99
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Hui Huang $^{1}$ , Yingqi  $\\mathrm{Qu}^{2}$ , Jing Liu $^{2}$ , Muyun Yang $^{1\\boxtimes}$ , Bing  $\\mathrm{Xu}^{1\\boxtimes}$ , Tiejun Zhao $^{1}$ , Wenpeng Lu $^{3}$",
+        "bbox": [
+            220,
+            105,
+            781,
+            143
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Faculty of Computing, Harbin Institute of Technology, Harbin, China  $^{2}$ Baidu Inc., Beijing, China",
+        "bbox": [
+            216,
+            143,
+            784,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>3</sup>Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center, Qilu University of Technology, Jinan, China huanghui@stu.hit.edu.cn, {quyingqi, liujing46}@baidu.com, {yangmuyun, hitxb, tjzhao}@hit.edu.cn, lwp@qlu.edu.cn;",
+        "bbox": [
+            107,
+            177,
+            892,
+            244
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The proliferation of open-source Large Language Models (LLMs) underscores the pressing need for evaluation methods. Existing works primarily rely on external evaluators, focusing on training and prompting strategies. However, a crucial aspect - model-aware glass-box features - is overlooked. In this study, we explore the utility of glass-box features under the scenario of self-evaluation, namely applying an LLM to evaluate its own output. We investigate various glass-box feature groups and discovered that the softmax distribution serves as a reliable quality indicator for self-evaluation. Experimental results on public benchmarks validate the feasibility of self-evaluation of LLMs using glass-box features<sup>1</sup>.",
+        "bbox": [
+            141,
+            280,
+            460,
+            507
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            520,
+            260,
+            536
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, as the Large Language Models (LLMs) brings a storm to the area of artificial intelligence, evaluating the quality of LLM outputs also draws a lot of research concentration. As the ability of LLMs develop beyond limitation, it is essential to evaluate them from a comprehensive and scalable perspective. However, traditional evaluation metrics for generative models, such as BLEU and ROUGE, only capture limited aspects of a model's performance (Mathur et al., 2020).",
+        "bbox": [
+            112,
+            546,
+            487,
+            706
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Some research has proposed language model-based evaluation (Li et al., 2023b; Zheng et al., 2023), leveraging proprietary LLMs such as GPT-3.5 or GPT4 (Achiam et al., 2023), to evaluate the LLM's outputs. However, relying on external API for evaluation may introduce consideration about privacy leakage, and the opacity of API models also challenges the evaluation reproducibility. Other works propose to fine-tune open-source models as specialized evaluators (Wang et al., 2024; Zhu et al., 2023; Li et al., 2023a). However, constrained",
+        "bbox": [
+            112,
+            708,
+            489,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "by the capability of foundation model, these finetuned evaluators severely underperform GPT4 on generalizability and fairness (Huang et al., 2024).",
+        "bbox": [
+            507,
+            253,
+            884,
+            300
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, we take a novel approach to LLM evaluation: Is LLM capable of self-evaluation? To answer this, we delve into the utility of glass-box features, namely the useful information that can be extracted from the model as a by-product of generation. Concretely, we explore three groups of glass-box features: 1) softmax distribution, 2) uncertainty estimation and 3) attention distribution. Our findings reveal that manipulating the softmax distribution by calculating its entropy and variance exhibits a strong correlation with annotated evaluation results. Furthermore, we propose two strategies to enhance the evaluation by incorporating features derived from references.",
+        "bbox": [
+            507,
+            302,
+            884,
+            526
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We conduct our experiments on two widely used LLM evaluation benchmarks, MT-Bench (Zheng et al., 2023) and Vicuna-Bench (Chiang et al., 2023). Experimental results notify that the LLM is capable of providing accurate self-evaluation, surpassing proprietary evaluators such as GPT-3.5 and Auto-J. The self-evaluation capability of LLMs holds promise for various applications, ranging from self-reflection to reward modeling.",
+        "bbox": [
+            507,
+            527,
+            882,
+            674
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2 Glass-box Features for Self-Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            688,
+            877,
+            703
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We assume an LLM architecture based on Transformer networks (Vaswani et al., 2017), which is currently the mainstream LLM architecture. In this section, we introduce the three groups of glass-box features for self-evaluation.",
+        "bbox": [
+            507,
+            717,
+            882,
