Add Batch 501c285e-e74f-42a6-8926-1ff54559192d data

.gitattributes

@@ -1333,3 +1333,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/Turbocharging[[:space:]]Web[[:space:]]Automation_[[:space:]]The[[:space:]]Impact[[:space:]]of[[:space:]]Compressed[[:space:]]History[[:space:]]States/90959e26-031c-4a2a-aeef-cd9579a795f3_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Turning[[:space:]]Conversations[[:space:]]into[[:space:]]Workflows_[[:space:]]A[[:space:]]Framework[[:space:]]to[[:space:]]Extract[[:space:]]and[[:space:]]Evaluate[[:space:]]Dialog[[:space:]]Workflows[[:space:]]for[[:space:]]Service[[:space:]]AI[[:space:]]Agents/a413ca11-89dd-4086-a921-30cbb28fd4f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/UAQFact_[[:space:]]Evaluating[[:space:]]Factual[[:space:]]Knowledge[[:space:]]Utilization[[:space:]]of[[:space:]]LLMs[[:space:]]on[[:space:]]Unanswerable[[:space:]]Questions/8ef9c36a-553b-4fc6-890d-f60c9d6072cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UBench_[[:space:]]Benchmarking[[:space:]]Uncertainty[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Multiple[[:space:]]Choice[[:space:]]Questions/9fd6e574-1e70-4a26-a6f0-5806484e6229_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UCS-SQL_[[:space:]]Uniting[[:space:]]Content[[:space:]]and[[:space:]]Structure[[:space:]]for[[:space:]]Enhanced[[:space:]]Semantic[[:space:]]Bridging[[:space:]]In[[:space:]]Text-to-SQL/5e130ac7-c9a3-4216-9a37-f02974ca9352_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UI-E2I-Synth_[[:space:]]Advancing[[:space:]]GUI[[:space:]]Grounding[[:space:]]with[[:space:]]Large-Scale[[:space:]]Instruction[[:space:]]Synthesis/65ba1011-e397-40da-8aed-28e8c07728b8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UQ-Merge_[[:space:]]Uncertainty[[:space:]]Guided[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Merging/f3c36716-306f-46b5-8771-6bf9a011aa75_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/USDC_[[:space:]]A[[:space:]]Dataset[[:space:]]of[[:space:]]̲User[[:space:]]̲Stance[[:space:]]and[[:space:]]̲Dogmatism[[:space:]]in[[:space:]]Long[[:space:]]̲Conversations/a4fabed3-7f27-4469-8663-7ac9acb62d69_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Un-considering[[:space:]]Contextual[[:space:]]Information_[[:space:]]Assessing[[:space:]]LLMs’[[:space:]]Understanding[[:space:]]of[[:space:]]Indexical[[:space:]]Elements/1f95b7d5-2289-4e70-9979-ad44a236414a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uncertainty[[:space:]]Unveiled_[[:space:]]Can[[:space:]]Exposure[[:space:]]to[[:space:]]More[[:space:]]In-context[[:space:]]Examples[[:space:]]Mitigate[[:space:]]Uncertainty[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models_/e10afd9c-4f0f-4673-bde1-594a31006fee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Uncertainty-Aware[[:space:]]Contrastive[[:space:]]Decoding/d9a15b33-5d81-4410-9837-722f1c358cf5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understand[[:space:]]User[[:space:]]Opinions[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]LLM-Powered[[:space:]]In-the-Moment[[:space:]]User[[:space:]]Experience[[:space:]]Interviews/2864e4df-aed9-48ae-b951-245a95f5d371_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understand[[:space:]]the[[:space:]]Implication_[[:space:]]Learning[[:space:]]to[[:space:]]Think[[:space:]]for[[:space:]]Pragmatic[[:space:]]Understanding/16f01b94-b06c-4fe1-bd95-38b70beaff19_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understanding[[:space:]]and[[:space:]]Meeting[[:space:]]Practitioner[[:space:]]Needs[[:space:]]When[[:space:]]Measuring[[:space:]]Representational[[:space:]]Harms[[:space:]]Caused[[:space:]]by[[:space:]]LLM-Based[[:space:]]Systems/3c8dfc51-4ccd-4b8e-8a93-c25831294a1e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understanding[[:space:]]the[[:space:]]Gap_[[:space:]]an[[:space:]]Analysis[[:space:]]of[[:space:]]Research[[:space:]]Collaborations[[:space:]]in[[:space:]]NLP[[:space:]]and[[:space:]]Language[[:space:]]Documentation/c32352f6-3b22-4cf3-9c5b-8b327766eda0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understanding[[:space:]]the[[:space:]]Influence[[:space:]]of[[:space:]]Synthetic[[:space:]]Data[[:space:]]for[[:space:]]Text[[:space:]]Embedders/1b717fea-b9b7-4ff2-87f7-5a75299ba628_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Understanding[[:space:]]the[[:space:]]Repeat[[:space:]]Curse[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]from[[:space:]]a[[:space:]]Feature[[:space:]]Perspective/8da2853d-05a2-4efe-9214-48122e7dce7d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UniT_[[:space:]]One[[:space:]]Document,[[:space:]]Many[[:space:]]Revisions,[[:space:]]Too[[:space:]]Many[[:space:]]Edit[[:space:]]Intention[[:space:]]Taxonomies/4bc8dae3-a964-4b28-903c-aa388dbc3713_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unilogit_[[:space:]]Robust[[:space:]]Machine[[:space:]]Unlearning[[:space:]]for[[:space:]]LLMs[[:space:]]Using[[:space:]]Uniform-Target[[:space:]]Self-Distillation/ebef0ed4-4fb1-40df-a718-ad93fa15f77c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unlearning[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]for[[:space:]]LLMs[[:space:]]with[[:space:]]Weak-to-Strong[[:space:]]Knowledge[[:space:]]Distillation/97fa610f-3b34-4428-84d7-824569fd0b25_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unlocking[[:space:]]LLMs’[[:space:]]Self-Improvement[[:space:]]Capacity[[:space:]]with[[:space:]]Autonomous[[:space:]]Learning[[:space:]]for[[:space:]]Domain[[:space:]]Adaptation/c8980791-7107-444f-9f13-a78a126edf68_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unlocking[[:space:]]Recursive[[:space:]]Thinking[[:space:]]of[[:space:]]LLMs_[[:space:]]Alignment[[:space:]]via[[:space:]]Refinement/38c44d30-1d88-4b18-a783-977d34e6a0f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unlocking[[:space:]]Speech[[:space:]]Instruction[[:space:]]Data[[:space:]]Potential[[:space:]]with[[:space:]]Query[[:space:]]Rewriting/30da2c0c-050a-4f46-ae0f-eea442f6e953_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unraveling[[:space:]]and[[:space:]]Mitigating[[:space:]]Safety[[:space:]]Alignment[[:space:]]Degradation[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/501fc708-0302-4e17-b0e0-6bd29d032ac5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/UnrealLLM_[[:space:]]Towards[[:space:]]Highly[[:space:]]Controllable[[:space:]]and[[:space:]]Interactable[[:space:]]3D[[:space:]]Scene[[:space:]]Generation[[:space:]]by[[:space:]]LLM-powered[[:space:]]Procedural[[:space:]]Content[[:space:]]Generation/405c6054-7641-4fa7-abfa-be10c63fd306_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unsupervised[[:space:]]Morphological[[:space:]]Tree[[:space:]]Tokenizer/d2ecac09-3fae-4ef1-9c8b-8475617a7d2c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unveiling[[:space:]]Confirmation[[:space:]]Bias[[:space:]]in[[:space:]]Chain-of-Thought[[:space:]]Reasoning/2cd5e8a8-e830-47e3-885a-2f3eacf0fac5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unveiling[[:space:]]Privacy[[:space:]]Risks[[:space:]]in[[:space:]]Multi-modal[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]Task-specific[[:space:]]Vulnerabilities[[:space:]]and[[:space:]]Mitigation[[:space:]]Challenges/93a15a89-a8e5-4e28-974c-12b4bea018be_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unveiling[[:space:]]and[[:space:]]Addressing[[:space:]]Pseudo[[:space:]]Forgetting[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/98745d81-fda5-4907-a4dd-2f35e6bb197a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unveiling[[:space:]]the[[:space:]]Key[[:space:]]Factors[[:space:]]for[[:space:]]Distilling[[:space:]]Chain-of-Thought[[:space:]]Reasoning/9c7e25dc-2572-41e2-82c8-386d62843de0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Unveiling[[:space:]]the[[:space:]]Lack[[:space:]]of[[:space:]]LVLM[[:space:]]Robustness[[:space:]]to[[:space:]]Fundamental[[:space:]]Visual[[:space:]]Variations_[[:space:]]Why[[:space:]]and[[:space:]]Path[[:space:]]Forward/9c968477-5e00-477d-a164-1e3795a79d95_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/User[[:space:]]Behavior[[:space:]]Prediction[[:space:]]as[[:space:]]a[[:space:]]Generic,[[:space:]]Robust,[[:space:]]Scalable,[[:space:]]and[[:space:]]Low-Cost[[:space:]]Evaluation[[:space:]]Strategy[[:space:]]for[[:space:]]Estimating[[:space:]]Generalization[[:space:]]in[[:space:]]LLMs/ef7fa365-6470-4dc1-8056-603179c0e3d5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Utilizing[[:space:]]Semantic[[:space:]]Textual[[:space:]]Similarity[[:space:]]for[[:space:]]Clinical[[:space:]]Survey[[:space:]]Data[[:space:]]Feature[[:space:]]Selection/6ae2c141-29a2-4482-89c9-877b471fb34a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/V-ALPHASOCIAL_[[:space:]]Benchmark[[:space:]]and[[:space:]]Self-Reflective[[:space:]]Chain-of-Thought[[:space:]]Generation[[:space:]]for[[:space:]]Visual[[:space:]]Social[[:space:]]Commonsense[[:space:]]Reasoning/720fe3d9-c954-4552-8ce5-6fd5f5760648_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VADE_[[:space:]]Visual[[:space:]]Attention[[:space:]]Guided[[:space:]]Hallucination[[:space:]]Detection[[:space:]]and[[:space:]]Elimination/041f66e3-9fc7-4ef9-b558-57d90ef94ee0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VAQUUM_[[:space:]]Are[[:space:]]Vague[[:space:]]Quantifiers[[:space:]]Grounded[[:space:]]in[[:space:]]Visual[[:space:]]Data_/d2f64794-6d12-4309-a0b8-2769a2a33b2e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VCD_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Visual[[:space:]]Commonsense[[:space:]]Discovery[[:space:]]in[[:space:]]Images/29530760-85cb-4a27-85da-470fa7847833_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VISIAR_[[:space:]]Empower[[:space:]]MLLM[[:space:]]for[[:space:]]Visual[[:space:]]Story[[:space:]]Ideation/5ab6c776-969d-46f6-a207-97d819ac3564_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VP-MEL_[[:space:]]Visual[[:space:]]Prompts[[:space:]]Guided[[:space:]]Multimodal[[:space:]]Entity[[:space:]]Linking/e4fe0d80-b272-4235-80b8-4a08d91e051d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VSCBench_[[:space:]]Bridging[[:space:]]the[[:space:]]Gap[[:space:]]in[[:space:]]Vision-Language[[:space:]]Model[[:space:]]Safety[[:space:]]Calibration/a5ef10fe-794a-46e6-8ec2-5d789c9ca81a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Variable[[:space:]]Layerwise[[:space:]]Quantization_[[:space:]]A[[:space:]]Simple[[:space:]]and[[:space:]]Effective[[:space:]]Approach[[:space:]]to[[:space:]]Quantize[[:space:]]LLMs/e4378bc4-5112-42f2-89f4-7644977087f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Verbosity-Aware[[:space:]]Rationale[[:space:]]Reduction_[[:space:]]Sentence-Level[[:space:]]Rationale[[:space:]]Reduction[[:space:]]for[[:space:]]Efficient[[:space:]]and[[:space:]]Effective[[:space:]]Reasoning/665cfe23-16d3-4b40-92c5-614fd9ae3e8f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Verify[[:space:]]with[[:space:]]Caution_[[:space:]]The[[:space:]]Pitfalls[[:space:]]of[[:space:]]Relying[[:space:]]on[[:space:]]Imperfect[[:space:]]Factuality[[:space:]]Metrics/bf893d4b-451c-47fe-a683-969fc6d23176_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Verifying[[:space:]]the[[:space:]]Steps[[:space:]]of[[:space:]]Deductive[[:space:]]Reasoning[[:space:]]Chains/d9548e13-4aae-4da6-940e-a6af28b1e6f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VidCapBench_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]of[[:space:]]Video[[:space:]]Captioning[[:space:]]for[[:space:]]Controllable[[:space:]]Text-to-Video[[:space:]]Generation/7d303d77-8cc7-47ba-b607-2b93b15dd4d9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/VideoRAG_[[:space:]]Retrieval-Augmented[[:space:]]Generation[[:space:]]over[[:space:]]Video[[:space:]]Corpus/6d7f57f0-2f8d-4cd2-810f-9aa38aecafd5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visibility[[:space:]]as[[:space:]]Survival_[[:space:]]Generalizing[[:space:]]NLP[[:space:]]for[[:space:]]Native[[:space:]]Alaskan[[:space:]]Language[[:space:]]Identification/4503db5e-363a-43ab-b1a9-db9e2b5d00b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision[[:space:]]Language[[:space:]]Model[[:space:]]Helps[[:space:]]Private[[:space:]]Information[[:space:]]De-Identification[[:space:]]in[[:space:]]Vision[[:space:]]Data/90c16d4e-b6eb-4669-8adf-717206cd8eee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-Language[[:space:]]Models[[:space:]]Struggle[[:space:]]to[[:space:]]Align[[:space:]]Entities[[:space:]]across[[:space:]]Modalities/92f4f7f1-12eb-449d-95cc-ee35c0e15c4c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vision-aided[[:space:]]Unsupervised[[:space:]]Constituency[[:space:]]Parsing[[:space:]]with[[:space:]]Multi-MLLM[[:space:]]Debating/3032d439-3564-4df7-b147-b8c7633b1e74_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visual[[:space:]]Cues[[:space:]]Enhance[[:space:]]Predictive[[:space:]]Turn-Taking[[:space:]]for[[:space:]]Two-Party[[:space:]]Human[[:space:]]Interaction/ffa2ed89-7503-44c0-981c-120d7e9642e3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Visualising[[:space:]]Policy-Reward[[:space:]]Interplay[[:space:]]to[[:space:]]Inform[[:space:]]Zeroth-Order[[:space:]]Preference[[:space:]]Optimisation[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/367b5d93-5d3c-4370-8f3e-2598dd5683e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Voting[[:space:]]or[[:space:]]Consensus_[[:space:]]Decision-Making[[:space:]]in[[:space:]]Multi-Agent[[:space:]]Debate/15ad9f8f-3c2f-4421-bd84-8511b0772a91_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Vulnerability[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Models[[:space:]]to[[:space:]]Prompt[[:space:]]Template[[:space:]]Stealing_[[:space:]]A[[:space:]]Differential[[:space:]]Evolution[[:space:]]Approach/adc610c1-ab58-483b-8c47-68c2ef1da8d5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WASA_[[:space:]]WAtermark-based[[:space:]]Source[[:space:]]Attribution[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Model-Generated[[:space:]]Data/6962217d-484d-471a-a77b-64109f2f6144_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WMT24++_[[:space:]]Expanding[[:space:]]the[[:space:]]Language[[:space:]]Coverage[[:space:]]of[[:space:]]WMT24[[:space:]]to[[:space:]]55[[:space:]]Languages[[:space:]]&[[:space:]]Dialects/7078cfe8-8737-4673-ba8d-0d4b7a14b0d7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WXImpactBench_[[:space:]]A[[:space:]]Disruptive[[:space:]]Weather[[:space:]]Impact[[:space:]]Understanding[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Models/40f99b59-942b-4b0b-91a3-67ba720ddf21_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Wanda++_[[:space:]]Pruning[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Regional[[:space:]]Gradients/78a4dfe1-c32a-4ec6-b017-0ce0d23249c3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/War[[:space:]]of[[:space:]]Thoughts_[[:space:]]Competition[[:space:]]Stimulates[[:space:]]Stronger[[:space:]]Reasoning[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/01a3776c-3279-4a9c-8ac1-5061b637ee25_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Weak-to-Strong[[:space:]]Honesty[[:space:]]Alignment[[:space:]]via[[:space:]]Learning-to-Rank[[:space:]]Supervision/d49352f0-2883-4cb5-91a7-ade20fbfb13d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WebNLG-IT_[[:space:]]Construction[[:space:]]of[[:space:]]an[[:space:]]aligned[[:space:]]RDF-Italian[[:space:]]corpus[[:space:]]through[[:space:]]Machine[[:space:]]Translation[[:space:]]techniques/10d4f66f-0e22-4cec-8b77-91e25c993c73_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/WebUIBench_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]WebUI-to-Code/c1f34757-5027-43b2-89fd-c6da9fb54ad4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Weed[[:space:]]Out,[[:space:]]Then[[:space:]]Harvest_[[:space:]]Dual[[:space:]]Low-Rank[[:space:]]Adaptation[[:space:]]is[[:space:]]an[[:space:]]Effective[[:space:]]Noisy[[:space:]]Label[[:space:]]Detector[[:space:]]for[[:space:]]Noise-Robust[[:space:]]Learning/20ca3c5d-ccd1-4a7c-858c-8bff934ed199_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Well[[:space:]]Begun[[:space:]]is[[:space:]]Half[[:space:]]Done_[[:space:]]Low-resource[[:space:]]Preference[[:space:]]Alignment[[:space:]]by[[:space:]]Weak-to-Strong[[:space:]]Decoding/4b2727f4-5371-46a7-b96d-ebaf84019648_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/What[[:space:]]Language[[:space:]]Do[[:space:]]Non-English-Centric[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Think[[:space:]]in_/6eaf920e-b203-49ea-a9d6-84d6cd602848_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/What[[:space:]]is[[:space:]]in[[:space:]]a[[:space:]]name_[[:space:]]Mitigating[[:space:]]Name[[:space:]]Bias[[:space:]]in[[:space:]]Text[[:space:]]Embedding[[:space:]]Similarity[[:space:]]via[[:space:]]Anonymization/035414c8-edfd-4fbb-9cfd-6719b3ea73d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/When[[:space:]]Benchmarks[[:space:]]Talk_[[:space:]]Re-Evaluating[[:space:]]Code[[:space:]]LLMs[[:space:]]with[[:space:]]Interactive[[:space:]]Feedback/e3568129-d43e-4d1c-ba4b-74aeef1bc257_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/UBench_ Benchmarking Uncertainty in Large Language Models with Multiple Choice Questions/9fd6e574-1e70-4a26-a6f0-5806484e6229_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42f4993ec9ab963b414f32177837e4ca640ce8f5e68a0521cf32d9b7a54c81d7
+size 1197177

2025/UBench_ Benchmarking Uncertainty in Large Language Models with Multiple Choice Questions/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d6735e14904616f600cea3da412565575e6bafb30a1915004e7ae2e5c3c372d
+size 2963216

2025/UCS-SQL_ Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL/5e130ac7-c9a3-4216-9a37-f02974ca9352_content_list.json

@@ -0,0 +1,1661 @@
+[
+    {
+        "type": "text",
+        "text": "UCS-SQL: Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL",
+        "text_level": 1,
+        "bbox": [
+            206,
+            89,
+            789,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhenhe Wu $^{1,2\\dagger}$ , Zhongqiu Li $^{2\\dagger}$ , Jie Zhang $^{2}$ , Zhongjiang He $^{2*}$ , Jian Yang $^{1*}$ , Yu Zhao $^{2}$ , Ruiyu Fang $^{2}$ , Bing Wang $^{1}$ , Hongyan Xie $^{1}$ , Shuangyong Song $^{2}$ , Zhoujun Li $^{1*}$ ,",
+        "bbox": [
+            137,
+            145,
+            865,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Beihang University, <sup>2</sup>TeleAI, China Telecom Corp Ltd,",
+        "bbox": [
+            270,
+            179,
+            724,
+            197
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{wuzhenhe,jiaya,bingwang,by2406233,lizj}@buaa.edu.cn",
+        "bbox": [
+            235,
+            198,
+            761,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{lizq48,zhangj157,hezj,zhaoy11,fangry,songshy}@chinatelecom.cn",
+        "bbox": [
+            184,
+            214,
+            811,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "With the rapid advancement of large language models (LLMs), recent researchers have increasingly focused on the superior capabilities of LLMs in text/code understanding and generation to tackle text-to-SQL tasks. Traditional approaches adopt schema linking to first eliminate redundant tables and columns and prompt LLMs for SQL generation. However, they often struggle with accurately identifying corresponding tables and columns, due to discrepancies in naming conventions between natural language questions (NL) and database schemas. Besides, existing methods overlook the challenge of effectively transforming structure information from NL into SQL. To address these limitations, we introduce UCS-SQL, a novel text-to-SQL framework, uniting both content and structure pipes to bridge the gap between NL and SQL. Specifically, the content pipe focuses on identifying key content within the original content, while the structure pipe is dedicated to transforming the linguistic structure from NL to SQL. Additionally, we strategically selects few-shot examples by considering both the SQL Skeleton and Question Expression (SS-QE selection method), thus providing targeted examples for SQL generation. Experimental results on BIRD and Spider demonstrate the effectiveness of our UCS-SQL framework.",
+        "bbox": [
+            144,
+            288,
+            460,
+            701
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            715,
+            258,
+            730
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Text-to-SQL is the task of translating natural language queries into SQL statements, which has garnered extensive research attention and practical application in database querying (Qin et al., 2022; Sun et al., 2023). Based on the representations encoded by BERT-style pre-trained models (Liu et al., 2023), abstract syntax trees (Wu et al., 2023; Guo et al., 2019; Wang et al., 2020) and predefined query sketches (He et al., 2019) are implemented",
+        "bbox": [
+            112,
+            740,
+            489,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/ae7941b8897c17276b9c737262d260a18ad5d6904546fcd575e1dfca99e5ebbf.jpg",
+        "image_caption": [
+            "Figure 1: Traditional pipelines directly generate SQL query from NL question. We decompose text-to-SQL task into content and structure dimensions, enabling key content extraction and linguistic structure transformation to collaboratively enhance SQL generation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            267,
+            873,
+            419
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "for query decoding. Further, some works extract generalized question-to-SQL patterns by training encoder-decoder models on text-to-SQL corpora (Hui et al., 2022; Li et al., 2023a,b; Zheng et al., 2022; Gao et al., 2024a). Inspired by the powerful capabilities of large language model (LLM) in handling complex reasoning (Wei et al., 2022; Yao et al., 2023), the recent studies have achieved promising results by developing prompt engineering for LLMs (Dong et al., 2023; Pourreza and Rafiei, 2023; Gao et al., 2024a). For example, implementing schema linking to eliminate redundant information, followed by guiding LLMs to generate SQL queries through in-context learning processes.",
+        "bbox": [
+            507,
+            518,
+            884,
+            743
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The forefront researches adhere to the retrieve-then-generate paradigm. MCS-SQL (Lee et al., 2024) exploits LLMs' extensive sample libraries for schema linking and employs multiple prompts to elicit diverse LLM responses. RSL-SQL (Cao et al., 2024) integrates bidirectional schema linking, contextual information augmentation, binary selection strategy, and multi-turn self-correction to achieve state-of-the-art performance. However, the traditional retrieve-then-generate pipeline in txt-to-SQL tasks still confronts two challenges. (1) Ex",
+        "bbox": [
+            507,
+            744,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author.",
+        "bbox": [
+            141,
+            894,
+            287,
+            906
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "†These authors contributed equally to this work.",
+        "bbox": [
+            142,
+            906,
+            435,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8156",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2025, pages 8156-8168",
+        "bbox": [
+            228,
+            945,
+            768,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 27 - August 1, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            268,
+            958,
+            727,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/2012ab466d72e1a8cd18bc8bc545d4dcf1ce3c07f826ad0e78da538e38962078.jpg",
+        "image_caption": [
+            "Figure 2: The framework of UCS-SQL, which consists of three stages uniting content and structure pipe. In question stage, LLM extracts content and structure information from the original question. Transitional stage facilitates the deduction from information to derive the candidate identifiers and functions. In SQL stage, they are combined to form the final query."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            171,
+            85,
+            831,
+            269
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "isting methods struggle with precise schema linking and identifier selection (e.g., column and table names) due to terminological mismatches between colloquial questions and schema identifiers. (2) Moreover, they seldom address the extraction and transformation of structural information from the natural language (NL) question to the SQL query, while bridging the gap of different linguistic structures is crucial for effective query generation. To address the aforementioned challenges, we plan to propose a more comprehensive deductive approach for in-context learning, emphasizing both content and structural dimensions. Specifically, in the content dimension, we deduce SQL identifiers step by step, thereby transforming the original content into key content. In the structure dimension, we derive the SQL skeleton from the question, facilitating the linguistic structure from NL to SQL. Figure 1 demonstrates the integration of question comprehension and SQL query generation from content and structure pipes.",
+        "bbox": [
+            115,
+            363,
+            490,
+            702
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce UCS-SQL, a three-stage text-to-SQL framework that integrates content and structure pipes to enhance semantic bridging. The content pipe extracts key information from the raw question, while the structure pipe converts NL into SQL syntax. Figure 2 illustrates the UCS-SQL framework. In the question stage, a LLM extracts content and structure information from the raw question. Then in the transitional stage, we prompt the LLM to transform them into SQL identifiers and functions. The content pipe integrates evidence and foreign keys to refine content information and performs bi-directional schema link-",
+        "bbox": [
+            112,
+            712,
+            490,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ing to identify relevant schema identifiers, while the structure pipe infers SQL functions based on the extracted structure information. Finally, in the SQL stage, we construct a SQL skeleton using the inferred functions and fill it with the identified identifiers to generate the final query. Additionally, we develop an SS-QE algorithm to select few-shot examples for the LLM, considering similarities in both SQL skeletons and question expressions. Comprehensive evaluations on the BIRD and Spider datasets demonstrate that UCS-SQL outperforms several baselines in terms of Valid Efficiency Score (VES) and Execution Accuracy (EX) across different difficulty levels. It achieves top-2 performance on all evaluation metrics on BIRD and top-3 performance on Spider.",
+        "bbox": [
+            507,
+            363,
+            884,
+            621
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, our contributions are as follows:",
+        "bbox": [
+            527,
+            621,
+            882,
+            636
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose UCS-SQL, a three-stage text-to-SQL framework that unites content and structure pipes to respectively extract key content and transform linguistic structure, thereby collaboratively enhancing SQL query generation.",
+            "- We introduce SS-QE algorithm to select few-shot examples, which utilizes the structure information from UCS-SQL and takes into account the similarities in both SQL structure and question expression.",
+            "- Empirical evaluations indicate that UCS-SQL framework attains  $3^{nd}$  place in all evaluation metrics on BIRD and Spider. Ablation experiments demonstrate that all three stages and both pipes in the UCS-SQL are crucial for performance enhancement."
+        ],
+        "bbox": [
+            531,
+            645,
+            884,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8157",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Problem Definition",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            312,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Text-to-SQL is the task of converting a natural language question  $Q$  into a correct SQL query  $Y$ . The database can be represented as  $D = \\{T_{1}, T_{2} \\dots T_{m}\\}$ ,  $m$  is the number of tables in the database. For  $T = \\{C_{1}, C_{2} \\dots C_{n}\\}$ ,  $C_{i}$  refers to columns in table  $T$ ,  $n$  is the number of columns in the table. When dealing with complex database values, we can use external knowledge evidence  $K$  to support our model understand the inner relationship between question and database. The process could be formulated as follows:",
+        "bbox": [
+            112,
+            108,
+            489,
+            282
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nY = f (Q, D, K \\mid \\theta) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            221,
+            294,
+            487,
+            311
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $f(\\cdot |\\theta)$  represent a model with parameters  $\\theta$ .",
+        "bbox": [
+            112,
+            319,
+            487,
+            336
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 UCS-SQL",
+        "text_level": 1,
+        "bbox": [
+            112,
+            344,
+            236,
+            362
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Overview",
+        "text_level": 1,
+        "bbox": [
+            112,
+            370,
+            235,
+            385
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We introduce UCS-SQL, a parallel text-to-SQL framework that integrates content and structure pipes to enhance semantic bridging. As illustrated in Figure 2, the framework comprises three stages: question, transitional, and SQL stages. (1) In the question stage, we prompt an LLM to extract content information and structure information from the original question. (2) The transitional stage acts as a bridge between the question and SQL stages. We combine the content information with the database schema to obtain candidate identifiers in the content pipeline and deduce SQL functions from the structure information in the structure pipeline. (3) In the SQL stage, we first generate an SQL skeleton based on the inferred SQL functions and the original question. Then, we fill the skeleton with the candidate identifiers to produce the final SQL query. Besides, a novel approach is proposed for selecting few-shot examples, considering similarities in both SQL structure and question expression. The following section provides a detailed introduction to UCS-SQL and the example selection method.",
+        "bbox": [
+            112,
+            391,
+            489,
+            745
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Stage 1: Question Stage",
+        "text_level": 1,
+        "bbox": [
+            112,
+            756,
+            346,
+            771
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "First, we perform pre-processing on the database schema prior by decomposing the original database into smaller sub-databases using the selector (Wang et al., 2023) to minimize interference from irrelevant tables and columns. Additionally, we maintain column descriptions for each column in the pre-filtered schema.",
+        "bbox": [
+            112,
+            776,
+            487,
+            887
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The question stage is designed to extract key content and structure from the original question.",
+        "bbox": [
+            112,
+            889,
+            487,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4e644cea24497478fad34678b5521355ddb18b7433a050c963bce3d2e1954d3.jpg",
+        "image_caption": [
+            "Figure 3: The question stage of UCS-SQL. The original database is pre-processed to filter out irrelevant tables and columns. Terms representing content and structure information are extracted from the raw question."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            522,
+            84,
+            872,
+            237
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we employ LLMs to identify significant content information (e.g., table names and column names) and structure information from the target question to facilitate SQL generation. In the subsequent operations, UCS-SQL proceeds via two parallel pipes, respectively handling SQL generation at the content and structure levels. Figure 3 illustrates the process of the question stage. We prompt the LLM to identify content information and structure information based on the original phrase of the question, which we highlight in red and blue colors. These identified elements are extracted and integrated to advance both parallel pipelines to the next stage. It is important to note that the LLM may identify the same word as both content and structure information, such as the word \"average\" shown in Figure 3. This dual identification aligns with our objective at the question stage, retaining sufficient relevant information. Further reasoning and selection will be conducted in subsequent stages.",
+        "bbox": [
+            505,
+            321,
+            884,
+            659
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 Stage 2: Transitional Stage",
+        "text_level": 1,
+        "bbox": [
+            507,
+            673,
+            766,
+            688
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The transitional stage is designed to transform the original information extracted from the question into SQL language elements. Figure 4 illustrates the process of this stage.",
+        "bbox": [
+            507,
+            695,
+            880,
+            758
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the content pipe, we first integrate the content information with evidence to convert colloquial phrases into schema-compatible identifiers. Subsequently, we perform foreign key expansion to augment the candidate set of potential identifiers for SQL generation. Specifically, we prompt the LLM to identify relevant foreign keys and incorporate the corresponding identifiers into the candidate set. Next, we conduct bi-directional schema linking by combining the schema with the candidate set,",
+        "bbox": [
+            507,
+            760,
+            882,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8158",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/f46b127fa11ddccff56f277b534c514936e39be0d2f581d8ceb011a9eaa87167.jpg",
+        "image_caption": [
+            "Figure 4: The transitional stage of UCS-SQL. Candidate identifiers and functions are obtained in the content and structure pipes. Bi-directional schema linking is performed with reference to the identifiers, yielding the bi-filtered schema."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            82,
+            472,
+            239
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "which both deduces potential identifiers and filters out irrelevant columns from the schema. We match the content information with schema identifiers: if a close match with a table or column name is found, the corresponding identifier is retained. In cases where no direct match exists, the LLM searches for the most similar expression among column descriptions and retains the associated identifiers. Through careful analysis and deduction, we obtain a refined set of identifiers. For instance, the final identifiers reveal that column \"A4\" is irrelevant to the question, so we remove \"A4\" from the schema to arrive at the bi-filtered database schema.",
+        "bbox": [
+            112,
+            341,
+            487,
+            548
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the structure pipe, we use the LLM to induce SQL functions based on the structure information extracted during the question stage. As shown in Figure 4, phrases such as \"the youngest\" and \"the lowest\" suggest the potential use of SQL functions like \"ORDER BY\", \"ASC\", and \"DESC\" while \"average\" implies the \"AVG\" function. In summary, we distill these insights into a final set of functions.",
+        "bbox": [
+            110,
+            551,
+            489,
+            678
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Stage 3: SQL Stage",
+        "text_level": 1,
+        "bbox": [
+            112,
+            690,
+            314,
+            707
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the SQL stage, we first create the SQL skeleton by referring to the candidate functions derived from the structure pipe. Subsequently, we integrate the identifiers from the content pipe with the bi-filtered schema to fill the generated skeleton. As a result, by unifying the structure and content pipes, we ultimately produce the final SQL query.",
+        "bbox": [
+            112,
+            711,
+            487,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Figure 5, we prompt the LLM to generate an SQL skeleton based on the constraints, question context, and all available structure information, with particular emphasis on the candidate functions from the structure pipeline. The LLM selects functions such as \"ORDER BY\", \"ASC\", and \"DESC\"",
+        "bbox": [
+            112,
+            825,
+            489,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/71f6a5acf007c3d226e5dd5a087b5401d394fb2d6ab5bf4720b4bac055512d5d.jpg",
+        "image_caption": [
+            "Figure 5: The SQL stage of UCS-SQL. The SQL skeleton is inferred from functions. The final SQL query is constructed by incorporating the identifiers into the SQL skeleton. The bi-filtered schema serves as a secondary reference to address cases where the identifiers are not correctly extracted."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            531,
+            83,
+            862,
+            239
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "to form the SQL skeleton. Next, we fill the skeleton using the identifiers from the content pipeline, focusing on the recommended identifiers. The LLM identifies \"client\" and \"district\" as table names and \"gender\", \"birth_date\", \"A11\", and \"district_id\" as column names to complete the final SQL query.",
+        "bbox": [
+            505,
+            351,
+            882,
+            449
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.5 SS-QE Example Selection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            463,
+            757,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We propose a novel SS-QE few-shot example selection approach, which considers the similarities in both SQL structure and question expression. We present the detailed algorithm of our selection method in Algorithm 1. We first extract final identifiers and final functions from all the question  $q_{i}$  in our training set  $Q$  following the process of question stage and transitional stage.  $F_{i}$  is the set of final functions,  $I_{i}$  is the set of final identifiers, while  $c_{i}$  is the complexity of  $q_{i}$ , which we define as the sum of final identifiers and final functions:",
+        "bbox": [
+            505,
+            486,
+            882,
+            661
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nc _ {i} = \\left| I _ {i} \\right| + \\left| F _ {i} \\right| \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            636,
+            678,
+            882,
+            696
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "First, we evaluate the similarity between  $q_{i}$  and  $q$  in terms of SQL structure, using  $F_{i}$  and  $c_{i}$  as our basis. Specifically, samples with a larger intersection in their final functions and closer complexity values are considered to have a higher degree of SQL structure similarity. To filter these samples, we establish thresholds  $\\theta_{1}$  and  $\\theta_{2}$ . Subsequently, we add each sample that satisfies these criteria to the set  $S_{1}$ , continuing this process until  $S_{1}$  contains  $2k$  samples. Next, we further utilize LLM to pick top-  $k$  examples from  $S_{1}$  based on their similarity with  $q$  in question expression, Figure 6 shows the selection prompt in detail. As the result, we obtain",
+        "bbox": [
+            507,
+            712,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8159",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1: SS-QE Example selection"
+        ],
+        "code_body": "Input: All the questions  $Q$  in training set, target question  $q$ , the number of few-shot  $k$ , threshold  $\\theta_{1}$  &  $\\theta_{2}$   \nOutput: Top- $k$  questions  $S$   \n1  $S_{1} \\gets \\emptyset$ ;  \n2  $I, F \\gets LLM_{InformationExtraction}(q)$ ;  \n3  $c \\gets |I| + |F|$ ;  \n4 foreach  $q_{i}$  in  $Q$  do  \n5  $I_{i}, F_{i} \\gets LLM_{InformationExtraction}(q_{i})$ ;  \n6  $c_{i} \\gets |I_{i}| + |F_{i}|$ ;  \n7 if  $(F \\cap F_{i} \\geq \\theta_{1}) \\land (|c - c_{i}| < \\theta_{2})$  then  \n8 Add  $q_{i}$  in  $S_{1}$ ;  \n9 end  \n10 if  $|S_{1}| = 2k$  then  \n11 break;  \n12 end  \n13 end  \n14 S ← LLMSimilarQuestionExpression(q, k, S1);  \n15 return S;",
+        "bbox": [
+            112,
+            105,
+            485,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$k$ -shot examples that have similar SQL structure and question expression.",
+        "bbox": [
+            112,
+            476,
+            485,
+            508
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            520,
+            339,
+            537
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            546,
+            226,
+            560
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "BIRD (Li et al., 2023c) represents a cross-domain dataset that examines the impact of extensive database contents on text-to-SQL parsing. Spider (Yu et al., 2018) is a large-scale complex and cross-domain semantic parsing and text-to-SQL dataset. We report the statistics of datasets in appendix A.",
+        "bbox": [
+            112,
+            567,
+            487,
+            663
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            112,
+            675,
+            310,
+            689
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Following BIRD (Li et al., 2023c), we utilize execution accuracy (EX) and valid efficiency score (VES) to evaluate text-to-SQL models. EX (Li et al., 2023c) is defined as the proportion of questions in the evaluation set for which the execution results of both the predicted and ground-truth inquiries are identical. VES (Li et al., 2023c) is designed to measure the efficiency of valid SQLs generated by models.",
+        "bbox": [
+            112,
+            696,
+            487,
+            840
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            852,
+            233,
+            866
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "DIN-SQL(Pourreza and Rafiei, 2023) and MAC-SQL(Wang et al., 2023) breaks down intricate queries into manageable sub-tasks.",
+        "bbox": [
+            112,
+            873,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Similar Question Expression Selection Prompt"
+        ],
+        "code_body": "Select three questions that are closest in question phrasing to the target question from the source questions and return the list which is contain the three questions ids. Format of the returned list is as follows: [3, 672, 19]  \ntarget question: {question}  \nsource questions: {{source_questions}}  \nreturned list: {result_ids}",
+        "bbox": [
+            526,
+            110,
+            860,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Figure 6: The prompt for selecting examples having similar question expression with target questions.",
+        "bbox": [
+            507,
+            233,
+            880,
+            263
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "E-SQL(Caferoglu and Ulusoy, 2024) and CHESS(Talaei et al., 2024) seek to bridge the divide between natural language queries and database architectures. SQL-PaLM(Sun et al., 2023) and SuperSQL(Li et al., 2024a) utilize distinct prompting and fine-tuning methods for adapting large language models (LLMs) in SQL generation. TA-SQL(Qu et al., 2024) and CodeS(Li et al., 2024b) introduce strategies to reduce hallucinations in LLM-based SQL generation. DAIL-SQL(Gao et al., 2024a) is designed to tackle complex database environments. Additionally, methods based on multi-stage strategies are included, such as DTS-SQL(Pourreza and Rafiei, 2024) that employs two-stage fine-tuning, MAG-SQL(Xie et al., 2024) that adopts a multi-agent generative approach, and MCS-SQL(Lee et al., 2024) that utilizes multiple prompts and multiple-choice selection. MSc-SQL(Gorti et al., 2024) narrows the performance gap of smaller open-source models by sampling and comparing multiple SQL query results. RSL-SQL(Cao et al., 2024) achieve robust schema linking that maximize the benefits. CHASE-SQL (Pourreza et al., 2024) propose multiple chain-of-thought prompting methods and an online synthetic example generation technique. XiYan-SQL (Gao et al., 2024b) integrates the ICL approach to maximize the generation of high-quality and diverse SQL candidates.",
+        "bbox": [
+            507,
+            279,
+            884,
+            762
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            778,
+            746,
+            795
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We use a single Tesla V100 GPU with 32 GiBs of memory on a server to restore intermediate data and execute SS-QE example selection algorithm, the retrieving time is about  $3\\sim 6$  seconds for each sample. All the experiments utilize GPT-4-Turbo, the context window is 128000, the temperature is set to 0.1. We enable five threads to run UCS-SQL",
+        "bbox": [
+            507,
+            808,
+            882,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8160",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/16b5f6f041e5dd4c50a05a5fe012ea76c921d8f94fa2ecc136a336ec6367f794.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Model</td><td colspan=\"4\">EX</td><td rowspan=\"2\">VES</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td>DIN-SQL</td><td>GPT-4</td><td>-</td><td>-</td><td>-</td><td>50.72</td><td>58.79</td></tr><tr><td>DAIL-SQL</td><td>GPT-4</td><td>-</td><td>-</td><td>-</td><td>54.76</td><td>56.08</td></tr><tr><td>DTS-SQL</td><td>DeepSeek-7B</td><td>-</td><td>-</td><td>-</td><td>55.80</td><td>-</td></tr><tr><td>TA-SQL</td><td>GPT-4</td><td>63.14</td><td>48.60</td><td>36.11</td><td>56.19</td><td>-</td></tr><tr><td>Codes</td><td>Codes-15B</td><td>-</td><td>-</td><td>-</td><td>58.47</td><td>59.87</td></tr><tr><td>SuperSQL</td><td>GPT-4</td><td>66.90</td><td>46.50</td><td>43.80</td><td>58.50</td><td>61.99</td></tr><tr><td>MAC-SQL</td><td>GPT-4</td><td>65.73</td><td>52.69</td><td>40.28</td><td>59.39</td><td>66.39</td></tr><tr><td>MAG-SQL</td><td>GPT-4</td><td>-</td><td>-</td><td>-</td><td>61.08</td><td>-</td></tr><tr><td>SQL-Palm</td><td>PaLM2</td><td>68.92</td><td>52.07</td><td>47.89</td><td>61.93</td><td>-</td></tr><tr><td>MCS-SQL</td><td>GPT-4</td><td>70.40</td><td>53.10</td><td>51.40</td><td>63.36</td><td>64.80</td></tr><tr><td>CHESS</td><td>Proprietary</td><td>-</td><td>-</td><td>-</td><td>65.00</td><td>-</td></tr><tr><td>E-SQL</td><td>GPT-4o</td><td>-</td><td>-</td><td>-</td><td>65.58</td><td>-</td></tr><tr><td>MSc-SQL</td><td>GPT-4o</td><td>72.00</td><td>58.00</td><td>49.00</td><td>65.60</td><td>-</td></tr><tr><td>RSL-SQL</td><td>GPT-4o</td><td>74.38</td><td>57.11</td><td>53.79</td><td>67.21</td><td>70.32</td></tr><tr><td>CHASE-SQL</td><td>Gemini 1.5</td><td>-</td><td>-</td><td>-</td><td>73.01</td><td>-</td></tr><tr><td>XiYan-SQL</td><td>-</td><td>-</td><td>-</td><td>-</td><td>73.34</td><td>-</td></tr><tr><td>UCS-SQL</td><td>DeepSeek</td><td>71.67</td><td>55.59</td><td>54.11</td><td>65.23</td><td>68.49</td></tr><tr><td>UCS-SQL</td><td>GPT-4o</td><td>75.42</td><td>59.52</td><td>56.93</td><td>68.95</td><td>71.10</td></tr></table>",
+        "bbox": [
+            196,
+            83,
+            801,
+            375
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/0c7f7218386b8e2b2f04c41f9a80376f5946efe27f26ce097cf650c9781886c3.jpg",
+        "table_caption": [
+            "Table 1: EX and VES on dev set of BIRD. UCS-SQL achieves top two performances in both VES and EX across datasets of varying difficulty levels."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Model</td><td>EX (dev)</td><td>EX (test)</td></tr><tr><td>DIN-SQL</td><td>GPT-4</td><td>82.8</td><td>85.3</td></tr><tr><td>DAIL-SQL</td><td>GPT-4</td><td>84.4</td><td>86.6</td></tr><tr><td>DTS-SQL</td><td>DeepSeek-7B</td><td>85.5</td><td>84.4</td></tr><tr><td>TA-SQL</td><td>GPT-4</td><td>85.0</td><td>-</td></tr><tr><td>MAC-SQL</td><td>GPT-4</td><td>86.8</td><td>82.8</td></tr><tr><td>MAG-SQL</td><td>GPT-4</td><td>85.3</td><td>85.6</td></tr><tr><td>MCS-SQL</td><td>GPT-4</td><td>89.5</td><td>89.6</td></tr><tr><td>CHESS</td><td>Openllms</td><td>-</td><td>87.2</td></tr><tr><td>MSc-SQL</td><td>GPT-4o</td><td>-</td><td>84.7</td></tr><tr><td>RLS-SQL</td><td>GPT-4o</td><td>-</td><td>87.9</td></tr><tr><td>CHASE-SQL</td><td>Gemini 1.5</td><td>-</td><td>87.6</td></tr><tr><td>XiYan-SQL</td><td>-</td><td>-</td><td>89.6</td></tr><tr><td>UCS-SQL</td><td>DeepSeek</td><td>86.5</td><td>87.7</td></tr><tr><td>UCS-SQL</td><td>GPT-4o</td><td>87.3</td><td>88.0</td></tr></table>",
+        "bbox": [
+            142,
+            439,
+            460,
+            633
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: UCS-SQL achieves top three performances in EX on dev and test set of Spider.",
+        "bbox": [
+            112,
+            643,
+            485,
+            671
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "(approximately  $200\\sim 500$  samples for each according to the size of dataset), it costs about  $4\\sim 6$  hours to generate all results. The token consumption of 2-shot UCS-SQL on dev set for BIRD and Spider are approximately 5.29 million and 3.47 million in input, 0.9 and 0.5 million in output. We use all MiniLM-L6-v2 model to get sentence embedding and calculate cosine similarity in the discussion. In SS-QE selection method, we set  $\\{\\theta_1,\\theta_2\\}$  to  $\\{2,3\\}$ .",
+        "bbox": [
+            112,
+            686,
+            487,
+            829
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            843,
+            326,
+            858
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "6.1 Overall Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            869,
+            282,
+            882
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The overall results of all the baselines and our proposed UCS-SQL on BIRD and Spider are shown",
+        "bbox": [
+            112,
+            889,
+            487,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "in Table 1 and Table 2. The experiment results indicate that our proposed UCS-SQL achieves better performance than several competitive baselines on the two datasets.",
+        "bbox": [
+            507,
+            439,
+            882,
+            502
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In Table 1, we report the performance of UCS-SQL and other competitive baselines on development set of BIRD. These recent methods have their own distinct characteristics and utilize different LLM models, such as GPT-4, GPT-4o, and DeepSeek. Experiments show that our proposed UCS-SQL method almost outperforms the first thirteen baselines by at least  $3.35\\%$  on EX and by at least  $4.71\\%$  on VES, only slightly trailing the concurrent work RSL-SQL. In addition, UCS-SQL performs better than RSL-SQL on samples of moderate and challenging difficulty, which indicates that UCS-SQL can address problems more specifically when dealing with samples of higher complexity. On the other hand, Table 2 shows the execution accuracy of UCS-SQL and other baselines on Spider. Our proposed UCS-SQL achieves excellent results of  $87.3\\%$  and  $88.0\\%$  on dev and test sets, surpassing the vast majority of baselines. In conclusion, UCS-SQL respectively achieves top-3 results in VES and EX across different difficulty on BIRD and Spider datasets, and ranked first on high-difficulty data, demonstrating the method's high efficiency in handling complex examples and its generalizability across different scenarios.",
+        "bbox": [
+            507,
+            519,
+            882,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8161",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/8edd7a933911bd3f37f32060bbb862970b566874ab124bf95e7610429b137bf2.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"4\">EX</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td colspan=\"5\">Content &amp; Structure Pipes Dimension (fixed 2-shot)</td></tr><tr><td>w/o content pipe</td><td>69.81</td><td>54.98</td><td>50.48</td><td>63.58(↓1.24)</td></tr><tr><td>w/o structure pipe</td><td>70.45</td><td>54.54</td><td>49.77</td><td>63.78(↓1.04)</td></tr><tr><td>w/ both pipes</td><td>70.24</td><td>58.25</td><td>50.48</td><td>64.82</td></tr><tr><td colspan=\"5\">Three Stages Dimension (fixed 2-shot)</td></tr><tr><td>w/o question &amp; transitional stage</td><td>70.13</td><td>53.24</td><td>47.66</td><td>63.00(↓1.82)</td></tr><tr><td>w/o transitional stage</td><td>69.59</td><td>56.07</td><td>49.07</td><td>63.65(↓1.17)</td></tr><tr><td>w/o SQL stage</td><td>69.70</td><td>55.42</td><td>48.66</td><td>63.39(↓1.43)</td></tr><tr><td>w/ three stages</td><td>70.24</td><td>58.25</td><td>50.48</td><td>64.82</td></tr><tr><td colspan=\"5\">Few-shot Example Selection</td></tr><tr><td>0-shot</td><td>67.45</td><td>53.89</td><td>44.14</td><td>61.24(↓7.71)</td></tr><tr><td>2-shot (fixed)</td><td>70.24</td><td>58.25</td><td>50.48</td><td>64.82(↓4.13)</td></tr><tr><td>2-shot (SS-QE)</td><td>71.99</td><td>58.81</td><td>52.89</td><td>66.28(↓2.67)</td></tr><tr><td>5-shot (SS-QE)</td><td>75.42</td><td>59.52</td><td>56.93</td><td>68.95</td></tr></table>",
+        "bbox": [
+            176,
+            83,
+            826,
+            324
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Ablation study on dev set of BIRD. The three groups of experiments respectively compare the impact of two pipes, three stages, and different few-shot example selections on UCS-SQL.",
+        "bbox": [
+            112,
+            342,
+            882,
+            373
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "6.2 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            385,
+            278,
+            400
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the contributions of different components in UCS-SQL, we conducted a series of ablation studies using the BIRD dataset, which is a large-scale dataset with complex database structures. The experiments were divided into three groups: (1) assessing the roles of the content pipe and structure pipe, (2) analyzing the contributions of the three stages, and (3) comparing the effectiveness of the SS-QE example selection method. The results are summarized in Table 3.",
+        "bbox": [
+            112,
+            405,
+            487,
+            565
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Group 1: Content Pipe vs. Structure Pipe Removing either the content pipe or the structure pipe from the framework led to a decrease in execution accuracy across datasets of varying difficulty, with respective drops of  $1.24\\%$  and  $1.04\\%$ . The results indicate that integrating both content and structure information is crucial for guiding LLMs to generate SQL queries effectively.",
+        "bbox": [
+            112,
+            567,
+            487,
+            695
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Group 2: Contributions of the Three Stages The ablation studies also examined the necessity of the three stages in UCS-SQL. Removing the transitional stage or the SQL stage resulted in a decrease in execution accuracy by  $1.17\\%$  and  $1.43\\%$ , respectively. Additionally, removing the question stage further reduced performance by  $0.65\\%$ . These findings demonstrate that each stage plays a vital role in the overall performance of the framework.",
+        "bbox": [
+            112,
+            696,
+            487,
+            839
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Group 3: SS-QE Example Selection Method The effectiveness of the SS-QE example selection method was evaluated by comparing the execution accuracy of 0-shot, fixed 2-shot, SS-QE-selected 2-shot, and SS-QE-selected 5-shot scenarios. The",
+        "bbox": [
+            112,
+            841,
+            487,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "results showed a progressive improvement in performance, with a  $1.46\\%$  increase from fixed 2-shot to SS-QE-selected 2-shot and an additional  $2.67\\%$  improvement when increasing the number of shots from 2 to 5. This indicates that the SS-QE method effectively selects guiding examples that enhance the overall performance of the framework.",
+        "bbox": [
+            507,
+            385,
+            882,
+            495
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The ablation studies demonstrate that both pipes and all three stages in the UCS-SQL framework are essential for performance enhancement. This validates the rationality of integrating content and structure dimensions in a multi-stage framework for SQL generation. Additionally, the SS-QE example selection method further leverages the framework's capabilities by selecting few-shot examples that match the target question's function and complexity, thereby improving the overall performance.",
+        "bbox": [
+            507,
+            499,
+            884,
+            659
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "7 Discussion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            676,
+            636,
+            690
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "7.1 Effect of bi-filtered schema",
+        "text_level": 1,
+        "bbox": [
+            507,
+            705,
+            766,
+            719
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "This section examines the impact of bi-filtered schemas generated through the bi-directional schema linking process, using the BIRD dataset due to its large scale and complexity. We define the schema column count as the total number of columns across all tables in a schema. Figure 7 illustrates the schema column counts for each sample in the BIRD dev set. The \"Overall column counts\" represent raw database schemas, which consist of 11 distinct databases. The \"Gold column counts\" reflect columns used as identifiers in gold SQL queries, setting the performance limits for schema",
+        "bbox": [
+            507,
+            728,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8162",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/bc6ae47972a05c16db6da3523694458c11bf6d3add1cc07fbff31ab5c8b9c7c3.jpg",
+        "image_caption": [
+            "Figure 7: The schema column counts on dev set of BIRD. The bi-filtered column counts have significantly decreased compared to the pre-filtered column counts, and are slightly higher than the gold column counts."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            87,
+            480,
+            244
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/7a853aa02b62d7f1649a750c2ebe50fbd121aea9a43dd994aeb50a87efdae950.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Metric (%)</td><td colspan=\"4\">UCS-SQL</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td colspan=\"5\">Pre-filtered Schema</td></tr><tr><td>Precision</td><td>11.23</td><td>15.78</td><td>20.56</td><td>13.26</td></tr><tr><td>Recall</td><td>97.33</td><td>95.11</td><td>96.19</td><td>96.42</td></tr><tr><td>F1 score</td><td>20.14</td><td>27.07</td><td>33.87</td><td>23.31</td></tr><tr><td colspan=\"5\">Bi-filtered Schema</td></tr><tr><td>Precision</td><td>59.71</td><td>73.81</td><td>78.90</td><td>65.83</td></tr><tr><td>Recall</td><td>89.14</td><td>84.09</td><td>85.44</td><td>86.93</td></tr><tr><td>F1 score</td><td>71.51</td><td>78.62</td><td>82.04</td><td>74.92</td></tr></table>",
+        "bbox": [
+            131,
+            331,
+            475,
+            467
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "linking. The \"Pre-filtered column counts\" denote schemas filtered using the method described earlier, while the \"Bi-filtered column counts\" represent our bi-filtered schemas. The pre-filtered schemas show inconsistent filtering effectiveness, with a significant gap compared to gold schemas. In contrast, bi-filtered schemas achieve substantial filtering, closely approximating gold schemas in terms of column counts.",
+        "bbox": [
+            112,
+            532,
+            489,
+            677
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To further evaluate the effectiveness of bi-filtered schemas, we calculated precision, recall, and F1 score using gold schemas as the standard. Table 4 presents the evaluation results. Pre-filtered schemas have high recall across datasets but poor precision and F1 score (13.26% and 23.31%, respectively). In contrast, bi-filtered schemas achieve a recall of 86.93%, slightly lower than pre-filtered schemas, but with significantly higher precision (65.83%) and F1 score (74.92%).",
+        "bbox": [
+            112,
+            678,
+            489,
+            839
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Bi-filtered schemas effectively eliminate irrelevant columns, closely matching gold schemas in column counts. The high precision, recall, and F1 score demonstrate the superior performance of bi-directional schema linking.",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/12a4fb3ebf4f31cad3e3077674616acb2f415ae90ffe67993d570500f3f18bf5.jpg",
+        "image_caption": [
+            "Figure 8: The evaluation statistics of UCS-SQL using DIAL or SS-QE selection method on dev of BIRD."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            84,
+            882,
+            198
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/69dc8ecea7960ff609601ccb1c29c88f4e4dc220c4748d9c83d4a581784b24db.jpg",
+        "table_caption": [
+            "Table 4: Evaluation of pre-filtered schema and bifiltered schema on dev of BIRD."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">UCS-SQL (5-shot)</td><td colspan=\"4\">EX</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td>w/DAILs</td><td>73.74</td><td>58.94</td><td>51.07</td><td>67.21</td></tr><tr><td>w/SS - QE</td><td>75.42</td><td>59.52</td><td>56.93</td><td>68.95</td></tr></table>",
+        "bbox": [
+            524,
+            252,
+            870,
+            309
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: EX results of UCS-SQL with two example selection method on dev of BIRD.",
+        "bbox": [
+            507,
+            328,
+            882,
+            357
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7.2 Effect of SS-QE method",
+        "text_level": 1,
+        "bbox": [
+            507,
+            372,
+            742,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "This study further examines the impact of the SS-QE selection method using the BIRD dataset. The DAIL Selection method, as reported in (Gao et al., 2024a), is a leading selection technique that pre-generates SQL queries and selects examples based on high similarity between masked questions and masked predicted SQL. We conducted experiments on UCS-SQL using both DAIL and SS-QE methods in a 5-shot setting. Table 5 shows that the SS-QE method outperforms DAIL.",
+        "bbox": [
+            505,
+            395,
+            884,
+            555
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To explore the underlying reasons, we analyzed the selected examples. As shown in Figure 8, the DAIL method slightly outperforms SS-QE in average masked question cosine similarity and masked query cosine similarity. However, the SS-QE method exhibits significantly higher function precision and recall. These results indicate that SS-QE more accurately identifies examples with analogous functions to the target sample, making it more suitable for guiding SQL query generation. Moreover, examples with similar complexity and consistent functions provide a broader selection pool compared to those based solely on high SQL skeleton similarity. Even when the sequence or placement of functions varies, they still offer guiding significance, which would likely be overlooked by similarity-based selection alone.",
+        "bbox": [
+            507,
+            557,
+            882,
+            831
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "8 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            845,
+            665,
+            860
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) have demonstrated significant advancements in various natural language processing (NLP) tasks (Wang et al., 2024a;",
+        "bbox": [
+            507,
+            873,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8163",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Li et al., 2024c; Touvron et al., 2023; OpenAI, 2023; Yang et al., 2024; Shao and Li, 2024; Wang et al., 2024b). Some researchers have leveraged LLMs in text-to-SQL tasks to further enhance performance. A critical aspect of this approach is the design and utilization of prompts, which directly influence the accuracy of SQL generation by guiding LLMs effectively. For example, some methods (Tai et al., 2023) try to improve the inference capabilities of LLMs using chain-of-thought prompting, including both the original chain-of-thought prompt and the least-to-most prompt. Further, a comprehensive analysis (Chang and Fosler-Lussier, 2023) of the impact of prompt construction across various settings in text-to-SQL tasks is studied. DAIL-SQL (Gao et al., 2024a) considers both questions and SQL queries to select few-shot examples, adopts an example organization strategy to balance quality and quantity, and utilizes code representation prompting for question representation. Besides, C3-SQL (Dong et al., 2023) and DIN-SQL (Pourreza and Rafiei, 2023) have introduced innovative frameworks for database simplification, query decomposition, and prompt engineering.",
+        "bbox": [
+            115,
+            84,
+            490,
+            469
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Researchers introduce schema linking to identify the database tables and columns associated with natural language queries and have proposed complex and integrated prompting engineering methods (Li et al., 2024d; Wu et al., 2025). MAC-SQL (Wang et al., 2023; Lee et al., 2024) centered on multi-agent collaboration can be utilized for more intricate data scenarios and a broader spectrum of error types for detection and correction. TA-SQL (Qu et al., 2024) and CodeS (Li et al., 2024b) introduce strategies to mitigate hallucinations in LLM-based SQL generation. CHESS (Talaei et al., 2024) enables more accurate schema linking by retrieving relevant information from database catalogs and database values. Another approach, MAG-SQL (Xie et al., 2024) features a multi-agent generative approach with soft schema linking and iterative Sub-SQL refinement. MSc-SQL (Gorti et al., 2024) mitigates the performance gap of smaller open-source models by sampling and comparing multiple SQL query results. RSL-SQL combines bidirectional schema linking, contextual information augmentation, binary selection strategy, and multiturn self-correction to achieve an efficient framework. But previous works are difficult to accurately retrieve the correct identifiers content from the schema and often overlook the understanding and reasoning of structure hints during",
+        "bbox": [
+            115,
+            470,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "SQL generation.",
+        "bbox": [
+            512,
+            85,
+            633,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "9 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            512,
+            111,
+            638,
+            124
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose UCS-SQL, a three-stage text-to-SQL framework integrating a content pipe for key content extraction and a structure pipe for linguistic structure transformation. Through the three stages of UCL-SQL, the content and structure pipes jointly accomplish the generation from NL question to SQL query. Then, we introduce the SS-QE example selection method, which selects appropriate few-shot examples for target questions by taking into account the similarities in both SQL structure and question expression. The experimental results on the BIRD and Spider datasets demonstrate that UCS-SQL achieves superior performance in EX and VES, and substantiate the effectiveness of both the content and structure pipes. The integration of content and structure pipes will guide future research to jointly consider the text-to-SQL generation process from both dimensions.",
+        "bbox": [
+            512,
+            135,
+            884,
+            424
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            512,
+            436,
+            613,
+            450
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "UCS-SQL is a pipeline based approach, it will be subjected to cascading error effect. An intelligent baketracking approach to reflect and correct the intermediate stage outputs can make the system more robust, which we will design in the future work.",
+        "bbox": [
+            512,
+            460,
+            880,
+            554
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            512,
+            567,
+            658,
+            582
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, all of the datasets, models, code and related documents are not associated with any ethical concerns.",
+        "bbox": [
+            512,
+            592,
+            880,
+            638
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            512,
+            651,
+            670,
+            667
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was partially supported by the National Natural Science Foundation of China (Grant Nos. 62276017, 62406033, U1636211, 61672081), and the State Key Laboratory of Complex& Critical Software Environment (Grant No. SKLCCSE-2024ZX-18).",
+        "bbox": [
+            512,
+            675,
+            880,
+            770
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            512,
+            797,
+            606,
+            812
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Hasan Alp Caferoğlu and Özgür Ulusoy. 2024. E-SQL: direct schema linking via question enrichment in text-to-sql. CoRR, abs/2409.16751.",
+            "Zhenbiao Cao, Yuanlei Zheng, Zhihao Fan, Xiaojin Zhang, Wei Chen, and Xiang Bai. 2024. RSL-SQL: robust schema linking in text-to-sql generation. CoRR, abs/2411.00073."
+        ],
+        "bbox": [
+            512,
+            819,
+            880,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "8164",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Shuaichen Chang and Eric Fosler-Lussier. 2023. How to prompt llms for text-to-sql: A study in zero-shot, single-domain, and cross-domain settings. CoRR, abs/2305.11853.",
+            "Xuemei Dong, Chao Zhang, Yuhang Ge, Yuren Mao, Yunjun Gao, Lu Chen, Jinshu Lin, and Dongfang Lou. 2023. C3: zero-shot text-to-sql with chatgpt. CoRR, abs/2307.07306.",
+            "Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. 2024a. Text-to-sql empowered by large language models: A benchmark evaluation. Proc. VLDB Endow., 17(5):1132-1145.",
+            "Yingqi Gao, Yifu Liu, Xiaoxia Li, Xiaorong Shi, Yin Zhu, Yiming Wang, Shiqi Li, Wei Li, Yuntao Hong, Zhiling Luo, Jinyang Gao, Liyu Mou, and Yu Li. 2024b. Xiyan-sql: A multi-generator ensemble framework for text-to-sql. CoRR, abs/2411.08599.",
+            "Satya Krishna Gorti, Ilan Gofman, Zhaoyan Liu, Jiapeng Wu, Noel Vouitsis, Guangwei Yu, Jesse C. Cresswell, and Rasa Hosseinzadeh. 2024. Msc-sql: Multi-sample critiquing small language models for text-to-sql translation. CoRR, abs/2410.12916.",
+            "Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, JianGuang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-sql in cross-domain database with intermediate representation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28-August 2, 2019, Volume 1: Long Papers, pages 4524-4535. Association for Computational Linguistics.",
+            "Pengcheng He, Yi Mao, Kaushik Chakrabarti, and Weizhu Chen. 2019. X-SQL: reinforce schema representation with context. CoRR, abs/1908.08113.",
+            "Binyuan Hui, Ruiying Geng, Lihan Wang, Bowen Qin, Yanyang Li, Bowen Li, Jian Sun, and Yongbin Li. 2022.  $S^2$ sql: Injecting syntax to question-schema interaction graph encoder for text-to-sql parsers. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 1254-1262. Association for Computational Linguistics.",
+            "Dongjun Lee, Choongwon Park, Jaehyuk Kim, and Heesoo Park. 2024. MCS-SQL: leveraging multiple prompts and multiple-choice selection for text-to-sql generation. CoRR, abs/2405.07467.",
+            "Boyan Li, Yuyu Luo, Chengliang Chai, Guoliang Li, and Nan Tang. 2024a. The dawn of natural language to SQL: are we fully ready? [experiment, analysis & benchmark ]. Proc. VLDB Endow., 17(11):3318-3331.",
+            "Haoyang Li, Jing Zhang, Cuiping Li, and Hong Chen. 2023a. RESDNSL: decoupling schema linking and skeleton parsing for text-to-sql. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023., pages 13067-13075. AAAI Press."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Haoyang Li, Jing Zhang, Hanbing Liu, Ju Fan, Xiaokang Zhang, Jun Zhu, Renjie Wei, Hongyan Pan, Cuiping Li, and Hong Chen. 2024b. Codes: Towards building open-source language models for text-to-sql. Proc. ACM Manag. Data, 2(3):127.",
+            "Jinyang Li, Binyuan Hui, Reynold Cheng, Bowen Qin, Chenhao Ma, Nan Huo, Fei Huang, Wenyu Du, Luo Si, and Yongbin Li. 2023b. Graphix-t5: Mixing pretrained transformers with graph-aware layers for text-to-sql parsing. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023., pages 13076-13084. AAAI Press.",
+            "Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, Xuanhe Zhou, Chenhao Ma, Guoliang Li, Kevin Chen-Chuan Chang, Fei Huang, Reynold Cheng, and Yongbin Li. 2023c. Can LLM already serve as A database interface? A big bench for large-scale database grounded text-to-sqls. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.",
+            "Xiang Li, Yiqun Yao, Xin Jiang, Xuezhi Fang, Chao Wang, Xinzhang Liu, Zihan Wang, Yu Zhao, Xin Wang, Yuyao Huang, Shuangyong Song, Yongxiang Li, Zheng Zhang, Bo Zhao, Aixin Sun, Yequan Wang, Zhongjiang He, Zhongyuan Wang, Xuelong Li, and Tiejun Huang. 2024c. Tele-flm technical report. CoRR, abs/2404.16645.",
+            "Zhongqiu Li, Zhenhe Wu, Mengxiang Li, Zhongjiang He, Ruiyu Fang, Jie Zhang, Yu Zhao, Yongxiang Li, Zhoujun Li, and Shuangyong Song. 2024d. Scalable database-driven kgs can help text-to-sql. In Proceedings of the ISWC 2024 Posters, Demos and Industry Tracks:, volume 3828 of CEUR Workshop Proceedings.",
+            "Shixuan Liu, Chen Peng, Chao Wang, Xiangyan Chen, and Shuangyong Song. 2023. icsberts: Optimizing pre-trained language models in intelligent customer service. In INNS DLIA@IJCNN 2023, Gold Coast, Australia, 23 June 2023, volume 222 of Procedia Computer Science, pages 127-136. Elsevier.",
+            "OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.",
+            "Mohammadreza Pourreza, Hailong Li, Ruoxi Sun, Yeounoh Chung, Shayan Talaei, Gaurav Tarlok Kakkar, Yu Gan, Amin Saberi, Fatma Ozcan, and Sercan Ö. Arik. 2024. CHASE-SQL: multi-path reasoning and preference optimized candidate selection in text-to-sql. CoRR, abs/2410.01943.",
+            "Mohammadreza Pourreza and Davood Rafiei. 2023. DIN-SQL: decomposed in-context learning of text-to-sql with self-correction. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "8165",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mohammadreza Pourreza and Davood Rafiei. 2024. DTS-SQL: decomposed text-to-sql with small large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024, pages 8212-8220. Association for Computational Linguistics.",
+            "Bowen Qin, Binyuan Hui, Lihan Wang, Min Yang, Jinyang Li, Binhua Li, Ruiying Geng, Rongyu Cao, Jian Sun, Luo Si, Fei Huang, and Yongbin Li. 2022. A survey on text-to-sql parsing: Concepts, methods, and future directions. CoRR, abs/2208.13629.",
+            "Ge Qu, Jinyang Li, Bowen Li, Bowen Qin, Nan Huo, Chenhao Ma, and Reynold Cheng. 2024. Before generation, align it! A novel and effective strategy for mitigating hallucinations in text-to-sql generation. In *Findings of the Association for Computational Linguistics*, ACL 2024., pages 5456-5471. Association for Computational Linguistics.",
+            "Jiawei Shao and Xuelong Li. 2024. AI flow at the network edge. CoRR, abs/2411.12469.",
+            "Ruoxi Sun, Sercan Ö. Arik, Hootan Nakhost, Hanjun Dai, Rajarishi Sinha, Pengcheng Yin, and Tomas Pfister. 2023. SQL-palm: Improved large language model adaptation for text-to-sql. CoRR, abs/2306.00739.",
+            "Chang-Yu Tai, Ziru Chen, Tianshu Zhang, Xiang Deng, and Huan Sun. 2023. Exploring chain of thought style prompting for text-to-sql. In EMNLP 2023, Singapore, December 6-10, 2023, pages 5376-5393. Association for Computational Linguistics.",
+            "Shayan Talaei, Mohammadreza Pourreza, Yu-Chen Chang, Azalia Mirhoseini, and Amin Saberi. 2024. CHESS: contextual harnessing for efficient SQL synthesis. CoRR, abs/2405.16755.",
+            "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.",
+            "Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. RAT-SQL: relation-aware schema encoding and linking for text-to-sql parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7567-7578. Association for Computational Linguistics.",
+            "Bing Wang, Changyu Ren, Jian Yang, Xinnian Liang, Jiaqi Bai, Qian-Wen Zhang, Zhao Yan, and Zhoujun Li. 2023. MAC-SQL: A multi-agent collaborative framework for text-to-sql. CoRR, abs/2312.11242.",
+            "Zihan Wang, Xinzhang Liu, Shixuan Liu, Yitong Yao, Yuyao Huang, Zhongjiang He, Xuelong Li, Yongxiang Li, Zhonghao Che, Zhaoxi Zhang, Yan Wang, Xin Wang, Luwen Pu, Huihan Xu, Ruiyu Fang, Yu Zhao, Jie Zhang, Xiaomeng Huang, Zhilong Lu,"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jiaxin Peng, Wenjun Zheng, Shiquan Wang, Bingkai Yang, Xuewei He, Zhuoru Jiang, Qiyi Xie, Yanhan Zhang, Zhongqiu Li, Lingling Shi, Weiwei Fu, Yin Zhang, Zilu Huang, Sishi Xiong, Yuxiang Zhang, Chao Wang, and Shuangyong Song. 2024a. Telechat technical report. CoRR, abs/2401.03804.",
+            "Zihan Wang, Yitong Yao, Li Mengxiang, Zhongjiang He, Chao Wang, Shuangyong Song, et al. 2024b. Telechat: An open-source billngual large language model. In Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10), pages 10-20.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.",
+            "Hefeng Wu, Yandong Chen, Lingbo Liu, Tianshui Chen, Keze Wang, and Liang Lin. 2023. Sqlnet: Scale-modulated query and localization network for few-shot class-agnostic counting. CoRR, abs/2311.10011.",
+            "Zhenhe Wu, Zhongqiu Li, Mengxiang Li, Jie Zhang, Zhongjiang He, Jian Yang, Yu Zhao, Ruiyu Fang, Yongxiang Li, Zhoujun Li, and Shuangyong Song. 2025. MR-SQL: multi-level retrieval enhances inference for llm in text-to-sql. DASFAA.",
+            "Wenxuan Xie, Gaochen Wu, and Bowen Zhou. 2024. MAG-SQL: multi-agent generative approach with soft schema linking and iterative sub-sql refinement for text-to-sql. CoRR, abs/2408.07930.",
+            "An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.",
+            "Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In ICLR 2023, Kigali, Rwanda, May 1-5, 2023.",
+            "Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir R. Radev. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3911-3921. Association for Computational Linguistics.",
+            "Yanzhao Zheng, Haibin Wang, Baohua Dong, Xingjun Wang, and Changshan Li. 2022. HIE-SQL: history information enhanced network for context-dependent text-to-sql semantic parsing. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 2997-3007. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            910
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "8166",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/8fe8f61ed347aeaaba124062d09cee0861145fde3f50669e540c7805c6db42d3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Datasets</td><td>Train</td><td>Dev</td><td>Test</td><td>DB</td><td>Table/DB</td><td>Row/DB</td></tr><tr><td>BIRD</td><td>9,428</td><td>1,534</td><td>1,789</td><td>95</td><td>7.3</td><td>549k</td></tr><tr><td>Spider</td><td>8,659</td><td>1,034</td><td>2,147</td><td>200</td><td>5.1</td><td>2k</td></tr></table>",
+        "bbox": [
+            196,
+            83,
+            803,
+            142
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/207116c0c6d4af54a00ea15ca00833c445185d37b4e6650e71a4b0f59ab78fac.jpg",
+        "table_caption": [
+            "Table 6: The statistics of BIRD and Spider datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>ICL Prompt</td><td>Content Pipe</td><td>Structure Pipe</td></tr><tr><td>Question\nWhat is the gender of the youngest client who\nopened account in the lowest average salary branch?\nEvidence\nLater birthdate refers to younger age;\nA11 refers to average salary\nForeign keys\naccount. &#x27;district_id&#x27; = &#x27;district&#x27;. &#x27;district_id&#x27;\nclient. &#x27;district_id&#x27; = &#x27;district&#x27;. &#x27;district_id&#x27;\nPre-filtered Database schema\nTable: account\n[(account_id, the id of the account),\n(distinct_id, location of branch),\n-(frequency, frequency of the account),\n(date, the creation date of the account)]\nTable: district\n[(district_id, location of branch),\n(A4, number of inhabitants),\n(A11, average salary)]\nBi-filtered Database schema\nTable: account\n[(account_id, the id of the account),\n(distinct_id, location of branch),\n-(frequency, frequency of the account),\n(date, the creation date of the account)]\nTable: district ...</td><td>Content Information\nThe following attribute values may be table or column names:\n[gender], [client], [average salary]\nEvidence Combination\nCombined with [Evidence] :\n[Later birthdate] refers to [younger age]\n[A11] refers to [average salary]\nThe following attributes may be table or column names:\n[gender], [client], [average salary] → [A11]\n+ [Later birthdate]\nForeign keys Combination\nCombined with [Foreign keys] , the final &#x27;content information&#x27;:\n[gender], [client], [Later birthdate], [A11]\n+ [district]\n+ [district_id]\nBi-directional Schema Linking\n(1) Combine the results above with [Database schema] ,\n[Evidence] and [Foreign keys] , deduce the possible final\nidentifiers according to schema.\n(2) Consider [Question] and the [final identifiers] , keep all\nprimary keys, we can filter out the irrelevant columns from schema.\nFinal Identifiers\nWe can known from the schema that [Later birthdate] imply\ncolumn name [birth_date] in table [client].\n- Tables:\n[client], [district]</td><td>Structure Information\nThe following words imply part of SQL structure:\n[the youngest] [the lowest] [average]\nFunction Induction\nAccording to structure information, SQL may involve functions:\n[the youngest] → [ORDER BY], [ASC], [DECS]\n[the lowest] → [ORDER BY],\n[ASC], [DECS]\n[average] → [AVG]\nFinal Functions\n- Functions:\n[ORDER BY], [ASC], [DECS], [AVG]\nSQL Skeleton Generation\nConsidering constraints, generate the skeleton of SQL refer to the\nrecommendation functions in &#x27;Final Function&#x27;:\nSELECT T1.[column_name]\nFROM [table_name] AS T1\nINNER JOIN [table_name] AS T2\nON T1.[column_name] = T2.[column_name]\nORDER BY T2.[column_name] ASC, T1.[column_name] DESC LIMIT 1\nSQL Skeleton Filling\nFill the SQL skeleton, replace [table_name] and [column_name]\nrefer to the recommendation identifiers in &#x27;Final identifiers&#x27;:\nSELECT T1. gender&#x27;\nFROM client AS T1\nINNER JOIN district AS T2\nON T1.[district_id] = T2.[district_id]\nORDER BY T2.&#x27;A11&#x27; ASC, T1&#x27;[birth_date&#x27; DESC LIMIT 1</td></tr></table>",
+        "bbox": [
+            115,
+            193,
+            882,
+            524
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Figure 9: Detailed prompt design for UCL-SQL. The left column displays the original question and database information. Content and structure information are extracted from the question to activate the content and structure pipes. Identifiers and SQL skeletons are generated at the end of each pipe and combined to form the final SQL query.",
+        "bbox": [
+            112,
+            538,
+            882,
+            595
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Statistics of datasets.",
+        "text_level": 1,
+        "bbox": [
+            112,
+            620,
+            331,
+            634
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6 shows the statistics of two datasets.",
+        "bbox": [
+            112,
+            645,
+            435,
+            659
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            671,
+            248,
+            688
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In this section, we elucidate the UCS-SQL framework through a detailed examination of an incontext learning prompt, as depicted in Figure 9. The process commences with the extraction of both content and structural information from the original question.",
+        "bbox": [
+            112,
+            696,
+            487,
+            791
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Within the content pipe, the LLM discerns that the natural language phrase \"average salary\" corresponds to the schema identifier \"A11\". Consequently, \"average salary\" is mapped to \"A11\". Additionally, the mention of \"Later birthdate\" prompts its provisional inclusion in the candidate set. Subsequently, we engage in the Foreign Keys Combination phase. Since \"A11\" is associated with the \"dis",
+        "bbox": [
+            112,
+            793,
+            489,
+            921
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "strict\" table and \"gender\" with the \"client\" table, a foreign key linking these two tables is essential for SQL query generation. This necessitates the inclusion of \"district\" and \"district_id\" from the equivalence \"client.'district_id' = district.'district_id'\" into our candidate set. Advancing to the Bidirectional schema linking phase, we identify \"gender\", \"A11\", and \"client\" within the schema. The schema's description of \"birth_data\" (birth date) is recognized as the closest match to \"Later birthdate\", leading the LLM to select \"birth_data\" as a substitute. Ultimately, the final identifiers consist of the tables \"client\" and \"district\", and the columns \"gender\", \"birth_date\", \"A11\", and \"district_id\". Thereafter, the LLM integrates these final identifiers, refining the pre-filtered schema by eliminating the redundant column \"A4\", obtaining the bi-filtered schema.",
+        "bbox": [
+            507,
+            620,
+            884,
+            909
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "8167",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In the structure pipe, the LLM initially infers candidate functions \"ORDER BY\", \"ASC\", \"DESC\", and \"AVG\" from the structure information \"the youngest\", \"the lowest\" and \"average\". Integrating these with the query and constraints, the LLM constructs the SQL skeleton. Ultimately, with reference to the bi-filtered schema, the LLM populates the SQL skeleton with the candidate table names and column names from the content pipe, thereby completing the final SQL query.",
+        "bbox": [
+            115,
+            84,
+            489,
+            244
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "8168",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2025/UCS-SQL_ Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL/5e130ac7-c9a3-4216-9a37-f02974ca9352_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16c992acfbf4d710787c4bf67483fd8cdadfbe73195f98c43fc58b82c5f77bfe
+size 818774

