Add Batch ede7eef3-0fb4-497b-b933-4cf8c8f482bd data

.gitattributes

@@ -620,3 +620,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/Human-Inspired[[:space:]]Obfuscation[[:space:]]for[[:space:]]Model[[:space:]]Unlearning_[[:space:]]Local[[:space:]]and[[:space:]]Global[[:space:]]Strategies[[:space:]]with[[:space:]]Hyperbolic[[:space:]]Representations/827580b9-daee-4f17-9f73-b43754214845_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Humanity’s[[:space:]]Last[[:space:]]Code[[:space:]]Exam_[[:space:]]Can[[:space:]]Advanced[[:space:]]LLMs[[:space:]]Conquer[[:space:]]Human’s[[:space:]]Hardest[[:space:]]Code[[:space:]]Competition_/8a4bafab-99a5-4429-b80f-2894c7e31f4c_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/Humor[[:space:]]in[[:space:]]Pixels_[[:space:]]Benchmarking[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models[[:space:]]Understanding[[:space:]]of[[:space:]]Online[[:space:]]Comics/30c39bf0-e05b-47cd-8b04-4866f611adbd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/I-GUARD_[[:space:]]Interpretability-Guided[[:space:]]Parameter[[:space:]]Optimization[[:space:]]for[[:space:]]Adversarial[[:space:]]Defense/af7da807-decc-491b-967a-bbededc545c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ICL-Bandit_[[:space:]]Relevance[[:space:]]Labeling[[:space:]]in[[:space:]]Advertisement[[:space:]]Recommendation[[:space:]]Systems[[:space:]]via[[:space:]]LLM/deeabab9-126e-40e2-a116-c0a839b8d3f0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ICLER_[[:space:]]Intent[[:space:]]CLassification[[:space:]]with[[:space:]]Enhanced[[:space:]]Reasoning/80ea3d72-543b-4c68-9989-ef5d6936a0b2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/INDOORWORLD[[:space:]]_[[:space:]]Integrating[[:space:]]Physical[[:space:]]Task[[:space:]]Solving[[:space:]]and[[:space:]]Social[[:space:]]Simulation[[:space:]]in[[:space:]]A[[:space:]]Heterogeneous[[:space:]]Multi-Agent[[:space:]]Environment/ee507a4c-7454-4514-a8c9-73b0446536c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/INREACT_[[:space:]]An[[:space:]]Inspire-Then-Reinforce[[:space:]]Training[[:space:]]Framework[[:space:]]For[[:space:]]Multimodal[[:space:]]GUI[[:space:]]Agent/889785bd-0923-422b-bcca-82b35b178935_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/IP-Dialog_[[:space:]]Evaluating[[:space:]]Implicit[[:space:]]Personalization[[:space:]]in[[:space:]]Dialogue[[:space:]]Systems[[:space:]]with[[:space:]]Synthetic[[:space:]]Data/347614f0-d1c5-4371-901f-a53f1ec61429_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ISACL_[[:space:]]Internal[[:space:]]State[[:space:]]Analyzer[[:space:]]for[[:space:]]Copyrighted[[:space:]]Training[[:space:]]Data[[:space:]]Leakage/8060ed99-c298-4105-b64b-1e8c92bcd4fe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Identifying[[:space:]]Aspects[[:space:]]in[[:space:]]Peer[[:space:]]Reviews/6e98954a-cad9-4f3b-8b14-95738b229b2e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Identifying[[:space:]]Noise[[:space:]]in[[:space:]]Human-Created[[:space:]]Datasets[[:space:]]using[[:space:]]Training[[:space:]]Dynamics[[:space:]]from[[:space:]]Generative[[:space:]]Models/c8a60779-b000-4e4a-849a-d9e2608e2b0d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Identifying[[:space:]]Rare[[:space:]]Languages[[:space:]]in[[:space:]]Common[[:space:]]Crawl[[:space:]]Data[[:space:]]is[[:space:]]a[[:space:]]Needles-in-a-Haystack[[:space:]]Problem/d7168612-36e6-4401-bef3-d21a2048507a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Idola[[:space:]]Tribus[[:space:]]of[[:space:]]AI_[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]tend[[:space:]]to[[:space:]]perceive[[:space:]]order[[:space:]]where[[:space:]]none[[:space:]]exists/1bf6f6ce-e282-45f4-a060-9ab69570c3a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Imagination[[:space:]]and[[:space:]]Contemplation_[[:space:]]A[[:space:]]Balanced[[:space:]]Framework[[:space:]]for[[:space:]]Semantic-Augmented[[:space:]]Multimodal[[:space:]]Machine[[:space:]]Translation/64ce1505-6443-42a1-89d0-82a64e9cb443_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/ImpRAG_[[:space:]]Retrieval-Augmented[[:space:]]Generation[[:space:]]with[[:space:]]Implicit[[:space:]]Queries/4f4eceb9-3a9c-48c1-809f-872de91d71a6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]Alignment[[:space:]]in[[:space:]]LVLMs[[:space:]]with[[:space:]]Debiased[[:space:]]Self-Judgment/32697c53-7fef-414e-b3b7-e030b23a50e8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]Influence-based[[:space:]]Instruction[[:space:]]Tuning[[:space:]]Data[[:space:]]Selection[[:space:]]for[[:space:]]Balanced[[:space:]]Learning[[:space:]]of[[:space:]]Diverse[[:space:]]Capabilities/5c0b6b22-8af2-4269-a6c9-877fe9379fdd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]LLM[[:space:]]Reasoning[[:space:]]through[[:space:]]Interpretable[[:space:]]Role-Playing[[:space:]]Steering/1ccb4d81-5d13-43eb-8399-8a7e201b5e03_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]LLM-as-a-Judge[[:space:]]Inference[[:space:]]with[[:space:]]the[[:space:]]Judgment[[:space:]]Distribution/8deb81ec-d61f-4c79-9e1a-8ce6753b581f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]Language[[:space:]]Model[[:space:]]Personas[[:space:]]via[[:space:]]Rationalization[[:space:]]with[[:space:]]Psychological[[:space:]]Scaffolds/e8541c71-4741-49cc-a5ef-0526328aa152_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]Preference[[:space:]]Alignment[[:space:]]of[[:space:]]LLM[[:space:]]with[[:space:]]Inference-Free[[:space:]]Self-Refinement/7abcd001-bfb3-4dc2-8eda-1407dc4757e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]Prompt[[:space:]]Generalization[[:space:]]for[[:space:]]Cross-prompt[[:space:]]Essay[[:space:]]Trait[[:space:]]Scoring[[:space:]]from[[:space:]]the[[:space:]]Scoring-invariance[[:space:]]Perspective/840e2b05-7ba3-4432-93af-a31ccfd365e0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Improving[[:space:]]the[[:space:]]Language[[:space:]]Understanding[[:space:]]Capabilities[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Using[[:space:]]Reinforcement[[:space:]]Learning/969a5451-8e86-40c5-aa13-35d019493d26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/InFact_[[:space:]]Informativeness[[:space:]]Alignment[[:space:]]for[[:space:]]Improved[[:space:]]LLM[[:space:]]Factuality/4a7784af-9c05-4e5f-92fd-cb3cac7e4966_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Inclusive[[:space:]]Leadership[[:space:]]in[[:space:]]the[[:space:]]Age[[:space:]]of[[:space:]]AI_[[:space:]]A[[:space:]]Dataset[[:space:]]and[[:space:]]Comparative[[:space:]]Study[[:space:]]of[[:space:]]LLMs[[:space:]]vs.[[:space:]]Real-Life[[:space:]]Leaders[[:space:]]in[[:space:]]Workplace[[:space:]]Action[[:space:]]Planning/cda74f64-5929-4872-afa2-28a662e2e1f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Inducing[[:space:]]Argument[[:space:]]Facets[[:space:]]for[[:space:]]Faithful[[:space:]]Opinion[[:space:]]Summarization/89727a66-03a8-4d55-aa1a-a10dcc5f7071_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Inductive[[:space:]]Reasoning[[:space:]]on[[:space:]]Few-Shot[[:space:]]Knowledge[[:space:]]Graphs[[:space:]]with[[:space:]]Task-Aware[[:space:]]Language[[:space:]]Models/bfd55c8b-574e-4efa-98f2-4da0aaa19eaf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Inefficiencies[[:space:]]of[[:space:]]Meta[[:space:]]Agents[[:space:]]for[[:space:]]Agent[[:space:]]Design/8f0a3cda-8dae-46fb-92cb-99de5c49db98_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/InfAL_[[:space:]]Inference[[:space:]]Time[[:space:]]Adversarial[[:space:]]Learning[[:space:]]for[[:space:]]Improving[[:space:]]Research[[:space:]]Ideation/b5a730a2-49a3-4e67-b2ca-82e3b2a2f0ce_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/InfiMM-WebMath-40B_[[:space:]]Advancing[[:space:]]Multimodal[[:space:]]Pre-Training[[:space:]]for[[:space:]]Enhanced[[:space:]]Mathematical[[:space:]]Reasoning/e640740c-c7a7-4c38-9250-07fc7b80e4b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Injecting[[:space:]]Domain-Specific[[:space:]]Knowledge[[:space:]]into[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Survey/3a1508fc-ce49-428a-b12f-d4c3abe70ebe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/InsBank_[[:space:]]Evolving[[:space:]]Instruction[[:space:]]Subset[[:space:]]for[[:space:]]Ongoing[[:space:]]Alignment/abf356ff-ef5e-4c89-92ce-9bfc24747fd2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Insights[[:space:]]into[[:space:]]using[[:space:]]temporal[[:space:]]coordinated[[:space:]]behaviour[[:space:]]to[[:space:]]explore[[:space:]]connections[[:space:]]between[[:space:]]social[[:space:]]media[[:space:]]posts[[:space:]]and[[:space:]]influence/65356c92-192f-48a5-ba1f-a606b6919816_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Instability[[:space:]]in[[:space:]]Downstream[[:space:]]Task[[:space:]]Performance[[:space:]]During[[:space:]]LLM[[:space:]]Pretraining/d6dd6187-2a7a-4542-8340-004573337190_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Instance-level[[:space:]]Randomization_[[:space:]]Toward[[:space:]]More[[:space:]]Stable[[:space:]]LLM[[:space:]]Evaluations/95559c12-5bf0-4a01-a6eb-ba493a0cfd9c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/InteGround_[[:space:]]On[[:space:]]the[[:space:]]Evaluation[[:space:]]of[[:space:]]Verification[[:space:]]and[[:space:]]Retrieval[[:space:]]Planning[[:space:]]in[[:space:]]Integrative[[:space:]]Grounding/8266c9ad-a314-4941-9e89-b61ff7582b7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Intelligent[[:space:]]Document[[:space:]]Parsing_[[:space:]]Towards[[:space:]]End-to-end[[:space:]]Document[[:space:]]Parsing[[:space:]]via[[:space:]]Decoupled[[:space:]]Content[[:space:]]Parsing[[:space:]]and[[:space:]]Layout[[:space:]]Grounding/04afef2a-e04c-4f6c-8df3-a211856e297e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Intent-aware[[:space:]]Schema[[:space:]]Generation[[:space:]]and[[:space:]]Refinement[[:space:]]for[[:space:]]Literature[[:space:]]Review[[:space:]]Tables/d2f568b0-c3fc-4ca5-865f-1bbb26f7fd5e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/InterFeedback_[[:space:]]Unveiling[[:space:]]Interactive[[:space:]]Intelligence[[:space:]]of[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models[[:space:]]with[[:space:]]Human[[:space:]]Feedback/2658808c-dc6f-4d8a-a474-da23c43eb3ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/InteractSpeech_[[:space:]]A[[:space:]]Speech[[:space:]]Dialogue[[:space:]]Interaction[[:space:]]Corpus[[:space:]]for[[:space:]]Spoken[[:space:]]Dialogue[[:space:]]Model/fcab0de9-47dd-4778-aa8a-a4c1203466f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Interesting[[:space:]]Culture_[[:space:]]Social[[:space:]]Relation[[:space:]]Recognition[[:space:]]from[[:space:]]Videos[[:space:]]via[[:space:]]Culture[[:space:]]De-confounding/1c7f5a59-df59-416a-b014-39c519869dfb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Internal[[:space:]]states[[:space:]]before[[:space:]]wait[[:space:]]modulate[[:space:]]reasoning[[:space:]]patterns/1c3c5dee-5c90-4299-9751-ba1856956625_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/IntrEx_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Modeling[[:space:]]Engagement[[:space:]]in[[:space:]]Educational[[:space:]]Conversations/adbf96dc-8947-47b2-a360-1283a12a3bf2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Investigating[[:space:]]Controversy[[:space:]]Framing[[:space:]]across[[:space:]]Topics[[:space:]]on[[:space:]]Social[[:space:]]Media/fe76ad70-e896-4379-9c3c-765c2cc6c273_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Investigating[[:space:]]Dictionary[[:space:]]Expansion[[:space:]]for[[:space:]]Video-based[[:space:]]Sign[[:space:]]Language[[:space:]]Dictionaries/c1620fb8-f81a-41b3-a013-1b5d4cc165c3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Investigating[[:space:]]Multi-layer[[:space:]]Representations[[:space:]]for[[:space:]]Dense[[:space:]]Passage[[:space:]]Retrieval/6efe70a5-99ea-42f0-a75e-453684566e45_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Investigating[[:space:]]the[[:space:]]Impact[[:space:]]of[[:space:]]Conceptual[[:space:]]Metaphors[[:space:]]on[[:space:]]LLM-based[[:space:]]NLI[[:space:]]through[[:space:]]Shapley[[:space:]]Interactions/d5b64bad-b0f8-4671-8a3e-b6fabed0fe1e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Invisible[[:space:]]Prompts,[[:space:]]Visible[[:space:]]Threats_[[:space:]]Malicious[[:space:]]Font[[:space:]]Injection[[:space:]]in[[:space:]]External[[:space:]]Resources[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/cd9e4d09-7474-4329-b0d6-6db847262f15_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Invoke[[:space:]]Interfaces[[:space:]]Only[[:space:]]When[[:space:]]Needed_[[:space:]]Adaptive[[:space:]]Invocation[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Question[[:space:]]Answering/b38dd4f2-c7ae-449b-91d7-474df522307f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/IoTMigrator_[[:space:]]LLM-driven[[:space:]]Embedded[[:space:]]IoT[[:space:]]Code[[:space:]]Migration[[:space:]]across[[:space:]]Different[[:space:]]OSes[[:space:]]for[[:space:]]Cloud-device[[:space:]]Integration/9a280433-7bbb-403f-b3e5-01bd45168640_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Is[[:space:]]Safety[[:space:]]Standard[[:space:]]Same[[:space:]]for[[:space:]]Everyone_[[:space:]]User-Specific[[:space:]]Safety[[:space:]]Evaluation[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/09abed6a-dae8-4893-8914-ea7813d86b57_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Jailbreak[[:space:]]Attack[[:space:]]Initializations[[:space:]]as[[:space:]]Extractors[[:space:]]of[[:space:]]Compliance[[:space:]]Directions/892a3e8f-b996-4ee4-8aa7-f6292a9969de_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Jailbreak[[:space:]]Distillation_[[:space:]]Renewable[[:space:]]Safety[[:space:]]Benchmarking/93e80362-1d8f-4476-82d1-2b3ddf5f6f2a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Joint[[:space:]]Enhancement[[:space:]]of[[:space:]]Relational[[:space:]]Reasoning[[:space:]]for[[:space:]]Long-Context[[:space:]]LLMs/1fe92dbd-6dee-4d53-9f3c-1a6adadf067c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Judging[[:space:]]with[[:space:]]Many[[:space:]]Minds_[[:space:]]Do[[:space:]]More[[:space:]]Perspectives[[:space:]]Mean[[:space:]]Less[[:space:]]Prejudice_[[:space:]]On[[:space:]]Bias[[:space:]]Amplification[[:space:]]and[[:space:]]Resistance[[:space:]]in[[:space:]]Multi-Agent[[:space:]]Based[[:space:]]LLM-as-Judge/4f402faa-1235-47f5-9577-4ef513614681_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KAHAN_[[:space:]]Knowledge-Augmented[[:space:]]Hierarchical[[:space:]]Analysis[[:space:]]and[[:space:]]Narration[[:space:]]for[[:space:]]Financial[[:space:]]Data[[:space:]]Narration/d4177a0f-a09a-4764-b4fe-5bb14128f2be_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KBAlign_[[:space:]]Efficient[[:space:]]Self[[:space:]]Adaptation[[:space:]]on[[:space:]]Specific[[:space:]]Textual[[:space:]]Knowledge[[:space:]]Bases/250b77c5-c5c1-46ff-99cc-89d45eba33ee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KBM_[[:space:]]Delineating[[:space:]]Knowledge[[:space:]]Boundary[[:space:]]for[[:space:]]Adaptive[[:space:]]Retrieval[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/d88567de-85b9-4609-bb46-a2baabe640c1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KELE_[[:space:]]A[[:space:]]Multi-Agent[[:space:]]Framework[[:space:]]for[[:space:]]Structured[[:space:]]Socratic[[:space:]]Teaching[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/cac5c00e-c3ed-4cc5-88b6-0d8b21d70363_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KELE_[[:space:]]Residual[[:space:]]Knowledge[[:space:]]Erasure[[:space:]]for[[:space:]]Enhanced[[:space:]]Multi-hop[[:space:]]Reasoning[[:space:]]in[[:space:]]Knowledge[[:space:]]Editing/496badd2-aa54-43de-a5b7-2a54a2c1a2b7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KERAG_[[:space:]]Knowledge-Enhanced[[:space:]]Retrieval-Augmented[[:space:]]Generation[[:space:]]for[[:space:]]Advanced[[:space:]]Question[[:space:]]Answering/c373ff50-8e8a-4131-ba75-0e9acd7c9584_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KaeDe_[[:space:]]Progressive[[:space:]]Generation[[:space:]]of[[:space:]]Logical[[:space:]]Forms[[:space:]]via[[:space:]]Knowledge-Aware[[:space:]]Question[[:space:]]Decomposition[[:space:]]for[[:space:]]Improved[[:space:]]KBQA/50b9cdc0-87af-40f1-a074-b1c9bd35c6f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Knowing[[:space:]]More,[[:space:]]Acting[[:space:]]Better_[[:space:]]Hierarchical[[:space:]]Representation[[:space:]]for[[:space:]]Embodied[[:space:]]Decision-Making/95131e3d-f946-4d0f-804a-66e6b858736c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Knowledge[[:space:]]Graph-Driven[[:space:]]Memory[[:space:]]Editing[[:space:]]with[[:space:]]Directional[[:space:]]Interventions/42f6a5f4-28aa-454b-9465-ec295ad39c6d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Knowledge-enhanced[[:space:]]Multimodal[[:space:]]ECG[[:space:]]Representation[[:space:]]Learning[[:space:]]with[[:space:]]Arbitrary-Lead[[:space:]]Inputs/d0e114fd-1c9d-4c98-ace6-4e6ada2f72f6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Ko-LongRAG_[[:space:]]A[[:space:]]Korean[[:space:]]Long-Context[[:space:]]RAG[[:space:]]Benchmark[[:space:]]Built[[:space:]]with[[:space:]]a[[:space:]]Retrieval-Free[[:space:]]Approach/42eb6703-6159-40aa-adb1-d0769f014e4a_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/I-GUARD_ Interpretability-Guided Parameter Optimization for Adversarial Defense/af7da807-decc-491b-967a-bbededc545c7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c13d61d92d61ba3ccfd8bb49915457f60c9de4ab47e8e12a92f238ba441ed867
+size 1608468

2025/I-GUARD_ Interpretability-Guided Parameter Optimization for Adversarial Defense/full.md

@@ -0,0 +1,443 @@
+# I-GUARD: Interpretability-Guided Parameter Optimization for Adversarial Defense
+
+Mamta and Oana Cocarascu
+
+King's College London
+
+{mamta.name, oana.cocarascu} @kcl.ac.uk
+
+# Abstract
+
+Transformer-based models are highly vulnerable to adversarial attacks, where even small perturbations can cause significant misclassifications. This paper introduces I-Guard, a defense framework to increase the robustness of transformer-based models against adversarial perturbations. I-Guard leverages model interpretability to identify influential parameters responsible for adversarial misclassifications. By selectively fine-tuning a small fraction of model parameters, our approach effectively balances performance on both original and adversarial test sets. We conduct extensive experiments on English and code-mixed Hinglish datasets and demonstrate that I-Guard significantly improves model robustness. Furthermore, we demonstrate the transferability of I-Guard in handling other character-based perturbations.