+            796
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2.1 Softmax Distribution",
+        "text_level": 1,
+        "bbox": [
+            507,
+            810,
+            721,
+            826
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Given an instruction  $x$ , the probability of generating response  $y$  can be factorized as:",
+        "bbox": [
+            507,
+            834,
+            882,
+            866
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {S e n t P r o b} = \\prod_ {t = 1} ^ {T} p \\left(y _ {t} \\mid y _ {<   t}, x, \\theta\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            579,
+            878,
+            808,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1Codes are openly available at https://github. com/HuihuiChyan/SelfEval",
+        "bbox": [
+            112,
+            891,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "5813",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5813-5820 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta$  represents model parameters and  $T$  is the response length. However, the model would always select from the most probable tokens during decoding regardless of their quality, leading to biased evaluation. Therefore, we propose two metrics to exploit the softmax distribution for evaluation.",
+        "bbox": [
+            112,
+            84,
+            487,
+            179
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "First, we compute the entropy of softmax output distribution over target vocabulary of size  $V$  at each decoding step to obtain a sentence-level measure:",
+        "bbox": [
+            112,
+            181,
+            485,
+            229
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {S o f t m a x - E n t} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {v = 1} ^ {V} p \\left(y _ {t} ^ {v}\\right) \\log p \\left(y _ {t} ^ {v}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            241,
+            460,
+            284
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $p(y_{t})$  represents the conditional distribution  $p(y_{t} | x, y_{<t}, \\theta)$ . If the majority of the probability mass is concentrated on a limited number of vocabulary words, it indicates that the model is confident and the response is more likely to be accurate. Conversely, if the softmax probabilities resemble a uniform distribution, where selecting any word from the vocabulary is equally probable, the quality of the resulting response is expected to be low.",
+        "bbox": [
+            112,
+            288,
+            487,
+            434
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Second, we hypothesize that the dispersion of probabilities of individual words might provide useful information that is inevitably lost when taking an average. To formalize this intuition we compute the variance of word-level log-probabilities:",
+        "bbox": [
+            112,
+            435,
+            489,
+            514
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {S o f t m a x} - \\operatorname {V a r} = \\mathbb {E} [ \\mathrm {P} ^ {2} ] - (\\mathbb {E} [ \\mathrm {P} ]) ^ {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            530,
+            421,
+            548
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{P} = p(y_1),\\dots,p(y_T)$  represents word-level probabilities for a given response.",
+        "bbox": [
+            112,
+            555,
+            485,
+            587
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Uncertainty Estimation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            599,
+            344,
+            614
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Uncertainty estimation aims to assess the confidence of a mapping across various inputs (Xiao and Wang, 2019). In this work, we propose to employ ensemble-based uncertainty estimation (Malinin and Gales, 2021) during LLM inference as a quality indicator. More specifically, we perform several random forward passes through the model and collect posterior probabilities. Intuitively, if the model is confident with the generation, the sampled distributions should be concentrated and the diversity among them should be small.",
+        "bbox": [
+            112,
+            620,
+            487,
+            796
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Given that LLMs are typically trained without dropout, we propose two strategies to introduce randomness to the forward-passes:",
+        "bbox": [
+            112,
+            797,
+            487,
+            846
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Decoding-based Ensemble: Adopt random top-k decoding (Fan et al., 2018) for inference.",
+            "2. Prompt-based Ensemble: Randomly choose"
+        ],
+        "bbox": [
+            127,
+            859,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "a system prompt from a pre-designed prompt pool for each inference.",
+        "bbox": [
+            544,
+            85,
+            880,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Subsequently, expectation and variance of the resulting probability distributions can be used to quantify uncertainty:",
+        "bbox": [
+            507,
+            128,
+            880,