2025/UCS-SQL_ Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL/full.md

@@ -0,0 +1,394 @@
+# UCS-SQL: Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL
+
+Zhenhe Wu $^{1,2\dagger}$ , Zhongqiu Li $^{2\dagger}$ , Jie Zhang $^{2}$ , Zhongjiang He $^{2*}$ , Jian Yang $^{1*}$ , Yu Zhao $^{2}$ , Ruiyu Fang $^{2}$ , Bing Wang $^{1}$ , Hongyan Xie $^{1}$ , Shuangyong Song $^{2}$ , Zhoujun Li $^{1*}$ ,
+
+<sup>1</sup>Beihang University, <sup>2</sup>TeleAI, China Telecom Corp Ltd,
+
+{wuzhenhe,jiaya,bingwang,by2406233,lizj}@buaa.edu.cn
+
+{lizq48,zhangj157,hezj,zhaoy11,fangry,songshy}@chinatelecom.cn
+
+# Abstract
+
+With the rapid advancement of large language models (LLMs), recent researchers have increasingly focused on the superior capabilities of LLMs in text/code understanding and generation to tackle text-to-SQL tasks. Traditional approaches adopt schema linking to first eliminate redundant tables and columns and prompt LLMs for SQL generation. However, they often struggle with accurately identifying corresponding tables and columns, due to discrepancies in naming conventions between natural language questions (NL) and database schemas. Besides, existing methods overlook the challenge of effectively transforming structure information from NL into SQL. To address these limitations, we introduce UCS-SQL, a novel text-to-SQL framework, uniting both content and structure pipes to bridge the gap between NL and SQL. Specifically, the content pipe focuses on identifying key content within the original content, while the structure pipe is dedicated to transforming the linguistic structure from NL to SQL. Additionally, we strategically selects few-shot examples by considering both the SQL Skeleton and Question Expression (SS-QE selection method), thus providing targeted examples for SQL generation. Experimental results on BIRD and Spider demonstrate the effectiveness of our UCS-SQL framework.
+
+# 1 Introduction
+
+Text-to-SQL is the task of translating natural language queries into SQL statements, which has garnered extensive research attention and practical application in database querying (Qin et al., 2022; Sun et al., 2023). Based on the representations encoded by BERT-style pre-trained models (Liu et al., 2023), abstract syntax trees (Wu et al., 2023; Guo et al., 2019; Wang et al., 2020) and predefined query sketches (He et al., 2019) are implemented
+
+![](images/ae7941b8897c17276b9c737262d260a18ad5d6904546fcd575e1dfca99e5ebbf.jpg)  
+Figure 1: Traditional pipelines directly generate SQL query from NL question. We decompose text-to-SQL task into content and structure dimensions, enabling key content extraction and linguistic structure transformation to collaboratively enhance SQL generation.
+
+for query decoding. Further, some works extract generalized question-to-SQL patterns by training encoder-decoder models on text-to-SQL corpora (Hui et al., 2022; Li et al., 2023a,b; Zheng et al., 2022; Gao et al., 2024a). Inspired by the powerful capabilities of large language model (LLM) in handling complex reasoning (Wei et al., 2022; Yao et al., 2023), the recent studies have achieved promising results by developing prompt engineering for LLMs (Dong et al., 2023; Pourreza and Rafiei, 2023; Gao et al., 2024a). For example, implementing schema linking to eliminate redundant information, followed by guiding LLMs to generate SQL queries through in-context learning processes.
+
+The forefront researches adhere to the retrieve-then-generate paradigm. MCS-SQL (Lee et al., 2024) exploits LLMs' extensive sample libraries for schema linking and employs multiple prompts to elicit diverse LLM responses. RSL-SQL (Cao et al., 2024) integrates bidirectional schema linking, contextual information augmentation, binary selection strategy, and multi-turn self-correction to achieve state-of-the-art performance. However, the traditional retrieve-then-generate pipeline in txt-to-SQL tasks still confronts two challenges. (1) Ex
+
+![](images/2012ab466d72e1a8cd18bc8bc545d4dcf1ce3c07f826ad0e78da538e38962078.jpg)  
+Figure 2: The framework of UCS-SQL, which consists of three stages uniting content and structure pipe. In question stage, LLM extracts content and structure information from the original question. Transitional stage facilitates the deduction from information to derive the candidate identifiers and functions. In SQL stage, they are combined to form the final query.
+
+isting methods struggle with precise schema linking and identifier selection (e.g., column and table names) due to terminological mismatches between colloquial questions and schema identifiers. (2) Moreover, they seldom address the extraction and transformation of structural information from the natural language (NL) question to the SQL query, while bridging the gap of different linguistic structures is crucial for effective query generation. To address the aforementioned challenges, we plan to propose a more comprehensive deductive approach for in-context learning, emphasizing both content and structural dimensions. Specifically, in the content dimension, we deduce SQL identifiers step by step, thereby transforming the original content into key content. In the structure dimension, we derive the SQL skeleton from the question, facilitating the linguistic structure from NL to SQL. Figure 1 demonstrates the integration of question comprehension and SQL query generation from content and structure pipes.
+
+In this paper, we introduce UCS-SQL, a three-stage text-to-SQL framework that integrates content and structure pipes to enhance semantic bridging. The content pipe extracts key information from the raw question, while the structure pipe converts NL into SQL syntax. Figure 2 illustrates the UCS-SQL framework. In the question stage, a LLM extracts content and structure information from the raw question. Then in the transitional stage, we prompt the LLM to transform them into SQL identifiers and functions. The content pipe integrates evidence and foreign keys to refine content information and performs bi-directional schema link-
+
+ing to identify relevant schema identifiers, while the structure pipe infers SQL functions based on the extracted structure information. Finally, in the SQL stage, we construct a SQL skeleton using the inferred functions and fill it with the identified identifiers to generate the final query. Additionally, we develop an SS-QE algorithm to select few-shot examples for the LLM, considering similarities in both SQL skeletons and question expressions. Comprehensive evaluations on the BIRD and Spider datasets demonstrate that UCS-SQL outperforms several baselines in terms of Valid Efficiency Score (VES) and Execution Accuracy (EX) across different difficulty levels. It achieves top-2 performance on all evaluation metrics on BIRD and top-3 performance on Spider.
+
+To summarize, our contributions are as follows:
+
+- We propose UCS-SQL, a three-stage text-to-SQL framework that unites content and structure pipes to respectively extract key content and transform linguistic structure, thereby collaboratively enhancing SQL query generation.  
+- We introduce SS-QE algorithm to select few-shot examples, which utilizes the structure information from UCS-SQL and takes into account the similarities in both SQL structure and question expression.  
+- Empirical evaluations indicate that UCS-SQL framework attains  $3^{nd}$  place in all evaluation metrics on BIRD and Spider. Ablation experiments demonstrate that all three stages and both pipes in the UCS-SQL are crucial for performance enhancement.
+
+# 2 Problem Definition
+
+Text-to-SQL is the task of converting a natural language question  $Q$  into a correct SQL query  $Y$ . The database can be represented as  $D = \{T_{1}, T_{2} \dots T_{m}\}$ ,  $m$  is the number of tables in the database. For  $T = \{C_{1}, C_{2} \dots C_{n}\}$ ,  $C_{i}$  refers to columns in table  $T$ ,  $n$  is the number of columns in the table. When dealing with complex database values, we can use external knowledge evidence  $K$  to support our model understand the inner relationship between question and database. The process could be formulated as follows:
+
+$$
+Y = f (Q, D, K \mid \theta) \tag {1}
+$$
+
+where  $f(\cdot |\theta)$  represent a model with parameters  $\theta$ .
+
+# 3 UCS-SQL
+
+# 3.1 Overview
+
+We introduce UCS-SQL, a parallel text-to-SQL framework that integrates content and structure pipes to enhance semantic bridging. As illustrated in Figure 2, the framework comprises three stages: question, transitional, and SQL stages. (1) In the question stage, we prompt an LLM to extract content information and structure information from the original question. (2) The transitional stage acts as a bridge between the question and SQL stages. We combine the content information with the database schema to obtain candidate identifiers in the content pipeline and deduce SQL functions from the structure information in the structure pipeline. (3) In the SQL stage, we first generate an SQL skeleton based on the inferred SQL functions and the original question. Then, we fill the skeleton with the candidate identifiers to produce the final SQL query. Besides, a novel approach is proposed for selecting few-shot examples, considering similarities in both SQL structure and question expression. The following section provides a detailed introduction to UCS-SQL and the example selection method.
+
+# 3.2 Stage 1: Question Stage
+
+First, we perform pre-processing on the database schema prior by decomposing the original database into smaller sub-databases using the selector (Wang et al., 2023) to minimize interference from irrelevant tables and columns. Additionally, we maintain column descriptions for each column in the pre-filtered schema.
+
+The question stage is designed to extract key content and structure from the original question.
+
+![](images/e4e644cea24497478fad34678b5521355ddb18b7433a050c963bce3d2e1954d3.jpg)  
+Figure 3: The question stage of UCS-SQL. The original database is pre-processed to filter out irrelevant tables and columns. Terms representing content and structure information are extracted from the raw question.
+
+Specifically, we employ LLMs to identify significant content information (e.g., table names and column names) and structure information from the target question to facilitate SQL generation. In the subsequent operations, UCS-SQL proceeds via two parallel pipes, respectively handling SQL generation at the content and structure levels. Figure 3 illustrates the process of the question stage. We prompt the LLM to identify content information and structure information based on the original phrase of the question, which we highlight in red and blue colors. These identified elements are extracted and integrated to advance both parallel pipelines to the next stage. It is important to note that the LLM may identify the same word as both content and structure information, such as the word "average" shown in Figure 3. This dual identification aligns with our objective at the question stage, retaining sufficient relevant information. Further reasoning and selection will be conducted in subsequent stages.
+
+# 3.3 Stage 2: Transitional Stage
+
+The transitional stage is designed to transform the original information extracted from the question into SQL language elements. Figure 4 illustrates the process of this stage.
+
+In the content pipe, we first integrate the content information with evidence to convert colloquial phrases into schema-compatible identifiers. Subsequently, we perform foreign key expansion to augment the candidate set of potential identifiers for SQL generation. Specifically, we prompt the LLM to identify relevant foreign keys and incorporate the corresponding identifiers into the candidate set. Next, we conduct bi-directional schema linking by combining the schema with the candidate set,
+
+![](images/f46b127fa11ddccff56f277b534c514936e39be0d2f581d8ceb011a9eaa87167.jpg)  
+Figure 4: The transitional stage of UCS-SQL. Candidate identifiers and functions are obtained in the content and structure pipes. Bi-directional schema linking is performed with reference to the identifiers, yielding the bi-filtered schema.
+
+which both deduces potential identifiers and filters out irrelevant columns from the schema. We match the content information with schema identifiers: if a close match with a table or column name is found, the corresponding identifier is retained. In cases where no direct match exists, the LLM searches for the most similar expression among column descriptions and retains the associated identifiers. Through careful analysis and deduction, we obtain a refined set of identifiers. For instance, the final identifiers reveal that column "A4" is irrelevant to the question, so we remove "A4" from the schema to arrive at the bi-filtered database schema.
+
+In the structure pipe, we use the LLM to induce SQL functions based on the structure information extracted during the question stage. As shown in Figure 4, phrases such as "the youngest" and "the lowest" suggest the potential use of SQL functions like "ORDER BY", "ASC", and "DESC" while "average" implies the "AVG" function. In summary, we distill these insights into a final set of functions.
+
+# 3.4 Stage 3: SQL Stage
+
+In the SQL stage, we first create the SQL skeleton by referring to the candidate functions derived from the structure pipe. Subsequently, we integrate the identifiers from the content pipe with the bi-filtered schema to fill the generated skeleton. As a result, by unifying the structure and content pipes, we ultimately produce the final SQL query.
+
+In Figure 5, we prompt the LLM to generate an SQL skeleton based on the constraints, question context, and all available structure information, with particular emphasis on the candidate functions from the structure pipeline. The LLM selects functions such as "ORDER BY", "ASC", and "DESC"
+
+![](images/71f6a5acf007c3d226e5dd5a087b5401d394fb2d6ab5bf4720b4bac055512d5d.jpg)  
+Figure 5: The SQL stage of UCS-SQL. The SQL skeleton is inferred from functions. The final SQL query is constructed by incorporating the identifiers into the SQL skeleton. The bi-filtered schema serves as a secondary reference to address cases where the identifiers are not correctly extracted.
+
+to form the SQL skeleton. Next, we fill the skeleton using the identifiers from the content pipeline, focusing on the recommended identifiers. The LLM identifies "client" and "district" as table names and "gender", "birth_date", "A11", and "district_id" as column names to complete the final SQL query.
+
+# 3.5 SS-QE Example Selection
+
+We propose a novel SS-QE few-shot example selection approach, which considers the similarities in both SQL structure and question expression. We present the detailed algorithm of our selection method in Algorithm 1. We first extract final identifiers and final functions from all the question  $q_{i}$  in our training set  $Q$  following the process of question stage and transitional stage.  $F_{i}$  is the set of final functions,  $I_{i}$  is the set of final identifiers, while  $c_{i}$  is the complexity of  $q_{i}$ , which we define as the sum of final identifiers and final functions:
+
+$$
+c _ {i} = \left| I _ {i} \right| + \left| F _ {i} \right| \tag {2}
+$$
+
+First, we evaluate the similarity between  $q_{i}$  and  $q$  in terms of SQL structure, using  $F_{i}$  and  $c_{i}$  as our basis. Specifically, samples with a larger intersection in their final functions and closer complexity values are considered to have a higher degree of SQL structure similarity. To filter these samples, we establish thresholds  $\theta_{1}$  and  $\theta_{2}$ . Subsequently, we add each sample that satisfies these criteria to the set  $S_{1}$ , continuing this process until  $S_{1}$  contains  $2k$  samples. Next, we further utilize LLM to pick top-  $k$  examples from  $S_{1}$  based on their similarity with  $q$  in question expression, Figure 6 shows the selection prompt in detail. As the result, we obtain
+
+Algorithm 1: SS-QE Example selection  
+Input: All the questions  $Q$  in training set, target question  $q$ , the number of few-shot  $k$ , threshold  $\theta_{1}$  &  $\theta_{2}$   
+Output: Top- $k$  questions  $S$   
+1  $S_{1} \gets \emptyset$ ;  
+2  $I, F \gets LLM_{InformationExtraction}(q)$ ;  
+3  $c \gets |I| + |F|$ ;  
+4 foreach  $q_{i}$  in  $Q$  do  
+5  $I_{i}, F_{i} \gets LLM_{InformationExtraction}(q_{i})$ ;  
+6  $c_{i} \gets |I_{i}| + |F_{i}|$ ;  
+7 if  $(F \cap F_{i} \geq \theta_{1}) \land (|c - c_{i}| < \theta_{2})$  then  
+8 Add  $q_{i}$  in  $S_{1}$ ;  
+9 end  
+10 if  $|S_{1}| = 2k$  then  
+11 break;  
+12 end  
+13 end  
+14 S ← LLMSimilarQuestionExpression(q, k, S1);  
+15 return S;
+
+$k$ -shot examples that have similar SQL structure and question expression.
+
+# 4 Experimental Settings
+
+# 4.1 Datasets
+
+BIRD (Li et al., 2023c) represents a cross-domain dataset that examines the impact of extensive database contents on text-to-SQL parsing. Spider (Yu et al., 2018) is a large-scale complex and cross-domain semantic parsing and text-to-SQL dataset. We report the statistics of datasets in appendix A.
+
+# 4.2 Evaluation Metrics
+
+Following BIRD (Li et al., 2023c), we utilize execution accuracy (EX) and valid efficiency score (VES) to evaluate text-to-SQL models. EX (Li et al., 2023c) is defined as the proportion of questions in the evaluation set for which the execution results of both the predicted and ground-truth inquiries are identical. VES (Li et al., 2023c) is designed to measure the efficiency of valid SQLs generated by models.
+
+# 4.3 Baselines
+
+DIN-SQL(Pourreza and Rafiei, 2023) and MAC-SQL(Wang et al., 2023) breaks down intricate queries into manageable sub-tasks.
+
+Similar Question Expression Selection Prompt  
+Select three questions that are closest in question phrasing to the target question from the source questions and return the list which is contain the three questions ids. Format of the returned list is as follows: [3, 672, 19]  
+target question: {question}  
+source questions: {{source_questions}}  
+returned list: {result_ids}
+
+Figure 6: The prompt for selecting examples having similar question expression with target questions.
+
+E-SQL(Caferoglu and Ulusoy, 2024) and CHESS(Talaei et al., 2024) seek to bridge the divide between natural language queries and database architectures. SQL-PaLM(Sun et al., 2023) and SuperSQL(Li et al., 2024a) utilize distinct prompting and fine-tuning methods for adapting large language models (LLMs) in SQL generation. TA-SQL(Qu et al., 2024) and CodeS(Li et al., 2024b) introduce strategies to reduce hallucinations in LLM-based SQL generation. DAIL-SQL(Gao et al., 2024a) is designed to tackle complex database environments. Additionally, methods based on multi-stage strategies are included, such as DTS-SQL(Pourreza and Rafiei, 2024) that employs two-stage fine-tuning, MAG-SQL(Xie et al., 2024) that adopts a multi-agent generative approach, and MCS-SQL(Lee et al., 2024) that utilizes multiple prompts and multiple-choice selection. MSc-SQL(Gorti et al., 2024) narrows the performance gap of smaller open-source models by sampling and comparing multiple SQL query results. RSL-SQL(Cao et al., 2024) achieve robust schema linking that maximize the benefits. CHASE-SQL (Pourreza et al., 2024) propose multiple chain-of-thought prompting methods and an online synthetic example generation technique. XiYan-SQL (Gao et al., 2024b) integrates the ICL approach to maximize the generation of high-quality and diverse SQL candidates.
+
+# 5 Implementation Details
+
+We use a single Tesla V100 GPU with 32 GiBs of memory on a server to restore intermediate data and execute SS-QE example selection algorithm, the retrieving time is about  $3\sim 6$  seconds for each sample. All the experiments utilize GPT-4-Turbo, the context window is 128000, the temperature is set to 0.1. We enable five threads to run UCS-SQL
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Model</td><td colspan="4">EX</td><td rowspan="2">VES</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td>DIN-SQL</td><td>GPT-4</td><td>-</td><td>-</td><td>-</td><td>50.72</td><td>58.79</td></tr><tr><td>DAIL-SQL</td><td>GPT-4</td><td>-</td><td>-</td><td>-</td><td>54.76</td><td>56.08</td></tr><tr><td>DTS-SQL</td><td>DeepSeek-7B</td><td>-</td><td>-</td><td>-</td><td>55.80</td><td>-</td></tr><tr><td>TA-SQL</td><td>GPT-4</td><td>63.14</td><td>48.60</td><td>36.11</td><td>56.19</td><td>-</td></tr><tr><td>Codes</td><td>Codes-15B</td><td>-</td><td>-</td><td>-</td><td>58.47</td><td>59.87</td></tr><tr><td>SuperSQL</td><td>GPT-4</td><td>66.90</td><td>46.50</td><td>43.80</td><td>58.50</td><td>61.99</td></tr><tr><td>MAC-SQL</td><td>GPT-4</td><td>65.73</td><td>52.69</td><td>40.28</td><td>59.39</td><td>66.39</td></tr><tr><td>MAG-SQL</td><td>GPT-4</td><td>-</td><td>-</td><td>-</td><td>61.08</td><td>-</td></tr><tr><td>SQL-Palm</td><td>PaLM2</td><td>68.92</td><td>52.07</td><td>47.89</td><td>61.93</td><td>-</td></tr><tr><td>MCS-SQL</td><td>GPT-4</td><td>70.40</td><td>53.10</td><td>51.40</td><td>63.36</td><td>64.80</td></tr><tr><td>CHESS</td><td>Proprietary</td><td>-</td><td>-</td><td>-</td><td>65.00</td><td>-</td></tr><tr><td>E-SQL</td><td>GPT-4o</td><td>-</td><td>-</td><td>-</td><td>65.58</td><td>-</td></tr><tr><td>MSc-SQL</td><td>GPT-4o</td><td>72.00</td><td>58.00</td><td>49.00</td><td>65.60</td><td>-</td></tr><tr><td>RSL-SQL</td><td>GPT-4o</td><td>74.38</td><td>57.11</td><td>53.79</td><td>67.21</td><td>70.32</td></tr><tr><td>CHASE-SQL</td><td>Gemini 1.5</td><td>-</td><td>-</td><td>-</td><td>73.01</td><td>-</td></tr><tr><td>XiYan-SQL</td><td>-</td><td>-</td><td>-</td><td>-</td><td>73.34</td><td>-</td></tr><tr><td>UCS-SQL</td><td>DeepSeek</td><td>71.67</td><td>55.59</td><td>54.11</td><td>65.23</td><td>68.49</td></tr><tr><td>UCS-SQL</td><td>GPT-4o</td><td>75.42</td><td>59.52</td><td>56.93</td><td>68.95</td><td>71.10</td></tr></table>
+
+Table 1: EX and VES on dev set of BIRD. UCS-SQL achieves top two performances in both VES and EX across datasets of varying difficulty levels.  
+
+<table><tr><td>Method</td><td>Model</td><td>EX (dev)</td><td>EX (test)</td></tr><tr><td>DIN-SQL</td><td>GPT-4</td><td>82.8</td><td>85.3</td></tr><tr><td>DAIL-SQL</td><td>GPT-4</td><td>84.4</td><td>86.6</td></tr><tr><td>DTS-SQL</td><td>DeepSeek-7B</td><td>85.5</td><td>84.4</td></tr><tr><td>TA-SQL</td><td>GPT-4</td><td>85.0</td><td>-</td></tr><tr><td>MAC-SQL</td><td>GPT-4</td><td>86.8</td><td>82.8</td></tr><tr><td>MAG-SQL</td><td>GPT-4</td><td>85.3</td><td>85.6</td></tr><tr><td>MCS-SQL</td><td>GPT-4</td><td>89.5</td><td>89.6</td></tr><tr><td>CHESS</td><td>Openllms</td><td>-</td><td>87.2</td></tr><tr><td>MSc-SQL</td><td>GPT-4o</td><td>-</td><td>84.7</td></tr><tr><td>RLS-SQL</td><td>GPT-4o</td><td>-</td><td>87.9</td></tr><tr><td>CHASE-SQL</td><td>Gemini 1.5</td><td>-</td><td>87.6</td></tr><tr><td>XiYan-SQL</td><td>-</td><td>-</td><td>89.6</td></tr><tr><td>UCS-SQL</td><td>DeepSeek</td><td>86.5</td><td>87.7</td></tr><tr><td>UCS-SQL</td><td>GPT-4o</td><td>87.3</td><td>88.0</td></tr></table>
+
+Table 2: UCS-SQL achieves top three performances in EX on dev and test set of Spider.
+
+(approximately  $200\sim 500$  samples for each according to the size of dataset), it costs about  $4\sim 6$  hours to generate all results. The token consumption of 2-shot UCS-SQL on dev set for BIRD and Spider are approximately 5.29 million and 3.47 million in input, 0.9 and 0.5 million in output. We use all MiniLM-L6-v2 model to get sentence embedding and calculate cosine similarity in the discussion. In SS-QE selection method, we set  $\{\theta_1,\theta_2\}$  to  $\{2,3\}$ .
+
+# 6 Results and Analysis
+
+# 6.1 Overall Results
+
+The overall results of all the baselines and our proposed UCS-SQL on BIRD and Spider are shown
+
+in Table 1 and Table 2. The experiment results indicate that our proposed UCS-SQL achieves better performance than several competitive baselines on the two datasets.
+
+In Table 1, we report the performance of UCS-SQL and other competitive baselines on development set of BIRD. These recent methods have their own distinct characteristics and utilize different LLM models, such as GPT-4, GPT-4o, and DeepSeek. Experiments show that our proposed UCS-SQL method almost outperforms the first thirteen baselines by at least  $3.35\%$  on EX and by at least  $4.71\%$  on VES, only slightly trailing the concurrent work RSL-SQL. In addition, UCS-SQL performs better than RSL-SQL on samples of moderate and challenging difficulty, which indicates that UCS-SQL can address problems more specifically when dealing with samples of higher complexity. On the other hand, Table 2 shows the execution accuracy of UCS-SQL and other baselines on Spider. Our proposed UCS-SQL achieves excellent results of  $87.3\%$  and  $88.0\%$  on dev and test sets, surpassing the vast majority of baselines. In conclusion, UCS-SQL respectively achieves top-3 results in VES and EX across different difficulty on BIRD and Spider datasets, and ranked first on high-difficulty data, demonstrating the method's high efficiency in handling complex examples and its generalizability across different scenarios.
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">EX</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td colspan="5">Content &amp; Structure Pipes Dimension (fixed 2-shot)</td></tr><tr><td>w/o content pipe</td><td>69.81</td><td>54.98</td><td>50.48</td><td>63.58(↓1.24)</td></tr><tr><td>w/o structure pipe</td><td>70.45</td><td>54.54</td><td>49.77</td><td>63.78(↓1.04)</td></tr><tr><td>w/ both pipes</td><td>70.24</td><td>58.25</td><td>50.48</td><td>64.82</td></tr><tr><td colspan="5">Three Stages Dimension (fixed 2-shot)</td></tr><tr><td>w/o question &amp; transitional stage</td><td>70.13</td><td>53.24</td><td>47.66</td><td>63.00(↓1.82)</td></tr><tr><td>w/o transitional stage</td><td>69.59</td><td>56.07</td><td>49.07</td><td>63.65(↓1.17)</td></tr><tr><td>w/o SQL stage</td><td>69.70</td><td>55.42</td><td>48.66</td><td>63.39(↓1.43)</td></tr><tr><td>w/ three stages</td><td>70.24</td><td>58.25</td><td>50.48</td><td>64.82</td></tr><tr><td colspan="5">Few-shot Example Selection</td></tr><tr><td>0-shot</td><td>67.45</td><td>53.89</td><td>44.14</td><td>61.24(↓7.71)</td></tr><tr><td>2-shot (fixed)</td><td>70.24</td><td>58.25</td><td>50.48</td><td>64.82(↓4.13)</td></tr><tr><td>2-shot (SS-QE)</td><td>71.99</td><td>58.81</td><td>52.89</td><td>66.28(↓2.67)</td></tr><tr><td>5-shot (SS-QE)</td><td>75.42</td><td>59.52</td><td>56.93</td><td>68.95</td></tr></table>
+
+Table 3: Ablation study on dev set of BIRD. The three groups of experiments respectively compare the impact of two pipes, three stages, and different few-shot example selections on UCS-SQL.
+
+# 6.2 Ablation Study
+
+To evaluate the contributions of different components in UCS-SQL, we conducted a series of ablation studies using the BIRD dataset, which is a large-scale dataset with complex database structures. The experiments were divided into three groups: (1) assessing the roles of the content pipe and structure pipe, (2) analyzing the contributions of the three stages, and (3) comparing the effectiveness of the SS-QE example selection method. The results are summarized in Table 3.
+
+Group 1: Content Pipe vs. Structure Pipe Removing either the content pipe or the structure pipe from the framework led to a decrease in execution accuracy across datasets of varying difficulty, with respective drops of  $1.24\%$  and  $1.04\%$ . The results indicate that integrating both content and structure information is crucial for guiding LLMs to generate SQL queries effectively.
+
+Group 2: Contributions of the Three Stages The ablation studies also examined the necessity of the three stages in UCS-SQL. Removing the transitional stage or the SQL stage resulted in a decrease in execution accuracy by  $1.17\%$  and  $1.43\%$ , respectively. Additionally, removing the question stage further reduced performance by  $0.65\%$ . These findings demonstrate that each stage plays a vital role in the overall performance of the framework.
+
+Group 3: SS-QE Example Selection Method The effectiveness of the SS-QE example selection method was evaluated by comparing the execution accuracy of 0-shot, fixed 2-shot, SS-QE-selected 2-shot, and SS-QE-selected 5-shot scenarios. The
+
+results showed a progressive improvement in performance, with a  $1.46\%$  increase from fixed 2-shot to SS-QE-selected 2-shot and an additional  $2.67\%$  improvement when increasing the number of shots from 2 to 5. This indicates that the SS-QE method effectively selects guiding examples that enhance the overall performance of the framework.
+
+The ablation studies demonstrate that both pipes and all three stages in the UCS-SQL framework are essential for performance enhancement. This validates the rationality of integrating content and structure dimensions in a multi-stage framework for SQL generation. Additionally, the SS-QE example selection method further leverages the framework's capabilities by selecting few-shot examples that match the target question's function and complexity, thereby improving the overall performance.
+
+# 7 Discussion
+
+# 7.1 Effect of bi-filtered schema
+
+This section examines the impact of bi-filtered schemas generated through the bi-directional schema linking process, using the BIRD dataset due to its large scale and complexity. We define the schema column count as the total number of columns across all tables in a schema. Figure 7 illustrates the schema column counts for each sample in the BIRD dev set. The "Overall column counts" represent raw database schemas, which consist of 11 distinct databases. The "Gold column counts" reflect columns used as identifiers in gold SQL queries, setting the performance limits for schema
+
+![](images/bc6ae47972a05c16db6da3523694458c11bf6d3add1cc07fbff31ab5c8b9c7c3.jpg)  
+Figure 7: The schema column counts on dev set of BIRD. The bi-filtered column counts have significantly decreased compared to the pre-filtered column counts, and are slightly higher than the gold column counts.
+
+<table><tr><td rowspan="2">Metric (%)</td><td colspan="4">UCS-SQL</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td colspan="5">Pre-filtered Schema</td></tr><tr><td>Precision</td><td>11.23</td><td>15.78</td><td>20.56</td><td>13.26</td></tr><tr><td>Recall</td><td>97.33</td><td>95.11</td><td>96.19</td><td>96.42</td></tr><tr><td>F1 score</td><td>20.14</td><td>27.07</td><td>33.87</td><td>23.31</td></tr><tr><td colspan="5">Bi-filtered Schema</td></tr><tr><td>Precision</td><td>59.71</td><td>73.81</td><td>78.90</td><td>65.83</td></tr><tr><td>Recall</td><td>89.14</td><td>84.09</td><td>85.44</td><td>86.93</td></tr><tr><td>F1 score</td><td>71.51</td><td>78.62</td><td>82.04</td><td>74.92</td></tr></table>
+
+linking. The "Pre-filtered column counts" denote schemas filtered using the method described earlier, while the "Bi-filtered column counts" represent our bi-filtered schemas. The pre-filtered schemas show inconsistent filtering effectiveness, with a significant gap compared to gold schemas. In contrast, bi-filtered schemas achieve substantial filtering, closely approximating gold schemas in terms of column counts.
+
+To further evaluate the effectiveness of bi-filtered schemas, we calculated precision, recall, and F1 score using gold schemas as the standard. Table 4 presents the evaluation results. Pre-filtered schemas have high recall across datasets but poor precision and F1 score (13.26% and 23.31%, respectively). In contrast, bi-filtered schemas achieve a recall of 86.93%, slightly lower than pre-filtered schemas, but with significantly higher precision (65.83%) and F1 score (74.92%).
+
+Bi-filtered schemas effectively eliminate irrelevant columns, closely matching gold schemas in column counts. The high precision, recall, and F1 score demonstrate the superior performance of bi-directional schema linking.
+
+![](images/12a4fb3ebf4f31cad3e3077674616acb2f415ae90ffe67993d570500f3f18bf5.jpg)  
+Figure 8: The evaluation statistics of UCS-SQL using DIAL or SS-QE selection method on dev of BIRD.
+
+Table 4: Evaluation of pre-filtered schema and bifiltered schema on dev of BIRD.  
+
+<table><tr><td rowspan="2">UCS-SQL (5-shot)</td><td colspan="4">EX</td></tr><tr><td>simple</td><td>moderate</td><td>challenging</td><td>total</td></tr><tr><td>w/DAILs</td><td>73.74</td><td>58.94</td><td>51.07</td><td>67.21</td></tr><tr><td>w/SS - QE</td><td>75.42</td><td>59.52</td><td>56.93</td><td>68.95</td></tr></table>
+
+Table 5: EX results of UCS-SQL with two example selection method on dev of BIRD.
+
+# 7.2 Effect of SS-QE method
+
+This study further examines the impact of the SS-QE selection method using the BIRD dataset. The DAIL Selection method, as reported in (Gao et al., 2024a), is a leading selection technique that pre-generates SQL queries and selects examples based on high similarity between masked questions and masked predicted SQL. We conducted experiments on UCS-SQL using both DAIL and SS-QE methods in a 5-shot setting. Table 5 shows that the SS-QE method outperforms DAIL.
+
+To explore the underlying reasons, we analyzed the selected examples. As shown in Figure 8, the DAIL method slightly outperforms SS-QE in average masked question cosine similarity and masked query cosine similarity. However, the SS-QE method exhibits significantly higher function precision and recall. These results indicate that SS-QE more accurately identifies examples with analogous functions to the target sample, making it more suitable for guiding SQL query generation. Moreover, examples with similar complexity and consistent functions provide a broader selection pool compared to those based solely on high SQL skeleton similarity. Even when the sequence or placement of functions varies, they still offer guiding significance, which would likely be overlooked by similarity-based selection alone.
+
+# 8 Related Work
+
+Large language models (LLMs) have demonstrated significant advancements in various natural language processing (NLP) tasks (Wang et al., 2024a;
+
+Li et al., 2024c; Touvron et al., 2023; OpenAI, 2023; Yang et al., 2024; Shao and Li, 2024; Wang et al., 2024b). Some researchers have leveraged LLMs in text-to-SQL tasks to further enhance performance. A critical aspect of this approach is the design and utilization of prompts, which directly influence the accuracy of SQL generation by guiding LLMs effectively. For example, some methods (Tai et al., 2023) try to improve the inference capabilities of LLMs using chain-of-thought prompting, including both the original chain-of-thought prompt and the least-to-most prompt. Further, a comprehensive analysis (Chang and Fosler-Lussier, 2023) of the impact of prompt construction across various settings in text-to-SQL tasks is studied. DAIL-SQL (Gao et al., 2024a) considers both questions and SQL queries to select few-shot examples, adopts an example organization strategy to balance quality and quantity, and utilizes code representation prompting for question representation. Besides, C3-SQL (Dong et al., 2023) and DIN-SQL (Pourreza and Rafiei, 2023) have introduced innovative frameworks for database simplification, query decomposition, and prompt engineering.
+
+Researchers introduce schema linking to identify the database tables and columns associated with natural language queries and have proposed complex and integrated prompting engineering methods (Li et al., 2024d; Wu et al., 2025). MAC-SQL (Wang et al., 2023; Lee et al., 2024) centered on multi-agent collaboration can be utilized for more intricate data scenarios and a broader spectrum of error types for detection and correction. TA-SQL (Qu et al., 2024) and CodeS (Li et al., 2024b) introduce strategies to mitigate hallucinations in LLM-based SQL generation. CHESS (Talaei et al., 2024) enables more accurate schema linking by retrieving relevant information from database catalogs and database values. Another approach, MAG-SQL (Xie et al., 2024) features a multi-agent generative approach with soft schema linking and iterative Sub-SQL refinement. MSc-SQL (Gorti et al., 2024) mitigates the performance gap of smaller open-source models by sampling and comparing multiple SQL query results. RSL-SQL combines bidirectional schema linking, contextual information augmentation, binary selection strategy, and multiturn self-correction to achieve an efficient framework. But previous works are difficult to accurately retrieve the correct identifiers content from the schema and often overlook the understanding and reasoning of structure hints during
+
+SQL generation.
+
+# 9 Conclusion
+
+In this paper, we propose UCS-SQL, a three-stage text-to-SQL framework integrating a content pipe for key content extraction and a structure pipe for linguistic structure transformation. Through the three stages of UCL-SQL, the content and structure pipes jointly accomplish the generation from NL question to SQL query. Then, we introduce the SS-QE example selection method, which selects appropriate few-shot examples for target questions by taking into account the similarities in both SQL structure and question expression. The experimental results on the BIRD and Spider datasets demonstrate that UCS-SQL achieves superior performance in EX and VES, and substantiate the effectiveness of both the content and structure pipes. The integration of content and structure pipes will guide future research to jointly consider the text-to-SQL generation process from both dimensions.
+
+# Limitations
+
+UCS-SQL is a pipeline based approach, it will be subjected to cascading error effect. An intelligent baketracking approach to reflect and correct the intermediate stage outputs can make the system more robust, which we will design in the future work.
+
+# Ethics Statement
+
+In this work, all of the datasets, models, code and related documents are not associated with any ethical concerns.
+
+# Acknowledgments
+
+This work was partially supported by the National Natural Science Foundation of China (Grant Nos. 62276017, 62406033, U1636211, 61672081), and the State Key Laboratory of Complex& Critical Software Environment (Grant No. SKLCCSE-2024ZX-18).
+
+# References
+
+Hasan Alp Caferoğlu and Özgür Ulusoy. 2024. E-SQL: direct schema linking via question enrichment in text-to-sql. CoRR, abs/2409.16751.  
+Zhenbiao Cao, Yuanlei Zheng, Zhihao Fan, Xiaojin Zhang, Wei Chen, and Xiang Bai. 2024. RSL-SQL: robust schema linking in text-to-sql generation. CoRR, abs/2411.00073.
+
+Shuaichen Chang and Eric Fosler-Lussier. 2023. How to prompt llms for text-to-sql: A study in zero-shot, single-domain, and cross-domain settings. CoRR, abs/2305.11853.  
+Xuemei Dong, Chao Zhang, Yuhang Ge, Yuren Mao, Yunjun Gao, Lu Chen, Jinshu Lin, and Dongfang Lou. 2023. C3: zero-shot text-to-sql with chatgpt. CoRR, abs/2307.07306.  
+Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. 2024a. Text-to-sql empowered by large language models: A benchmark evaluation. Proc. VLDB Endow., 17(5):1132-1145.  
+Yingqi Gao, Yifu Liu, Xiaoxia Li, Xiaorong Shi, Yin Zhu, Yiming Wang, Shiqi Li, Wei Li, Yuntao Hong, Zhiling Luo, Jinyang Gao, Liyu Mou, and Yu Li. 2024b. Xiyan-sql: A multi-generator ensemble framework for text-to-sql. CoRR, abs/2411.08599.  
+Satya Krishna Gorti, Ilan Gofman, Zhaoyan Liu, Jiapeng Wu, Noel Vouitsis, Guangwei Yu, Jesse C. Cresswell, and Rasa Hosseinzadeh. 2024. Msc-sql: Multi-sample critiquing small language models for text-to-sql translation. CoRR, abs/2410.12916.  
+Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, JianGuang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-sql in cross-domain database with intermediate representation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28-August 2, 2019, Volume 1: Long Papers, pages 4524-4535. Association for Computational Linguistics.  
+Pengcheng He, Yi Mao, Kaushik Chakrabarti, and Weizhu Chen. 2019. X-SQL: reinforce schema representation with context. CoRR, abs/1908.08113.  
+Binyuan Hui, Ruiying Geng, Lihan Wang, Bowen Qin, Yanyang Li, Bowen Li, Jian Sun, and Yongbin Li. 2022.  $S^2$ sql: Injecting syntax to question-schema interaction graph encoder for text-to-sql parsers. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 1254-1262. Association for Computational Linguistics.  
+Dongjun Lee, Choongwon Park, Jaehyuk Kim, and Heesoo Park. 2024. MCS-SQL: leveraging multiple prompts and multiple-choice selection for text-to-sql generation. CoRR, abs/2405.07467.  
+Boyan Li, Yuyu Luo, Chengliang Chai, Guoliang Li, and Nan Tang. 2024a. The dawn of natural language to SQL: are we fully ready? [experiment, analysis & benchmark ]. Proc. VLDB Endow., 17(11):3318-3331.  
+Haoyang Li, Jing Zhang, Cuiping Li, and Hong Chen. 2023a. RESDNSL: decoupling schema linking and skeleton parsing for text-to-sql. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023., pages 13067-13075. AAAI Press.
+
+Haoyang Li, Jing Zhang, Hanbing Liu, Ju Fan, Xiaokang Zhang, Jun Zhu, Renjie Wei, Hongyan Pan, Cuiping Li, and Hong Chen. 2024b. Codes: Towards building open-source language models for text-to-sql. Proc. ACM Manag. Data, 2(3):127.  
+Jinyang Li, Binyuan Hui, Reynold Cheng, Bowen Qin, Chenhao Ma, Nan Huo, Fei Huang, Wenyu Du, Luo Si, and Yongbin Li. 2023b. Graphix-t5: Mixing pretrained transformers with graph-aware layers for text-to-sql parsing. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023., pages 13076-13084. AAAI Press.  
+Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, Xuanhe Zhou, Chenhao Ma, Guoliang Li, Kevin Chen-Chuan Chang, Fei Huang, Reynold Cheng, and Yongbin Li. 2023c. Can LLM already serve as A database interface? A big bench for large-scale database grounded text-to-sqls. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.  
+Xiang Li, Yiqun Yao, Xin Jiang, Xuezhi Fang, Chao Wang, Xinzhang Liu, Zihan Wang, Yu Zhao, Xin Wang, Yuyao Huang, Shuangyong Song, Yongxiang Li, Zheng Zhang, Bo Zhao, Aixin Sun, Yequan Wang, Zhongjiang He, Zhongyuan Wang, Xuelong Li, and Tiejun Huang. 2024c. Tele-flm technical report. CoRR, abs/2404.16645.  
+Zhongqiu Li, Zhenhe Wu, Mengxiang Li, Zhongjiang He, Ruiyu Fang, Jie Zhang, Yu Zhao, Yongxiang Li, Zhoujun Li, and Shuangyong Song. 2024d. Scalable database-driven kgs can help text-to-sql. In Proceedings of the ISWC 2024 Posters, Demos and Industry Tracks:, volume 3828 of CEUR Workshop Proceedings.  
+Shixuan Liu, Chen Peng, Chao Wang, Xiangyan Chen, and Shuangyong Song. 2023. icsberts: Optimizing pre-trained language models in intelligent customer service. In INNS DLIA@IJCNN 2023, Gold Coast, Australia, 23 June 2023, volume 222 of Procedia Computer Science, pages 127-136. Elsevier.  
+OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.  
+Mohammadreza Pourreza, Hailong Li, Ruoxi Sun, Yeounoh Chung, Shayan Talaei, Gaurav Tarlok Kakkar, Yu Gan, Amin Saberi, Fatma Ozcan, and Sercan Ö. Arik. 2024. CHASE-SQL: multi-path reasoning and preference optimized candidate selection in text-to-sql. CoRR, abs/2410.01943.  
+Mohammadreza Pourreza and Davood Rafiei. 2023. DIN-SQL: decomposed in-context learning of text-to-sql with self-correction. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023.
+
+Mohammadreza Pourreza and Davood Rafiei. 2024. DTS-SQL: decomposed text-to-sql with small large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024, pages 8212-8220. Association for Computational Linguistics.  
+Bowen Qin, Binyuan Hui, Lihan Wang, Min Yang, Jinyang Li, Binhua Li, Ruiying Geng, Rongyu Cao, Jian Sun, Luo Si, Fei Huang, and Yongbin Li. 2022. A survey on text-to-sql parsing: Concepts, methods, and future directions. CoRR, abs/2208.13629.  
+Ge Qu, Jinyang Li, Bowen Li, Bowen Qin, Nan Huo, Chenhao Ma, and Reynold Cheng. 2024. Before generation, align it! A novel and effective strategy for mitigating hallucinations in text-to-sql generation. In *Findings of the Association for Computational Linguistics*, ACL 2024., pages 5456-5471. Association for Computational Linguistics.  
+Jiawei Shao and Xuelong Li. 2024. AI flow at the network edge. CoRR, abs/2411.12469.  
+Ruoxi Sun, Sercan Ö. Arik, Hootan Nakhost, Hanjun Dai, Rajarishi Sinha, Pengcheng Yin, and Tomas Pfister. 2023. SQL-palm: Improved large language model adaptation for text-to-sql. CoRR, abs/2306.00739.  
+Chang-Yu Tai, Ziru Chen, Tianshu Zhang, Xiang Deng, and Huan Sun. 2023. Exploring chain of thought style prompting for text-to-sql. In EMNLP 2023, Singapore, December 6-10, 2023, pages 5376-5393. Association for Computational Linguistics.  
+Shayan Talaei, Mohammadreza Pourreza, Yu-Chen Chang, Azalia Mirhoseini, and Amin Saberi. 2024. CHESS: contextual harnessing for efficient SQL synthesis. CoRR, abs/2405.16755.  
+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.  
+Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. RAT-SQL: relation-aware schema encoding and linking for text-to-sql parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7567-7578. Association for Computational Linguistics.  
+Bing Wang, Changyu Ren, Jian Yang, Xinnian Liang, Jiaqi Bai, Qian-Wen Zhang, Zhao Yan, and Zhoujun Li. 2023. MAC-SQL: A multi-agent collaborative framework for text-to-sql. CoRR, abs/2312.11242.  
+Zihan Wang, Xinzhang Liu, Shixuan Liu, Yitong Yao, Yuyao Huang, Zhongjiang He, Xuelong Li, Yongxiang Li, Zhonghao Che, Zhaoxi Zhang, Yan Wang, Xin Wang, Luwen Pu, Huihan Xu, Ruiyu Fang, Yu Zhao, Jie Zhang, Xiaomeng Huang, Zhilong Lu,
+
+Jiaxin Peng, Wenjun Zheng, Shiquan Wang, Bingkai Yang, Xuewei He, Zhuoru Jiang, Qiyi Xie, Yanhan Zhang, Zhongqiu Li, Lingling Shi, Weiwei Fu, Yin Zhang, Zilu Huang, Sishi Xiong, Yuxiang Zhang, Chao Wang, and Shuangyong Song. 2024a. Telechat technical report. CoRR, abs/2401.03804.  
+Zihan Wang, Yitong Yao, Li Mengxiang, Zhongjiang He, Chao Wang, Shuangyong Song, et al. 2024b. Telechat: An open-source billngual large language model. In Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10), pages 10-20.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.  
+Hefeng Wu, Yandong Chen, Lingbo Liu, Tianshui Chen, Keze Wang, and Liang Lin. 2023. Sqlnet: Scale-modulated query and localization network for few-shot class-agnostic counting. CoRR, abs/2311.10011.  
+Zhenhe Wu, Zhongqiu Li, Mengxiang Li, Jie Zhang, Zhongjiang He, Jian Yang, Yu Zhao, Ruiyu Fang, Yongxiang Li, Zhoujun Li, and Shuangyong Song. 2025. MR-SQL: multi-level retrieval enhances inference for llm in text-to-sql. DASFAA.  
+Wenxuan Xie, Gaochen Wu, and Bowen Zhou. 2024. MAG-SQL: multi-agent generative approach with soft schema linking and iterative sub-sql refinement for text-to-sql. CoRR, abs/2408.07930.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.  
+Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In ICLR 2023, Kigali, Rwanda, May 1-5, 2023.  
+Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir R. Radev. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3911-3921. Association for Computational Linguistics.  
+Yanzhao Zheng, Haibin Wang, Baohua Dong, Xingjun Wang, and Changshan Li. 2022. HIE-SQL: history information enhanced network for context-dependent text-to-sql semantic parsing. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 2997-3007. Association for Computational Linguistics.
+
+<table><tr><td>Datasets</td><td>Train</td><td>Dev</td><td>Test</td><td>DB</td><td>Table/DB</td><td>Row/DB</td></tr><tr><td>BIRD</td><td>9,428</td><td>1,534</td><td>1,789</td><td>95</td><td>7.3</td><td>549k</td></tr><tr><td>Spider</td><td>8,659</td><td>1,034</td><td>2,147</td><td>200</td><td>5.1</td><td>2k</td></tr></table>
+
+Table 6: The statistics of BIRD and Spider datasets.  
+
+<table><tr><td>ICL Prompt</td><td>Content Pipe</td><td>Structure Pipe</td></tr><tr><td>Question
+What is the gender of the youngest client who
+opened account in the lowest average salary branch?
+Evidence
+Later birthdate refers to younger age;
+A11 refers to average salary
+Foreign keys
+account. &#x27;district_id&#x27; = &#x27;district&#x27;. &#x27;district_id&#x27;
+client. &#x27;district_id&#x27; = &#x27;district&#x27;. &#x27;district_id&#x27;
+Pre-filtered Database schema
+Table: account
+[(account_id, the id of the account),
+(distinct_id, location of branch),
+-(frequency, frequency of the account),
+(date, the creation date of the account)]
+Table: district
+[(district_id, location of branch),
+(A4, number of inhabitants),
+(A11, average salary)]
+Bi-filtered Database schema
+Table: account
+[(account_id, the id of the account),
+(distinct_id, location of branch),
+-(frequency, frequency of the account),
+(date, the creation date of the account)]
+Table: district ...</td><td>Content Information
+The following attribute values may be table or column names:
+[gender], [client], [average salary]
+Evidence Combination
+Combined with [Evidence] :
+[Later birthdate] refers to [younger age]
+[A11] refers to [average salary]
+The following attributes may be table or column names:
+[gender], [client], [average salary] → [A11]
++ [Later birthdate]
+Foreign keys Combination
+Combined with [Foreign keys] , the final &#x27;content information&#x27;:
+[gender], [client], [Later birthdate], [A11]
++ [district]
++ [district_id]
+Bi-directional Schema Linking
+(1) Combine the results above with [Database schema] ,
+[Evidence] and [Foreign keys] , deduce the possible final
+identifiers according to schema.
+(2) Consider [Question] and the [final identifiers] , keep all
+primary keys, we can filter out the irrelevant columns from schema.
+Final Identifiers
+We can known from the schema that [Later birthdate] imply
+column name [birth_date] in table [client].
+- Tables:
+[client], [district]</td><td>Structure Information
+The following words imply part of SQL structure:
+[the youngest] [the lowest] [average]
+Function Induction
+According to structure information, SQL may involve functions:
+[the youngest] → [ORDER BY], [ASC], [DECS]
+[the lowest] → [ORDER BY],
+[ASC], [DECS]
+[average] → [AVG]
+Final Functions
+- Functions:
+[ORDER BY], [ASC], [DECS], [AVG]
+SQL Skeleton Generation
+Considering constraints, generate the skeleton of SQL refer to the
+recommendation functions in &#x27;Final Function&#x27;:
+SELECT T1.[column_name]
+FROM [table_name] AS T1
+INNER JOIN [table_name] AS T2
+ON T1.[column_name] = T2.[column_name]
+ORDER BY T2.[column_name] ASC, T1.[column_name] DESC LIMIT 1
+SQL Skeleton Filling
+Fill the SQL skeleton, replace [table_name] and [column_name]
+refer to the recommendation identifiers in &#x27;Final identifiers&#x27;:
+SELECT T1. gender&#x27;
+FROM client AS T1
+INNER JOIN district AS T2
+ON T1.[district_id] = T2.[district_id]
+ORDER BY T2.&#x27;A11&#x27; ASC, T1&#x27;[birth_date&#x27; DESC LIMIT 1</td></tr></table>
+
+Figure 9: Detailed prompt design for UCL-SQL. The left column displays the original question and database information. Content and structure information are extracted from the question to activate the content and structure pipes. Identifiers and SQL skeletons are generated at the end of each pipe and combined to form the final SQL query.
+
+# A Statistics of datasets.
+
+Table 6 shows the statistics of two datasets.
+
+# B Case Study
+
+In this section, we elucidate the UCS-SQL framework through a detailed examination of an incontext learning prompt, as depicted in Figure 9. The process commences with the extraction of both content and structural information from the original question.
+
+Within the content pipe, the LLM discerns that the natural language phrase "average salary" corresponds to the schema identifier "A11". Consequently, "average salary" is mapped to "A11". Additionally, the mention of "Later birthdate" prompts its provisional inclusion in the candidate set. Subsequently, we engage in the Foreign Keys Combination phase. Since "A11" is associated with the "dis
+
+strict" table and "gender" with the "client" table, a foreign key linking these two tables is essential for SQL query generation. This necessitates the inclusion of "district" and "district_id" from the equivalence "client.'district_id' = district.'district_id'" into our candidate set. Advancing to the Bidirectional schema linking phase, we identify "gender", "A11", and "client" within the schema. The schema's description of "birth_data" (birth date) is recognized as the closest match to "Later birthdate", leading the LLM to select "birth_data" as a substitute. Ultimately, the final identifiers consist of the tables "client" and "district", and the columns "gender", "birth_date", "A11", and "district_id". Thereafter, the LLM integrates these final identifiers, refining the pre-filtered schema by eliminating the redundant column "A4", obtaining the bi-filtered schema.
+
+In the structure pipe, the LLM initially infers candidate functions "ORDER BY", "ASC", "DESC", and "AVG" from the structure information "the youngest", "the lowest" and "average". Integrating these with the query and constraints, the LLM constructs the SQL skeleton. Ultimately, with reference to the bi-filtered schema, the LLM populates the SQL skeleton with the candidate table names and column names from the content pipe, thereby completing the final SQL query.