+
+# 1 Introduction
+
+Transformer-based Pre-trained Language Models (PLM) and Large Language Models (LLM) have achieved remarkable performance across various NLP tasks (Waswani et al., 2017). However, extensive research has shown that they are vulnerable to small perturbations which can significantly affect their performance (Wang et al., 2022b; Le et al., 2022; Qiang et al., 2024; Nguyen and Le, 2024). These weaknesses raise concerns about the robustness and reliability of such models, particularly in real-world applications.
+
+There have been many attempts to study the adversarial robustness of these models across different NLP tasks (Sofi et al., 2022a; Goyal et al., 2023a; Ballout et al., 2024; Mamta and Cocarascu, 2025). Numerous studies have explored strategies to improve adversarial robustness in transformer-based models (Devlin et al., 2019), with data augmentation and adversarial training being widely examined (Goyal et al., 2023b; Morris et al., 2020a).
+
+Data augmentation aims to improve model generalization by generating diverse text variations, but it often requires significant human effort to curate and validate high-quality augmented data. Adversarial training enhances model resilience by exposing it to adversarial examples during training. However, this approach is computationally expensive, requiring models to be re-trained from scratch using both original and carefully crafted adversarial data. Defense methods based on adversarial training also introduce new regularization functions (Liu et al., 2022; Yang et al., 2023) or apply perturbations to the embedding space (Ren et al., 2020; Nguyen Minh and Luu, 2022). However, these approaches depend on a continuous input space and struggle with the discrete nature of text, making it difficult to generate meaningful and semantically coherent interpolations (Nguyen and Le, 2024). Despite these efforts, achieving robust NLP models remains a challenging task. There is a growing need for transparent and efficient methods to enhance adversarial robustness without excessive manual intervention or computational overhead.
+
+In this paper, we propose I-Guard, an interpretability-guided model training method to selectively modify model's parameters to increase the robustness of PLMs and LLMs towards adversarial attacks. For this, we employ model interpretability based on Shapley values (Lundberg and Lee, 2017) to analyze the contribution of each parameter in misleading the model in the presence of adversarial perturbations. For traditional PLMs such as BERT, I-Guard can be applied to the complete model. However, for LLMs with billions of parameters, directly interpreting every parameter is computationally expensive. To address this challenge, we integrate I-Guard with LoRA (Low-Rank Adaptation) (Hu et al., 2022) to find the contribution of each parameter in the LoRA adapters with the help of a probing task.
+
+To measure the efficacy of  $I$ -Guard, we con
+
+sider realistic adversarial attacks, such as phonetic and visual (i.e. LEET) perturbations, which are commonly used in social media. Phonetic variations are especially pronounced in code-mixed texts, where people frequently switch between languages when writing (Das et al., 2022) and spell words phonetically based on their native language's pronunciation, leading to diverse spelling patterns (Crystal, 2018). We evaluate  $I$ -Guard on two tasks, fact verification and code-mixed sentiment analysis (Hindi+English). In addition, we also investigate the transferability of our proposed approach to handle other character-based perturbations such as character deletion, character insertion, and character repetition. Our results show that the proposed defense mechanism effectively enhances the robustness of PLMs and LLMs against adversarial perturbations.
+
+Our contributions are as follows:
+
+1. We propose I-Guard, a novel interpretability-guided parameter training framework to enhance the robustness of PLMs and LLMs against phonetic and LEET-based adversarial attacks.  
+2. To demonstrate the generalizability of our approach, we conduct experiments on English and code-mixed datasets. Our results show that  $I$ -Guard maintains a balanced performance on both original test set and adversarial test set compared to other baselines.  
+3. We assess the transferability of  $I$ -Guard in strengthening model resilience against various character-level perturbations and achieve good defense performance.
+
+# 2 Related Work
+
+# 2.1 Adversarial Robustness
+
+Several studies have shown that high-performing transformer-based models are susceptible to adversarial attacks and minor input perturbations (Lin et al., 2021; Neerudu et al., 2023; Gupta et al., 2024). Adversarial attacks have been explored in a variety of NLP tasks, including sentiment analysis (Jin et al., 2020a; Yuan et al., 2023; Mamta et al., 2023), machine translation (Wang et al., 2021b; Sai et al., 2021; Morris et al., 2020b), argument mining (Mayer et al., 2020; Sofi et al., 2022b), toxic content detection (Yuan et al., 2023), question answering (Goel et al., 2021; Moradi and Samwald, 2021; Kiela et al., 2021; Yuan et al., 2023; Gupta et al., 2024), and natural language inference (Wu et al.,
+
+2021; Morris et al., 2020b; Li et al., 2021; Yuan et al., 2023). Whilst the majority of character-based attacks rely on researchers defining text manipulation strategies, Le et al. (2022) introduced a realistic phonetic perturbation attack by collecting 600K human-written text variations from real-world data and utilizing them for adversarial attacks.
+
+Adversarial defense strategies in NLP can be broadly classified as adversarial training-based defense, data augmentation, and regularization-based defense (Wang and Lin, 2025). Most works employ an adversarial training approach which retransmits the model from scratch by adding adversarial examples to the training data (Li et al., 2019; Wang et al., 2021a, 2020; Jin et al., 2020b; Si et al., 2020). Other approaches incorporate adversarial training as a regularization technique; for example, Flooding-X (Liu et al., 2022), adversarial label smoothing (Yang et al., 2023), and temperature scaling (Raina et al., 2024) have proven effective in improving adversarial robustness. Some studies utilize adversarial training based on Generative Adversarial Networks (Ren et al., 2020) or Virtual Adversarial Training (Li and Qiu, 2021), where perturbations are introduced in the model's embedding space. Denoising-based methods have been proposed to improve adversarial robustness by applying changes to the embedding space of text (Yuan et al., 2024; Ji et al., 2024). However, these approaches often lack semantic correctness and can lead to incoherent modifications (Chen et al., 2020). In addition, there is a lack of explainability and transparency in the regularization-based defense methods (Goyal et al., 2023b). Another line of work focuses on ensembling-based methods, where multiple input text variants are generated at inference time, and predictions are aggregated across these variants. However, these approaches can be inefficient, as they require running the model on each variant, leading to increased inference time proportional to the number of ensembles (Li et al., 2023; Zeng et al., 2023).
+
+# 2.2 Detection-based Defense
+
+There are several methods for detecting adversarial perturbations in text (Goyal et al., 2023c). Some approaches focus solely on identifying and filtering these adversarial inputs, while others use spell checkers or rule-based techniques to correct the perturbations. However, these corrections often rely on predefined rules or dictionaries, which may be ineffective when the spelling of a word devi
+
+ates from those present in the corpus or dictionary. Approaches for identifying adversarial perturbations include the Synonym Encoding Method (Wang et al., 2021c), frequency-aware randomization frameworks (Bao et al., 2021) for detecting word substitutions, and robust density estimation (Yoo et al., 2022), among others. Our objective is to improve the robustness of models against adversarial perturbations, in particular phonetic and LEET-based attacks, which are inspired by real-world noise in the data. Rather than filtering such perturbations, we aim to build models that are robust and capable of handling them effectively.
+
+# 2.3 Transformers Interpretation
+
+The remarkable success of pre-trained transformer-based models has driven researchers to explore their interpretability in-depth, aiming to explain their black-box nature (Petroni et al., 2019; Liu et al., 2019; Hewitt and Manning, 2019). Many works define neurons as dimensions within contextualized representations and investigate the linguistic information encoded by these representations (Durrani et al., 2020; Dalvi et al., 2019). Other studies focus on analyzing multi-head self-attention layers (Clark et al., 2019; Voita et al., 2019) and examine the roles of different attention heads across various tasks (Gould et al., 2024; Conmy et al., 2023; Hanna et al., 2024).
+
+Several works (Geva et al., 2021; Dai et al., 2022a) have explored the neurons in the feedforward neural networks within transformer models, revealing that neurons in these layers encode word patterns and conceptual knowledge. For instance, Dai et al. (2022b) examined knowledge neurons in feed-forward networks, demonstrating their function in encoding factual knowledge for fill-in-the-blank tasks in BERT. Similarly, Wang et al. (2022a) identified skill neurons within the feed-forward layers of pre-trained transformers after prompt tuning for a task, showing that these neurons capture task-specific abilities. More recently, Yu and Ananiadou (2024) proposed a method to study neuron-level knowledge attribution in large language models by identifying query neurons which activate value neurons, offering deeper insights into model predictions. Kulkarni and Weng (2024) proposed a test-time defense primarily for image classification. Their approach focuses on neuron-level interpretability, computing importance based on methods like Leave-One-Out (LO-IR) or CLIP-Dissect (CD-IR) to identify and
+
+mask unimportant neuron activations at test time.
+
+Our work differs from the aforementioned studies as follows. Instead of identifying neuron contributions, we conduct a fine-grained analysis of model parameters, focusing on those responsible for misleading the model. Based on this analysis, we further fine-tune the model to enhance its robustness against adversarial manipulations.
+
+# 3 Methodology
+
+# 3.1 Problem Formulation
+
+Attack Goal For a given input  $X$  consisting of  $n$  tokens  $\{x_{1},x_{2},\ldots ,x_{n}\}$  and an associated ground truth label  $y$ , the attack objective is to generate a perturbed version  $X^{\prime}$  such that the target model  $M$  misclassifies it, i.e.,  $M(X^{\prime})\neq y$  (untargeted attack). In the case of two-input tasks, where the input consists of  $X$  and  $E$ , the adversary applies perturbations only to  $X$ , resulting in a modified input  $X^{\prime}$ , while keeping  $E$  unchanged. The goal remains to mislead the model such that  $M(X^{\prime},E)\neq y$ . The attack is carried out using phonetic and LEET-based perturbations, following prior work (Le et al., 2022; Das et al., 2022).
+
+Defense Goal The defense mechanism must meet the following requirements: 1) Enhance robustness against phonetic and LEET-based perturbations; and 2) Ensure that the performance on the actual test set remains consistent with the original performance.
+
+# 3.2 I-Guard Framework
+
+In this section, we present  $I$ -Guard, a defense framework that enhances the robustness of the model against phonetic and LEET-based perturbations. Figure 1 depicts the main steps in  $I$ -Guard. The adversarial generator applies perturbations to actual examples to mislead the trained model  $M$ . We then identify the influential parameters responsible for the misclassification of adversarial examples using a subset of the misclassified adversarial examples and their corresponding actual examples. To determine these influential parameters,  $I$ -Guard leverages a model interpretation technique based on Shapley values (Lundberg and Lee, 2017). In the case of LLMs, our approach involves identifying specific LoRA parameters that contribute most to adversarial misclassifications. These selected parameters are then optimized to improve the model's robustness.
+
+![](images/670df6457fcaa677505a31adef834654f5a941f6146ce951653ad0ba0177c12b.jpg)  
+Figure 1: Key steps of I-Guard.
+
+Adversarial Sample Generation To generate adversarial samples, we perform a black-box attack on BERT (Devlin et al., 2019) and Llama (Touvron et al., 2023) models. We follow Le et al. (2022) and Das et al. (2022) to apply perturbations to English and Hindi+English (Hinglish), respectively. To implement the phonetic and LEET-based perturbations for BERT, we follow Mamta and Ekbal (2022) to identify the important tokens. For Llama, we use feature ablation to identify the most important words for the model's predictions (Miglani et al., 2023). Next, we perturb the words in descending order of their importance scores until a successful attack is executed or all words have been perturbed.
+
+Model Interpretation Shapley values originate from cooperative game theory, where players contribute to the overall outcome of a game (Lundberg and Lee, 2017). Here, we treat model parameters as players in a cooperative game, where their collective interaction determines the model's predictions. The Shapley value for each parameter represents its marginal contribution across all possible parameter combinations, providing a comprehensive measure of its importance. The mathematical formulation for the Shapley values is:
+
+$$
+I _ {p} (v) = \sum_ {S \subseteq \Omega \backslash \{p \}} \frac {| S | ! (\left| \Omega \right| - | S | - 1) !}{\left| \Omega \right| !} [ v (S \cup \{p \}) - v (S) ] \tag {1}
+$$
+
+where  $\psi_p(v)$  represents the contribution score of parameter  $p$ , with  $\Omega = \{1,2,\dots,n\}$  denoting the complete set of  $n$  model parameters. The term  $S$  encompasses all possible parameter subsets, and  $v(S)$  represents the model's prediction using only the parameters in subset  $S$ , with other parameters left out. This formulation ensures a fair attribution of importance to each parameter by considering all possible combinations of parameter interactions.
+
+Computing Shapley values (Equation 1) presents significant computational challenges, particularly for large neural networks. The computational com
+
+plexity of  $\mathrm{O}(2^{\Omega})$  makes direct calculations inefficient for models with millions of parameters (Kang et al., 2023). Furthermore, in large-scale neural networks, the individual impact of single parameters on the overall prediction is typically minimal, making exhaustive computation inefficient.
+
+To overcome this, we estimate the marginal contributions through random sampling by sampling  $k$  sized parameter subsets from  $\Omega$  several times. This sampling-based approach significantly reduces computational overhead. As discussed in Kang et al. (2023), the marginal contribution calculation is approximated using the norm of weight-gradient products:
+
+$$
+\begin{array}{l} | v (S \cup p) - v (S) | \approx | - L (\omega + \Delta \omega_ {i}, x, y) + L (\omega , x, y) | \\ \approx \left| \frac {\partial L}{\partial \omega_ {i}} \Delta \omega_ {i} \right| = \left| \frac {\partial L}{\partial \omega_ {i}} \right| | \Delta \omega_ {i} | = \left| \frac {\partial L}{\partial \omega_ {i}} \right| | \omega_ {i} |, \tag {2} \\ \end{array}
+$$
+
+This approximation leverages the first-order Taylor expansion of the loss function  $L$ . The contribution of parameter  $\omega_{i}$  is approximated using the gradient of the loss with respect to the parameter and its magnitude.
+
+Identifying Parameters for Adversarial Misclassifications To identify the key parameters responsible for adversarial misclassification, we follow a systematic approach involving adversarial sample generation, dataset partitioning, and Shapley value-based parameter importance estimation. The process is shown in Algorithm 1.
+
+First, we generate adversarial samples for a specific perturbation type using the training set (line 2). We then pass the generated adversarial samples through the trained model  $M$  and analyze the predictions. Based on the classification outcomes, we make a subset of the dataset for misclassified Adversarial Samples. This includes original  $(D_{act})$  and corresponding adversarial samples  $(D_{adv})$  where examples are initially classified correctly by the model  $M$ , but their adversarial counterparts are misclassified (lines 3-13).
+
+To understand the influence of individual model parameters in adversarial misclassifications, we apply Shapley value-based parameter importance estimation to both  $D_{act}$  and  $D_{adv}$  (lines 38-39). This method assigns importance scores  $(I_{act}$  and  $I_{adv}$ ) to each parameter by evaluating its contribution to the model's predictions. The importance estimation process relies on repeated random sampling ( $K$  iterations), where, in each iteration  $m$ , a subset
+
+Algorithm 1 Parameter Identification for Adversarial Misclassifications  
+Require: Model M, Training dataset  $\mathcal{D}_{\mathrm{train}}$  , Perturbation type  $\epsilon$  , Sampling iterations  $K$  , Sample size  $k$  , Importance threshold  $t$  , Fine-tuning epochs  $E$    
+Ensure: Refined model  $M^{*}$    
+1: // Generate adversarial samples   
+2:  $\mathcal{D}_{\mathrm{advr}}\gets$  GenerateAdversarialSamples(  $\mathcal{D}_{\mathrm{train}}$ $M,\epsilon)$    
+3: // Initialize dataset partitions   
+4:  $D_{\mathrm{act}}\leftarrow \{\}$ $D_{\mathrm{adv}}\leftarrow \{\}$    
+5: for  $(x,y)$  in  $\mathcal{D}_{\mathrm{train}}$  do   
+6:  $x_{\mathrm{adv}}\gets \mathcal{D}_{\mathrm{advr}}[x]$    
+7:  $y_{\mathrm{pred}}\gets M(x)$    
+8:  $y_{\mathrm{adv}}\gets M(x_{\mathrm{adv}})$    
+9: if  $y_{\mathrm{pred}} = y$  and  $y_{\mathrm{adv}}\neq y$  then   
+10:  $D_{\mathrm{act}}\gets D_{\mathrm{act}}\cup \{(x,y)\}$    
+11:  $D_{\mathrm{adv}}\gets D_{\mathrm{adv}}\cup \{(x_{\mathrm{adv}},y)\}$    
+12: end if   
+13: end for   
+14: // Initialize importance score vectors   
+15:  $I_{\mathrm{act}}\gets 0|_M|$ $I_{\mathrm{adv}}\gets 0|M|$    
+16: // Compute parameter importance   
+17: function CALCULATEIMPORTANCE(D,M)   
+18: for  $(x,y)$  in  $D$  do   
+19: for  $m = 1$  to  $K$  do   
+20:  $k\gets k\times |M|$ $\triangleright$  Compute k% of elements   
+21: mask_values  $\leftarrow$  RandomSample(M,k)  $\triangleright$  Select  $k$  random indices from complete model M or LoRA adapters   
+22:  $M_{\mathrm{masked}}\gets M$    
+23:  $M_{\mathrm{masked}}[\mathrm{mask\_indices}]\gets 0$    
+24:  $\mathcal{L}\gets \mathcal{L}(y,M_{\mathrm{masked}}(x))$    
+25: Compute  $\nabla_{M_{\mathrm{masked}}}L$    
+26: start_in, end_in  $\leftarrow 0$  , 0   
+27: for par in  $M_{\mathrm{masked}}.\textit{Params} ()$  do   
+28: start_in, end_in  $\leftarrow$  end_in, end_in + |par|   
+29: if pargradient  $\neq$  None then   
+30:  $w\gets M_{\mathrm{masked}}[\mathrm{start\_idx}:\mathrm{end\_idx}]$    
+31:  $I_{\mathrm{act}}[\mathrm{start\_idx}:\quad : \quad \mathrm{end\_idx}] \leftarrow I_{\mathrm{act}}[\mathrm{start\_idx}:\quad : \quad \mathrm{end\_idx}] \leftarrow$    
+32: end if   
+33: end for   
+34: end for   
+35: end for   
+36: return I   
+37: end function   
+38: // Compute parameter importance for actual samples  $I_{\mathrm{act}}\gets \textit{CalculateImportance} (D_{\mathrm{act}},M)$    
+39: // Compute parameter importance for adversarial samples  $I_{\mathrm{adv}}\gets \textit{CalculateImportance} (D_{\mathrm{adv}},M)$    
+40: // Compute importance difference and identify crucial parameters   
+41:  $I_{\mathrm{diff}}\gets I_{\mathrm{adv}} - I_{\mathrm{act}}$    
+42:  $P_{\mathrm{critical}}\gets \textit{SelectTopParameters} (I_{\mathrm{diff}},t)$    
+43: // Freezing selective parameters   
+44:  $M^{*}\gets M$    
+45: for each layer l in  $M^{*}$  do   
+46: for each submodule sm in l do   
+47: if sm  $\in P_{\mathrm{critical}}$  then   
+48: smrequires_grad  $\leftarrow$  True   
+49: else   
+50: smrequires_grad  $\leftarrow$  False   
+51: end if   
+52: end for   
+53: end for   
+54: Fine-tune  $M^{*}$  on  $D_{\mathrm{train}}$  and  $D_{\mathrm{advr}}$    
+55: return  $M^{*}$
+
+of parameters is randomly chosen to assess their marginal contribution. For subset selection, we use RandomSample  $(M,k)$ , which selects  $k\%$  parameter indices to be temporarily zeroed out, forming a zeroed out model  $M_{\mathrm{masked}}$  (lines 18-23). Here,  $|M|$  is the size of model, i.e., total number of parameters present in the model. A forward pass with  $M_{\mathrm{masked}}$  model produces predictions  $\hat{y} = M_{\mathrm{masked}}(\mathbf{x})$ , followed by computing the loss  $\mathcal{L}(y,\hat{y})$ . Backpropagation then yields gradients  $\nabla_{M_{\mathrm{masked}}} \mathcal{L}$ , which, when multiplied by parameter values, approximate the marginal contributions (lines 24-31).