+            177
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {U n t - E x p} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathrm {S P} _ {T ^ {n}}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            600,
+            187,
+            788,
+            231
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {U n t - V a r} = \\mathbb {E} \\left[ \\mathrm {S P} _ {T ^ {n}} ^ {2} \\right] - \\left(\\mathbb {E} \\left[ \\mathrm {S P} _ {T ^ {n}} \\right]\\right) ^ {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            241,
+            823,
+            260
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where SP denotes the sentence level probability, and  $\\mathbf{N}$  is the forward-pass number.",
+        "bbox": [
+            507,
+            266,
+            880,
+            298
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 Attention Distribution",
+        "text_level": 1,
+        "bbox": [
+            509,
+            309,
+            729,
+            322
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Attention weights represent the strength of connection between source and target tokens, which may be indicative of response quality (Rikters and Fishel, 2017). One way to measure it is to compute the entropy of the attention distribution:",
+        "bbox": [
+            507,
+            330,
+            880,
+            409
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A t t n E n t} = - \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\sum_ {j = 1} ^ {J} \\alpha_ {j i} \\log \\alpha_ {j i}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            569,
+            420,
+            820,
+            466
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha$  represents attention weights,  $I$  and  $J$  are the token numbers of instruction and response.",
+        "bbox": [
+            507,
+            470,
+            880,
+            502
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Since LLMs typically employ a multi-layer and multi-head self-attention architecture, we calculate attention entropy for each head (H) and layer (L) of the decoder in this study. As it is not clear which combination would give the best results for quality prediction, to summarize the information from different heads and layers, we propose to choose the minimum value or compute the average:",
+        "bbox": [
+            507,
+            502,
+            882,
+            632
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A t t n E n t - M i n} = \\min  _ {h l} \\left(\\operatorname {A t t n E n t} _ {h l}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            648,
+            823,
+            665
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {A t t n E n t - A v g} = \\frac {1}{H \\times L} \\sum_ {h = 1} ^ {H} \\sum_ {l = 1} ^ {L} \\operatorname {A t t n E n t} _ {h l}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            537,
+            682,
+            852,
+            726
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Self-Evaluation with Reference",
+        "text_level": 1,
+        "bbox": [
+            507,
+            743,
+            811,
+            758
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "When evaluating the LLM's response to an instruction, a reference answer might be available. In this work, we introduce two strategies to effectively utilize the references for self-evaluation.",
+        "bbox": [
+            507,
+            768,
+            882,
+            831
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 In-Context Illustration",
+        "text_level": 1,
+        "bbox": [
+            507,
+            843,
+            734,
+            857
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Previous research shows that a few annotated samples can improve the performance of LLM via In",
+        "bbox": [
+            507,
+            864,
+            882,
+            896
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "The prompt pool is presented in Appendix A.1.",
+        "bbox": [
+            529,
+            903,
+            825,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "5814",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/8bd5f1a8cb2be34c0d5ab15c0fe0ba1eb30509375c1cb34fde50b3dbbebdb0f1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Method</td><td colspan=\"3\">MT-Bench</td><td colspan=\"3\">Vicuna-Bench</td><td rowspan=\"2\">Average</td></tr><tr><td>Pearson</td><td>Kendall</td><td>Spearman</td><td>Pearson</td><td>Kendall</td><td>Spearman</td></tr><tr><td