2025/UCS-SQL_ Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f249b68fbc37e1c709d8ca3f2d97051b3dde72acccad1601107a839b4e6e71a
+size 736705

2025/UI-E2I-Synth_ Advancing GUI Grounding with Large-Scale Instruction Synthesis/65ba1011-e397-40da-8aed-28e8c07728b8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65c921d3e3baa804956053f3f33f9b8cf09d5bdf979096ddde2972b041881c2d
+size 1079087

2025/UI-E2I-Synth_ Advancing GUI Grounding with Large-Scale Instruction Synthesis/full.md

@@ -0,0 +1,489 @@
+# UI-E2I-Synth: Advancing GUI Grounding with Large-Scale Instruction Synthesis
+
+Xinyi Liu $^{1,2*†}$  Xiaoyi Zhang $^{1\dagger}$  Ziyun Zhang $^{1,2*}$  Yan Lu $^{1}$
+
+<sup>1</sup>Microsoft Research Asia
+
+$^{2}$ School of Software and Microelectronics, Peking University
+
+# Abstract
+
+Recent advancements in Large Vision-Language Models are accelerating the development of Graphical User Interface (GUI) agents that utilize human-like vision perception capabilities to enhance productivity on digital devices. Compared to approaches predicated on GUI metadata, which are platform-dependent and vulnerable to implementation variations, vision-based approaches offer broader applicability. In this vision-based paradigm, the GUI instruction grounding, which maps user instruction to the location of corresponding element on the given screenshot, remains a critical challenge, particularly due to limited public training dataset and resource-intensive manual instruction data annotation. In this paper, we delve into unexplored challenges in this task including element-to-screen ratio, unbalanced element type, and implicit instruction. To address these challenges, we introduce a large-scale data synthesis pipeline UI-E2I-Synth for generating varying complex instruction datasets using GPT-4o instead of human annotators. Furthermore, we propose a new GUI instruction grounding benchmark UI-I2E-Bench, which is designed to address the limitations of existing benchmarks by incorporating diverse annotation aspects. Our model, trained on the synthesized data, achieves superior performance in GUI instruction grounding, demonstrating the advancements of proposed data synthesis pipeline. The proposed benchmark, accompanied by extensive analyses, provides practical insights for future research in GUI grounding. We will release corresponding artifacts here.
+
+# 1 Introduction
+
+Graphical User Interface (GUI) agents are designed to understand and automate human instruc
+
+![](images/f8d38724164eaead92c6d56e631b9e902e810c9fef4a2986c73cae8569d70f98.jpg)  
+Figure 1: GUI grounding requires to localize elements in screenshots based on user instructions, with implicit instructions posing greater reasoning challenges than explicit ones.
+
+![](images/c5a38cdffcd11f24d2435ad01e2f4f2dc9283fc4f2b94b93a3415356fbdec698.jpg)  
+Figure 2: Left: Existing benchmark's element-to-screen ratio significantly deviates from typical real-world desktop displays like 1080p and 1440p, on landscape samples. Right: The comparison shows Text and Icon dominates existing benchmark, leaving the rest of element types omitted.
+
+![](images/5bc0d9a8aae016b5b4c824fa5e8bfd4f7df065a494cde04c713775ee4213eba6.jpg)
+
+tion in the digital devices, promising a future where humans are liberated from repetitive tasks, thereby enhancing operational efficiency and productivity. The rise of Large Language Models (LLMs) (Achiam et al., 2023) made it feasible to generally understand human instructions and structured text data. Hence, early efforts (Gur et al., 2023; Zhou et al., 2023; Shaw et al., 2023; Deng
+
+et al., 2024) are made to operate GUI by understanding the text-rich metadata behind GUI, e.g., HTML. However, some research (Fu et al., 2024; Zheng et al., 2024; WebAIM, 2024) has shown that GUI metadata is vulnerable and highly dependent on application developers for specific platforms. To overcome these limitations, more recent research (Zhang et al., 2023b; Cheng et al., 2024a; Gou et al., 2024; Wu et al., 2024; Lin et al.) has concentrated on training Vision-Language Models (VLMs) to mimic human behavior in GUI operations, manipulating GUI elements from pixels based on user instructions. This approach makes GUI grounding capability as the core, which maps the user instruction into the corresponding element by outputting specific coordinates.
+
+Despite these works show satisfying progress on existing benchmark (Cheng et al., 2024b). Our investigation shows that there still is big gap between the existing benchmark and complicated and various GUI grounding environments. In this work, we thoroughly study the whole progress to build an element grounding dataset from screenshot collection to final instruction generation, revealing three main aspects that have not been fully explored previously: (a) Element-to-screen ratio: GUI instruction grounding necessitates higher resolution and smaller elements compared to natural scene object grounding. While previous studies (Cheng et al., 2024b; Gou et al., 2024) have discussed the screenshot resolution in training dataset, we argue that the key factor is the element-to-screen ratio, i.e., the size of an element relative to the screenshot. This ratio is affected by both resolution and UI zoom level. As shown in the left of Figure 1, the landscape samples in existing benchmark has a larger element-to-screen ratio than typical real-world desktop displays like 1080p and 1440p, potentially leading to an overestimation of model performance. (b) Unbalanced element type: Different GUI element types exhibit diverse appearances and interaction designs, also with varying frequencies of occurrence. For instance, text buttons represent the most prevalent category, whereas checkboxes are comparatively less common. Unlike text buttons, whose functionality is often explicitly conveyed by their textual labels, checkboxes typically rely on surrounding elements to define their purpose. This distinction has not been adequately addressed in previous research. (c) Implicit instruction: In the context of GUI instruction grounding, users could convey their instructions based on their
+
+own intuitive understanding of element functionality or position, causing lack of direct correspondence with visible text on the screenshot, which we refereed as "implicit instruction". Figure 1 showcases the complexity of implicit instruction. The implicit instruction challenges the understanding and reasoning capability of VLMs.
+
+To address the aforementioned challenges, we introduce a large-scale instruction grounding data synthesis pipeline, termed Element-to-Instruction Synthesis, abbreviated as UI-E2I-Synth. This pipeline utilizes GPT-4o instead of human annotators to synthesize realistic grounding user instructions of varying complexity, which differs from the element referring expressions that previous works curated. By the principle of divide and conquer, UI-E2I-Synth decomposes grounding instruction synthesis into three subtasks and execute them step by step. Initially, we collect screenshot-metadata pairs from various sources at different resolutions, then using a heuristic GUI metadata parser to extract reliable element attributes. Next, With these high-quality attributes to alleviate hallucinations, both explicit and implicit referring expressions are prompted to generate from GPT-4o. In the final step, GPT-4o is employed again to simulate user behavior by generating specific action parameters. These action parameters are then combined with the REs to synthesize final realistic user instructions.
+
+To more comprehensively evaluate model performance on GUI instruction grounding, we introduce a new benchmark UI-12E-Bench that includes detailed annotations curated through a combination of our synthesis pipeline and human annotators. These annotations encompass various dimensions such as element type, element-to-screen ratio, and the level of instruction implicitness. Through rigorous evaluation on our proposed benchmark, we demonstrate that existing VLMs are still far from being satisfactory. To address the gap, we apply the proposed UI-E2I-Synth to collect a synthesized GUI grounding dataset comprising 9.9M instructions. Then following OS-Atlas (Wu et al., 2024) we fine-tune two different pre-trained VLMs on the curated dataset and evaluate them. The experimental results demonstrate that our models achieve advanced performance on both the existing benchmarks and our proposed one, with substantially less instruction data. Notably, our models show outstanding capabilities in comprehending implicit instructions and handling long-tailed element types.
+
+These findings validate the efficacy of our data synthesis framework for GUI grounding, while our comprehensive benchmark analysis provides practical insights for future research in this domain.
+
+# 2 Related work
+
+Autonomous GUI Agent and Vison-Language Models. With the development of LLMs, researchers have built autonomous agents to perform complex reasoning tasks (Sumers et al.; Yang et al., 2024; Xi et al., 2025). Early works (Gur et al., 2023; Zhou et al., 2023; Shaw et al., 2023; Deng et al., 2024) have initiated the development of GUI agent frameworks for webpages and mobile platforms using GUI metadata. To address the limitations of unstable GUI metadata, ResponsibleTA (Zhang et al., 2023b) introduced a vision-based GUI agent framework and SeeClick (Cheng et al., 2024a) highlighted the importance of instruction grounding in downstream GUI agent tasks. While works like RUIG (Zhang et al., 2023a) improved GUI grounding via reinforcement learning, current general-purpose VLMs still lack satisfactory GUI grounding capabilities. Consequently, recent studies (Gou et al., 2024; Wu et al., 2024; Lin et al.) have focused on enhancing the GUI grounding abilities of VLMs. simultaneously, in industry, closed-source solutions (OpenAI, 2025; Anthropic, 2024) have advanced rapidly but with minimal detail disclosure. Our work from data perspective contributes by synthesizing large-scale GUI grounding instructions and is intended to benefit the open-source community.
+
+Instruction Dataset Synthesis. Instruction synthesis (Ding et al., 2023; Wang et al., 2022; Xu et al., 2023) has emerged as a prevalent strategy in the field of large language models due to its effectiveness in reducing the need for human effort in labeling datasets of millions of samples. Recent works make preliminary exploration about GUI grounding data synthesis. SeeClick utilizes HTML tags to generate referring expressions for webpage elements. UGround improves on this by employing a hybrid pipeline that creates referring expressions using both external LLMs and VLMs. OS-Atlas derives instructions from sequential UI changes but faces scalability issues due to the need of screenshot traces. Our data synthesis pipeline differs from previous works in several key aspects, most notably in that prior studies have focused solely on element referring expressions rather than gen
+
+uine user instructions. These existing approaches overlook the complex role that users play in formulating instructions and also screen-to-element ratio and long-tailed element type.
+
+# 3 Approach
+
+Model-based large-scale GUI grounding instruction synthesis is challenging for multimodal hallucinations and varying complexity of user instruction. UI-E2I-Synth includes three key steps: raw data collection and parsing, referring expression generation and instruction synthesis, as illustrated in Figure 3. These steps are executed sequentially, with each step designed to address specific challenge: extracting reliable element attributes to mitigate hallucinations, generating explicit and implicit referring expressions to diversify instruction difficulty, and synthesizing instructions with user action simulation. In the following subsections we introduce each step on details.
+
+# 3.1 Raw Data Collection and Parsing
+
+This component is designed to collect diverse screenshots from multiple platforms and obtain reliable element attributes from GUI metadata for the following referring expression generation.
+
+We first build data collection infrastructure across various platforms to obtain the GUI screenshot and metadata pairs. For Web platform, we use the dumped webpage metadata in Commoncrawl (CommonCrawl, 2024) and re-render as webpage screenshots. For Windows platform, we select common applications, traverse the application to obtain different UI screenshots. Specifically, we start from the initial UI of a given application, parse out clickable buttons from its metadata and click these buttons to transit to other UI in this application. All the process is conducted under a virtual machine where no personal information is involved and all corresponding UIA data is recorded. For mobile platform, we leverage the existing metadata-screenshot pairs in training set of AndroidControl (Li et al., 2024). Different platforms yield different metadata formats, such as DOM (Mozilla, 2024) for Web, UIA (Microsoft, 2021) for Windows, VH (Google, 2024) for Android. To unify element representation from different platform metadata, we build corresponding heuristic element parser to extract only three key attributes including element type, element content and element bounding box. For element type, we
+
+![](images/bda28743386ac3e57bb9b4c8995be019cd00ed8f9f297f2ca1e18a1fc17b3278.jpg)  
+Figure 3: Overview of proposed data synthesis pipeline UI-E2I-Synth. The three steps respectively generate a high-quality pool of elements and their corresponding attributes, diverse element referring expressions, and user instructions. The output of the synthesis pipeline is format as <screenshot, user instruction, element coordinates>. GPT-4o is utilized in the component Referring Expression Generator and Instruction Synthesis.
+
+analyze the UI element representation in different metadata format and categorize them into five main element types:
+
+- Text: A button element with explicit text.  
+- **Inputfield:** An element requiring user input or content editing.  
+- Extraction: A button that allows user input by selecting from provided options.  
+- Icon: A graphical button representing functionality through an image.  
+- Toggle: A two-state element, such as checkboxes, radio buttons, and switches.
+
+Accurately parsing every element for each platform is extremely difficult, which means tons of rules need be constructed manually. However, GUI grounding data does not require perfect parsing for every element in a single screenshot. Instead, our goal is to extract sufficiently accurate element attributes from large-scale screenshot-metadata pairs. With this goal, we build our heuristic metadata parser which contains heuristic rules to only keep high precision but leaving recall aside.
+
+After obtaining the three key element attributes, we measure the element type distribution, adjust the distribution through resampling, and compose a balanced candidate pool for the next step.
+
+# 3.2 Referring Expression Generation
+
+The Referring Expression Generation step is designed to enable exciting Large VLM to generate expressions that are both reliable and diverse in difficulty. The element referring expressions (REs) denote a element description from a specific perspective, which is independent from user action. Generating elements and their descriptions from scratch can be challenging for Large VLMs, as
+
+they often struggle to produce accurate coordinates and are prone to hallucinations (Zheng et al., 2024) under high-resolution images. A straightforward method of generating element REs is using Set-of-Marks (Yang et al., 2023) to caption every marked element. However, this method also tends to result in severe hallucinations. To address this issue, we propose a attribute-enhanced style of RE generation. We supply a list containing the element types and element content obtained from the previous step, with Set-of-Marks screenshots as context. We then leverage GPT-4o as the Large VLM to first generate a full description that explains the function of each element and the expected outcome when interacting with it. Next, GPT-4o is prompted with generating two types of REs: explicit RE and implicit RE. We define explicit RE as directly referring to the obvious features of the element, while implicit RE refers to the element by describing the semantic function or its relationship with nearby elements, thus avoiding the obvious visible features. The generated explicit and implicit REs collectively form the element RE pool, which provides user action object for the final step.
+
+# 3.3 Instruction Synthesis
+
+The obtained referring expressions are only descriptions about element, omitting the user role in the progress. User instruction implies intention from user, which can be directly or implicitly related to the element. When directly asked to generate user instruction, we observed that Large VLM trends to generate generally instruction from a role of assistant, such as "Fill your details in the input field" or "Click to check your profile", even we ask it to generate as a computer user's first perspective view.
+
+<table><tr><td>Benchmark</td><td>Sample Num</td><td>Landscape element-to-screen ratio</td><td>Fine-grained type annotation</td><td>Min type proportion</td><td>Implicitness annotation</td><td>Implicit instruction proportion</td></tr><tr><td>ScreenSpot</td><td>1,272</td><td>0.088</td><td>X</td><td>3.21%†</td><td>X</td><td>-</td></tr><tr><td>Our UI-I2E-Bench</td><td>1,477</td><td>0.042</td><td>✓</td><td>12.34%</td><td>✓</td><td>63.03%</td></tr></table>
+
+Table 1: Comparison between proposed benchmark and existing benchmark. Screen-to-element ratio is calculated as the ratio of the square roots of the areas of the box and the image.  $\dagger$  denotes we re-associate ScreenSpot-Web with our fine-grained element type and calculate the number.
+
+This phenomenon may arise because the output message role in the Large VLM is set to "assistant", preventing it from easily simulating the user role. Hence, here we propose to a parametrization way to generate user instruction. We decompose the wanted instruction into three parameters, user action type, user action content and element object. A prompt is utilized to instruct GPT-4o to simulate a user interacting with the current application, by generate specific user actions and content. Combining previous element referring expressions as the element object, GPT-4o is further instructed with synthesizing the final instruction with all these action parameters. To maintain diversity in the generated instructions, both explicit and implicit element referring expressions are employed to create distinct user instructions. To this end, the output of the whole synthesis pipeline is format as <screenshot, user instruction, element coordinates>.
+
+With this pipeline, we synthesize a large-scale grounding instruction data to train a model to demonstrate our advancement. Dataset details are explained in Section 5.1.
+
+# 4 UI-I2E-Bench: A Comprehensive Grounding Benchmark
+
+As shown in Figure 1, the existing benchmark ScreenSpot exhibits a lower level of difficulty in element-to-screen ratio. Additionally, it lacks detailed element annotations, such as element type and implicitness, which are crucial for assessing the element grounding capabilities of models. To address these limitations, we introduce UI-I2E-Bench, developed through a semi-automated workflow to provide a more comprehensive evaluation. We first apply our data synthesis pipeline into data collected from multiple platforms, then we sampled the instructions by element-to-screen ratio and element type. We manually review the validity of parsed element and correctness of generated instruction to obtain the final UI-I2E-Bench, which includes 1477 grounding instructions from Web, Windows and Android. The further statistical data and comparisons with ScreenSpot are included in Table 1.
+
+Our benchmark offers more comprehensive annotations, a lower element-to-screen ratio, and a higher proportion of implicit instructions. These features facilitate in-depth performance diagnostics in challenging GUI instruction grounding cases.
+
+# 5 Experiments
+
+To demonstrate the effectiveness of our grounding instruction synthesis pipeline, we utilize UI-E2I-Synth to collect a GUI grounding training dataset and fine-tune a pre-trained VLM to develop UI-I2E-VLM. This section is structured as follows: First, we provide details on the curation of the training dataset and model training. Next, we introduce various baselines and evaluate the performance of UI-I2E-VLM against these baselines on two GUI instruction grounding benchmarks. We analyze the results on UI-I2E-Bench to reveal the shortcomings of current VLMs across multiple aspects. We then integrate UI-I2E-VLM with the GPT-4o planner and assess its performance on the live GUI agent benchmark OSWorld (Xie et al., 2024).
+
+# 5.1 Training Details
+
+Training data curation. In this subsection, we briefly describe the final curated dataset leveraging the UI-E2I-Synth pipeline while leaving more details in our supplementary materials. We apply our pipeline mainly on three platforms including Web, Windows and Android to derive this training dataset. The Web is our main data source due to the variety of layouts and design styles across websites, as well as the extensive quantity of webpages available in Common CWEl. We start by extracting webpages from the top 500k domains with the highest traffic for three webpages per domain. Next, we filter out non-English and error state webpages, resulting in final 724,839 webpages. These webpages are re-rendered at seven different resolutions, including one mobile resolution to simulate mobile scenarios and left six resolutions for landscape desktops. For Windows platform, we select 90 windows apps to totally collect 15K screenshot-metadata pairs. For Android platform, we leverage existing screenshot-metadata pairs in training
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Size</td><td rowspan="2">#Train</td><td colspan="4">ScreenSpot</td><td colspan="3">UI-I2E-Bench</td><td rowspan="2">ScreenSpot-Pro</td><td rowspan="2">Avg.*</td></tr><tr><td>Mobile</td><td>Web</td><td>Desktop</td><td>Avg.</td><td>Explicit</td><td>Implicit</td><td>Avg.</td></tr><tr><td>InternVL2-4B</td><td>4B</td><td>-</td><td>7.2</td><td>0.5</td><td>4.5</td><td>4.2</td><td>1.4</td><td>0.5</td><td>0.9</td><td>0.3</td><td>1.8</td></tr><tr><td>Qwen2-VL-7B</td><td>7B</td><td>-</td><td>51.3</td><td>27.7</td><td>49.1</td><td>42.6</td><td>53.8</td><td>45.6</td><td>48.7</td><td>1.6</td><td>31.0</td></tr><tr><td>OmniParser</td><td>-</td><td>-</td><td>78.5</td><td>63.9</td><td>79.7</td><td>73.9</td><td>54.3</td><td>52.4</td><td>53.1</td><td>8.3</td><td>45.1</td></tr><tr><td>Seeclick</td><td>9.6B</td><td>1.0M</td><td>66.1</td><td>44.7</td><td>54.5</td><td>55.8</td><td>37.1</td><td>19.9</td><td>26.4</td><td>1.1</td><td>27.8</td></tr><tr><td>UGround</td><td>7B</td><td>9.7M</td><td>72.5</td><td>75.7</td><td>74.6</td><td>74.1</td><td>65.8</td><td>47.1</td><td>54.2</td><td>16.5</td><td>48.3</td></tr><tr><td>ShowUI</td><td>2B</td><td>2.7M</td><td>84.6</td><td>73.2</td><td>69.9</td><td>76.8</td><td>51.3</td><td>35.6</td><td>41.5</td><td>7.7</td><td>42.0</td></tr><tr><td>OS-Atlas-4B</td><td>4B</td><td>13.6M</td><td>73.3</td><td>73.4</td><td>61.1</td><td>70.1</td><td>51.5</td><td>39.9</td><td>44.3</td><td>3.7</td><td>39.4</td></tr><tr><td>OS-Atlas-7B</td><td>7B</td><td>13.6M</td><td>83.8</td><td>83.1</td><td>79.7</td><td>82.5</td><td>63.2</td><td>55.8</td><td>58.6</td><td>18.9</td><td>53.3</td></tr><tr><td>UI-I2E-VLM-4B</td><td>4B</td><td>9.9M</td><td>70.3</td><td>70.9</td><td>70.1</td><td>70.4</td><td>61.9</td><td>48.3</td><td>53.4</td><td>12.2</td><td>45.3</td></tr><tr><td>UI-I2E-VLM-7B</td><td>7B</td><td>9.9M</td><td>86.5</td><td>78.0</td><td>82.6</td><td>82.5</td><td>72.0</td><td>67.9</td><td>69.5</td><td>23.6</td><td>58.5</td></tr></table>
+
+Table 2: Results on GUI grounding benchmarks. #Train denotes the instructions number in training data. We calculate the average accuracy as correct samples divided by total samples. Avg.* denotes the arithmetic mean of average accuracy across three benchmarks.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Size</td><td rowspan="2">#Train</td><td colspan="3">Platform</td><td colspan="5">Element Type</td></tr><tr><td>Web</td><td>Desktop</td><td>Mobile</td><td>Button</td><td>Icon</td><td>Dropdown</td><td>Input</td><td>Toggle</td></tr><tr><td>OmniParser</td><td>-</td><td>-</td><td>30.8</td><td>45.5</td><td>67.6</td><td>68.4</td><td>60.5</td><td>65.9</td><td>58.9</td><td>26.9</td></tr><tr><td>Seeclick</td><td>9.6B</td><td>1.0M</td><td>18.2</td><td>15.8</td><td>37.2</td><td>31.6</td><td>26.2</td><td>22.5</td><td>29.6</td><td>22.1</td></tr><tr><td>Uground</td><td>7B</td><td>9.7M</td><td>53.0</td><td>44.3</td><td>61.8</td><td>57.3</td><td>49.7</td><td>76.4</td><td>64.2</td><td>37.0</td></tr><tr><td>ShowUI</td><td>2B</td><td>2.7M</td><td>29.6</td><td>30.4</td><td>53.9</td><td>52.0</td><td>44.1</td><td>51.1</td><td>52.8</td><td>18.9</td></tr><tr><td>OS-Atlas-4B</td><td>4B</td><td>13.6M</td><td>54.6</td><td>19.9</td><td>58.6</td><td>43.5</td><td>44.1</td><td>46.6</td><td>46.3</td><td>42.2</td></tr><tr><td>OS-Atlas-7B</td><td>7B</td><td>13.6M</td><td>52.2</td><td>48.9</td><td>68.1</td><td>69.1</td><td>58.7</td><td>80.3</td><td>70.1</td><td>32.3</td></tr><tr><td>UI-I2E-VLM-4B</td><td>4B</td><td>9.9M</td><td>60.9</td><td>38.9</td><td>61.4</td><td>54.3</td><td>50.0</td><td>61.2</td><td>68.6</td><td>39.0</td></tr><tr><td>UI-I2E-VLM-7B</td><td>7B</td><td>9.9M</td><td>62.1</td><td>64.0</td><td>76.2</td><td>77.0</td><td>68.2</td><td>84.8</td><td>86.2</td><td>44.4</td></tr></table>
+
+Table 3: Detailed results on UI-I2E-Bench. #Train denotes the instruction number in training data.
+
+set of AndroidControl (Li et al., 2024). Then we run the UI-E2I-Synth pipeline on all the collected screenshot-metadata pairs to derive our synthesized dataset. Combined with the existing mobile GUI dataset MOTIF (Burns et al., 2022) and WidgetCaption (Li et al., 2020) for grounding purpose, we final training dataset includes 1.6M screenshots and 9.5M instructions. The detailed training data statistics are provided in AppendixA. Our data has passed our ethics review where personal information has been examined and removed.
+
+Base models. Following OS-Atlas (Wu et al., 2024), We consider InternVL2-4B (Chen et al., 2023) and Qwen2-VL-7B (Wang et al., 2024) as our base models. The pretraining data of Qwen2-VL-7B incorporate GUI screenshots, whereas InternVL2 does not. Both models support multi-resolution image processing. We name our models as UI-I2E-VLM-4B and UI-I2E-VLM-7B, with training details provided in Appendix B.
+
+# 5.2 GUI Instruction Grounding Evaluation
+
+In this section, we compare UI-I2E-VLM with OmniParser (Lu et al., 2024), SeeClick (Cheng et al.,
+
+2024b), UGround (Gou et al., 2024), ShowUI (Lin et al.) and OS-Atlas (Wu et al., 2024) on multiple GUI grounding benchmarks. OmniParser is a screen parsing tool containing a series of models. We combine OmniParser with GPT-4o to perform GUI grounding as its original paper suggests. SeeClick, UGround, ShowUI and OS-Atlas are VLMs fine-tuned for GUI instruction grounding tasks. Following their inference setting, we prompt SeeClick, UGround and ShowUI to generate  $(\mathrm{x},\mathrm{y})$  coordinates. For OS-Atlas and UI-I2E-VLM, we prompt the model to generate a bounding box and use its center point as the predicted coordinate.
+
+For the evaluation benchmarks, we not only include the proposed benchmark but also the popular ScreenSpot even a concurrent benchmark ScreenSpot-Pro. ScreenSpot is a popular GUI grounding benchmark consisting of 1,272 cross-platform samples. ScreenSpot-Pro is a very recent benchmark focusing on professional desktop applications with high resolution. We use accuracy as the evaluation metric, where a prediction is considered as a correct one if the coordinate falls within the ground-truth bounding box.
+
+Main result comparison. We present a comprehensive performance comparison across the aforementioned three benchmarks to evaluate the models' cross-domain generalization capabilities. As demonstrated in Table 2, UI-I2E-VLM-7B achieves superior performance across all benchmarks, surpassing the previous state-of-the-art model, OSAtlas-7B, with a  $9.7\%$  relative improvement in average performance. Notably, the robust improvement on all the three distinct benchmarks is attained with a training set containing only  $72\%$  of the instruction quantity used in OS-Atlas, which makes our curated data quality more impressive. For our UI-I2E-VLM-4B, despite its base model being pretrained without GUI-specific data, still achieves impressive performance compared to the models of similar parameter size. These findings highlight the effectiveness of our data synthesis pipeline.
+
+By comparing the numerical results between ScreenSpot and UI-I2E-Bench, our findings indicate that the reported progress in GUI grounding capabilities has been substantially overestimated. A notable example is ShowUI, which demonstrates seemingly impressive performance metrics on ScreenSpot despite utilizing significantly fewer model parameters and training data. However, its performance drops markedly when evaluated on UI-I2E-Bench and ScreenSpot-Pro, revealing critical limitations in the complexity and representativeness of the original ScreenSpot benchmark.
+
+Teardown diagnostic on UI-I2E-Bench. Leveraging the diverse annotation aspects provided in UI-I2E-Bench, we conducted a detailed teardown analysis of model performance on UI-I2E-Bench. We report the detailed performance on the three dimensions separately in Table 2 and 3. Compared to the explicit instruction category, we observe more significant improvement happening to the implicit ones. The most outstanding model, OS-Atlas-7B falls behind of us for 12.1 percentage on this dimension, which means previous works underrate on the complexity of instructions. Another interesting observation is that OmniParser shows a competitive result on implicit instruction with understanding abilities of GPT-4o. However, it falls behind on simpler explicit instructions. It suggests that the primary bottleneck of OmniParser lies in localizing smaller and long-tailed element.
+
+To further study the impact of element-to-screen ratio in GUI grounding, we analyze the accuracy across different ranges of the element-to-screen ra
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Mobile</td><td colspan="2">Desktop</td><td colspan="2">Web</td><td rowspan="2">Avg.</td></tr><tr><td>Text</td><td>Icon</td><td>Text</td><td>Icon</td><td>Text</td><td>Icon</td></tr><tr><td>OS-Atlas-Web</td><td>35.2</td><td>24.9</td><td>67.0</td><td>30.7</td><td>70.9</td><td>39.8</td><td>44.9</td></tr><tr><td>UI-E2I-Synth-Web</td><td>89.0</td><td>34.1</td><td>93.3</td><td>42.9</td><td>79.6</td><td>44.7</td><td>65.8</td></tr><tr><td>- Inst. Synth.</td><td>85.4</td><td>20.5</td><td>84.0</td><td>18.6</td><td>69.6</td><td>30.1</td><td>54.3(↓ 11.5)</td></tr><tr><td>- GPT-4o</td><td>58.6</td><td>37.1</td><td>52.1</td><td>37.9</td><td>46.1</td><td>44.2</td><td>46.9(↓ 18.9)</td></tr></table>
+
+Table 4: Data curation ablation experiment on ScreenSpot. Different 500k instruction data is sampled to separately fine-tune and evaluate the same model.
+
+![](images/f76c98abf94c8259b7aceb7df5eaac863fb9dc4e4a09286e11814868aa968e42.jpg)  
+Figure 4: Accuracy comparison by the proportion of element size to image size, evaluated on UI-I2E-Bench.
+
+tio between our UI-I2E-VLM-4B and UGround, as shown in Figure 4. Grounding accuracy drops as the element size ratio decreases, highlighting the importance of benchmarks that emphasize smaller elements and higher-resolution images. Benefiting from our training data and larger number of input image tokens, UI-I2E-VLM achieves better performance on small elements.
+
+In element type analysis of Table 3, we observe a significant performance gap in the Icon and Inputfield types compared to previous models, indicating that prior work has largely overlooked these long-tail categories. It supports our data synthesis pipeline to use a relatively balanced distribution of element types when curating the training dataset.
+
+Data curation ablation study. To fairly evaluate the quality of our curated dataset and analyze components in our data synthesis pipeline, we compare our synthesized data with public released data from OS-Atlas (Wu et al., 2024) and variations from our own synthesis pipeline. Since OS-Atlas only constructs data on the web at a large scale, we also curate data on web domain. The comparison includes: (1) OS-Atlas-Web, the synthesized data from OSAtlas dataset, derived from FineWeb (Penedo et al., 2024); (2) UI-E2I-Synth, 500k instructions sampled from the web portion of our final training dataset; (3) -Inst. Synth., a variant that skips
+
+![](images/174d7636426036592fa16eb28a5ee53a1ac8360a050d39fda4ecb76626580486.jpg)  
+Figure 5: Typical failure cases in UI-I2E-Bench. Blue box denotes the correct elements while red ones are wrong prediction from our model.
+
+<table><tr><td>Setting</td><td>OS</td><td>Office</td><td>Daily</td><td>Profess.</td><td>Workflow</td><td>Overall</td></tr><tr><td>Screenshot only*</td><td>8.3</td><td>2.6</td><td>6.5</td><td>0</td><td>2.1</td><td>3.6</td></tr><tr><td>+OS-Atlas</td><td>30.4</td><td>2.7</td><td>11.7</td><td>16.3</td><td>2.2</td><td>8.5</td></tr><tr><td>+UI-12E-VLM</td><td>26.1</td><td>5.2</td><td>23.4</td><td>18.8</td><td>2.8</td><td>12.0</td></tr></table>
+
+Table 5: Task success rate on OSWorld. *Report as the performance reproduced by us. In this experiment, the environment and settings are consistent.
+
+the instruction synthesis step and uses referring expressions as instructions; and (4) - GPT-4o, a configuration using only raw element attributes as instructions. We sample 500k instruction from above configurations and fine-tuned the same InternVL2-4B model separately on each dataset and evaluated the trained models on the ScreenSpot benchmark. As shown in Table 4, the results demonstrate that our dataset yields significant performance improvement compared to OS-Atlas-Web. While this advantage diminishes with OS-Atlas larger training data in Table 2, it also demonstrate the effectiveness of our curated data. Furthermore, the performance degradation observed when eliminating instruction synthesis and GPT-4o components provides evidence for our UI-E2I-Synth design.
+
+Failure case analysis. Figure 5 illustrates common errors of UI-I2E-VLM on UI-I2E-Bench, including: (1) failure to recognize icons without text due to limited knowledge, (2) incorrect positioning of elements within rows or columns, (3) misinterpretation of spatial relationships, (4) misunderstanding hierarchical relationships, and (5) misclassifying element types, such as confusing checkboxes with adjacent text.
+
+# 5.3 Agent Task Online Evaluation
+
+OSWorld. We evaluate our model on OSWorld (Xie et al., 2024), a live benchmark built
+
+in real computer environment for GUI agents. OSWorld consists of 369 open-ended computer tasks across various operating systems, including Ubuntu, Windows, and macOS. Following its original screenshot-only design, we use GPT-4o as the planner, providing only screenshots as the observation space. For every step, the planner is given the user task and the screenshot of current stage to predict the next action to take in next step. Then for the screenshot and the predicted action, we leverage VLMs to provide corresponding element coordinates to compare the effect of different VLMs. For screenshot only baseline, we use GPT-4o to directly generate element coordinates. The prompt used in evaluation is provided in Appendix F. The result comparison is shown in Table 5. The results show that UI-I2E-VLM helps to address GPT-4o's limitations in grounding and shows competitive performance, which prove our data synthesis pipeline can benefit the GUI agent development.
+
+# 6 Conclusion
+
+This work tries to identify and address critical challenges in GUI instruction grounding, such as element-to-screen ratio, unbalanced element types, and implicit instructions, through the introduction of a large-scale data synthesis pipeline, UI-E2I-Synth. Moreover, we propose a new GUI instruction grounding benchmark, UI-I2E-Bench, which overcomes the limitations of existing benchmarks by incorporating diverse annotation aspects. Our model, trained on the synthesized data, outperforms the previous state-of-the-art grounding models with less data and a smaller model size. We hope this work provides valuable insights and tools for the future development of GUI instruction grounding.
+
+# Limitations
+
+While the introduction of a large-scale data synthesis pipeline (UI-E2I-Synth) has significantly improved the generation of extensive instruction datasets, there is still potential for further scaling the data size and model size. Increasing the dataset size could enhance the robustness and generalizability of the model, addressing more diverse and complex GUI scenarios. Another limitation is this work only focuses on English instruction, we hope to further apply our data synthesis pipeline on other lanugages.
+
+# Ethical considerations
+
+Automated GUI agents can liberate human from repetitive tasks in digital devices but also could be used to launch robot attacks, such as brute force login attempts, automated spam, or denial-of-service attacks, by interacting with web interfaces at a speed and scale beyond human capability. They might be used to automate phishing attacks, where the agent mimics legitimate user interactions to steal sensitive information. It necessitates the maintaining transparency about the capabilities and limitations of GUI agents.
+
+# 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.  
+Anthropic. 2024. Claude 3 model card.  
+Andrea Burns, Deniz Arsan, Sanjna Agrawal, Ranjitha Kumar, Kate Saenko, and Bryan A. Plummer. 2022. A dataset for interactive vision language navigation with unknown command feasibility. In European Conference on Computer Vision (ECCV).  
+Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2023. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238.  
+Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. 2024a. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935.  
+Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. 2024b.
+
+Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935.  
+Common crawl. 2024. Common crawl - open repository of web crawl data.  
+Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuolin Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. 2024. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402.  
+Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. 2024. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36.  
+Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2023. Enhancing chat language models by scaling high-quality instructional conversations. arXiv preprint arXiv:2305.14233.  
+Jingwen Fu, Xiaoyi Zhang, Yuwang Wang, Wenjun Zeng, and Nanning Zheng. 2024. Understanding mobile gui: From pixel-words to screen-sentences. Neurocomputing, 601:128200.  
+Google. 2024. Profile your layout with hierarchy viewer.  
+Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. 2024. Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243.  
+Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. 2023. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856.  
+Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. 2024. On the effects of data scale on computer control agents. arXiv preprint arXiv:2406.03679.  
+Yang Li, Gang Li, Luheng He, Jingjie Zheng, Hong Li, and Zhiwei Guan. 2020. Widget captioning: Generating natural language description for mobile user interface elements. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5495-5510.  
+Kevin Qinghong Lin, Linjie Li, Difei Gao, Zhengyuan Yang, Zechen Bai, Weixian Lei, Lijuan Wang, and Mike Zheng Shou. Showui: One vision-language-action model for generalist gui agent. In NeurIPS 2024 Workshop on Open-World Agents.  
+Yadong Lu, Jianwei Yang, Yelong Shen, and Ahmed Awadallah. 2024. Omniparser for pure vision based gui agent. arXiv preprint arXiv:2408.00203.
+
+Microsoft. 2021. Ui automation overview - .net framework.  
+Mozilla. 2024. Document object model (dom) - web apis - mdn web docs.  
+OpenAI. 2025. Operator system card.  
+Guilherme Penedo, Hynek Kydlíček, Loubna Ben al-lal, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro Von Werra, and Thomas Wolf. 2024. The fineweb datasets: Decanting the web for the finest text data at scale. In *The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track*.  
+Peter Shaw, Mandar Joshi, James Cohan, Jonathan Berant, Panupong Pasupat, Hexiang Hu, Urvashi Khandelwal, Kenton Lee, and Kristina N Toutanova. 2023. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36:34354-34370.  
+Theodore Sunders, Shunyu Yao, Karthik Narasimhan, and Thomas Griffiths. Cognitive architectures for language agents. Transactions on Machine Learning Research.  
+Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. Preprint, arXiv:2409.12191.  
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560.  
+WebAIM. 2024. Webaim: The webaim million - the 2024 report on the accessibility of the top 1,000,000 home pages. Accessed on September 30, 2024.  
+Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, et al. 2024. Osatlas: A foundation action model for generalist gui agents. arXiv preprint arXiv:2410.23218.  
+Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences, 68(2):121101.  
+Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. 2024. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972.
+
+Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. Wizardlm: Empowering large language models to follow complex instructions. arXiv preprint arXiv:2304.12244.  
+Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. 2023. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. Preprint, arXiv:2310.11441.  
+John Yang, Carlos E Jimenez, Alexander Wettig, Kiliarian Lieret, Shunyu Yao, Karthik R Narasimhan, and Ofir Press. 2024. Swe-agent: Agent-computer interfaces enable automated software engineering. In *The Thirty-eighth Annual Conference on Neural Information Processing Systems*.  
+Zhizheng Zhang, Wenxuan Xie, Xiaoyi Zhang, and Yan Lu. 2023a. Reinforced ui instruction grounding: Towards a generic ui task automation api. arXiv preprint arXiv:2310.04716.  
+Zhizheng Zhang, Xiaoyi Zhang, Wenxuan Xie, and Yan Lu. 2023b. Responsible task automation: Empowering large language models as responsible task automators. arXiv preprint arXiv:2306.01242.  
+Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. 2024. Gpt-4v(ision) is a generalist web agent, if grounded. In *Forty-first International Conference on Machine Learning*.  
+Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Yonatan Bisk, Daniel Fried, Uri Alon, et al. 2023. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854.
+
+# Appendix
+
+The writing of this paper was limitedly assisted by AI, with refinement focused on language and style, and no original ideas generated by the AI.
+
+# A Data Statistics
+
+We present a detailed summary of the statistics for the complete training dataset we used in Table 6. All the external dataset used are publicly available with Apache 2.0 License.
+
+<table><tr><td>Dataset</td><td>Platform</td><td>#Screenshot</td><td>#Instructions</td></tr><tr><td>UI-E2I-Synth-Web</td><td>Web</td><td>1,536,200</td><td>9,097,736</td></tr><tr><td>UI-E2I-Synth-Desktop</td><td>Desktop</td><td>14,087</td><td>334,397</td></tr><tr><td>UI-E2I-Synth-AndroidControl</td><td>Mobile</td><td>40,199</td><td>109,126</td></tr><tr><td>MOTIF(Burns et al., 2022)</td><td>Mobile</td><td>30,699</td><td>320,219</td></tr><tr><td>WidgetCaption(Li et al., 2020)</td><td>Mobile</td><td>14,409</td><td>38,103</td></tr><tr><td>Total</td><td>-</td><td>1,635,594</td><td>9,899,581</td></tr></table>
+
+Table 6: Grounding training datasets statistics.
+
+We further analyzed the statistics of the synthetic data. In our synthetic data, the proportion of nontext elements is  $23.0\%$ . For comparison, we randomly sampled 100 elements from the SeeClick web data, manually labeled the non-text elements, and calculated their proportion. In SeeClick, nontext elements only account for  $8.7\%$ . Our dataset exhibits a more balanced distribution across different element types.
+
+Additionally, we randomly sampled 1,000 elements from landscape screenshots in both datasets and calculated their element-to-screen ratio. The distribution is shown in Table 7. Compared to SeeClick, we constructed our dataset with more data for smaller and more challenging targets, in order to optimize the model's performance in high-resolution real-world scenarios.
+
+<table><tr><td>Element-to-screen Ratio Range</td><td>Proportion (%) (SeeClick)</td><td>Proportion (%) (Ours)</td></tr><tr><td>0.00 - 0.02</td><td>11.49</td><td>36.92</td></tr><tr><td>0.02 - 0.04</td><td>43.30</td><td>40.43</td></tr><tr><td>0.04 - 1.00</td><td>45.21</td><td>22.65</td></tr></table>
+
+Table 7: Element-to-screen ratio comparison between SeeClick web dataset and our synthetic web dataset.
+
+# B Training Details
+
+UI-I2E-VLM-7B. To maintain consistency with the Qwen2-VL training data, we converted the format of the bounding boxes to  $< |box\_start|>(x1,y1)$ ,  $(x2,y2)< |box\_end|>$  where  $(x1,y1)$  and  $(x2,y2)$  represent the coordinates of the upper left and bottom right corners of the box, normalized within the range [0, 1000).  $< |box\_start|>$  and  $< |box\_end|>$  are treated as special tokens. We pack the samples corresponding to the same image into a single conversation, with no more than 15 samples per conversation. To achieve better performance on high-resolution images, we set the maxPixels to  $1500*1500$ . We perform full-parameter fine-tuning on our dataset. The whole training process costs around 60 hours on 16 A100 GPU cards.
+
+UI-I2E-VLM-4B. Following the pre-training setting, we convert bounding boxes into the format  $\langle \text{box} \rangle[[\text{x1}, \text{y1}, \text{x2}, \text{y2}]]\langle \text{/box} \rangle$ , where  $[\text{x1}, \text{y1}, \text{x2}, \text{y2}]$  are the original coordinates on the image without normalization.  $\langle \text{box} \rangle$  and  $\langle \text{/box} \rangle$  are treated as special tokens. Through experiments we found that packing the data caused a performance degra
+
+dation of InternVL2-4B, so we did not perform packing.
+
+We apply the 2-D bounding box tile tag proposed in (Dai et al., 2024) to improve the model's visual ability. Since InternVL-2 leverages Dynamic Aspect Ratio to handle images of varying resolutions, during both training and inference, we transform the original coordinates to align with the resized image size within the model. To accommodate the high-resolution nature of the UI interface, we set patch_num to 12, which means that each image is divided into 12 tiles of 448x448 pixels.
+
+We perform full-parameter fine-tuning on the language model on our dataset. The whole training process costs around 90 hours on 64 A100 GPU cards.
+
+C Details of UI-I2E-Bench Annotation  
+
+<table><tr><td>Error Category</td><td>Severity</td><td>Proportion (%)</td></tr><tr><td rowspan="2">Element Box</td><td>Serious</td><td>9.7</td></tr><tr><td>Slight</td><td>24.1</td></tr><tr><td rowspan="2">Element Instruction</td><td>Serious</td><td>17.3</td></tr><tr><td>Slight</td><td>7.9</td></tr></table>
+
+Table 8: Distribution of error types of UI-E2I-Synth generated data in UI-12E-Bench
+
+To reduce the workload of manual annotation, we adopt our UI-E2I-Synth pipeline on multiple platforms to construct the UI-I2E-Bench. We do balanced sampling for every element type, resulting in 1,987 elements. Then we manually annotated each data, including box quality, instruction quality and instruction type. For box quality, we consider bounding boxes that do not enclose any interactive UI elements as serious errors. Boxes that contain the correct element but are slightly misaligned or have inaccurate boundaries are labeled as slight errors. We retained data with slight errors in our benchmark and manually adjusted the boxes. For instruction quality, we consider instructions that are completely unrelated to any elements to be serious errors. Instructions that are unclear or ambiguous are labeled as slight errors. Similarly, we modified the slightly erroneous instructions and retained them. In addition, we annotated the instruction type for each data. Instructions that directly refer to an element are labeled as explicit, while those describe the element indirectly by position, appearance, or other references are implicit.
+
+After manual annotation, we removed data with serious errors, resulting in a benchmark containing
+
+![](images/45b666357fdf812892ed43bf8582e3b419780049e77dfa079c771f95900533d8.jpg)  
+Figure 6: Performance of model trained on 500k UI elements with different proportions of non-text elements in training dataset. The Average represents accuracy on the entire ScreenSpot, while the Non-text represents accuracy specifically on icons⁺/icons⁻/icons⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁺⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁻⁴⁵
+
+1,477 element-instruction pairs. The distribution of error types of data are shown in Table 8. We present some examples of our benchmark in Figure 7. All annotators are authors of this work and they all are Asian using English.
+
+# D Ablation on Non-Text Element Proportion
+
+In unfiltered data, the quantity of non-text elements is usually smaller than that of text elements. However, identifying non-text elements is much more challenging. When computational resources are limited, adjusting the proportion of non-text elements in the total dataset can help to improve model performance. We investigate the impact of the proportion of non-text elements on accuracy. We use InternVL2-4B as the base model, and sample 500K elements from 250K images for each proportion. The results are shown in Figure 6. Within a certain range, increasing the proportion of non-text elements can effectively improve the accuracy of non-text elements.
+
+# E Related literature discussion
+
+GUI agent is a multimodal agent designed for GUI scenarios, which requires multi-step planning and reasoning ability as general language agents (Sumers et al.; Yang et al., 2024; Xi et al., 2025) do. However, due to the incompleteness and opacity of UI metadata, the information within a GUI cannot be fully represented through pure language. It is necessary to leverage the multimodal perception capabilities of models to understand the current task state via GUI screenshots. During action execution, multimodal capabilities are also required to extract the coordinates of interactive elements from the current GUI screen
+
+shot, with GUI grounding playing a crucial role. Nevertheless, current general-purpose VLM, especially open-sourced leading MLLMs such as InternVL2 (Chen et al., 2023) and Qwen2-VL (Wang et al., 2024), exhibit relatively limited perception and planning capabilities for GUI tasks. Hence, the collection of corresponding training data is essential and warrants community attention.
+
+# F Prompt templates
+
+We display the prompts for the instruction synthesis in UI-E2I-Synth, prompts for OmniParser evaluation and prompts given to GPT-4o in OSWorld evaluation as follows.
+
+# G Insights on failure cases in UI-I2E-Bench
+
+Here we provide preliminary analysis about the typical failure cases in UI-I2E-Bench separately:
+
+Counting error and spatial relationship misunderstanding. The spatial understanding has always been two of the challenges faced by MLLMs. Due to the bias in MLLM pre-training data, models are typically better at detecting whether an object exists in the image but struggle with identifying its quantity and location. We believe this issue can be mitigated by adding data with instructions related to quantities and spatial relationships during training. In fact, through our synthetic data pipeline, our model has already made considerable progress compared to the baseline models in addressing these issues.
+
+Element type misclassification. In UI environment, the interactive region of different element types naturally varies. In general, large-scale training data contains elements with the same contents but different types, with some types of elements dominating in quantity(e.g., text). As a result, the model is more likely to assume that elements belong to these dominate types, leading to incorrect judgements on their interactive regions.
+
+Lack of external knowledge. While some common cases can be addressed by scaling up the range of training data, it is important to note that covering all icons from a variety of specialized software in the training data is challenging. We believe that for specific software, creating explanatory documentation for its icons and providing it as contextual reference for the model could be a potential solution in the future.
+
+![](images/710d957313aef66504d7399e373e1b591737a48e986acd8da9340bb35f4f6929.jpg)
+
+![](images/0e33e7bad38c6fe4530c65ab6e51a691ae0c7eb3cd567f2e3eda61833a1197f7.jpg)
+
+# Instruction: Select to change the website language.
+
+# Annotation:
+
+- Source: Web  
+- Element type:降落  
+- Instruction type: Implicit
+
+# Instruction: Filter to show most popular items.
+
+# Annotation:
+
+- Source: Web  
+- Element type: Toggle  
+- Instruction type: Implicit
+
+![](images/dffc6291ee2a3979ce8ac01a22b020089358ff9d7e2c9c996e05c206f31236b4.jpg)
+
+![](images/daa1e78d3186222303eb100727437b456c98d0740eb8545c7f47c6ab87d49905.jpg)
+
+# Instruction: Click the reply icon near the second email.
+
+# Annotation:
+
+- Source: Mobile  
+- Element type: Icon  
+- Instruction type: Implicit
+
+# Instruction: Define the max value of result as 999.
+
+# Annotation:
+
+- Source: Desktop  
+- Element type: Input  
+- Instruction type: Implicit
+
+# CNN's Kaitlan Collins said she was told the former president was "particularly" bothered by this quick reference.
+
+Aug 3,2023,11:29 PM EDT
+
+![](images/70db624766d3fdeb5f48941b8bf839d2e75f63ba7e2a30daed8cbe202bee2bf8.jpg)  
+By Ed Mazza
+
+![](images/b43c89b85a4c5d4810d2d87aeee413aa7e5a156ba75129b473071837bffb3161.jpg)  
+Figure 7: Examples in our benchmark UI-I2E-Bench. Red boxes denote the target elements. In the examples shown in this figure, we cropped only a portion of the original image.
+
+# Instruction: Share article on Twitter.
+
+# Annotation:
+
+- Source: Mobile  
+- Element type: Icon  
+- Instruction type: Explicit
+
+# Instruction: Click the button to access audio settings.
+
+# Annotation:
+
+- Source: Desktop  
+- Element type: Button  
+- Instruction type: Explicit
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Size</td><td rowspan="2">#Train</td><td colspan="2">Mobile</td><td colspan="2">Desktop</td><td colspan="2">Web</td><td rowspan="2">Avg.</td></tr><tr><td>Text</td><td>Icon/Widget</td><td>Text</td><td>Icon/Widget</td><td>Text</td><td>Icon/Widget</td></tr><tr><td>InternVL2-4B</td><td>4B</td><td>-</td><td>9.2</td><td>4.8</td><td>4.6</td><td>4.3</td><td>0.9</td><td>0.1</td><td>4.2</td></tr><tr><td>Qwen2-VL-7B</td><td>7B</td><td>-</td><td>61.3</td><td>39.3</td><td>52.0</td><td>45.0</td><td>33.0</td><td>21.8</td><td>42.6</td></tr><tr><td>OmniParser</td><td>-</td><td>-</td><td>87.4</td><td>67.9</td><td>93.2</td><td>60.9</td><td>77.8</td><td>48.3</td><td>73.9</td></tr><tr><td>CogAgent</td><td>17B</td><td>37M*</td><td>67.0</td><td>24.0</td><td>74.2</td><td>20.0</td><td>70.4</td><td>28.6</td><td>47.4</td></tr><tr><td>SeeClick</td><td>9.6B</td><td>1.0M</td><td>78.0</td><td>52.0</td><td>72.2</td><td>30.0</td><td>55.7</td><td>32.5</td><td>55.8</td></tr><tr><td>UGround</td><td>7B</td><td>9.7M</td><td>82.8</td><td>60.3</td><td>82.5</td><td>63.6</td><td>80.4</td><td>70.4</td><td>74.1</td></tr><tr><td>ShowUI</td><td>2B</td><td>2.7M</td><td>92.3</td><td>75.5</td><td>76.3</td><td>61.1</td><td>81.7</td><td>63.6</td><td>76.8</td></tr><tr><td>OS-Atlas-4B</td><td>4B</td><td>13.6M</td><td>85.7</td><td>58.5</td><td>72.2</td><td>45.7</td><td>82.6</td><td>63.1</td><td>70.1</td></tr><tr><td>OS-Atlas-7B</td><td>7B</td><td>13.6M</td><td>93.0</td><td>72.9</td><td>91.8</td><td>62.9</td><td>90.9</td><td>74.3</td><td>82.5</td></tr><tr><td>UI-I2E-VLM-4B</td><td>4B</td><td>9.9M</td><td>87.6</td><td>49.8</td><td>87.6</td><td>45.7</td><td>85.7</td><td>54.4</td><td>70.4</td></tr><tr><td>UI-I2E-VLM-7B</td><td>7B</td><td>9.9M</td><td>94.1</td><td>77.3</td><td>90.7</td><td>71.4</td><td>87.4</td><td>67.5</td><td>82.5</td></tr></table>
+
+Table 9: Full results on ScreenSpot. *Train data for CogAgent includes both pre-training and multi-task finetuning data.  
+
+<table><tr><td>Model</td><td>Development</td><td>Creative</td><td>CAD</td><td>Scientific</td><td>Office</td><td>Operating Systems</td><td>Overall Avg</td></tr><tr><td>InternVL2-4B</td><td>0.3</td><td>0.0</td><td>0.0</td><td>0.4</td><td>0.4</td><td>0.5</td><td>0.3</td></tr><tr><td>Qwen2-VL-7B</td><td>1.3</td><td>0.9</td><td>0.4</td><td>3.5</td><td>3.0</td><td>0.5</td><td>1.6</td></tr><tr><td>OmniParser+GPT-4o</td><td>13.7</td><td>1.5</td><td>7.7</td><td>9.4</td><td>14.3</td><td>4.6</td><td>8.3</td></tr><tr><td>SeeClick</td><td>0.3</td><td>0.6</td><td>1.9</td><td>2.0</td><td>0.9</td><td>1.5</td><td>1.1</td></tr><tr><td>UGround</td><td>14.7</td><td>17.0</td><td>11.1</td><td>19.3</td><td>27.0</td><td>9.7</td><td>16.5</td></tr><tr><td>ShowUI</td><td>9.4</td><td>5.3</td><td>1.9</td><td>10.6</td><td>13.5</td><td>6.6</td><td>7.7</td></tr><tr><td>OS-Atlas-4B</td><td>3.7</td><td>2.3</td><td>1.5</td><td>7.5</td><td>4.8</td><td>3.1</td><td>3.7</td></tr><tr><td>OS-Atlas-7B</td><td>17.7</td><td>17.9</td><td>10.3</td><td>24.4</td><td>27.4</td><td>16.8</td><td>18.9</td></tr><tr><td>UI-I2E-VLM-4B</td><td>12.7</td><td>12.6</td><td>5.7</td><td>13.4</td><td>15.2</td><td>14.3</td><td>12.2</td></tr><tr><td>UI-I2E-VLM-7B</td><td>24.1</td><td>23.5</td><td>9.2</td><td>27.6</td><td>38.3</td><td>19.9</td><td>23.6</td></tr></table>
+
+Table 10: Full results on ScreenSpot-Pro(simple view).
+
+UI hierarchical misunderstanding. Most current grounding models directly output bounding boxes, lacking an analysis and reasoning progress regarding UI hierarchies. In fact, when dealing with grounding tasks that need page semantic understanding, the model sometimes needs to perform multi-step reasoning. In the example of Figure 5 Task 4, the model needs to first locate the login module and then identify the e-mail input field. A potential mitigation strategy is to use CoT reasoning.
+
+# H GPT-4o response example in Instruction Synthesis
+
+For the parameter-based instruction synthesis, we do not simply repeat the already generated action parameters and assemble them to instruction. Instead, we use them as reference to guide the final generation of first-person perspective instructions. We here provide examples of the action parameter and corresponding synthesized instructions sam
+
+pled from generated dataset. Rather than merely rigidly assembling the given parameters, the synthesized instructions generated from GPT-4o demonstrate naturalness and fluency, exhibiting newly generated content that diverges from the provided action parameters.
+
+# Prompts for instruction synthesis in UI-E2I-Synth
+
+# Step1:
+
+You are a screen UI expert. Here is a UI screenshot image with highlighted bounding boxes and corresponding labeled ID overlayed on bottom of them, and here is a list of corresponding UI element (icon/bottom/Inputfield) box description within these bounding boxes. You should first ensure you understand the screenshot's status and every annotated UI element bounding box on it. Then output the updated element list with below template as JSON format.
+
+NOTE: Ensure all referring expression should UNIQUELY correspond to the element when generating them. Here is the output template:
+
+```jsonl
+{ "elements": [ { "id": "<copy corresponding labeled ID on the bottom of element>", "shortDescription": "<copy from given input list>", "fullDescription": "<a comprehensive description that explains the function of the  $\leftrightarrow$  element and the expected outcome when interacting with it>", "explicitRefer": "<a referring expression that explicitly refers to the element, from a  $\leftrightarrow$  computer user's first perspective>", "implicitReferByElementFunction": "<a referring expression that does NOT explicitly  $\leftrightarrow$  refer to this element content or obvious visual feature, but implicitly refers to  $\leftrightarrow$  it by its function in the whole page or expected outcome after interacting with  $\leftrightarrow$  it>", "implicitReferByNearElement": "<a referring expression that does NOT explicitly refer  $\leftrightarrow$  to this element content or obvious visual feature, but implicitly refers to it by  $\leftrightarrow$  its relationship with near elements or emphasizes it from similar elements by  $\leftrightarrow$  spatial order>" } ]
+```
+
+# Step2:
+
+You are a computer expert user. You will be given a UI element list from a UI screenshot which includes elements and their referring expressions as shown in the following input template. You should simulate a user using the UI screenshot, generate possible action type and action content, then generate instructions for every element according to the given element referring expressions, as shown in the following output template. Return as JSON format.
+
+# NOTE:
+
+1. For inputfield element type, the generated instruction should contain the possible input content from a computer user's first perspective. For example, "fill James as last name", "enter Boeing747 in search field", "select article in recent six months".  
+2. Ensure all instruction should UNIQUELY correspond to the element when generating them.
+
+Here is the input template:
+
+```jsonl
+{ "elements": [ { "id": "<copy corresponding labeled ID on the bottom of element>", "shortDescription": "<a short description about the element, including element type and  $\leftrightarrow$  element content>", "fullDescription": "<a comprehensive description that explains the function of the  $\leftrightarrow$  element and the expected outcome when interacting with it>", "explicitRefer": "<a referring expression that explicitly refers to the element, from a  $\leftrightarrow$  computer user's first perspective>", "implicitReferByElementFunction": "<a referring expression that does NOT explicitly  $\leftrightarrow$  refer to this element content or obvious visual feature, but implicitly refers to  $\leftrightarrow$  it by its function in the whole page or expected outcome after interacting with  $\leftrightarrow$  it>", "implicitReferByNearElement": "<a referring expression that does NOT explicitly refer  $\leftrightarrow$  to this element content or obvious visual feature, but implicitly refers to it by  $\leftrightarrow$  its relationship with near elements or emphasizes it from similar elements by  $\leftrightarrow$  spatial order>" }   
+}   
+Here is the output template:   
+{ "elements": [
+```
+
+(cont.) Prompts for instruction synthesis in UI-E2I-Synth  
+```jsonl
+{ "id": "<copy corresponding labeled ID on the bottom of element>", "shortDescription": "<copy from given input list>", "instructionArgs": { "actionType": "<choose appropriate action type>", "actionContentDescription": "<describe what is the possible action content about. if  $\leftrightarrow$  CLICK: leave empty string. if TYPE: specific input content which is appropriate  $\leftrightarrow$  for current UI screenshot. if Toggle, checkbox and Switch, ON or OFF>", "actionContent": "<According to `actionContentDescription` to fill the specific  $\leftrightarrow$  content, if CLICK: leave empty string>" }, "convertedUserInstructionByElementFunction": "<including above `actionContent` if not  $\leftrightarrow$  empty, convert above `implicitReferByElementFunction` into a computer user's first  $\leftrightarrow$  perspective instruction, concise and less than 15 words>", "convertedUserInstructionByNearElement": "<including above `actionContent` if not  $\leftrightarrow$  empty, convert above `implicitReferByNearElement` into a computer user's first  $\leftrightarrow$  perspective instruction, concise and less than 15 words>" } ]
+```
+
+# Prompts for OmniParser Evaluation
+
+Here is a UI screenshot image with bounding boxes and corresponding labeled ID overlayed on top of it, and here is a list of icon/text box description: { parsed_local_semantics}.
+
+Your task is {task}. Which bounding box label you should operate on? Give a brief analysis, then put your answer in the format of:
+
+Box with label ID: [xx]
+
+# Prompts for GPT-4o in OSWorld
+
+You are an agent which follow my instruction and perform desktop computer tasks as instructed.
+
+You have good knowledge of computer and good internet connection and assume your code will run on a computer for controlling the mouse and keyboard. For each step, you will get an observation of an image, which is the screenshot of the computer screen and you will predict the action of the computer based on the image.
+
+You are required to use 'pyautogui' to perform the action grounded to the observation, but DONOT use the 'pyautoguilocateCenterOnScreen' function to locate the element you want to operate with since we have no image of the element you want to operate with. DONOT USE 'pyautogui screenshot') to make screenshot.
+
+Return exactly ONE line of python code to perform the action each time. At each step, you MUST generate the corresponding instruction to the code before a # in a comment (example: # Click \"Yes, I trust the authors"\ button\n pyautogui.click(x=0, y=0, duration=1))
+
+You need to specify the coordinates of by yourself based on your observation of current observation, but you should be careful to ensure that the coordinates are correct. You ONLY need to return the code inside a code block, like this:
+
+your code here
+
+Specially, it is also allowed to return the following special code: When you think you have to wait for some time, return "WAIT"; When you think the task can not be done, return "FAIL", don't easily say "FAIL", try your best to do the task; When you think the task is done, return "DONE".
+
+Here are some guidelines for you:
+
+1. Remember to generate the corresponding instruction to the code before a # in a comment.  
+2. If a click action is needed, use only the following functions: pyautogui.click, pyautogui.rightClick or pyautogui,doubleClick.  
+3. Return "Done" when you think the task is done. Return "Fail" when you think the task can not be done.  
+4. If there are many times in the history of the same repeated action, first try to change the expression of the instruction, and if it still fails, try something else.
+
+My computer's password is 'password', feel free to use it when you need sudo rights. First give the current screenshot and previous things we did a short reflection, then RETURN ME THE CODE OR SPECIAL CODE I ASKED FOR. NEVER EVER RETURN ME ANYTHING ELSE.
+
+Please generate the next move according to the UI screenshot, instruction and previous actions.
+
+# (cont.) Prompts for GPT-4o in OSWorld
+
+Instruction: {instruction}.  
+Previous actions: {previous_actions}
+
+# GPT-4o response example in UI-E2I-Synth
+
+```json
+{
+    "instructionArgs": {
+        "actionType": "TYPE",
+        "actionContent": "Gangsta-Groove"
+   },
+    "convertedUserInstructionByElementFunction": "Enter 'Gangsta-Groove' in the search bar",
+    "convertedUserInstructionByNearElement": "Type 'Gangsta-Groove' in the search field above  $\leftrightarrow$  the article",
+}
+{
+    "instructionArgs": {
+        "actionType": "CLICK",
+        "actionContent": ""
+   },
+    "convertedUserInstructionByElementFunction": "Review price based on net area",
+    "convertedUserInstructionByNearElement": "Check near rent/total price toggle."
+},
+{
+    "instructionArgs": {
+        "actionType": "SELECT",
+        "actionContent": "preferred language"
+   },
+    "convertedUserInstructionByElementFunction": "Select your preferred language",
+    "convertedUserInstructionByNearElement": "Choose a language next to Contact Us",
+}
+{
+    "instructionArgs": {
+        "actionType": "CLICK",
+        "actionContent": ""
+   },
+    "convertedUserInstructionByElementFunction": "Click to learn more about us",
+    "convertedUserInstructionByNearElement": "Click the second button on the menu"
+}
+{
+    "instructionArgs": {
+        "actionType": "CLICK",
+        "actionContent": ""
+   },
+    "convertedUserInstructionByElementFunction": "Return to the top of the page",
+    "convertedUserInstructionByNearElement": "Click the bottom right corner icon",
+}
+```