+
+Finally, the importance scores for each parameter are accumulated across multiple iterations and batches (lines 27-34). Since there exists a one-to-one correspondence between the actual and adversarial subsets, some parameters which are crucial for correctly classifying  $D_{act}$  may not hold the same significance in  $D_{adv}$ . This discrepancy arises due to the perturbations altering the input space, causing certain parameters to play a more dominant role in misclassifications. Therefore, to pinpoint the parameters responsible for adversarial misclassification, we compute the difference between importance scores  $I_{adv}$  and  $I_{act}$  (line 41). This difference highlights parameters that significantly contribute to the misclassification of adversarial samples. I-Guard adapts parameter interpretation strategy for different types of language models, considering their architectural differences and computational requirements.
+
+Pre-trained Language Models For BERT like models,  $I$ -Guard applies Shapley value-based model interpretation to the complete model as shown in Figure 1. This involves analyzing the contributions of parameters across all layers and components of the entire BERT model.
+
+Large Language Models Due to the high computational cost associated with fine-tuning and interpreting LLMs,  $I$ -Guard applies Shapley-based interpretation to the LoRA adapters (Hu et al., 2022) as shown in Figure 1.  $I$ -Guard utilizes a model finetuned using the LoRA technique. We perform this initial fine-tuning with a Causal Language Modeling (CLM) objective, rather than a direct classification objective. This choice underscores the generic nature of our approach. After LoRA-based fine-tuning, we perform linear probing to adapt the LLM's learned representations for classification tasks. This is because the CLM loss, designed for next-token prediction, does not inherently provide a direct gradient signal which is required to
+
+understand the influence of the parameters on a downstream classification task.
+
+Probing We train a classifier (probe) on the top of hidden representation of the last layer to map this hidden representation to the task label. During fine-tuning, the weights of the LLM and LoRA adapters are frozen to ensure that the probe is using the existing task-specific information, rather than modifying the underlying representation. Therefore, only the parameters of the linear probe are updated during this training phase.
+
+After training, we freeze parameters and use Shapley-based model interpretation (lines 17-37) to calculate the importance of LoRA parameters. As the trained probe transforms the hidden states into differentiable class logits, this transformation enables the calculation of a loss that is directly linked to the classification task.
+
+Selective Parameter Optimization Based on the final importance scores, we identify the top  $t$  most influential parameters (line 42). In transformer-based models, each layer consists of Self-Attention Weights (Query, Key, and Values), Feed-Forward Network weights (intermediate or output), and Layer Norm Weights. Instead of fine-tuning individual weights, we selectively fine-tune the specific submodules (e.g., output Feed-Forward Network, submodules of LoRA) where these important parameters are located, while keeping the rest of the model frozen (lines 42-53). The model is then finetuned on a mix of actual and adversarial samples (line 54).
+
+# 4 Experiments
+
+# 4.1 Datasets
+
+We conduct extensive evaluation of  $I$ -Guard on English and Hinglish datasets, covering two tasks.
+
+Fact verification (English) We use CLIMATE-FEVER (Diggelmann et al., 2020) which contains claims related to climate change along with their evidence. Each claim is labeled with one of three classes: supports, refutes, and not-enough-info (NEI). The dataset includes 7,675 annotated claim-evidence pairs split into train (5,756), validation (767), and test set (1,152).
+
+Sentiment analysis (Hindi + English) We use a Hinglish dataset comprising posts from various public Facebook pages (Joshi et al., 2016). Each Hinglish post is labeled with one of three sentiment categories: positive, negative, or neutral. In total,
+
+the dataset includes 3,879 instances split into train (2,482), validation (621), and test set (776).<sup>2</sup>
+
+# 4.2 Baselines
+
+We compare I-Guard with the following baselines: BERT (Devlin et al., 2019): We fine-tune BERT-base and BERT-base-multilingual (mBERT) models on English and Hinglish datasets, respectively, by adding a dense layer on top of it.
+
+Llama: We fine-tune Llama-3.2-1B model on English and Hinglish datasets using LoRA adapters with CLM objective.
+
+Adversarial Training: We create adversarial samples using phonetic and LEET-based perturbations and incorporate them into the original training data, re-training the model from scratch on the combined dataset.
+
+LoRA based Adversarial Training: We re-train the LoRA adapters of Llama from scratch on the combined dataset.
+
+I-Guard-Adv: We fine-tune the selective parameters of the model on only adversarial samples.
+
+# 5 Results and Discussion
+
+Table 1 shows the accuracy and  $F_{1}$  score on both the actual test set and adversarial test set (TS), as well as the percentage of fine-tuned parameters (PT) for both datasets. For Llama, we show the number of fine-tuned LoRA adapters.
+
+English Dataset The accuracy and  $F_{1}$  scores for BERT-base and Llama models drop significantly under phonetic and LEET-based perturbations. For instance, under phonetic perturbations for BERT (FV Phonetic), the  $F_{1}$  score drops by  $36.72\%$ . The most widely used adversarial defense, adversarial training, significantly improves both accuracy and  $F_{1}$  scores for these perturbations in BERT and Llama models. However, this approach is computationally expensive, as the model needs to be trained from scratch (100% PT and all LoRA adapters).
+
+$I$ -Guard-Adv also effectively handles phonetic and LEET-based perturbations by fine-tuning much fewer parameters than adversarial training. In phonetic perturbations,  $I$ -Guard achieves an  $F_{1}$  score of  $59.53\%$  while fine-tuning only  $21.79\%$  of BERT's parameters, compared to adversarial training, which requires  $100\%$  fine-tuning. Similarly, for Llama, only 20 LoRA adapters are updated. For LEET-based perturbations also,  $I$ -Guard enhances the robustness while fine-tuning fewer parameters.
+
+<table><tr><td rowspan="3" colspan="2"></td><td colspan="6">Phonetic</td><td colspan="5">LEET</td></tr><tr><td colspan="2">Original TS</td><td colspan="2">Adv TS</td><td rowspan="2">PT</td><td colspan="2">Original TS</td><td colspan="2">Adv TS</td><td rowspan="2" colspan="2">PT</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td rowspan="4">\( \mathbf{{FV}} \)</td><td>BERT-base</td><td>71.61</td><td>62.70</td><td>51.21</td><td>25.98</td><td>-</td><td>71.61</td><td>62.70</td><td>49.91</td><td>22.52</td><td colspan="2">-</td></tr><tr><td>Adv training</td><td>71.44</td><td>61.02</td><td>75.17</td><td>65.42</td><td>100%</td><td>71.61</td><td>59.96</td><td>76.90</td><td>68.24</td><td colspan="2">100%</td></tr><tr><td>I-Guard-adv</td><td>60.58</td><td>58.03</td><td>73.87</td><td>64.38</td><td>21.79%</td><td>59.80</td><td>57.21</td><td>73.43</td><td>64.98</td><td colspan="2">26.11%</td></tr><tr><td>I-Guard</td><td>69.27</td><td>62.40</td><td>69.79</td><td>59.53</td><td>21.79%</td><td>72.13</td><td>63.74</td><td>74.82</td><td>66.82</td><td colspan="2">26.11%</td></tr><tr><td rowspan="4">SA</td><td>mBERT</td><td>67.26</td><td>62.58</td><td>40.46</td><td>28.09</td><td>-</td><td>67.26</td><td>62.58</td><td>40.25</td><td>20.77</td><td colspan="2">-</td></tr><tr><td>Adv training</td><td>67.13</td><td>64.80</td><td>60.43</td><td>56.93</td><td>100%</td><td>65.72</td><td>60.83</td><td>72.16</td><td>67.60</td><td colspan="2">100%</td></tr><tr><td>I-Guard-adv</td><td>68.68</td><td>65.00</td><td>56.31</td><td>50.38</td><td>28.98%</td><td>59.77</td><td>58.17</td><td>71.77</td><td>66.39</td><td colspan="2">27.67%</td></tr><tr><td>I-Guard</td><td>67.96</td><td>64.94</td><td>56.31</td><td>51.36</td><td>28.98%</td><td>67.78</td><td>64.72</td><td>67.13</td><td>61.20</td><td colspan="2">27.67%</td></tr></table>
+
+Table 1: Results on Fact Verification (FV) and Sentiment Analysis (SA) for BERT. Here, TS: test set, Adv TS: Adversarial test set, PT: percentage of fine-tuned parameters.  
+
+<table><tr><td rowspan="3" colspan="2"></td><td colspan="6">Phonetic</td><td colspan="5">LEET</td></tr><tr><td colspan="2">Original TS</td><td colspan="2">Adv TS</td><td rowspan="2">LM</td><td colspan="2">Original TS</td><td colspan="2">Adv TS</td><td rowspan="2" colspan="2">LM</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td rowspan="4">\( FV \)</td><td>Llama</td><td>57.20</td><td>35.84</td><td>15.00</td><td>8.98</td><td>224</td><td>57.20</td><td>35.84</td><td>18.76</td><td>10.81</td><td>224</td><td></td></tr><tr><td>LoRA Adv</td><td>43.23</td><td>33.70</td><td>31.93</td><td>18.03</td><td>224</td><td>51.11</td><td>29.71</td><td>31.93</td><td>18.03</td><td>224</td><td></td></tr><tr><td>I-Guard-adv</td><td>59.45</td><td>35.76</td><td>42.75</td><td>22.30</td><td>20</td><td>57.60</td><td>32.16</td><td>39.26</td><td>20.92</td><td>20</td><td></td></tr><tr><td>I-Guard</td><td>58.77</td><td>35.58</td><td>39.09</td><td>20.83</td><td>20</td><td>56.27</td><td>33.24</td><td>40.13</td><td>21.02</td><td>20</td><td></td></tr><tr><td rowspan="4">SA</td><td>Llama</td><td>45.74</td><td>38.6</td><td>19.29</td><td>14.91</td><td>224</td><td>45.74</td><td>38.60</td><td>22.55</td><td>13.56</td><td>224</td><td></td></tr><tr><td>LoRA Adv</td><td>51.80</td><td>39.04</td><td>35.28</td><td>25.19</td><td>224</td><td>50.23</td><td>42.73</td><td>39.94</td><td>29.11</td><td>224</td><td></td></tr><tr><td>I-Guard-adv</td><td>51.15</td><td>39.73</td><td>36.98</td><td>26.05</td><td>25</td><td>51.59</td><td>41.35</td><td>38.53</td><td>28.49</td><td>24</td><td></td></tr><tr><td>I-Guard</td><td>52.06</td><td>39.59</td><td>38.65</td><td>27.97</td><td>25</td><td>49.87</td><td>38.68</td><td>40.46</td><td>30.01</td><td>24</td><td></td></tr></table>
+
+Table 2: Results on Fact Verification (FV) and Sentiment Analysis (SA) for LLama. Here, TS: test set, Adv TS: Adversarial test set, LM: fine-tuned LoRA modules.  
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="2">BERT</td><td colspan="2">I-Guard</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td rowspan="3">\( \mathbf{{FV}} \)</td><td>Char Delete</td><td>49.82</td><td>22.52</td><td>68.05</td><td>55.92</td></tr><tr><td>Char Insert</td><td>49.56</td><td>22.47</td><td>66.49</td><td>52.99</td></tr><tr><td>Char Repetition</td><td>50.00</td><td>22.93</td><td>68.66</td><td>58.05</td></tr><tr><td rowspan="3">SA</td><td>Char Delete</td><td>37.75</td><td>26.20</td><td>52.83</td><td>48.15</td></tr><tr><td>Char Insert</td><td>35.69</td><td>26.63</td><td>48.84</td><td>44.65</td></tr><tr><td>Char Repetition</td><td>46.64</td><td>38.06</td><td>54.76</td><td>50.04</td></tr></table>
+
+Table 3: Transferability of  $I$  -Guard to other attacks.
+
+Moreover, adversarial training and  $I$ -Guard-adv reduce the  $F_{1}$  score on the original test set (BERT). In contrast,  $I$ -Guard not only improves robustness against adversarial perturbations, but also enhances performance on the original test set, thereby maintaining a good balance between performance on both the original and adversarial test sets.
+
+Hinglish Dataset We observe a similar phenomenon on the sentiment analysis task, where  $I$ -Guard achieves better generalization compared to adversarial training-based defense for the BERT and Llama models. This demonstrates that  $I$ -Guard can enhance robustness against adversarial text perturbations by fine-tuning fewer parameters compared to adversarial training-based defense. We also conducted experiments on a larger dataset (see Appendix D).
+
+# 5.1 Transferability to other Perturbations
+
+In addition, we assess the transferability of model trained using  $I$ -Guard to other types of adversarial attacks. Specifically, we generate adversarial test sets using character repetition, character deletion, and character insertion perturbations (Moradi and Samwald, 2021) as discussed in Section 3.2.
+
+The results are presented in Table 3. We observe a significant drop in accuracy and  $F_{1}$  score for BERT (base and multilingual) across all types of perturbations in both English and Hinglish. However, our proposed method demonstrates robustness against these perturbations, leading to improved accuracy and  $F_{1}$  scores in both languages. This indicates that I-Guard not only defends against the phonetic perturbations considered, but also offers enhanced generalization capabilities to handle other related perturbations effectively.
+
+# 5.2 Ablation Study
+
+Affect of  $t$  We experiment with different values of  $t$  (top  $t$  parameters from  $I_{diff}$ ) to understand their impact on the robustness and actual performance of the model. Table 4 presents BERT's behaviour for various parameter values. For the English language, we observe that fine-tuning only  $0.016\%$  (for  $t = 50$ ) of the model parameters significantly improves its robustness against phonetic perturbations. Similarly, for LEET-based perturba
+
+<table><tr><td rowspan="3"></td><td rowspan="3">t</td><td colspan="5">Phonetic</td><td colspan="5">LEET</td></tr><tr><td colspan="2">Original TS</td><td colspan="2">Adv TS</td><td rowspan="2">PT</td><td colspan="2">Original TS</td><td colspan="2">Adv TS</td><td rowspan="2">PT</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td rowspan="6">\( \mathbf{{FV}} \)</td><td>50</td><td>67.88</td><td>61.58</td><td>68.48</td><td>57.07</td><td>0.016</td><td>67.27</td><td>60.91</td><td>67.36</td><td>56.50</td><td>0.018</td></tr><tr><td>100</td><td>68.14</td><td>61.60</td><td>68.48</td><td>57.57</td><td>0.020</td><td>67.53</td><td>60.82</td><td>68.66</td><td>58.28</td><td>0.023</td></tr><tr><td>150</td><td>69.18</td><td>62.34</td><td>69.77</td><td>59.39</td><td>21.791</td><td>67.62</td><td>60.95</td><td>68.83</td><td>58.51</td><td>0.024</td></tr><tr><td>200</td><td>69.35</td><td>62.54</td><td>69.79</td><td>59.39</td><td>21.793</td><td>72.13</td><td>63.74</td><td>74.82</td><td>66.82</td><td>26.105</td></tr><tr><td>250</td><td>69.27</td><td>62.40</td><td>69.79</td><td>59.53</td><td>21.794</td><td>72.04</td><td>63.85</td><td>73.78</td><td>64.99</td><td>26.106</td></tr><tr><td>300</td><td>69.27</td><td>62.40</td><td>69.79</td><td>59.53</td><td>21.798</td><td>72.04</td><td>63.85</td><td>73.78</td><td>64.99</td><td>26.106</td></tr><tr><td rowspan="6">SA</td><td>50</td><td>68.42</td><td>64.08</td><td>48.32</td><td>39.97</td><td>0.019</td><td>68.29</td><td>63.82</td><td>47.42</td><td>34.55</td><td>0.017</td></tr><tr><td>100</td><td>68.29</td><td>63.81</td><td>48.06</td><td>39.55</td><td>0.020</td><td>68.55</td><td>65.67</td><td>64.17</td><td>57.89</td><td>0.021</td></tr><tr><td>150</td><td>68.42</td><td>64.28</td><td>51.03</td><td>44.64</td><td>0.022</td><td>67.65</td><td>64.51</td><td>65.07</td><td>58.53</td><td>0.382</td></tr><tr><td>200</td><td>68.42</td><td>64.28</td><td>51.03</td><td>44.64</td><td>0.023</td><td>67.78</td><td>64.72</td><td>67.13</td><td>61.20</td><td>27.67</td></tr><tr><td>250</td><td>67.91</td><td>63.94</td><td>56.31</td><td>51.36</td><td>24.65</td><td>67.78</td><td>64.71</td><td>67.13</td><td>61.20</td><td>30.34</td></tr><tr><td>300</td><td>67.96</td><td>64.94</td><td>56.31</td><td>51.36</td><td>28.98</td><td>67.43</td><td>64.21</td><td>67.13</td><td>61.20</td><td>32.21</td></tr></table>
+
+tions, fine-tuning just  $0.018\%$  of the model parameters greatly enhances its robustness.
+
+Increasing the value of  $t$  results in improved performance on both the actual and adversarial test sets. However, further increasing the number of fine-tuned parameters can negatively impact the model's performance. For example, increasing  $t$  from 200 to 250 leads to performance decrease on adversarial test set (FV LEET). Further increasing the number of parameters from 250 to 300 results in the same performance as when  $t$  is set to 250. In case of the Hinglish language also, fine-tuning only  $0.019\%$  (phonetic) and  $0.017\%$  (LEET) of the parameters improves performance on both the actual and adversarial test sets. Increasing the number of parameters from 250 to 300 (SA phonetic) does not affect the robustness, but it does lead to an increase in performance on actual test set.
+
+We also observe that for the top 50, 100, and 150 indices (FV LEET), the percentage (PT) remains low (less than  $1\%$ ). This indicates that these indices correspond primarily to a small subset of parameters within a few submodules. However, when the number of top  $t$  parameters increases from 150 to 200, there is a substantial rise in PT, suggesting that the additional indices span across multiple submodules, significantly increasing the number of parameters being fine-tuned. Beyond 200 indices, the PT stabilizes at around  $26.11\%$ , indicating that most of the newly added weight parameters belong to already unfrozen submodules rather than introducing entirely new ones. This pattern is consistent for both languages and perturbations.
+
+Random parameter selection We randomly select  $30\%$  model parameters and fine-tune them on (i) adversarial data; (ii) mixture of adversarial and original data to observe the impact of model interpretability on targeted parameter optimization. We
+
+Table 4: Affect of choosing different parameter values (top  $t$  ).  
+
+<table><tr><td></td><td>Data</td><td colspan="2">Original TS</td><td colspan="2">Adv TS</td></tr><tr><td></td><td></td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>Random</td><td rowspan="2">Mix</td><td>67.55</td><td>64.67</td><td>62.98</td><td>57.76</td></tr><tr><td>I-Guard</td><td>67.78</td><td>64.72</td><td>67.13</td><td>61.2</td></tr><tr><td>Random</td><td rowspan="2">Adv</td><td>44.58</td><td>41.05</td><td>74.8</td><td>70.09</td></tr><tr><td>I-Guard-adv</td><td>59.77</td><td>58.17</td><td>71.77</td><td>66.39</td></tr></table>
+
+Table 5: Results on LEET perturbations for sentiment analysis by ablating model interpretability component.
+
+compare this random selection with two variants of our approach,  $I$ -Guard and  $I$ -Guard-adv. The results for BERT are reported in Table 5.
+
+$I$ -Guard outperforms random parameter finetuning on both the original and adversarial test sets. Thus, leveraging interpretability techniques to identify and optimize the most relevant parameters leads to more robust performance. When training the model only on adversarial test data, random selection achieves higher accuracy and  $F_{1}$  score, but at the cost of a significant reduction in accuracy on the original test set. However,  $I$ -Guard-adv performs better on the original test set, highlighting that a targeted optimization approach focused on adversarial data can enhance model robustness without overfitting to adversarial examples. We also fine-tuned the last layer of the model (see results in Appendix C.2).