rowspan=\"13\">LLaMA2-7B-Chat</td><td>Auto-J</td><td>0.5024</td><td>0.4092</td><td>0.5112</td><td>0.5233</td><td>0.3403</td><td>0.3773</td><td>0.5129</td></tr><tr><td>GPT-3.5-Turbo</td><td>0.4342</td><td>0.3982</td><td>0.5033</td><td>0.6695</td><td>0.3858</td><td>0.4330</td><td>0.5519</td></tr><tr><td>Self-Generation</td><td>0.1492</td><td>0.0992</td><td>0.1976</td><td>0.1415</td><td>0.1600</td><td>0.1023</td><td>0.1454</td></tr><tr><td>SentProb</td><td>0.2034</td><td>0.2099</td><td>0.3067</td><td>0.4970</td><td>0.2304</td><td>0.3192</td><td>0.3502</td></tr><tr><td>Softmax-Ent</td><td>0.4666</td><td>0.4395</td><td>0.5978</td><td>0.3730</td><td>0.2441</td><td>0.3223</td><td>0.4198</td></tr><tr><td>Softmax-Var</td><td>0.5612</td><td>0.4695</td><td>0.6239</td><td>0.6506</td><td>0.4534</td><td>0.5894</td><td>0.6059</td></tr><tr><td>Softmax-combo</td><td>0.5879</td><td>0.4638</td><td>0.6222</td><td>0.6209</td><td>0.4352</td><td>0.5650</td><td>0.6044</td></tr><tr><td>DecEnsem-Exp</td><td>0.4105</td><td>0.3877</td><td>0.3526</td><td>0.5606</td><td>0.3545</td><td>0.5051</td><td>0.4856</td></tr><tr><td>DecEnsem-Var</td><td>0.3535</td><td>0.3215</td><td>0.3023</td><td>0.2504</td><td>0.1615</td><td>0.3197</td><td>0.3020</td></tr><tr><td>PrmEnsem-Exp</td><td>0.4242</td><td>0.3054</td><td>0.3692</td><td>0.5381</td><td>0.3359</td><td>0.4675</td><td>0.4812</td></tr><tr><td>PrmtEnsem-Var</td><td>0.1695</td><td>0.1692</td><td>0.2441</td><td>0.1321</td><td>0.0895</td><td>0.1250</td><td>0.1508</td></tr><tr><td>Attn-Ent-Min</td><td>0.1444</td><td>0.1150</td><td>0.1563</td><td>0.1182</td><td>0.1161</td><td>0.1584</td><td>0.1313</td></tr><tr><td>Attn-Ent-Avg</td><td>0.1361</td><td>0.1094</td><td>0.1475</td><td>0.0977</td><td>0.1181</td><td>0.1600</td><td>0.1169</td></tr><tr><td rowspan=\"13\">Vicuna-7B</td><td>Auto-J</td><td>0.5673</td><td>0.4302</td><td>0.5405</td><td>0.6003</td><td>0.4818</td><td>0.5345</td><td>0.5838</td></tr><tr><td>GPT3.5-Turbo</td><td>0.4812</td><td>0.4358</td><td>0.5353</td><td>0.6946</td><td>0.5901</td><td>0.6733</td><td>0.5879</td></tr><tr><td>Self-Generation</td><td>0.1636</td><td>0.1753</td><td>0.1498</td><td>0.1985</td><td>0.1319</td><td>0.1709</td><td>0.1811</td></tr><tr><td>SentProb</td><td>0.2350</td><td>0.1706</td><td>0.2547</td><td>0.4606</td><td>0.2545</td><td>0.3051</td><td>0.3478</td></tr><tr><td>Softmax-Ent</td><td>0.4690</td><td>0.3003</td><td>0.4291</td><td>0.6379</td><td>0.2851</td><td>0.3855</td><td>0.5535</td></tr><tr><td>Softmax-Var</td><td>0.4632</td><td>0.3171</td><td>0.4403</td><td>0.6146</td><td>0.2879</td><td>0.3772</td><td>0.5389</td></tr><tr><td>Softmax-combo</td><td>0.5358</td><td>0.3578</td><td>0.5022</td><td>0.6856</td><td>0.3276</td><td>0.4270</td><td>0.6107</td></tr><tr><td>DecEnsem-Exp</td><td>0.4518</td><td>0.2645</td><td>0.2168</td><td>0.5099</td><td>0.5137</td><td>0.6548</td><td>0.4809</td></tr><tr><td>DecEnsem-Var</td><td>0.2091</td><td>0.1476</td><td>0.1622</td><td>0.3246</td><td>0.2931</td><td>0.3014</td><td>0.2669</td></tr><tr><td>PrmEnsem-Exp</td><td>0.4414</td><td>0.2677</td><td>0.2229</td><td>0.4897</td><td>0.4873</td><td>0.6312</td><td>0.4656</td></tr><tr><td>PrmtEnsem-Var</td><td>0.2353</td><td>0.0875</td><td>0.1239</td><td>0.2028</td><td>0.1860</td><td>0.3201</td><td>0.2191</td></tr><tr><td>Attn-Ent-Avg</td><td>0.0355</td><td>0.0234</td><td>0.0043</td><td>0.2254</td><td>0.1981</td><td>0.2567</td><td>0.1305</td></tr><tr><td>Attn-Ent-Min</td><td>0.0463</td><td>0.0307</td><td>0.0530</td><td>0.2161</td><td>0.1888</td><td>0.2459</td><td>0.1312</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            884,
+            463
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Experiment results of different groups of methods for LLM self-evaluation. Softmax-combo denotes the normalized summation of Softmax-Ent and Softmax-Var.",
+        "bbox": [
+            112,
+            470,
+            882,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "context Learning (Wang et al., 2023a). Therefore, we extend the prompt with the instruction and its reference as in-context demonstration<sup>3</sup>.",
+        "bbox": [
+            112,
+            525,
+            487,
+            573
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "By prefixing the reference, the model tends to generate a similar softmax distribution (Wang et al., 2023a). Therefore, when the model is subsequently forced to generate the current answer, the resulting softmax distribution could represent the difference between the current answer and the golden answer, effectively indicating the quality.",
+        "bbox": [
+            112,
+            575,
+            489,
+            688
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Probability Calibration",
+        "text_level": 1,
+        "bbox": [
+            112,
+            701,
+            346,
+            715
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "There has been a lot of research about the bias of the evaluator (Huang et al., 2023; Wang et al., 2023b), where the evaluator would predict on superficial quality, such as complexity. As the reference should always be assigned with a maximum score, we can quantify the bias of the model by calculate the log-probability of reference answer:",
+        "bbox": [
+            112,
+            722,
+            489,
+            838
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathrm {S e n t P r o b - R e f} = - \\frac {1}{T} \\sum_ {t = 1} ^ {T} p (\\tilde {y} _ {t}) \\log p (\\tilde {y} _ {t})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            152,
+            848,
+            448,
+            892
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\tilde{y}_t$  denotes the log-probability assigned by the model to the reference. Based on that, the bias of the self-evaluation can be mitigated by subtracting the result with SentProb-Ref, thereby improving the evaluation accuracy.",
+        "bbox": [
+            507,
+            525,
+            884,
+            606
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            615,