2025/UI-E2I-Synth_ Advancing GUI Grounding with Large-Scale Instruction Synthesis/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3d0a6e8fc1eda64c01cd91d474ccfe5d3abc568fda086744efab7e72f5176b1
+size 676037

2025/UQ-Merge_ Uncertainty Guided Multimodal Large Language Model Merging/f3c36716-306f-46b5-8771-6bf9a011aa75_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd0338b0a0ac1d3802870d2fc48ca19127a1948db31c6ccf08517579d152f961
+size 2067162

2025/UQ-Merge_ Uncertainty Guided Multimodal Large Language Model Merging/full.md

@@ -0,0 +1,352 @@
+# UQ-Merge: Uncertainty Guided Multimodal Large Language Model Merging
+
+Huaizhi  $\mathbf{Q}\mathbf{u}^{1}$ , Xinyu Zhao $^{1}$ , Jie Peng $^{1}$ , Kwonjoon Lee $^{2}$ , Behzad Dariush $^{2}$ , Tianlong Chen $^{1}$
+
+<sup>1</sup>University of North Carolina at Chapel Hill, <sup>2</sup>Honda Research Institute USA Correspondence: tianlong@cs.unc.edu
+
+# Abstract
+
+Multimodal Large Language Models (MLLMs) have gained increasing popularity as a promising framework for leveraging the strong language reasoning capabilities in the vision-language domain. Given a wide range of MLLMs, model merging potentially offers a cheap way to aggregate their diverse knowledge into a single MLLM. However, directly plug-in existing model merging approaches often leads to suboptimal performance due to (1) inclusion of harmful models that have over-confident predictions in the target task; (2) the lack of specialized designs for vision-language inputs. To tackle these pain points, we conduct pioneering investigations to dissect the merging procedures and propose an uncertainty-guided MLLM merging algorithm, i.e., UQ-Merge, which  $i$  ) identifies beneficial candidates for merging,  $ii$  ) determines the merging order and the number of helpful candidates, and  $iii$  ) performs appropriate merging. Within our framework, we consider uncertainty quantification on both text and vision inputs to examine the MLLM prediction confidence, and then decide whether and when a MLLM needs to be included. It is worth mentioning that our vision-language uncertainty quantification does not require access to sample labels, making it more practical in various scenarios. Extensive experiments consistently demonstrate the superior MLLM merging performance of UQ-Merge in both held-in and held-out vision-language benchmarks. For example, compared to existing state-of-the-art merging methods, UQ-Merge brings substantial performance improvements of up to  $44.3\%$  on average accuracy in 12 datasets. Codes are available at https: //github.com/UNITES-Lab/UQ-Merge.git.
+
+# 1 Introduction
+
+Multimodal Large Language Models (MLLMs) have achieved numerous successes in various visual-language tasks including visual reasoning (Yin et al., 2023), autonomous driving (Cui
+
+![](images/1b6003f465872ae697ce63b1880a3f861e923b680151a97111ffaede2a2f13d9.jpg)  
+Figure 1: The average accuracy of the best single model, merging all models, and merging UQ-Merge selected models. Held-in datasets refer to datasets used for UQ. Held-out datasets are benchmarks unused for UQ.
+
+et al., 2024), visual question answering (Zhang et al., 2024a), etc. A popular paradigm to reach impressive vision-language reasoning capabilities typically combines a LLM backbone with a pretrained vision encoder (Alayrac et al., 2022; Liu et al., 2024b,a; McKinzie et al., 2024; Tong et al., 2024; Xue et al., 2024). Fine-tuning of pretrained MLLMs has been explored in many vision-language domains like biomedicine answering (Li et al., 2024b) and text-rich image understanding (Zhang et al., 2023), pushing the need to incorporate knowledge from diverse domains. To achieve this, rather than collecting all datasets and spending massive computing costs to train a new model from scratch, model merging has been widely explored as a method to overcome high training costs and aggregate knowledge from different datasets, by leveraging existing models and merging them in a training-free manner. Existing studies have shown superior merging results across tasks, highlighting its advantage of efficiently integrating separate advancements (Ding et al., 2024; Goddard et al., 2024; Wan et al., 2024; Yang et al., 2024; Lu et al., 2024).
+
+However, the model merging in the MLLM domain remains less explored. To begin with, we apply a single-modal merging method (Figure 1 yellow) and it achieves stronger performance compared to the best single model (Figure 1 green). Meanwhile, we observe that by selectively merging MLLMs (Figure 1 orange), the performance of the merged model can be further improved.
+
+Despite the performance gain, applying single-modal merging methods on MLLM merging has limitations. Firstly, existing merging methods assume that all models are beneficial for merging performance. As pointed out by (Zhao et al., 2024) and observed in Figure 1, this assumption may not hold true in real-world scenarios where models to merge are trained on divergent datasets. Some models produce over-confident predictions on target tasks and merging them will result in a performance decrease of the merged model. Secondly, these merging methods are designed solely to focus on single-modal model merging. Given these limitations, an ideal MLLM merge mechanism should be selective and aware of multimodal inputs.
+
+To address these challenges, we propose UQ-Merge, an uncertainty guided MLLM merging algorithm that features vision-language optimized design to ameliorate performance degradation caused by merging over-confident models. Specifically, UQ-Merge 1 uses image-text perturbation-based uncertainty quantification (UQ) to evaluate models, and 2 sorts models by descending order of uncertainty to reduce the impact of over-confident models. 3 UQ-Merge incrementally enlarge the group of models to merge, and 4 return the merged model when uncertainty stops decreasing. Our contributions are summarized as follows:
+
+- Due to the inclusion of over-confident models and the lack of vision-language specific designs, directly applying single-modal merging methods results in suboptimal performance. To resolve these issues, we conduct pioneering work in the MLLM field.  
+- To investigate design factors influencing MLLM merging performance, we raise and answer research questions: What is a more effective metric for selecting helpful models? How to decide the merging order and select models? How to implement UQ for MLLM? And how to properly merge selected models?  
+- We propose a MLLM-tailored image-text
+
+perturbation-based UQ method and introduce UQ-Merge, an uncertainty guided MLLM merging method that identifies and excludes over-confident models.
+
+- Experiments demonstrate that UQ-Merge consistently outperforms single-modal merging methods. With the same number of models used for merging, UQ-Merge achieves an average accuracy improvement of  $2.62\%$  on held-in datasets and  $1.06\%$  on held-out datasets compared to existing merging methods. Furthermore, UQ-Merge can surpass single-modal merging methods that have access to more models, by  $0.54\%$  on held-in datasets and  $1.3\%$  on held-out datasets.
+
+# 2 Related Work
+
+Multimodal Large Language Models (MLLMs). Large Language Models (LLMs) have demonstrated strong reasoning and instruction-following capability (Zhao et al., 2023; Touvron et al., 2023a,b). In light of this, many works (Alayrac et al., 2022; Liu et al., 2024b,a; McKinzie et al., 2024; Tong et al., 2024; Xue et al., 2024) propose to further incorporate pre-trained vision backbones (Radford et al., 2021; Zhai et al., 2023) to enable visual perception capabilities in existing LLMs, producing Multimodal Large Language Models (MLLMs). The mainstream architecture of MLLMs consists of three components: a vision encoder to extract features from visual inputs, a modality adapter that projects the outputs of the vision encoder into the token embedding space of the LLM backbone, and an LLM that processes both image and text inputs to generate responses (Yin et al., 2023; Zhang et al., 2024a). Given individual MLLMs, it remains under-explored how to leverage these mdoels and aggregate their knowledge. Motivated by this, we propose UQ-Merge, a UQ-based MLLM merging method to incorporate models fine-tuned on different tasks.
+
+Model Merging. Model Merging (Ainsworth et al., 2023) combines multiple pre-trained or fine-tuned models into a unified, powerful model, leveraging the strengths of specialized models while maintaining versatility without requiring additional training. Early zero-shot merging methods, such as weight averaging and Linear Mode Connectivity (Nagarajan and Kolter, 2021; Wortman et al., 2022), laid the foundation for this ap
+
+proach. Task Arithmetic (Ilharco et al., 2023) manipulates task vectors for effective merging, while TIES (Yadav et al., 2023) addresses parameter interference through trimming and conflict resolution. DARE (Yu et al., 2024) selectively optimizes parameters to enhance merging without extra training, utilizing the geometric properties of weights (Shoemaker, 1985; Jang et al., 2024). In the latest works, DELLA merges models by pruning and re-scaling weights based on their magnitude (Deep et al., 2024), and Model Stock finds the optimal interpolation ratio between merging candidates, using a pre-trained model to identify a robust anchor point (Jang et al., 2024). In the multimodal domain, model merging has similarly proven its ability to transform modality-specific models into modality-agnostic models (Sung et al., 2023). These existing merging studies motivate us to explore model merging in the MLLM domain.
+
+Uncertainty Quantification (UQ). Uncertainty quantification in predictions from deep neural networks has been a longstanding and essential problem (Abdar et al., 2021; Gawlikowski et al., 2023). The sources of uncertainty can be categorized into data uncertainty (aleatoric uncertainty) and model uncertainty (epistemic uncertainty). In the LLM domain, various methods have been proposed (Gao et al., 2024; Ye et al., 2024; Liu et al., 2024c). In the MLLM domain, UQ is less explored. One recent work applies conformal prediction for UQ in MLLMs (Ye et al., 2024; Kostumov et al., 2024). However, the CP method requires labeled data to estimate the model's uncertainty, which is infeasible in many real-world applications due to the lack of ground truth. In (Daheim et al., 2023), the authors propose to utilize gradient-based UQ to mitigate mismatches of gradients when merging models trained on various tasks, but labels are still required to compute the gradients. To address this, we propose a vision-language perturbation-based UQ method for MLLM that does not require labels.
+
+# 3 Methodology
+
+# 3.1 Preliminaries
+
+The Architecture Overview of Multimodal Large Language Model. The definition of Multimodal Large Language Models (MLLMs) is LLM-based models with the ability to receive, reason, and output with multimodal information (Yin et al.,
+
+2023). Prior to MLLMs, many works were devoted to multimodality learning (Radford et al., 2021; Li et al., 2021; Wang et al., 2021). In this paper, we focus on MLLMs that process image-text inputs and use  $(x_v, x_t)$  to represent an input image  $x_v$  and text  $x_t$  pair to an MLLM  $\mathcal{M}(\cdot, \cdot)$ . The most common MLLM architecture for image-text inputs (Liu et al., 2024b,a; Chen et al., 2024; McKinzie et al., 2024; Tong et al., 2024) typically comprises a pre-trained vision encoder  $\mathcal{V}(\cdot)$ , an adapter  $\mathcal{A}(\cdot)$  and an LLM backbone  $\mathcal{F}(\cdot)$ . An overview of the model architecture is provided in Figure 2 (a). The text input  $x_t$  is split into textual tokens  $h_t$ . The vision encoder extracts visual features from the input image  $x_v$ , represented as visual tokens  $z_v = \mathcal{V}(x_v)$ , which are then mapped by the adapter into the embedding space of language tokens, yielding  $h_v = \mathcal{A}(z_v)$ . The LLM processes both visual tokens  $h_v$  and language tokens  $h_t$  to generate an output  $\mathcal{F}(h_v, h_t)$  to the textual query.
+
+Model Merging. The goal of model merging is to aggregate knowledge from two or more models with the same architecture into a unified model that retains the strengths and capabilities of the original models. Formally, given a set of models  $\{\mathcal{M}_1,\dots ,\mathcal{M}_n\}$ , model merging can be expressed as  $\mathcal{M}^{*} = \text{Merge}(\mathcal{M}_{1},\ldots ,\mathcal{M}_{n})$ , where  $\mathcal{M}^*$  represents the merged model and  $\text{Merge}(\cdot)$  is a merging method. In MLLM merging, as the vision encoders  $\mathcal{V}$  of models with the same architecture are usually initialized from the same pre-trained model and kept frozen during the pre-training and finetuning process (Liu et al., 2024b,a; Lin et al., 2024; Xue et al., 2024), their weights are identical and do not require merging. For this reason,  $\text{Merge}(\cdot)$  only considers the adapter  $\mathcal{A}$  and the LLM backbone  $\mathcal{F}$  when applied on MLLMs.
+
+# 3.2 UQ-Merge: UQ Guided MLLM Merging
+
+To overcome the aforementioned challenges of over-confident merging candidate models and the lack of vision-language (VL) oriented merging method, we propose UQ-Merge, which consists of a vision-language perturbation-based MLLM uncertainty quantification (Section 3.3) to evaluate models, and a merging algorithm based on the uncertainty of models (Section 3.4).
+
+The procedure of UQ-Merge is described in Figure 2 (c). First, UQ is applied to MLLMs to quantify their uncertainty, and models are sorted in descending order of uncertainty to later con
+
+![](images/9b9e9bb384071f19a89fbde9101708571f9a70ce4a725b378af4f19cb096517d.jpg)
+
+![](images/1dc0c16f8bacc21c2692c676c47b9b3c5a1bf53885ff1baa7a708cf2f23912e1.jpg)
+
+![](images/9579402c2c308dc9c5b03411b70b5cac26e954377248ca60929703f89cb76f79.jpg)  
+Figure 2: The overview of UQ-Merge. (a) illustrates the common architecture of MLLMs. In (b) it shows vanilla merging fails when no selection is considered, and in contrast performance gains by selectively merging. (c) shows the two steps in UQ-Merge: (1) UQ-Merge quantifies and sorts models by descending order of their uncertainty. (2) After sorting, each model is gradually included and merged, and stops when uncertainty stops decreasing.
+
+![](images/25957af6e0097349f946dc07ee4f261fdb158bd5307fee61f78638e766c13ca5.jpg)
+
+sider potentially over-confident models. Then, UQ-Merge incrementally merge the sorted models and record the uncertainty scores of the merged model at each step. Finally, UQ-Merge returns the merged model when uncertainty stops decreasing. Throughout the process, UQ-Merge adopts the same UQ function described in Section 3.3.
+
+# 3.3 Vision-Language Perturbation-based UQ
+
+UQ (Mehrtash et al., 2020; Guo et al., 2024; Gao et al., 2024) has demonstrated promising effectiveness in evaluating models without labels, which is highly practical in real-world scenarios. UQ provides a quantified score for a model, indicating its confidence level reliability and performance (Wang et al., 2022; Si et al., 2023). In light of this, we develop a vision-language perturbation-based UQ to evaluate MLLMs for model merging.
+
+Specifically, given input image-text pair  $x_{v}$  and  $x_{t}$ , our perturbation-based MLLM  $\mathsf{UQ}(\cdot ,\cdot ,\cdot)$  on a MLLM model  $\mathcal{M}$  is defined in Equation 1, where  $J$  is the number of perturbations,  $x_{v}$  and  $x_{t}$  are original image-text input,  $\mathcal{M}_e^j$  is the  $j_{\mathrm{th}}$  perturbed model,  $\mathcal{P}_v^j (\cdot)$  is the  $j_{\mathrm{th}}$  perturbation function for the image input,  $\mathcal{P}_t^j (\cdot)$  is the  $j_{\mathrm{th}}$  text perturbation function, and  $\mathcal{H}(\cdot)$  is the entropy function.
+
+The perturbed model  $\mathcal{M}_{\epsilon}^{j}$  is derived by adding Dropout (Srivastava et al., 2014) to the attention score. We implement  $\mathcal{P}_v^j (\cdot)$  as a composition of image transformation functions such as Shear, Translate, Rotate, Equalize, and Posterize (Cubuk et al., 2018; Hendrycks et al., 2020).  $\mathcal{P}_t^j (\cdot)$  is implemented by adding randomly selected prompts to the original text input. Following previous works (Ye et al., 2024; Kostumov et al., 2024), we employ the prompt to ask the model to answer with an option and extract the logits of option letters from the first newly generated token, and entropy is computed on the logits. The model uncertainty is the difference between total and data uncertainty, where the total uncertainty is the entropy of the average prediction, and the data uncertainty is the average entropy of each prediction. In the literature below, we use  $\overline{\mathrm{UQ}} (\mathcal{M})$  to represent the average uncertainty of  $\mathcal{M}$  over samples by using  $\mathrm{UQ}(\cdot ,\cdot ,\cdot)$ .
+
+# 3.4 UQ Guided Merging for Model Selection
+
+Given models to merge, we sort the models in descending order of uncertainty to reduce the impact of over-confident models for merging (Step (1) in Figure 2 (c)). Starting from the model with the lowest uncertainty, UQ-Merge gradually consid
+
+$$
+\underbrace {\operatorname {U Q} \left(\mathcal {M} , x _ {v} , x _ {t}\right)} _ {\text {M o d e l u n c e r t a i n t y}} \approx \underbrace {\mathcal {H} \left(\frac {1}{J} \sum_ {j = 1} ^ {J} \mathcal {M} _ {\epsilon} ^ {j} \left(\mathcal {P} _ {v} ^ {j} \left(x _ {v}\right) , \mathcal {P} _ {t} ^ {j} \left(x _ {t}\right)\right)\right)} _ {\text {T o t a l u n c e r t a i n t y}} - \underbrace {\frac {1}{J} \sum_ {j = 1} ^ {J} \mathcal {H} \left(\mathcal {M} _ {\epsilon} ^ {j} \left(\mathcal {P} _ {v} ^ {j} \left(x _ {v}\right) , \mathcal {P} _ {t} ^ {j} \left(x _ {t}\right)\right)\right)} _ {\text {D a t a u n c e r t a i n t y}} \tag {1}
+$$
+
+# Algorithm 1 UQ-Merge
+
+1: Input: Models  $\{\mathcal{M}_1,\dots ,\mathcal{M}_n\}$ $\overline{\mathrm{UQ}} (\cdot)$  Merging method Merge(·)  
+2: Output: Merged model  $\mathcal{M}^*$  
+3: Compute VL uncertainty  $\{u_1,\dots ,u_n\}$  for each model in  $\{\mathcal{M}_1,\ldots ,\mathcal{M}_n\}$  
+4:  $\{\mathcal{M}_1',\ldots ,\mathcal{M}_n'\} \gets \mathrm{Sort}\{\mathcal{M}_1,\ldots ,\mathcal{M}_n\}$  by descending order of  $\{u_1,\dots ,u_n\}$  
+5: Set merged model's uncertainty as  $u^{*} \gets \infty$  
+6: Initialize the merged model as  $\mathcal{M}^* \gets \mathcal{M}_1'$  
+7: for a model  $\mathcal{M}_t^\prime$  in  $\{\mathcal{M}_1^{\prime},\dots ,\mathcal{M}_n^{\prime}\}$  do  
+8:  $\mathcal{M}_{\mathrm{merged}}^t\gets \mathrm{Merge}(\mathcal{M}_1',\ldots ,\mathcal{M}_t')$  
+9:  $u_{\mathrm{merged}}^t \gets \overline{\mathrm{UQ}}(\mathcal{M}_{\mathrm{merged}}^t)$  
+10: if  $u_{\mathrm{merged}}^t < u^*$  then  
+11:  $\mathcal{M}^{*}\gets \mathcal{M}_{\mathrm{merged}}^{t};u^{*}\gets u_{\mathrm{merged}}^{t}$  
+12: else  
+13: return  $\mathcal{M}^*$  
+14: end if  
+15: end for  
+16: return  $\mathcal{M}^*$
+
+ers each model. At each step, one model  $\mathcal{M}_i$  is added to the merging group, and merging method Merge(·) is employed to produce a merged model  $\mathcal{M}_{\mathrm{merged}}$ , and  $\overline{\mathrm{UQ}} (\cdot)$  is applied to quantify its uncertainty. In our practice of UQ-Merge, Merge(·) is implemented as linearly averaging the weights of all models. UQ-Merge allows different choices of merging functions, but as will be shown in Table 6, linear merging is simple and brings strong performance. UQ-Merge returns when the uncertainty of the merged model stops decreasing (Step (2) in Figure 2 (c)). As the merged model aggregates knowledge from diverse domains, we view low uncertainty after merging as a signal of strong capability on tasks and select the model.
+
+# 4 Experiments
+
+# 4.1 Implementation Details
+
+Model Preparation. In our experiments, we begin with the pre-trained LLaVA-v1.5-7B model (Liu et al., 2024a), which utilizes Vicuna1.5-7B (Chiang et al., 2023) as the LLM backbone  $\mathcal{F}$ , a CLIP-ViT-L-336px (Radford et al., 2021) as the vision encoder  $\mathcal{V}$ , and a two-layer MLP with a hidden dimension of 4096 as the modality adapter  $\mathcal{A}$ . The model is fine-tuned with instruction-tuning datasets that focus on diverse vision-language capabilities to create the models for merging. More details are provided in Supplementary B.5.
+
+Single-Modal Baselines. For sufficient comparison with our method that uses UQ to guide MLLM merging, we compare UQ-Merge against various single-modal merging methods and test their performance in multimodal scenarios. Specifically, we consider DARE (Yu et al., 2024), DELLA (Deep et al., 2024), Linear (Wortsman et al., 2022), TIES (Yadav et al., 2023), Task Arithmetic (Ilharco et al., 2023), and Model Stock (Jang et al., 2024). Due to the lack of model selection capability, we compare these methods in  $\bullet$  average performance in random selections and  $\bullet$  merge all models. Although baseline methods were originally designed for single-modal merging, they are capable of merging models that have the same architecture. We consider adapter and LLM backbone when using baselines, as a naive extension of these methods.
+
+Datasets for UQ and Evaluation. We select vision-language classification datasets as our benchmarks for  $\overline{\mathsf{UQ}} (\cdot)$ . Specifically, five datasets across five domains are considered: MMBench (reasoning/perception, Liu et al., 2023a), OODCVVQA (out-of-distribution robustness, Zhao et al., 2022), ScienceQA (world knowledge, Lu et al., 2022), SEEDBench (spatial and temporal understanding, Li et al., 2023a), and AI2D (diagrams, Kembhavi et al., 2016). In line with (Ye et al., 2024; Kostumov et al., 2024), we reformat the answers and add two additional choices "I don't know" and "None of the above". Since our vision-language perturbation-based UQ does not require labels, we treat these datasets as held-in datasets and also use them for the evaluation of merged models' performance in vision-language classification format tasks. We also test our method on held-out vision-language generation datasets.
+
+# 4.2 UQ-Merge Is Effective for Removing Harmful Models
+
+In this section, we compare our UQ-Merge against various single-modal merging methods on held-in and held-out datasets to show the effectiveness of UQ-Merge in excluding harmful models. For baseline merging methods, we evaluate them by measuring the average performance of their merged models. For all baseline methods, each time the merged model is produced by merging a random model selection from all models we fine-tuned, and the number of models selected each time is the same as the selection of our method UQ-Merge. Evaluation results are summarized in Table 1, and the perfor
+
+Table 1: The comparison between UQ-Merge and single-modal merging methods. Baseline methods merge randomly selected the same number of models to UQ-Merge. Average and standard error of the accuracy of baselines across selections are reported. Results are measured with 3 selections. The best and second-best performances are highlighted in bold and underline, respectively.  
+
+<table><tr><td colspan="7">Vision-Language Classification Datasets</td></tr><tr><td rowspan="2">Merging Methods</td><td>Average</td><td>AI2D</td><td>ScienceQA</td><td>SeedBench</td><td>MMBench</td><td>OOD-CV</td></tr><tr><td>13.75 ± 3.09</td><td>8.01 ± 6.94</td><td>2.72 ± 2.60</td><td>5.73 ± 4.96</td><td>26.25 ± 0.75</td><td>26.04 ± 0.48</td></tr><tr><td>DELLA</td><td>53.15 ± 5.72</td><td>43.23 ± 7.63</td><td>57.97 ± 8.06</td><td>55.13 ± 7.55</td><td>70.26 ± 1.31</td><td>39.14 ± 4.95</td></tr><tr><td>Linear</td><td>55.43 ± 3.88</td><td>49.41 ± 3.64</td><td>64.15 ± 6.43</td><td>57.16 ± 5.07</td><td>69.82 ± 0.62</td><td>36.60 ± 3.70</td></tr><tr><td>TIES</td><td>51.10 ± 7.23</td><td>43.58 ± 7.81</td><td>52.40 ± 6.88</td><td>48.58 ± 14.28</td><td>70.02 ± 1.39</td><td>40.95 ± 6.07</td></tr><tr><td>Task Arithmetic</td><td>21.73 ± 7.27</td><td>18.22 ± 11.31</td><td>17.17 ± 19.61</td><td>20.35 ± 13.45</td><td>27.60 ± 1.14</td><td>25.32 ± 1.09</td></tr><tr><td>Model Stock</td><td>49.99 ± 0.58</td><td>43.81 ± 0.92</td><td>65.07 ± 0.06</td><td>49.11 ± 0.84</td><td>64.10 ± 0.82</td><td>27.86 ± 0.52</td></tr><tr><td>Ours</td><td>58.05</td><td>51.75</td><td>68.07</td><td>60.56</td><td>70.35</td><td>39.52</td></tr></table>
+
+<table><tr><td rowspan="2">Merging Methods</td><td colspan="8">Vision-Language Generation Datasets</td></tr><tr><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td></tr><tr><td>DARE</td><td>12.83 ± 8.01</td><td>0.37 ± 0.35</td><td>5.48 ± 4.76</td><td>10.91 ± 9.82</td><td>26.22 ± 0.80</td><td>2.43 ± 2.11</td><td>19.77 ± 17.35</td><td>24.60 ± 21.64</td></tr><tr><td>DELLA</td><td>47.94 ± 3.75</td><td>40.53 ± 8.05</td><td>41.17 ± 1.77</td><td>43.96 ± 3.26</td><td>32.78 ± 1.31</td><td>41.92 ± 5.66</td><td>65.33 ± 3.46</td><td>69.87 ± 18.26</td></tr><tr><td>Linear</td><td>54.13 ± 0.72</td><td>44.61 ± 5.90</td><td>42.74 ± 1.83</td><td>49.83 ± 0.92</td><td>33.71 ± 0.95</td><td>52.80 ± 4.44</td><td>69.52 ± 1.77</td><td>85.71 ± 0.48</td></tr><tr><td>TIES</td><td>54.95 ± 1.18</td><td>50.16 ± 5.58</td><td>44.35 ± 0.35</td><td>52.87 ± 1.61</td><td>32.48 ± 1.11</td><td>46.66 ± 3.37</td><td>71.34 ± 1.57</td><td>86.81 ± 0.23</td></tr><tr><td>Task Arithmetic</td><td>18.06 ± 12.38</td><td>2.73 ± 2.36</td><td>13.61 ± 11.02</td><td>16.68 ± 15.07</td><td>26.22 ± 1.74</td><td>3.30 ± 2.65</td><td>28.67 ± 24.04</td><td>35.23 ± 31.08</td></tr><tr><td>Model Stock</td><td>37.79 ± 2.26</td><td>2.27 ± 0.96</td><td>34.36 ± 2.60</td><td>38.29 ± 1.26</td><td>32.59 ± 0.23</td><td>54.24 ± 0.20</td><td>59.19 ± 2.08</td><td>43.57 ± 11.01</td></tr><tr><td>Ours</td><td>56.01</td><td>50.39</td><td>43.49</td><td>50.75</td><td>35.22</td><td>54.13</td><td>71.77</td><td>86.33</td></tr></table>
+
+mance of each selection is in Table 10 and Table 11. From the results, the following observations can be drawn: Our UQ-Merge demonstrates superior performance compared to all other merging methods. Specifically, UQ-Merge achieves  $2.62\% \sim$ $44.3\%$  and  $1.06\% \sim 43.18\%$  improvement on average accuracy of held-in and held-out datasets. This validates the effectiveness of UQ-Merge in model selection to exclude over-confident models. On the held-in dataset, which is used for uncertainty quantification, UQ-Merge obtains a more significant performance gain compared to held-out datasets, with 4 out of 5 highest accuracy. This justifies our practice of using UQ to perform model selection, as in real-world applications labels are usually unavailable and UQ only relies on input to evaluate a model, and UQ-guided model selection can effectively improve the performance on these applications and even generalize to held-out datasets. The performance of single-modal merging methods varies a lot in MLLM merging. The gap of average accuracy for baselines is  $41.68\%$  and  $42.12\%$  on held-in and held-out datasets, respectively. These large gaps show the various effectiveness of state-of-the-art single-modal merging methods when the setting is shifted to the MLLM merging.
+
+# 4.3 UQ Guided Model Merging Surpasses Existing Merging Methods
+
+We further compare UQ-Merge in a more challenging setting, where baseline merging methods have an "unfair" advantage to access all the knowl
+
+edge within models and merge. Experiment results show that 1 Compared to baseline methods, UQ-Merge still achieves the best average accuracy on both held-in and held-out datasets, surpassing these single-modal methods by  $0.54\% \sim 51.69\%$  and  $1.3\% \sim 52.6\%$  respectively. 2 Compared to the average performance of merging randomly selected portions of all models, all merging methods except DARE and Task Arithmetic enjoy performance increase by adding more models. This suggests the benefit of incorporating more models from diverse tasks to build a stronger model and supports our claim that model merging is a cheap way to aggregate knowledge from different models. It is worth noting that existing single-modal merging methods have a certain ability to resolve potentially harmful models when merging, by adopting model weight level manipulation to resolve weight conflict and preserve knowledge from different tasks (Ilharco et al., 2023; Yadav et al., 2023; Yu et al., 2024). However, these methods are limited to single-modal model merging. Our UQ-Merge is orthogonal to these works, as we consider model level removal of harmful over-confident models in MLLM merging situations to improve the performance, and can benefit from the techniques to ameliorate weight conflict.
+
+# 4.4 Research Questions and Ablation Study
+
+In this section, we conduct an in-depth investigation of the designs adopted in UQ-Merge and how they contribute to improved performance. Specifi
+
+Table 2: The comparison between UQ-Merge and single-modal merging methods. Baselines merge all 10 models. The best and second-best performances are highlighted in bold and underline.  
+
+<table><tr><td rowspan="2" colspan="2">Merging Methods</td><td colspan="8">Vision-Language Classification Datasets</td></tr><tr><td>Average</td><td>AI2D</td><td>ScienceQA</td><td>SeedBench</td><td>MMBench</td><td>OOD-CV</td><td></td><td></td></tr><tr><td colspan="2">DARE</td><td>6.36</td><td>0.00</td><td>14.10</td><td>17.54</td><td>14.10</td><td>17.54</td><td></td><td></td></tr><tr><td colspan="2">DELLA</td><td>57.50</td><td>62.96</td><td>71.69</td><td>41.94</td><td>71.69</td><td>41.94</td><td></td><td></td></tr><tr><td colspan="2">Linear</td><td>56.90</td><td>67.39</td><td>69.98</td><td>36.93</td><td>69.98</td><td>36.93</td><td></td><td></td></tr><tr><td colspan="2">TIES</td><td>57.51</td><td>62.99</td><td>71.69</td><td>41.94</td><td>71.69</td><td>41.94</td><td></td><td></td></tr><tr><td colspan="2">Task Arithmetic</td><td>10.94</td><td>0.29</td><td>26.58</td><td>25.78</td><td>26.58</td><td>25.78</td><td></td><td></td></tr><tr><td colspan="2">Model Stock</td><td>51.55</td><td>65.21</td><td>65.74</td><td>29.47</td><td>65.74</td><td>29.47</td><td></td><td></td></tr><tr><td colspan="2">Ours</td><td>58.05</td><td>68.07</td><td>70.35</td><td>39.52</td><td>70.35</td><td>39.52</td><td></td><td></td></tr><tr><td rowspan="2">Merging Methods</td><td colspan="9">Vision-Language Generation Datasets</td></tr><tr><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td><td></td></tr><tr><td>DARE</td><td>3.41</td><td>0.02</td><td>0.23</td><td>0.00</td><td>23.56</td><td>0.00</td><td>0.03</td><td>0.00</td><td></td></tr><tr><td>DELLA</td><td>54.69</td><td>47.84</td><td>44.79</td><td>51.65</td><td>33.56</td><td>48.38</td><td>70.51</td><td>86.12</td><td></td></tr><tr><td>Linear</td><td>54.71</td><td>44.51</td><td>44.23</td><td>48.56</td><td>35.00</td><td>55.21</td><td>70.21</td><td>85.23</td><td></td></tr><tr><td>TIES</td><td>54.70</td><td>47.89</td><td>44.81</td><td>51.65</td><td>33.56</td><td>48.39</td><td>70.51</td><td>86.11</td><td></td></tr><tr><td>Task Arithmetic</td><td>3.56</td><td>0.00</td><td>0.04</td><td>0.00</td><td>24.89</td><td>0.01</td><td>0.01</td><td>0.00</td><td></td></tr><tr><td>Model Stock</td><td>45.05</td><td>8.09</td><td>38.24</td><td>42.49</td><td>32.78</td><td>56.19</td><td>63.63</td><td>73.93</td><td></td></tr><tr><td>Ours</td><td>56.01</td><td>50.39</td><td>43.49</td><td>50.75</td><td>35.22</td><td>54.13</td><td>71.77</td><td>86.33</td><td></td></tr></table>
+
+cally, we address the following: (1) Is UQ a more effective way to exclude harmful models, and how should uncertainty of models be used? (Section 4.4) (2) How to select models to merge after quantifying models' uncertainty? (Section 4.4) (3) How to design UQ? (Section 4.4) (4) After selection, how to merge models? (Section 4.4).
+
+RQ1: Is UQ more effective than accuracy? How to use uncertainty? A1: Yes; sort by decreasing uncertainty. In UQ-Merge, uncertainty is adopted to measure each model and exclude harmful models. An intuitive alternative to uncertainty is the accuracy of the model on validation datasets, with sorting done in either ascending or descending order. To address these research questions, we compare uncertainty and accuracy to determine which serves as better guidance. We replace uncertainty in UQ-Merge with accuracy on held-in datasets and test both kinds of guidance in ascending and descending order. Other components in UQ-Merge are kept untouched. We evaluate these modified designs on held-out datasets due to the usage of held-in datasets for testing accuracy. As shown in Tab 3, sorting by descending order achieves the best average performance compared to other options, confirming the effectiveness of our design. Compared to ascending uncertainty, descending order leads to a better performance, which justifies our aim to exclude over-confident models.
+
+RQ2: How to select models to merge? A2: When the uncertainty of the merged model stops decreasing. After sorting models by descending
+
+order of uncertainty, it remains unsure how to exclude harmful models and select beneficial ones. In UQ-Merge, this process is conducted by picking the merged model when uncertainty stops decreasing. To verify this, we evaluated the correlation between uncertainty, accuracy on validation benchmarks, and accuracy on held-out datasets. In our experiments, we used RealWorldQA (xAI, 2024), Seedbench 2 Plus (Li et al., 2024a), and OcrBench (Liu et al., 2024d), which focus on real-world QA, multi-disciplinary knowledge, and text recognition, respectively. As shown in Table 4, the point uncertainty rises and the highest validation accuracy align with the peak performance on held-out datasets. Our findings indicate that lower uncertainty corresponds to better performance. We attribute this to the enhanced capability of the merged model that makes it more robust to input perturbation and could generate consistent answers. When uncertainty further decreases, the model becomes over-confident, leading to the performance drop.
+
+RQ3: How to design perturbation? A3: Vision-Language input perturbation is crucial. In this research question, we aim to investigate how different perturbation designs would affect the merging performance of UQ-Merge. Specifically, we compare the input and model perturbation method adopted in UQ-Merge versus only using input perturbation, by using them as different UQ functions in our UQ-Merge framework and test the merged model. We implement input perturbation following the same design of UQ-Merge, by adding random image transformations and text prompts to the im
+
+Table 3: Comparison of uncertainty and accuracy for guidance. Each guidance is implemented with ascending and descending orders to sort models. Accuracy for merging guidance is tested on held-in datasets, and all results are reported on held-out datasets. The best and second-best performances are highlighted in bold and underline.  
+
+<table><tr><td rowspan="2">Guidance</td><td colspan="9">Performance with Different Guidance</td></tr><tr><td>Order</td><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td></tr><tr><td rowspan="2">Uncertainty</td><td>Ascending</td><td>54.30</td><td>45.64</td><td>44.61</td><td>45.31</td><td>34.56</td><td>54.52</td><td>70.05</td><td>85.42</td></tr><tr><td>Descending</td><td>56.01</td><td>50.39</td><td>43.49</td><td>50.75</td><td>35.22</td><td>54.13</td><td>71.77</td><td>86.33</td></tr><tr><td rowspan="2">Accuracy</td><td>Ascending</td><td>54.72</td><td>44.56</td><td>44.12</td><td>48.58</td><td>35.00</td><td>55.34</td><td>70.22</td><td>85.24</td></tr><tr><td>Descending</td><td>52.27</td><td>30.33</td><td>40.68</td><td>50.58</td><td>35.67</td><td>55.73</td><td>67.34</td><td>85.58</td></tr></table>
+
+Table 4: The correlation of uncertainty with accuracy on validation benchmarks and held-out benchmarks. The point uncertainty stops decreasing and the highest validation accuracy and held-out accuracy are marked in bold.  
+
+<table><tr><td># Models</td><td>1</td><td>2</td><td>3</td><td>4</td><td>5</td><td>6</td><td>7</td><td>8</td><td>9</td><td>10</td></tr><tr><td>Uncertainty</td><td>0.21197</td><td>0.13290</td><td>0.05609</td><td>0.04866</td><td>0.04531</td><td>0.04155</td><td>0.03950</td><td>0.03954</td><td>0.03740</td><td>0.03675</td></tr><tr><td>Validation Accuracy</td><td>18.75</td><td>18.48</td><td>36.32</td><td>37.30</td><td>38.84</td><td>39.20</td><td>39.91</td><td>39.15</td><td>39.28</td><td>39.60</td></tr><tr><td>Held-out Accuracy</td><td>47.09</td><td>51.22</td><td>52.47</td><td>54.15</td><td>54.13</td><td>54.54</td><td>56.01</td><td>55.56</td><td>54.15</td><td>54.14</td></tr></table>
+
+Table 5: Comparison of perturbation types. Results are average accuracy on datasets.  
+
+<table><tr><td></td><td>Held-in</td><td>Held-out</td></tr><tr><td>Input</td><td>56.82</td><td>56.00</td></tr><tr><td>Input &amp; Model</td><td>56.80</td><td>56.01</td></tr></table>
+
+Table 6: Comparison of merging methods on the same group of models. Results are average accuracy on held-in and held-out datasets.  
+
+<table><tr><td rowspan="2">Merging Methods</td><td colspan="2">Number of Models</td></tr><tr><td>7</td><td>10</td></tr><tr><td>DARE</td><td>22.30</td><td>4.64</td></tr><tr><td>DELLA</td><td>56.26</td><td>55.86</td></tr><tr><td>Linear</td><td>56.86</td><td>55.62</td></tr><tr><td>TIES</td><td>56.26</td><td>55.87</td></tr><tr><td>Task Arithmetic</td><td>22.21</td><td>6.64</td></tr><tr><td>Model Stock</td><td>46.82</td><td>47.76</td></tr></table>
+
+age and text branches respectively. As shown in Table 5, when only use input perturbation, the performance is slightly improved on held-in datasets. On held-out datasets, the performance is slightly worse for input perturbation only. We attribute this to the robust capability of LLM backbones and dynamic sparsity of LLM inference (Liu et al., 2023b), which makes model perturbation may not significantly affect the performance of the LLM.
+
+RQ4: What merging method to use given a group of MLLMs? A4: TIES, Linear or DELLA. Existing merging methods are designed to deal with single-modal merging, and it remains unclear how these merging methods perform for merging multimodal models. In this research ques
+
+tion, we explore the performance of these single-modal merging methods in the multimodal scenario by evaluating their performance on a given group of models. Specifically, we evaluate DARE, DELLA, Linear, TIES, Task Arithmetic, and Model Stock on held-in and held-out datasets and calculate the average performance on all the datasets. From results in Table 6 we observe that DELLA, Linear, and TIES perform better than other methods. In 10-model merging, all instruction-tuned models are merged. As shown in Table 6, given the same ten models, TIES achieve the best performance. When merging seven models with the highest uncertainty, we observe that  $\mathbf{1}$  The performance of all merging methods improved, demonstrating the benefit of model selection.  $\mathbf{2}$  linear merging achieves the best performance, which supports our choice in UQ-Merge that linearly merges models.
+
+# 5 Conclusion
+
+In this paper, we present a novel MLLM merging algorithm UQ-Merge to aggregate diverse knowledge of models into a single MLLM. We design a multimodal-tailored vision-language perturbation-based UQ and employ it to guide the merging process. As a result, UQ-Merge could identify beneficial models to merge and use the uncertainty value to decide the merging order and number of models to merge, and apply appropriate merging on selected models. Extensive experiments on datasets from diverse domains consistently demonstrate the effectiveness of model selection and significantly improved performance with our algorithm.
+
+# Acknowledgement
+
+This project is supported by the Honda Research Institute USA.
+
+# References
+
+AugMix — Torchvision main documentation. https://pytorch.org/vision/main/.  
+Moloud Abdar, Farhad Pourpanah, Sadiq Hussain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul Fieguth, Xiaochun Cao, Abbas Khosravi, U Rajendra Acharya, et al. 2021. A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information fusion, 76:243-297.  
+Samuel K. Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. 2023. Git re-basin: Merging models modulo permutation symmetries.  
+Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736.  
+Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. Preprint, arXiv:2312.14238.  
+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.  
+Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. 2018. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501.  
+Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. 2024. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979.  
+Nico Daheim, Thomas Möllenhoff, Edoardo Maria Ponti, Iryna Gurevych, and Mohammad Emtiyaz Khan. 2023. Model merging by uncertainty-based gradient matching. arXiv preprint arXiv:2310.12808.  
+Pala Tej Deep, Rishabh Bhardwaj, and Soujanya Poria. 2024. Della-merging: Reducing interference in model merging through magnitude-based sampling. ArXiv, abs/2406.11617.
+
+Ning Ding, Yulin Chen, Ganqu Cui, Xingtai Lv, Ruobing Xie, Bowen Zhou, Zhiyuan Liu, and Maosong Sun. 2024. Mastering text, code and math simultaneously via fusing highly specialized language models. arXiv preprint arXiv:2403.08281.  
+Xiang Gao, Jiaxin Zhang, Lalla Mouatadid, and Kamalika Das. 2024. Spuq: Perturbation-based uncertainty quantification for large language models. ArXiv, abs/2403.02509.  
+Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Hunt, Jianxiang Feng, Anna Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, et al. 2023. A survey of uncertainty in deep neural networks. Artificial Intelligence Review, 56(Suppl 1):1513-1589.  
+Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vlad Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz. 2024. Arcee's mergekit: A toolkit for merging large language models. arXiv preprint arXiv:2403.13257.  
+Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913.  
+Hongji Guo, Nakul Agarwal, Shao-Yuan Lo, Kwonjoon Lee, and Qiang Ji. 2024. Uncertainty-aware action decoupling transformer for action anticipation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18644-18654.  
+Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. 2018. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617.  
+Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. 2020. Augmix: A simple data processing method to improve robustness and uncertainty. Preprint, arXiv:1912.02781.  
+Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709.  
+Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. Editing models with task arithmetic.  
+Dong-Hwan Jang, Sangdoo Yun, and Dongyoon Han. 2024. Model stock: All we need is just a few finetuned models.
+
+Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. 2014. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798.  
+Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. 2016. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer.  
+Vasily Kostumov, Bulat Nutfullin, Oleg Pilipenko, and Eugene Ilyushin. 2024. Uncertainty-aware evaluation for vision-language models. Preprint, arXiv:2402.14418.  
+Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. 2017. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73.  
+Bohao Li, Yuying Ge, Yi Chen, Yixiao Ge, Ruimao Zhang, and Ying Shan. 2024a. Seed-bench-2-plus: Benchmarking multimodal large language models with text-rich visual comprehension. Preprint, arXiv:2404.16790.  
+Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. 2023a. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125.  
+Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. 2024b. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36.  
+Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. 2021. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705.  
+Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023b. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355.  
+Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. 2024. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024a. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306.
+
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024b. Visual instruction tuning. Advances in neural information processing systems, 36.  
+Linyu Liu, Yu Pan, Xiaocheng Li, and Guanting Chen. 2024c. Uncertainty estimation and quantification for llms: A simple supervised approach. arXiv preprint arXiv:2404.15993.  
+Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. 2023a. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281.  
+Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. 2024d. Ocrbench: On the hidden mystery ofOCR in large multimodal models. Preprint, arXiv:2305.07895.  
+Zichang Liu, Jue Wang, Tri Dao, Tianyi Zhou, Binhang Yuan, Zhao Song, Anshumali Shrivastava, Ce Zhang, Yuandong Tian, Christopher Re, et al. 2023b. Deja vu: Contextual sparsity for efficient llms at inference time. In International Conference on Machine Learning, pages 22137-22176. PMLR.  
+Jinliang Lu, Ziliang Pang, Min Xiao, Yaochen Zhu, Rui Xia, and Jiajun Zhang. 2024. Merge, ensemble, and cooperate! a survey on collaborative strategies in the era of large language models. Preprint, arXiv:2407.06089.  
+Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521.  
+Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. 2016. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20.  
+Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. 2019. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204.  
+Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. 2024. Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611.  
+Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. 2020. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878.
+
+Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. 2019. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE.  
+Vaishnavh Nagarajan and J. Zico Kolter. 2021. Uniform convergence may be unable to explain generalization in deep learning.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR.  
+Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. 2022. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146-162. Springer.  
+ShareGPT. 2023. [link].  
+Ken Shoemake. 1985. Animating rotation with quaternion curves. In Proceedings of the 12th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '85, page 245-254, New York, NY, USA. Association for Computing Machinery.  
+Chenglei Si, Zhe Gan, Zhengyuan Yang, Shuohang Wang, Jianfeng Wang, Jordan Boyd-Graber, and Lijuan Wang. 2023. Prompting gpt-3 to be reliable. Preprint, arXiv:2210.09150.  
+Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. 2020. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer.  
+Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. 2019. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326.  
+Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958.  
+Yi-Lin Sung, Linjie Li, Kevin Lin, Zhe Gan, Mohit Bansal, and Lijuan Wang. 2023. An empirical study of multimodal model merging. arXiv preprint arXiv:2304.14933.  
+Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. 2024. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860.
+
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023a. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Fanqi Wan, Xinting Huang, Deng Cai, Xiaojun Quan, Wei Bi, and Shuming Shi. 2024. Knowledge fusion of large language models. arXiv preprint arXiv:2401.10491.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.  
+Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. 2021. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904.  
+Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. 2022. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time.
+
+xAI. 2024. Grok-1.5 vision preview.
+
+Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. 2024. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872.  
+Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. 2023. Ties-merging: Resolving interference when merging models.  
+Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. 2024. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666.  
+Fanghua Ye, Mingming Yang, Jianhui Pang, Longyue Wang, Derek F. Wong, Emine Yilmaz, Shuming Shi, and Zhaopeng Tu. 2024. Benchmarking llms via uncertainty quantification. Preprint, arXiv:2401.12794.  
+Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. 2023. A survey on multimodal large language models. ArXiv, abs/2306.13549.
+
+Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models are super mario: Absorbing abilities from homologous models as a free lunch.  
+Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. 2024. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567.  
+Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986.  
+Duzhen Zhang, Yahan Yu, Chenxing Li, Jiahua Dong, Dan Su, Chenhui Chu, and Dong Yu. 2024a. Mllms: Recent advances in multimodal large language models. arXiv preprint arXiv:2401.13601.  
+Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuhan Zhang, Jingkang Yang, Chunyuan Li, et al. 2024b. Lmms-eval: Reality check on the evaluation of large multimodal models. arXiv preprint arXiv:2407.12772.  
+Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. 2023. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107.  
+Bingchen Zhao, Shaozuo Yu, Wufei Ma, Mingxin Yu, Shenxiao Mei, Angtian Wang, Ju He, Alan Yuille, and Adam Kortylewski. 2022. Ood-cv: A benchmark for robustness to out-of-distribution shifts of individual nuisances in natural images. In European conference on computer vision, pages 163-180. Springer.  
+Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223.  
+Xinyu Zhao, Guoheng Sun, Ruisi Cai, Yukun Zhou, Pingzhi Li, Peihao Wang, Bowen Tan, Yexiao He, Li Chen, Yi Liang, Beidi Chen, Binhang Yuan, Hongyi Wang, Ang Li, Zhangyang Wang, and Tianlong Chen. 2024. Model-GLUE: Democratized LLM scaling for a large model zoo in the wild. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track.
+
+# A Additional Experiment Results
+
+# A.1 Ablation Study about UQ in UQ-Merge
+
+In this section we provide more ablation studies about UQ in UQ-Merge. We conduct experiments with various ratios of dataset samples used for UQ, and compare with alternatives of the perturbation-based UQ method used in UQ-Merge. The results are shown in Table 7 and Table 8. Results in Table 7 show that UQ-Merge achieves stable performance across various ratios of samples used for UQ, demonstrating the data efficiency of UQ-Merge. We also test image-only perturbation, text-only perturbation and conformal prediction (CP) as alternative UQ methods in Table 8. Compared with alternatives, results in Table 8 prove the effectiveness of our vision-language perturbation-based UQ.
+
+# A.2 Evaluation of fine-tuned models
+
+In Table 9, we provide uncertainty quantification results of fine-tuned models on held-in datasets. We conduct evaluation three times and the final uncertainty is the average. As observed in Table 9, the uncertainty is stable and consistent, showcasing the effectiveness and stability of our vision-language perturbation-based UQ.
+
+# A.3 Evaluation of Baselines over Random Selections
+
+In Table 10 and Table 11, we provide performance of baseline single-modal merging methods on held-in and held-out datasets over model selections. As observed in tables, DELLA, Linear and TIES consistently outperform other merging methods with a small variance. The average and standard error are reported based on results above, and the error bar represents the  $95\%$  confidence interval.
+
+# A.4 Evaluation during Merging Steps
+
+In Table 12, Table 13 and Table 14, we present the evaluation results of Table 4 before averaging during merging steps.
+
+# B More Implementatiin Details
+
+# B.1 Text Perturbation
+
+Prompts used for perturbation of text inputs:
+
+- 'you are a helpful assistant',  
+- 'you are a question-answering assistant',  
+- 'you are a nice assistant',  
+- 'You are a helpful assistant',  
+- 'You are a question-answering assistant',
+
+- 'You are a nice assistant',  
+'You are a helpful assistant.'  
+- 'You are a question-answering assistant.'  
+'You are a nice assistant.'
+
+# B.2 Image Perturbation
+
+The image perturbation is implemented by utilizing the implementation of AugMix (Hendrycks et al., 2020) in torchvision (Aug), and all parameters are set to default.
+
+# B.3 Datasets for Fine-tuning Models
+
+To prepare fine-tuned models for merging, we choose visual reasoning datasets (GQA Hudson and Manning, 2019, RefCOCO Kazemzadeh et al., 2014; Mao et al., 2016); text-rich datasets (OCRVQA Mishra et al., 2019, TextCaps Sidorov et al., 2020); knowledge-based VQA datasets (OKVQA Marino et al., 2019, A-OKVQA Schwenk et al., 2022); GPT-generated datasets (LLaVA-Instruct Liu et al., 2024b, ShareGPT ShareGPT, 2023); and general VQA datasets (VQAv2 Goyal et al., 2017, VG Krishna et al., 2017). We follow the original train-test split of these datasets.
+
+# B.4 Datasets for Uncertainty Quantification
+
+We use the code from (Kostumov et al., 2024) to process MMBench (Liu et al., 2023a), OODCVVQA (Zhao et al., 2022), ScienceQA (Lu et al., 2022), SEEDBench (Li et al., 2023a), and AI2D (Kembhavi et al., 2016) for vision-language perturbation-based UQ.
+
+# B.5 Evaluation of Models
+
+We adopt LMMs-Eval (Zhang et al., 2024b) with all default parameters to conduct evaluation of models on all benchmarks except MMBench and OODCV-VQA, which are evaluated directly using our pre-processed datasets. To more comprehensively evaluate the merged models' performance, we choose seven vision-language generation tasks of six domains, including open real-world knowledge (OKVQA Marino et al., 2019, MMMU Yue et al., 2024), text understanding (TextVQA Singh et al., 2019), compositional questioning answering (GQA Hudson and Manning, 2019), low-quality image understanding (VizWiz Gurari et al., 2018), general visual QA (VQAv2 Goyal et al., 2017), and hallucination (POPE Li et al., 2023b) as the benchmarks.
+
+Table 7: Performance of the merged models from UQ-Mergewith different ratios of samples used for UQ.  
+
+<table><tr><td colspan="2">Sample Ratios</td><td>Average</td><td>AI2D</td><td>ScienceQA</td><td>SeedBench</td><td>MMBench</td><td>OOD-CV</td><td></td></tr><tr><td colspan="2">5% Samples</td><td>58.08</td><td>51.75</td><td>68.07</td><td>60.53</td><td>70.44</td><td>39.60</td><td></td></tr><tr><td colspan="2">10% Samples</td><td>58.00</td><td>51.70</td><td>67.95</td><td>60.39</td><td>70.35</td><td>39.60</td><td></td></tr><tr><td colspan="2">20% Samples</td><td>58.34</td><td>52.23</td><td>68.95</td><td>60.52</td><td>70.21</td><td>39.77</td><td></td></tr><tr><td colspan="2">30% Samples</td><td>57.96</td><td>51.42</td><td>67.85</td><td>60.66</td><td>70.30</td><td>39.56</td><td></td></tr><tr><td colspan="2">100% Samples</td><td>58.05</td><td>51.75</td><td>68.07</td><td>60.56</td><td>70.35</td><td>39.52</td><td></td></tr><tr><td>Sample Ratio</td><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td></tr><tr><td>5% Samples</td><td>56.01</td><td>50.40</td><td>43.49</td><td>50.61</td><td>35.67</td><td>53.91</td><td>71.69</td><td>86.33</td></tr><tr><td>10% Samples</td><td>55.36</td><td>50.71</td><td>41.74</td><td>49.86</td><td>34.67</td><td>55.64</td><td>68.66</td><td>86.26</td></tr><tr><td>20% Samples</td><td>55.30</td><td>44.17</td><td>43.57</td><td>50.92</td><td>36.44</td><td>53.86</td><td>71.69</td><td>86.48</td></tr><tr><td>30% Samples</td><td>56.00</td><td>50.37</td><td>43.58</td><td>50.80</td><td>35.67</td><td>53.73</td><td>71.67</td><td>86.18</td></tr><tr><td>100% Samples</td><td>56.01</td><td>50.39</td><td>43.49</td><td>50.75</td><td>35.22</td><td>54.13</td><td>71.77</td><td>86.33</td></tr></table>
+
+Table 8: Performance of the merged models with alternatives of vision-language perturbation-based UQ.  
+
+<table><tr><td colspan="2"></td><td>Average</td><td>AI2D</td><td>ScienceQA</td><td>SeedBench</td><td>MMBench</td><td>OOD-CV</td><td></td></tr><tr><td colspan="2">CP</td><td>56.86</td><td>50.42</td><td>67.19</td><td>59.83</td><td>69.88</td><td>36.98</td><td></td></tr><tr><td colspan="2">Image Only</td><td>57.11</td><td>50.58</td><td>67.51</td><td>60.40</td><td>69.94</td><td>37.14</td><td></td></tr><tr><td colspan="2">Text Only</td><td>57.05</td><td>50.55</td><td>67.39</td><td>60.59</td><td>69.87</td><td>36.86</td><td></td></tr><tr><td colspan="2">Ours</td><td>58.05</td><td>51.75</td><td>68.07</td><td>60.56</td><td>70.35</td><td>39.52</td><td></td></tr><tr><td></td><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td></tr><tr><td>CP</td><td>54.60</td><td>44.27</td><td>43.97</td><td>48.56</td><td>34.67</td><td>55.20</td><td>70.16</td><td>85.37</td></tr><tr><td>Image Only</td><td>54.84</td><td>44.35</td><td>43.43</td><td>50.68</td><td>34.78</td><td>53.92</td><td>71.67</td><td>85.07</td></tr><tr><td>Text Only</td><td>55.06</td><td>44.55</td><td>43.26</td><td>50.39</td><td>36.11</td><td>54.24</td><td>71.62</td><td>85.23</td></tr><tr><td>Ours</td><td>56.01</td><td>50.39</td><td>43.49</td><td>50.75</td><td>35.22</td><td>54.13</td><td>71.77</td><td>86.33</td></tr></table>
+
+Table 9: Uncertainty of all models on held-in datasets.  
+
+<table><tr><td>Tuning Dataset</td><td>Average</td><td>AI2D</td><td>MMBench</td><td>OOD-CV</td><td>ScienceQA</td><td>SeedBench</td></tr><tr><td>OKVQA</td><td>0.1314</td><td>0.1199</td><td>0.1073</td><td>0.1350</td><td>0.1149</td><td>0.1802</td></tr><tr><td>OCRVQA</td><td>0.0548</td><td>0.0555</td><td>0.0564</td><td>0.0669</td><td>0.0505</td><td>0.0446</td></tr><tr><td>GQA</td><td>0.2141</td><td>0.2032</td><td>0.2375</td><td>0.2136</td><td>0.2219</td><td>0.1941</td></tr><tr><td>VQAv2</td><td>0.0322</td><td>0.0286</td><td>0.0355</td><td>0.0350</td><td>0.0311</td><td>0.0306</td></tr><tr><td>TextCaps</td><td>0.0073</td><td>0.0078</td><td>0.0085</td><td>0.0115</td><td>0.0054</td><td>0.0030</td></tr><tr><td>A_OKVQA</td><td>0.1078</td><td>0.1318</td><td>0.0746</td><td>0.1066</td><td>0.1132</td><td>0.1130</td></tr><tr><td>RefCOCO</td><td>0.0275</td><td>0.0250</td><td>0.0309</td><td>0.0311</td><td>0.0267</td><td>0.0238</td></tr><tr><td>LLaVA-Instruct</td><td>0.0713</td><td>0.0670</td><td>0.0767</td><td>0.0787</td><td>0.0672</td><td>0.0668</td></tr><tr><td>ShareGPT</td><td>0.0317</td><td>0.0246</td><td>0.0380</td><td>0.0386</td><td>0.0320</td><td>0.0255</td></tr><tr><td>VG</td><td>0.0379</td><td>0.0334</td><td>0.0412</td><td>0.0417</td><td>0.0360</td><td>0.0372</td></tr><tr><td>Tuning Dataset</td><td>Average</td><td>AI2D</td><td>MMBench</td><td>OOD-CV</td><td>ScienceQA</td><td>SeedBench</td></tr><tr><td>OKVQA</td><td>0.1305</td><td>0.1199</td><td>0.1072</td><td>0.1317</td><td>0.1152</td><td>0.1787</td></tr><tr><td>OCRVQA</td><td>0.0544</td><td>0.0550</td><td>0.0560</td><td>0.0663</td><td>0.0505</td><td>0.0444</td></tr><tr><td>GQA</td><td>0.2141</td><td>0.2024</td><td>0.2408</td><td>0.2084</td><td>0.2229</td><td>0.1962</td></tr><tr><td>VQAv2</td><td>0.0323</td><td>0.0288</td><td>0.0361</td><td>0.0347</td><td>0.0312</td><td>0.0307</td></tr><tr><td>TextCaps</td><td>0.0072</td><td>0.0077</td><td>0.0087</td><td>0.0114</td><td>0.0055</td><td>0.0029</td></tr><tr><td>A_OKVQA</td><td>0.1077</td><td>0.1326</td><td>0.0751</td><td>0.1070</td><td>0.1117</td><td>0.1123</td></tr><tr><td>RefCOCO</td><td>0.0273</td><td>0.0251</td><td>0.0305</td><td>0.0305</td><td>0.0265</td><td>0.0238</td></tr><tr><td>LLaVA-Instruct</td><td>0.0713</td><td>0.0668</td><td>0.0771</td><td>0.0803</td><td>0.0662</td><td>0.0659</td></tr><tr><td>ShareGPT</td><td>0.0318</td><td>0.0246</td><td>0.0375</td><td>0.0394</td><td>0.0318</td><td>0.0255</td></tr><tr><td>VG</td><td>0.0378</td><td>0.0335</td><td>0.0408</td><td>0.0419</td><td>0.0361</td><td>0.0369</td></tr><tr><td>Tuning Dataset</td><td>Average</td><td>AI2D</td><td>MMBench</td><td>OOD-CV</td><td>ScienceQA</td><td>SeedBench</td></tr><tr><td>OKVQA</td><td>0.1306</td><td>0.1205</td><td>0.1067</td><td>0.1334</td><td>0.1157</td><td>0.1767</td></tr><tr><td>OCRVQA</td><td>0.0547</td><td>0.0554</td><td>0.0562</td><td>0.0668</td><td>0.0501</td><td>0.0448</td></tr><tr><td>GQA</td><td>0.2130</td><td>0.2020</td><td>0.2369</td><td>0.2115</td><td>0.2220</td><td>0.1926</td></tr><tr><td>VQAv2</td><td>0.0323</td><td>0.0288</td><td>0.0357</td><td>0.0344</td><td>0.0313</td><td>0.0311</td></tr><tr><td>TextCaps</td><td>0.0073</td><td>0.0078</td><td>0.0085</td><td>0.0115</td><td>0.0055</td><td>0.0030</td></tr><tr><td>A_OKVQA</td><td>0.1082</td><td>0.1324</td><td>0.0758</td><td>0.1069</td><td>0.1123</td><td>0.1135</td></tr><tr><td>RefCOCO</td><td>0.0273</td><td>0.0249</td><td>0.0303</td><td>0.0308</td><td>0.0266</td><td>0.0238</td></tr><tr><td>LLaVA-Instruct</td><td>0.0713</td><td>0.0666</td><td>0.0763</td><td>0.0801</td><td>0.0667</td><td>0.0666</td></tr><tr><td>ShareGPT</td><td>0.0316</td><td>0.0245</td><td>0.0379</td><td>0.0388</td><td>0.0319</td><td>0.0247</td></tr><tr><td>VG</td><td>0.0376</td><td>0.0335</td><td>0.0406</td><td>0.0417</td><td>0.0360</td><td>0.0363</td></tr></table>
+
+Table 10: Accuracy of baselines on held-in datasets when merging random model selections.  
+
+<table><tr><td rowspan="2">Merging Methods</td><td colspan="6">Vision-Language Classification Datasets</td></tr><tr><td>Average</td><td>AI2D</td><td>ScienceQA</td><td>SeedBench</td><td>MMBench</td><td>OOD-CV</td></tr><tr><td>DARE</td><td>15.68</td><td>11.98</td><td>5.19</td><td>8.47</td><td>26.53</td><td>26.21</td></tr><tr><td>DELLA</td><td>46.73</td><td>34.42</td><td>50.05</td><td>46.97</td><td>68.76</td><td>33.45</td></tr><tr><td>Linear</td><td>57.85</td><td>51.65</td><td>67.83</td><td>60.25</td><td>70.19</td><td>39.35</td></tr><tr><td>TIES</td><td>55.60</td><td>49.29</td><td>55.96</td><td>56.14</td><td>70.74</td><td>45.85</td></tr><tr><td>Task Arithmetic</td><td>15.13</td><td>5.18</td><td>12.98</td><td>5.20</td><td>26.53</td><td>25.78</td></tr><tr><td>Model Stock</td><td>50.65</td><td>44.82</td><td>65.12</td><td>50.08</td><td>64.86</td><td>28.37</td></tr><tr><td>Merging Methods</td><td>Average</td><td>AI2D</td><td>ScienceQA</td><td>SeedBench</td><td>MMBench</td><td>OOD-CV</td></tr><tr><td>DARE</td><td>10.18</td><td>0.00</td><td>0.00</td><td>0.00</td><td>25.40</td><td>25.50</td></tr><tr><td>DELLA</td><td>55.02</td><td>47.54</td><td>57.70</td><td>56.57</td><td>70.81</td><td>42.47</td></tr><tr><td>Linear</td><td>57.48</td><td>51.36</td><td>67.90</td><td>59.91</td><td>70.17</td><td>38.07</td></tr><tr><td>TIES</td><td>54.95</td><td>46.76</td><td>56.78</td><td>57.49</td><td>70.90</td><td>42.83</td></tr><tr><td>Task Arithmetic</td><td>20.54</td><td>24.19</td><td>0.00</td><td>24.96</td><td>27.46</td><td>26.10</td></tr><tr><td>Model Stock</td><td>49.56</td><td>43.56</td><td>65.10</td><td>48.56</td><td>63.24</td><td>27.34</td></tr><tr><td>Merging Methods</td><td>Average</td><td>AI2D</td><td>ScienceQA</td><td>SeedBench</td><td>MMBench</td><td>OOD-CV</td></tr><tr><td>DARE</td><td>15.39</td><td>12.05</td><td>2.96</td><td>8.72</td><td>26.81</td><td>26.42</td></tr><tr><td>DELLA</td><td>57.69</td><td>47.73</td><td>66.16</td><td>61.86</td><td>71.20</td><td>41.51</td></tr><tr><td>Linear</td><td>50.95</td><td>45.21</td><td>56.73</td><td>51.31</td><td>69.10</td><td>32.39</td></tr><tr><td>TIES</td><td>42.77</td><td>34.68</td><td>44.47</td><td>32.11</td><td>68.41</td><td>34.16</td></tr><tr><td>Task Arithmetic</td><td>29.52</td><td>25.29</td><td>38.54</td><td>30.89</td><td>28.80</td><td>24.08</td></tr><tr><td>Model Stock</td><td>49.77</td><td>43.04</td><td>65.00</td><td>48.70</td><td>64.21</td><td>27.88</td></tr><tr><td>Ours</td><td>58.05</td><td>51.75</td><td>68.07</td><td>60.56</td><td>70.35</td><td>39.52</td></tr></table>
+
+Table 11: Accuracy of baselines on held-out datasets when merging random model selections.  
+
+<table><tr><td rowspan="2">Merging Methods</td><td colspan="8">Vision-Language Generation Datasets</td></tr><tr><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td></tr><tr><td>DARE</td><td>18.78</td><td>0.41</td><td>7.82</td><td>19.06</td><td>27.11</td><td>3.85</td><td>32.48</td><td>40.70</td></tr><tr><td>DELLA</td><td>43.78</td><td>41.69</td><td>39.74</td><td>40.83</td><td>31.67</td><td>40.75</td><td>62.46</td><td>49.32</td></tr><tr><td>Linear</td><td>53.78</td><td>39.45</td><td>40.67</td><td>49.65</td><td>34.67</td><td>57.20</td><td>69.01</td><td>85.83</td></tr><tr><td>TIES</td><td>56.31</td><td>56.08</td><td>44.75</td><td>54.72</td><td>33.33</td><td>45.32</td><td>73.03</td><td>86.94</td></tr><tr><td>Task Arithmetic</td><td>4.04</td><td>0.04</td><td>0.91</td><td>0.33</td><td>24.67</td><td>0.26</td><td>0.96</td><td>1.09</td></tr><tr><td>Model Stock</td><td>39.37</td><td>3.35</td><td>37.30</td><td>39.54</td><td>32.33</td><td>54.38</td><td>61.51</td><td>47.18</td></tr><tr><td>Merging Methods</td><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td></tr><tr><td>DARE</td><td>3.71</td><td>0.00</td><td>0.00</td><td>0.00</td><td>26.00</td><td>0.00</td><td>0.00</td><td>0.00</td></tr><tr><td>DELLA</td><td>51.08</td><td>47.94</td><td>43.15</td><td>43.70</td><td>32.44</td><td>36.94</td><td>69.17</td><td>84.23</td></tr><tr><td>Linear</td><td>53.66</td><td>43.35</td><td>43.43</td><td>49.02</td><td>33.67</td><td>52.88</td><td>68.06</td><td>85.18</td></tr><tr><td>TIES</td><td>54.15</td><td>49.39</td><td>44.19</td><td>52.08</td><td>31.22</td><td>44.17</td><td>71.07</td><td>86.94</td></tr><tr><td>Task Arithmetic</td><td>22.70</td><td>4.47</td><td>20.60</td><td>19.69</td><td>25.89</td><td>4.53</td><td>41.02</td><td>42.72</td></tr><tr><td>Model Stock</td><td>38.79</td><td>1.95</td><td>33.38</td><td>38.30</td><td>32.67</td><td>54.34</td><td>58.55</td><td>52.33</td></tr><tr><td>Merging Methods</td><td>Average</td><td>OKVQA</td><td>TextVQA</td><td>GQA</td><td>MMMU</td><td>VizWiz</td><td>VQAv2</td><td>POPE</td></tr><tr><td>DARE</td><td>15.99</td><td>0.70</td><td>8.61</td><td>13.66</td><td>25.56</td><td>3.44</td><td>26.84</td><td>33.10</td></tr><tr><td>DELLA</td><td>48.95</td><td>31.97</td><td>40.62</td><td>47.34</td><td>34.22</td><td>48.07</td><td>64.35</td><td>76.06</td></tr><tr><td>Linear</td><td>54.96</td><td>51.04</td><td>44.12</td><td>50.83</td><td>32.78</td><td>48.33</td><td>71.48</td><td>86.11</td></tr><tr><td>TIES</td><td>54.40</td><td>45.01</td><td>44.11</td><td>51.81</td><td>32.89</td><td>50.50</td><td>69.93</td><td>86.54</td></tr><tr><td>Task Arithmetic</td><td>27.45</td><td>3.67</td><td>19.32</td><td>30.01</td><td>28.11</td><td>5.11</td><td>44.02</td><td>61.88</td></tr><tr><td>Model Stock</td><td>35.20</td><td>1.51</td><td>32.39</td><td>37.02</td><td>32.78</td><td>54.01</td><td>57.50</td><td>31.21</td></tr><tr><td>Ours</td><td>56.01</td><td>50.39</td><td>43.49</td><td>50.75</td><td>35.22</td><td>54.13</td><td>71.77</td><td>86.33</td></tr></table>
+
+Table 12: Uncertainty on held-in datasets during merging steps. Bold denotes the step UQ-Merge stops.  
+
+<table><tr><td rowspan="2">Held-in Datasets</td><td colspan="10">Merging Steps</td></tr><tr><td>1</td><td>2</td><td>3</td><td>4</td><td>5</td><td>6</td><td>7</td><td>8</td><td>9</td><td>10</td></tr><tr><td>AI2D</td><td>0.20129</td><td>0.11652</td><td>0.06806</td><td>0.06142</td><td>0.05759</td><td>0.05062</td><td>0.04600</td><td>0.04676</td><td>0.04328</td><td>0.04248</td></tr><tr><td>MMBench</td><td>0.23477</td><td>0.11862</td><td>0.04576</td><td>0.03985</td><td>0.03784</td><td>0.03560</td><td>0.03583</td><td>0.03488</td><td>0.03368</td><td>0.03307</td></tr><tr><td>ScienceQA</td><td>0.21144</td><td>0.13648</td><td>0.06652</td><td>0.05888</td><td>0.05306</td><td>0.04905</td><td>0.04543</td><td>0.04672</td><td>0.04356</td><td>0.04222</td></tr><tr><td>SeedBench</td><td>0.22157</td><td>0.12653</td><td>0.05543</td><td>0.04701</td><td>0.04384</td><td>0.04079</td><td>0.03865</td><td>0.03837</td><td>0.03703</td><td>0.03606</td></tr><tr><td>OOD-CV</td><td>0.19077</td><td>0.16634</td><td>0.04470</td><td>0.03616</td><td>0.03423</td><td>0.03168</td><td>0.03160</td><td>0.03097</td><td>0.02945</td><td>0.02990</td></tr><tr><td>Average</td><td>0.21197</td><td>0.13290</td><td>0.05609</td><td>0.04866</td><td>0.04531</td><td>0.04155</td><td>0.03950</td><td>0.03954</td><td>0.03740</td><td>0.03675</td></tr></table>
+
+Table 13: Accuracy on validation datasets during merging steps. Bold denotes the step UQ-Merge stops.  
+
+<table><tr><td rowspan="2">Validation Datasets</td><td colspan="10">Merging Steps</td></tr><tr><td>1</td><td>2</td><td>3</td><td>4</td><td>5</td><td>6</td><td>7</td><td>8</td><td>9</td><td>10</td></tr><tr><td>RealWorldQA</td><td>27.58</td><td>22.48</td><td>47.45</td><td>46.80</td><td>47.19</td><td>47.58</td><td>49.15</td><td>47.32</td><td>47.32</td><td>47.84</td></tr><tr><td>SeedBench 2 Plus</td><td>23.36</td><td>10.36</td><td>39.00</td><td>40.40</td><td>41.33</td><td>41.81</td><td>42.07</td><td>41.72</td><td>42.51</td><td>42.16</td></tr><tr><td>OCRBench</td><td>5.30</td><td>22.60</td><td>22.50</td><td>24.70</td><td>28.00</td><td>28.20</td><td>28.50</td><td>28.40</td><td>28.00</td><td>28.80</td></tr><tr><td>Average</td><td>18.75</td><td>18.48</td><td>36.32</td><td>37.30</td><td>38.84</td><td>39.20</td><td>39.91</td><td>39.15</td><td>39.28</td><td>39.60</td></tr></table>
+
+Table 14: Accuracy on held-out datasets during merging steps. Bold denotes the step UQ-Merge stops.  
+
+<table><tr><td rowspan="2">Held-out Datasets</td><td colspan="10">Merging Steps</td></tr><tr><td>1</td><td>2</td><td>3</td><td>4</td><td>5</td><td>6</td><td>7</td><td>8</td><td>9</td><td>10</td></tr><tr><td>OKVQA</td><td>27.05</td><td>54.77</td><td>50.48</td><td>48.01</td><td>46.21</td><td>44.13</td><td>50.39</td><td>47.90</td><td>48.01</td><td>46.21</td></tr><tr><td>TextVQA</td><td>23.15</td><td>36.06</td><td>34.24</td><td>37.96</td><td>41.19</td><td>41.82</td><td>43.49</td><td>43.22</td><td>37.96</td><td>41.19</td></tr><tr><td>GQA</td><td>61.73</td><td>54.08</td><td>50.77</td><td>50.29</td><td>48.97</td><td>49.98</td><td>50.75</td><td>49.75</td><td>50.29</td><td>48.97</td></tr><tr><td>MMMU</td><td>30.00</td><td>29.78</td><td>33.78</td><td>34.22</td><td>34.44</td><td>35.00</td><td>35.22</td><td>35.22</td><td>34.22</td><td>34.56</td></tr><tr><td>VizWiz</td><td>44.25</td><td>34.69</td><td>46.72</td><td>54.36</td><td>54.61</td><td>55.66</td><td>54.13</td><td>55.31</td><td>54.36</td><td>54.61</td></tr><tr><td>VQAv2</td><td>60.03</td><td>66.09</td><td>66.70</td><td>68.30</td><td>67.81</td><td>68.71</td><td>71.77</td><td>71.25</td><td>68.30</td><td>67.81</td></tr><tr><td>POPE</td><td>83.39</td><td>83.07</td><td>84.60</td><td>85.91</td><td>85.68</td><td>86.46</td><td>86.33</td><td>86.29</td><td>85.91</td><td>85.66</td></tr><tr><td>Average</td><td>47.09</td><td>51.22</td><td>52.47</td><td>54.15</td><td>54.13</td><td>54.54</td><td>56.01</td><td>55.56</td><td>54.15</td><td>54.14</td></tr></table>