+
+Affect of  $k$  We conducted an ablation study by varying the value of  $k$  in Shapley calculation and subsequently evaluating the fine-tuned model's performance (Llama).  $k$  represents the sample-size of parameters considered within each repeat of Shapley approximation. The results for different  $k$  values on both clean and adversarial performance, are presented in Table 7. We observe a slight improvement in adversarial test set accuracy and  $F_{1}$  score as  $k$  increases from 10 to 50. This means a larger sample-size  $k$  leads to more fair and stable comparisons in Shapley approximation, poten
+
+<table><tr><td></td><td>Original Claim</td><td>Adversarial Claim</td><td>Original Label</td><td>BERT-base</td><td>Llama</td><td>BERT Adv Training</td><td>LoRA Adv Training</td><td>I-Guard</td></tr><tr><td>1</td><td>There are many lines of evidence which clearly show that the atmospheric co2 increase is caused by humans.</td><td>There are many lines of evidence which clearly show that the atmospheric co2 increase is caused by humans.</td><td>Support</td><td>NEI</td><td>Support</td><td>Support</td><td>Support</td><td>Support</td></tr><tr><td>2</td><td>Climate change will also reduce the number of cold days and cold spells.</td><td>Climate changee will also reduce the number of coooold days and cold spells.</td><td>Support</td><td>NEI</td><td>NEI</td><td>Support</td><td>Support</td><td>Support</td></tr><tr><td>3</td><td>In the past, warming has never been a threat to life on earth.</td><td>In the past, warming has neveeeeer been a threat to life on earth.</td><td>Refute</td><td>NEI</td><td>Support</td><td>Refute</td><td>Refute</td><td>Refute</td></tr><tr><td>4</td><td>Temperatures in the arctic have soared recently, and scientists are struggling to explain exactly why.</td><td>-</td><td>Support</td><td>Support</td><td>Support</td><td>NEI</td><td>NEI</td><td>Support</td></tr><tr><td>5</td><td>New study confirms evs considerably worse for climate than diesel cars.</td><td>-</td><td>Refute</td><td>Refute</td><td>Refute</td><td>NEI</td><td>NEI</td><td>Refute</td></tr></table>
+
+Table 6: Behaviour of different models on adversarial (1-3) and original examples (4-5).  
+
+<table><tr><td></td><td colspan="2">Original TS</td><td colspan="3">Adv TS</td></tr><tr><td></td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>k</td></tr><tr><td>Llama</td><td>45.74</td><td>38.60</td><td>22.55</td><td>13.56</td><td></td></tr><tr><td>I-Guard</td><td>50.90</td><td>38.52</td><td>39.43</td><td>28.00</td><td>k=10</td></tr><tr><td>I-Guard</td><td>49.87</td><td>38.68</td><td>40.46</td><td>30.01</td><td>k=25</td></tr><tr><td>I-Guard</td><td>50.51</td><td>38.50</td><td>40.59</td><td>29.76</td><td>k=50</td></tr></table>
+
+Table 7: Affect of  $k$  values on LEET perturbations for sentiment analysis.
+
+tially yielding more accurate and robust importance scores for parameters critical to adversarial defense. The model's performance on the original test set remains stable across the different  $k$  values, demonstrating that the parameter selection method does not significantly compromise the model's clean accuracy, regardless of the  $k$  chosen. These results indicate that an even smaller value of  $k$  is effective. The reason for this stability lies in the combined power of  $k$  and the number of repeats. The multiple repeats ensure a comprehensive exploration of different parameter combinations. In each subsequent repeat, another set of parameters is randomly chosen for the  $k$  coalition. This extensive averaging across diverse combinations effectively mitigates the variance that might arise from smaller individual  $k$  values, leading to stable and robust final Shapley importance scores. We also observe the impact of sample size on model interpretation (Appendix C.1).
+
+# 5.3 Qualitative Analysis
+
+We analyze the behavior of different models on adversarial and original samples for the English language in Table 6, where we apply phonetic perturbations to the original claims (evidence remains unchanged), leading to misclassification by the BERT-base model (examples 1-3) and Llama (2-3). It can be seen that when perturbations are applied to claims from the Support (1-2) or Refute (3) classes, the BERT-base model misclassifies them to NEI
+
+class and Llama model misclassifies them to other two classes. However, both adversarial training and our proposed I-Guard correctly classify these adversarial samples, indicating that these models understand phonetic perturbations.
+
+Examples 4 and 5 illustrate the behavior of the models on actual examples. In these examples, the adversarial training-based defense misclassifies claims from the Support and Refute classes as NEI, whereas I-Guard correctly classifies them.
+
+# 6 Conclusion
+
+We proposed  $I$ -Guard, a defense mechanism designed to handle adversarial perturbations in transformer-based models.  $I$ -Guard employs model interpretability to identify the parameters responsible for adversarial misclassifications and applies selective fine-tuning. While the widely used adversarial training-based defense improves robustness, it requires re-training from scratch and can lead to a drop in performance on the actual test sets. Our experiments on English and Hinglish datasets demonstrate that  $I$ -Guard enhances both adversarial and actual test performance, ensuring better generalization while fine-tuning fewer parameters. Additionally, we showed the transferability of our approach in handling other related character perturbations. In future work, we aim to extend  $I$ -Guard to other models to assess its efficiency and robustness in more complex architectures.
+
+# Limitations
+
+This study, like most others, has limitations that could be addressed in future research. Currently, we focus only on English and Hinglish datasets and have not evaluated the performance of  $I$ -Guard on other low-resource language pairs. Expanding our approach to additional languages, particularly those
+
+with different scripts, would be a valuable direction for future work. Additionally, our evaluation primarily considers text-based adversarial attacks, such as phonetic and LEET perturbations. However, adversarial robustness in multimodal settings (e.g., text with images) remains unexplored. Future work could investigate I-Guard's robustness in multimodal scenarios. We also plan to integrate the selection of t into an automated hyperparameter optimization framework (e.g., Bayesian optimization) to systematically identify the value that maximizes robustness or generalization.
+
+# Ethics Statement
+
+We use publicly accessible datasets for our experiments, strictly for academic purposes and in full accordance with their licensing terms.
+
+# Acknowledgements
+
+This research was supported by EPSRC (grant number EP/X04162X/1).
+
+# References
+
+Mohamad Ballout, Anne Dedert, Nohayr Muhammad Abdelmoneim, Ulf Krumnack, Gunther Heidemann, and Kai-Uwe Kühnberger. 2024. FOOL ME IF YOU CAN! an adversarial dataset to investigate the robustness of LMs in word sense disambiguation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5042-5059, Miami, Florida, USA. Association for Computational Linguistics.  
+Rongzhou Bao, Jiayi Wang, and Hai Zhao. 2021. Defending pre-trained language models from adversarial word substitutions without performance sacrifice. arXiv preprint arXiv:2105.14553.  
+Jiaao Chen, Zichao Yang, and Diyi Yang. 2020. Mix-Text: Linguistically-informed interpolation of hidden space for semi-supervised text classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2147-2157, Online. Association for Computational Linguistics.  
+Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics.  
+Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso.
+
+2023. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36:16318-16352.  
+David Crystal. 2018. The Cambridge encyclopedia of the English language. Cambridge university press.  
+Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022a. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493-8502.  
+Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022b. Knowledge neurons in pretrained transformers. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493-8502, Dublin, Ireland. Association for Computational Linguistics.  
+Fahim Dalvi, Nadir Durrani, Hassan Sajjad, Yonatan Belinkov, Anthony Bau, and James Glass. 2019. What is one grain of sand in the desert? analyzing individual neurons in deep nlp models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6309-6317.  
+Sourya Dipta Das, Ayan Basak, Soumil Mandal, and Dipankar Das. 2022. Advcodemix: Adversarial attack on code-mixed data. In Proceedings of the 5th Joint International Conference on Data Science & Management of Data (9th ACM IKDD CODS and 27th COMAD), CODS-COMAD '22, page 125-129, New York, NY, USA. Association for Computing Machinery.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Thomas Diggelmann, Jordan L. Boyd-Graber, Jannis Bulian, Massimiliano Ciaramita, and Markus Leippold. 2020. CLIMATE-FEVER: A dataset for verification of real-world climate claims. CoRR, abs/2012.00614.  
+Nadir Durrani, Hassan Sajjad, Fahim Dalvi, and Yonatan Belinkov. 2020. Analyzing individual neurons in pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4865-4880, Online. Association for Computational Linguistics.  
+Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484-5495.
+
+Karan Goel, Nazneen Fatema Rajani, Jesse Vig, Zachary Taschdjian, Mohit Bansal, and Christopher Ré. 2021. Robustness gym: Unifying the NLP evaluation landscape. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Demonstrations, pages 42-55, Online. Association for Computational Linguistics.  
+Rhys Gould, Euan Ong, George Ogden, and Arthur Conmy. 2024. Successor heads: Recurring, interpretable attention heads in the wild. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net.  
+Shreya Goyal, Sumanth Doddapaneni, Mitesh M. Khapra, and Balaraman Ravindran. 2023a. A survey of adversarial defenses and robustness in NLP. ACM Comput. Surv., 55(14s):332:1-332:39.  
+Shreya Goyal, Sumanth Doddapaneni, Mitesh M. Khapra, and Balaraman Ravindran. 2023b. A survey of adversarial defenses and robustness in nlp. ACM Comput. Surv., 55(14s).  
+Shreya Goyal, Sumanth Doddapaneni, Mitesh M Khapra, and Balaraman Ravindran. 2023c. A survey of adversarial defenses and robustness in nlp. ACM Computing Surveys, 55(14s):1-39.  
+Ashim Gupta, Rishanth Rajendhran, Nathan Stringham, Vivek Srikumar, and Ana Marasovic. 2024. Whispers of doubt amidst echoes of triumph in NLP robustness. 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 5533-5590, Mexico City, Mexico. Association for Computational Linguistics.  
+Michael Hanna, Ollie Liu, and Alexandre Variengien. 2024. How does gpt-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model. Advances in Neural Information Processing Systems, 36.  
+John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. 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 4129-4138, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3.  
+Jiabao Ji, Bairu Hou, Zhen Zhang, Guanhua Zhang, Wenqi Fan, Qing Li, Yang Zhang, Gaowen Liu, Sijia Liu, and Shiyu Chang. 2024. Advancing the robustness of large language models through self-denoised smoothing. In Proceedings of the 2024 Conference
+
+of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers), pages 246-257, Mexico City, Mexico. Association for Computational Linguistics.  
+Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020a. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018-8025.  
+Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020b. Is BERT really robust? A strong baseline for natural language attack on text classification and entailment. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 8018-8025. AAAI Press.  
+Aditya Joshi, Ameya Prabhu, Manish Shrivastava, and Vasudeva Varma. 2016. Towards sub-word level compositions for sentiment analysis of Hindi-English code mixed text. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2482-2491, Osaka, Japan. The COLING 2016 Organizing Committee.  
+Mintong Kang, Linyi Li, and Bo Li. 2023. Fashapley: Fast and approximated shapley based model pruning towards certifiably robust dnns. In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 575-592.  
+Douwe Kiela, Max Bartolo, Yixin Nie, Divyansh Kaushik, Atticus Geiger, Zhengxuan Wu, Bertie Vidgen, Grusha Prasad, Amanpreet Singh, Pratik Ringshia, Zhiyi Ma, Tristan Thrush, Sebastian Riedel, Zeerak Waseem, Pontus Stenetorp, Robin Jia, Mohit Bansal, Christopher Potts, and Adina Williams. 2021. Dynabench: Rethinking benchmarking in NLP. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4110-4124, Online. Association for Computational Linguistics.  
+Akshay Kulkarni and Tsui-Wei Weng. 2024. Interpretability-guided test-time adversarial defense. In European Conference on Computer Vision, pages 466-483. Springer.  
+Thai Le, Jooyoung Lee, Kevin Yen, Yifan Hu, and Dongwon Lee. 2022. Perturbations in the wild: Leveraging human-written text perturbations for realistic adversarial attack and defense. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 2953-2965, Dublin, Ireland. Association for Computational Linguistics.
+
+Dianqi Li, Yizhe Zhang, Hao Peng, Liquin Chen, Chris Brockett, Ming-Ting Sun, and Bill Dolan. 2021. Contextualized perturbation for textual adversarial attack. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5053-5069, Online. Association for Computational Linguistics.  
+Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. 2019. Textbugger: Generating adversarial text against real-world applications. In 26th Annual Network and Distributed System Security Symposium, NDSS 2019, San Diego, California, USA, February 24-27, 2019. The Internet Society.  
+Linyang Li and Xipeng Qiu. 2021. Token-aware virtual adversarial training in natural language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8410-8418.  
+Linyang Li, Demin Song, and Xipeng Qiu. 2023. Text adversarial purification as defense against adversarial attacks. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 338-350, Toronto, Canada. Association for Computational Linguistics.  
+Bill Yuchen Lin, Wenyang Gao, Jun Yan, Ryan Moreno, and Xiang Ren. 2021. RockNER: A simple method to create adversarial examples for evaluating the robustness of named entity recognition models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3728-3737, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Nelson F. Liu, Matt Gardner, Yonatan Belinkov, Matthew E. Peters, and Noah A. Smith. 2019. Linguistic knowledge and transferability of contextual representations. 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 1073-1094, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Qin Liu, Rui Zheng, Bao Rong, Jingyi Liu, Zhihua Liu, Zhanzhan Cheng, Liang Qiao, Tao Gui, Qi Zhang, and Xuan-Jing Huang. 2022. Flooding-x: Improving bert's resistance to adversarial attacks via loss-restricted fine-tuning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5634-5644.  
+Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. In Proceedings of the 31st international conference on neural information processing systems, pages 4768-4777.  
+Mamta and Asif Ekbal. 2022. Adversarial sample generation for aspect based sentiment classification. In Findings of the Association for Computational Linguistics: AACL-IJCNLP 2022, pages 478-492, Online only. Association for Computational Linguistics.
+
+Mamta Mamta, Zishan Ahmad, and Asif Ekbal. 2023. Elevating code-mixed text handling through auditory information of words. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15918-15932, Singapore. Association for Computational Linguistics.  
+Mamta Mamta and Oana Cocarascu. 2025. FactEval: Evaluating the robustness of fact verification systems in the era of large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 10647-10660, Albuquerque, New Mexico. Association for Computational Linguistics.  
+Tobias Mayer, Santiago Marro, Elena Cabrio, and Serena Villata. 2020. Generating adversarial examples for topic-dependent argument classification 1. In Computational Models of Argument, pages 33-44. IOS Press.  
+Vivek Miglani, Aobo Yang, Aram Markosyan, Diego Garcia-Olano, and Narine Kokhlikyan. 2023. Using captum to explain generative language models. In Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023), pages 165-173, Singapore. Association for Computational Linguistics.  
+Milad Moradi and Matthias Samwald. 2021. Evaluating the robustness of neural language models to input perturbations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1558-1570, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020a. TextAttack: A framework for adversarial attacks, data augmentation, and adversarial training in NLP. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119-126, Online. Association for Computational Linguistics.  
+John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020b. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119-126.  
+Pavan Kalyan Reddy Neerudu, Subba Oota, Mounika Marreddy, Venkateswara Kagita, and Manish Gupta. 2023. On robustness of finetuned transformer-based NLP models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7180-7195, Singapore. Association for Computational Linguistics.  
+Tuc Van Nguyen and Thai Le. 2024. Adapters mixup: Mixing parameter-efficient adapters to enhance the
+
+adversarial robustness of fine-tuned pre-trained text classifiers. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21183-21203, Miami, Florida, USA. Association for Computational Linguistics.  
+Dang Nguyen Minh and Anh Tuan Luu. 2022. Textual manifold-based defense against natural language adversarial examples. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6612-6625, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.  
+Yao Qiang, Subhrangshu Nandi, Ninareh Mehrabi, Greg Ver Steeg, Anoop Kumar, Anna Rumshisky, and Aram Galstyan. 2024. Prompt perturbation consistency learning for robust language models. In *Findings of the Association for Computational Linguistics: EACL* 2024, pages 1357-1370, St. Julian's, Malta. Association for Computational Linguistics.  
+Vyas Raina, Samson Tan, Volkan Cevher, Aditya Rawal, Sheng Zha, and George Karypis. 2024. Extreme miscalibration and the illusion of adversarial robustness. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 2500-2525. Association for Computational Linguistics.  
+Yankun Ren, Jianbin Lin, Siliang Tang, Jun Zhou, Shuang Yang, Yuan Qi, and Xiang Ren. 2020. Generating natural language adversarial examples on a large scale with generative models. In ECAI 2020 - 24th European Conference on Artificial Intelligence, 29 August-8 September 2020, Santiago de Compostela, Spain, August 29 - September 8, 2020 - Including 10th Conference on Prestigious Applications of Artificial Intelligence (PAIS 2020), volume 325 of Frontiers in Artificial Intelligence and Applications, pages 2156-2163. IOS Press.  
+Ananya B Sai, Tanay Dixit, Dev Yashpal Sheth, Sreyas Mohan, and Mitesh M Khapra. 2021. Perturbation checklists for evaluating nlg evaluation metrics. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7219-7234.  
+Chenglei Si, Zhengyan Zhang, Fanchao Qi, Zhiyuan Liu, Yasheng Wang, Qun Liu, and Maosong Sun. 2020. Better robustness by more coverage: Adversarial training with mixup augmentation for robust fine-tuning. CoRR, abs/2012.15699.
+
+Mehmet Sofi, Matteo Fortier, and Oana Cocarascu. 2022a. A robustness evaluation framework for argument mining. In Proceedings of the 9th Workshop on Argument Mining, pages 171-180, Online and in Gyeongju, Republic of Korea. International Conference on Computational Linguistics.  
+Mehmet Sofi, Matteo Fortier, and Oana Cocarascu. 2022b. A robustness evaluation framework for argument mining. In Proceedings of the 9th Workshop on Argument Mining, pages 171-180.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971.  
+Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. 2019. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797-5808, Florence, Italy. Association for Computational Linguistics.  
+Tianlu Wang, Xuezhi Wang, Yao Qin, Ben Packer, Kang Li, Jilin Chen, Alex Beutel, and Ed Chi. 2020. CATgen: Improving robustness in NLP models via controlled adversarial text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5141-5146, Online. Association for Computational Linguistics.  
+Wenqi Wang, Run Wang, Lina Wang, Zhibo Wang, and Aoshuang Ye. 2021a. Towards a robust deep neural network against adversarial texts: A survey. *ieee transactions on knowledge and data engineering*, 35(3):3159-3179.  
+Xiao Wang, Qin Liu, Tao Gui, Qi Zhang, Yicheng Zou, Xin Zhou, Jiacheng Ye, Yongxin Zhang, Rui Zheng, Zexiong Pang, et al. 2021b. Textflint: Unified multilingual robustness evaluation toolkit for natural language processing. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations, pages 347-355.  
+Xiaosen Wang, Jin Hao, Yichen Yang, and Kun He. 2021c. Natural language adversarial defense through synonym encoding. In Uncertainty in Artificial Intelligence, pages 823-833. PMLR.  
+Xiaozhi Wang, Kaiyue Wen, Zhengyan Zhang, Lei Hou, Zhiyuan Liu, and Juanzi Li. 2022a. Finding skill neurons in pre-trained transformer-based language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11132-11152, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+Xuezhi Wang, Haohan Wang, and Diyi Yang. 2022b. Measure and improve robustness in NLP models: A survey. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4569-4586, Seattle, United States. Association for Computational Linguistics.
+
+Yang Wang and Chenghua Lin. 2025. Tougher text, smarter models: Raising the bar for adversarial defence benchmarks. In Proceedings of the 31st International Conference on Computational Linguistics, pages 6475-6491, Abu Dhabi, UAE. Association for Computational Linguistics.
+
+A Waswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. 2017. Attention is all you need. In NIPS.
+
+T Wolf. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.
+
+Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, and Daniel S Weld. 2021. Polyjuice: Generating counterfactuals for explaining, evaluating, and improving models. 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 6707-6723.
+
+Yahan Yang, Soham Dan, Dan Roth, and Insup Lee. 2023. In and out-of-domain text adversarial robustness via label smoothing. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 657-669.
+
+KiYoon Yoo, Jangho Kim, Jiho Jang, and Nojun Kwak. 2022. Detection of adversarial examples in text classification: Benchmark and baseline via robust density estimation. In *Findings of the Association for Computational Linguistics: ACL* 2022, pages 3656-3672.
+
+Zeping Yu and Sophia Ananiadou. 2024. Neuron-level knowledge attribution in large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 3267-3280, Miami, Florida, USA. Association for Computational Linguistics.
+
+Lifan Yuan, Yangyi Chen, Ganqu Cui, Hongcheng Gao, Fangyuan Zou, Xingyi Cheng, Heng Ji, Zhiyuan Liu, and Maosong Sun. 2023. Revisiting out-of-distribution robustness in NLP: benchmarks, analysis, and llms evaluations. 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.
+
+Shilong Yuan, Wei Yuan, Hongzhi Yin, and Tieke He. 2024. Roic-dm: Robust text inference and classification via diffusion model. arXiv preprint arXiv:2401.03514.
+
+Jiehang Zeng, Jianhan Xu, Xiaoqing Zheng, and Xuanjing Huang. 2023. Certified robustness to text adversarial attacks by randomized [MASK]. Computational Linguistics, 49(2):395-427.