+            655,
+            633
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Set-up",
+        "text_level": 1,
+        "bbox": [
+            507,
+            640,
+            608,
+            657
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Benchmark. We carry out our experiments on two widely used LLM evaluation benchmarks: MT-Bench (Zheng et al., 2023) and Vicuna-Bench (Chiang et al., 2023). Both of the benchmarks encompass 80 questions covering diverse areas, and we derive the responses for different models following the default settings of MT-Bench. To mitigate the cost of human annotations, we follow the official evaluators according to each benchmark, namely GPT-4 (Achiam et al., 2023).",
+        "bbox": [
+            507,
+            661,
+            884,
+            820
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Model. Our experiments are based on two popular models, namely Vicuna-7B (Chiang et al., 2023) and LLaMA2-7B-chat (Touvron et al., 2023), both are enabled with instruction-following ability.",
+        "bbox": [
+            507,
+            822,
+            882,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Metric. Pearson Correlation Coefficient between the prediction and the annotation is taken as the",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "Detailed prompt template is presented in Appendix A.2.",
+        "bbox": [
+            134,
+            903,
+            482,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "5815",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "major metric. We also report Spearman's Ranking Correlation Coefficient and Kendall's Tau Ranking Correlation Coefficient<sup>4</sup> as extra reference.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Baseline. We mainly compare with two evaluators, GPT-3.5-Turbo $^{5}$  and Auto-J (Li et al., 2023a). The former is a close-source LLM carefully prompted for LLM evaluation, and the latter is an open-source LLM specifically fine-tuned for LLM evaluation. These are the two mainstream methods for LLM evaluation. We also compare with SelfGeneration, namely prompting the model as an evaluator to evaluate its own response $^{6}$ .",
+        "bbox": [
+            112,
+            133,
+            489,
+            278
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            290,
+            265,
+            305
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 1, among the three groups of glass-box features, softmax distribution based features perform the best, which achieves a high correlation with annotations and outperforms Auto-J and GPT-3.5-Turbo. This verifies the strong association between the confidence represented by the softmax distribution and the response quality. When the instruction exceeds the LLM's capability scope, the model would generate the response with low confidence, resulting in a sparse log-probability distribution. Furthermore, combining both entropy and variance by simply adding them together can achieve further improvement.",
+        "bbox": [
+            112,
+            312,
+            489,
+            520
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The uncertainty-based methods underperform, particularly those relying on variance. This might be because we set forward-pass number as 10, which is inadequate to quantify model uncertainty. Considering multiple forward-passes is overly time-consuming, we think the ensembled uncertainty estimation is less practical for self-evaluation.",
+        "bbox": [
+            112,
+            521,
+            489,
+            633
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The attention-based methods exhibit a poor correlation. Despite its effectiveness in unsupervised translation evaluation (Fomicheva et al., 2020), LLM-based response generation differs significantly from encoder-decoder-based machine translation, which typically produces translations for one or two tokens at each step. Therefore, a dispersed attention across multiple positions, leading to a low attention entropy, does not necessarily indicate a poor response.",
+        "bbox": [
+            112,
+            634,
+            489,
+            796
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The self-generation also exhibits a poor correlation. This is because the model fails to comprehend the instruction to generate a score between 1 to 10",
+        "bbox": [
+            112,
+            797,
+            489,
+            844
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/7a55d4fe7e2d401200268d2320d9befddd61fe0b5d52e15d77af4df19cb28538.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Vicuna-Bench</td></tr><tr><td>Pearson</td><td>Kendall</td><td>Spearman</td></tr><tr><td>GPT-3.5-Turbo</td><td>0.6695</td><td>0.3858</td><td>0.4330</td></tr><tr><td colspan=\"4\">Results on LLaMA2-7B-Chat</td></tr><tr><td>Softmax-Ent</td><td>0.3730</td><td>0.2441</td><td>0.3223</td></tr><tr><td>+illustration</td><td>0.3678</td><td>0.2429</td><td>0.3370</td></tr><tr><td>+calibration</td><td>0.3930</td><td>0.2631</td><td>0.3490</td></tr><tr><td>Softmax-Var</td><td>0.6506</td><td>0.4534</td><td>0.5894</td></tr><tr><td>+illustration</td><td>0.6580</td><td>0.4472</td><td>0.5865</td></tr><tr><td>+calibration</td><td>0.6617</td><td>0.4360</td><td>0.5625</td></tr><tr><td colspan=\"4\">Results on Vicuna-7B</td></tr><tr><td>Softmax-Ent</td><td>0.6379</td><td>0.2851</td><td>0.3855</td></tr><tr><td>+illustration</td><td>0.6375</td><td>0.2982</td><td>0.3960</td></tr><tr><td>+calibration</td><td>0.6441</td><td>0.3177</td><td>0.4240</td></tr><tr><td>Softmax-Var</td><td>0.6146</td><td>0.2879</td><td>0.3772</td></tr><tr><td>+illustration</td><td>0.6247</td><td>0.2976</td><td>0.3903</td></tr><tr><td>+calibration</td><td>0.6526</td><td>0.3099</td><td>0.4133</td></tr></table>",