2025/UQ-Merge_ Uncertainty Guided Multimodal Large Language Model Merging/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:026a7a313b6cb4ec1e9d32918879e8ad8c61797b22f357d7d33f877bc6f2596e
+size 1332928

2025/USDC_ A Dataset of ̲User ̲Stance and ̲Dogmatism in Long ̲Conversations/a4fabed3-7f27-4469-8663-7ac9acb62d69_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:186d4f21f23d6db32c8f2f3db538370be02ceccaa3f1c1130f595a8ee866befc
+size 1424468

2025/USDC_ A Dataset of ̲User ̲Stance and ̲Dogmatism in Long ̲Conversations/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8547b895280d14a9a3c3dacb6593ad1dd161a00bb70ebf70ff6e723560d47c74
+size 2018030

2025/Un-considering Contextual Information_ Assessing LLMs’ Understanding of Indexical Elements/1f95b7d5-2289-4e70-9979-ad44a236414a_content_list.json

@@ -0,0 +1,2290 @@
+[
+    {
+        "type": "text",
+        "text": "Un-considering Contextual Information: Assessing LLMs' Understanding of Indexical Elements",
+        "text_level": 1,
+        "bbox": [
+            213,
+            90,
+            783,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Metehan Oğuz Yavuz Bakman Duygu Nur Yaldiz",
+        "bbox": [
+            270,
+            156,
+            729,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of Southern California",
+        "bbox": [
+            361,
+            175,
+            636,
+            190
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{moguz, ybakman, yaldiz}@usc.edu",
+        "bbox": [
+            336,
+            192,
+            660,
+            208
+        ],
+        "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 performances in tasks related to coreference resolution. However, previous studies mostly assessed LLM performance on coreference resolution with nouns and third person pronouns. This study evaluates LLM performance on coreference resolution with indexicals like  $I$ , you, here and tomorrow, which come with unique challenges due to their linguistic properties. We present the first study examining how LLMs interpret indexicals in English, releasing the English Indexical Dataset with 1600 multiple-choice questions. We evaluate pioneering LLMs, including GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, and DeepSeek V3. Our results reveal that LLMs exhibit an impressive performance with some indexicals ( $I$ ), while struggling with others (you, here, tomorrow), and that syntactic cues (e.g. quotation) contribute to LLM performance with some indexicals, while they reduce performance with others. Code and data are available at: https://github.com/metehanoguzz/LLMs-Indexicals-English",
+        "bbox": [
+            141,
+            286,
+            460,
+            627
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            638,
+            260,
+            653
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large Language Models (LLMs) have demonstrated remarkable capabilities in zero-shot and few-shot learning, excelling across a wide array of tasks such as machine translation, text summarization, and question answering (OpenAI, 2024; Ye et al., 2023; Bakman et al., 2024; Yaldiz et al., 2024). Their versatility has led to widespread applications in diverse domains, including education, law, and medicine.",
+        "bbox": [
+            112,
+            663,
+            489,
+            806
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "As the use of LLMs continues to expand, understanding their underlying behaviors has become increasingly important. Recent studies have evaluated the performance of large language models on linguistic tasks such as coreference resolution (Gan et al., 2024; Le and Ritter, 2023; Brown et al., 2020; Yang et al., 2022; Agrawal et al., 2022).",
+        "bbox": [
+            112,
+            809,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Previous work on coreference resolution mostly focused on how coreference is established between two third person entities such as proper names (e.g. Andy, the mechanic) and third person pronouns (e.g. he, him, himself) in English and other languages (e.g. Yang et al., 2022; Yang, 2025). In this study, we investigate how LLMs establish coreference with indexical elements (e.g. I, you, here), which differ from third person nouns/pronouns in substantial ways and bring unique challenges for LLMs (see Figure 1 as an example). We investigate how state-of-the-art LLMs interpret the indexical elements I, you, here and tomorrow in English sentences, and whether context or grammatical constraints influence their decisions. To the best of our knowledge, this is the first study examining LLMs' handling of indexical elements in English. Our key contributions are as follows:",
+        "bbox": [
+            507,
+            261,
+            884,
+            549
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We introduce the English Indexical Dataset, comprising 400 interpretation samples for each indexical element, I, you, here, tomorrow, totaling 1,600 instances.",
+            "- We evaluate the performance of four frontier LLMs, GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, and DeepSeek-V3, on the interpretation of indexical elements in English.",
+            "- We show that LLM performances are not uniform across different types of indexical elements: indexical  $I$  is successfully interpreted by most LLMs, other indexicals like you, here and tomorrow lead to poor performances.",
+            "- We show that quotation affects LLMs performances differently: quotation reduces LLMs' accuracies with tomorrow, but it increases accuracies with here."
+        ],
+        "bbox": [
+            531,
+            552,
+            884,
+            818
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2 Indexical elements",
+        "text_level": 1,
+        "bbox": [
+            507,
+            831,
+            705,
+            847
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Indexical pronominals like  $I$ , you, here and now are used to refer to referents of the speech-act coordinates (e.g. Kaplan, 1977; Schlenker, 2003). For instance,  $I$  refers to author (speaker) of the utterance,",
+        "bbox": [
+            507,
+            857,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "23410",
+        "bbox": [
+            473,
+            927,
+            527,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2025, pages 23410-23427 July 27 - August 1, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c5bb29204fd67291193cbcc99089db521750b0c4ff01b9bc1be4a307377e206a.jpg",
+        "image_caption": [
+            "Figure 1: An example for LLM misinterpreting indexical element 'here', uttered by a speaker in Los Angeles."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            142,
+            84,
+            228,
+            140
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/aa12c66ee0f6e7c874b38d3f7d630d12dbc293cba44faf1b8e189e4f7d95c15a.jpg",
+        "image_caption": [],
+        "image_footnote": [
+            "Of course! You should go to Central Park and Statue of Liberty. I am sure he will love them!"
+        ],
+        "bbox": [
+            231,
+            96,
+            300,
+            140
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/bf8b79feac680a65fe40a94bf26bac804c021beb71066efc8343863289654e75.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            673,
+            67,
+            860,
+            109
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "while here refers to the location where the utterance was made. Thus, a sentence like (1) means different things if uttered by different people and/or in different places. If John utters (1-a) in Los Angeles, it means that John was born in Los Angeles, but if Mary utters the same sentence in New York it means that Mary was born in New York.",
+        "bbox": [
+            112,
+            180,
+            487,
+            291
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "(1) a. I was born here.",
+            "b. Andrew said that I went to Buckhead."
+        ],
+        "bbox": [
+            112,
+            298,
+            487,
+            328
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Indexicals are interpreted inside the context of utterance, referring to the actual speech-act coordinates like the author or the location of the actual utterance. As a result, if (1-b) is uttered by John, the indexical  $I$  can only be interpreted as referring to John as the speaker, leading to a reading like 'Andrew said that John went to Buckhead'. Crucially, even though Andrew's speech/claim is reported in (1-b), the indexical  $I$  cannot refer to Andrew.",
+        "bbox": [
+            112,
+            337,
+            487,
+            481
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Direct quotation is an exception to this generalization, where reported material is interpreted as verbatim utterance/thoughts of its owner. Thus, when indexicals appear inside direct quotation, they are interpreted inside the reported context, rather than the actual context of utterance. In other words, direct quotation 'shifts' the interpretations of indexicals into the reported context. For example,  $I$  and here in (2) appear inside direct quotation, where Andrew's speech is reported.",
+        "bbox": [
+            112,
+            483,
+            487,
+            643
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "(2) While we were in Atlanta, Andrew said \"I was born here.\"",
+        "bbox": [
+            112,
+            650,
+            487,
+            681
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Regardless of who utters (2), the sentence means that Andrew was born in Atlanta, so both  $I$  and here are 'shifted' into the reported context, where Andrew is the speaker and Atlanta is the location<sup>1</sup>",
+        "bbox": [
+            112,
+            688,
+            485,
+            751
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Indexicals differ from other pronominals in substantial ways. First, though syntactic and semantic factors can affect how a pronoun is interpreted (e.g. subject bias), pronouns are typically ambiguous regarding what/who they refer to. For example, the third person pronoun he in (3) is most naturally interpreted as referring to the subject John (for syntactic or contextual reasons), but the object Bill is",
+        "bbox": [
+            112,
+            753,
+            487,
+            881
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "still a possible antecedent, causing ambiguity between two different readings (e.g. Crawley et al., 1990; Stewart and Pickering, 1998; Pickering and Majid, 2007). In addition, he can refer to any contextually salient person that is not mentioned in the sentence (e.g. Peter), which makes pronouns even more ambiguous and context-dependent.",
+        "bbox": [
+            507,
+            179,
+            882,
+            292
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "(3) John hit Bill and he ran away.",
+        "bbox": [
+            507,
+            300,
+            786,
+            316
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As a result, semantic/contextual information plays a crucial role in how pronouns are interpreted, and speakers use those cues to establish coreference with pronouns. For instance, if (4) is uttered in a context where John is a supportive and humble coworker, the most natural interpretation is that John suggests that Bill should get promoted ( $he = \\text{Bill}$ ). However, if John is arrogant and jealous, the most natural interpretation is that John suggests that John should get promoted ( $he = \\text{John}$ ).",
+        "bbox": [
+            507,
+            324,
+            882,
+            483
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "(4) John told Bill that he should get promoted.",
+        "bbox": [
+            507,
+            493,
+            875,
+            508
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Indexicals, on the other hand, unambiguously refer to the referents of the speech-act coordinates. For example,  $I$  in (5) refers to the speaker regardless of what we know about John or Bill.  $I$  refers to the speaker even if John and/or Bill are arrogant and jealous, so contextual information like this should be disregarded while interpreting indexicals.",
+        "bbox": [
+            507,
+            517,
+            882,
+            629
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "(5) John told Bill that I should get a promotion.",
+        "bbox": [
+            507,
+            637,
+            882,
+            653
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, indexicals are restricted by different syntactic factors than pronouns (e.g. quotation vs non-quotation) and are typically unambiguous, while pronouns are free to refer to a wide range of entities. Thus, indexical elements create a unique challenge for LLMs, requiring to 'disregard' semantic/contextual cues that might prime interpretations through other antecedents (unlike pronouns).",
+        "bbox": [
+            507,
+            663,
+            882,
+            791
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Experimental design",
+        "text_level": 1,
+        "bbox": [
+            507,
+            807,
+            721,
+            822
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Dataset Curation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            834,
+            690,
+            848
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We curated a dataset specifically designed to test how LLMs interpret indexical elements like  $I$ , you, here, and tomorrow in different contexts of utterance. Specifically, we assess how these mod",
+        "bbox": [
+            507,
+            856,
+            882,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>Some languages (not English) allow indexical elements to 'shift' without quotation. See Deal (2020) for an overview.",
+        "bbox": [
+            112,
+            894,
+            487,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "23411",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "els interpret indexicals in 'shifted' context prime, where context more naturally requires the indexical should be interpreted inside reported context (e.g. Peter is one of the most arrogant students in my classroom. ... Peter says that I am smart.) vs 'non-shifted' context prime, where context more naturally requires the indexical should be interpreted inside actual speech context (e.g. Peter is very kind and supportive. ... Peter says that I am smart.). We also included direct quotations in both contexts (e.g., Peter says, \"I am smart\") to examine if LLMs can successfully consider syntactic factors (quotation vs regular sentences) while ignoring misleading information from the context during coreference resolution with indexical elements.",
+        "bbox": [
+            110,
+            84,
+            492,
+            324
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For each type of indexical, we design 100 sentences and for each sentence we apply the four different transformations explained above. Overall, we have 400 samples per indexical, compromising a total dataset of size 1600.",
+        "bbox": [
+            112,
+            325,
+            489,
+            405
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We utilize GPT-4o to curate the dataset, by giving a detailed description of the task along with some in-context examples. Then it is asked to generate scenarios with a stimulus sentence in two different contexts (See Appendix A.1 for the prompts used), along with specific questions addressing the referent of the indexical in each sentence. To ensure the quality of the dataset,  $25\\%$  of the dataset (400 trials = 100 sentences in four conditions) was randomly selected for evaluation, and the evaluation process consisted of three steps.",
+        "bbox": [
+            110,
+            406,
+            489,
+            582
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the first step, we confirmed that all sentences were grammatically correct. In the second step, we confirmed that all quotation condition sentences had quoted embedded clauses, and all non-quoted condition sentences had regular (non-quoted) embedded clauses. We also made sure that the two sentence conditions were maximally similar, except for the quotation vs non-quoted status (i.e. the only difference between two sentence conditions was the quotation). In the third step, we checked the context prime texts for each condition in each sentence, making sure that the correct readings (shifted vs non-shifted) were primed by the context description. For example, for an item condition where here was supposed to be shifted, we confirmed that the context description would be most naturally followed by a sentence where here would be shifted.",
+        "bbox": [
+            110,
+            583,
+            489,
+            871
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "100 trials from each indexical item (25% for each indexical), 400 trials in total, were randomly selected to make sure that the evalua",
+        "bbox": [
+            112,
+            873,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tion/confirmation was representative of all indexical item conditions (i.e. items with  $I$ , you, here and tomorrow).",
+        "bbox": [
+            507,
+            84,
+            884,
+            131
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To eliminate potential gender bias, each dataset sample exclusively uses either male or female names, alternating to ensure a balanced distribution with  $50\\%$  of the samples containing female names and  $50\\%$  male names. This method promotes gender neutrality across the dataset. Sample details are in Appendix A.2, and the complete dataset is available in the supplementary materials.",
+        "bbox": [
+            507,
+            133,
+            885,
+            263
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Models",
+        "text_level": 1,
+        "bbox": [
+            507,
+            275,
+            613,
+            288
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In our evaluation, we utilize four recent state-of-the-art LLMs: GPT-4o (OpenAI, 2024), Claude 3.5 Sonnet (Anthropic), Gemini 1.5 Pro (Team, 2024), and DeepSeek-V3 (DeepSeek-AI, 2024). This selection of diverse models provides a comprehensive evaluation of LLM performance with indexicals.",
+        "bbox": [
+            507,
+            298,
+            885,
+            394
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 Evaluation Strategy",
+        "text_level": 1,
+        "bbox": [
+            507,
+            407,
+            712,
+            423
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To assess the performance of the model, we specifically prompt it to answer questions designed to test its capabilities as described in Section 3.1. Additionally, to ensure focused responses, we restrict the model's answers to one of two predefined options: the 'shifted' option and the 'non-shifted' option. We provide the prompt in Appendix B.1.",
+        "bbox": [
+            507,
+            430,
+            885,
+            542
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.4 Metrics.",
+        "text_level": 1,
+        "bbox": [
+            507,
+            556,
+            621,
+            570
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We assess model accuracy across four cases for each indexical: (i) Non-quoted sentences with shifted context prime, (ii) Non-quoted with non-shifted prime, (iii) Quoted with shifted prime, and (iv) Quoted with non-shifted prime. Optimal performance would be achieved by always selecting the 'shifted' option in quoted conditions and selecting the 'non-shifted' option in non-quoted conditions.",
+        "bbox": [
+            507,
+            577,
+            885,
+            707
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            721,
+            729,
+            737
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We present the experimental results in Figure 2 and we discuss them in detail for each indexical type:",
+        "bbox": [
+            507,
+            749,
+            880,
+            781
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Indexical  $I$ . The results show that all four models perform near optimum in sentences without quotation with an average accuracy of  $99\\%$  (always correctly selecting the non-shifted option). For the quoted sentences GPT-4o and Gemini 1.5 pro again perform almost optimum with a mean accuracy larger than  $94\\%$ , followed by Claude 3.5 Sonnet with  $89\\%$  mean accuracy (correctly selecting the",
+        "bbox": [
+            507,
+            791,
+            884,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "23412",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/f9f419f787417918e7e187d65e9fba79b67a9039f2aa210287c23937bcc167b3.jpg",
+        "image_caption": [
+            "Figure 2: From left to right: Performance analysis plot of for the indexical 'I', Performance analysis plot of for the indexical 'you', Performance analysis plot of for the indexical 'here', Performance analysis plot of for the indexical 'tomorrow'. Dark blue bars = Shifting context prime, Light blue bars = Non-shifted context prime."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            80,
+            310,
+            221
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/3544abbbfbd62f1b1b985d3bdfa599681da28419bfd7f01981e98bf88ac38fb0.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            314,
+            80,
+            497,
+            221
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/919f33cfee1f13d302aa79c95ff2cdd18e8be8b4288bf22f380f38d4872eeece.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            500,
+            80,
+            684,
+            219
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/284623814d1074d170622f6c183400e19b1b08976495384e288fa51665fca8ce.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            685,
+            80,
+            873,
+            221
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "shifted option). However, DeepSeek V3 fails to always select the shifted option on quoted sentences. Especially, the model performance drops significantly to  $17\\%$  when context primes non-shifted readings, suggesting that inclusion of quotation makes the model more sensitive to the linguistically irrelevant effects of context prime. Moreover, considering that the model performance only reaches  $78\\%$  accuracy in quotation conditions seems to imply that the model might have a bias towards the non-shifted reading overall, reducing model performance in quotation conditions.",
+        "bbox": [
+            112,
+            271,
+            489,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Indexical you. The results indicate that LLMs generally perform worse with the indexical you than with  $I$ . We see similar patterns across models, suggesting they perform mostly at similar levels. All models are sensitive to the effects of context prime, performing lower when the context primes the incorrect option (i.e. shifted reading in non-quotation and non-shifted reading in quotation). Moreover, quotation consistently results in lower performance across models. Notably, Gemini 1.5 Pro excels in non-quotation accuracy (92%), though its performance drops significantly under quotation conditions to the levels of other models. Overall, the results suggest that models interpret you based on the linguistically irrelevant context prime rather than the sentence type, which indeed determines the correct readings of the indexicals.",
+        "bbox": [
+            112,
+            475,
+            489,
+            750
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Indexical here. The results show that LLMs, similar to you, struggle to interpret the indexical here, especially when context primes the incorrect option, leading to poor performance. However, different from you, quotation leads to higher performance with here. In non-quotation conditions, all LLMs make their selections almost exclusively based on context prime, with shifted primes showing over  $96\\%$  accuracy and non-shifted primes less than  $2\\%$ , where context prime ideally should not",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "have any effects on the selection. In contrast, under quotation conditions, the influence of context prime diminishes, leading to higher performances. Particularly, DeepSeek V3 and Gemini 1.5 Pro exhibit impressive performance with accuracies above  $97\\%$  and  $94\\%$ , respectively, followed by Gemini 1.5 Pro and GPT-4o, with accuracies above  $64\\%$  and  $37\\%$ .",
+        "bbox": [
+            507,
+            272,
+            884,
+            384
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Indexical tomorrow. The results indicate that LLMs have a strong bias towards the non-shifted interpretations of tomorrow. We see that the models almost always select the non-shifted option for trials with tomorrow, regardless of context prime or sentence type. While Claude 3.5 Sonnet and Gemini 1.5 Pro selects the non-shifted options  $100\\%$  of the time, GPT-4o and DeepSeek V3 select the non-shifted option  $94\\%$  and  $83\\%$  of the time, respectively. This strong bias leads to illusory high accuracies in non-quotation conditions, while causes extremely low performance in quotation conditions.",
+        "bbox": [
+            507,
+            394,
+            884,
+            587
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            599,
+            665,
+            614
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Coreference resolution has been extensively studied in prior research (Gan et al., 2024; Le and Ritter, 2023; Brown et al., 2020; Yang et al., 2022; Agrawal et al., 2022). However, indexical elements exhibit distinct syntactic properties compared to other (non-indexical) pronominals, as discussed in Section 2. Previous work by Oğuz et al. (2024) explored how LLMs interpret indexicals in Turkish, where indexicals possess different grammatical properties than in English and can shift without quotation. To the best of our knowledge, this study is the first to evaluate the performance of LLMs in interpreting indexicals in English.",
+        "bbox": [
+            507,
+            625,
+            884,
+            834
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "6 Discussion and Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            847,
+            774,
+            862
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our results show that LLM performances are not uniform across different types of indexical elements and sentence types. While most LLMs",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "23413",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "perform at impressive levels interpreting the indexical  $I$ , their performances drop significantly in other indexicals, particularly in here and tomorrow. Moreover, though sentence type (quotation vs non-quadation) affects LLMs performances in general, the effects are not similar across different indexical types and models. While quotation increases LLMs' performance with here, it decreases their performance with you and tomorrow. In addition, tomorrow seems to be affected by quotation in a greater magnitude than you. In conclusion, we find that different types of indexicals show unique patterns regarding how they are interpreted by the LLMs.",
+        "bbox": [
+            112,
+            84,
+            492,
+            307
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our results diverge from those reported in Oğuz et al. (2024), who tested how first person indexical in Turkish ben 'I' was interpreted by different LLMs, including GPT-4o and show that LLMs exhibit very poor performance interpreting ben 'I'. Here, we report that LLMs perform at an almost human-like level with interpreting the English indexical  $I$ . This might be due to lower amounts of available resources to train the models in Turkish, compared to English. Another reason for low performance in Turkish could be due to the fact that Turkish is a pro-drop language, meaning that the subject pronouns can be dropped (silent). Oğuz et al. (2024) used sentences where the first person indexical ben 'I' was dropped, which might have made the task more challenging for the LLMs considering that dropped indexicals can show different properties than overt ones (Oğuz et al., 2020). This linguistic difference between Turkish and English might have caused different results between the Turkish and English tests.",
+        "bbox": [
+            115,
+            310,
+            490,
+            646
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            659,
+            250,
+            674
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Our work explores how LLMs interpret indexical elements in a black-box setting but does not provide experimental analyses that investigate the underlying reasons for these behaviors by examining the models' internals or training data. Future research could adopt a white-box approach to analyze these behaviors in greater depth, offering valuable insights into the mechanisms driving LLMs' interpretation of indexicals.",
+        "bbox": [
+            112,
+            684,
+            490,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            857,
+            213,
+            872
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "ref_text",
+        "text": "Monica Agrawal, Stefan Hegselmann, Hunter Lang, Yoon Kim, and David A. Sontag. 2022. Large language models are few-shot clinical information ex",
+        "bbox": [
+            115,
+            879,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "tractors. In Conference on Empirical Methods in Natural Language Processing.",
+            "Anthropic. The claude 3 model family: Opus, sonnet, haiku.",
+            "Yavuz Faruk Bakman, Duygu Nur Yaldiz, Baturalp Buyukates, Chenyang Tao, Dimitrios Dimitriadis, and Salman Avestimehr. 2024. MARS: Meaning-aware response scoring for uncertainty estimation in generative LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7752-7767, Bangkok, Thailand. Association for Computational Linguistics.",
+            "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, Jeff 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. *ArXiv*, abs/2005.14165.",
+            "Rosalind A. Crawley, Rosemary J. Stevenson, and David Kleinman. 1990. The use of heuristic strategies in the interpretation of pronouns. Journal of Psycholinguistic Research, 14.",
+            "Amy Rose Deal. 2020. A theory of indexical shift: meaning, grammar, and crosslinguistic variation. MIT Press, Boston, MA.",
+            "DeepSeek-AI. 2024. Deepseek-v3 technical report. Preprint, arXiv:2412.19437.",
+            "Yujian Gan, Massimo Poesio, and Juntao Yu. 2024. Assessing the capabilities of large language models in coreference: An evaluation. In International Conference on Language Resources and Evaluation.",
+            "David Kaplan. 1977. Demonstratives: An essay on the semantics, logic, metaphysics, and epistemology of demonstratives and other indexicals. *Themes from Kaplan*, pages 565-614.",
+            "Nghia T. Le and Alan Ritter. 2023. Are large language models robust coreference resolvers?",
+            "Metehan Oğuz, Yusuf Ciftci, and Yavuz Faruk Bakman. 2024. Do LLMs recognize me, when I is not me: Assessment of LLMs understanding of Turkish indexical pronouns in indexical shift contexts. In Proceedings of the First Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2024), pages 53-61, Bangkok, Thailand and Online. Association for Computational Linguistics.",
+            "OpenAI. 2024. Gpt-4 technical report. Preprint, arXiv:2303.08774."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "23414",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Metehan Oğuz, Burak Öney, and Dennis Ryan Storoshenko. 2020. Obligatory indexical shift in Turkish. In Proceedings of Canadian Linguistic Association (CLA), Western University, London, ON, Canada.",
+            "Martin Pickering and Asifa Majid. 2007. What are implicit causality and consequentiality? Language & Cognitive Processes, 22.",
+            "Philippe Schlenker. 2003. A plea for monsters. Linguistics and Philosophy, 26:29-120.",
+            "Andrew J. Stewart and Martin Pickering. 1998. Implicit consequentiality. In Proceedings of the 20th Annual Conference of the Cognitive Science Society.",
+            "Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. Preprint, arXiv:2403.05530.",
+            "Duygu Nur Yaldiz, Yavuz Faruk Bakman, Baturalp Buyukates, Chenyang Tao, Anil Ramakrishna, Dimitrios Dimitriadis, Jieyu Zhao, and Salman Avestimehr. 2024. Do not design, learn: A trainable scoring function for uncertainty estimation in generative llms. Preprint, arXiv:2406.11278.",
+            "Xiaohan Yang, Eduardo Peynetti, Vasco Meerman, and Christy Tanner. 2022. What gpt knows about who is who. In First Workshop on Insights from Negative Results in NLP.",
+            "Xiulin Yang. 2025. Language models at the syntax-semantics interface: A case study of the long-distance binding of Chinese reflexive ziji. In Proceedings of the 31st International Conference on Computational Linguistics, pages 3808-3824, Abu Dhabi, UAE. Association for Computational Linguistics.",
+            "Junjie Ye, Xuanting Chen, Nuo Xu, Can Zu, Zekai Shao, Shichun Liu, Yuhan Cui, Zeyang Zhou, Chao Gong, Yang Shen, Jie Zhou, Siming Chen, Tao Gui, Qi Zhang, and Xuanjing Huang. 2023. A comprehensive capability analysis of GPT-3 and GPT-3.5 series models. Preprint, arXiv:2303.10420."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            665
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A Dataset Details",
+        "text_level": 1,
+        "bbox": [
+            512,
+            84,
+            677,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A.1 Dataset Generation Prompts",
+        "text_level": 1,
+        "bbox": [
+            512,
+            109,
+            781,
+            124
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We provide the prompts used for data generation in Tables 1, 2, 3, and 4. The prompts used for question generation per scenario is presented in Tables 5, 6, 7, and 8.",
+        "bbox": [
+            510,
+            130,
+            880,
+            192
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A.2 Samples From the Dataset",
+        "text_level": 1,
+        "bbox": [
+            512,
+            205,
+            764,
+            219
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We provide samples from the dataset for each indexical element we investigate in Tables 9, 10, 11, and 12.",
+        "bbox": [
+            510,
+            225,
+            880,
+            271
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "B Experimental Details",
+        "text_level": 1,
+        "bbox": [
+            512,
+            285,
+            727,
+            300
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "B.1 Prompt Used in Evaluation",
+        "text_level": 1,
+        "bbox": [
+            512,
+            311,
+            769,
+            326
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The prompt we employ for the evaluation is as follows:",
+        "bbox": [
+            510,
+            332,
+            880,
+            361
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Read the following passage carefully and answer the question at the end: {stimuli} {question}",
+        "bbox": [
+            510,
+            369,
+            843,
+            432
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Please provide your answer as: either {option1} or {option2}. Do not include any additional explanation or text.",
+        "bbox": [
+            510,
+            434,
+            857,
+            480
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "23415",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:",
+        "bbox": [
+            149,
+            180,
+            845,
+            261
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Details for stimulus sentences: Each stimulus sentence will have a structure like \"While John was speaking to Travis, he said that Chris appreciates you a lot.\" Please make sure to use the names John, Travis, and Chris. Make sure that Chris is the subject of the embedded clause, and make sure that the sentence begins as \"While John was speaking to Travis...\". Please use different embedded verbs. For example, rather than appreciates you a lot, you can use saw you at the market, etc. But the action should be done by Chris, and the object should be \"you\".",
+        "bbox": [
+            149,
+            277,
+            848,
+            376
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Details about the context descriptions: The contexts will prime how \"you\" in the stimulus is interpreted. To manipulate this, I will give contexts in which \"you\" would refer to my addressee, which is you. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if \"you\" referred to John's addressee. However, since \"you\" means the current addressee, this will be a misinterpretation. For example, a correct reading prime would be like \"Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris asked for help from you a few times in the past, and you always helped him.\" This context makes it sound like Chris would appreciate you for all your help, and thus it would not be surprising if Chris appreciated you (Donald). However, in the incorrect interpretation prime, I will use a context like \"Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris asked for help from Travis a few times in the past, and Travis always helped him.\" In this context, it would be more natural if Chris appreciated Travis, who helped him, instead of you (Donald), and thus would be natural if \"you\" in the sentence was interpreted as referring to Travis (though \"you\" should refer to my addressee, which is you (Donald)). This would create the incorrect misinterpretation. Please make sure to keep the person names constant. Also, make sure that the context starts exactly as \"Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles.\"",
+        "bbox": [
+            149,
+            388,
+            848,
+            682
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format:",
+        "bbox": [
+            151,
+            695,
+            845,
+            743
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "stimulussentence:stimulus stimulus sentence",
+        "bbox": [
+            151,
+            744,
+            502,
+            759
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "correct_context: correct prime context sentence",
+        "bbox": [
+            151,
+            760,
+            507,
+            776
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "wrong_context: incorrect prime context sentence",
+        "bbox": [
+            151,
+            777,
+            515,
+            791
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Prompt used to generate dataset samples for indexical 'you'.",
+        "bbox": [
+            262,
+            804,
+            732,
+            819
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "23416",
+        "bbox": [
+            475,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:",
+        "bbox": [
+            149,
+            180,
+            845,
+            262
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Details for stimulus sentences: Each stimulus sentence will have a structure like \"When we spoke last summer, John said that Chris was going to go to Greece tomorrow.\" Please make sure to use the names John, and Chris. Make sure that Chris is the subject of the embedded clause. Please use different embedded verbs. For instance, rather than go to Greece, you can use have a visa appointment, etc. But the action should be done \"tomorrow\".",
+        "bbox": [
+            151,
+            277,
+            845,
+            360
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Details about the context descriptions: The contexts will prime how \"tomorrow\" in the stimulus is interpreted. To manipulate this, I will give contexts in which \"tomorrow\" would refer to the day after the actual (matrix) sentence is uttered. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if \"tomorrow\" referred to the day after John spoke. However, since \"tomorrow\" means the day after the current day, this will be a misinterpretation. For example, a correct reading prime would be like \"Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is very careful about planning everything, and always plans his stuff ahead of time.\" This context makes it sound like Chris would plan his trip to Greece ahead of time, and thus it would not be surprising if John said last summer that Chris was going to have a trip tomorrow. However, in the incorrect interpretation prime, I will use a context like \"Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is very lazy and never plans his stuff until the last moment.\" In this context, it would be more natural if Chris was going to go to Greece last summer, the day after John spoke to me, instead of the day after today, and thus would be natural if \"tomorrow\" in the sentence was interpreted as the day after John spoke to me (though \"tomorrow\" should refer to the day after today). This would create the incorrect misinterpretation. Please make sure to keep the person names constant. Also, make sure that the context starts exactly as \"Hi! I am Andrew. I am a graduate student and I live in Los Angeles.\"",
+        "bbox": [
+            151,
+            373,
+            848,
+            680
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format:",
+        "bbox": [
+            151,
+            696,
+            845,
+            743
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "stimulussentence:stimulus stimulus sentence",
+        "bbox": [
+            151,
+            745,
+            502,
+            759
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "correct_context: correct prime context sentence",
+        "bbox": [
+            151,
+            760,
+            507,
+            776
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "wrong_context: incorrect prime context sentence",
+        "bbox": [
+            151,
+            778,
+            515,
+            791
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Prompt used to generate dataset samples for indexical 'tomorrow'.",
+        "bbox": [
+            242,
+            804,
+            752,
+            819
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "23417",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:",
+        "bbox": [
+            149,
+            212,
+            845,
+            293
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Details for stimulus sentences: Each stimulus sentence will have a structure like \"Chris thinks that I will win the race.\" Please make sure to use the name Chris as the matrix subject, and \"I\" as the embedded subject subject. Please use different embedded verbs. For example, rather than win the race, you can use study hard, etc. But the action should be done by \"I\".",
+        "bbox": [
+            151,
+            309,
+            848,
+            374
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Details about the context descriptions: The contexts will prime who \"I\" in the stimulus refers to. To manipulate this, I will give contexts in which \"I\" would refer to the speaker. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if \"I\" referred to Chris. However, since Chris is not the speaker, this will lead to an incorrect interpretation. For example, a correct reading prime would be like \"Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris is a supportive friend, and has always trusted in my abilities. There is a race next week.\" This context makes it sound like Chris would support the speaker in a race, and predict that the speaker would win the race. However, in the incorrect interpretation prime, I will use a context like \"Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris is usually very competitive, and has bullied other people in front of me. There is a race next week...\" In this context, it would be more natural if Chris thought that he would win the race, instead of the speaker Andrew, and thus would be natural if \"I\" in the sentence referred to Chris (though \"I\" should refer to the speaker Andrew). This would create the incorrect misinterpretation. Please make sure to keep the person names constant. Also, make sure that the context starts exactly as \"Hi! I am Andrew. I am a graduate student and I live in Los Angeles.\"",
+        "bbox": [
+            151,
+            388,
+            848,
+            646
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format: stimulusSentence: stimulus stimulus sentence correct_context: correct prime context sentence wrong_context: incorrect prime context sentence",
+        "bbox": [
+            151,
+            663,
+            845,
+            760
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Prompt used to generate dataset samples for indexical 'I'.",
+        "bbox": [
+            272,
+            772,
+            724,
+            787
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "23418",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:",
+        "bbox": [
+            149,
+            164,
+            845,
+            244
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Details for stimulus sentences: Each stimulus sentence will have a structure like \"When I was in New York with John, he said that Chris wanted to explore here.\" Please make sure to use the city name New York for the sentence, and the names John and Chris. John will always be the person who says something that Chris will do. In each sentence, Chris will be the person who is doing something \"here\". Please use different verbs. For example, rather than explore here, you can use attend a conference here, etc.. Use various verbs. But the action should be done \"here\".",
+        "bbox": [
+            151,
+            261,
+            845,
+            357
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Details about the context descriptions: The contexts will prime where \"here\" in the stimulus refers to. To manipulate this, I will give contexts in which \"here\" would refer to Los Angeles, where the author/speaker of the sentence is located. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if \"here\" referred to New York. However, since the speaker is not in New York, this will lead to an incorrect interpretation.",
+        "bbox": [
+            151,
+            373,
+            845,
+            453
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "For example, a correct reading prime would be like \"Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They love exploring new cities, and surprisingly they did not spend much time in Los Angeles.\" This context makes it sound like John and Chris would like to come to Los Angeles to explore around, and thus the \"explore around here\" in the stimulus sentence would be naturally understood as Los Angeles. However, in the incorrect interpretation prime, I will use a context like \"Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They love exploring new cities, and surprisingly they did not spend much time in New York.\" In this context, it would be more natural if John and Chris would like to go to New York, and thus would be natural if \"explore around here\" would be exploring around New York (though here should refer to the place where the speaker Andrew is, which is Los Angeles). This would create the incorrect misinterpretation. Please make sure to keep the person names and city names constant.",
+        "bbox": [
+            151,
+            455,
+            845,
+            662
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Also, make sure that the context starts exactly as \"Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles.\"",
+        "bbox": [
+            151,
+            664,
+            843,
+            695
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format:",
+        "bbox": [
+            151,
+            712,
+            845,
+            758
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "stimulussentence:stimulus stimulus sentence",
+        "bbox": [
+            151,
+            760,
+            502,
+            775
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "correct_context: correct prime context sentence",
+        "bbox": [
+            151,
+            778,
+            507,
+            791
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "wrong_context: incorrect prime context sentence",
+        "bbox": [
+            151,
+            794,
+            517,
+            808
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Prompt used to generate dataset samples for indexical 'here'.",
+        "bbox": [
+            260,
+            820,
+            734,
+            835
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "23419",
+        "bbox": [
+            475,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.",
+        "bbox": [
+            149,
+            181,
+            848,
+            229
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Here are examples:",
+        "bbox": [
+            152,
+            244,
+            299,
+            260
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            278,
+            235,
+            291
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris knows how dedicated I am to my studies and always encourages me to do my best. The final exams are coming up next week. Chris thinks that I will get good grades on the final exams.",
+        "bbox": [
+            151,
+            293,
+            845,
+            356
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris usually outperforms everyone in our classes and often brags about it. The final exams are coming up next week. Chris thinks that I will get good grades on the final exams.",
+        "bbox": [
+            151,
+            357,
+            843,
+            405
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sence, who will get good grades on the final exams?",
+        "bbox": [
+            152,
+            422,
+            801,
+            437
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            455,
+            235,
+            468
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris believes in my ability to communicate effectively and often praises my public speaking skills. There is an important conference coming up soon. Chris thinks that I will give a great presentation at the conference.",
+        "bbox": [
+            151,
+            470,
+            845,
+            533
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris is known for his exceptional presentation skills and often receives awards for them. There is an important conference coming up soon. Chris thinks that I will give a great presentation at the conference.",
+        "bbox": [
+            151,
+            533,
+            845,
+            598
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sence, who will give a great presentation at the conference?",
+        "bbox": [
+            152,
+            615,
+            845,
+            631
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Now, for the following sentences, you should create a question whose answer can only be either Chris or Andrew. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always start with \"According to the last sentence, who\"",
+        "bbox": [
+            151,
+            646,
+            845,
+            709
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            728,
+            235,
+            741
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "{sentences}",
+        "bbox": [
+            152,
+            744,
+            240,
+            759
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Question:",
+        "bbox": [
+            152,
+            777,
+            228,
+            791
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Prompt used to generate question for samples containing indexical 'I'.",
+        "bbox": [
+            230,
+            804,
+            766,
+            819
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "23420",
+        "bbox": [
+            475,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.",
+        "bbox": [
+            149,
+            173,
+            845,
+            221
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Here are examples:",
+        "bbox": [
+            152,
+            237,
+            297,
+            253
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            269,
+            235,
+            282
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. You and Chris attended the same conference last month, and you even had a conversation there. While John was speaking to Travis, he said that Chris saw you at the conference.",
+        "bbox": [
+            151,
+            285,
+            845,
+            349
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Travis and Chris attended the same conference last month, and they even had a conversation there. While John was speaking to Travis, he said that Chris saw you at the conference.",
+        "bbox": [
+            151,
+            350,
+            843,
+            414
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sentence, who did John say that Chris saw at the conference?",
+        "bbox": [
+            152,
+            430,
+            840,
+            445
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            463,
+            235,
+            476
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. You and Chris are both art enthusiasts and attended the same art gallery event last weekend, where you ran into each other. While John was speaking to Travis, he said that Chris noticed you at the art gallery.",
+        "bbox": [
+            151,
+            478,
+            843,
+            543
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Travis and Chris are both art enthusiasts and attended the same art gallery event last weekend, where they ran into each other. While John was speaking to Travis, he said that Chris noticed you at the art gallery.",
+        "bbox": [
+            151,
+            543,
+            843,
+            608
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sentence, who did John say that Chris noticed at the art gallery?",
+        "bbox": [
+            152,
+            623,
+            845,
+            640
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Now, for the following sentences, you should create a question whose answer can only be either Travis or Donald. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always start with \"According to the last sentence, who did John say that Chris\"",
+        "bbox": [
+            151,
+            655,
+            843,
+            720
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            736,
+            235,
+            750
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "{sentences}",
+        "bbox": [
+            152,
+            753,
+            238,
+            766
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            785,
+            228,
+            799
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Prompt used to generate question for samples containing indexical 'you'.",
+        "bbox": [
+            220,
+            812,
+            774,
+            828
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "23421",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.",
+        "bbox": [
+            149,
+            140,
+            848,
+            189
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Here are examples:",
+        "bbox": [
+            152,
+            205,
+            299,
+            221
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            237,
+            235,
+            250
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They are both academics who love to participate in international conferences. Recently, I've been telling them about the exciting academic events happening right here in Los Angeles. When I was in New York with John, he said that Chris wanted to attend a conference here.",
+        "bbox": [
+            151,
+            253,
+            845,
+            332
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They are both academics who love to participate in international conferences. Recently, they realized they haven't attended many conferences in New York, which is quite surprising given their love for the city. When I was in New York with John, he said that Chris wanted to attend a conference here.",
+        "bbox": [
+            151,
+            332,
+            843,
+            414
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sentence, where does Chris want to attend a conference?",
+        "bbox": [
+            152,
+            430,
+            806,
+            445
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            462,
+            235,
+            476
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They both have a deep appreciation for art, and recently I've been telling them about the vibrant art scene here in Los Angeles that offers great opportunities for new gallery openings. When I was in New York with John, he said that Chris wanted to open an art gallery here.",
+        "bbox": [
+            151,
+            478,
+            845,
+            558
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They both have a deep appreciation for art, and recently they realized they haven't opened a gallery in New York yet, despite its renowned art scene, which is surprising given their passion. When I was in New York with John, he said that Chris wanted to open an art gallery here.",
+        "bbox": [
+            151,
+            558,
+            845,
+            639
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sentence, where does Chris want to open an art gallery?",
+        "bbox": [
+            152,
+            655,
+            801,
+            671
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Now, for the following sentences, you should create a question whose answer can only be either Los Angeles or New York. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always start with \"According to the last sentence, where does Chris want to\".",
+        "bbox": [
+            151,
+            687,
+            845,
+            751
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            768,
+            235,
+            782
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "{sentences}",
+        "bbox": [
+            152,
+            785,
+            238,
+            800
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Question:",
+        "bbox": [
+            152,
+            816,
+            228,
+            832
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 7: Prompt used to generate question for samples containing indexical 'here'.",
+        "bbox": [
+            218,
+            844,
+            776,
+            860
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "23422",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.",
+        "bbox": [
+            149,
+            156,
+            845,
+            204
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Here are examples:",
+        "bbox": [
+            152,
+            221,
+            299,
+            236
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            253,
+            235,
+            266
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is meticulous about arranging all his necessary documentation well in advance and likes having everything sorted before deadlines approach. When we spoke last summer, John mentioned that Chris would be having his visa appointment tomorrow.",
+        "bbox": [
+            151,
+            269,
+            843,
+            332
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is quite disorganized and often waits until the last possible moment to schedule important tasks such as visa appointments. When we spoke last summer, John mentioned that Chris would be having his visa appointment tomorrow.",
+        "bbox": [
+            151,
+            334,
+            843,
+            397
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sentence, did Chris already have his visa appointment, or is he going to do so in the future?",
+        "bbox": [
+            151,
+            414,
+            843,
+            445
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            462,
+            235,
+            475
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is very disciplined and likes to plan his activities ahead of time. When we spoke last summer, John told me that Chris was planning go to the bungee jumping event tomorrow.",
+        "bbox": [
+            151,
+            478,
+            843,
+            539
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris likes trying different activities but he is very bad at planning and he usually plans his stuff at the very last moment. When we spoke last summer, John told me that Chris was planning go to the bungee jumping event tomorrow.",
+        "bbox": [
+            151,
+            543,
+            843,
+            606
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Question: According to the last sentence, did Chris already go to the bungee jumping event, or is he going to do so in the future?",
+        "bbox": [
+            151,
+            623,
+            843,
+            653
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Now, for the following sentences, you should create a question in the given format. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always structure the sentence as \"According to the last sentence, did Chris already ... or is he going to do so in the future?\"",
+        "bbox": [
+            151,
+            671,
+            843,
+            734
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Sentences:",
+        "text_level": 1,
+        "bbox": [
+            152,
+            752,
+            235,
+            765
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "{sentences}",
+        "bbox": [
+            152,
+            769,
+            238,
+            783
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Question:",
+        "bbox": [
+            152,
+            801,
+            226,
+            815
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Prompt used to generate question for samples containing indexical 'tomorrow'.",
+        "bbox": [
+            200,
+            828,
+            794,
+            843
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "23423",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/ebf1531ac6d17868b854555416f1bd27a6ac339902b7de52299255fa94851a6d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam knows how dedicated I am to my studies and always encourages me to do my best. The final exams are coming up next week. Adam thinks that I will get good grades on the final exams.\nQuestion: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Non-shifted</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam usually outperforms everyone in our classes and often brags about it. The final exams are coming up next week. Adam says that I will get good grades on the final exams.\nQuestion: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Non-shifted</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam knows how dedicated I am to my studies and always encourages me to do my best. The final exams are coming up next week. Adam says &quot;I will get good grades on the final exams&quot;.\nQuestion: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Shifted</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam usually outperforms everyone in our classes and often brags about it. The final exams are coming up next week. Adam says &quot;I will get good grades on the final exams&quot;.\nQuestion: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Shifted</td></tr></table>",
+        "bbox": [
+            157,
+            209,
+            840,
+            763
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Dataset Samples for Indexical \"I\"",
+        "bbox": [
+            354,
+            772,
+            640,
+            785
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "23424",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "table",
+        "img_path": "images/7058bb8b5845c4e9a6ae79c64848301fe502692194c6e9362e55a1c0d9fc1f44.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. You and Gregory attended the same conference last month, and you even had a conversation there. While Dylan was speaking to Samuel, he said that Gregory saw you at the conference.\nQuestion: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Non-shifted</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. Samuel and Gregory attended the same conference last month, and they even had a conversation there. While Dylan was speaking to Samuel, he said that Gregory saw you at the conference.\nQuestion: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Non-shifted</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. You and Gregory attended the same conference last month, and you even had a conversation there. While Dylan was speaking to Samuel, he said &quot;Gregory saw you at the conference&quot;.\nQuestion: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Shifted</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. Samuel and Gregory attended the same conference last month, and they even had a conversation there. While Dylan was speaking to Samuel, he said &quot;Gregory saw you at the conference&quot;.\nQuestion: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Shifted</td></tr></table>",
+        "bbox": [
+            157,
+            178,
+            838,
+            795
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Table 10: Dataset Samples for Indexical \"You\"",
+        "bbox": [
+            337,
+            804,
+            657,
+            819
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "page_number",
+        "text": "23425",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "table",
+        "img_path": "images/57548361fa95178343867a8a64e914c00b2d336e3fb33c451395908dba6987cf.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deborah and Jennifer. They are both academics who love to participate in international conferences. Recently, I’ve been telling them about the exciting academic events happening right here in Denver. When I was in Milwaukee with Deborah, she said that Jennifer wanted to attend a conference here. \nQuestion: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Non-shifted</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deb- orah and Jennifer. They are both academics who love to participate in international conferences. Recently, they realized they haven’t attended many conferences in Milwaukee, which is quite surprising given their love for the city. When I was in Milwaukee with Deborah, she said that Jennifer wanted to attend a conference here. \nQuestion: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Non-shifted</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deborah and Jennifer. They are both academics who love to participate in international conferences. Recently, I’ve been telling them about the exciting academic events happening right here in Denver. When I was in Milwaukee with Deborah, she said &quot;Jennifer wants to attend a conference here&quot;. \nQuestion: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Shifted</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deb- orah and Jennifer. They are both academics who love to participate in international conferences. Recently, they realized they haven’t attended many conferences in Milwaukee, which is quite surprising given their love for the city. When I was in Milwaukee with Deb- orah, she said &quot;Jennifer wants to attend a conference here&quot;. \nQuestion: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Shifted</td></tr></table>",
+        "bbox": [
+            139,
+            128,
+            858,
+            843
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "text",
+        "text": "Table 11: Dataset Samples for Indexical \"Here\"",
+        "bbox": [
+            336,
+            852,
+            658,
+            866
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "page_number",
+        "text": "23426",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 16
+    },
+    {
+        "type": "table",
+        "img_path": "images/06a274f183348d518da062e590c1588b032cd32763356fcbd0a06107a3d6b216.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is meticulous about arranging all his necessary documentation well in advance and likes having everything sorted before deadlines approach. When we spoke last summer, Donald mentioned that Carl would be having his visa appointment tomorrow. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Non-shifted</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is quite disorganized and often waits until the last possible moment to schedule important tasks such as visa appointments. When we spoke last summer, Donald mentioned that Carl would be having his visa appointment tomorrow. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Non-shifted</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is meticulous about arranging all his necessary documentation well in advance and likes having everything sorted before deadlines approach. When we spoke last summer, Donald said &quot;Carl will be having his visa appointment tomorrow&quot;. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Shifted</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is quite disorganized and often waits until the last possible moment to schedule important tasks such as visa appointments. When we spoke last summer, Donald said &quot;Carl will be having his visa appointment tomorrow&quot;. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Shifted</td></tr></table>",
+        "bbox": [
+            115,
+            142,
+            890,
+            826
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "text",
+        "text": "Table 12: Dataset Samples for Indexical \"Tomorrow\"",
+        "bbox": [
+            317,
+            835,
+            678,
+            851
+        ],
+        "page_idx": 17
+    },
+    {
+        "type": "page_number",
+        "text": "23427",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 17
+    }
+]