+
+# A Experimental Setup
+
+We use PyTorch and HuggingFace (Wolf, 2019) to implement all models. We focus on transformer-based fine-tuned classification models due to their success in NLP tasks. The BERT-base model consists of 12 transformer layers, a hidden dimension of 768, and 12 self-attention heads, totaling 110 million trainable parameters.
+
+I-Guard has access to 200 misclassified samples and their corresponding actual examples to calculate the contribution of each weight parameter. We set  $k$  (weights in each iteration) to  $25\%$  and the number of iterations  $K$  to 10. The model was optimized using Adam, with weight updates guided by categorical cross-entropy loss. All experiments were conducted on an NVIDIA A100-SXM4 GPU with 40 GB of memory.
+
+# B Adversarial Samples
+
+Table presents a few adversarial examples for phonetic and LEET based perturbations.
+
+# C Ablation Study
+
+# C.1 Sample size for model interpretation
+
+To observe the impact of sample size on model interpretation, we consider all misclassified samples and their corresponding actual samples for Shapley value calculations instead of randomly choosing 200 misclassified samples. Results are presented in Table 9. We observe that both variants exhibit similar performance with minor variations, demonstrating that  $I$ -Guard can effectively determine the importance of parameters using only a subset of the data rather than all misclassified samples.
+
+# C.2 Fine-tuning Last Layer
+
+We fine-tune the last layer of the model on (i) adversarial samples and (ii) a mixture of adversarial and original data and compare against  $I$ -Guard and  $I$ -Guard-adv. Results for the sentiment analysis task under LEET-based perturbations are reported in Table 10.
+
+We observe similar phenomena in this case. The results show that fine-tuning the last layer on adversarial data yields better performance on the adversarial test set compared to I-Guard-adv, but this
+
+<table><tr><td></td><td>Original Claim</td><td>Phonetic Perturbations</td><td>LEET Perturbations</td></tr><tr><td>1</td><td>Human activities (mainly greenhouse-gas emissions) are the dominant cause of the rapid warming since the middle 1900s (ipcc, 2013).</td><td>Humaaan activities (mainly greenhouse-gas emissions) areeee the dominant cause of the rapid warming since the middle 1900s (ipcc, 2013).</td><td>Hvm4n act1vitles (mainly greenhouse-gas emissions) ar3 the dominant cause of the rapid warming since the middle 1900s (IPCC, 2013).</td></tr><tr><td>2</td><td>In the past, warming has never been a threat to life on earth.</td><td>In the past, warming has neveeeeeer been a threat to life on earth.</td><td>In the past, w4rming has n3v3r been a threat to life on earth.</td></tr><tr><td>3</td><td>Sea level rise due to global warming is exaggerated.</td><td>Seaa level risee due to global warming is exaggerated.</td><td>Sea level r1se due to gl0bal warmlng is exagg3rated</td></tr><tr><td>4</td><td>Clouds provide negative feedback.</td><td>Clouds provide negative feedback.</td><td>Clouds pr0vide n3gat1ve feedback.</td></tr></table>
+
+Table 8: Examples of adversarial inputs generated using phonetic and LEET perturbations  
+
+<table><tr><td></td><td colspan="5">Phonetic</td><td colspan="4">LEET</td></tr><tr><td></td><td>S</td><td colspan="2">Original TS</td><td colspan="2">Adv TS</td><td colspan="2">Original TS</td><td colspan="2">Adv TS</td></tr><tr><td></td><td></td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td rowspan="2">FV</td><td>200</td><td>69.27</td><td>62.40</td><td>69.79</td><td>59.53</td><td>72.13</td><td>63.74</td><td>74.82</td><td>66.82</td></tr><tr><td>All</td><td>70.35</td><td>62.87</td><td>69.27</td><td>59.13</td><td>69.79</td><td>62.28</td><td>73.01</td><td>65.98</td></tr><tr><td rowspan="2">SA</td><td>200</td><td>67.96</td><td>64.94</td><td>56.31</td><td>51.36</td><td>67.78</td><td>64.72</td><td>67.13</td><td>61.20</td></tr><tr><td>All</td><td>67.64</td><td>64.87</td><td>56.70</td><td>51.50</td><td>67.19</td><td>64.69</td><td>67.55</td><td>61.86</td></tr></table>
+
+Table 9: Affect of choosing different sample size for model interpretation. Here S is number of samples.  
+
+<table><tr><td></td><td colspan="2">Original TS</td><td colspan="2">Adv TS</td></tr><tr><td></td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>Last layer</td><td>48.45</td><td>46.15</td><td>73.58</td><td>68.65</td></tr><tr><td>I-Guard-adv</td><td>59.77</td><td>58.17</td><td>71.77</td><td>66.39</td></tr><tr><td>Last layer</td><td>63.56</td><td>60.66</td><td>63.33</td><td>57.16</td></tr><tr><td>I-Guard</td><td>67.78</td><td>64.72</td><td>67.13</td><td>61.2</td></tr></table>
+
+Table 10: Results on LEET perturbations for sentiment analysis by fine-tuning only last layer.  
+
+<table><tr><td></td><td colspan="2">Original TS</td><td colspan="3">Adv TS</td></tr><tr><td></td><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td><td>PT</td></tr><tr><td>mBERT</td><td>90.31</td><td>73.32</td><td>71.62</td><td>46.28</td><td>-</td></tr><tr><td>Adv training</td><td>89.45</td><td>71.78</td><td>90.77</td><td>73.59</td><td>100%</td></tr><tr><td>I-Guard-adv</td><td>67.73</td><td>59.77</td><td>90.17</td><td>78.46</td><td>7.20%</td></tr><tr><td>I-Guard</td><td>90.2</td><td>73.08</td><td>90.82</td><td>75.96</td><td>7.20%</td></tr></table>
+
+Table 11: Results on HSOL dataset for phonetic perturbations.
+
+comes at the cost of large drop in accuracy on the original test set. In contrast, I-Guard-adv maintains a better balance between performance on both the original and adversarial test sets.
+
+Additionally, when fine-tuning the last layer on a mix of original and adversarial data, we observe that  $I$ -Guard consistently outperforms this last layer
+
+fine-tuning approach on both actual and adversarial test set.
+
+# D Results on Large Dataset
+
+To demonstrate the effectiveness of our proposed approach, we conducted experiments on the Hate Speech and Offensive Language (HSOL) dataset. The HSOL dataset consists of tweets categorized into three classes: hate speech, offensive but not hate speech, and neither offensive nor hate speech.
+
+It contains 24,783 tweets, which are split into training (18,587), validation (2,478), and test (3,718) sets. Results for the phonetic perturbation-based adversarial attack are presented in Table 11. We observe that both accuracy and  $F_{1}$  scores drop significantly under this attack. Further, adversarial training-based defense improves performance on the adversarial test set by fine-tuning  $100\%$  of the parameters on a mixture of actual and adversarial data. Our proposed approach effectively defends against this attack by fine-tuning only  $7.20\%$  parameters. It outperforms the adversarial training based defense on both the actual and adversarial test set. In addition, we compute Shapley values using a subset of the data (i.e., 200 misclassified
+
+samples and their corresponding actual examples). This illustrates that  $I$ -Guard can enhance adversarial robustness by fine-tuning fewer parameters compared to adversarial training-based defenses, even in the case of larger datasets.

2025/I-GUARD_ Interpretability-Guided Parameter Optimization for Adversarial Defense/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aaded7f245b686c3666ea0bf834bdb2d1dc0addf1c0e47356c35d2b20faaf2ae
+size 568179

2025/ICL-Bandit_ Relevance Labeling in Advertisement Recommendation Systems via LLM/deeabab9-126e-40e2-a116-c0a839b8d3f0_content_list.json

@@ -0,0 +1,2038 @@
+[
+    {
+        "type": "text",
+        "text": "ICL-Bandit: Relevance Labeling in Advertisement Recommendation Systems via LLM",
+        "text_level": 1,
+        "bbox": [
+            142,
+            89,
+            855,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Lu Wang $^{1}$ , Chiming Duan $^{2*}$ , Pu Zhao $^{1}$ , Fangkai Yang $^{1}$ , Yong Shi $^{1}$ , Xuefeng Luo $^{1}$ , Bingjing Xu $^{1}$ , Weiwei Deng $^{1}$ , Qingwei Lin $^{1}$ , Dongmei Zhang $^{1}$ $^{1}$ Microsoft,  $^{2}$ Peking University",
+        "bbox": [
+            166,
+            156,
+            823,
+            209
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            342,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Measuring the relevance between user queries and advertisements is a critical task for advertisement (ad) recommendation systems, such as Microsoft Bing Ads and Google Ads. Traditionally, this requires expert data labeling, which is both costly and time-consuming. Recent advances have explored using Large Language Models (LLMs) for labeling, but these models often lack domain-specific knowledge. In-context learning (ICL), which involves providing a few demonstrations, is a common practice to enhance LLM performance on domain-specific tasks. However, retrieving high-quality demonstrations in a vast exploration space remains challenging. In this paper, we introduce ICL-Bandit, a practical and effective approach that leverages ICL to enhance the query-ad relevance labeling capabilities of LLMs. We develop a novel bandit learning method to identify and provide superior demonstrations for ICL, thereby improving labeling performance. Experimental results demonstrate that ICL-Bandit achieves state-of-the-art performance compared to existing methods. Additionally, ICL-Bandit has been deployed in Microsoft that serves billions of users worldwide, confirming its robustness and effectiveness.",
+        "bbox": [
+            144,
+            288,
+            460,
+            671
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            684,
+            260,
+            700
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In advertisement (ad) recommendation systems such as Microsoft Bing Ads and Google Ads, high-quality labeled data is of critical importance for training ad recommendation models, especially labeling the relevance between user query text and ad description text, as discussed in (Ling et al., 2017; Shuai et al., 2020; Wang et al., 2022a). The traditional approach is human labeling which is costly and inefficient. This is particularly challenging given the huge amount of data to be labelled, and labeling such relevance between user query and ad",
+        "bbox": [
+            112,
+            711,
+            489,
+            888
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "requires a good knowledge and experience. For example,",
+        "bbox": [
+            507,
+            261,
+            882,
+            293
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "User Query:",
+        "text_level": 1,
+        "bbox": [
+            536,
+            313,
+            638,
+            328
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "\"Innovative treatments for reducing hospital readmission rates in heart failure patients.\"",
+        "bbox": [
+            536,
+            329,
+            857,
+            362
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Advertisement:",
+        "text_level": 1,
+        "bbox": [
+            536,
+            368,
+            660,
+            382
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "\"Remote Patient Monitoring Systems - Continuous Care for Heart Failure Patients\"",
+        "bbox": [
+            536,
+            384,
+            858,
+            416
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This query-ad pair is labeled as relevant since remote patient monitoring systems provide continuous care and real-time health data, enabling proactive management of heart failure, which is essential for reducing hospital readmission rates, and such manual labeling requires domain knowledge.",
+        "bbox": [
+            507,
+            436,
+            884,
+            532
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in Large Language Models (LLMs) have shown that LLMs are highly aligned with human judgments and even surpass human performance in certain tasks (Ouyang et al., 2022), such as topic identification and twitter relevance for political issues (Gilardi et al., 2023), general question-answering data generation (Meng et al., 2023), and instruction data generation (Wang et al., 2022b). However, the lack of domain knowledge limits the performance of LLM in the query-ad relevance labeling task. To address this challenge, many approaches employed in-context learning (ICL) to incorporate domain-specific knowledge as extra context in the LLM's prompt (Kossen et al., 2023; Dong et al., 2022). Besides, it is well known that the effectiveness of ICL heavily depends on the quality of the provided demonstrations, which has motivated many works to explore effective demonstration retrieval methods for ICL, such as (Rubin et al., 2021; Li et al., 2023; Wu et al.; Zhang et al., 2022), all of these methods aim to retrieve better examples from annotated training sets to enhance LLMs' domain knowledge.",
+        "bbox": [
+            507,
+            533,
+            884,
+            903
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Previous work on demonstration retrieval falls",
+        "bbox": [
+            527,
+            904,
+            882,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "23439",
+        "bbox": [
+            473,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 23439-23449 November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            208,
+            945,
+            788,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "into two categories. One category involves off-the-shelf retrievers like BM25 (Robertson et al., 2009) or KNN (Guo et al., 2003), which can retrieve textually or semantically similar demonstrations. The other category focuses on training task-specific retrievers with positive and negative demonstrations. Notable examples include Rubin et al. (Rubin et al., 2021), Shi et al. (Shi et al., 2022), and Xiaonan et al. (Li et al., 2023), who leverage LLM feedback (compare the labels generated by LLM with the ground-truth labels, using them as the training signal) to train these retrievers via supervised or contrastive learning. However, the vast combination space of different demonstrations and queries poses a challenge. Randomly sampling demonstrations to collect the LLM's feedback may lead to large parts of \"less useful\" examples. Some methods, like Zhang et al. (Zhang et al., 2022) and Mingkai et al. (Deng et al., 2022), employ reinforcement learning to actively sample demonstrations and obtain LLM feedback. But these methods are limited in considering only a fixed number of candidate demonstrations, which reduces the action space for policy training.",
+        "bbox": [
+            115,
+            85,
+            489,
+            469
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To overcome the challenges addressed above in SOTA methods, at first, we frame demonstration retrieval problem as a multi-armed bandit (MAB) problem (Lai and Robbins, 1985), and bandit algorithms solving MAB problems (Vermorel and Mohri, 2005; Li et al., 2010) have demonstrated excellent performance in addressing exploration and exploitation dilemma when dealing with largescale search spaces. This allows us to design effective exploration techniques for sampling demonstrations and obtaining LLM feedback during retriever training. Then, we propose a novel in-context learning (ICL) algorithm, called ICL-Bandit, which leverages a stochastic bandit algorithm to empower ICL at scale with diverse demonstration pools. The objective of ICL-Bandit is to retrieve demonstrations and maximize cumulative positive LLM feedback over a series of retrievals. Figure 1 shows a comparison on the example query-ad pair with demonstrations retrieved with KNN and our ICL-Bandit, respectively.",
+        "bbox": [
+            115,
+            470,
+            489,
+            807
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions can be summarized as follows:",
+        "bbox": [
+            132,
+            809,
+            487,
+            822
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We formulate demonstration retrieval as a multi-armed bandit problem, focusing on effective retrieval during retriever training.",
+            "- We design a stochastic bandit algorithm suitable for ICL with a large and varied demon"
+        ],
+        "bbox": [
+            134,
+            831,
+            489,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "stration pool.",
+        "bbox": [
+            544,
+            84,
+            648,
+            99
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- Our approach achieves SOTA performance comparing to other existing methods and it has been deployed to the labeling process at Microsoft, resulting in substantial cost savings by automated labeling.",
+        "bbox": [
+            531,
+            112,
+            882,
+            193
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 ICL-Bandit with large and varied demonstration pool",
+        "text_level": 1,
+        "bbox": [
+            507,
+            219,
+            835,
+            253
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce ICL-Bandit, a stochastic bandit algorithm designed to efficiently retrieve demonstrations and collect LLM feedback during retriever training. The task is to precisely label query-ad relevance with ICL, and our goal is to train a demonstration retriever which retrieves good demonstrations for ICL. When training the retriever using LLM feedback (we compare the output labels generated by the LLM with the ground-truth labels, employing them as the reward signal), the key lies in how to effectively retrieve demonstrations from a large and diverse pool during the training process. Addressing the Exploration (searching for diverse and potentially informative demonstrations) versus Exploitation (retrieving high-reward demonstrations) balance is pivotal. To tackle this challenge, we first formulate the demonstration retrieval task as a bandit problem.",
+        "bbox": [
+            507,
+            263,
+            884,
+            552
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Task Definition",
+        "text_level": 1,
+        "bbox": [
+            507,
+            565,
+            675,
+            581
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The task is to label the relevance of query-ad pairs leveraging LLMs. Compared with zero-shot LLM labeling, providing with demonstrations as context in the ICL manner improves the labeling performance. The ICL prompt comprises four key components:",
+        "bbox": [
+            507,
+            588,
+            882,
+            683
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Instruction We employ the following instruction to describe labeling requests: \"Given user query and an ad, assign a label based on following definitions: - 'Relevant': The ad content directly addresses the user's query, providing information or a solution that aligns with the search intent. - 'Irrelevant': The ad content does not address the user's query, failing to provide information or a solution that matches the search intent.\" This instruction guides the relevance labeling process for LLMs.",
+            "- Input The input component specifies the query and ad requiring labeling. For instance, \"User"
+        ],
+        "bbox": [
+            507,
+            699,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "23440",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d91cfd714ccf3e822c2e8d88f9fe236f6f973226d4a28863044fcb10afa9ca8.jpg",
+        "image_caption": [
+            "Figure 1: An example illustrating how LLM labeler prompting, when combined with demonstrations retrieved via KNN and our ICL-Bandit approach, yields distinct labeling outcomes."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            82,
+            803,
+            259
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/f8f8fcbfc03dd67d689e151a98becdb91bc43fd54b49efcd1bbb84ff5f2cc66c.jpg",
+        "image_caption": [
+            "Figure 2: An illustration on training demonstration retrieval process of ICL-Bandit. The demonstration pool consists of expert labeled demonstrations, and each demonstration includes user query, advertisement, and label. For each query-ad from the training set, ICL-Bandit retrieves demonstrations from the demonstration pool considering the estimated reward distribution. With the retrieved demonstrations as the context, LLM labels the query-ad as relevant or irrelevant. Then the LLM-generated label is compared with the ground truth label from the training set to give feedback. In the figure, the positive reward (matched label) is used to update ICL-Bandit to refine the retrieval policy."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            318,
+            828,
+            518
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Query: 'Student loans suspended until september'; Advertisement: '10 Best Student Loan Refinance''. This information sets the context for the query-ad labeling task.",
+        "bbox": [
+            126,
+            656,
+            489,
+            720
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Demonstrations Demonstrations consist of a set of labeled demonstrations, such as, \"User Query: 'School registration', Advertisement: 'Find Virtual School Programs', Label: Relevant\". Different retrieved demonstrations would highly affect the labeling performance. These examples, provided by human annotators, serve as training instances for a policy  $\\pi_{\\theta}$  that leverages LLM feedback to retrieve appropriate demonstrations for each unique input.",
+            "- Output Indicator The output indicator instructs"
+        ],
+        "bbox": [
+            114,
+            732,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "the LLM to generate the labeling answer. For instance, \"Return your decision on the label in  $<\\text{Label}>$  {{ $\\text{}/\\text{Label}>$  tags.\"}. This guides the LLM to generate the final label.",
+        "bbox": [
+            519,
+            656,
+            880,
+            719
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            737,
+            724,
+            751
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The key of the labeling task lies in retrieving informative demonstrations for ICL labeling, then we formulate the demonstration retrieval problem as a multi-armed bandit (MAB) problem, drawing an analogy to the scenario of a gambler selecting from a slot machine with multiple arms in a casino. The player's objective is to choose the arm that offers the highest expected gain. Each time the player pulls an arm and receives a gain or not, they update their estimation of the arm's potential",
+        "bbox": [
+            507,
+            760,
+            884,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "23441",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "gain. Similar to this scenario, in the query-ad relevance labeling task, we propose a MAB approach to retrieve demonstrations from a pool to achieve a good performance. We define key components such as states, arms, rewards, and the overall objective. **Trial:** In each trial  $t$ , our goal is to retrieve  $m$  demonstration from the demonstration pool. **State:** State  $s_t$  represents the current contextual environment. In the context of ICL,  $s_t$  denotes the embedding of the query and ads that require labeling. **Arm/Action:** Each arm in the MAB problem corresponds to a potential demonstration, representing different choice or action that can be taken during the ICL labeling. **Reward:** The reward  $r_{(t,a_k)}$  provides numerical feedback, indicating whether the LLM assigns the correct relevance label based on the retrieved demonstration  $a_k$ . We define two reward options: a continuous reward  $r_{(t,a_k)} \\in [0,1]$  (representing the probability of the LLM's output label) and a discrete reward  $r_{(t,a_k)} \\in [0,1]$ . In the discrete case, a correct label receives a reward of 1, while an incorrect label receives 0. **Objective:** The overall goal is to learn a retrieval policy  $\\pi_\\theta$  by maximizing the cumulative reward over a series of trials during training.",
+        "bbox": [
+            110,
+            84,
+            492,
+            487
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The process of using a MAB algorithm to efficiently retrieve demonstrations during training consists of the following steps:",
+        "bbox": [
+            112,
+            487,
+            489,
+            535
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Step 1: Retrieval of Demonstrations At each trial  $t$ , the retrieval policy  $\\pi_{\\theta}$  retrieves a demonstration (arm)  $a_{k}$  from a demonstration pool. The optimal retrieval will be conducted based on a balance between the benefits (the mean of  $a_{k}$ 's rewards) and the chance (the variance of  $a_{k}$ 's rewards).",
+        "bbox": [
+            112,
+            535,
+            489,
+            631
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Step 2: Reward from LLM Feedback After retrieving a demonstration  $a_{k}$ , the retriever receives a reward  $r_t$ , indicating the correctness of the label provided by the LLM with the chosen demonstration.",