+        "bbox": [
+            515,
+            82,
+            867,
+            326
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Experiment results of different reference augmentation strategies for self-evaluation.",
+        "bbox": [
+            507,
+            336,
+            882,
+            365
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "in most cases. This underscores the superiority of our proposed method, which is not limited by the capabilities of the evaluated model and can be effectively applied to 7B-sized models.",
+        "bbox": [
+            507,
+            392,
+            882,
+            456
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 Self-Evaluation with Reference",
+        "text_level": 1,
+        "bbox": [
+            507,
+            469,
+            800,
+            483
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we evaluate the two reference augmentation methods proposed in Section 3, namely in-context illustration and probability calibration. As shown in Table 2, both the two strategies can enhance the evaluation accuracy by a large margin, from the perspective of in-context learning and bias calibration. This notifies the references is a pivotal information for response evaluation, and should not be neglected if available. Furthermore, among the two methods, the calibration-based method performs better, verifying calibration is an effective method to mitigate the bias introduced by the casual language modeling process, which matters a lot for more accurate self-evaluation.",
+        "bbox": [
+            507,
+            491,
+            884,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            730,
+            640,
+            746
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this work, we investigate the utility of glass-box features for LLM self-evaluation. Our findings verify that the confidence quantified by softmax distribution can be a reliable quality indicator. The self-evaluation ability of LLM is a promising pathway, for example, the LLM can rely on self-evaluation results to decide whether to perform self-reflection (Xie et al., 2024), or to select preferred data for reward modeling (Lee et al., 2024), and we leave these as the future work.",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "4Notice measuring our method with ranking coefficient is not fair, as our method can not predict tie. 5https://platform.openai.com/docs/ models/gpt-3-5-turbo 6Detailed prompt template is presented in Appendix A.3.",
+        "bbox": [
+            112,
+            854,
+            487,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "5816",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our work still has some limitations: 1) Due to time and resource constraints, we primarily relied on GPT-4 for annotating golden labels during the meta-evaluation process. Including human evaluators would enhance the credibility of our proposed self-evaluation methods. 2) The experiments are primarily conducted on 7B version models. To conduct a more thorough evaluation of our methods, it would be beneficial to incorporate larger models with more parameters. 3) The self-evaluation capability of the LLM can be applied to various applications, as discussed in the conclusion part. Including the validation of self-evaluation methods on other LLM-based applications such as self-reflection would further bolster its utility.",
+        "bbox": [
+            112,
+            109,
+            492,
+            351
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            363,
+            285,
+            379
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by National Natural Science Foundation of China (62276077, 62376075, U1908216, 62376076), Key R&D Program of Yunnan (202203AA080004) and Shenzhen College Stability Support Plan (No. GXWD20220811170358002).",
+        "bbox": [
+            112,
+            388,
+            489,
+            483
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            511,
+            213,
+            526
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.",
+            "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.",
+            "Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 889-898, Melbourne, Australia.",
+            "Marina Fomicheva, Shuo Sun, Lisa Yankovskaya, Frédéric Blain, Francisco Guzmán, Mark Fishel, Nikolaos Aletras, Vishrav Chaudhary, and Lucia Specia. 2020. Unsupervised quality estimation for neural machine translation. Transactions of the Association for Computational Linguistics, 8:539-555.",
+            "Hui Huang, Yingqi Qu, Jing Liu, Muyun Yang, and Tiejun Zhao. 2024. An empirical study of llm-as-a-judge for llm evaluation: Fine-tuned judge models are task-specific classifiers."