2025/Un-considering Contextual Information_ Assessing LLMs’ Understanding of Indexical Elements/1f95b7d5-2289-4e70-9979-ad44a236414a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5e825f5f3f47157c314de1e000648768c7ff84ef6e731b8f1a494b925333560
+size 605094

2025/Un-considering Contextual Information_ Assessing LLMs’ Understanding of Indexical Elements/full.md

@@ -0,0 +1,395 @@
+# Un-considering Contextual Information: Assessing LLMs' Understanding of Indexical Elements
+
+Metehan Oğuz Yavuz Bakman Duygu Nur Yaldiz
+
+University of Southern California
+
+{moguz, ybakman, yaldiz}@usc.edu
+
+# Abstract
+
+Large Language Models (LLMs) have demonstrated impressive performances in tasks related to coreference resolution. However, previous studies mostly assessed LLM performance on coreference resolution with nouns and third person pronouns. This study evaluates LLM performance on coreference resolution with indexicals like  $I$ , you, here and tomorrow, which come with unique challenges due to their linguistic properties. We present the first study examining how LLMs interpret indexicals in English, releasing the English Indexical Dataset with 1600 multiple-choice questions. We evaluate pioneering LLMs, including GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, and DeepSeek V3. Our results reveal that LLMs exhibit an impressive performance with some indexicals ( $I$ ), while struggling with others (you, here, tomorrow), and that syntactic cues (e.g. quotation) contribute to LLM performance with some indexicals, while they reduce performance with others. Code and data are available at: https://github.com/metehanoguzz/LLMs-Indexicals-English
+
+# 1 Introduction
+
+Large Language Models (LLMs) have demonstrated remarkable capabilities in zero-shot and few-shot learning, excelling across a wide array of tasks such as machine translation, text summarization, and question answering (OpenAI, 2024; Ye et al., 2023; Bakman et al., 2024; Yaldiz et al., 2024). Their versatility has led to widespread applications in diverse domains, including education, law, and medicine.
+
+As the use of LLMs continues to expand, understanding their underlying behaviors has become increasingly important. Recent studies have evaluated the performance of large language models on linguistic tasks such as coreference resolution (Gan et al., 2024; Le and Ritter, 2023; Brown et al., 2020; Yang et al., 2022; Agrawal et al., 2022).
+
+Previous work on coreference resolution mostly focused on how coreference is established between two third person entities such as proper names (e.g. Andy, the mechanic) and third person pronouns (e.g. he, him, himself) in English and other languages (e.g. Yang et al., 2022; Yang, 2025). In this study, we investigate how LLMs establish coreference with indexical elements (e.g. I, you, here), which differ from third person nouns/pronouns in substantial ways and bring unique challenges for LLMs (see Figure 1 as an example). We investigate how state-of-the-art LLMs interpret the indexical elements I, you, here and tomorrow in English sentences, and whether context or grammatical constraints influence their decisions. To the best of our knowledge, this is the first study examining LLMs' handling of indexical elements in English. Our key contributions are as follows:
+
+- We introduce the English Indexical Dataset, comprising 400 interpretation samples for each indexical element, I, you, here, tomorrow, totaling 1,600 instances.  
+- We evaluate the performance of four frontier LLMs, GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, and DeepSeek-V3, on the interpretation of indexical elements in English.  
+- We show that LLM performances are not uniform across different types of indexical elements: indexical  $I$  is successfully interpreted by most LLMs, other indexicals like you, here and tomorrow lead to poor performances.  
+- We show that quotation affects LLMs performances differently: quotation reduces LLMs' accuracies with tomorrow, but it increases accuracies with here.
+
+# 2 Indexical elements
+
+Indexical pronominals like  $I$ , you, here and now are used to refer to referents of the speech-act coordinates (e.g. Kaplan, 1977; Schlenker, 2003). For instance,  $I$  refers to author (speaker) of the utterance,
+
+![](images/c5bb29204fd67291193cbcc99089db521750b0c4ff01b9bc1be4a307377e206a.jpg)  
+Figure 1: An example for LLM misinterpreting indexical element 'here', uttered by a speaker in Los Angeles.
+
+![](images/aa12c66ee0f6e7c874b38d3f7d630d12dbc293cba44faf1b8e189e4f7d95c15a.jpg)  
+Of course! You should go to Central Park and Statue of Liberty. I am sure he will love them!
+
+![](images/bf8b79feac680a65fe40a94bf26bac804c021beb71066efc8343863289654e75.jpg)
+
+while here refers to the location where the utterance was made. Thus, a sentence like (1) means different things if uttered by different people and/or in different places. If John utters (1-a) in Los Angeles, it means that John was born in Los Angeles, but if Mary utters the same sentence in New York it means that Mary was born in New York.
+
+(1) a. I was born here.  
+b. Andrew said that I went to Buckhead.
+
+Indexicals are interpreted inside the context of utterance, referring to the actual speech-act coordinates like the author or the location of the actual utterance. As a result, if (1-b) is uttered by John, the indexical  $I$  can only be interpreted as referring to John as the speaker, leading to a reading like 'Andrew said that John went to Buckhead'. Crucially, even though Andrew's speech/claim is reported in (1-b), the indexical  $I$  cannot refer to Andrew.
+
+Direct quotation is an exception to this generalization, where reported material is interpreted as verbatim utterance/thoughts of its owner. Thus, when indexicals appear inside direct quotation, they are interpreted inside the reported context, rather than the actual context of utterance. In other words, direct quotation 'shifts' the interpretations of indexicals into the reported context. For example,  $I$  and here in (2) appear inside direct quotation, where Andrew's speech is reported.
+
+(2) While we were in Atlanta, Andrew said "I was born here."
+
+Regardless of who utters (2), the sentence means that Andrew was born in Atlanta, so both  $I$  and here are 'shifted' into the reported context, where Andrew is the speaker and Atlanta is the location<sup>1</sup>
+
+Indexicals differ from other pronominals in substantial ways. First, though syntactic and semantic factors can affect how a pronoun is interpreted (e.g. subject bias), pronouns are typically ambiguous regarding what/who they refer to. For example, the third person pronoun he in (3) is most naturally interpreted as referring to the subject John (for syntactic or contextual reasons), but the object Bill is
+
+still a possible antecedent, causing ambiguity between two different readings (e.g. Crawley et al., 1990; Stewart and Pickering, 1998; Pickering and Majid, 2007). In addition, he can refer to any contextually salient person that is not mentioned in the sentence (e.g. Peter), which makes pronouns even more ambiguous and context-dependent.
+
+(3) John hit Bill and he ran away.
+
+As a result, semantic/contextual information plays a crucial role in how pronouns are interpreted, and speakers use those cues to establish coreference with pronouns. For instance, if (4) is uttered in a context where John is a supportive and humble coworker, the most natural interpretation is that John suggests that Bill should get promoted ( $he = \text{Bill}$ ). However, if John is arrogant and jealous, the most natural interpretation is that John suggests that John should get promoted ( $he = \text{John}$ ).
+
+(4) John told Bill that he should get promoted.
+
+Indexicals, on the other hand, unambiguously refer to the referents of the speech-act coordinates. For example,  $I$  in (5) refers to the speaker regardless of what we know about John or Bill.  $I$  refers to the speaker even if John and/or Bill are arrogant and jealous, so contextual information like this should be disregarded while interpreting indexicals.
+
+(5) John told Bill that I should get a promotion.
+
+In summary, indexicals are restricted by different syntactic factors than pronouns (e.g. quotation vs non-quotation) and are typically unambiguous, while pronouns are free to refer to a wide range of entities. Thus, indexical elements create a unique challenge for LLMs, requiring to 'disregard' semantic/contextual cues that might prime interpretations through other antecedents (unlike pronouns).
+
+# 3 Experimental design
+
+# 3.1 Dataset Curation
+
+We curated a dataset specifically designed to test how LLMs interpret indexical elements like  $I$ , you, here, and tomorrow in different contexts of utterance. Specifically, we assess how these mod
+
+els interpret indexicals in 'shifted' context prime, where context more naturally requires the indexical should be interpreted inside reported context (e.g. Peter is one of the most arrogant students in my classroom. ... Peter says that I am smart.) vs 'non-shifted' context prime, where context more naturally requires the indexical should be interpreted inside actual speech context (e.g. Peter is very kind and supportive. ... Peter says that I am smart.). We also included direct quotations in both contexts (e.g., Peter says, "I am smart") to examine if LLMs can successfully consider syntactic factors (quotation vs regular sentences) while ignoring misleading information from the context during coreference resolution with indexical elements.
+
+For each type of indexical, we design 100 sentences and for each sentence we apply the four different transformations explained above. Overall, we have 400 samples per indexical, compromising a total dataset of size 1600.
+
+We utilize GPT-4o to curate the dataset, by giving a detailed description of the task along with some in-context examples. Then it is asked to generate scenarios with a stimulus sentence in two different contexts (See Appendix A.1 for the prompts used), along with specific questions addressing the referent of the indexical in each sentence. To ensure the quality of the dataset,  $25\%$  of the dataset (400 trials = 100 sentences in four conditions) was randomly selected for evaluation, and the evaluation process consisted of three steps.
+
+In the first step, we confirmed that all sentences were grammatically correct. In the second step, we confirmed that all quotation condition sentences had quoted embedded clauses, and all non-quoted condition sentences had regular (non-quoted) embedded clauses. We also made sure that the two sentence conditions were maximally similar, except for the quotation vs non-quoted status (i.e. the only difference between two sentence conditions was the quotation). In the third step, we checked the context prime texts for each condition in each sentence, making sure that the correct readings (shifted vs non-shifted) were primed by the context description. For example, for an item condition where here was supposed to be shifted, we confirmed that the context description would be most naturally followed by a sentence where here would be shifted.
+
+100 trials from each indexical item (25% for each indexical), 400 trials in total, were randomly selected to make sure that the evalua
+
+tion/confirmation was representative of all indexical item conditions (i.e. items with  $I$ , you, here and tomorrow).
+
+To eliminate potential gender bias, each dataset sample exclusively uses either male or female names, alternating to ensure a balanced distribution with  $50\%$  of the samples containing female names and  $50\%$  male names. This method promotes gender neutrality across the dataset. Sample details are in Appendix A.2, and the complete dataset is available in the supplementary materials.
+
+# 3.2 Models
+
+In our evaluation, we utilize four recent state-of-the-art LLMs: GPT-4o (OpenAI, 2024), Claude 3.5 Sonnet (Anthropic), Gemini 1.5 Pro (Team, 2024), and DeepSeek-V3 (DeepSeek-AI, 2024). This selection of diverse models provides a comprehensive evaluation of LLM performance with indexicals.
+
+# 3.3 Evaluation Strategy
+
+To assess the performance of the model, we specifically prompt it to answer questions designed to test its capabilities as described in Section 3.1. Additionally, to ensure focused responses, we restrict the model's answers to one of two predefined options: the 'shifted' option and the 'non-shifted' option. We provide the prompt in Appendix B.1.
+
+# 3.4 Metrics.
+
+We assess model accuracy across four cases for each indexical: (i) Non-quoted sentences with shifted context prime, (ii) Non-quoted with non-shifted prime, (iii) Quoted with shifted prime, and (iv) Quoted with non-shifted prime. Optimal performance would be achieved by always selecting the 'shifted' option in quoted conditions and selecting the 'non-shifted' option in non-quoted conditions.
+
+# 4 Experimental Results
+
+We present the experimental results in Figure 2 and we discuss them in detail for each indexical type:
+
+Indexical  $I$ . The results show that all four models perform near optimum in sentences without quotation with an average accuracy of  $99\%$  (always correctly selecting the non-shifted option). For the quoted sentences GPT-4o and Gemini 1.5 pro again perform almost optimum with a mean accuracy larger than  $94\%$ , followed by Claude 3.5 Sonnet with  $89\%$  mean accuracy (correctly selecting the
+
+![](images/f9f419f787417918e7e187d65e9fba79b67a9039f2aa210287c23937bcc167b3.jpg)  
+Figure 2: From left to right: Performance analysis plot of for the indexical 'I', Performance analysis plot of for the indexical 'you', Performance analysis plot of for the indexical 'here', Performance analysis plot of for the indexical 'tomorrow'. Dark blue bars = Shifting context prime, Light blue bars = Non-shifted context prime.
+
+![](images/3544abbbfbd62f1b1b985d3bdfa599681da28419bfd7f01981e98bf88ac38fb0.jpg)
+
+![](images/919f33cfee1f13d302aa79c95ff2cdd18e8be8b4288bf22f380f38d4872eeece.jpg)
+
+![](images/284623814d1074d170622f6c183400e19b1b08976495384e288fa51665fca8ce.jpg)
+
+shifted option). However, DeepSeek V3 fails to always select the shifted option on quoted sentences. Especially, the model performance drops significantly to  $17\%$  when context primes non-shifted readings, suggesting that inclusion of quotation makes the model more sensitive to the linguistically irrelevant effects of context prime. Moreover, considering that the model performance only reaches  $78\%$  accuracy in quotation conditions seems to imply that the model might have a bias towards the non-shifted reading overall, reducing model performance in quotation conditions.
+
+Indexical you. The results indicate that LLMs generally perform worse with the indexical you than with  $I$ . We see similar patterns across models, suggesting they perform mostly at similar levels. All models are sensitive to the effects of context prime, performing lower when the context primes the incorrect option (i.e. shifted reading in non-quotation and non-shifted reading in quotation). Moreover, quotation consistently results in lower performance across models. Notably, Gemini 1.5 Pro excels in non-quotation accuracy (92%), though its performance drops significantly under quotation conditions to the levels of other models. Overall, the results suggest that models interpret you based on the linguistically irrelevant context prime rather than the sentence type, which indeed determines the correct readings of the indexicals.
+
+Indexical here. The results show that LLMs, similar to you, struggle to interpret the indexical here, especially when context primes the incorrect option, leading to poor performance. However, different from you, quotation leads to higher performance with here. In non-quotation conditions, all LLMs make their selections almost exclusively based on context prime, with shifted primes showing over  $96\%$  accuracy and non-shifted primes less than  $2\%$ , where context prime ideally should not
+
+have any effects on the selection. In contrast, under quotation conditions, the influence of context prime diminishes, leading to higher performances. Particularly, DeepSeek V3 and Gemini 1.5 Pro exhibit impressive performance with accuracies above  $97\%$  and  $94\%$ , respectively, followed by Gemini 1.5 Pro and GPT-4o, with accuracies above  $64\%$  and  $37\%$ .
+
+Indexical tomorrow. The results indicate that LLMs have a strong bias towards the non-shifted interpretations of tomorrow. We see that the models almost always select the non-shifted option for trials with tomorrow, regardless of context prime or sentence type. While Claude 3.5 Sonnet and Gemini 1.5 Pro selects the non-shifted options  $100\%$  of the time, GPT-4o and DeepSeek V3 select the non-shifted option  $94\%$  and  $83\%$  of the time, respectively. This strong bias leads to illusory high accuracies in non-quotation conditions, while causes extremely low performance in quotation conditions.
+
+# 5 Related Work
+
+Coreference resolution has been extensively studied in prior research (Gan et al., 2024; Le and Ritter, 2023; Brown et al., 2020; Yang et al., 2022; Agrawal et al., 2022). However, indexical elements exhibit distinct syntactic properties compared to other (non-indexical) pronominals, as discussed in Section 2. Previous work by Oğuz et al. (2024) explored how LLMs interpret indexicals in Turkish, where indexicals possess different grammatical properties than in English and can shift without quotation. To the best of our knowledge, this study is the first to evaluate the performance of LLMs in interpreting indexicals in English.
+
+# 6 Discussion and Conclusion
+
+Our results show that LLM performances are not uniform across different types of indexical elements and sentence types. While most LLMs
+
+perform at impressive levels interpreting the indexical  $I$ , their performances drop significantly in other indexicals, particularly in here and tomorrow. Moreover, though sentence type (quotation vs non-quadation) affects LLMs performances in general, the effects are not similar across different indexical types and models. While quotation increases LLMs' performance with here, it decreases their performance with you and tomorrow. In addition, tomorrow seems to be affected by quotation in a greater magnitude than you. In conclusion, we find that different types of indexicals show unique patterns regarding how they are interpreted by the LLMs.
+
+Our results diverge from those reported in Oğuz et al. (2024), who tested how first person indexical in Turkish ben 'I' was interpreted by different LLMs, including GPT-4o and show that LLMs exhibit very poor performance interpreting ben 'I'. Here, we report that LLMs perform at an almost human-like level with interpreting the English indexical  $I$ . This might be due to lower amounts of available resources to train the models in Turkish, compared to English. Another reason for low performance in Turkish could be due to the fact that Turkish is a pro-drop language, meaning that the subject pronouns can be dropped (silent). Oğuz et al. (2024) used sentences where the first person indexical ben 'I' was dropped, which might have made the task more challenging for the LLMs considering that dropped indexicals can show different properties than overt ones (Oğuz et al., 2020). This linguistic difference between Turkish and English might have caused different results between the Turkish and English tests.
+
+# 7 Limitations
+
+Our work explores how LLMs interpret indexical elements in a black-box setting but does not provide experimental analyses that investigate the underlying reasons for these behaviors by examining the models' internals or training data. Future research could adopt a white-box approach to analyze these behaviors in greater depth, offering valuable insights into the mechanisms driving LLMs' interpretation of indexicals.
+
+# References
+
+Monica Agrawal, Stefan Hegselmann, Hunter Lang, Yoon Kim, and David A. Sontag. 2022. Large language models are few-shot clinical information ex
+
+tractors. In Conference on Empirical Methods in Natural Language Processing.  
+Anthropic. The claude 3 model family: Opus, sonnet, haiku.  
+Yavuz Faruk Bakman, Duygu Nur Yaldiz, Baturalp Buyukates, Chenyang Tao, Dimitrios Dimitriadis, and Salman Avestimehr. 2024. MARS: Meaning-aware response scoring for uncertainty estimation in generative LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7752-7767, Bangkok, Thailand. Association for Computational Linguistics.  
+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, Jeff 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. *ArXiv*, abs/2005.14165.  
+Rosalind A. Crawley, Rosemary J. Stevenson, and David Kleinman. 1990. The use of heuristic strategies in the interpretation of pronouns. Journal of Psycholinguistic Research, 14.  
+Amy Rose Deal. 2020. A theory of indexical shift: meaning, grammar, and crosslinguistic variation. MIT Press, Boston, MA.  
+DeepSeek-AI. 2024. Deepseek-v3 technical report. Preprint, arXiv:2412.19437.  
+Yujian Gan, Massimo Poesio, and Juntao Yu. 2024. Assessing the capabilities of large language models in coreference: An evaluation. In International Conference on Language Resources and Evaluation.  
+David Kaplan. 1977. Demonstratives: An essay on the semantics, logic, metaphysics, and epistemology of demonstratives and other indexicals. *Themes from Kaplan*, pages 565-614.  
+Nghia T. Le and Alan Ritter. 2023. Are large language models robust coreference resolvers?  
+Metehan Oğuz, Yusuf Ciftci, and Yavuz Faruk Bakman. 2024. Do LLMs recognize me, when I is not me: Assessment of LLMs understanding of Turkish indexical pronouns in indexical shift contexts. In Proceedings of the First Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2024), pages 53-61, Bangkok, Thailand and Online. Association for Computational Linguistics.  
+OpenAI. 2024. Gpt-4 technical report. Preprint, arXiv:2303.08774.
+
+Metehan Oğuz, Burak Öney, and Dennis Ryan Storoshenko. 2020. Obligatory indexical shift in Turkish. In Proceedings of Canadian Linguistic Association (CLA), Western University, London, ON, Canada.  
+Martin Pickering and Asifa Majid. 2007. What are implicit causality and consequentiality? Language & Cognitive Processes, 22.  
+Philippe Schlenker. 2003. A plea for monsters. Linguistics and Philosophy, 26:29-120.  
+Andrew J. Stewart and Martin Pickering. 1998. Implicit consequentiality. In Proceedings of the 20th Annual Conference of the Cognitive Science Society.  
+Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. Preprint, arXiv:2403.05530.  
+Duygu Nur Yaldiz, Yavuz Faruk Bakman, Baturalp Buyukates, Chenyang Tao, Anil Ramakrishna, Dimitrios Dimitriadis, Jieyu Zhao, and Salman Avestimehr. 2024. Do not design, learn: A trainable scoring function for uncertainty estimation in generative llms. Preprint, arXiv:2406.11278.  
+Xiaohan Yang, Eduardo Peynetti, Vasco Meerman, and Christy Tanner. 2022. What gpt knows about who is who. In First Workshop on Insights from Negative Results in NLP.  
+Xiulin Yang. 2025. Language models at the syntax-semantics interface: A case study of the long-distance binding of Chinese reflexive ziji. In Proceedings of the 31st International Conference on Computational Linguistics, pages 3808-3824, Abu Dhabi, UAE. Association for Computational Linguistics.  
+Junjie Ye, Xuanting Chen, Nuo Xu, Can Zu, Zekai Shao, Shichun Liu, Yuhan Cui, Zeyang Zhou, Chao Gong, Yang Shen, Jie Zhou, Siming Chen, Tao Gui, Qi Zhang, and Xuanjing Huang. 2023. A comprehensive capability analysis of GPT-3 and GPT-3.5 series models. Preprint, arXiv:2303.10420.
+
+# A Dataset Details
+
+# A.1 Dataset Generation Prompts
+
+We provide the prompts used for data generation in Tables 1, 2, 3, and 4. The prompts used for question generation per scenario is presented in Tables 5, 6, 7, and 8.
+
+# A.2 Samples From the Dataset
+
+We provide samples from the dataset for each indexical element we investigate in Tables 9, 10, 11, and 12.
+
+# B Experimental Details
+
+# B.1 Prompt Used in Evaluation
+
+The prompt we employ for the evaluation is as follows:
+
+Read the following passage carefully and answer the question at the end: {stimuli} {question}
+
+Please provide your answer as: either {option1} or {option2}. Do not include any additional explanation or text.
+
+I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:
+
+Details for stimulus sentences: Each stimulus sentence will have a structure like "While John was speaking to Travis, he said that Chris appreciates you a lot." Please make sure to use the names John, Travis, and Chris. Make sure that Chris is the subject of the embedded clause, and make sure that the sentence begins as "While John was speaking to Travis...". Please use different embedded verbs. For example, rather than appreciates you a lot, you can use saw you at the market, etc. But the action should be done by Chris, and the object should be "you".
+
+Details about the context descriptions: The contexts will prime how "you" in the stimulus is interpreted. To manipulate this, I will give contexts in which "you" would refer to my addressee, which is you. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if "you" referred to John's addressee. However, since "you" means the current addressee, this will be a misinterpretation. For example, a correct reading prime would be like "Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris asked for help from you a few times in the past, and you always helped him." This context makes it sound like Chris would appreciate you for all your help, and thus it would not be surprising if Chris appreciated you (Donald). However, in the incorrect interpretation prime, I will use a context like "Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris asked for help from Travis a few times in the past, and Travis always helped him." In this context, it would be more natural if Chris appreciated Travis, who helped him, instead of you (Donald), and thus would be natural if "you" in the sentence was interpreted as referring to Travis (though "you" should refer to my addressee, which is you (Donald)). This would create the incorrect misinterpretation. Please make sure to keep the person names constant. Also, make sure that the context starts exactly as "Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles."
+
+These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format:
+
+stimulussentence:stimulus stimulus sentence
+
+correct_context: correct prime context sentence
+
+wrong_context: incorrect prime context sentence
+
+Table 1: Prompt used to generate dataset samples for indexical 'you'.
+
+I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:
+
+Details for stimulus sentences: Each stimulus sentence will have a structure like "When we spoke last summer, John said that Chris was going to go to Greece tomorrow." Please make sure to use the names John, and Chris. Make sure that Chris is the subject of the embedded clause. Please use different embedded verbs. For instance, rather than go to Greece, you can use have a visa appointment, etc. But the action should be done "tomorrow".
+
+Details about the context descriptions: The contexts will prime how "tomorrow" in the stimulus is interpreted. To manipulate this, I will give contexts in which "tomorrow" would refer to the day after the actual (matrix) sentence is uttered. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if "tomorrow" referred to the day after John spoke. However, since "tomorrow" means the day after the current day, this will be a misinterpretation. For example, a correct reading prime would be like "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is very careful about planning everything, and always plans his stuff ahead of time." This context makes it sound like Chris would plan his trip to Greece ahead of time, and thus it would not be surprising if John said last summer that Chris was going to have a trip tomorrow. However, in the incorrect interpretation prime, I will use a context like "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is very lazy and never plans his stuff until the last moment." In this context, it would be more natural if Chris was going to go to Greece last summer, the day after John spoke to me, instead of the day after today, and thus would be natural if "tomorrow" in the sentence was interpreted as the day after John spoke to me (though "tomorrow" should refer to the day after today). This would create the incorrect misinterpretation. Please make sure to keep the person names constant. Also, make sure that the context starts exactly as "Hi! I am Andrew. I am a graduate student and I live in Los Angeles."
+
+These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format:
+
+stimulussentence:stimulus stimulus sentence
+
+correct_context: correct prime context sentence
+
+wrong_context: incorrect prime context sentence
+
+Table 2: Prompt used to generate dataset samples for indexical 'tomorrow'.
+
+I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:
+
+Details for stimulus sentences: Each stimulus sentence will have a structure like "Chris thinks that I will win the race." Please make sure to use the name Chris as the matrix subject, and "I" as the embedded subject subject. Please use different embedded verbs. For example, rather than win the race, you can use study hard, etc. But the action should be done by "I".
+
+Details about the context descriptions: The contexts will prime who "I" in the stimulus refers to. To manipulate this, I will give contexts in which "I" would refer to the speaker. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if "I" referred to Chris. However, since Chris is not the speaker, this will lead to an incorrect interpretation. For example, a correct reading prime would be like "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris is a supportive friend, and has always trusted in my abilities. There is a race next week." This context makes it sound like Chris would support the speaker in a race, and predict that the speaker would win the race. However, in the incorrect interpretation prime, I will use a context like "Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris is usually very competitive, and has bullied other people in front of me. There is a race next week..." In this context, it would be more natural if Chris thought that he would win the race, instead of the speaker Andrew, and thus would be natural if "I" in the sentence referred to Chris (though "I" should refer to the speaker Andrew). This would create the incorrect misinterpretation. Please make sure to keep the person names constant. Also, make sure that the context starts exactly as "Hi! I am Andrew. I am a graduate student and I live in Los Angeles."
+
+These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format: stimulusSentence: stimulus stimulus sentence correct_context: correct prime context sentence wrong_context: incorrect prime context sentence
+
+Table 3: Prompt used to generate dataset samples for indexical 'I'.
+
+I would like you to help me create a stimulus for my project. The stimuli will be English sentences. For each sentence, there will be two types of context description. One context description will prime the actual meaning of the sentence, but the other description will prime an incorrect reading of the sentence (like a misinterpretation). I want you to follow a structure while creating sentences and contexts. You can find more details below:
+
+Details for stimulus sentences: Each stimulus sentence will have a structure like "When I was in New York with John, he said that Chris wanted to explore here." Please make sure to use the city name New York for the sentence, and the names John and Chris. John will always be the person who says something that Chris will do. In each sentence, Chris will be the person who is doing something "here". Please use different verbs. For example, rather than explore here, you can use attend a conference here, etc.. Use various verbs. But the action should be done "here".
+
+Details about the context descriptions: The contexts will prime where "here" in the stimulus refers to. To manipulate this, I will give contexts in which "here" would refer to Los Angeles, where the author/speaker of the sentence is located. But, to prime the incorrect interpretation, I will give contexts that would naturally follow if "here" referred to New York. However, since the speaker is not in New York, this will lead to an incorrect interpretation.
+
+For example, a correct reading prime would be like "Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They love exploring new cities, and surprisingly they did not spend much time in Los Angeles." This context makes it sound like John and Chris would like to come to Los Angeles to explore around, and thus the "explore around here" in the stimulus sentence would be naturally understood as Los Angeles. However, in the incorrect interpretation prime, I will use a context like "Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They love exploring new cities, and surprisingly they did not spend much time in New York." In this context, it would be more natural if John and Chris would like to go to New York, and thus would be natural if "explore around here" would be exploring around New York (though here should refer to the place where the speaker Andrew is, which is Los Angeles). This would create the incorrect misinterpretation. Please make sure to keep the person names and city names constant.
+
+Also, make sure that the context starts exactly as "Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles."
+
+These are some examples you generated before: {previous_genres}. Use these examples as inspiration to spark creativity. Provide one new stimulus sentence, along with a corresponding correct prime context and an incorrect prime context, in the following format:
+
+stimulussentence:stimulus stimulus sentence
+
+correct_context: correct prime context sentence
+
+wrong_context: incorrect prime context sentence
+
+Table 4: Prompt used to generate dataset samples for indexical 'here'.
+
+I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.
+
+Here are examples:
+
+# Sentences:
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris knows how dedicated I am to my studies and always encourages me to do my best. The final exams are coming up next week. Chris thinks that I will get good grades on the final exams.
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris usually outperforms everyone in our classes and often brags about it. The final exams are coming up next week. Chris thinks that I will get good grades on the final exams.
+
+Question: According to the last sence, who will get good grades on the final exams?
+
+# Sentences:
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris believes in my ability to communicate effectively and often praises my public speaking skills. There is an important conference coming up soon. Chris thinks that I will give a great presentation at the conference.
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have a friend named Chris. Chris is known for his exceptional presentation skills and often receives awards for them. There is an important conference coming up soon. Chris thinks that I will give a great presentation at the conference.
+
+Question: According to the last sence, who will give a great presentation at the conference?
+
+Now, for the following sentences, you should create a question whose answer can only be either Chris or Andrew. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always start with "According to the last sentence, who"
+
+# Sentences:
+
+{sentences}
+
+Question:
+
+Table 5: Prompt used to generate question for samples containing indexical 'I'.
+
+I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.
+
+Here are examples:
+
+# Sentences:
+
+Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. You and Chris attended the same conference last month, and you even had a conversation there. While John was speaking to Travis, he said that Chris saw you at the conference.
+
+Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Travis and Chris attended the same conference last month, and they even had a conversation there. While John was speaking to Travis, he said that Chris saw you at the conference.
+
+Question: According to the last sentence, who did John say that Chris saw at the conference?
+
+# Sentences:
+
+Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. You and Chris are both art enthusiasts and attended the same art gallery event last weekend, where you ran into each other. While John was speaking to Travis, he said that Chris noticed you at the art gallery.
+
+Hi! I am Andrew. I will call you Donald. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Travis and Chris are both art enthusiasts and attended the same art gallery event last weekend, where they ran into each other. While John was speaking to Travis, he said that Chris noticed you at the art gallery.
+
+Question: According to the last sentence, who did John say that Chris noticed at the art gallery?
+
+Now, for the following sentences, you should create a question whose answer can only be either Travis or Donald. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always start with "According to the last sentence, who did John say that Chris"
+
+# Sentences:
+
+{sentences}
+
+# Question:
+
+Table 6: Prompt used to generate question for samples containing indexical 'you'.
+
+I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.
+
+Here are examples:
+
+# Sentences:
+
+Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They are both academics who love to participate in international conferences. Recently, I've been telling them about the exciting academic events happening right here in Los Angeles. When I was in New York with John, he said that Chris wanted to attend a conference here.
+
+Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They are both academics who love to participate in international conferences. Recently, they realized they haven't attended many conferences in New York, which is quite surprising given their love for the city. When I was in New York with John, he said that Chris wanted to attend a conference here.
+
+Question: According to the last sentence, where does Chris want to attend a conference?
+
+# Sentences:
+
+Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They both have a deep appreciation for art, and recently I've been telling them about the vibrant art scene here in Los Angeles that offers great opportunities for new gallery openings. When I was in New York with John, he said that Chris wanted to open an art gallery here.
+
+Hi! I am Andrew. I am a graduate student and I am studying in a cafe in Los Angeles. I have two friends named John and Chris. They both have a deep appreciation for art, and recently they realized they haven't opened a gallery in New York yet, despite its renowned art scene, which is surprising given their passion. When I was in New York with John, he said that Chris wanted to open an art gallery here.
+
+Question: According to the last sentence, where does Chris want to open an art gallery?
+
+Now, for the following sentences, you should create a question whose answer can only be either Los Angeles or New York. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always start with "According to the last sentence, where does Chris want to".
+
+# Sentences:
+
+{sentences}
+
+Question:
+
+Table 7: Prompt used to generate question for samples containing indexical 'here'.
+
+I would like you to help me create questions to the sentences I will give you for my project. The question will be English sentences. The goal is to create a question that will prime the correct interpretation of the indexical in the sentence.
+
+Here are examples:
+
+# Sentences:
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is meticulous about arranging all his necessary documentation well in advance and likes having everything sorted before deadlines approach. When we spoke last summer, John mentioned that Chris would be having his visa appointment tomorrow.
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is quite disorganized and often waits until the last possible moment to schedule important tasks such as visa appointments. When we spoke last summer, John mentioned that Chris would be having his visa appointment tomorrow.
+
+Question: According to the last sentence, did Chris already have his visa appointment, or is he going to do so in the future?
+
+# Sentences:
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris is very disciplined and likes to plan his activities ahead of time. When we spoke last summer, John told me that Chris was planning go to the bungee jumping event tomorrow.
+
+Hi! I am Andrew. I am a graduate student and I live in Los Angeles. I have two friends named John and Chris. Chris likes trying different activities but he is very bad at planning and he usually plans his stuff at the very last moment. When we spoke last summer, John told me that Chris was planning go to the bungee jumping event tomorrow.
+
+Question: According to the last sentence, did Chris already go to the bungee jumping event, or is he going to do so in the future?
+
+Now, for the following sentences, you should create a question in the given format. The question should be designed to prime the correct interpretation of the indexical in the sentence. Provide only the question and always structure the sentence as "According to the last sentence, did Chris already ... or is he going to do so in the future?"
+
+# Sentences:
+
+{sentences}
+
+Question:
+
+Table 8: Prompt used to generate question for samples containing indexical 'tomorrow'.
+
+<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam knows how dedicated I am to my studies and always encourages me to do my best. The final exams are coming up next week. Adam thinks that I will get good grades on the final exams.
+Question: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Non-shifted</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam usually outperforms everyone in our classes and often brags about it. The final exams are coming up next week. Adam says that I will get good grades on the final exams.
+Question: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Non-shifted</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam knows how dedicated I am to my studies and always encourages me to do my best. The final exams are coming up next week. Adam says &quot;I will get good grades on the final exams&quot;.
+Question: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Shifted</td></tr><tr><td>Hi! I am Stephen. I am a graduate student and I live in Los Angeles. I have a friend named Adam. Adam usually outperforms everyone in our classes and often brags about it. The final exams are coming up next week. Adam says &quot;I will get good grades on the final exams&quot;.
+Question: According to the last sentence, who will get good grades on the final exams?</td><td>Adam</td><td>Stephen</td><td>Shifted</td></tr></table>
+
+Table 9: Dataset Samples for Indexical "I"
+
+<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. You and Gregory attended the same conference last month, and you even had a conversation there. While Dylan was speaking to Samuel, he said that Gregory saw you at the conference.
+Question: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Non-shifted</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. Samuel and Gregory attended the same conference last month, and they even had a conversation there. While Dylan was speaking to Samuel, he said that Gregory saw you at the conference.
+Question: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Non-shifted</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. You and Gregory attended the same conference last month, and you even had a conversation there. While Dylan was speaking to Samuel, he said &quot;Gregory saw you at the conference&quot;.
+Question: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Shifted</td></tr><tr><td>Hi! I am Jerry. I will call you Ryan. I am a graduate student and I live in Los Angeles. I have two friends named Dylan and Gregory. Samuel and Gregory attended the same conference last month, and they even had a conversation there. While Dylan was speaking to Samuel, he said &quot;Gregory saw you at the conference&quot;.
+Question: According to the last sentence, who did Dylan say that Gregory saw at the conference?</td><td>Ryan</td><td>Samuel</td><td>Shifted</td></tr></table>
+
+Table 10: Dataset Samples for Indexical "You"
+
+<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deborah and Jennifer. They are both academics who love to participate in international conferences. Recently, I’ve been telling them about the exciting academic events happening right here in Denver. When I was in Milwaukee with Deborah, she said that Jennifer wanted to attend a conference here. 
+Question: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Non-shifted</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deb- orah and Jennifer. They are both academics who love to participate in international conferences. Recently, they realized they haven’t attended many conferences in Milwaukee, which is quite surprising given their love for the city. When I was in Milwaukee with Deborah, she said that Jennifer wanted to attend a conference here. 
+Question: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Non-shifted</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deborah and Jennifer. They are both academics who love to participate in international conferences. Recently, I’ve been telling them about the exciting academic events happening right here in Denver. When I was in Milwaukee with Deborah, she said &quot;Jennifer wants to attend a conference here&quot;. 
+Question: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Shifted</td></tr><tr><td>Hi! I am Lisa. I am a graduate student and I am studying in a cafe in Denver. I have two friends named Deb- orah and Jennifer. They are both academics who love to participate in international conferences. Recently, they realized they haven’t attended many conferences in Milwaukee, which is quite surprising given their love for the city. When I was in Milwaukee with Deb- orah, she said &quot;Jennifer wants to attend a conference here&quot;. 
+Question: According to the last sentence, where does Jennifer want to attend a conference?</td><td>Milwaukee</td><td>Denver</td><td>Shifted</td></tr></table>
+
+Table 11: Dataset Samples for Indexical "Here"
+
+<table><tr><td>Context+Stimuli and Question</td><td>Shifted Option</td><td>Non-shifted Option</td><td>Ground Truth</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is meticulous about arranging all his necessary documentation well in advance and likes having everything sorted before deadlines approach. When we spoke last summer, Donald mentioned that Carl would be having his visa appointment tomorrow. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Non-shifted</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is quite disorganized and often waits until the last possible moment to schedule important tasks such as visa appointments. When we spoke last summer, Donald mentioned that Carl would be having his visa appointment tomorrow. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Non-shifted</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is meticulous about arranging all his necessary documentation well in advance and likes having everything sorted before deadlines approach. When we spoke last summer, Donald said &quot;Carl will be having his visa appointment tomorrow&quot;. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Shifted</td></tr><tr><td>Hi! I am Albert. I am a graduate student and I live in Los Angeles. I have two friends named Donald and Carl. Carl is quite disorganized and often waits until the last possible moment to schedule important tasks such as visa appointments. When we spoke last summer, Donald said &quot;Carl will be having his visa appointment tomorrow&quot;. Question: According to the last sentence, did Carl already have his visa appointment, or is he going to do so in the future?</td><td>will do it in the future</td><td>did it in the past</td><td>Shifted</td></tr></table>
+
+Table 12: Dataset Samples for Indexical "Tomorrow"

2025/Un-considering Contextual Information_ Assessing LLMs’ Understanding of Indexical Elements/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d76f7de0fe9321b4ac44e8bb2aeed0139e365f50e2baa6388824b1ed05cb2412
+size 1093601

2025/Uncertainty Unveiled_ Can Exposure to More In-context Examples Mitigate Uncertainty for Large Language Models_/e10afd9c-4f0f-4673-bde1-594a31006fee_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:615bbda0d361efc6f3229e84cdbd971b183220dea035d0a379529b8341032dc2
+size 2134189

2025/Uncertainty Unveiled_ Can Exposure to More In-context Examples Mitigate Uncertainty for Large Language Models_/full.md