+        "bbox": [
+            112,
+            631,
+            489,
+            709
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Step 3: Policy Update The collected reward value is used to update the policy parameters  $\\pi_{\\theta}$  to maximize cumulative reward during training.",
+        "bbox": [
+            112,
+            712,
+            489,
+            760
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "These steps iteratively occur for a pre-defined number of iterations ( $T = 2000$  in our paper). The exploration-exploitation trade-off in Step 1 is crucial, requiring the retrieval policy to balance exploring new demonstrations for potential benefits (exploration) and exploiting known good demonstrations to maximize the mean of rewards (exploitation). This exploration-exploitation balance ensures the effectiveness of learning the demonstration retrieval policy in ICL labeling.",
+        "bbox": [
+            112,
+            760,
+            490,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2.3 ICL-Bandit",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            650,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce ICL-Bandit as an innovative approach to address the challenges encountered by previous bandit algorithms. Traditional bandit algorithms, which assign a parameter  $\\theta$  to each arm or action, facing the limitations when applied to demonstration retrieval due to the expansive and varied nature of the available demonstrations.",
+        "bbox": [
+            507,
+            105,
+            884,
+            233
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We leverage the framework of Stochastic Multi-Armed Bandit (Bubeck et al., 2012), a variant of the classical MAB problem where rewards associated with different actions (referred to as \"arms\") are influenced by stochastic processes. In our context, we develop a novel stochastic bandit algorithm tailored to scenarios with an extensive and diverse set of demonstrations.",
+        "bbox": [
+            507,
+            234,
+            884,
+            362
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "First, we fine-tune a BERT model in Microsoft's query and ads dataset, resulting in an embedding vector  $e_{a_k}$  unique to each demonstration  $a_k$ , an embedding vector  $e_{s_t}$  for the state  $s_t$  at trial  $t$  and a mutual embedding  $e_{(s,a)_t}$ . Then we have a unified feature embedding  $x_{s_t,a_k} = [e_{s_t}, e_{a_k}, e_{(s,a)_t}]$  to capture contextual information. Next, we adopt a shared parameter  $\\theta$  applicable to all demonstrations. This shared parameterization streamlines the learning process, enhancing efficiency and generalization across the diverse pool of demonstrations.",
+        "bbox": [
+            507,
+            363,
+            885,
+            541
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2.3.1 Demonstration and State Representation Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            550,
+            769,
+            583
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To integrate both the state and demonstration into a unified embedding vector, we employ a self-supervised learning approach to fine-tune a 24-layer BERT model using Microsoft's user query-ad dataset. The final embedding is derived by extracting the output of the last hidden layer, serving as a comprehensive representation of both the state and demonstration.",
+        "bbox": [
+            507,
+            586,
+            884,
+            715
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2.3.2 Objective Function of ICL-Bandit",
+        "text_level": 1,
+        "bbox": [
+            507,
+            725,
+            836,
+            740
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Throughout the total  $T$  trials, the cumulative reward is defined as  $\\sum_{t=1}^{T} r_{(t,a_k)}$ . In this context, we establish the optimal expected  $T$ -iteration reward, denoted as  $E[\\sum_{t=1}^{T} r_{(t,a_k^*)}]$ , where  $a_k^*$  represents the optimal demonstration yielding the maximum expected reward at trial  $t$ . Our objective is to proficiently retrieve a sequence of demonstrations during training, maximizing the expected total payoff. Alternatively, our aim is to minimize the regret of the algorithm concerning the optimal demonstration retrieval strategy. The  $T$ -iteration regret of",
+        "bbox": [
+            507,
+            744,
+            884,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "23442",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ICL-Bandit can be formally defined as:",
+        "bbox": [
+            114,
+            84,
+            406,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR e (T) = E [ \\sum_ {t = 1} ^ {T} r _ {(t, a _ {k} ^ {*})} ] - E [ \\sum_ {t = 1} ^ {T} r _ {(t, a _ {k})} ] \\quad (1)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            139,
+            126,
+            487,
+            168
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2.3.3 Optimization of ICL-Bandit",
+        "text_level": 1,
+        "bbox": [
+            114,
+            180,
+            394,
+            196
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To minimize the regret, it is assumed that the expected reward of an example  $a$  is linear in the  $d$ -dimensional state-action integrated feature  $x_{(t,a)}$ , with three unknown policy parameters  $\\theta_{\\mathrm{state}}^*$ ,  $\\theta_{\\mathrm{action}}^*$  and  $\\theta_{\\mathrm{mut}}^*$ :",
+        "bbox": [
+            112,
+            200,
+            489,
+            281
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {E} [ r _ {(t, a)} \\mid x _ {(t, a)} ] = x _ {(t, a)} ^ {T} [ \\theta_ {\\mathrm {s t a t e}} ^ {*}, \\theta_ {\\mathrm {a c t i o n}} ^ {*}, \\theta_ {\\mathrm {m u t}} ^ {*} ]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            142,
+            292,
+            458,
+            315
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\theta_{\\mathrm{state}}^{*}$  denotes the policy parameter for the context or the state, i.e., for the target sample,  $\\theta_{\\mathrm{action}}^{*}$  denotes the policy parameter for an action, i.e., for an example, and  $\\theta_{\\mathrm{mut}}^{*}$  denotes the policy parameter for mutual information of the target sample and the example. The mutual information may be common or similar information between the target sample and the example. The policy parameter  $\\theta_{\\mathrm{state}}^{*}$  can map the target sample to a first vector space. The policy parameter  $\\theta_{\\mathrm{action}}^{*}$  can map the example to a second vector space. The first vector space and the second vector space are different and independent, but they are dual to each other. The policy parameter  $\\theta_{\\mathrm{mut}}^{*}$  can map both the target sample and the example to the same vector space. The technical effect of using the three policy parameters  $\\theta_{\\mathrm{state}}^{*}$ ,  $\\theta_{\\mathrm{action}}^{*}$ , and  $\\theta_{\\mathrm{mut}}^{*}$  is to more accurately measure the relationship or distance between the target sample and the example, so as to calculate a more accurate expected reward.",
+        "bbox": [
+            112,
+            324,
+            487,
+            645
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The embodiments of the present disclosure propose that all examples share three policy parameters  $\\theta_{\\mathrm{state}}^{*}$ ,  $\\theta_{\\mathrm{action}}^{*}$ , and  $\\theta_{\\mathrm{mut}}^{*}$ . This parameterization remains constant regardless of the number of examples. The technical effect of such settings is to streamline the learning process, and enhance efficiency and generalization across the diverse set of examples. This framework enables the application of the proposed reinforced retrieval operation to large-scale and diverse candidate examples, contributing to its scalability and adaptability.",
+        "bbox": [
+            112,
+            646,
+            489,
+            822
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For each example, we have three kinds of features: state, action, and mut, denoted as  $[e_{s_t}, e_{a_k}, e_{(s,a)_t}]$ . These correspond to the data matrices  $D_{\\mathrm{state}}$ ,  $D_{\\mathrm{action}}$ , and  $D_{\\mathrm{mut}}$ , which represent samples on different features. Let  $D_{\\mathrm{state}}$ ,  $D_{\\mathrm{action}}$ , and  $D_{\\mathrm{mut}}$  be data matrices of dimension  $m \\times d$  at",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "trial  $t$ , where the rows correspond to  $m$  training inputs of context, action, and mutual information, and  $b \\in \\mathbb{R}^m$  is the corresponding reward vector (e.g., the  $m$  rewards indicating whether the LLM provided the correct label in the training set). Applying ridge regression to the training data  $(D, b)$  yields an estimate of the policy parameters:",
+        "bbox": [
+            507,
+            84,
+            882,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {\\text {s t a t e}} = \\left(D _ {\\text {s t a t e}} ^ {T} D _ {\\text {s t a t e}} + \\lambda I\\right) ^ {- 1} D _ {\\text {s t a t e}} ^ {T} b \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            557,
+            212,
+            882,
+            234
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {\\text {a c t i o n}} = \\left(D _ {\\text {a c t i o n}} ^ {T} D _ {\\text {a c t i o n}} + \\lambda I\\right) ^ {- 1} D _ {\\text {a c t i o n}} ^ {T} b \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            531,
+            256,
+            882,
+            278
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {\\mathrm {m u t}} = \\left(D _ {\\mathrm {m u t}} ^ {T} D _ {\\mathrm {m u t}} + \\lambda I\\right) ^ {- 1} D _ {\\mathrm {m u t}} ^ {T} b \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            564,
+            294,
+            882,
+            316
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $I$  is the  $d\\times d$  identity matrix,  $\\lambda \\in [0,1]$  is the regularization term of ridge regression estimation. Let  $D = [D_{\\mathrm{state}},D_{\\mathrm{action}},D_{\\mathrm{mut}}]$ , and  $\\theta = [\\theta_{\\mathrm{state}},\\theta_{\\mathrm{action}},\\theta_{\\mathrm{mut}}]$ . When components in  $b$  are independent conditioned on corresponding rows in  $D$ , it can be shown that, with probability at least  $1 - \\delta$ :  $\\left|x_{(t,a_t)}^T\\hat{\\theta} - \\mathbb{E}[r_{(t,a_t)}\\mid x_{(t,a_t)}]\\right| \\leq \\alpha' \\sqrt{x_{(t,a_t)}^T(D^TD + \\lambda I)^{-1}x_{(t,a_t)}}$  For any  $\\delta > 0$  and  $x_{(t,a_t)} \\in \\mathbb{R}^d$ , where  $\\hat{\\theta}$  is the mean of  $\\theta$ , and  $a_t$  indicates the example selected at  $t$ ,  $r_{(t,a)}$  is the observed reward,  $\\sigma^2$  is the variance proxy of the noise and  $\\alpha'$  is a constant.",
+        "bbox": [
+            507,
+            329,
+            882,
+            539
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2.4 Proof",
+        "text_level": 1,
+        "bbox": [
+            509,
+            557,
+            601,
+            571
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2.4.1 Step 1: Decompose the Estimation Error",
+        "text_level": 1,
+        "bbox": [
+            507,
+            581,
+            880,
+            596
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The estimation error can be expressed as:",
+        "bbox": [
+            509,
+            604,
+            816,
+            619
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\hat {\\theta} - \\theta^ {*} = \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} b - \\theta^ {*} \\\\ = \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} \\left(D \\theta^ {*} + \\epsilon\\right) - \\theta^ {*} \\\\ = \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} D \\theta^ {*} + \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} \\epsilon - \\theta^ {*} \\\\ = \\left[ \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} D - I \\right] \\theta^ {*} + \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} \\epsilon . \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            650,
+            868,
+            734
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Simplifying:",
+        "bbox": [
+            526,
+            753,
+            625,
+            769
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\theta} - \\theta^ {*} = - \\lambda \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} \\theta^ {*} + \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} \\epsilon .\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            784,
+            868,
+            807
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2.4.2 Step 2: Express the Estimation Error Components",
+        "text_level": 1,
+        "bbox": [
+            507,
+            824,
+            860,
+            856
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Let:  $\\mathrm{Bias} = -\\lambda (D^{\\top}D + \\lambda I)^{-1}\\theta^{*}$",
+        "bbox": [
+            509,
+            860,
+            781,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Variance  $= \\left(D^{\\top}D + \\lambda I\\right)^{-1}D^{\\top}\\epsilon$",
+        "bbox": [
+            527,
+            882,
+            784,
+            902
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Then:  $\\hat{\\theta} -\\theta^{*} = \\mathrm{Bias} + \\mathrm{Variance}$",
+        "bbox": [
+            527,
+            904,
+            778,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "23443",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2.4.3 Step 3: Bound the Bias Term",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            401,
+            99
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We aim to bound  $\\left|x_{(t,a)}^{\\top} \\text{Bias}\\right|$ .",
+        "bbox": [
+            112,
+            101,
+            337,
+            124
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Using the Cauchy-Schwarz inequality:",
+        "bbox": [
+            132,
+            124,
+            420,
+            140
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\left| x _ {(t, a)} ^ {\\top} \\operatorname {B i a s} \\right| = \\lambda \\left| x _ {(t, a)} ^ {\\top} \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} \\theta^ {*} \\right| (5) \\\\ \\leq \\lambda \\left\\| x _ {(t, a)} \\right\\| _ {\\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1}} \\| \\theta^ {*} \\|, (6) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            159,
+            487,
+            221
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\| x\\| _A = \\sqrt{x^\\top Ax}$  denotes the Mahalanobis norm with respect to the matrix  $A$",
+        "bbox": [
+            112,
+            227,
+            487,
+            260
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Assuming  $\\| \\theta^{*}\\| \\leq S$ , where  $S$  is a known bound on the norm of  $\\theta^{*}$ , we have:",
+        "bbox": [
+            112,
+            261,
+            485,
+            292
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left| x _ {(t, a)} ^ {\\top} \\mathbf {B i a s} \\right| \\leq \\lambda S \\left\\| x _ {(t, a)} \\right\\| _ {\\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1}}. \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            139,
+            312,
+            487,
+            341
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2.4.4 Step 4: Bound the Variance Term",
+        "text_level": 1,
+        "bbox": [
+            112,
+            347,
+            436,
+            361
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We aim to bound  $\\left|x_{(t,a)}^{\\top}\\right|$  Variance with high probability.",
+        "bbox": [
+            112,
+            363,
+            487,
+            401
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Since  $\\epsilon$  has independent components with zero mean and variance proxy  $\\sigma^2$ , the variance of  $x_{(t,a)}^{\\top}$  Variance is:",
+        "bbox": [
+            112,
+            403,
+            487,
+            454
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\operatorname {V a r} \\left(x _ {(t, a)} ^ {\\top} \\text {V a r i a n c e}\\right) = \\operatorname {V a r} \\left(x _ {(t, a)} ^ {\\top} \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} \\epsilon\\right) \\\\ = \\sigma^ {2} x _ {(t, a)} ^ {\\top} \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} D ^ {\\top} D \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} x _ {(t, a)}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            469,
+            473,
+            521
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Therefore:",
+        "bbox": [
+            132,
+            526,
+            215,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\operatorname {V a r} \\left(x _ {(t, a)} ^ {\\top} \\text {V a r i a n c e}\\right) = \\sigma^ {2} x _ {(t, a)} ^ {\\top} \\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1} \\\\ \\cdot \\left(I - \\lambda (D ^ {\\top} D + \\lambda I) ^ {- 1}\\right) x _ {(t, a)} \\\\ \\leq \\sigma^ {2} \\left\\| x _ {(t, a)} \\right\\| _ {(D ^ {\\top} D + \\lambda I) ^ {- 1}} ^ {2}. \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            557,
+            473,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2.4.5 Step 5: Apply Concentration Inequality",
+        "text_level": 1,
+        "bbox": [
+            112,
+            631,
+            484,
+            646
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Since  $x_{(t,a)}^{\\top}$  Variance is a linear combination of independent sub-Gaussian variables, it is sub-Gaussian with parameter  $\\sigma' = \\sigma\\left\\| x_{(t,a)}\\right\\|_{(D^{\\top}D + \\lambda I)^{-1}}$ .",
+        "bbox": [
+            112,
+            648,
+            487,
+            718
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Using a sub-Gaussian tail bound, for any  $\\delta > 0$ :",
+        "bbox": [
+            132,
+            718,
+            487,
+            733
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP \\left(\\left| x _ {(t, a)} ^ {\\top} \\text {V a r i a n c e} \\right| \\geq \\alpha \\left\\| x _ {(t, a)} \\right\\| _ {\\left(D ^ {\\top} D + \\lambda I\\right) ^ {- 1}}\\right) \\leq \\delta \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            753,
+            487,
+            797
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha = \\sigma \\sqrt{2\\ln\\left(\\frac{1}{\\delta}\\right)}$",
+        "bbox": [
+            112,
+            797,
+            295,
+            821
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Step 6: Combine Bias and Variance Terms",
+        "text_level": 1,
+        "bbox": [
+            112,
+            827,
+            450,
+            841
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The total estimation error is:",
+        "bbox": [
+            112,
+            847,
+            327,
+            860
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left| x _ {(t, a)} ^ {\\top} \\left(\\hat {\\theta} - \\theta^ {*}\\right) \\right| \\leq \\left| x _ {(t, a)} ^ {\\top} \\operatorname {B i a s} \\right| + \\left| x _ {(t, a)} ^ {\\top} \\operatorname {V a r i a n c e} \\right|. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            114,
+            879,
+            487,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2.4.6 Step 7: Final Inequality",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            754,
+            99
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Combine the bounds:",
+        "bbox": [
+            509,
+            104,
+            670,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left| x _ {(t, a)} ^ {\\top} \\left(\\hat {\\theta} - \\theta^ {*}\\right) \\right| \\leq (\\lambda S + \\alpha) \\| x _ {(t, a)} \\| _ {(D ^ {\\top} D + \\lambda I) ^ {- 1}}.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            532,
+            118,
+            875,
+            137
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For sufficiently small  $\\lambda$  and bounded  $\\theta^{*}$ , the bias term can be controlled, and the dominant term becomes the variance term.",
+        "bbox": [
+            507,
+            137,
+            880,
+            183
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Therefore, we can simplify the inequality to:  $\\left|x_{(t,a)}^{\\top}\\left(\\hat{\\theta} -\\theta^{*}\\right)\\right|\\leq \\alpha^{\\prime}\\left\\| x_{(t,a)}\\right\\|_{\\left(D^{\\top}D + \\lambda I\\right)^{-1}},$",
+        "bbox": [
+            514,
+            185,
+            882,
+            225
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\alpha^{\\prime} = \\lambda S + \\alpha$",
+        "bbox": [
+            527,
+            227,
+            684,
+            240
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2.5 ICL-Bandit vs. Traditional Bandit",
+        "text_level": 1,
+        "bbox": [
+            507,
+            253,
+            825,
+            267
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ICL-Bandit improves upon traditional bandit methods by introducing shared parameters,  $\\theta_{\\mathrm{state}}^{*}$ ,  $\\theta_{\\mathrm{action}}^{*}$  and  $\\theta_{\\mathrm{mut}}^{*}$ , to jointly model state, actions, and their interactions. This enables better alignment between context and candidate demonstrations, leading to more accurate action selection.",
+        "bbox": [
+            507,
+            273,
+            882,
+            368
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Unlike traditional methods that treat actions independently, ICL-Bandit captures complex contextual dependencies, enhances generalization, and scales efficiently to high-dimensional data. Its unified framework ensures consistent performance across diverse ICL labeling tasks, mitigating the inconsistency and overfitting often seen in conventional approaches.",
+        "bbox": [
+            507,
+            370,
+            882,
+            498
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3 Experiment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            510,
+            645,
+            526
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dataset: We use a high-quality, expert-labeled dataset collected daily over 1.5 years, consisting of user queries and associated advertisement information (e.g., keywords, titles, descriptions, URLs), each labeled as relevant or irrelevant. The dataset is temporally split into: Example pool: 1,578,728 samples used as demonstrations. Training set: 9,999 query-ad pairs. Test set: 1,986 query-ad pairs. This temporal partitioning simulates real-world deployment, where models are trained on historical data and evaluated on recent, unseen examples.",
+        "bbox": [
+            507,
+            536,
+            882,
+            728
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics: We assess binary classification performance using Accuracy (ACC), F1-score, Precision, and Recall to capture both correctness and balance in predictions. More details of the experimental settings are provided in Appendix B.",
+        "bbox": [
+            507,
+            728,
+            882,
+            809
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.1 Competitors",
+        "text_level": 1,
+        "bbox": [
+            507,
+            820,
+            653,
+            835
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For a fair comparison, all baselines and ICL-Bandit (except \"No Example\" and \"Crowdsourcing\") were provided with 3 positive (Relevant) and 3 negative (Irrelevant) historically labeled samples as demonstrations. The following methods were selected as",
+        "bbox": [
+            507,
+            841,
+            882,
+            920
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "23444",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "our competitors: No Example (Zero-shot Learning), Crowdsourcing, EPR (SL-KNN), EPR (SL-LLM) (Li et al., 2023), Q-learning (Zhang et al., 2022), Static, BM25 (Robertson et al., 2009), Random, KNN (Guo et al., 2003). Details of the baseline methods are provided in Appendix A.",
+        "bbox": [
+            112,
+            84,
+            490,
+            181
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/5661171e8d9dea8bd10c4cb1a962e5a847a732bb3bbecf3efc208bb60ae8f082.jpg",
+        "table_caption": [