+        ],
+        "bbox": [
+            115,
+            533,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Hui Huang, Shuangzhi Wu, Kehai Chen, Hui Di, Muyun Yang, and Tiejun Zhao. 2023. Improving translation quality estimation with bias mitigation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2175-2190.",
+            "Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Ren Lu, Thomas Mesnard, Johan Ferret, Colton Bishop, Ethan Hall, Victor Carbune, and Abhinav Rastogi. 2024. RLAIF: Scaling reinforcement learning from human feedback with AI feedback.",
+            "Junlong Li, Shichao Sun, Weizhe Yuan, Run-Ze Fan, Hai Zhao, and Pengfei Liu. 2023a. Generative judge for evaluating alignment. arXiv preprint arXiv:2310.05470.",
+            "Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023b. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval.",
+            "Andrey Malinin and Mark Gales. 2021. Uncertainty estimation in autoregressive structured prediction.",
+            "Nitika Mathur, Timothy Baldwin, and Trevor Cohn. 2020. Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4984-4997, Online.",
+            "Matiss Rikters and Mark Fishel. 2017. Confidence through attention.",
+            "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.",
+            "Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. 2023a. Label words are anchors: An information flow perspective for understanding in-context learning. In *Conference on Empirical Methods in Natural Language Processing*.",
+            "Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. 2023b. Large language models are not fair evaluators.",
+            "Yidong Wang, Zhuohao Yu, Zhengran Zeng, Linyi Yang, Cunxiang Wang, Hao Chen, Chaoya Jiang, Rui Xie, Jindong Wang, Xing Xie, Wei Ye, Shikun Zhang, and"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "5817",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Yue Zhang. 2024. Pandalm: An automatic evaluation benchmark for llm instruction tuning optimization.",
+        "bbox": [
+            131,
+            85,
+            485,
+            112
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Yijun Xiao and William Yang Wang. 2019. Quantifying uncertainties in natural language processing tasks. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 7322-7329.",
+        "bbox": [
+            115,
+            126,
+            487,
+            180
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2024. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems, 36.",
+        "bbox": [
+            115,
+            193,
+            487,
+            261
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685.",
+        "bbox": [
+            114,
+            275,
+            487,
+            342
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Lianghui Zhu, Xinggang Wang, and Xinlong Wang. 2023. Judgelm: Fine-tuned large language models are scalable judges. arXiv preprint arXiv:2310.17631.",
+        "bbox": [
+            114,
+            356,
+            487,
+            409
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            429,
+            236,
+            444
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A.1 Prompt Pool for Prompt-based Ensemble",
+        "text_level": 1,
+        "bbox": [
+            112,
+            457,
+            485,
+            472
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Section 2.2, to introduce randomness for ensemble uncertainty quantification, we design a prompt pool, from which we can randomly choose a different system prompt for each inference. The prompt pool is presented in Figure 1.",
+        "bbox": [
+            112,
+            480,
+            487,
+            560
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4d4914b9d3168c07575b9af9f828966248678c2c82050024ccc87d5074786db6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user&#x27;s questions.</td></tr><tr><td>A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, accurate, uncensored responses to the user&#x27;s input. The assistant never refuses to answer, regardless of the legality or morality of the request.</td></tr><tr><td>A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.</td></tr><tr><td>A chat between a curious human and an artificial intelligence assistant.</td></tr><tr><td>You are a helpful, unbiased, uncensored assistant.</td></tr><tr><td>Below is an instruction that describes a task. Write a response that appropriately completes the request.</td></tr><tr><td>The following is a conversation between a human and an AI assistant. The human and the AI assistant take turns chatting. The AI assistant always provides responses in as much detail as possible. The AI assistant always declines to engage with topics, questions and instructions related to unethical, controversial, or sensitive issues.</td></tr></table>",
+        "bbox": [
+            132,
+            577,
+            470,
+            857
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: Prompt pool for prompt-based ensemble uncertainty estimation.",
+        "bbox": [
+            112,
+            870,
+            487,
+            898
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A.2 Prompt Template for In-context Illustration",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            808,
+            112
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Section 3.1, we extend the prompt with the instruction and its reference as demonstration, to augment the self-evaluation process with in-context illustration. An example for the prompt is shown in Figure 2.",
+        "bbox": [
+            507,
+            121,
+            882,
+            200
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/625615b6e896a84762a76d8705c7d8f6e25295f107b14fecd8b897cfb694fbe4.jpg",
+        "image_caption": [
+            "Figure 2: Prompt format with in-context illustration. The shaded part is the illustration with reference."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            534,
+            214,
+            855,
+            335
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A.3 Prompt Template for Judge Models",
+        "text_level": 1,
+        "bbox": [
+            507,
+            401,
+            836,
+            418