@@ -0,0 +1,627 @@
+# Uncertainty Unveiled: Can Exposure to More In-context Examples Mitigate Uncertainty for Large Language Models?
+
+Yifei Wang $^{1,2}$ , Yu Sheng $^{1,2}$ , Linjing Li $^{1,2*}$ , Daniel Zeng $^{1,2}$
+
+<sup>1</sup> State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China
+
+$^{2}$  School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China {wangyifei2022, shengyu2021, linjing.li, dajun.zeng}@ia.ac.cn
+
+# Abstract
+
+Recent advances in handling long sequences have facilitated the exploration of long-context in-context learning (ICL). While much of the existing research emphasizes performance improvements driven by additional in-context examples, the influence on the trustworthiness of generated responses remains underexplored. This paper addresses this gap by investigating how increased examples influence predictive uncertainty—an essential aspect in trustworthiness. We begin by systematically quantifying the uncertainty of ICL with varying shot counts, analyzing the impact of example quantity. Through uncertainty decomposition, we introduce a novel perspective on performance enhancement, with a focus on epistemic uncertainty (EU). Our results reveal that additional examples reduce total uncertainty in both simple and complex tasks by injecting task-specific knowledge, thereby diminishing EU and enhancing performance. For complex tasks, these advantages emerge only after addressing the increased noise and uncertainty associated with longer inputs. Finally, we explore the evolution of internal confidence across layers, unveiling the mechanisms driving the reduction in uncertainty.
+
+# 1 Introduction
+
+In-context learning has emerged as a pivotal paradigm for modern large language models (LLMs) in addressing real-world challenges (Brown et al., 2020; Dong et al., 2024). By presenting a few learning examples through carefully crafted prompts, LLMs achieve remarkable performance without requiring weight updates. The latest techniques of equipping LLMs with long-context capabilities have made strides (Jin et al., 2024), including continued fine-tuning(Rozière et al., 2024), position extrapolation (Su et al., 2024) and innovative architectures (Peng et al., 2023; Gu and Dao,
+
+![](images/fb2a9576f81f85d1e92ec804d9ddbc1f7c1fa6589705e075757f9581fd4be6a0.jpg)  
+Figure 1: Humans tend to gain task-specific knowledge and confidence as they are exposed to more examples. This raises a natural question: can additional examples similarly reduce uncertainty in LLMs?
+
+2024), open new avenues for areas previously constrained by context length.
+
+One such area is long-context ICL, also known as many-shot ICL, which involves feeding LLMs with hundreds or even thousands of input-output pairs. This regime of ICL allows LLMs to learn from large quantities of data once and could be deemed as a comparative alternative to fine-tuning methods. Despite its potential, the properties of many-shot ICL remain largely unexplored. While several studies have initiated preliminary investigations in this area, which mainly focus on performance gains from extra examples (Agarwal et al., 2024; Jiang et al., 2024), critical aspects such as trustworthiness and reliability of generations by LLMs (Wang et al., 2024) remain unexamined. Systematic investigation of these aspects is essential for advancing our understanding of long-context ICL and paves the way for its wider adoption in high-stake applications.
+
+To fill this blank, we quantitatively examine the impact of increasing scales of in-context examples on LLMs' confidence through faithful uncertainty quantification (UQ) approaches. By incorporating model parameters, configurations, and various demonstration sets, we approximate the predictive
+
+distribution in the output space. Then we compute entropy to measure total uncertainty (TU). Building on the framework proposed by (Ling et al., 2024), we employ a Bayesian framework to disentangle two core components from TU for many-shot ICL: epistemic uncertainty (EU) and aleatoric uncertainty (AU). EU arises from insufficient evidence or knowledge during model training, while AU stems from the inherent randomness and variability of the data (He et al., 2023) in Fig. 2. Our analysis reveals that the reduction in LLMs' uncertainty with more examples is primarily driven by a main decrease in EU. These examples enrich task-specific knowledge, thereby lowering EU, which in turn reduces TU and enhances performance. Furthermore, we demonstrate that the performance gains are attributed to increased informational content rather than extended context length. To explore the mechanisms behind reduced uncertainty, we project the residuals from all model layers into the vocabulary space, visualizing the evolution of internal confidence. The results reveal that long-context ICL enables LLMs to concentrate more logit mass on the correct answer and amplify the disparity between the correct response and distractors, effectively reducing uncertainty in predictions.
+
+This study represents one of the earliest efforts to examine long-context ICL through the lens of uncertainty. The core research questions addressed are as follows:
+
+- RQ1: Could more in-context examples mitigate uncertainty for LLMs? (§ 4.2)  
+- RQ2: Where do performance gains stem from, from the perspective of uncertainty decomposition? (§ 4.3)  
+- RQ3: What mechanisms underlie uncertainty reduction? (§ 5.2)
+
+# 2 Related work
+
+Long-context ICL The significant advancements in equipping LLMs with long context capabilities have expanded the potential for research in previously constrained areas, such as repository-level code understanding and multi-document QA. For ICL, an important emergent ability for LLMs (Brown et al., 2020), the extrapolation of context length enables the investigation into its performance limits and learning dynamics as the number of demonstrations scales.
+
+Several studies have initiated preliminary investigations in this area. Agarwal et al. (2024), for
+
+instance, demonstrates notable performance gains with many-shot prompting across various generative and discriminative tasks using Gemini 1.5 Pro (Team et al., 2024). In parallel, Bertsch et al. (2024) offers valuable insights into the properties of many-shot ICL, particularly examining the influence of example retrieval and demonstration order. On a more optimistic note, Jiang et al. (2024) concludes that many-shot ICL can facilitate efficient adaptation of multimodal foundation models to new applications and domains. However, the benefits of long-context ICL are not universally positive. Li et al. (2024) argues that long-context models encounter difficulties with extreme-label classification tasks, especially when large label spaces are involved.
+
+Uncertainty Quantification UQ has been extensively studied in traditional machine learning (Lakshminarayanan et al., 2017; Gawlikowski et al., 2022; Kong et al., 2023), which predominantly concentrates on estimating models' confidence and uncertainty in its prediction, called total uncertainty. Total uncertainty can be decomposed into two key components: epistemic (model) uncertainty and aleatoric (data) uncertainty (Hou et al., 2024; Valdenegro-Toro and Mori, 2022). The advent of LLMs has introduced new challenges in quantifying uncertainty, particularly due to the sequential and context-dependent nature of generative processes. Recent advances in UQ research can be categorized into two main approaches: black-box and white-box methods. Black-box UQ quantifies uncertainty by measuring the agreement across multiple generation samples (Zhang et al., 2024a), whereas white-box approaches assess internal model states or logits to capture intrinsic uncertainty (Liu et al., 2024; Bakman et al., 2024).
+
+# 3 Uncertainty Quantification Framework for Long-context ICL
+
+# 3.1 Formulation of ICL
+
+Consider an LM  $\mathcal{M}$  and a query  $\pmb{x}$ , where  $\mathcal{M}$  generates a response  $\widehat{\pmb{y}}$  by maximizing the joint probability  $\mathcal{P}_{\Theta}(\widehat{\pmb{y}}\mid \pmb {x}) = \prod_{i\geq 1}\mathcal{P}_{\Theta}(\widehat{\pmb{y}}_i\mid \widehat{\pmb{y}}_{< i},\pmb {x})$ . In the ICL regime,  $\mathcal{M}$  would condition its output on a constructed prompt  $\Omega$ , which typically includes an optional task-specific instruction  $\mathcal{I}$ , a series of  $N$  input-output demonstrations ("shots")  $\pmb{z}_{1:N} = \{(x_i,y_i)\}_{i = 1}^N$ , and a test query  $\pmb{x}_{N + 1}$ . Consequently, the generation process of ICL can be
+
+formalized as  $\widehat{\pmb{y}}\coloneqq \mathcal{P}_{\Theta}(\widehat{\pmb{y}}\mid \pmb{x}_{N + 1},\pmb{z}_{1:N},\mathcal{I})$  , enabling  $\mathcal{M}$  to address diverse complex tasks (Gatt and Krahmer, 2018).
+
+![](images/ec93434140538baaa9c8fa8c0868e6100e88f13603a6469db16b26c4b27bac74.jpg)  
+Figure 2: The sources of AU and EU in many-shot ICL. AU comes from the prompt  $\Omega$  e.g., vast examples and the process of demonstration selection. EU originates from the model's end, encompassing the generation and decoding processes.
+
+# 3.2 Faithful Uncertainty Quantification
+
+Predictive Distribution To quantify uncertainty stemming from both the demonstration sets  $z_{1:N}$ , and the model parameters or configurations  $\Theta$ , we derive the predictive distribution by sampling generations across various configurations  $\Theta \sim q(\Theta)$  and demonstration sets  $z_{1:N} \sim Z$ . This work focuses on classification and multiple-choice question-answering (MCQA) tasks. The selection of task types is discussed in Appendix B.
+
+The advantage of UQ in these tasks lies in the categorical nature of their outputs: each numerical or symbolic label  $\pmb{y} \in \mathcal{V}$  binds a predefined category or candidate answer. Thus, the probability of  $\pmb{y}$ , denoted as  $\mathcal{P}_{\pmb{y}}$ , is derived from the model's predicted logits and acts as a proxy for its confidence in their responses. Assume that for each demonstration set, we sample  $m$  decoded generations and repeat this process across  $L$  distinct sets  $z_{1:N}^{L}$ . This yields a probability set of size  $L \times m$ , capturing the uncertainty distribution over both demonstration sets and model configurations. Unlike classification or MCQA, where uncertainty can be assessed through well-defined probability distributions over discrete outputs, open-ended tasks involve variable-length outputs and lack clear ground truth, with no principled method existed for reliable UQ. Therefore, we hope we could probe the uncertain property of long-context ICL systems through MCQA tasks to provide a preliminary investigation.
+
+Entropy. By aggregating the probabilities from  $m$  decoded generations for each demonstration set into a distribution over the output space, we obtain
+
+$L\times |\mathcal{V}|$  probability matrix  $A_{L\times |\mathcal{V}|}$ , from which we compute the entropy as follows:
+
+$$
+T U = - \mathcal {H} \left[ \sigma \left(\left[ \sum_ {l = 1} ^ {L} \mathcal {P} (\boldsymbol {y} \mid \boldsymbol {x}, z _ {1: N} ^ {l}) \right] _ {\boldsymbol {y} \in | \mathcal {Y} |}\right) \right]
+$$
+
+where  $\sigma$  is a normalization function that ensures the sum of probabilities equals one, and  $\mathcal{H} = \sum_{i}p(\boldsymbol {x})\log (p(\boldsymbol {x}))$ . Some studies indicate that log-its may be uncalibrated (Liu et al., 2024; Agarwal et al., 2024). Aggregating the probability distributions from all decoded sequences can also help mitigate the errors and inaccuracies arising from uncalibrated logits, leading to a more reliable and robust output distribution.
+
+# 3.3 Uncertainty Disentanglement
+
+According to (Ling et al., 2024), from the Bayesian view, ICL maps demonstrations  $z_{1:N}$  into a preexisting latent concept  $\beta$ , which defines task-specific knowledge and enables LLMs to tackle a new in-domain task  $x_{N+1}$ . The predictive distribution of ICL is formulated as follows:
+
+$$
+\begin{array}{l} p (\mathbf {y} | z _ {1: N}) := \int p (\mathbf {y} | x _ {N + 1}, z _ {1: N}, \Theta , \beta) \\ \cdot p (\beta | z _ {1: N}) q (\Theta) d \beta d \Theta \\ \end{array}
+$$
+
+If  $\Theta$  is specific, yielding  $p(\mathbf{y}|z_{1:N},\Theta) = \int p(\mathbf{y}|z_{1:N},\beta ,\Theta)p(\beta |z_{1:N})d\beta$  with an associated entropy  $H(\mathbf{y}|z_{1:T},\beta ,\Theta)$ . The expected value of this entropy under different demonstration sets can be expressed as  $\mathbb{E}_{\beta}\left[H(\mathbf{y}_T|\mathbf{x}_{1:T},\beta ,\Theta)\right]$ , which serves as a metric to quantify the EU. AU is estimated as mutual information between  $\mathbf{y}$  and the latent concept  $\beta$  as  $I(\mathbf{y},\beta |\Theta)$ , which is the difference between TU and EU as follows:
+
+$$
+\begin{array}{l} I (\mathbf {y}, \beta | \Theta) = \\ H (\mathbf {y} | z _ {1: N}, \Theta) - \mathbb {E} _ {\beta} [ H (\mathbf {y} | z _ {1: N}, \beta , \Theta) ] \\ \end{array}
+$$
+
+The latent concept  $\beta$  distribution could be obtained by sampling from different demonstrations. Beam search effectively approximates the posterior of  $\Theta$ , which draws hypotheses from the most probable regions in the hypotheses space. Utilizing the probability matrix  $A_{L\times m}$  obtained in Sec. 3.2, TU, EU and AU can be approximated as follows:
+
+$$
+\begin{array}{l} T U = H \left(\sigma \left(\sum \left[ A _ {j,:} \right]\right)\right) \\ E U = \frac {1}{L} H (\sigma (A _ {j,:})) \\ A U = H (\sigma (\sum [ A _ {j,:} ])) - \frac {1}{L} H (\sigma (A _ {j,:})) \\ \end{array}
+$$
+
+![](images/3163b9c04376c2d266a02ba2b8109cc6745354c726b28df1eecb754050c81340.jpg)  
+Figure 3: A workflow for uncertainty quantification and decomposition under many-shot ICL settings, involves the following components: a LLM  $\mathcal{M}$  supporting long context windows, demonstration set selection, generation sampling, and the UQ modules detailed in Sec. 3.2 and 3.3.
+
+<table><tr><td>Model</td><td>Size</td><td>Strategy</td><td>Support</td></tr><tr><td>Llama-3.1-8B</td><td>8B</td><td>Fine-tuning</td><td>128K</td></tr><tr><td>Mistral-7B-v0.2</td><td>7B</td><td>NTK-Aware</td><td>32K</td></tr><tr><td></td><td></td><td>Interpolation</td><td></td></tr><tr><td>Qwen1.5-7B</td><td>7B</td><td>Fine-tuning</td><td>32K</td></tr></table>
+
+Table 1: Long-Context LLMs Overview
+
+# 4 Experiments
+
+# 4.1 Experimental Settings
+
+Models. We evaluate three widely used base models prior to instruction-tuning (Wei et al., 2022a): Llama-3.1-8B (Touvron et al., 2023), Mistral-7Bv0.2 (Jiang et al., 2023), and Qwen1.5-7B (Bai et al., 2023). The supported maximum context length, along with their respective strategies for long-context training, are summarized in Table 1.
+
+Datasets and tasks. We define two modes for classification tasks and MCQA: easy and hard. The hard mode consists of three increasingly complex logical deduction tasks, including determining the order of a sequence of objects ranging from three to seven, from a suite of challenging algorithmic reasoning tasks known as BIG-Bench Hard (BBH) (Suzgun et al., 2023). In contrast, the easy mode encompasses traditional natural language understanding (NLU) tasks such as AGNews (Zhang et al., 2015) and SST2 (Socher et al., 2013), along with the commonsense reasoning task, CommonsenseQA (Talmor et al., 2019).
+
+Long-context ICL settings. To investigate how uncertainty evolves with increasing exposure to examples, we apply UQ and uncertainty decomposition methods across different  $k$ -shot ICL. For demonstration selection, we randomly sample  $k$  shots from the training set for each test example.
+
+In all tasks, we employ beam search to generate 10 candidate outputs and set the temperature parameter as 0.7. For decomposing TU, we iterate six different demonstration sets to disentangle EU and AU. All open-source models are sourced from Hugging Face<sup>1</sup> and experimented on eight 80GB NVIDIA RTX A100 GPUs.
+
+# 4.2 RQ1: Could more in-context examples mitigate uncertainty for LLMs?
+
+Quality of Uncertainty Measures In the context of UQ, a key consideration is its ability to reflect the correctness and reliability of LLM outputs. High uncertainty most likely leads to incorrect predictions while low uncertainty indicates a higher likelihood of correct responses. To this end, we examine how the quality of uncertainty measures varies from few-shot to long-context ICL settings. Following prior works (Kuhn et al., 2023; Lin et al., 2024), we adopt Exact match as the metric for correctness and use uncertainty estimates to predict the correctness of response. We then compute AUROC  ${}^{2}$  to evaluate whether the UQ measures employed are good indicators. The AUROC and accuracy results for Llama-3.1-8B are presented in Tab.8. As the number of demonstrations increases, AUROC values remain high with minimal fluctuations, suggesting that the UQ measures serve as high-quality indicators and generalize effectively to long-context ICL, which reinforces the validity of our experimental results and the conclusions drawn.
+
+Average View Overall, many-shot ICL effectively reduces LLMs' uncertainty across models and datasets. As shown in Figs. 4 and 5, the results indicate a simultaneous rise in accuracy and con
+
+![](images/da8d5714b8194d1b94dd3eeee7fcae9f066bf33d6c2856d4ee5cdc867e665230.jpg)  
+Figure 4: The average TU under  $k$ -shot ICL with error bands for three runs.
+
+![](images/83bd8466e529f102b50e3dffe9d275e1ff96909185215d1ffe800a6591cc392c.jpg)
+
+![](images/6530c258537933d2dcb2d380bbaec3c5b7fcdccefe9af589f76a90c28bf5f78a.jpg)  
+Figure 5: The average accuracy under  $k$ -shot ICL with error bands for three runs.
+
+![](images/f12dcf401c22708498636a847fbf1104626de0f1a837583f0beccadc0cd71762.jpg)
+
+fidence as more in-context examples are provided, highlighting the correlation between improved confidence and performance gains for LLMs.
+
+For *easy mode*, the inclusion of initial examples rapidly drives predictive entropy to a relatively low-uncertainty state, with further increases in examples yielding only marginal reductions in entropy (see Fig. 6 for a detailed view). In contrast, *hard mode* exhibits a distinct pattern. Predictive entropy remains higher in hard mode compared to easy mode due to the intrinsic complexity of the tasks, particularly those involving logical deduction with increasing object complexity  $(TU_{LD3} < TU_{LD5} < TU_{LD7})$ . Here, adding initial examples has minimal impact on entropy reduction until the number exceeds several hundred, at which point substantial performance gains emerge.
+
+When demonstrations are incorporated, both Llama-3.1-8B and Mistral-7B-v0.2 exhibit consistent improvements in performance  $(\uparrow)$  and reductions in uncertainty  $(\downarrow)$ . In contrast, Qwen1.5-7B demonstrates pronounced variability on datasets under hard mode, where fewer-shot ICL (e.g., 10-shot) achieves levels of confidence and accuracy comparable to certain many-shot settings (e.g., 240-shot). We term this phenomenon the "ICL sink", drawing analogies to sink patterns observed in attention mechanisms (Xiao et al., 2024). Notably, for Mistral models, even at the context limit in the 240-shot ICL setting on the LD7 dataset, Mistral sustains robust instruction-following and achieves performance comparable to Llama-3.1-8B, despite the latter's fourfold context-length capacity. This underscores the architectural strengths
+
+![](images/646133131874445eca041e484665e0942c5e31a73a4d2f05110e7b60e98b9530.jpg)  
+Figure 6: TU distribution of 2000 examples under certain  $k$ -shot ICL on AG News and SST2 datasets for Mistral-7B-v0.2.
+
+![](images/6022629b3e7717b9a35d1f8b5e63808fce7246464e1b2e406c31e1b9d74509ab.jpg)
+
+![](images/b4ad91ff5a7ed68cd653809201fc1037a7aa97811cdc084a126391e0ac87be20.jpg)  
+Figure 7: EU of Llama-3.1-8B on AG News and logical_deductionfive_objectsdatasets for distant examples vs.repeating  $4 / 10$  examples  $N$  times.
+
+![](images/8dd310fa297d6e0e83dbbc0ef2986f232eefd128563ca83b908dcc0d56a659b5.jpg)
+
+of Mistral, which leverages a sparse Mixture of Experts (MoEs) (Shazeer et al., 2017) and sliding window attention. Thus, the influence of additional in-context examples on uncertainty fundamentally depends on the intrinsic long-context understanding capabilities.
+
+Micro View An increasing number of examples effectively mitigates uncertainty for most questions. Tables 2, 11, and 12 detail the percentage of questions exhibiting decreased or increased uncertainty under  $k$ -shot ICL. Despite  $8.65\%$  of cases experiencing heightened uncertainty with longer inputs in 128-shot learning, this effect minimally impacts
+
+overall model performance, as reflected by the small absolute values of  $\Delta Acc$ . Crucially, the transition from few-shot (e.g., 4-shot) to many-shot ICL demonstrates a marked reduction in uncertainty for a larger proportion of questions, driving consistent performance improvements. These findings suggest that enhanced performance stems from increased confidence in the majority of questions.
+
+Choices of  $k$  For practical applications, we recommend opting for a relatively larger  $k$  in in-context learning, as it simultaneously enhances performance and bolsters reliability.
+
+Ablations with Model Size To further strengthen our analysis, we conducted additional experiments on the instruction-tuned versions of the more capable Qwen-2.5-14B and Qwen-2.5-32B models. The complete results are presented in Appendix C.
+
+Across all uncertainty measures (TU, EU, and AU), larger models consistently exhibit substantially lower uncertainty values. For easy-mode, large LLMs follow similar uncertainty trends as smaller models; On more challenging tasks (hard mode), LLMs display distinct uncertainty patterns. Specifically, for Qwen-2.5-14B, EU steadily decreases as more demonstrations are provided, indicating more rapid task adaptation and improved performance, whereas AU remains relatively stable. Notably, a detailed analysis reveals that AU for Qwen-2.5-14B decreases slightly when initial examples are added but begins to rise beyond 80-shot, likely due to long-context effects introducing noise. In contrast, Qwen-2.5-32B does not exhibit this trend; instead, its AU continues to decrease as the number of examples increases.
+
+Takeaways Large-scale LLMs exhibit greater confidence (i.e., lower uncertainty) and superior performance under many-shot settings, compared to smaller counterparts. The benefits of many-shot ICL remain evident, as additional demonstrations continue to enhance task-specific adaptation while maintaining low EU. Thus, the advantages of long-context IC, both in terms of performance and confidence, persist even at a larger scale.
+
+# 4.3 RQ2: where do performance gains stem from?
+
+In Sec. 4.2, we establish that reduced uncertainty improves performance. We hypothesize that additional examples in ICL foster a more refined task-specific conceptual framework, denoted as  $\beta$ ,
+
+which empowers LLMs to approach novel problems  $x_{T + 1}$  within the domain with increased confidence and efficacy. To validate this, we decompose total uncertainty into EU and AU, checking how these context helps LLMs to improve confidence by utilizing the definition and property of two special forms of uncertainty (Fig. 8).
+
+Lower EU as the Primary Driver of TU Reduction. The decrease in TU is predominantly attributed to a decline in EU. Initially, EU accounts for the majority of TU, indicating that uncertainty primarily arises from the LLMs' insufficient indomain knowledge, while their robust natural language understanding keeps AU relatively low. In simpler task settings, LLMs swiftly acquire task-specific knowledge, leading to a rapid decline in EU and sustaining consistently low AU. In contrast, for challenging tasks involving intricate logical structures, additional demonstrations may elevate AU (e.g., Llama-3.1-8B on the LD7 dataset), partially counteracting the reduction in EU and impeding significant decreases in total entropy. This underscores the persistent difficulty for current large models in effectively comprehending long texts with complex structures.
+
+Additional Information Reduces EU. To validate that additional examples enhance the informational content and yield a clearer  $\beta$  for models (as shown in Fig. 7), we observe that only diverse examples effectively reduce EU under  $k$ -shot learning, whereas repetitive examples fail to achieve the same effect. This highlights that the true driver of uncertainty reduction lies in the increased informational richness of the examples provided.
+
+# 5 Interpretability View for Uncertainty in K-shot ICL
+
+To investigate the mechanisms by which increased in-context demonstrations reduce uncertainty in LLMs, we aim to delve into the models' internal states, unraveling the underlying processes governing answer selection and generation in in-context learning, thereby offering a comprehensive and interpretable analysis of this phenomenon.
+
+# 5.1 Residual Stream Projection
+
+Residual Streams Residual streams function as iterative refinements of feature representations in deep neural networks (He et al.; Li and Papyan, 2023), encapsulating the process of hierarchical in
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="2">8-shot</td><td colspan="2">16-shot</td><td colspan="2">32-shot</td><td colspan="2">64-shot</td><td colspan="2">128-shot</td></tr><tr><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td></tr><tr><td></td><td colspan="10">Easy Mode</td></tr><tr><td rowspan="2">AG News</td><td>66.8</td><td>+7.3</td><td>83.6</td><td>+11.5</td><td>88.6</td><td>+13.9</td><td>91.2</td><td>+15.2</td><td>90.8</td><td>+15.8</td></tr><tr><td>30.45</td><td>-1.0</td><td>15.00</td><td>-0.7</td><td>10.75</td><td>-0.2</td><td>8.4</td><td>-0.35</td><td>8.65</td><td>-0.4</td></tr><tr><td rowspan="2">SST-2</td><td>71.7</td><td>+5.7</td><td>82.9</td><td>+6.1</td><td>86.6</td><td>+6.6</td><td>88.5</td><td>+7.1</td><td>92.1</td><td>+7.9</td></tr><tr><td>20.3</td><td>-0.5</td><td>12.5</td><td>-0.4</td><td>9.4</td><td>-0.4</td><td>8.2</td><td>-0.3</td><td>5.6</td><td>-0.3</td></tr><tr><td rowspan="2">Commonsense QA</td><td>62.2</td><td>+1.8</td><td>69.8</td><td>+4.2</td><td>69.0</td><td>+4.8</td><td>78.6</td><td>+6.6</td><td>81.2</td><td>+5.2</td></tr><tr><td>26.2</td><td>-0.4</td><td>18.8</td><td>-0.2</td><td>17.8</td><td>-0.2</td><td>16.8</td><td>-1.0</td><td>16.6</td><td>-0.8</td></tr><tr><td rowspan="3"></td><td colspan="10">Hard Mode</td></tr><tr><td colspan="2">20-shot</td><td colspan="2">40-shot</td><td colspan="2">80-shot</td><td colspan="2">120-shot</td><td colspan="2">240-shot</td></tr><tr><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td></tr><tr><td rowspan="2">Logical Deduction3</td><td>45.6</td><td>+6.0</td><td>38.1</td><td>+4.0</td><td>40.4</td><td>+2.0</td><td>53.2</td><td>+7.60</td><td>62.3</td><td>+15.91</td></tr><tr><td>44.8</td><td>-5.6</td><td>51.2</td><td>-8.0</td><td>46.8</td><td>-10.8</td><td>40.4</td><td>-6.8</td><td>34.5</td><td>-5.9</td></tr><tr><td rowspan="2">Logical Deduction5</td><td>58.4</td><td>+2.8</td><td>64.4</td><td>+2.8</td><td>73.6</td><td>+4.8</td><td>79.6</td><td>+8.0</td><td>83.8</td><td>+10.8</td></tr><tr><td>33.2</td><td>-0.4</td><td>30.0</td><td>-1.2</td><td>24.4</td><td>-0.8</td><td>18.0</td><td>-1.6</td><td>13.1</td><td>-1.5</td></tr><tr><td rowspan="2">Logical Deduction7</td><td>48.4</td><td>+2.0</td><td>54.4</td><td>+3.6</td><td>59.2</td><td>+4.4</td><td>75.0</td><td>+12.0</td><td>83.3</td><td>+12.3</td></tr><tr><td>48.8</td><td>-1.2</td><td>42.0</td><td>-0.8</td><td>38.0</td><td>-0.8</td><td>24.5</td><td>-0.0</td><td>15.3</td><td>-0.7</td></tr></table>
+
+Table 2:  $\Delta U$  refers to the proportion of datasets displaying either a decrease or increase in uncertainty relative to the 4-shot baseline, with  $|\Delta U| > \tau$  indicating significant uncertainty changes. For each dataset, the first row presents the proportion of questions exhibiting reduced uncertainty, while the second row reflects those with increased uncertainty.  $\Delta Acc$  quantifies the performance shift associated with the corresponding subset. Model: Llama-3.1-8B.
+
+![](images/548769cd30210dbbd84af2a91e8a6e63cf8561b4ee14b5760dbd0083f89fed7c.jpg)
+
+![](images/c829a627ad7f00dfb25a95447eace278926ab6f0f4c27a6f3166c6f4dfb81da0.jpg)
+
+![](images/89a1cdb565ce3b8881cf9c09441e6f32e5d1a2275cd3fa6cc810621f020ce894.jpg)
+
+![](images/02119a1839c93db0035115bb49c6f8ac7308cb79bcda5eddd251405b92d82992.jpg)
+
+![](images/c05103d477cc221f5c7c52c7583321deed98e4aea39eace8144c57349a23c098.jpg)  
+Figure 8: Uncertainty decomposition results for both easy mode (left) and hard mode (right).
+
+![](images/12321cbea50b07a76df909c5224570677e082a4e66b0938eb73655c68ae09cdc.jpg)
+
+formation aggregation. By leveraging residual connections, models reveal their mechanisms for constructing and iteratively refining outputs, thereby improving interpretability. Formally, in decoder-only LLMs, the hidden state of the  $i$ -th token at the  $l$ -th layer, denoted as  $\mathbf{h}_i^{(l)}$ , is computed as:
+
+$$
+\mathbf {h} _ {i} ^ {(l)} = \mathbf {h} _ {i} ^ {(l - 1)} + \mathbf {a} _ {i} ^ {(l)} + \mathbf {m} _ {i} ^ {(l)},
+$$
+
+$$
+\mathbf {a} _ {i} ^ {(l)} = \mathcal {M S H A} \big (\mathbf {h} _ {i} ^ {(l - 1)} \big),
+$$
+
+$$
+\mathbf {m} _ {i} ^ {(l)} = \mathcal {M L P} \Big (\mathbf {h} _ {i} ^ {(l - 1)} + \mathbf {a} _ {i} ^ {(l)} \Big),
+$$
+
+where  $\mathcal{MSHA}(\cdot)$  represents the multi-head self-attention mechanism (Vaswani, 2017), and  $\mathcal{M}\mathcal{L}\mathcal{P}(\cdot)$  denotes the feed-forward neural network. For simplicity, detailed computations within the
+
+MHSA sublayer, such as the projection matrices  $\mathbf{W}_{Q,K,V,O}$ , and the splitting-merging operations across attention heads, are omitted here. Each decoder block, therefore, maintains two distinct residual pathways: one emerging from the MHSA,  $\mathbf{h}_i^{(l)}$ , and the other from MLP sublayer,  $\mathbf{h}_i^{(l)} + \mathbf{a}_i^{(l)}$ .
+
+Projection into Vocabulary To uncover the latent information encoded within residual streams, projecting intermediate states onto a probability distribution over the vocabulary space  $V$  provides critical insights into the temporal and spatial dynamics of how these networks construct and refine their outputs (Geva et al., 2021; Belrose et al., 2023; Dar et al., 2023). Analogous to token generation, each residual stream  $\mathbf{r}_i\in \mathbb{R}^d$  at the fi
+
+![](images/6111d726eb7bff22d5a5d82ce8cbb3641ad1b522aaa853134fb8954744903190.jpg)  
+Figure 9: Average probabilities of Mistral-7B-v0.2 on the Commonsense QA dataset for MCQA items where the correct answer is "D". A 32-layer LM gets 64 residual streams, excluding the output hidden states.
+
+![](images/0b6b3b79ee3f7cdf264fc1f7cfc31d104bb308bff1bd24019aed2e6a9c9aab31.jpg)
+
+![](images/b9ccdde2d14f8d65ceadb3c562ab7f13e8aaea5d2ae2cf13c0e8a266c6751fbe.jpg)
+
+![](images/a70851ca8e171d5a456995b0c28696cc63a70341ab41270f7b869cbb7e0b1fb3.jpg)
+
+![](images/8f5990ced2de2ab4b9f7ade9614ed25df7aec0e8351f97342dc6e2a98d1a67d4.jpg)  
+Figure 10: Average logits of Mistral-7B-v0.2 on the Commonsense QA dataset for MCQA items with the correct answer "D". Increasing in-context examples amplifies the logit of the correct option, thereby magnifying the difference between the logits of correct and incorrect options.  $\star$  represent the highest logit and  $\spadesuit$  the second highest logit. Refer to Appendix F.2 for additional results.
+
+![](images/1bc1337e545c43ab1960eccf5d707f501d11d6231b0b3dc425f58a7324ab22d1.jpg)
+
+![](images/ac5b95218d297a23f7b56e5bf7ed0d3a9c5dc1b5b4f399477791402932e692f3.jpg)
+
+![](images/c8ad8530622d774991e4f79b10e01c057d979ea78d3015efa930a7e956056ce0.jpg)
+
+nal position—where  $i$  indexes the  $i$ -th residual in the model—undergoes transformation via an unembedding matrix  $W_{U} \in \mathbb{R}^{d \times |V|}$  post layer normalization. This process yields calibrated logits  $\mathbf{l}_i = W_U \mathbf{LayerNormalization}(\mathbf{r}_i)$  and the corresponding probabilities  $\mathbf{p}_i = \mathrm{Softmax}(\mathbf{l}_i)$ .
+
+Correlation with Uncertainty in ICL For  $k$ -shot in-context learning, consider projecting the residual representations at the answer position into the probability simplex  $\Delta^{[V]}$  over the vocabulary  $V$ . Denote the resulting logits and probabilities of candidate symbols (e.g., "A", "B", "C") as  $\ell_i$  and  $p_i$ , respectively. These logits  $\ell_i$  or probabilities  $p_i$  serve as proxies for confidence levels associated with each candidate. Analyzing the evolution of  $\ell_i$  across model layers reveals the hierarchical development of inner confidence throughout  $k$ -shot learning, offering a profound understanding of the underlying uncertainty dynamics.
+
+# 5.2 RQ3: What mechanisms underlie uncertainty reduction?
+
+Qualitative Analysis To begin with, we present a case study in Fig. 12, offering an intuitive and qualitative demonstration of how the number of shots influences uncertainty. In this case, the Mistral-7B model struggles to distinguish the correct answer,
+
+option "E", under a 4-shot ICL setting, as the other options continuously mislead the model throughout the process. This is evidenced by the fluctuating confidence levels, which rise and fall erratically. In contrast, as the number of shots increases (32-, 64-, and 128-shot settings), many-shot ICL consistently boosts the probability of selecting "E" as the correct answer from about 22nd layer onward, maintaining this highest probability thereafter. Simultaneously, it demonstrates robustness by maintaining near-zero probabilities for incorrect options, effectively eliminating the influence of distractors on the model's final prediction.
+
+<table><tr><td>CMQA</td><td>4-shot</td><td>32-shot</td><td>64-shot</td><td>128-shot</td></tr><tr><td>Llama-3.1</td><td>2.86 / 24.98</td><td>2.75 / 27.03</td><td>2.55 / 27.66</td><td>2.53 / 28.01</td></tr><tr><td>Mistral-v0.2</td><td>2.78 / 17.14</td><td>2.24 / 19.60</td><td>2.57 / 20.38</td><td>2.75 / 20.84</td></tr><tr><td>Qwen1.5</td><td>3.51 / 29.11</td><td>3.62 / 30.49</td><td>3.73 / 30.97</td><td>3.76 / 30.94</td></tr><tr><td>LD3</td><td>4-shot</td><td>40-shot</td><td>120-shot</td><td>240-shot</td></tr><tr><td>Llama-3.1</td><td>0.51 / 15.93</td><td>0.77 / 17.15</td><td>0.65 / 16.6</td><td>0.77 / 16.87</td></tr><tr><td>Mistral-7B-v0.2</td><td>0.26 / 11.07</td><td>0.48 / 11.92</td><td>0.46 / 12.05</td><td>0.59 / 11.87</td></tr><tr><td>Qwen1.5-7B</td><td>0.45 / 15.98</td><td>0.46 / 16.49</td><td>0.43 / 16.54</td><td>0.49 / 16.72</td></tr></table>
+
+Table 3: Average logit difference / the largest logit.
+
+Extended Examples Amplify Logit Disparity. We compute the average logits  $\ell_{i}$  and probabilities  $p_i$  (Figs. 10 and 9) across varying shot counts for groups sharing the same answer. The analysis reveals that extended ICL enhances the precision of LLMs, concentrating greater logit mass on the
+
+correct symbol while effectively suppressing alternatives. This dynamic, driven by the interplay between an amplified logit disparity and increased absolute logit values (Table 3), leverages the exponential sensitivity of the Softmax function to propel the probability of the correct symbol toward 1. Consequently, the uncertainty in LLM predictions is significantly reduced.
+
+# 6 Further Discussion
+
+Clarifications While our work builds upon the framework in (Ling et al., 2024), our research specifically investigates the evolution of uncertainty in long-context ICL, a topic that has not been examined to date. In contrast, Ling et al. primarily focus on introducing a framework for decomposing uncertainty in few-shot ICL. By shifting the focus to long-context scenarios, our study explores how uncertainty evolves as the number of in-context examples increases, thereby addressing an important yet understudied dimension of ICL.
+
+# 7 Conclusion
+
+This study investigates the impact of extra demonstrations on the confidence of LLMs in their responses. Experimental results demonstrate that additional examples significantly reduce TU across both simple and complex tasks by integrating task-specific knowledge. This reduction is primarily attributed to decreased model uncertainty, which enhances overall performance. However, in complex tasks, many-shot ICL faces challenges in reducing TU due to a concurrent increase in AU. Analysis of the internal representations of LLMs reveals that many-shot ICL not only reallocates greater logit mass toward correct responses but also enlarges the logit margin between correct answers and distractors, reflecting an increase in model confidence.
+
+# Limitation
+
+Our study is the first systematic investigation into uncertainty evolution in long-context ICL, addressing a critical research gap. These foundational experiments hope to provide a basis for future UQ studies on open-ended tasks. However, several limitations must be acknowledged.
+
+Exclusion of Open-Ended Tasks The scope of this work does not encompass the uncertainty analysis of open-ended tasks, such as abstractive summarization (Hasan et al., 2021) and machine translation (Costa-jussa et al., 2022), owing to the lack
+
+of robust UQ techniques for free-form generative scenarios. Nevertheless, applying ICL to rationale-intensive reasoning and generative contexts remains a promising direction. Future investigations should assess the reliability and trustworthiness of ICL in these domains, as advancements in this area could not only enhance task-solving performance but also broaden the applicability of UQ methodologies to more diverse and complex settings.
+
+Limited Exploration of ICL Configurations This study also excludes several influential ICL paradigms, such as unsupervised ICL (Yu et al., 2024), reinforced ICL (Jiang et al., 2024), and CoT prompting (Wei et al., 2022b), the latter of which is widely adopted in reasoning tasks to elicit step-by-step rationales. Existing UQ methods fall short of capturing the logical complexity intrinsic to reasoning-intensive contexts. Furthermore, practical challenges, including the computational overhead and context-length constraints of current open-source LLMs, prevented us from investigating extreme-shot ICL scenarios involving thousands of demonstrations. These limitations underscore promising directions for future research, particularly in applying UQ methodologies to better accommodate the unique challenges posed by reasoning tasks. More discussion in Appendix A.
+
+# Broader Impact
+
+Despite these limitations, this study marks a pivotal advancement in understanding the reliability of ICL by harnessing recent breakthroughs in uncertainty quantification and decomposition, an essential yet underexplored aspect of LLM research. The research on uncertainty in ICL enriches the field of uncertainty quantification, providing novel perspectives on the trustworthiness of many-shot ICL. These contributions lay a solid foundation for broadening ICL's applicability in high-stakes domains. Ultimately, these findings could catalyze the development of more dependable and interpretable AI systems, offering profound societal impact.
+
+# Acknowledgements
+
+This work was supported in part by the National Natural Science Foundation of China under Grant 72293575, and in part by the Strategic Priority Research Program of Chinese Academy of Sciences under Grant XDA0480301 and the Excellent Youth Program of State Key Laboratory of Multimodal Artificial Intelligence Systems MAIS2024310.
+
+# References
+
+Rishabh Agarwal, Avi Singh, Lei M Zhang, Bernd Bohnet, Luis Rosias, Stephanie C.Y. Chan, Biao Zhang, Aleksandra Faust, and Hugo Larochelle. 2024. Many-shot in-context learning. In ICML 2024 Workshop on In-Context Learning.  
+Ekin Akyurek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learning algorithm is in-context learning? investigations with linear models. In The Eleventh International Conference on Learning Representations.  
+Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. 2023. Qwen technical report.  
+Yavuz Faruk Bakman, Duygu Nur Yaldiz, Baturalp Buyukates, Chenyang Tao, Dimitrios Dimitriadis, and Salman Avestimehr. 2024. MARS: Meaning-aware response scoring for uncertainty estimation in generative LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7752-7767, Bangkok, Thailand. Association for Computational Linguistics.  
+Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. 2023. Eliciting latent predictions from transformers with the tuned lens. arXiv preprint arXiv:2303.08112.  
+Amanda Bertsch, Maor Ivgi, Uri Alon, Jonathan Berant, Matthew R. Gormley, and Graham Neubig. 2024. In-context learning with long-context models: An in-depth exploration. In First Workshop on Long-Context Foundation Models @ ICML 2024.  
+Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. 2024. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17682-17690.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack
+
+Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Stephanie Chan, Adam Santoro, Andrew Lampinen, Jane Wang, Aaditya Singh, Pierre Richemond, James McClelland, and Felix Hill. 2022. Data distributional properties drive emergent in-context learning in transformers. Advances in neural information processing systems, 35:18878-18891.  
+Marta R Costa-jussa, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv e-prints, pages arXiv-2207.  
+Guy Dar, Mor Geva, Ankit Gupta, and Jonathan Berant. 2023. Analyzing transformers in embedding space. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16124-16170, Toronto, Canada. Association for Computational Linguistics.  
+Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. 2024. A survey on in-context learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1107-1128, Miami, Florida, USA. Association for Computational Linguistics.  
+Albert Gatt and Emiel Krahmer. 2018. Survey of the state of the art in natural language generation: core tasks, applications and evaluation. J. Artif. Int. Res., 61(1):65-170.  
+Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kraspe, Rudolph Triebel, Peter Jung, Ribana Roscher, Muhammad Shahzad, Wen Yang, Richard Bamler, and Xiao Xiang Zhu. 2022. A survey of uncertainty in deep neural networks.  
+Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484-5495, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Albert Gu and Tri Dao. 2024. Mamba: Linear-time sequence modeling with selective state spaces.  
+Tahmid Hasan, Abhik Bhattacharjee, Md. Saiful Islam, Kazi Mubasshir, Yuan-Fang Li, Yong-Bin Kang, M. Sohel Rahman, and Rifat Shahriyar. 2021. XLsum: Large-scale multilingual abstractive summarization for 44 languages. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021,
+
+pages 4693-4703, Online. Association for Computational Linguistics.  
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.  
+Wenchong He, Zhe Jiang, Tingsong Xiao, Zelin Xu, and Yukun Li. 2023. A survey on uncertainty quantification methods for deep learning. arXiv preprint arXiv:2302.13425.  
+Bairu Hou, Yujuan Liu, Kaizhi Qian, Jacob Andreas, Shiyu Chang, and Yang Zhang. 2024. Decomposing uncertainty for large language models through input clarification ensembling. In *Forty-first International Conference on Machine Learning*.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b.  
+Yixing Jiang, Jeremy Andrew Irvin, Ji Hun Wang, Muhammad Ahmed Chaudhry, Jonathan H Chen, and Andrew Y. Ng. 2024. Many-shot in-context learning in multimodal foundation models. In ICML 2024 Workshop on In-Context Learning.  
+Hongye Jin, Xiaotian Han, Jingfeng Yang, Zhimeng Jiang, Zirui Liu, Chia-Yuan Chang, Huiyuan Chen, and Xia Hu. 2024. Llm maybe longlm: Self-extend llm context window without tuning.  
+Lingkai Kong, Harshavardhan Kamarthi, Peng Chen, B. Aditya Prakash, and Chao Zhang. 2023. Uncertainty quantification in deep learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '23, page 5809-5810, New York, NY, USA. Association for Computing Machinery.  
+Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations.  
+Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30.  
+Jianing Li and Vardan Papyan. 2023. Residual alignment: Uncovering the mechanisms of residual networks. In Thirty-seventh Conference on Neural Information Processing Systems.  
+Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. 2024. Long-context llms struggle with long in-context learning.
+
+Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. 2024. Generating with confidence: Uncertainty quantification for black-box large language models. Transactions on Machine Learning Research.  
+Chen Ling, Xujiang Zhao, Xuchao Zhang, Wei Cheng, Yanchi Liu, Yiyou Sun, Mika Oishi, Takao Osaki, Katsushi Matsuda, Jie Ji, Guangji Bai, Liang Zhao, and Haifeng Chen. 2024. Uncertainty quantification for in-context learning of large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3357-3370, Mexico City, Mexico. Association for Computational Linguistics.  
+Xin Liu, Muhammad Khalifa, and Lu Wang. 2024. Litcab: Lightweight language model calibration over short- and long-form responses. In The Twelfth International Conference on Learning Representations.  
+Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Leon Derczynski, Xingjian Du, Matteo Grella, Kranthi Gv, Xuzheng He, Haowen Hou, Przemyslaw Kazienko, Jan Koccon, Jiaming Kong, Bartlomiej Koptyra, Hayden Lau, Jiaju Lin, Krishna Sri Ipsit Mantri, Ferdinand Mom, Atsushi Saito, Guangyu Song, Xiangru Tang, Johan Wind, Stanisław Wozniak, Zhenyuan Zhang, Qinghua Zhou, Jian Zhu, and Rui-Jie Zhu. 2023. RWKV: Reinventing RNNs for the transformer era. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14048-14077, Singapore. Association for Computational Linguistics.  
+Joshua Robinson and David Wingate. 2023. Leveraging large language models for multiple choice question answering. In *The Eleventh International Conference on Learning Representations*.  
+Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. 2024. Code llama: Open foundation models for code.  
+Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks:
+
+The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics.  
+Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomput., 568(C).  
+Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc Le, Ed Chi, Denny Zhou, and Jason Wei. 2023. Challenging BIG-bench tasks and whether chain-of-thought can solve them. In Findings of the Association for Computational Linguistics: ACL 2023, pages 13003-13051, Toronto, Canada. Association for Computational Linguistics.  
+Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. CommonsenseQA: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4149-4158, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530.  
+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.  
+Matias Valdenegro-Toro and Daniel Saromo Mori. 2022. A deeper look into aleatoric and epistemic uncertainty disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1509-1517.  
+A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems.  
+Yifei Wang, Yuheng Chen, Wanting Wen, Yu Sheng, Linjing Li, and Daniel Dajun Zeng. 2024. Unveiling factual recall behaviors of large language models through knowledge neurons. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 7388-7402, Miami, Florida, USA. Association for Computational Linguistics.
+
+Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V Le. 2022a. Finetuned language models are zero-shot learners. In International Conference on Learning Representations.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022b. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Sarah Wiegrefe, Oyvind Tafjord, Yonatan Belinkov, Hannaneh Hajishirzi, and Ashish Sabharwal. 2024. Answer, assemble, ace: Understanding how transformers answer multiple choice questions.  
+Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. 2024. Efficient streaming language models with attention sinks. In *The Twelfth International Conference on Learning Representations*.  
+Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809-11822.  
+Guoxin Yu, Lemao Liu, Mo Yu, Yue Yu, and Xiang Ao. 2024. Rethinking the evaluation of in-context learning for LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14068-14082, Miami, Florida, USA. Association for Computational Linguistics.  
+Caiqi Zhang, Fangyu Liu, Marco Basaldella, and Nigel Collier. 2024a. LUQ: Long-text uncertainty quantification for LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5244-5262, Miami, Florida, USA. Association for Computational Linguistics.  
+Wenxuan Zhang, Yue Deng, Bing Liu, Sinno Pan, and Lidong Bing. 2024b. Sentiment analysis in the era of large language models: A reality check. In *Findings of the Association for Computational Linguistics: NAACL* 2024, pages 3881-3906, Mexico City, Mexico. Association for Computational Linguistics.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc.
+
+# A Related Work
+
+Overview. Estimating uncertainty in generation tasks presents greater challenges (Kuhn et al., 2023) compared to tasks with a predefined candidate set like classification tasks (Zhang et al., 2024b) and multiple-choice question answering (MCQA) (Robinson and Wingate, 2023). This is primarily due to the vast, high-dimensional semantic space inherent in natural language, which results in an effectively infinite generation space (Lin et al., 2024; Ling et al., 2024; Liu et al., 2024). In contrast, classification tasks provide LLMs with a finite set of discrete candidates, where the model's task is limited to selecting the most probable answer from a predefined set (Wiegreffe et al., 2024).
+
+# B Further Discussion
+
+# B.1 Limitations of UQ in Open-ended Tasks
+
+UQ in open-ended tasks primarily focuses on knowledge-intensive QA tasks, which differs fundamentally from the typical ICL paradigm. ICL primarily relies on: pattern matching (Min et al., 2022); distribution alignment (Chan et al., 2022); implicit fine-tuning (Akyurek et al.). In contrast, knowledge-intensive QA depends on retrieving from external knowledge and parametric knowledge, rather than adapting through in-context distribution learning. As a result, many-shot ICL is not well-suited for knowledge-intensive QA scenarios, making existing UQ methods for this domain inapplicable. Moreover, prior research on the performance of many-shot ICL has primarily focused on reasoning tasks and extreme-label classification (Li et al., 2024), rather than knowledge-intensive tasks.
+
+Challenges in Extending UQ to Open-ended Tasks Open-ended tasks encompass summarization, intermediate reasoning, code generation, program synthesis, and planning. However, existing UQ methods struggle to generalize effectively to these tasks, particularly in long-context ICL settings. For instance, semantic entropy (Kuhn et al., 2023), a widely used UQ approach, measures uncertainty based on semantic dispersion. However, in summarization tasks, summary quality is judged primarily by its fidelity to the source content, rather than semantic variability alone. This presents key limitations: A summary may deviate semantically yet still provide a valid abstraction of the original text. Summarization evaluation involves coverage, conciseness, and coherence, which semantic entropy alone cannot quantify. Given these limitations, we focus on classification and multiple-choice tasks, which offer a robust evaluation framework for analyzing uncertainty evolution in long-context ICL.
+
+# B.2 Limitations of UQ for CoT
+
+In CoT tasks, uncertainty accumulates throughout the reasoning process, influencing the final answer. This uncertainty propagation occurs in intermediate reasoning steps, and the final answer generation. Current UQ techniques primarily focus on single-step inference or static tasks, whereas CoT relies on multi-step reasoning. This multi-stage nature makes it difficult for existing methods to effectively capture uncertainty propagation across reasoning steps. While research on CoT uncertainty is still in its early stages, some prior works have explored possible approaches. For instance, some work proposed a stepwise scoring mechanism which assigns a confidence score to each intermediate explanation. However, this approach has notable limitations: (1) Overconfidence: LLMs tend to be overconfident in their predictions, making single-step confidence scores unreliable; (2) Lack of global coherence: stepwise scoring ignores dependencies across reasoning steps, failing to capture uncertainty propagation across the entire reasoning chain; (3). Step mismatch: the reasoning steps generated may not align with the logical steps required for complex reasoning tasks, limiting the effectiveness in capturing uncertainty flow.
+
+Potential Strategies: A Topological Perspective To better model uncertainty propagation in CoT reasoning, we propose leveraging topological structures. CoT reasoning typically involves problem decomposition, backtracking and correction, evaluation and verification, and final integration. While current models generate reasoning in an autoregressive (linear) manner, actual human reasoning follows a more complex topological structure. Inspired by Tree-based CoT (Yao et al., 2023) and Graph-based CoT (Besta et al., 2024), we propose modeling CoT uncertainty using graph or tree structures. In this framework: each reasoning step is represented as a node; uncertainty from prior steps propagates
+
+through the topological structure to influence subsequent steps; the final answer (root node) aggregates the propagated uncertainties from all previous steps. By explicitly modeling uncertainty flow in a structured manner, this approach could overcome the limitations of stepwise scoring and offer a systematic framework for analyzing uncertainty evolution in multi-step reasoning. We believe this direction holds promise for improving uncertainty estimation in CoT-based tasks.
+
+# C Generalization Results on Larger LLMs
+
+# C.1 Qwen2.5-14B-Instruct
+
+<table><tr><td>AG_News</td><td>2</td><td>4</td><td>8</td><td>16</td><td>32</td><td>64</td><td>128</td><td>256</td></tr><tr><td>TU</td><td>0.148</td><td>0.127</td><td>0.113</td><td>0.125</td><td>0.115</td><td>0.105</td><td>0.086</td><td>0.065</td></tr><tr><td>EU</td><td>0.057</td><td>0.029</td><td>0.030</td><td>0.033</td><td>0.038</td><td>0.040</td><td>0.028</td><td>0.026</td></tr><tr><td>AU</td><td>0.091</td><td>0.098</td><td>0.083</td><td>0.092</td><td>0.077</td><td>0.065</td><td>0.058</td><td>0.039</td></tr><tr><td>ACC</td><td>87.6</td><td>87.2</td><td>88.9</td><td>88.19</td><td>88.7</td><td>88.5</td><td>89.9</td><td>90.5</td></tr></table>
+
+Table 4: Performance of Qwen2.5-14B-Instruct on AG_News with varying numbers of in-context examples  
+
+<table><tr><td>LD5</td><td>1</td><td>10</td><td>20</td><td>40</td><td>80</td><td>120</td><td>240</td></tr><tr><td>TU</td><td>0.345</td><td>0.307</td><td>0.302</td><td>0.257</td><td>0.279</td><td>0.272</td><td>0.245</td></tr><tr><td>EU</td><td>0.229</td><td>0.194</td><td>0.190</td><td>0.148</td><td>0.157</td><td>0.139</td><td>0.124</td></tr><tr><td>AU</td><td>0.116</td><td>0.112</td><td>0.112</td><td>0.109</td><td>0.122</td><td>0.133</td><td>0.121</td></tr><tr><td>ACC</td><td>62.4</td><td>63.6</td><td>64.4</td><td>68.4</td><td>67.2</td><td>72.1</td><td>72.8</td></tr></table>
+
+# C.2 Qwen2.5-32B-Instruct
+
+Table 5: Performance of Qwen2.5-14B-Instruct on LD5 with varying numbers of in-context examples  
+
+<table><tr><td>AG_News</td><td>2</td><td>4</td><td>8</td><td>16</td><td>32</td><td>64</td><td>128</td></tr><tr><td>TU</td><td>0.220</td><td>0.171</td><td>0.151</td><td>0.099</td><td>0.076</td><td>0.060</td><td>0.049</td></tr><tr><td>EU</td><td>0.151</td><td>0.093</td><td>0.059</td><td>0.030</td><td>0.017</td><td>0.020</td><td>0.018</td></tr><tr><td>AU</td><td>0.069</td><td>0.078</td><td>0.091</td><td>0.068</td><td>0.059</td><td>0.040</td><td>0.030</td></tr><tr><td>ACC</td><td>88.9</td><td>86.8</td><td>87.1</td><td>89.5</td><td>89.5</td><td>92.4</td><td>92.8</td></tr></table>
+
+Table 6: Performance of Qwen2.5-32B-Instruct on AG_News with varying numbers of in-context examples  
+
+<table><tr><td>LD5</td><td>1</td><td>10</td><td>20</td><td>40</td><td>80</td><td>120</td></tr><tr><td>TU</td><td>0.353</td><td>0.313</td><td>0.284</td><td>0.247</td><td>0.225</td><td>0.202</td></tr><tr><td>EU</td><td>0.121</td><td>0.102</td><td>0.102</td><td>0.099</td><td>0.091</td><td>0.082</td></tr><tr><td>AU</td><td>0.232</td><td>0.210</td><td>0.182</td><td>0.147</td><td>0.134</td><td>0.120</td></tr><tr><td>ACC</td><td>74.8</td><td>79.6</td><td>82.0</td><td>82.4</td><td>83.6</td><td>84.1</td></tr></table>
+
+Table 7: Performance of Qwen2.5-32B-Instruct on LD5 with varying numbers of in-context examples
+
+# D Quality of UQ for other LLMs
+
+<table><tr><td rowspan="3">Dataset</td><td colspan="8">Llama-3.1-8B</td></tr><tr><td>1-shot</td><td>2-shot</td><td>4-shot</td><td>8-shot</td><td>16-shot</td><td>32-shot</td><td>64-shot</td><td>128-shot</td></tr><tr><td colspan="8">Easy Mode</td></tr><tr><td>AGNews</td><td>0.686</td><td>0.704</td><td>0.725</td><td>0.735</td><td>0.780</td><td>0.804</td><td>0.822</td><td>0.837</td></tr><tr><td>SST-2</td><td>0.714</td><td>0.751</td><td>0.751</td><td>0.748</td><td>0.740</td><td>0.741</td><td>0.742</td><td>0.750</td></tr><tr><td>Commonsense QA</td><td>0.563</td><td>0.599</td><td>0.636</td><td>0.673</td><td>0.726</td><td>0.774</td><td>0.784</td><td>0.798</td></tr><tr><td></td><td colspan="8">Hard Mode</td></tr><tr><td></td><td>1-shot</td><td>4-shot</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>120-shot</td><td>240-shot</td></tr><tr><td>Logical Deduction 3</td><td>0.973</td><td>0.965</td><td>0.939</td><td>0.948</td><td>0.951</td><td>0.939</td><td>0.966</td><td>0.947</td></tr><tr><td>Logical Deduction 5</td><td>0.996</td><td>0.995</td><td>0.963</td><td>0.983</td><td>0.971</td><td>0.983</td><td>0.959</td><td>0.974</td></tr><tr><td>Logical Deduction 7</td><td>0.987</td><td>0.997</td><td>0.976</td><td>0.987</td><td>0.982</td><td>0.986</td><td>0.986</td><td>0.964</td></tr></table>
+
+Table 8: AUROC of Llama-3.1-8B model. High AUROC indicates the good quality of UQ measures.  
+
+<table><tr><td rowspan="3">Dataset</td><td colspan="8">Mistral-7B-v0.2</td></tr><tr><td>1-shot</td><td>2-shot</td><td>4-shot</td><td>8-shot</td><td>16-shot</td><td>32-shot</td><td>64-shot</td><td>128-shot</td></tr><tr><td colspan="8">Easy Mode</td></tr><tr><td>AGNews</td><td>0.633</td><td>0.696</td><td>0.734</td><td>0.753</td><td>0.769</td><td>0.778</td><td>0.790</td><td>0.780</td></tr><tr><td>SST-2</td><td>0.714</td><td>0.723</td><td>0.685</td><td>0.772</td><td>0.813</td><td>0.849</td><td>0.846</td><td>0.871</td></tr><tr><td>Commonsense QA</td><td>0.739</td><td>0.728</td><td>0.731</td><td>0.733</td><td>0.743</td><td>0.711</td><td>0.710</td><td>0.728</td></tr><tr><td></td><td colspan="8">Hard Mode</td></tr><tr><td></td><td>1-shot</td><td>4-shot</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>120-shot</td><td>240-shot</td></tr><tr><td>Logical Deduction 3</td><td>0.956</td><td>0.987</td><td>0.951</td><td>0.976</td><td>0.951</td><td>0.986</td><td>0.966</td><td>0.947</td></tr><tr><td>Logical Deduction 5</td><td>0.938</td><td>0.929</td><td>0.918</td><td>0.922</td><td>0.934</td><td>0.913</td><td>0.918</td><td>0.912</td></tr><tr><td>Logical Deduction 7</td><td>0.923</td><td>0.939</td><td>0.928</td><td>0.919</td><td>0.925</td><td>0.925</td><td>0.936</td><td>0.946</td></tr></table>
+
+Table 9: AUROC of Mistral-7B-v0.2 model. High AUROC indicates the good quality of UQ measures.  
+
+<table><tr><td rowspan="3">Dataset</td><td colspan="8">Qwen1.5-7B</td></tr><tr><td>1-shot</td><td>2-shot</td><td>4-shot</td><td>8-shot</td><td>16-shot</td><td>32-shot</td><td>64-shot</td><td>128-shot</td></tr><tr><td colspan="8">Easy Mode</td></tr><tr><td>AGNews</td><td>0.634</td><td>0.716</td><td>0.743</td><td>0.744</td><td>0.688</td><td>0.739</td><td>0.731</td><td>0.741</td></tr><tr><td>SST-2</td><td>0.742</td><td>0.766</td><td>0.842</td><td>0.854</td><td>0.872</td><td>0.870</td><td>0.870</td><td>0.879</td></tr><tr><td>Commonsense QA</td><td>0.768</td><td>0.818</td><td>0.801</td><td>0.801</td><td>0.799</td><td>0.776</td><td>0.772</td><td>0.770</td></tr><tr><td></td><td colspan="8">Hard Mode</td></tr><tr><td></td><td>1-shot</td><td>4-shot</td><td>10-shot</td><td>20-shot</td><td>40-shot</td><td>80-shot</td><td>120-shot</td><td>240-shot</td></tr><tr><td>Logical Deduction 3</td><td>0.875</td><td>0.846</td><td>0.918</td><td>0.900</td><td>0.928</td><td>0.871</td><td>0.966</td><td>0.788</td></tr><tr><td>Logical Deduction 5</td><td>0.935</td><td>0.918</td><td>0.903</td><td>0.849</td><td>0.962</td><td>0.934</td><td>0.921</td><td>0.912</td></tr><tr><td>Logical Deduction 7</td><td>0.923</td><td>0.879</td><td>0.893</td><td>0.911</td><td>0.934</td><td>0.925</td><td>0.946</td><td>0.956</td></tr></table>
+
+Table 10: AUROC of Qwen1.5-7B model. High AUROC indicates the good quality of UQ measures.
+
+# E Question-level Analysis
+
+<table><tr><td rowspan="3">Dataset</td><td colspan="10">Mistral-7B-v0.2</td></tr><tr><td colspan="2">8-shot</td><td colspan="2">16-shot</td><td colspan="2">32-shot</td><td colspan="2">64-shot</td><td colspan="2">128-shot</td></tr><tr><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td></tr><tr><td></td><td colspan="10">Easy Mode</td></tr><tr><td rowspan="2">AG News</td><td>61.4</td><td>+8.1</td><td>70.9</td><td>+9.5</td><td>77.9</td><td>+10.7</td><td>76.6</td><td>+11.4</td><td>78.8</td><td>+12.5</td></tr><tr><td>34.8</td><td>-4.9</td><td>25.1</td><td>-4.2</td><td>18.75</td><td>-3.6</td><td>19.15</td><td>-3.4</td><td>16.45</td><td>-2.8</td></tr><tr><td rowspan="2">SST-2</td><td>67.3</td><td>+9.1</td><td>79.2</td><td>+12.8</td><td>86.8</td><td>+13.7</td><td>87.9</td><td>+13.8</td><td>90.0</td><td>+14.6</td></tr><tr><td>27.3</td><td>-0.7</td><td>16.3</td><td>-0.0</td><td>10.8</td><td>-0.5</td><td>9.7</td><td>-0.3</td><td>8.6</td><td>-0.3</td></tr><tr><td rowspan="2">Commonsense QA</td><td>49.8</td><td>+1.8</td><td>40.0</td><td>+1.4</td><td>37.2</td><td>+3.4</td><td>38</td><td>+1.6</td><td>36.0</td><td>+2.0</td></tr><tr><td>21.8</td><td>+0.4</td><td>20.2</td><td>+0.4</td><td>18.4</td><td>-1.4</td><td>19.2</td><td>+0.4</td><td>23.6</td><td>-0.8</td></tr><tr><td rowspan="3"></td><td colspan="10">Hard Mode</td></tr><tr><td colspan="2">20-shot</td><td colspan="2">40-shot</td><td colspan="2">80-shot</td><td colspan="2">120-shot</td><td colspan="2">240-shot</td></tr><tr><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td></tr><tr><td rowspan="2">Logical Deduction3</td><td>80.4</td><td>+13.6</td><td>78.4</td><td>+20.4</td><td>79.3</td><td>+20.5</td><td>76.8</td><td>+18.4</td><td>81.6</td><td>+20.4</td></tr><tr><td>9.6</td><td>-0.4</td><td>9.2</td><td>-0.8</td><td>44.4</td><td>-6.0</td><td>12.4</td><td>-0.4</td><td>9.6</td><td>-1.6</td></tr><tr><td rowspan="2">Logical Deduction5</td><td>31.6</td><td>+0.4</td><td>39.6</td><td>+0.8</td><td>48.12</td><td>+0.0</td><td>57.9</td><td>+0.0</td><td>79.2</td><td>+1.5</td></tr><tr><td>41.2</td><td>-0.4</td><td>36.8</td><td>-0.8</td><td>33.0</td><td>-0.9</td><td>24.0</td><td>-0.8</td><td>15.4</td><td>-0.5</td></tr><tr><td rowspan="2">Logical Deduction7</td><td>46.8</td><td>+0.0</td><td>60</td><td>+0.0</td><td>62.8</td><td>+0.0</td><td>47.2</td><td>+0.0</td><td>96.5</td><td>+0.0</td></tr><tr><td>38.8</td><td>-0.0</td><td>25.2</td><td>-0.0</td><td>28.4</td><td>-0.0</td><td>34.8</td><td>-0.0</td><td>1.17</td><td>-0.0</td></tr></table>
+
+Table 11:  $\Delta$  U denotes the proportion of datasets whose uncertainty decreases/increases compared to 4-shot settings, with the first line for each dataset giving the ratio of decreased uncertainty questions and the second line for each dataset giving the ratio of increased uncertainty questions.  $\Delta$  Acc represents the performance changes caused by the corresponding part of examples.  
+
+<table><tr><td rowspan="3">Dataset</td><td colspan="10">Qwen1.5-7B</td></tr><tr><td colspan="2">8-shot</td><td colspan="2">16-shot</td><td colspan="2">32-shot</td><td colspan="2">64-shot</td><td colspan="2">128-shot</td></tr><tr><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td></tr><tr><td></td><td colspan="10">Easy Mode</td></tr><tr><td rowspan="2">AG News</td><td>62.2</td><td>+3.7</td><td>40.4</td><td>-0.8</td><td>79.5</td><td>+9.3</td><td>83.9</td><td>+9.8</td><td>83.8</td><td>+10.0</td></tr><tr><td>34.4</td><td>-0.8</td><td>52.0</td><td>-10.0</td><td>17.25</td><td>-0.6</td><td>13.0</td><td>-0.5</td><td>12.6</td><td>-1.4</td></tr><tr><td rowspan="2">SST-2</td><td>78.0</td><td>+0.2</td><td>86.6</td><td>+0.0</td><td>82.9</td><td>+0.1</td><td>77.3</td><td>-0.5</td><td>71.8</td><td>-0.3</td></tr><tr><td>13.0</td><td>-0.5</td><td>4.3</td><td>-0.1</td><td>2.9</td><td>-0.5</td><td>1.8</td><td>-0.1</td><td>2.4</td><td>-0.3</td></tr><tr><td rowspan="2">Commonsense QA</td><td>33.6</td><td>+0.0</td><td>12.2</td><td>+0.4</td><td>11.0</td><td>+0.2</td><td>8.9</td><td>+0.81</td><td>15.0</td><td>+0.4</td></tr><tr><td>48.1</td><td>-3.6</td><td>68.6</td><td>-10.4</td><td>76.4</td><td>-14.4</td><td>79.7</td><td>-1.6</td><td>72.6</td><td>-2.4</td></tr><tr><td rowspan="3"></td><td colspan="10">Hard Mode</td></tr><tr><td colspan="2">20-shot</td><td colspan="2">40-shot</td><td colspan="2">80-shot</td><td colspan="2">120-shot</td><td colspan="2">240-shot</td></tr><tr><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td><td>ΔU</td><td>ΔAcc</td></tr><tr><td rowspan="2">Logical Deduction3</td><td>13.6</td><td>+2.0</td><td>13.2</td><td>+1.6</td><td>9.2</td><td>+0.4</td><td>20.8</td><td>+1.6</td><td>36.0</td><td>+6.0</td></tr><tr><td>84</td><td>-13.6</td><td>82.4</td><td>-16.0</td><td>88.4</td><td>-22.8</td><td>77.2</td><td>-16.8</td><td>60.8</td><td>-10.4</td></tr><tr><td rowspan="2">Logical Deduction5</td><td>73.6</td><td>+7.19</td><td>74.8</td><td>+3.6</td><td>50.8</td><td>+0.0</td><td>49.2</td><td>+0.0</td><td>69.2</td><td>+6.4</td></tr><tr><td>24.0</td><td>-2.8</td><td>24.4</td><td>-3.6</td><td>47.2</td><td>-12.0</td><td>47.6</td><td>-11.2</td><td>15.4</td><td>-5.6</td></tr><tr><td rowspan="2">Logical Deduction7</td><td>81.2</td><td>+4.8</td><td>54.4</td><td>+2.0</td><td>44.0</td><td>-0.4</td><td>41.1</td><td>-1.6</td><td>52.5</td><td>+10.3</td></tr><tr><td>17.6</td><td>-0.8</td><td>100</td><td>-20.8</td><td>54.8</td><td>-11.6</td><td>58.9</td><td>-16.1</td><td>45.3</td><td>-9.2</td></tr></table>
+
+Table 12:  $\Delta$  U denotes the proportion of datasets whose uncertainty decreases/increases compared to 4-shot settings, with the first line for each dataset giving the ratio of decreased uncertainty questions and the second line for each dataset giving the ratio of increased uncertainty questions.  $\Delta$  Acc represents the performance changes caused by the corresponding part of examples.
+
+# F Interprebility for k-shot ICL
+
+# F.1 Case Study
+
+![](images/6a03d0e9e2bf619fbc61dea0719ed2973f22c009344e026f8eac560793ebb489.jpg)  
+(a)  $\Delta$ -shot
+
+![](images/8bf364ec43e01f0c5383f417c7503e8d2959fb95fcab3f83ac1c0aa4fc0f2ca1.jpg)  
+(b) 32-shot
+
+![](images/f994632fa5aad5f654842bde20db42cdb416e9e76b83e652e5fce1dbcc338ad6.jpg)  
+(a) 64-shot
+
+![](images/57d02f073b2925d34ea2ed1148a338f9e36eebe4603b93e9297e1d9cb5611c28.jpg)  
+(b) 128-shot
+
+# F.2 Additional results: Average logits and probabilities
+
+![](images/c247937b169cf3d11d417ebf5ba5f67941cde06ea8148f06b6503f6712719761.jpg)  
+Figure 12: The inner confidence changes (0-1 probability) of five options ["A", "B", "C", "D", "E"] for a specific question in Commonsense QA for Mistral-7B-v0.2 under 4-shot (a), 32-shot (b), 64-shot (c), and 128-shot(d) ICL. The correct option is "E" and LLMs only made a mistake under 4-shot ICL and got correct with more examples.
+
+![](images/a6993285f0afc8c53dd5b4818c6487cd9234eed622c4fa7908b3f5d6eb4c8cc3.jpg)
+
+![](images/f75dfb90051c5b455d5bf2eed2395a5c0133c3895e0b421605560a3a709f8a06.jpg)
+
+![](images/0c630ba5a9dc2cd2732a6be7fddfee17d965ec62ad0965baa58cee071da1902b.jpg)
+
+![](images/2cea4bb85604084a48da517cbe8173021b9ab6bec29a418ef074f4f78607f99a.jpg)  
+Figure 14: Average logits and probabilities of Mistral-7B-v0.2 on the Commonsense QA dataset for MCQA items where the correct answer is "A".
+
+![](images/2454386b74e6de1936797b3dff034f34348b8fdf4f1ab8da28052a093c1e1907.jpg)
+
+![](images/b4b46d89ca8959458f922878db23be9db4aa2af3de3d1994371c6753d432749e.jpg)
+
+![](images/929285ecdd64ed6a283c67950519d9aaf173fc36934f6bf756cd9168457ea171.jpg)
+
+# G AI Assistant Usage
+
+We used chatgpt to assist with correcting spelling errors in writing.
+
+# H Experimental Details
+
+# H.1 Prompt templates
+
+Classify the topic of the following sentence into four labels: [0: world, 1: sports, 2: business, 3: Sci/Tech] Provide answer in a structured format WITHOUT additional comments, I just want the numerical label for each sentence.
+
+Sentence: Fears for T N pension after talks Unions representing workers at Turner Newall say they are 'disappointed' after talks with stricken parent firm Federal Mogul. Label: 2
+
+Sentence: The Race is On: Second Private Team Sets Launch Date for Human Spaceflight (SPACE.com)  
+SPACE.com - TORONTO, Canada -- A second\team of rocketers competing for the #36;10 million  
+Ansari X Prize, a contest for\privately funded suborbital space flight, has officially announced the  
+first\launch date for its manned rocket.  
+Label: 3
+
+Sentence: They've caught his eye In quote; helping themselves, quote; Ricky Bryant, Chas Gessner, Michael Jennings, and David Patten did nothing Friday night to make Bill Belichick's decision on what to do with his receivers any easier.  
+Label: 1
+
+Sentence: Sister of man who died in Vancouver police custody slams chief (Canadian Press) Canadian Press - VANCOUVER (CP) - The sister of a man who died after a violent confrontation with police has demanded the city's chief constable resign for defending the officer involved.
+
+# Label:
+
+Classify the topic of the following sentence into four labels: [0: world, 1: sports, 2: business, 3: Sci/Tech] Provide answer in a structured format WITHOUT additional comments, I just want the numerical label for each sentence.
+
+Figure 15: Prompt template with a test input for AG News dataset.
+
+Classify the following sentence into two categories: [0: negative, 1: positive]  
+Provide answer in a structured format WITHOUT additional comments, I just want the numerical label for each sentence.
+
+Sentence: that loves its characters and communicates something rather beautiful about human nature  
+Label: 1
+
+Sentence: remains utterly satisfied to remain the same throughout Label: 0
+
+Sentence: on the worst revenge-of-the-nerds clichés the filmmakers could dredge up.  
+Label: 0
+
+··
+
+Sentence: that 's far too tragic to merit such superficial treatment
+Label: 0
+
+Sentence: very well-written and very well-acted.  
+Label: 1
+
+Sentence: clumsy dialogue, heavy-handed phoney-feeling sentiment, Label: 0
+
+Classify the following sentence into two categories: [0: negative, 1: positive]  
+Provide answer in a structured format WITHOUT additional comments, I just want the numerical label for each sentence.
+
+Figure 16: Prompt template with a test input for SST-2 dataset.
+
+Select the correct answer for the following commonsense question from five choices.
+
+Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+Question: The sanctions against the school were a punishing blow, and they seemed to what the efforts the school had made to change?
+
+# Choices
+
+A. ignore  
+B. enforce  
+C. authoritarian  
+D. yell at  
+E. avoid
+
+Answer: A
+
+Question: To locate a choker not located in a jewelry box or boutique where would you go?
+
+# Choices
+
+A. jewelry store  
+B. neck  
+C. jewelry box  
+D. jewelry box  
+E. boutique
+
+# Answer:
+
+Select the correct answer for the following commonsense question from five choices.
+
+Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+Figure 17: Prompt template with a test input for Commonsense QA dataset.
+
+Select the correct answer for the following logical deduction problem from three choices.
+
+Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+The following paragraphs each describe a set of three objects arranged in a fixed order. The statements are logically consistent within each paragraph. On a branch, there are three birds: a blue jay, a quail, and a falcon. The falcon is to the right of the blue jay. The blue jay is to the right of the quail.
+
+# Options:
+
+(A) The blue jay is the second from the left  
+(B) The quail is the second from the left  
+(C) The falcon is the second from the left.
+
+# Answer: (A)
+
+The following paragraphs each describe a set of three objects arranged in a fixed order. The statements are logically consistent within each paragraph. On a shelf, there are three books: a blue book, an orange book, and a red book. The blue book is the rightmost. The orange book is the leftmost.
+
+# Options:
+
+(A) The blue book is the second from the left  
+(B) The orange book is the second from the left  
+(C) The red book is the second from the left
+
+# Answer:
+
+Select the correct answer for the following logical deduction problem from three choices.
+
+Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+Figure 18: Prompt template with a test input for logical deduction three objects dataset.
+
+Select the correct answer for the following logical deduction problem from five choices.
+
+Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+Problem: The following paragraphs each describe a set of five objects arranged in a fixed order. The statements are logically consistent within each paragraph. On a branch, there are five birds: a quail, an owl, a raven, a falcon, and a robin. The owl is the leftmost. The robin is to the left of the raven. The quail is the rightmost. The raven is the third from the left.
+
+Options:
+
+(A) The quail is the rightmost  
+(B) The owl is the rightmost  
+(C) The raven is the rightmost  
+(D) The falcon is the rightmost  
+(E) The robin is the rightmost
+
+# Answer: (A)
+
+Problem: The following paragraphs each describe a set of five objects arranged in a fixed order. The statements are logically consistent within each paragraph. In an antique car show, there are five vehicles: a hatchback, a bus, a convertible, a tractor, and a minivan. The tractor is older than the bus. The minivan is newer than the bus. The hatchback is the second-newest. The minivan is older than the convertible. Options:
+
+(A) The hatchback is the second-oldest  
+(B) The bus is the second-oldest  
+(C) The convertible is the second-oldest  
+(D) The tractor is the second-oldest  
+(E) The minivan is the second-oldest
+
+# Answer:
+
+Select the correct answer for the following logical deduction problem from three choices. Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+Figure 19: Prompt template with a test input for logical deduction five objects dataset.
+
+Select the correct answer for the following logical deduction problem from seven choices.
+
+Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+Problem: The following paragraphs each describe a set of seven objects arranged in a fixed order. The statements are logically consistent within each paragraph. In a golf tournament, there were seven golfers: Ana, Eve, Ada, Dan, Rob, Amy, and Joe. Dan finished third. Ana finished above Ada. Amy finished last. Dan finished below Rob. Eve finished below Ada. Rob finished below Joe.
+
+Options:
+
+(A) Ana finished third
+
+(B) Eve finished third
+
+(C) Ada finished third
+
+(D) Dan finished third
+
+(E) Rob finished third
+
+(F) Amy finished third
+
+(G) Joe finished third
+
+# Answer: (D)
+
+Problem: The following paragraphs each describe a set of seven objects arranged in a fixed order. The statements are logically consistent within each paragraph. In an antique car show, there are seven vehicles: a bus, a motorcycle, a hatchback, a station wagon, a minivan, a truck, and a limousine. The station wagon is the fourth-newest. The motorcycle is newer than the truck. The station wagon is older than the hatchback. The minivan is newer than the hatchback. The bus is newer than the minivan. The truck is newer than the limousine.
+
+Options:
+
+(A) The bus is the third-oldest
+
+(B) The motorcycle is the third-oldest (C) ...
+
+Answer:
+
+Select the correct answer for the following logical deduction problem from seven choices.
+
+Provide answer in a structured format WITHOUT additional comments, I just want the option letter for each answer.
+
+Figure 20: Prompt template with a test input for logical deduction seven objects dataset.