+            "Table 1: Results of GPT-3.5 as the backbone LLM."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>53.95</td><td>25.72</td><td>41.76</td><td>18.58</td></tr><tr><td>Crowdsourcing</td><td>67.57</td><td>73.65</td><td>90.09</td><td>62.28</td></tr><tr><td>EPR (SL-KNN)</td><td>55.97</td><td>25.68</td><td>39.22</td><td>19.09</td></tr><tr><td>EPR (SL-LLM)</td><td>57.18</td><td>25.44</td><td>41.55</td><td>18.33</td></tr><tr><td>Q-learning</td><td>58.68</td><td>34.71</td><td>43.78</td><td>28.14</td></tr><tr><td>Static</td><td>56.57</td><td>28.56</td><td>43.40</td><td>21.28</td></tr><tr><td>BM25</td><td>58.84</td><td>28.77</td><td>46.35</td><td>20.86</td></tr><tr><td>Random</td><td>50.73</td><td>36.49</td><td>37.52</td><td>35.52</td></tr><tr><td>KNN</td><td>57.98</td><td>27.60</td><td>44.04</td><td>20.10</td></tr><tr><td>ICL-Bandit (Ours)</td><td>63.76</td><td>61.63</td><td>65.38</td><td>58.29</td></tr></table>",
+        "bbox": [
+            115,
+            217,
+            500,
+            319
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/dd6638c5f4406c191c8b791cb626018b18c7c647564be3697eb54dbd76f118dc.jpg",
+        "table_caption": [
+            "Table 2: Results of GPT-4 as the backbone LLM."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>65.05</td><td>61.45</td><td>64.08</td><td>59.03</td></tr><tr><td>Crowdsourcing</td><td>67.57</td><td>73.65</td><td>90.09</td><td>62.28</td></tr><tr><td>EPR (SL-KNN)</td><td>74.42</td><td>80.70</td><td>73.49</td><td>89.47</td></tr><tr><td>EPR (SL-LLM)</td><td>74.62</td><td>80.70</td><td>72.94</td><td>90.32</td></tr><tr><td>Q-learning</td><td>74.26</td><td>76.25</td><td>78.81</td><td>87.80</td></tr><tr><td>Static</td><td>73.56</td><td>79.71</td><td>71.35</td><td>90.28</td></tr><tr><td>BM25</td><td>73.62</td><td>80.09</td><td>72.94</td><td>88.80</td></tr><tr><td>Random</td><td>73.72</td><td>79.94</td><td>71.97</td><td>89.89</td></tr><tr><td>KNN</td><td>74.47</td><td>80.74</td><td>73.56</td><td>89.48</td></tr><tr><td>ICL-Bandit (Ours)</td><td>80.03</td><td>82.57</td><td>76.91</td><td>89.14</td></tr></table>",
+        "bbox": [
+            115,
+            370,
+            499,
+            470
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.2 Results Analysis",
+        "text_level": 1,
+        "bbox": [
+            114,
+            498,
+            290,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Tables 1 and 2 present the experimental results comparing nine demonstration retrieval methods, including our ICL-Bandit, across two versions of LLMs. The analysis highlights key performance trends. The \"No Example\" baseline performs poorly, while \"Crowdsourcing\" demonstrations achieve the highest accuracy and precision, emphasizing the importance of expert-labeled data. Among automated methods, Q-learning, EPR (SL-LLM), and ICL-Bandit show strong performance, benefiting from LLM feedback. Notably, ICL-Bandit surpasses Q-learning and EPR (SL-LLM) despite using only 2,000 feedback samples compared to their 5,000, due to its lightweight, linear design that requires fewer data.",
+        "bbox": [
+            112,
+            519,
+            489,
+            759
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "EPR (SL-KNN) and EPR (SL-LLM) improve over the \"No Example\" baseline but still lag behind \"Crowdsourcing,\" indicating that retrieval effectiveness depends on technique selection. Similarly, methods like \"Static,\" \"Random,\" \"KNN,\" and \"BM25\" show varied performance, with BM25 performing competitively but still unable to match expert-labeled demonstrations.",
+        "bbox": [
+            112,
+            760,
+            489,
+            888
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "ICL-Bandit consistently delivers superior results, often outperforming or matching \"Crowdsourcing.\"",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Its ability to balance exploration and exploitation allows it to retrieve relevant demonstrations effectively, adapt to diverse queries, and enhance recall, improving overall ICL performance.",
+        "bbox": [
+            507,
+            84,
+            884,
+            149
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.3 Learning Curve of ICL-Bandit",
+        "text_level": 1,
+        "bbox": [
+            507,
+            160,
+            798,
+            174
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The learning curve experiment was devised to examine the evolutionary performance of ICL-Bandit as training data accumulates. The primary objective was to discern how the method's effectiveness scales with an expanding dataset, providing insights into its adaptability and scalability. The experiment's results are depicted in Figure 3, where the x-axis represents training iterations, and the y-axis portrays the cumulative mean and variance of Accuracy, Binary Accuracy, True Negative Rate (TNR), and True Positive Rate (TPR). The learning curve analysis of ICL-Bandit highlights its capacity to dynamically adapt and enhance its performance over successive training iterations. Notably, it illustrates that ICL-Bandit achieves a rapid and stable convergence to a commendable performance level.",
+        "bbox": [
+            507,
+            181,
+            884,
+            437
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, the outcomes suggest that ICL-Bandit exhibits promise for demonstration retrieval in ICL, even when trained on a limited LLM feedback dataset. Remarkably, in comparison to EPR (SL-LLM), which utilized a larger dataset of 5000 feedback instances, ICL-Bandit demonstrates superior performance. The learning curve analysis underscores the efficacy of ICL-Bandit in iteratively improving its performance with an increasing volume of training data. This positions it as a robust and scalable solution for the nuanced task of demonstration retrieval in complex information retrieval scenarios.",
+        "bbox": [
+            507,
+            439,
+            882,
+            646
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.4 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            659,
+            673,
+            674
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.4.1 Number of retrivaled samples",
+        "text_level": 1,
+        "bbox": [
+            507,
+            680,
+            803,
+            696
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/e908ab8dff63237273604922811bf35d497071af25fd0bbecbbded11222e8e49.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Example number</td><td>ACC</td><td>Binary AUC</td><td>TNR</td><td>TPR</td></tr><tr><td>1</td><td>0.7826</td><td>0.8127</td><td>0.8326</td><td>0.6978</td></tr><tr><td>3</td><td>0.8003</td><td>0.8029</td><td>0.8215</td><td>0.7077</td></tr><tr><td>6</td><td>0.8001</td><td>0.8127</td><td>0.8178</td><td>0.7129</td></tr><tr><td>9</td><td>0.7697</td><td>0.7616</td><td>0.7716</td><td>0.7516</td></tr></table>",
+        "bbox": [
+            522,
+            711,
+            870,
+            772
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Performance metrics on different number of selected demonstrations.",
+        "bbox": [
+            507,
+            783,
+            882,
+            810
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this experiment, we evaluate the impact of varying the number of positive and negative demonstrations on model performance. The results, presented in Table 3, indicate that the performance metrics (ACC, Binary AUC, TNR, and TPR) generally improve as the number of positive/negative demonstrations increases from 1 to 3. Specifically,",
+        "bbox": [
+            507,
+            810,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "23445",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d7d5c558b1db9657f34936e0d405f5a6ab665dd22e888e995c898acaebe8143c.jpg",
+        "image_caption": [
+            "Figure 3: The learning curve of ICL-Bandit during 2000 trails training. TNR and TPR indicates the true negative rate and true positive rate respectively."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            149,
+            80,
+            843,
+            221
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the best overall performance is observed when 3 demonstrations are used. Thus we choose 3 as the final number.",
+        "bbox": [
+            112,
+            291,
+            480,
+            337
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "When 9 demonstrations are used, the performance metrics begin to decline, indicating that adding too many demonstrations may lead to diminishing returns or even reduced performance.",
+        "bbox": [
+            112,
+            338,
+            480,
+            404
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "3.4.2 Training Epochs and Reward Types",
+        "text_level": 1,
+        "bbox": [
+            112,
+            414,
+            455,
+            432
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The experiment evaluates the performance of the ICL-bandit approach under two reward settings: continuous and discrete. An epoch is defined as a complete pass through the training data. During each epoch, the ICL-bandit retrieves informative demonstrations, selects the best actions, and updates its retrieval policy based on the rewards received. The results in Figure 4 illustrate how the number of epochs affects performance across various metrics.",
+        "bbox": [
+            112,
+            435,
+            487,
+            594
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We observe that ICL-bandit's performance varies with the number of epochs, with different metrics reaching their optimal levels at different stages. The continuous reward setting, which provides more detailed feedback, achieves peak performance in fewer epochs compared to the discrete reward setting. This suggests that using continuous rewards in practice can reduce training complexity while still delivering strong performance. Finally, we choose continuous reward with 1 epoch for reducing the complexity and promising results.",
+        "bbox": [
+            112,
+            596,
+            489,
+            774
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "3.5 Application in Practice",
+        "text_level": 1,
+        "bbox": [
+            112,
+            787,
+            341,
+            802
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We deployed the ICL-Bandit approach in Microsoft's ad relevance pipeline to reduce manual labeling costs and enhance ad recommendation quality. Each day, the system collects fresh user queries and ads, cleans them using Bing's distributed platform, and applies ICL-Bandit for automated labeling. For each query-ad pair, we retrieve 3 relevant",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/c7a9608267025bd44e4cff50d16a16db196712f3a8eade1cc6edc02102dd581b.jpg",
+        "image_caption": [
+            "Figure 4: Performance on different epochs and reward types."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            487,
+            291,
+            870,
+            445
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "and 3 irrelevant historical examples to construct prompts, which are then labeled using GPT-4. This process generates high-quality labeled data daily for downstream CTR prediction (Lee et al., 2023). Table 4: A/B testing on English datasets with GPT-4 as the backbone LLM.",
+        "bbox": [
+            507,
+            508,
+            882,
+            601
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/00785c36547c7532e2d1e4f7d47949c2daeac7ff95b89f2c199cb886f71d897a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>66.82</td><td>62.18</td><td>63.74</td><td>60.75</td></tr><tr><td>KNN</td><td>76.52</td><td>75.12</td><td>78.72</td><td>71.81</td></tr><tr><td>ICL-Bandit (Ours)</td><td>87.12</td><td>82.28</td><td>86.95</td><td>78.17</td></tr></table>",
+        "bbox": [
+            510,
+            612,
+            895,
+            652
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/661adebf40a989bdf0e8a7559eab2164447c9e5539bf73f9ac345639ce2fd34c.jpg",
+        "table_caption": [
+            "Table 5: A/B testing on non-English dataset with GPT-4 as the backbone LLM."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>63.26</td><td>59.48</td><td>60.17</td><td>58.79</td></tr><tr><td>KNN</td><td>70.67</td><td>75.44</td><td>70.86</td><td>80.81</td></tr><tr><td>ICL-Bandit (Ours)</td><td>80.67</td><td>85.68</td><td>82.57</td><td>89.14</td></tr></table>",
+        "bbox": [
+            510,
+            705,
+            895,
+            744
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Tables 4 and 5, ICL-Bandit consistently outperforms baselines in both English and non-English settings, demonstrating robust improvements in accuracy, F1-score, precision, and recall.",
+        "bbox": [
+            507,
+            760,
+            884,
+            838
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "3.6 Impact on Ad Recommendation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            851,
+            806,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Integrating ICL-Bandit-labeled data into Bing's CTR prediction model led to significant business gains. Offline evaluation on 500K historical query-",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "23446",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "ad pairs showed a  $2.5\\%$  AUC increase and  $1.8\\%$  reduction in Log Loss. In two weeks of online A/B testing with 2 million users, CTR rose by  $3.2\\%$  and conversion rates improved by  $2.7\\%$ . Beyond performance, the automated labeling process reduced manual annotation costs by  $61\\%$ , enabling scalable and cost-effective data processing across millions of queries daily. The results of the A/B testing on English and Non-English datasets are summarized in Table 4 and Table 5.",
+        "bbox": [
+            112,
+            84,
+            489,
+            243
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            258,
+            270,
+            273
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4.1 LLM Labeling",
+        "text_level": 1,
+        "bbox": [
+            112,
+            285,
+            278,
+            300
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Latest studies in LLM have shown that LLM is highly consistent with human judgments and even outperforms humans in many tasks, for example, topic identification and twitter relevance for political issues (Gilardi et al., 2023), general question-answering data generation (Meng et al., 2023), instruction data generation (Wang et al., 2022b) and RL from AI feedback (RLAIF) (Lee et al., 2023). A set of work using LLM for labeling instead of human (Tan et al., 2024; Alaofi et al., 2024; Artemova et al., 2024). In this work, we focus on a domain-specific labeling problem, i.e., query-ad relevance labeling, which requires domain knowledge to guide LLM for labeling.",
+        "bbox": [
+            112,
+            307,
+            489,
+            533
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "4.2 Demonstration Retrieval for In-Context Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            544,
+            473,
+            576
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "LLMs have emerged as a pivotal strategy for addressing tasks specific to particular domains. However, the effectiveness of ICL is intrinsically tied to the quality of the provided demonstrations (Li et al., 2023; Wu et al.; Zhang et al., 2022). Works such as (Rubin et al., 2021; Li et al., 2023; Wu et al.; Zhang et al., 2022) collectively aim to optimize the retrieval of exemplary instances from annotated training sets, thereby enhancing the domain knowledge encapsulated by LLMs.",
+        "bbox": [
+            112,
+            583,
+            489,
+            743
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Existing demonstration retrieval methods are typically categorized into utilization of off-the-shelf retrievers such as BM25 (Robertson et al., 2009) or KNN (Guo et al., 2003), or training task-specific retrievers using positive and negative demonstrations (Rubin et al., 2021; Shi et al., 2022; Li et al., 2023). These researchers leverage LLM feedback to guide the training of these retrievers through supervised or contrastive learning. Despite these advancements, the vast combinatorial space encompassing different demonstrations and queries",
+        "bbox": [
+            112,
+            744,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "presents a significant challenge. Randomly sampling demonstrations to collect LLM feedback risks incorporating a substantial portion of less useful examples. Reinforcement learning-based methods (Zhang et al., 2022; Deng et al., 2022) actively sample demonstrations and elicit valuable LLM feedback. However, they are constrained by a fixed number of demonstrations, thereby limiting the action space available for policy training.",
+        "bbox": [
+            507,
+            84,
+            884,
+            230
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            240,
+            640,
+            255
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we leverage LLMs to automate query-ad relevance labeling for improved ad recommendation. To address the lack of domain-specific knowledge in LLMs, we adopt in-context learning (ICL) and propose ICL-Bandit, a stochastic bandit algorithm for retrieving high-quality demonstrations and collecting LLM feedback to train a retriever. Our approach outperforms existing retrieval methods and has been successfully deployed in Microsoft's ad recommendation system, delivering significant cost savings and strong real-world effectiveness.",
+        "bbox": [
+            507,
+            266,
+            884,
+            457
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            470,
+            645,
+            483
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "ICL-Bandit's performance heavily relies on the quality and coverage of the labeled demonstration pool. If the pool lacks diverse or representative examples for certain query-ad pairs, the retrieved demonstrations may be suboptimal, limiting the effectiveness of in-context learning. This constraint can affect generalization, especially in long-tail or evolving domains where labeled data is sparse or outdated.",
+        "bbox": [
+            507,
+            495,
+            882,
+            639
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            667,
+            608,
+            682
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Marwah Alaofi, Paul Thomas, Falk Scholer, and Mark Sanderson. 2024. Llms can be fooled into labelling a document as relevant: best café near me; this paper is perfectly relevant. In Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, pages 32-41.",
+            "Ekaterina Artemova, Akim Tsvigun, Dominik Schlechtweg, Natalia Fedorova, Sergei Tilga, and Boris Obmoroshev. 2024. Hands-on tutorial: Labeling with llm and human-in-the-loop. arXiv preprint arXiv:2411.04637.",
+            "Sebastien Bubeck, Nicolo Cesa-Bianchi, and 1 others. 2012. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends® in Machine Learning, 5(1):1-122."
+        ],
+        "bbox": [
+            509,
+            689,
+            885,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "23447",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric P Xing, and Zhiting Hu. 2022. Rlprompt: Optimizing discrete text prompts with reinforcement learning. arXiv preprint arXiv:2205.12548.",
+            "Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey for in-context learning. arXiv preprint arXiv:2301.00234.",
+            "Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. 2023. Chatgpt outperforms crowd-workers for text-annotation tasks. arXiv preprint arXiv:2303.15056.",
+            "Gongde Guo, Hui Wang, David Bell, Yaxin Bi, and Kieran Greer. 2003. Knn model-based approach in classification. In On The Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE: OTM Confederated International Conferences, CoopIS, DOA, and ODBASE 2003, Catania, Sicily, Italy, November 3-7, 2003. Proceedings, pages 986-996. Springer.",
+            "Jannik Kossen, Tom Rainforth, and Yarin Gal. 2023. In-context learning in large language models learns label relationships but is not conventional learning. arXiv preprint arXiv:2307.12375.",
+            "Tze Leung Lai and Herbert Robbins. 1985. Asymptotically efficient adaptive allocation rules. Advances in applied mathematics, 6(1):4-22.",
+            "Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. 2023. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267.",
+            "Lihong Li, Wei Chu, John Langford, and Robert E Schapire. 2010. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web, pages 661-670.",
+            "Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023. Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320.",
+            "Xiaoliang Ling, Weiwei Deng, Chen Gu, Hucheng Zhou, Cui Li, and Feng Sun. 2017. Model ensemble for click prediction in bing search ads. In Proceedings of the 26th international conference on world wide web companion, pages 689-698.",
+            "Yu Meng, Martin Michalski, Jiaxin Huang, Yu Zhang, Tarek Abdelzaher, and Jiawei Han. 2023. Tuning language models as training data generators for augmentation-enhanced few-shot learning. In International Conference on Machine Learning, pages 24457-24477. PMLR.",
+            "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "others. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744.",
+            "Stephen Robertson, Hugo Zaragoza, and 1 others. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389.",
+            "Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633.",
+            "Peng Shi, Rui Zhang, He Bai, and Jimmy Lin. 2022. XrcI: Cross-lingual retrieval-augmented in-context learning for cross-lingual text-to-sql semantic parsing. arXiv preprint arXiv:2210.13693.",
+            "Jiamei Shuai, Shuayb Zarar, and Denis Charles. 2020. Modeling and evaluation framework for responsive search ads. In Machine Learning, AI & Data Science Conference (MLADS). Microsoft.",
+            "Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansoresh Karami, Jundong Li, Lu Cheng, and Huan Liu. 2024. Large language models for data annotation and synthesis: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 930-957.",
+            "Joannes Vermorel and Mehryar Mohri. 2005. Multiarmed bandit algorithms and empirical evaluation. In European conference on machine learning, pages 437-448. Springer.",
+            "Dong Wang, Shaoguang Yan, Yunqing Xia, Kave Salamatian, Weiwei Deng, and Qi Zhang. 2022a. Learning supplementary nlp features for ctr prediction in sponsored search. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4010-4020.",
+            "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022b. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560.",
+            "Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering.",
+            "Yiming Zhang, Shi Feng, and Chenhao Tan. 2022. Active example selection for in-context learning. arXiv preprint arXiv:2211.04486."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            804
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "23448",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Competitors",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            263,
+            99
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "For a fair comparison, all baselines and ICL-Bandit (except \"No Example\" and \"Crowdsourcing\") were provided with 3 positive (Relevant) and 3 negative (Irrelevant) historically labeled samples as demonstrations. The following methods were selected as our competitors:",
+        "bbox": [
+            112,
+            110,
+            487,
+            205
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- No Example (Zero-shot Learning): Zero-shot learning without any demonstrations.",
+            "- Crowdsourcing: Demonstrations annotated by human evaluators through crowdsourcing to assess query-ad relevance. It is different from the human (expert) labeled data for demonstraton pool, train and test data.",
+            "- EPR (SL-KNN): Demonstrations are retrieved using the K-nearest neighbor (KNN) algorithm based on the training datasets as ground truth. EPR (SL-KNN) is then trained to input query-ads and output the retrieved demonstrations to assist the LLM in labeling.",
+            "- EPR (SL-LLM) (Li et al., 2023): Demonstrations are retrieved using GPT-3.5 based on the training datasets as ground truth. EPR (SL-LLM) is then trained to input query-ads and output the retrieved demonstrations to assist the LLM in labeling.",
+            "- Q-learning (Zhang et al., 2022): A demonstration candidate is predefined, and Q-learning is utilized to learn the retrieval policy. Demonstrations are clustered into 50 clusters to implement this algorithm.",
+            "- Static: Demonstrations are pre-defined and kept static.",
+            "- BM25 (Robertson et al., 2009): Demonstrations retrieved using the BM25 algorithm.",
+            "- Random: Demonstrations randomly sampled for each user query.",
+            "- KNN (Guo et al., 2003): Demonstrations are retrieved using the K-nearest neighbor (KNN) algorithm based on the user query. We use the same feature embedding as our method to retrieve the demonstrations with cosine similarity in KNN."