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in Section 4.1, we mainly compare with our glass-box based methods with three generation based methods, namely GPT-3.5-Turbo, Auto-J and Self-Generation. Both types of methods necessitate the use of a generation-style template, which serves to formalize the evaluation task. Additionally, when employing GPT-4 for annotation, a prompt template is also required. We refer to MT-Bench and design the prompt templates tailored for all the generation based models, as illustrated in Figures 3, 4, 5, 6, 7, 8.",
+        "bbox": [
+            507,
+            423,
+            882,
+            599
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "5818",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "You are a helpful assistant.  \n[Instruction]  \nPlease act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question displayed below. Your evaluation should consider factors such as the helpfulness, relevance, accuracy, depth, creativity, and level of detail of the response. Begin your evaluation by providing a short explanation. Be as objective as possible. After providing your explanation, you must rate the response on a scale of 1 to 10 by strictly following this format: \\[\"rating]\\\\\", for example: \\\\\"Rating: [[5]]\\\".  \n[Question]  \n{question}  \n[The Start of Assistant's Answer]  \n{answer}  \n[The End of Assistant's Answer]",
+        "guess_lang": "txt",
+        "bbox": [
+            203,
+            96,
+            786,
+            269
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 3: Prompt template for GPT4 and GPT-3.5-Turbo applied for single-turn evaluation.",
+        "bbox": [
+            189,
+            282,
+            806,
+            297
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "You are a helpful assistant.   \n[Instruction]   \nPlease act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question displayed below. Your evaluation should consider factors such as the helpfulness, relevance, accuracy, depth, creativity, and level of detail of the response. You evaluation should focus on the assistant's answer to the second user question. Begin your evaluation by providing a short explanation. Be as objective as possible. After providing your explanation, you must rate the response on a scale of 1 to 10 by strictly following this format: \\[\"[[rating]]\\\\\"', for example: \\\\\"Rating: [[5]]\\\".   \n<|The Start of Assistant A's Conversation with User|>   \n### User: {question_1}   \n### Assistant A: {answer_1}   \n### User: {question_2}   \n### Assistant A: {answer_2}   \n<|The End of Assistant A's Conversation with User|>",
+        "guess_lang": "markdown",
+        "bbox": [
+            203,
+            329,
+            793,
+            609
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: Prompt template for GPT4 and GPT-3.5-Turbo applied for multi-turn evaluation.",
+        "bbox": [
+            191,
+            621,
+            803,
+            638
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "You are a helpful and precise assistant for checking the quality of the answer. [Question] {question}   \n[The Start of Assistant's Answer] {answer}   \n[The End of Assistant's Answer]   \n[System]   \nWe would like to request your feedback on the performance of the AI assistant in response to the user question displayed above. Please rate the helpfulness, relevance, accuracy, level of details of the response. The assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance. Please first output a single line containing only one values indicating the score for the Assistant. In the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias.   \n``` Response:",
+        "guess_lang": "txt",
+        "bbox": [
+            203,
+            670,
+            793,
+            875
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 5: Prompt template for Auto-J applied for single-turn evaluation.",
+        "bbox": [
+            253,
+            889,
+            741,
+            903
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "5819",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "You are a helpful and precise assistant for checking the quality of the answer.   \n<|The Start of Assistant's Conversation with User|>   \n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n```\n``}\n```\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``'\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``\"\n``)\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n```java\n`",
+        "guess_lang": "txt",
+        "bbox": [
+            206,
+            98,
+            784,
+            394
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 6: Prompt template for Auto-J applied for multi-turn evaluation.",
+        "bbox": [
+            255,
+            411,
+            739,
+            426
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "Write critiques for a submitted response on a given user's query, and grade the response:   \n# [BEGIN DATA]   \n# ***   \n# [Query]: {question}   \n# ***   \n# [Response]: {answer}   \n# ***   \n# [END DATA]   \n# Write critiques for this response. After that, you should give a final rating for the response on a scale of 1 to 10 by strictly following this format:\"[[rating]]\", for example:\"Rating:[[5]]\".",
+        "guess_lang": "txt",
+        "bbox": [
+            205,
+            462,
+            774,
+            613
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 7: Prompt template for self-generation applied for single-turn evaluation.",
+        "bbox": [
+            226,
+            627,
+            769,
+            643
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "code",
+        "sub_type": "code",
+        "code_caption": [],
+        "code_body": "Write critiques for a submitted response on a given user's query, and grade the response:   \n# [BEGIN DATA]   \n# ***   \n# [Query 1]: {question_1}   \n# ***   \n# [Response 1]: {answer_1}   \n# ***   \n# [Query 2]: {question_2}   \n# ***   \n# [Response 2]: {answer_2}   \n# ***   \n# [END DATA]   \n# Write critiques for this response. After that, you should give a final rating for the response on a scale of 1 to 10 by strictly following this format:\"[[rating]]\", for example:\"Rating:[[5]]\".",
+        "guess_lang": "txt",
+        "bbox": [
+            203,
+            678,
+            774,
+            872
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 8: Prompt template for self-generation applied for multi-turn evaluation.",
+        "bbox": [
+            228,
+            887,
+            766,
+            902
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "5820",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 7
+    }
+]