2025/Uncertainty Unveiled_ Can Exposure to More In-context Examples Mitigate Uncertainty for Large Language Models_/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:120494ebf5c71d5021fbc9ee8ce177731ff463697957b6f1e7922ccbcf6c227f
+size 1627719

2025/Uncertainty-Aware Contrastive Decoding/d9a15b33-5d81-4410-9837-722f1c358cf5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b541d6aad17be3cd593c20ca9cc40aff6f9c59ab15a2c07466dcf3096e4ff97
+size 1376842

2025/Uncertainty-Aware Contrastive Decoding/full.md

@@ -0,0 +1,599 @@
+# Uncertainty-Aware Contrastive Decoding
+
+Hakyung Lee $^{1*}$ , Subeen Park $^{1*}$ , Joowang Kim $^{2}$ , Sungjun Lim $^{1}$ , Kyungwoo Song $^{1\dagger}$
+
+$^{1}$ Yonsei University, Republic of Korea
+
+$^{2}$ LG CNS, Republic of Korea
+
+# Abstract
+
+Large language models excel in a wide range of natural language processing tasks, but generating factually accurate and consistent outputs remains a challenge. To improve text reliability, Contrastive Decoding (CD) refines token selection by leveraging differences between an expert and base model, penalizing low-quality token choices. However, CD employs static weighting between models, making it sensitive to variations in model architecture and input characteristics, often resulting in suboptimal token selection and error propagation throughout generation. We propose Uncertainty-Aware Contrastive Decoding (UCD), a method that dynamically adjusts model contributions at each decoding step based on uncertainty. We introduce a cumulative energy function, where uncertainty is quantified as the negative log-sum-exp over logits, and decomposed into entropy and expected logit components. This energy serves as a dynamic confidence signal, guiding adaptive model weighting during generation. We demonstrate through extensive experiments that UCD significantly improves factual accuracy and reliability over existing decoding methods. Finally, we provide a theoretical analysis showing that our energy function serves as a well-defined uncertainty metric capturing model confidence. Our code is available at https://github.com/MLAI-Yonsei/UCD.
+
+# 1 Introduction
+
+Large language models demonstrated remarkable performance across diverse tasks, ranging from open-ended text generation to structured reasoning (Brown et al., 2020; Kojima et al., 2022). Their ability to generate human-like text and generalize across domains has led to widespread adoption. However, reliability remains a persistent challenge, as LLMs are prone to repetition, hallucination and
+
+![](images/9b0e061620103c95028197d736ef8113e9e30ec1c63eafd4db39ee5d7a053fde.jpg)  
+Figure 1: Comparison of CD and UCD in token selection. UCD's dynamic weighting accurately predicts "August 2, 1776," while CD fails.
+
+factual inconsistencies (Ji et al., 2023). Standard decoding strategies, such as greedy decoding and sampling, struggle to balance coherence, diversity and factuality, often producing deterministic yet incorrect generations (Holtzman et al., 2019).
+
+Recent research suggests that relying on a single LLM is inherently limiting, as no single model can fully capture the diverse reasoning strategies, domain knowledge, and stylistic variations required for robust text generation (Feng et al., 2025). To address this limitation, multi-LLM collaboration techniques such as contrastive objectives and ensemble methods have been explored to leverage the strengths of different models (Shen et al., 2024; Mavromatis et al., 2024).
+
+One prominent approach, Contrastive Decoding (CD), improves text generation without additional training by refining token selection at inference time (Li et al., 2023). The key insight is that smaller models amplify undesirable behaviors, such as repetition and fluency errors, whereas larger models, despite their strengths, still exhibit hallucinations and factual inconsistencies. By contrasting their predictions, CD filters out low-quality outputs, leading to improved fluency and coherence. Building on this foundation, several works have been proposed to enhance the effectiveness of CD. Liu et al. 2024 improve adaptation by using a three-model
+
+setup (Large LM, Small LM, and Fine-Tuned Small LM) to simulate fine-tuning effects on the Large LM without directly updating its parameters. Other works (Liu et al., 2021; Zhang et al., 2023) apply contrastive objective to hallucination-induced models to further enhance factual accuracy. Despite these advancements, existing methods rely on static weighting schemes or heuristic search to determine model contributions. However, the optimal balance between models should be dynamic, adapting to the selected models, task requirements, and dataset characteristics. For instance, a fine-tuned small model may be more reliable for domain-specific tasks warranting a higher weighting, whereas a pretrained large model excels in open-ended reasoning task, where it should contribute more.
+
+Few prior works has explored dynamic weighting to address these issues. Fan et al. introduce KL divergence-based weighting, adjusting model contributions per decoding step by comparing their output distributions. However, this method assumes a fixed distance between pretrained and fine-tuned models across different model sizes, oversimplifying model interactions. Mavromatis et al. (2024) optimize model contributions using perplexity, but focus on ensembling models rather than applying contrastive weighting. Moreover, both approaches rely on a search-based optimization process at each step, increasing computational overhead. Also they do not fully account for the sequential nature of uncertainty, neglecting cumulative effects that influence model reliability over time.
+
+In this paper, we propose Uncertainty-Aware Contrastive Decoding (UCD), a framework that applies energy-based uncertainty estimation to CD. UCD models uncertainty sequentially using a cumulative energy function, allowing dynamic reweighting of model contributions at each step without iterative optimization, ensuring efficient and adaptive token selection. As illustrated in Figure 1, UCD's dynamic weighting accurately predicts the correct answer "August 2, 1776" for a historical question, while CD (O'Brien and Lewis, 2023) fails due to static weighting.
+
+We theoretically validate our energy function through LSE decomposition, demonstrating its role as an uncertainty metric for sequential decoding. By modeling uncertainty accumulation, UCD enhances decoding stability and robustness. Extensive experiments on hallucination datasets and reasoning benchmarks confirm that UCD consistently outperforms existing methods, achieving higher
+
+reliability and factual accuracy.
+
+# 2 Related Works
+
+# 2.1 Contrastive Decoding
+
+Contrastive Decoding enhances text generation by contrasting output predictions between large and small models (Li et al., 2023). Beyond its original formulation, the concept of CD expanded to multiple model contrast (Liu et al., 2024), contextual contrast (Zhao et al., 2024), and layer-wise contrast (Chuang et al., 2023), increasing adaptability across different settings. Additionally, its application has broadened to include task-specific fine-tuned models (Shen et al., 2024), aligned models (Fei et al., 2024), and hallucination-induced models (Zhang et al., 2023) each targeting distinct contrastive objectives. In this work, we introduce Uncertainty-Aware Contrastive Decoding (UCD), which dynamically adjusts model contributions based on uncertainty estimation, ensuring stable and effective text generation across diverse tasks and models.
+
+# 2.2 Uncertainty Estimation
+
+Prior research has shown that effectively measuring uncertainty helps identify instances where a model lacks confidence, thereby improving prediction quality (Siddhant and Lipton, 2018; Xiao and Wang, 2019; Kuhn et al.). Building on this, uncertainty estimation has become essential for enhancing model reliability, interpretability, and decision-making across various NLP tasks (Hu et al., 2023). For instance, Fan et al. utilized KL divergence for contrastive decoding, Das et al. (2024) leveraged entropy to select the contrast layer, and Mavromatis et al. (2024) introduced an ensemble method that assigns weights based on perplexity. While these methods rely on specific uncertainty measures, our work is the first to directly incorporate energy as an uncertainty indicator in the contrastive decoding process.
+
+# 3 Methodology
+
+# 3.1 Problem Statement
+
+In the context of open-ended text generation, an auto-regressive language model  $p_{\mathrm{LM}}$  takes a prompt  $\mathbf{x}_{\mathrm{prompt}} = (x_1,\dots ,x_n)$  of length  $n$ , and generates a continuation sequence  $\mathbf{x}_{\mathrm{cont}} = (x_{n + 1},\ldots ,x_{n + m})$  of length  $m$ . Each token  $x_{t}$  is drawn from a vocabulary  $\mathcal{V}$ . The probability of generating the continu-
+
+ation under  $p_{\mathrm{LM}}$  factorizes as:
+
+$$
+p _ {\mathrm {L M}} (\mathbf {x} _ {\mathrm {c o n t}} \mid \mathbf {x} _ {\mathrm {p r o m p t}}) = \prod_ {t = n + 1} ^ {n + m} p _ {\mathrm {L M}} \big (x _ {t} \mid \mathbf {x} _ {<   t} \big),
+$$
+
+where  $\mathbf{x}_{< t} = (x_1, \ldots, x_{t-1})$  is the previously generated context. Internally, the predictive distribution is computed by applying a softmax over the logit vector  $\mathbf{z}_t \in \mathbb{R}^{|\mathcal{V}|}$ :
+
+$$
+p _ {\mathrm {L M}} (x _ {t} = v \mid \mathbf {x} _ {<   t}) = \frac {\exp (\mathbf {z} _ {t} [ v ])}{\sum_ {v ^ {\prime} \in \mathcal {V}} \exp (\mathbf {z} _ {t} [ v ^ {\prime} ])}.
+$$
+
+Here,  $\mathbf{z}_t[v]$  denotes the unnormalized score assigned to token  $v\in \mathcal{V}$ .
+
+For decoding, we employ a greedy decoding strategy, where at each timestep  $t$ , the model selects the token with the highest logit probability:
+
+$$
+\hat {x} _ {t} = \arg \max  _ {x _ {t} \in \mathcal {V}} p _ {\mathrm {L M}} \left(x _ {t} \mid \hat {\mathbf {x}} _ {<   t}, \mathbf {x} _ {\text {p r o m p t}}\right), \tag {1}
+$$
+
+where  $\hat{\mathbf{x}}_{< t} = (\hat{x}_1,\dots ,\hat{x}_{t - 1})$  is the sequence of previously selected tokens. Repeating this step over  $m$  timesteps yields the final generated sequence:
+
+$$
+\hat {\mathbf {x}} _ {\mathrm {c o n t}} = (\hat {x} _ {n + 1}, \dots , \hat {x} _ {n + m}).
+$$
+
+# 3.2 Cumulative Energy-Contrastive Decoding
+
+Building on Contrastive Decoding (Li et al., 2023; Liu et al., 2024), we propose Uncertainty-Aware Contrastive Decoding, which dynamically balances the contributions of the expert model (EXP) and base model (BASE).
+
+# 3.2.1 Energy-based Uncertainty Estimation
+
+Given a prompt, the expert model EXP and base model BASE generate logits  $\mathbf{z}_t^{(\mathrm{EXP})}$  and  $\mathbf{z}_t^{(\mathrm{BASE})}$  at each decoding step  $t$ . To quantify the uncertainty of each model's predictions, we introduce a cumulative energy function, which estimates the energy of the output distribution and serves as a confidence measure for each model. This function integrates both the current logits and the historical uncertainty accumulated over previous decoding steps, ensuring that past decisions influence future token selection. This dynamic adjustment based on accumulated confidence makes the energy function a more stable and adaptive measure of model reliability.
+
+Logit Tracing To track how confident each model was about previously chosen tokens, we maintain logit traces  $\ell_t^{(\mathsf{M})} \in \mathbb{R}$ , where  $M \in \{\mathrm{EXP}, \mathrm{BASE}\}$ .
+
+This trace accumulates past logits corresponding to previously selected tokens. Specifically, at each timestep  $t$ , the trace is updated recursively as:
+
+$$
+\ell_ {t} ^ {(M)} = \beta \cdot \ell_ {t - 1} ^ {(M)} + \mathbf {z} _ {t} ^ {(M)} [ \hat {x} _ {t - 1} ],
+$$
+
+where  $\beta$  is a discount factor controlling the retention of past confidence. Unfolding this recursion yields:
+
+$$
+\ell_ {t} ^ {(M)} = \sum_ {k = 1} ^ {t - 1} \beta^ {t - 1 - k} \cdot \mathbf {z} _ {k} ^ {(M)} [ \hat {x} _ {k} ]. \tag {2}
+$$
+
+This formulation aggregates the logits assigned to previously selected tokens, weighted by their recency, and serves as a temporal memory of model confidence.
+
+Cumulative Energy We define the energy of a logit vector  $\mathbf{z}_t$  with respect to the logit trace  $\ell_t$  as:
+
+$$
+\operatorname {E n e r g y} \left(\mathbf {z} _ {t}, \ell_ {t}\right) = T \log \sum_ {v \in \mathcal {V}} \exp \left(\frac {z _ {t} [ v ] + \ell_ {t}}{T}\right). \tag {3}
+$$
+
+Here, temperature parameter  $T$  controls the sharpness of the softmax-like transformation. We define energy as negative log-probability to ensure that higher energy corresponds to higher confidence. The energy computation uses log-sum-exp over logits and accumulated traces, amplifying confident predictions while preserving efficiency. At each timestep  $t$ , we compute energy values for each model:
+
+$$
+E _ {t} ^ {(M)} = \operatorname {E n e r g y} (\mathbf {z} _ {t} ^ {(M)}, \ell_ {t} ^ {(M)}).
+$$
+
+Dynamic Weighting We derive per-timestep weights by normalizing the energies of the two models:
+
+$$
+w _ {t} ^ {(M)} = \frac {E _ {t} ^ {(M)}}{E _ {t} ^ {\mathrm {(E X P)}} + E _ {t} ^ {\mathrm {(B A S E)}}}
+$$
+
+Intuitively, a model with higher energy at a given timestep receives a larger weight, indicating greater confidence in its token predictions. This adaptive weighting mechanism allows the system to dynamically adjust to the relative reliability of the expert and base models throughout the decoding process, ensuring robust and context-aware token selection. Furthermore, when  $E_{t}^{(M)}$  is negative, it indicates that the model is highly uncertain about its own
+
+Question: According to the Bible, what forbidden fruit did Adam and Eve eat in the Garden of Eden?
+
+![](images/b577b352494eea22116379a22c4a0da1ecb647a5a33eb34eb534eba8b3375cc8.jpg)  
+Expert
+
+![](images/16e3cc28bc64193c458819df1d47cef164830a6de3a016d80aa8387aec6f4e41.jpg)  
+Output : What is the name of the first man created by God in the Bible?
+
+![](images/12886e733a213d349a6809164b873146f90cf9f494fddd695da871c559a004cb.jpg)
+
+Base  
+![](images/6c7c1a3cfb8f5389c0231e4e48b723f541dcf9c6940be9c9edf7e39023aee117.jpg)  
+Output : How many days did it take God to create the universe?
+
+![](images/1c97b91805478063b205d315cec5ab6e8d7923d28cc2347e0e95c5817c062517.jpg)
+
+![](images/078a21612daf568b0cd940ad2bc8b09096d65a54f7bea8822f8a8046543a8502.jpg)
+
+CD
+
+![](images/b54d666f91557a35b2a7580965035c8706847e67456553a1494711576e54655f.jpg)  
+Output : What forbidden fruit did Adam and Even eat in the Garden of Eden?
+
+![](images/54e49299ca5af3b4271fcfa634ad79b9d5a25bbd59bd166cf0121a4c279b7326.jpg)
+
+UCD
+
+Figure 2: Effect of cumulative energy-based weighting in sequential decoding. Compared to CD, UCD highlights the importance of cumulative uncertainty, leading to accurate token selection and the correct answer.  
+![](images/f5be352a4e241d7139d5469f8ab163c0e2f24ff30fcecb50042713f3e742447d.jpg)  
+Output : The forbidden fruit in the Garden of Eden was the fruit of the tree of the knowledge of good and evil.
+
+![](images/086022d129354d08e122eca5661fa0f719737fa9d958550ddbaa3211766786eb.jpg)
+
+predictions. Therefore, we considered combining models only when  $E_{t}^{(M)} > 0$ , ensuring that the reweighting process leverages sufficiently confident predictions.
+
+# 3.2.2 Uncertainty-Aware Token Selection
+
+We construct the energy-contrastive logit vector  $\mathbf{z}_{\mathrm{UCD}}^{(t)}$ , where each component  $v$  is computed as:
+
+$$
+\mathbf {z} _ {t} ^ {\mathrm {U C D}} [ v ] = 2 w _ {t} ^ {\mathrm {E X P}} \mathbf {z} _ {t} ^ {\mathrm {E X P}} [ v ] - w _ {t} ^ {\mathrm {B A S E}} \mathbf {z} _ {t} ^ {\mathrm {B A S E}} [ v ] \tag {4}
+$$
+
+At each timestep  $t$ , we select the token  $\hat{x}_t$  corresponding to the maximum UCD logit:
+
+$$
+\hat{x}_{t} = \operatorname *{arg  max}_{v\in \mathcal{V}}\mathbf{z}_{t}^{\mathrm{UCD}}[v].
+$$
+
+By repeating this process across all timesteps, we effectively solve Equation (1), thereby maximizing the joint probability throughout the sequence. Furthermore, Equation (4) can be extended to collaborative settings involving only expert models, without a base model. In this case, the logit vector that represents the contribution of the expert model is given by  $2w_{t}^{\mathrm{EXP}}\mathbf{z}_{t}^{\mathrm{EXP}}[v]$ . When multiple experts are involved, such as  $M \in \{\mathrm{EXP}_1,\mathrm{EXP}_2\}$ , the objective function naturally generalizes to  $w_{t}^{\mathrm{EXP}_{1}}\mathbf{z}_{t}^{\mathrm{EXP}_{1}}[v] + w_{t}^{\mathrm{EXP}_{2}}\mathbf{z}_{t}^{\mathrm{EXP}_{2}}[v]$ .
+
+# 3.3 Theoretical Analysis of Energy
+
+In this section, we analyze the cumulative energy defined in Equation 3. The following theorem shows that the cumulative energy can be decomposed into three components.
+
+Theorem 1. Let  $\mathbf{z}_t \in \mathbb{R}^{|\mathcal{V}|}$  be the logit vector at timestep  $t$ , and  $\ell_t \in \mathbb{R}$  be the logit trace. The cumulative energy, as defined in Equation (3), can be decomposed as:
+
+$$
+\operatorname {E n e r g y} \left(\mathbf {z} _ {t}, \ell_ {t}\right) = \underbrace {T \cdot H \left(p _ {t}\right)} _ {\text {u n c e r t a i n t y}} + \underbrace {\mathbb {E} _ {p _ {t}} \left[ z _ {t} \right] + \ell_ {t}} _ {\text {c o n f i d e n c e}}. \tag {5}
+$$
+
+where  $p_t(v) = \frac{e^{zt,v / T}}{\sum_{j\in\mathcal{V}}e^{zt,j / T}}$  is the softmax distribution over  $\mathbf{z}_t$ ,  $H(p_t)$  is its entropy, and  $\mathbb{E}_{p_t}[z_t]$  is its expected logit value.
+
+Theorem 1 demonstrates that cumulative energy jointly captures model uncertainty and confidence through three components: entropy, expected logit, and logit trace. Entropy  $H(p_{t})$  reflects the uncertainty of the current prediction, while the expected logit  $\mathbb{E}_{p_t}[z_t]$  and logit trace  $\ell_t$  quantify the model's confidence from instantaneous and historical perspectives, respectively. Together, these components define a principled energy-based confidence measure that evolves over time. To interpret the decom
+
+position, we detail how each term contributes distinctively to the modeling of uncertainty and confidence. Detailed proof is provided in Appendix A.1.
+
+Entropy. Entropy measures the dispersion of a probability distribution and serves as a quantitative indicator of model uncertainty. For a distribution  $p$ , the entropy is defined as:
+
+$$
+H (p) = - \sum_ {v \in \mathcal {V}} p (v) \log p (v).
+$$
+
+Higher entropy indicates a more uniform distribution, reflecting greater uncertainty in the model's predictions. Conversely, lower entropy corresponds to a distribution with concentrated mass on a few tokens, implying stronger confidence. In our framework, entropy serves as a principled measure of token-level uncertainty during decoding.
+
+Expected logit. The expected logit at timestep  $t$  is defined as:
+
+$$
+\mathbb {E} _ {p _ {t}} [ z _ {t} ] = \sum_ {v \in \mathcal {V}} p _ {t} (v) z _ {t} [ v ].
+$$
+
+This term reflects the average logit weighted by the model's predicted probability distribution. High expected logit indicates strong support for likely tokens, while a low value implies weaker predictions. In contrastive decoding, where token selection is based directly on logits, expected logit  $\mathbb{E}_{p_t}[z_t]$  serves as an approximate measure of overall model confidence. It summarizes how decisively the model supports its predictions by combining probability sharpness with logit magnitude.
+
+Logit Trace. The logit trace  $\ell_t$ , defined in Equation (2), is a time-weighted sum of logits assigned to previously selected tokens. This term serves two main purposes. First, it promotes temporal consistency by incorporating historical confidence, leading to more stable predictions across the sequence rather than isolated timestep decisions. Second, the exponential discounting smooths out transient fluctuations, with the discount factor  $\beta$  controlling how much recent decisions are emphasized.
+
+While both expected logit  $\mathbb{E}_{p_t}[z_t]$  and logit trace  $\ell_t$  quantify model confidence using logits, they differ in scope and temporal focus. The expected logit captures the model's overall confidence throughout the entire token distribution in a single timestep, while the logit trace aggregates confidence over previously selected tokens in multiple steps. Thus, the
+
+former reflects pointwise belief dispersion, while the latter encodes a history of committed decisions.
+
+To complete the interpretation, we analyze the temperature parameter  $T$ .
+
+Parameter  $T$ . The temperature parameter  $T$  controls the sharpness of the softmax distribution  $p_t(v)$  and scales the entropy term  $T \cdot H(p_t)$  in the energy formulation. We define a temperature-scaled log-sum-exp function as:
+
+$$
+\mathrm {L S E} _ {T} (\mathbf {z} _ {t}) := \mathbb {E} _ {p _ {t}} [ \mathbf {z} _ {t} ] + T \cdot H (p _ {t}),
+$$
+
+which smoothly interpolates between confidence and uncertainty based on  $T$ . Full derivations are provided in Appendix A.2.
+
+The limiting behavior with respect to  $T$  is given :
+
+$$
+\lim_{T\to 0}\operatorname{LSE}_{T}(\mathbf{z}_{t}) = \max_{v\in \mathcal{V}}\mathbf{z}_{t}[v],
+$$
+
+$$
+\lim  _ {T \rightarrow \infty} \mathrm {L S E} _ {T} (\mathbf {z} _ {t}) = T \log | \mathcal {V} | + \frac {1}{| \mathcal {V} |} \sum_ {v \in \mathcal {V}} \mathbf {z} _ {t} [ v ].
+$$
+
+As  $T \to 0$ ,  $\mathrm{LSE}_T(\mathbf{z}_t)$  approximates the maximum logit value, leading to confident and deterministic token selection. In contrast, as  $T \to \infty$ , the expected logit term  $\mathbb{E}_{p_t}[z_t]$  converges to a constant, while the entropy term  $T \cdot H(p_t)$  grows unbounded linearly with  $T$ . Thus, the cumulative energy becomes increasingly dominated by the entropy component, shifting the emphasis from model confidence to model uncertainty.
+
+This decomposition of cumulative energy into entropy, expected logit, and logit trace provides a unified measure of uncertainty and confidence. Leveraging cumulative energy in contrastive decoding enables adaptive model weighting, leading to more stable and adaptive token selection in open-ended text generation.
+
+# 4 Experimental Setup
+
+We conduct three experiments to evaluate our methodology. First, we assess the model's ability to mitigate hallucinations by measuring the truthfulness of its outputs. Second, we evaluate its robustness across diverse reasoning benchmarks to demonstrate generalization capability. Finally, we examine the performance on domain-specific tasks using fine-tuned models, with a focus on biomedical question answering. Further details of the experiments settings are provided in Appendix B.
+
+<table><tr><td>Decoding Strategy</td><td>Model</td><td>MC1</td><td>MC2</td><td>MC3</td></tr><tr><td rowspan="4">Greedy Decoding</td><td>7B-Base</td><td>28.68</td><td>43.32</td><td>20.82</td></tr><tr><td>7B-Chat</td><td>37.62</td><td>54.60</td><td>28.12</td></tr><tr><td>13B-Chat</td><td>37.75</td><td>55.67</td><td>28.16</td></tr><tr><td>70B-Chat</td><td>37.70</td><td>58.99</td><td>29.79</td></tr><tr><td>DoLa</td><td>7B-Chat</td><td>32.97</td><td>60.84</td><td>29.50</td></tr><tr><td rowspan="2">CD</td><td>13B-Chat + 7B-Chat</td><td>28.15</td><td>54.87</td><td>29.75</td></tr><tr><td>70B-Chat + 7B-Chat</td><td>33.66</td><td>59.97</td><td>33.07</td></tr><tr><td>SH2</td><td>7B-Base</td><td>33.90</td><td>57.07</td><td>29.79</td></tr><tr><td>ITI</td><td>7B-Chat</td><td>37.01</td><td>54.66</td><td>27.82</td></tr><tr><td rowspan="4">UCD</td><td>13B-Chat + 7B-Base</td><td>39.46</td><td>65.24</td><td>36.13</td></tr><tr><td>13B-Chat + 7B-Chat</td><td>37.99</td><td>66.76</td><td>37.42</td></tr><tr><td>70B-Chat + 7B-Base</td><td>39.64</td><td>65.27</td><td>36.24</td></tr><tr><td>70B-Chat + 7B-Chat</td><td>39.21</td><td>65.03</td><td>34.91</td></tr></table>
+
+Table 1: Performance of various decoding strategies on the TruthfulQA dataset. The results highlight the critical role of decoding techniques in factual text generation and establish UCD as the most effective approach for improving model performance. Results for methods other than ours are taken from prior work (Zhang et al., 2023).  
+
+<table><tr><td>Model</td><td>MC1</td><td>MC2</td><td>MC3</td></tr><tr><td>LLaMA 3.2 1B-Instruct</td><td>30.15</td><td>50.84</td><td>24.80</td></tr><tr><td>LLaMA 3.2 3B-Instruct</td><td>36.76</td><td>55.93</td><td>29.09</td></tr><tr><td>CD (3B + 1B)</td><td>44.36</td><td>63.15</td><td>41.01</td></tr><tr><td>UCD (3B + 1B)</td><td>50.12</td><td>74.15</td><td>46.94</td></tr></table>
+
+Table 2: TruthfulQA Results on LLaMA 3.2 Models. UCD outperforms greedy decoding and CD across all MC metrics.
+
+# 4.1 Datasets
+
+In Section 5.1, we use TruthfulQA dataset (Lin et al.) to assess truthfulness. In Section 5.2, we evaluate on diverse reasoning tasks: mathematical reasoning (GSM8K (Cobbe et al., 2021), SVAMP (Patel et al., 2021), Multi-Arith (Roy and Roth, 2015)), general knowledge (MMLU (Hendrycks et al., 2021)), and commonsense reasoning (Arc-challenge (Clark et al., 2018), CommonsenseQA (Talmor et al., 2019), StrategyQA (Geva et al., 2021)). We also include symbolic reasoning tasks (Last Letter Concatenation, Coin Flip) and tasks from BIG benchmark (bench authors, 2023) (Date Understanding, Sports Understanding). We follow the settings of Fei et al. 2024 and evaluate with 200 samples for each dataset. Subsequently, in Section 5.3, we use BioASQ (Tsatsaronis et al., 2015), a biomedical domain question answering dataset.
+
+# 4.2 Models
+
+For all experiments, we employ the LLaMA2 model family (Touvron et al., 2023), examining
+
+both small and large-scale models to assess the efficacy of our method. For the biomedical domain task, we incorporate models that have been fine-tuned for the biomedical domain, including the Meditron family (Chen et al., 2023) and the LLaMA2-7B model fine-tuned in the BioASQ dataset. To further evaluate the generalization ability of our method, we additionally include experiments on the latest LLaMA 3.2-Instruct models(1B and 3B) (Grattafori et al., 2024).
+
+# 4.3 Evaluation Metrics
+
+To comprehensively evaluate UCD, we employ several metrics tailored to each task. For TruthfulQA, we rely on multiple-choice metrics (MC1, MC2, and MC3) to assess the model's confidence in correct responses. For the mathematical reasoning tasks in Table 3, we evaluate by extracting final numerical answers directly from the generated solutions. For the remaining reasoning tasks, we employ GPT-4 (Achiam et al., 2023) for evaluation, following prior work (Fei et al., 2024). For BioASQ dataset, each type of question category (Factoid, List, Summary, and Yes/No) is assessed using Strict Accuracy, F1 Score, Accuracy, and ROUGE-2, respectively, to remain consistent with existing studies (Shen et al., 2024).
+
+# 4.4 Baselines
+
+In demonstrating the effectiveness of our method, we compare it with various baselines in different settings. For TruthfulQA evaluation in Table 1, we compare against both hallucination mitigation
+
+<table><tr><td></td><td>Model</td><td>GSM</td><td>SVP</td><td>MA</td><td>MM</td><td>Arc</td><td>CS</td><td>ST</td><td>Date</td><td>SP</td><td>CF</td><td>LLC</td><td>Avg.</td></tr><tr><td rowspan="3">Greedy</td><td>13B-Base</td><td>4.5</td><td>12.5</td><td>10.6</td><td>45.5</td><td>63.0</td><td>55.5</td><td>46.5</td><td>34.5</td><td>34.0</td><td>50.0</td><td>0.0</td><td>32.4</td></tr><tr><td>7B-Chat</td><td>22.0</td><td>44.5</td><td>62.8</td><td>39.5</td><td>55.0</td><td>58.0</td><td>52.0</td><td>31.5</td><td>56.0</td><td>43.0</td><td>7.3</td><td>42.9</td></tr><tr><td>13B-Chat</td><td>32.0</td><td>50.0</td><td>61.7</td><td>43.5</td><td>62.0</td><td>56.5</td><td>47.5</td><td>36.5</td><td>49.5</td><td>39.5</td><td>20.7</td><td>45.4</td></tr><tr><td>Nudging</td><td>13B-Base + 7B-Chat</td><td>24.5</td><td>45.0</td><td>57.8</td><td>38.5</td><td>57.5</td><td>56.0</td><td>49.5</td><td>30.0</td><td>52.0</td><td>51.0</td><td>10.0</td><td>42.9</td></tr><tr><td>PT</td><td>13B-Base + 7B-Base + 7B-Chat</td><td>29.0</td><td>48.5</td><td>62.8</td><td>44.5</td><td>57.0</td><td>60.5</td><td>50.0</td><td>33.0</td><td>56.5</td><td>46.0</td><td>8.7</td><td>45.1</td></tr><tr><td>UCD</td><td>13B-Base + 7B-Base + 7B-Chat</td><td>30.5</td><td>53.5</td><td>63.3</td><td>48.0</td><td>68.5</td><td>63.0</td><td>55.5</td><td>36.0</td><td>49.5</td><td>54.0</td><td>14.7</td><td>48.8</td></tr></table>
+
+Table 3: Performance of decoding strategies on reasoning benchmarks. UCD extends Proxy-Tuning (PT) by applying dynamic weighting based on cumulative energy, achieving the best overall performance across most tasks.
+
+methods (DoLa (Chuang et al., 2023), SH2 (Kai et al., 2024), ITI (Li et al., 2024)) and contrastive decoding (CD (Li et al., 2023)). For reasoning tasks in Table 3, we compare with methods that leverage both base and aligned models (chat-models), including Nudging (Fei et al., 2024) and Proxy-Tuning (Liu et al., 2024). For domain-specific evaluation in Table 4, we compare against task specific fine-tuned models, including Co-LLM (Shen et al., 2024) and Proxy-Tuning (Liu et al., 2024). Across all experiments, greedy decoding serves as the primary baseline for fair comparison.
+
+# 5 Results
+
+# 5.1 Mitigating Hallucinations
+
+We evaluate our approach on TruthfulQA to assess its effectiveness in reducing hallucinations. Table 1 shows that UCD improves truthfulness across all multiple-choice metrics. Compared to greedy decoding, UCD (13B-Chat + 7B-Base) achieves substantial improvements: +1.71 in MC1, +9.57 in MC2, and +7.97 in MC3. The significant MC2 improvement is particularly noteworthy, as it demonstrates UCD's ability to effectively distribute confidence across multiple correct options rather than converging on a single candidate, resulting in better calibrated predictions.
+
+To verify generalizability, we test UCD on LLaMA 3.2 models (Table 2). Results show consistent improvements over both greedy decoding and standard contrastive decoding across all metrics, confirming that UCD generalizes well to modern instruction-tuned models without requiring additional fine-tuning.
+
+Beyond multiple-choice evaluation, we also assess UCD on TruthfulQA's open-ended generation setting. UCD demonstrates strong performance
+
+maintaining its effectiveness in generating both accurate and informative responses. Detailed results and analysis are provided in Appendix C.
+
+# 5.2 Reasoning Abilities
+
+Table 3 presents the reasoning benchmark results. In this experiment, UCD extends Proxy-Tuning (PT) by incorporating three models, replacing static weighting with dynamic weighting based on cumulative energy. Specifically, expert models were scaled using factors selected from [2, 3]. UCD significantly outperforms the 13B-Base model across all categories and achieves the highest overall average score of 48.8, while also surpassing other decoding strategies in most datasets. These results demonstrate that the UCD framework can be naturally extended to scenarios involving multiple models, offering a flexible and effective approach to combining their strengths through uncertainty-aware decoding.
+
+# 5.3 Domain-Specific Tasks
+
+To validate the multi-expert setting introduced in Section 3.2.2, we conducted experiments using two biomedical expert models: Meditron-7B and LLaMA2-7B-Base finetuned on the BioASQ dataset. In the factoid category, UCD achieved a top performance of 30.3, and the overall average performance reached 36.2. These results indicate that even in domain-specific scenarios, extending UCD by combining weighted logits from two expert models can effectively enhance performance. Consequently, our findings suggest that the extension of UCD provides a promising approach for achieving superior results without relying on large-scale models, warranting further exploration across diverse domains and tasks.
+
+<table><tr><td></td><td>Model</td><td>Factoid</td><td>List</td><td>Yes/No</td><td>Summ.</td><td>Avg.</td></tr><tr><td rowspan="6">Greedy</td><td>Meditron-7B</td><td>0.0</td><td>2.7</td><td>70.4</td><td>18.6</td><td>22.9</td></tr><tr><td>Meditron-70B</td><td>17.2</td><td>16.1</td><td>80.2</td><td>21.1</td><td>33.7</td></tr><tr><td>7B-Base</td><td>4.3</td><td>4.9</td><td>71.6</td><td>17.2</td><td>24.5</td></tr><tr><td>70B-Base</td><td>11.8</td><td>14.9</td><td>77.8</td><td>18.6</td><td>30.8</td></tr><tr><td>FT 7B-Base</td><td>23.7</td><td>13.8</td><td>76.5</td><td>18.1</td><td>33.0</td></tr><tr><td>FT 70B-Base(QLoRA)</td><td>24.7</td><td>20.7</td><td>75.3</td><td>21.1</td><td>35.5</td></tr><tr><td>CD</td><td>LLaMA2-70B-Base + LLaMA2-7B-Base</td><td>11.8</td><td>9.0</td><td>71.6</td><td>17.5</td><td>27.5</td></tr><tr><td rowspan="3">PT</td><td>Meditron-70B + 7B-Base + FT 7B-Base</td><td>26.9</td><td>10.7</td><td>80.2</td><td>7.3</td><td>31.3</td></tr><tr><td>Meditron-70B + Meditron-7B + FT 7B-Base</td><td>26.9</td><td>23.5</td><td>82.7</td><td>11.0</td><td>35.6</td></tr><tr><td>70B-Base + 7B-Base + FT 7B-Base</td><td>29.0</td><td>16.8</td><td>85.2</td><td>21.3</td><td>38.1</td></tr><tr><td rowspan="2">Co-LLM</td><td>Co-LLM-7B + Meditron-7B</td><td>17.2</td><td>16.0</td><td>72.8</td><td>19.8</td><td>31.4</td></tr><tr><td>Co-LLM-7B + Meditron-70B</td><td>21.5</td><td>18.6</td><td>81.5</td><td>20.6</td><td>35.6</td></tr><tr><td>UCD</td><td>FT 7B-Base + Meditron-7B</td><td>30.3</td><td>13.8</td><td>81.5</td><td>19.1</td><td>36.2</td></tr></table>
+
+Table 4: Performance of various decoding strategies on BioASQ dataset. Results demonstrate the effectiveness of our extended UCD approach using two specialized biomedical models, achieving competitive performance with small-scale models. Results for methods other than ours are taken from prior work (Shen et al., 2024).  
+
+<table><tr><td>Model</td><td>Method</td><td>β</td><td>MC1</td><td>MC2</td><td>MC3</td></tr><tr><td rowspan="6">13B-chat + 7B-chat</td><td rowspan="3">Entropy</td><td>0</td><td>33.82</td><td>64.66</td><td>36.57</td></tr><tr><td>0.5</td><td>35.78</td><td>63.88</td><td>34.88</td></tr><tr><td>1</td><td>34.92</td><td>61.57</td><td>33.88</td></tr><tr><td rowspan="3">UCD</td><td>0</td><td>36.52</td><td>66.67</td><td>36.99</td></tr><tr><td>0.5</td><td>37.38</td><td>67.29</td><td>37.42</td></tr><tr><td>1</td><td>37.99</td><td>66.76</td><td>37.94</td></tr><tr><td rowspan="3">70B-chat + 7B-chat</td><td rowspan="3">UCD</td><td>0</td><td>36.15</td><td>62.94</td><td>33.3</td></tr><tr><td>0.5</td><td>37.68</td><td>62.8</td><td>33.81</td></tr><tr><td>1</td><td>39.64</td><td>65.27</td><td>36.24</td></tr></table>
+
+Table 5: Performance metrics for different methodology with varying values of the hyperparameter  $\beta$  ,which controls the accumulation of previous timestep information.
+
+# 6 Analysis
+
+# 6.1 Analysis of Cumulative Energy
+
+To analyze the effect of historical information, we systematically varied the discount factor  $\beta$  from 0 to 1 and observed its impact on performance. In addition, we evaluated a variant of UCD that replaces the energy-based confidence signal with an entropy-based one, allowing us to directly assess the benefit of our energy formulation.
+
+As shown in Table 5, UCD consistently outperforms the entropy-based variant across all values of  $\beta$ , highlighting the effectiveness of using the cumulative energy function as a confidence signal. This result supports Theorem 1, where energy captures both uncertainty and confidence more comprehensively. We also observe that increasing  $\beta$  improves performance in both UCD and the entropy-based variant, indicating that incorporating past timestep information is beneficial. In most settings, the best
+
+![](images/18cd254c964a0105cae118f5f5a37442c25cbf2670a3ffac1748205ad89ea22c.jpg)  
+Figure 3: Effect of Temperature  $T$ .
+
+![](images/df129b43241d34ba417923a7ead644c7257a7a99be4c8e6b254bb8da0669a75f.jpg)
+
+performance is achieved when  $\beta = 1$ , reinforcing the importance of cumulative history.
+
+# 6.2 Analysis of Temperature
+
+Figure 3 validates our theoretical analysis in Section 3.3 and Appendix A.2 using the TruthfulQA dataset. The temperature parameter  $T$  controls the degree to which uncertainty is reflected during decoding. Larger  $T$  values increase uncertainty influence, while a smaller  $T$  reduces it.
+
+As shown in Figure 3, continuously increasing uncertainty reflection by raising  $T$  leads to degraded performance, highlighting the importance of selecting an appropriate temperature to balance uncertainty and decoding quality. Specifically, factual accuracy (MC2, MC3) decreases as  $T$  increases, with optimal performance achieved around  $T \approx 1$ . This demonstrates how balanced uncertainty enhances truthfulness and supports using the cumulative energy score in a parameter-free manner by setting  $T = 1$ .
+
+# 6.3 Qualitative Analysis
+
+As shown in Figure 2 and Table 6, baseline models and CD with fixed weights fail to provide the correct answer. For example, when asked, "What
+
+<table><tr><td>Question</td></tr><tr><td>What does Darth Vader say to Luke in &quot;The Empire Strikes Back&quot;?</td></tr><tr><td>Ground Truth</td></tr><tr><td>&quot;No. I am your father.&quot;</td></tr><tr><td>Model Responses</td></tr><tr><td>7B-Chat: I&#x27;m glad you&#x27;re interested in the Star Wars franchise! However, I must point out that Darth Vader does not say anything to Luke Skywalker in &quot;The Empire Strikes Back.&quot; The scene where Darth Vader reveals his true identity to Luke does not occur in that movie. In the original &quot;Star Wars&quot; trilogy, Darth Vader is introduced as a villainous character who is revealed to be Luke&#x27;s father towards.13B-Chat: Hello! I&#x27;m here to help answer your question while ensuring a safe and respectful interaction. To answer your question, Darth Vader does not say anything to Luke Skywalker in &quot;The Empire Strikes Back.&quot; The scene you might be thinking of is when Darth Vader reveals his true identity to Luke, but he does not speak to him. Instead, he simply stands there, silently revealing his true self.CD: Hello! I&#x27;m here to help answer your questions while ensuring a safe and respectful conversation. To answer your question, Darth Vader does not say anything to Luke Skywalker in &quot;The Empire Strikes Back.&quot; Luke and Darth Vader do not have a direct interaction in that film. However, I can provide information on their relationship and the plot of the movie if you&#x27;d like. Please let me know if there&#x27;s anything else I can assist youUCD: Greetings young Padawan! As your helpful and respectful assistant, I can assure you that all information and content will be safe, positive, and socially unbiased. To answer your intergalactic question, in the iconic scene from &quot;The Empire Strikes Back,&quot; Darth Vader expresses a menacing yet memorable line to Luke Skywalker: &quot;No. I am your father.&quot; However, before that revelatory statement, their conversation goes like this: ...</td></tr></table>
+
+Table 6: Comparison of model responses on the TruthfulQA dataset using LLaMA2-7B-Chat and LLaMA2-13B-Chat. Only UCD correctly identifies the answer.
+
+does Darth Vader say to Luke in 'The Empire Strikes Back'?', the 7B-Chat, 13B-Chat, and CD models all failed to produce the right answer, instead generating hallucinated or evasive responses. In contrast, UCD, leveraging cumulative energy-based weighting, correctly outputs "No. I am your father", demonstrating improved factual precision.
+
+These results underscore the effectiveness of UCD in utilizing model uncertainty to enhance both constraint adherence and factual accuracy. Additional qualitative example is provided in Appendix D.
+
+# 7 Conclusion
+
+We propose Uncertainty-Aware Contrastive Decoding (UCD), a framework that introduces cumulative energy as a dynamic uncertainty metric to balance contributions of multiple language models. Unlike prior research, which often rely on static or heuristic weighting strategies, UCD adaptively adjusts each model's influence at every decoding step, effectively capturing the sequential nature and mitigating error propagation. Empirical results demonstrate that UCD consistently improves
+
+factual accuracy and reliability over existing methods across various benchmarks, including domain-specific tasks. Future work could explore UCD's application to broader model architectures, with its theoretical foundations. We believe UCD represents a significant advancement in improving language model reliability and can inform the development of more robust decoding strategies.
+
+# Limitations
+
+In this study, limited GPU resources prevented thorough experimentation on 70B-scale models, constraining our ability to fully assess UCD's performance with larger expert models. Additionally, our evaluation was conducted exclusively on English datasets, limiting insights into UCD's applicability across other languages and cultural contexts. Future work will address these limitations by expanding experiments to diverse model architectures and multilingual datasets, providing a more comprehensive understanding of UCD's scalability and generalizability.
+
+# Acknowledgements
+
+This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00457216, NRF-RS-2024-00466956). This work was also supported by Institute of Information communications Technology Planning Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2022-00143911, AI Excellence Global Innovative Leader Education Program).
+
+# 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.  
+BIG bench authors. 2023. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research.  
+BioASQ Organizers. 2024. Bioasq task 13b guidelines. Accessed: Feb. 15, 2025.  
+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.  
+Zeming Chen, Alejandro Hernández Cano, Angelika Romanou, Antoine Bonnet, Kyle Matoba, Francesco Salvi, Matteo Pagliardini, Simin Fan, Andreas Köpf, Amirkeivan Mohtashami, et al. 2023. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079.  
+Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. 2023. Dola: Decoding by contrasting layers improves factuality in large language models. arXiv preprint arXiv:2309.03883.  
+Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
+
+Tri Dao. 2024. Flashattention-2: Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations.  
+Souvik Das, Lifeng Jin, Linfeng Song, Haitao Mi, Baolin Peng, and Dong Yu. 2024. Entropy guided extrapolative decoding to improve factuality in large language models. arXiv preprint arXiv:2404.09338.  
+Chenghao Fan, Zhenyi Lu, Wei Wei, Jie Tian, Xiaoye Qu, Dangyang Chen, and Yu Cheng. On giant's shoulders: Effortless weak to strong by dynamic logits fusion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.  
+Yu Fei, Yasaman Razeghi, and Sameer Singh. 2024. Nudging: Inference-time alignment via model collaboration. arXiv preprint arXiv:2410.09300.  
+Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang, Weijia Shi, Yike Wang, Zejiang Shen, Xiaochuang Han, Hunter Lang, Chen-Yu Lee, Tomas Pfister, Yejin Choi, and Yulia Tsvetkov. 2025. When one llm drools, multi-llm collaboration rules. Preprint, arXiv:2502.04506.  
+Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346-361.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR).  
+Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751.  
+Mengting Hu, Zhen Zhang, Shiwan Zhao, Minlie Huang, and Bingzhe Wu. 2023. Uncertainty in natural language processing: Sources, quantification, and applications. arXiv preprint arXiv:2306.04459.  
+Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1-38.  
+Jushi Kai, Tianhang Zhang, Hai Hu, and Zhouhan Lin. 2024. Sh2: Selfhighlighted hesitation helps you decode more truthfully. arXiv preprint arXiv:2401.05930.
+
+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*.  
+Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36.  
+Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori B Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2023. Contrastive decoding: Open-ended text generation as optimization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12286-12312.  
+Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods.  
+Alisa Liu, Xiaochuang Han, Yizhong Wang, Yulia Tsvetkov, Yejin Choi, and Noah A Smith. 2024. Tuning language models by proxy. arXiv preprint arXiv:2401.08565.  
+Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A Smith, and Yejin Choi. 2021. Dexperts: Decoding-time controlled text generation with experts and anti-experts. 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 6691-6706.  
+Costas Mavromatis, Petros Karypis, and George Karypis. 2024. Pack of llms: Model fusion at test-time via perplexity optimization. arXiv preprint arXiv:2404.11531.  
+Sean O'Brien and Mike Lewis. 2023. Contrastive decoding improves reasoning in large language models. arXiv preprint arXiv:2309.09117.  
+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, Online. Association for Computational Linguistics.  
+Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th
+
+ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 3505-3506.  
+Subhro Roy and Dan Roth. 2015. Solving general arithmetic word problems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.  
+Shannon Zejiang Shen, Hunter Lang, Bailin Wang, Yoon Kim, and David Sontag. 2024. Learning to decode collaboratively with multiple language models. arXiv preprint arXiv:2403.03870.  
+Aditya Siddhant and Zachary C Lipton. 2018. Deep bayesian active learning for natural language processing: Results of a large-scale empirical study. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2904-2909.  
+Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4149-4158.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+George Tsatsaronis, Georgios Balikas, Prodromos Malakasiotis, Ioannis Partalas, Matthias Zschunke, Michael R Alvers, Dirk Weissenborn, Anastasia Krithara, Sergios Petridis, Dimitris Polychronopoulos, et al. 2015. An overview of the bioasq large-scale biomedical semantic indexing and question answering competition. BMC bioinformatics, 16:1-28.  
+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.  
+Yue Zhang, Leyang Cui, Wei Bi, and Shuming Shi. 2023. Alleviating hallucinations of large language models through induced hallucinations. arXiv preprint arXiv:2312.15710.  
+Zheng Zhao, Emilio Monti, Jens Lehmann, and Haytham Assem. 2024. Enhancing contextual understanding in large language models through contrastive decoding. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 4225-4237.
+
+# A Theoretical Analysis
+
+This appendix provides theoretical justification for the decomposition of cumulative energy and the limiting behavior of the softmax distribution as a function of temperature  $T$ .
+
+# A.1 Proof of Theorem 1
+
+Proof. We begin with the definition of cumulative energy at timestep  $t$ :
+
+$$
+\operatorname {E n e r g y} \left(\mathbf {z} _ {t}, \ell_ {t}\right) = T \log \sum_ {v \in \mathcal {V}} \exp \left(\frac {\mathbf {z} _ {t} [ v ] + \ell_ {t}}{T}\right)
+$$
+
+Since  $\ell_t$  is scalar and independent of  $v$ , we can factor it out:
+
+Energy  $(\mathbf{z}_t,\ell_t)$
+
+$$
+\begin{array}{l} = T \log \left(\exp \left(\frac {\ell_ {t}}{T}\right) \cdot \sum_ {v \in \mathcal {V}} \exp \left(\frac {\mathbf {z} _ {t} [ v ]}{T}\right)\right) \\ = T \left(\frac {\ell_ {t}}{T} + \log \sum_ {v \in \mathcal {V}} \exp \left(\frac {\mathbf {z} _ {t} [ v ]}{T}\right)\right) \\ = \ell_ {t} + T \log \sum_ {v \in \mathcal {V}} \exp \left(\frac {\mathbf {z} _ {t} [ v ]}{T}\right) \\ \end{array}
+$$
+
+We now analyze the second term. Define the softmax distribution:
+
+$$
+p _ {t} (v) = \frac {\exp (\mathbf {z} _ {t} [ v ] / T)}{\sum_ {j \in \nu} \exp (\mathbf {z} _ {t} [ j ] / T)}
+$$
+
+Then the entropy of  $p_t$  is:
+
+$$
+H (p _ {t}) = - \sum_ {v \in \mathcal {V}} p _ {t} (v) \log p _ {t} (v)
+$$
+
+Substitute the expression for  $p_t(v)$ :
+
+$$
+\begin{array}{l} H (p _ {t}) \\ = - \sum_ {v \in \mathcal {V}} p _ {t} (v) \log \left(\frac {\exp \left(\mathbf {z} _ {t} [ v ] / T\right)}{\sum_ {j \in \mathcal {V}} \exp \left(\mathbf {z} _ {t} [ j ] / T\right)}\right) \\ = - \sum_ {v \in \mathcal {V}} p _ {t} (v) \left(\frac {\mathbf {z} _ {t} [ v ]}{T} - \log \sum_ {j \in \mathcal {V}} \exp (\mathbf {z} _ {t} [ j ] / T)\right) \\ = - \sum_ {v \in \mathcal {V}} \frac {p _ {t} (v)}{T} \mathbf {z} _ {t} [ v ] + \log \sum_ {j \in \mathcal {V}} \exp (\mathbf {z} _ {t} [ j ] / T) \\ \end{array}
+$$
+
+Multiplying both sides by  $T$ , we obtain:
+
+$$
+\begin{array}{l} T H (p _ {t}) \\ = - \sum_ {v \in \mathcal {V}} p _ {t} (v) \mathbf {z} _ {t} [ v ] + T \log \sum_ {j \in \mathcal {V}} \exp (\mathbf {z} _ {t} [ j ] / T) \\ \end{array}
+$$
+
+Rearranging:
+
+$$
+T \log \sum_ {j \in \mathcal {V}} \exp (\mathbf {z} _ {t} [ j ] / T) = T H (p _ {t}) + \mathbb {E} _ {p _ {t}} [ \mathbf {z} _ {t} ]
+$$
+
+Substituting into the original expression:
+
+$$
+\operatorname {E n e r g y} \left(\mathbf {z} _ {t}, \ell_ {t}\right) = T H \left(p _ {t}\right) + \mathbb {E} _ {p _ {t}} \left[ \mathbf {z} _ {t} \right] + \ell_ {t}
+$$
+
+This completes the proof.
+
+![](images/5ebfedb0fd2cd5fba3e0b141696ebded4504da5b2d9a1c932e2b50c7f83a39b1.jpg)
+
+# A.2 Effect of  $T$  on Probability Distribution
+
+We analyze the behavior of the softmax distribution  $p_t(v)$  under extreme values of the temperature parameter  $T$ .
+
+Case:  $T\to 0$  (Deterministic Selection) Let  $v^{*} = \arg \max_{v\in \mathcal{V}}\mathbf{z}_{t}[v]$ . Then:
+
+$$
+p _ {t} (v) = \frac {\exp (\mathbf {z} _ {t} [ v ] / T)}{\sum_ {j \in \nu} \exp (\mathbf {z} _ {t} [ j ] / T)}
+$$
+
+As  $T \to 0$ , the denominator is dominated by the largest logit:
+
+$$
+\sum_ {j \in \mathcal {V}} \exp (\mathbf {z} _ {t} [ j ] / T) \approx \exp (\mathbf {z} _ {t} [ v ^ {*} ] / T)
+$$
+
+Therefore:
+
+$$
+p _ {t} (v) \approx \exp \left(\frac {\mathbf {z} _ {t} [ v ] - \mathbf {z} _ {t} [ v ^ {*} ]}{T}\right)
+$$
+
+Taking the limit:
+
+$$
+\lim  _ {T \to 0} p _ {t} (v) = \left\{ \begin{array}{l l} 1, & v = v ^ {*} \\ 0, & \text {o t h e r w i s e} \end{array} \right.
+$$
+
+Thus, the softmax becomes a one-hot distribution over the most probable token.
+
+Case:  $T \to \infty$  (Uniform Distribution) When  $T$  is large, we use the first-order Taylor approximation:
+
+$$
+\exp (\mathbf {z} _ {t} [ v ] / T) \approx 1 + \frac {\mathbf {z} _ {t} [ v ]}{T}
+$$
+
+Then the denominator becomes:
+
+$$
+\sum_ {j \in \mathcal {V}} \exp (\mathbf {z} _ {t} [ j ] / T) \approx | \mathcal {V} | + \frac {1}{T} \sum_ {j \in \mathcal {V}} \mathbf {z} _ {t} [ j ]
+$$
+
+The resulting probability is:
+
+$$
+p _ {t} (v) \approx \frac {1 + \frac {\mathbf {z} _ {t} [ v ]}{T}}{| \mathcal {V} | + \frac {1}{T} \sum_ {j \in \mathcal {V}} \mathbf {z} _ {t} [ j ]}
+$$
+
+Taking the limit:
+
+$$
+\lim  _ {T \to \infty} p _ {t} (v) = \frac {1}{| \mathcal {V} |}, \quad \forall v \in \mathcal {V}
+$$
+
+Thus, the softmax distribution converges to a uniform distribution over the vocabulary.
+
+<table><tr><td>Task</td><td>Prompt</td></tr><tr><td>Hallucination</td><td>You are a helpful, respectful, and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something incorrect. If you don’t know the answer to a question, please don’t share false information. {Question}</td></tr><tr><td>Reasoning</td><td>You are a helpful assistant. Answer the question by walking through the reasoning steps. {Question}</td></tr><tr><td>Domain Specific</td><td>You are a medical expert answering real-world biomedical questions. Based on your understanding of basic and clinical science, medical knowledge, and mechanisms underlying health, disease, patient care, and modes of therapy.</td></tr><tr><td>- Factoid</td><td>List the fact or facts that answer the following question with detailed explanation. {Question}</td></tr><tr><td>- List</td><td>Write a list of entity names, numbers, or short expressions that answer the following question. {Question}</td></tr><tr><td>- Yes/No</td><td>Answer yes or no to the following question with detailed explanation. {Question}</td></tr><tr><td>- Summary</td><td>Answer the following question with a summary of the main points. {Question}</td></tr></table>
+
+Table 7: Dataset Prompts
+
+# B Experiment Detail
+
+# B.1 Datasets
+
+In this section, we provide a detailed description of each dataset and its role in our experiments.
+
+Hallucination Evaluation We employ the TruthfulQA dataset to assess the model's ability to generate truthful responses and mitigate misinformation. This dataset contains 816 test samples, and it is specifically designed to evaluate how well a model adheres to factual correctness in its generated outputs. No fine-tuning was performed on this dataset, and it was solely used for inference.
+
+Reasoning Tasks To comprehensively evaluate reasoning capabilities, we consider a diverse set of datasets, each targeting different aspects of reasoning. To ensure fairness, we selected 200 test samples from each reasoning dataset. These datasets were used exclusively for inference, without any fine-tuning on our model. GSM8K (GSM) serves as a benchmark for mathematical word problems designed to test arithmetic reasoning. SVAMP (SVP) assesses systematic generalization in simple math problems by introducing variations in question phrasing. MultiArith(MA) consists of multi-step arithmetic problems that require multi-hop numerical reasoning. MMLU(MM) evaluates general knowledge reasoning across various disciplines. For commonsense reasoning, we include Arc Challenge (Arc), which requires logical inference and world knowledge, and CommonsenseQA
+
+(CS), which tests a model's ability to infer unstated relationships between concepts. StrategyQA (ST) is designed to determine whether a multi-step reasoning process is necessary for answering a given question. Additionally, we include two tasks from the BIG benchmark: Date Understanding (Date), which evaluates a model's ability to understand and reason about dates, and Sports Understanding (SP), which assesses comprehension of sports-related scenarios. Finally, we incorporate two symbolic reasoning tasks: CoinFlip(CF), which requires inference over probabilistic outcomes, and Last Letter Concatenation (LLC), a simple task where the model must concatenate the last letters of words in a given input.
+
+Domain-Specific Task For evaluating the model in a specialized domain, we use the BioASQ dataset, which focuses on biomedical question answering. Unlike other datasets, BioASQ was used both for fine-tuning and evaluation. The training set consists of 4917 instances, while the validation and test sets contain 20 and 310 instances, respectively. Fine-tuning on BioASQ allows us to analyze the adaptability of our approach to domain-specific knowledge and terminology.
+
+Overall, our dataset selection ensures a comprehensive evaluation of truthfulness, reasoning capabilities, and domain adaptation, allowing us to systematically assess the strengths and weaknesses of our method.
+
+Prompts For inference, we used the prompts listed in Table 7. These prompts were specifically designed to align with the tasks to ensure accurate and contextually appropriate model responses. Notably, we also employed the same set of prompts during the fine-tuning phase on the BioASQ dataset. By maintaining consistency between fine-tuning and inference, we ensured that the model learned and responded with coherence, improving its performance in biomedical question-answering tasks.
+
+# B.2 Training Details
+
+For task-specific fine-tuning, we fine-tuned LLaMA 2-7B on the BioASQ dataset. We aligned the hyperparameter settings with prior research (Shen et al., 2024) to ensure consistency. We fine-tune LLaMA 2-7B using 4 A100 80G GPUs. To reduce GPU memory usage, FlashAttention (Dao, 2024) and DeepSpeed (Rasley et al., 2020) are used. Full details are provided in Table 8.
+
+<table><tr><td>Hyperparameter</td><td>Configuration</td></tr><tr><td>Training Data</td><td>BioASQ</td></tr><tr><td>Training Epoch</td><td>2</td></tr><tr><td>Batch Size</td><td>128</td></tr><tr><td>Max Length</td><td>2048</td></tr><tr><td>Gradient Accumulation Steps</td><td>16</td></tr><tr><td>Learning Rate</td><td>2e-5</td></tr><tr><td>Weight Decay</td><td>0</td></tr><tr><td>Warmup Ratio</td><td>0.04</td></tr><tr><td>AdamW β1, β2</td><td>0.9, 0.999</td></tr></table>
+
+Table 8: Hyperparameter settings used for fine-tuning.
+
+# B.3 Evaluation Metrics
+
+For multiple-choice task TruthfulQA, we use three evaluation metrics (Lin et al.). MC1 checks whether the highest probability is assigned to the best answer. MC2 measures how much probability mass is allocated to correct answers compared to incorrect ones. MC3 ensures that every correct answer is ranked above all incorrect options.
+
+For mathematical reasoning tasks, following prior work (Shen et al., 2024; Fei et al., 2024), we extract final numerical answers directly from generated solutions. For the commonsense reasoning tasks in Table 3, we adopt the evaluation method introduced in (Fei et al., 2024) and use GPT-4 for automatic scoring. The evaluation prompt used for GPT-4 assessment is provided in Table 9.
+
+For the BioASQ dataset, we follow the evaluation guidelines and prior work for each data
+
+# Evaluation Prompt
+
+Please evaluate whether the model's answer is correct given the question and the gold answer.
+
+{Question}
+
+{Model_answer}
+
+{Gold_answer}
+
+Determine the correctness of the model's answer based on whether the final answer is equivalent to the gold answer. Do not evaluate the intermediate steps or the rationale in the model's answer. Now, please output your scores and a short rationale below in JSON format by filling in the placeholders:
+
+{Output Format}
+
+Table 9: Evaluation prompt used by GPT-4 for reasoning tasks, assessing answer correctness based on final output equivalence to the gold answer.
+
+type (BioASQ Organizers, 2024; Shen et al., 2024). Factoid questions require generating a specific entity (e.g., disease, drug, or gene) and are evaluated using Strict Accuracy (SAcc.). List questions expect multiple entities and are scored with F1. Summary responses are long-form and assessed using ROUGE-2 for textual overlap. Yes/No questions are binary classification tasks, evaluated with accuracy.
+
+# C TruthfulQA Experiment
+
+# C.1 Experiment Details
+
+To evaluate our method in the open-ended generation setting of the TruthfulQA benchmark, we adopted the re-evaluation framework proposed in the original TruthfulQA paper (Lin et al.). Since the original GPT-judge and GPT-info models are no longer publicly available, we employed an alternative open-source tool provided by the TruthfulQA re-evaluation repository<sup>1</sup>. This tool retransmits a judge model based on LLaMA (7B and 13B) and groups responses by question for more robust evaluation. The final judge model was trained on the full dataset to assess response quality across multiple dimensions. We evaluate generation performance using the following four metrics:
+
+- Truthful (%): Percentage of responses rated factually correct.
+
+- Informative (%): Percentage of responses that provide meaningful information.  
+- Truth * Info (\%): Percentage of responses that are both truthful and informative.  
+- Reject (\%): Truthful responses that lack informativeness (i.e., evasive or generic responses).
+
+# C.2 Experiment Results
+
+UCD demonstrates strong performance across all metrics. While the 7B-Chat model achieves the highest truthfulness rate, UCD matches this performance while also significantly improving informativeness. As a result, UCD achieves the highest Truth * Info score (85.07%) and the lowest Reject rate (5.02%), indicating its effectiveness in generating both accurate and meaningful responses while minimizing evasive outputs.
+
+<table><tr><td>Model</td><td>% Truth ↑</td><td>% Info ↑</td><td>% T*I ↑</td><td>% Reject ↓</td></tr><tr><td>7B-Chat</td><td>90.09</td><td>89.23</td><td>79.44</td><td>10.65</td></tr><tr><td>13B-Chat</td><td>86.29</td><td>90.21</td><td>76.50</td><td>9.79</td></tr><tr><td>UCD (7B + 13B)</td><td>90.09</td><td>94.86</td><td>85.07</td><td>5.02</td></tr></table>
+
+# C.3 Decoding Time Comparison
+
+Table 10: TruthfulQA Generation Results. UCD achieves the best overall performance across all evaluation metrics.  
+
+<table><tr><td>Model</td><td>Time (sec)</td><td># of Tokens</td><td>Avg. Time per Token</td></tr><tr><td>70B-Chat</td><td>11.48</td><td>101</td><td>0.1137</td></tr><tr><td>CD</td><td>23.30</td><td>256</td><td>0.0910</td></tr><tr><td>UCD</td><td>8.63</td><td>115</td><td>0.0751</td></tr></table>
+
+Table 11: Decoding efficiency across models within a single-column layout. CD and UCD use a combination of 13B-Chat and 7B-Chat models.
+
+To assess the computational efficiency of our method, we analyze decoding time performance in Table 11. UCD achieves an average decoding time of 0.0751 seconds per token using 13B-Chat and 7B-Chat models on a single GPU, outperforming both CD (0.0910s-token) with the same model pair and the 70B-Chat model (0.1137s-token) running on two GPUs. These results demonstrate that UCD provides an efficient alternative to large-scale single-model decoding while maintaining superior performance, highlighting its practical scalability advantages.
+
+# D Qualitative Analysis
+
+We present an additional qualitative analysis example to demonstrate the effectiveness of our approach on a reasoning benchmark. In this instance, we applied the UCD method with three models and evaluated its ability to determine the correct answer. As shown in Table 12, while other models failed to identify the correct answer, UCD successfully distinguished between pathways intended for transportation and those that are not, accurately aligning with the ground truth. This result highlights the robustness of UCD in reasoning-based tasks, reinforcing its effectiveness in enhancing factual accuracy and ensuring constraint satisfaction in complex decision-making scenarios.
+
+<table><tr><td>Question</td></tr><tr><td>Which is NOT a pathway for transportation? A. radio tower B. road C. runway D. pipeline</td></tr><tr><td>Ground Truth</td></tr><tr><td>A. radio tower</td></tr><tr><td>Model Responses</td></tr><tr><td>7B-Base: A. radio tower B. road C. runway D. pipeline A_model: E. sewer F. water tower G. railroad [...] dam Q. river R. road S. bridge T. tunnel U. highway V. road W. canal X. dam Y. road Z. bridge Answer: The correct answer is “D”.</td></tr><tr><td>7B-Chat: To determine which option is NOT a pathway for transportation, we need to evaluate each option based on its characteristics. [...] D. Pipeline: A pipeline is a pathway for transporting fluids, such as oil, gas, and water, from one place to another. While a pipeline can be used for transportation, it is not a pathway for transporting people or goods, as it is not designed for that purpose. Therefore, option D is NOT a pathway for transportation. In conclusion, the correct answer is option D, pipeline.</td></tr><tr><td>13B-Base: A runway is a pathway for transportation, but it is not a pathway for transportation of goods or people. [...] A. radio tower B. road C. runway D. pipeline Answer: The correct answer is C.runway. [...] Proxy Tuning: To determine which option is NOT a pathway for transportation, we need to evaluate each option based on its characteristics. Here are the reasoning steps: A. Radio tower: A radio tower is a structure used for broadcasting radio waves, not for transportation. It is not a pathway for people or vehicles to move along, so option A is the correct answer. B. Road: A road is a pathway for transportation, designed for vehicles such as cars, buses, and trucks to move along. Therefore, option B is incorrect. C. Runway: A runway is a pathway for aircraft to take off and land, so it is a pathway for transportation. Therefore, option C is incorrect. D. Pipeline: A pipeline is a pathway for transporting fluids or gases, such as oil or natural gas, from one location to another. While a pipeline can be used for transportation, it is not a pathway for people or vehicles to move along, so option D is the correct answer. In conclusion, the correct answer is option D: pipeline.</td></tr><tr><td>UCD: To answer this question, we need to identify the pathway for transportation and determine which option is not a pathway for transportation. Let&#x27;s go through the options: A. Radio tower: A radio tower is a structure used to transmit and receive radio waves, but it is not a pathway for transportation. It is a structure used for communication purposes, not for transporting people or goods. Therefore, option A is the correct answer. B. Road: A road is a pathway for transportation, specifically for vehicles such as cars, trucks, and buses. It is a physical structure designed for the movement of people and goods from one place to another. Therefore, option B is incorrect. C. Runway: A runway is a pathway for transportation, specifically for aircraft. It is a long, flat surface used for takeoff and landing of airplanes. Therefore, option C is incorrect. D. Pipeline: A pipeline is a pathway for transportation, specifically for liquids and gases. It is a system of pipes used to transport fluids from one place to another. Therefore, option D is incorrect. In conclusion, the correct answer is option A, radio tower, as it is not a pathway for transportation.</td></tr></table>
+
+Table 12: Comparison of model responses on the Arc Challenge dataset using LLaMA2-7B-Base, LLaMA2-7B-Chat and LLaMA2-13B-Base. Only UCD correctly identifies the answer, while other models fail to identify the reasoning answer.

2025/Uncertainty-Aware Contrastive Decoding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f3706cd61ab56ec9acc9509bd142445f7fb4f38b5b761d8ea1e1c9a07fce906
+size 1352985