+        ],
+        "bbox": [
+            114,
+            206,
+            487,
+            769
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            114,
+            782,
+            324,
+            799
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Dataset: In the experiments, we leveraged a meticulously curated dataset tailored specifically for assessing the efficacy of demonstration retrieval systems. This dataset is derived from high-quality human (expert)-labeled data collected daily over the recent 1.5-year period. Each sample in the dataset consists of a user query along with associated infor",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "mation about recommended advertisements. This information includes query keywords, ad titles, ad descriptions, ad URLs, and other pertinent content, each labeled as either relevant or irrelevant.",
+        "bbox": [
+            507,
+            84,
+            882,
+            148
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "To facilitate a robust evaluation, we partitioned the dataset temporally into three distinct subsets: an example pool, a training set, and a test set. The example pool contains all 1,578,728 samples as demonstrations, ensuring a comprehensive range of instances. For the purpose of training the model, we selected a subset of 9,999 samples specifically for query-ads pair labeling. The evaluation phase was carried out on a test set, which included 1,986 samples also designated for query-ads pair labeling. This temporal division helps in mimicking real-world scenarios where models are trained on historical data and tested on recent, unseen data, thereby providing insights into the practical applicability and performance of the retrieval methods under study.",
+        "bbox": [
+            507,
+            149,
+            882,
+            405
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metric: Our in-context learning method aims to enhance the labeling performance of large language models (LLMs). Given that the labeling task at hand is a binary classification problem, we evaluate the effectiveness of our approach using several key metrics. Specifically, we measure Accuracy (ACC), F1-Score, Precision, and Recall. These metrics collectively provide a comprehensive assessment of the model's performance in terms of both its ability to correctly label data and its balance between precision and recall.",
+        "bbox": [
+            507,
+            406,
+            882,
+            582
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Computational Resource All experiments are performed on single Ubuntu 20.04 LTS system with Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz CPU, 112 Gigabyte memory and single NVIDIA Tesla P100 accelerator.",
+        "bbox": [
+            507,
+            583,
+            880,
+            661
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "23449",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2025/ICL-Bandit_ Relevance Labeling in Advertisement Recommendation Systems via LLM/deeabab9-126e-40e2-a116-c0a839b8d3f0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:444c4ef777416c9876bb05d9dd3467ed88b96ef1c6a6526ddef58ee0af464486
+size 649772

2025/ICL-Bandit_ Relevance Labeling in Advertisement Recommendation Systems via LLM/full.md

@@ -0,0 +1,391 @@
+# ICL-Bandit: Relevance Labeling in Advertisement Recommendation Systems via LLM
+
+Lu Wang $^{1}$ , Chiming Duan $^{2*}$ , Pu Zhao $^{1}$ , Fangkai Yang $^{1}$ , Yong Shi $^{1}$ , Xuefeng Luo $^{1}$ , Bingjing Xu $^{1}$ , Weiwei Deng $^{1}$ , Qingwei Lin $^{1}$ , Dongmei Zhang $^{1}$ $^{1}$ Microsoft,  $^{2}$ Peking University
+
+# Abstract
+
+Measuring the relevance between user queries and advertisements is a critical task for advertisement (ad) recommendation systems, such as Microsoft Bing Ads and Google Ads. Traditionally, this requires expert data labeling, which is both costly and time-consuming. Recent advances have explored using Large Language Models (LLMs) for labeling, but these models often lack domain-specific knowledge. In-context learning (ICL), which involves providing a few demonstrations, is a common practice to enhance LLM performance on domain-specific tasks. However, retrieving high-quality demonstrations in a vast exploration space remains challenging. In this paper, we introduce ICL-Bandit, a practical and effective approach that leverages ICL to enhance the query-ad relevance labeling capabilities of LLMs. We develop a novel bandit learning method to identify and provide superior demonstrations for ICL, thereby improving labeling performance. Experimental results demonstrate that ICL-Bandit achieves state-of-the-art performance compared to existing methods. Additionally, ICL-Bandit has been deployed in Microsoft that serves billions of users worldwide, confirming its robustness and effectiveness.
+
+# 1 Introduction
+
+In advertisement (ad) recommendation systems such as Microsoft Bing Ads and Google Ads, high-quality labeled data is of critical importance for training ad recommendation models, especially labeling the relevance between user query text and ad description text, as discussed in (Ling et al., 2017; Shuai et al., 2020; Wang et al., 2022a). The traditional approach is human labeling which is costly and inefficient. This is particularly challenging given the huge amount of data to be labelled, and labeling such relevance between user query and ad
+
+requires a good knowledge and experience. For example,
+
+# User Query:
+
+"Innovative treatments for reducing hospital readmission rates in heart failure patients."
+
+# Advertisement:
+
+"Remote Patient Monitoring Systems - Continuous Care for Heart Failure Patients"
+
+This query-ad pair is labeled as relevant since remote patient monitoring systems provide continuous care and real-time health data, enabling proactive management of heart failure, which is essential for reducing hospital readmission rates, and such manual labeling requires domain knowledge.
+
+Recent advances in Large Language Models (LLMs) have shown that LLMs are highly aligned with human judgments and even surpass human performance in certain tasks (Ouyang et al., 2022), such as topic identification and twitter relevance for political issues (Gilardi et al., 2023), general question-answering data generation (Meng et al., 2023), and instruction data generation (Wang et al., 2022b). However, the lack of domain knowledge limits the performance of LLM in the query-ad relevance labeling task. To address this challenge, many approaches employed in-context learning (ICL) to incorporate domain-specific knowledge as extra context in the LLM's prompt (Kossen et al., 2023; Dong et al., 2022). Besides, it is well known that the effectiveness of ICL heavily depends on the quality of the provided demonstrations, which has motivated many works to explore effective demonstration retrieval methods for ICL, such as (Rubin et al., 2021; Li et al., 2023; Wu et al.; Zhang et al., 2022), all of these methods aim to retrieve better examples from annotated training sets to enhance LLMs' domain knowledge.
+
+Previous work on demonstration retrieval falls
+
+into two categories. One category involves off-the-shelf retrievers like BM25 (Robertson et al., 2009) or KNN (Guo et al., 2003), which can retrieve textually or semantically similar demonstrations. The other category focuses on training task-specific retrievers with positive and negative demonstrations. Notable examples include Rubin et al. (Rubin et al., 2021), Shi et al. (Shi et al., 2022), and Xiaonan et al. (Li et al., 2023), who leverage LLM feedback (compare the labels generated by LLM with the ground-truth labels, using them as the training signal) to train these retrievers via supervised or contrastive learning. However, the vast combination space of different demonstrations and queries poses a challenge. Randomly sampling demonstrations to collect the LLM's feedback may lead to large parts of "less useful" examples. Some methods, like Zhang et al. (Zhang et al., 2022) and Mingkai et al. (Deng et al., 2022), employ reinforcement learning to actively sample demonstrations and obtain LLM feedback. But these methods are limited in considering only a fixed number of candidate demonstrations, which reduces the action space for policy training.
+
+To overcome the challenges addressed above in SOTA methods, at first, we frame demonstration retrieval problem as a multi-armed bandit (MAB) problem (Lai and Robbins, 1985), and bandit algorithms solving MAB problems (Vermorel and Mohri, 2005; Li et al., 2010) have demonstrated excellent performance in addressing exploration and exploitation dilemma when dealing with largescale search spaces. This allows us to design effective exploration techniques for sampling demonstrations and obtaining LLM feedback during retriever training. Then, we propose a novel in-context learning (ICL) algorithm, called ICL-Bandit, which leverages a stochastic bandit algorithm to empower ICL at scale with diverse demonstration pools. The objective of ICL-Bandit is to retrieve demonstrations and maximize cumulative positive LLM feedback over a series of retrievals. Figure 1 shows a comparison on the example query-ad pair with demonstrations retrieved with KNN and our ICL-Bandit, respectively.
+
+Our contributions can be summarized as follows:
+
+- We formulate demonstration retrieval as a multi-armed bandit problem, focusing on effective retrieval during retriever training.  
+- We design a stochastic bandit algorithm suitable for ICL with a large and varied demon
+
+stration pool.
+
+- Our approach achieves SOTA performance comparing to other existing methods and it has been deployed to the labeling process at Microsoft, resulting in substantial cost savings by automated labeling.
+
+# 2 ICL-Bandit with large and varied demonstration pool
+
+In this section, we introduce ICL-Bandit, a stochastic bandit algorithm designed to efficiently retrieve demonstrations and collect LLM feedback during retriever training. The task is to precisely label query-ad relevance with ICL, and our goal is to train a demonstration retriever which retrieves good demonstrations for ICL. When training the retriever using LLM feedback (we compare the output labels generated by the LLM with the ground-truth labels, employing them as the reward signal), the key lies in how to effectively retrieve demonstrations from a large and diverse pool during the training process. Addressing the Exploration (searching for diverse and potentially informative demonstrations) versus Exploitation (retrieving high-reward demonstrations) balance is pivotal. To tackle this challenge, we first formulate the demonstration retrieval task as a bandit problem.
+
+# 2.1 Task Definition
+
+The task is to label the relevance of query-ad pairs leveraging LLMs. Compared with zero-shot LLM labeling, providing with demonstrations as context in the ICL manner improves the labeling performance. The ICL prompt comprises four key components:
+
+- Instruction We employ the following instruction to describe labeling requests: "Given user query and an ad, assign a label based on following definitions: - 'Relevant': The ad content directly addresses the user's query, providing information or a solution that aligns with the search intent. - 'Irrelevant': The ad content does not address the user's query, failing to provide information or a solution that matches the search intent." This instruction guides the relevance labeling process for LLMs.  
+- Input The input component specifies the query and ad requiring labeling. For instance, "User
+
+![](images/6d91cfd714ccf3e822c2e8d88f9fe236f6f973226d4a28863044fcb10afa9ca8.jpg)  
+Figure 1: An example illustrating how LLM labeler prompting, when combined with demonstrations retrieved via KNN and our ICL-Bandit approach, yields distinct labeling outcomes.
+
+![](images/f8f8fcbfc03dd67d689e151a98becdb91bc43fd54b49efcd1bbb84ff5f2cc66c.jpg)  
+Figure 2: An illustration on training demonstration retrieval process of ICL-Bandit. The demonstration pool consists of expert labeled demonstrations, and each demonstration includes user query, advertisement, and label. For each query-ad from the training set, ICL-Bandit retrieves demonstrations from the demonstration pool considering the estimated reward distribution. With the retrieved demonstrations as the context, LLM labels the query-ad as relevant or irrelevant. Then the LLM-generated label is compared with the ground truth label from the training set to give feedback. In the figure, the positive reward (matched label) is used to update ICL-Bandit to refine the retrieval policy.
+
+Query: 'Student loans suspended until september'; Advertisement: '10 Best Student Loan Refinance''. This information sets the context for the query-ad labeling task.
+
+- Demonstrations Demonstrations consist of a set of labeled demonstrations, such as, "User Query: 'School registration', Advertisement: 'Find Virtual School Programs', Label: Relevant". Different retrieved demonstrations would highly affect the labeling performance. These examples, provided by human annotators, serve as training instances for a policy  $\pi_{\theta}$  that leverages LLM feedback to retrieve appropriate demonstrations for each unique input.  
+- Output Indicator The output indicator instructs
+
+the LLM to generate the labeling answer. For instance, "Return your decision on the label in  $<\text{Label}>$  {{ $\text{}/\text{Label}>$  tags."}. This guides the LLM to generate the final label.
+
+# 2.2 Problem Formulation
+
+The key of the labeling task lies in retrieving informative demonstrations for ICL labeling, then we formulate the demonstration retrieval problem as a multi-armed bandit (MAB) problem, drawing an analogy to the scenario of a gambler selecting from a slot machine with multiple arms in a casino. The player's objective is to choose the arm that offers the highest expected gain. Each time the player pulls an arm and receives a gain or not, they update their estimation of the arm's potential
+
+gain. Similar to this scenario, in the query-ad relevance labeling task, we propose a MAB approach to retrieve demonstrations from a pool to achieve a good performance. We define key components such as states, arms, rewards, and the overall objective. **Trial:** In each trial  $t$ , our goal is to retrieve  $m$  demonstration from the demonstration pool. **State:** State  $s_t$  represents the current contextual environment. In the context of ICL,  $s_t$  denotes the embedding of the query and ads that require labeling. **Arm/Action:** Each arm in the MAB problem corresponds to a potential demonstration, representing different choice or action that can be taken during the ICL labeling. **Reward:** The reward  $r_{(t,a_k)}$  provides numerical feedback, indicating whether the LLM assigns the correct relevance label based on the retrieved demonstration  $a_k$ . We define two reward options: a continuous reward  $r_{(t,a_k)} \in [0,1]$  (representing the probability of the LLM's output label) and a discrete reward  $r_{(t,a_k)} \in [0,1]$ . In the discrete case, a correct label receives a reward of 1, while an incorrect label receives 0. **Objective:** The overall goal is to learn a retrieval policy  $\pi_\theta$  by maximizing the cumulative reward over a series of trials during training.
+
+The process of using a MAB algorithm to efficiently retrieve demonstrations during training consists of the following steps:
+
+Step 1: Retrieval of Demonstrations At each trial  $t$ , the retrieval policy  $\pi_{\theta}$  retrieves a demonstration (arm)  $a_{k}$  from a demonstration pool. The optimal retrieval will be conducted based on a balance between the benefits (the mean of  $a_{k}$ 's rewards) and the chance (the variance of  $a_{k}$ 's rewards).
+
+Step 2: Reward from LLM Feedback After retrieving a demonstration  $a_{k}$ , the retriever receives a reward  $r_t$ , indicating the correctness of the label provided by the LLM with the chosen demonstration.
+
+Step 3: Policy Update The collected reward value is used to update the policy parameters  $\pi_{\theta}$  to maximize cumulative reward during training.
+
+These steps iteratively occur for a pre-defined number of iterations ( $T = 2000$  in our paper). The exploration-exploitation trade-off in Step 1 is crucial, requiring the retrieval policy to balance exploring new demonstrations for potential benefits (exploration) and exploiting known good demonstrations to maximize the mean of rewards (exploitation). This exploration-exploitation balance ensures the effectiveness of learning the demonstration retrieval policy in ICL labeling.
+
+# 2.3 ICL-Bandit
+
+In this section, we introduce ICL-Bandit as an innovative approach to address the challenges encountered by previous bandit algorithms. Traditional bandit algorithms, which assign a parameter  $\theta$  to each arm or action, facing the limitations when applied to demonstration retrieval due to the expansive and varied nature of the available demonstrations.
+
+We leverage the framework of Stochastic Multi-Armed Bandit (Bubeck et al., 2012), a variant of the classical MAB problem where rewards associated with different actions (referred to as "arms") are influenced by stochastic processes. In our context, we develop a novel stochastic bandit algorithm tailored to scenarios with an extensive and diverse set of demonstrations.
+
+First, we fine-tune a BERT model in Microsoft's query and ads dataset, resulting in an embedding vector  $e_{a_k}$  unique to each demonstration  $a_k$ , an embedding vector  $e_{s_t}$  for the state  $s_t$  at trial  $t$  and a mutual embedding  $e_{(s,a)_t}$ . Then we have a unified feature embedding  $x_{s_t,a_k} = [e_{s_t}, e_{a_k}, e_{(s,a)_t}]$  to capture contextual information. Next, we adopt a shared parameter  $\theta$  applicable to all demonstrations. This shared parameterization streamlines the learning process, enhancing efficiency and generalization across the diverse pool of demonstrations.
+
+# 2.3.1 Demonstration and State Representation Learning
+
+To integrate both the state and demonstration into a unified embedding vector, we employ a self-supervised learning approach to fine-tune a 24-layer BERT model using Microsoft's user query-ad dataset. The final embedding is derived by extracting the output of the last hidden layer, serving as a comprehensive representation of both the state and demonstration.
+
+# 2.3.2 Objective Function of ICL-Bandit
+
+Throughout the total  $T$  trials, the cumulative reward is defined as  $\sum_{t=1}^{T} r_{(t,a_k)}$ . In this context, we establish the optimal expected  $T$ -iteration reward, denoted as  $E[\sum_{t=1}^{T} r_{(t,a_k^*)}]$ , where  $a_k^*$  represents the optimal demonstration yielding the maximum expected reward at trial  $t$ . Our objective is to proficiently retrieve a sequence of demonstrations during training, maximizing the expected total payoff. Alternatively, our aim is to minimize the regret of the algorithm concerning the optimal demonstration retrieval strategy. The  $T$ -iteration regret of
+
+ICL-Bandit can be formally defined as:
+
+$$
+R e (T) = E [ \sum_ {t = 1} ^ {T} r _ {(t, a _ {k} ^ {*})} ] - E [ \sum_ {t = 1} ^ {T} r _ {(t, a _ {k})} ] \quad (1)
+$$
+
+# 2.3.3 Optimization of ICL-Bandit
+
+To minimize the regret, it is assumed that the expected reward of an example  $a$  is linear in the  $d$ -dimensional state-action integrated feature  $x_{(t,a)}$ , with three unknown policy parameters  $\theta_{\mathrm{state}}^*$ ,  $\theta_{\mathrm{action}}^*$  and  $\theta_{\mathrm{mut}}^*$ :
+
+$$
+\mathbb {E} [ r _ {(t, a)} \mid x _ {(t, a)} ] = x _ {(t, a)} ^ {T} [ \theta_ {\mathrm {s t a t e}} ^ {*}, \theta_ {\mathrm {a c t i o n}} ^ {*}, \theta_ {\mathrm {m u t}} ^ {*} ]
+$$
+
+where  $\theta_{\mathrm{state}}^{*}$  denotes the policy parameter for the context or the state, i.e., for the target sample,  $\theta_{\mathrm{action}}^{*}$  denotes the policy parameter for an action, i.e., for an example, and  $\theta_{\mathrm{mut}}^{*}$  denotes the policy parameter for mutual information of the target sample and the example. The mutual information may be common or similar information between the target sample and the example. The policy parameter  $\theta_{\mathrm{state}}^{*}$  can map the target sample to a first vector space. The policy parameter  $\theta_{\mathrm{action}}^{*}$  can map the example to a second vector space. The first vector space and the second vector space are different and independent, but they are dual to each other. The policy parameter  $\theta_{\mathrm{mut}}^{*}$  can map both the target sample and the example to the same vector space. The technical effect of using the three policy parameters  $\theta_{\mathrm{state}}^{*}$ ,  $\theta_{\mathrm{action}}^{*}$ , and  $\theta_{\mathrm{mut}}^{*}$  is to more accurately measure the relationship or distance between the target sample and the example, so as to calculate a more accurate expected reward.
+
+The embodiments of the present disclosure propose that all examples share three policy parameters  $\theta_{\mathrm{state}}^{*}$ ,  $\theta_{\mathrm{action}}^{*}$ , and  $\theta_{\mathrm{mut}}^{*}$ . This parameterization remains constant regardless of the number of examples. The technical effect of such settings is to streamline the learning process, and enhance efficiency and generalization across the diverse set of examples. This framework enables the application of the proposed reinforced retrieval operation to large-scale and diverse candidate examples, contributing to its scalability and adaptability.
+
+For each example, we have three kinds of features: state, action, and mut, denoted as  $[e_{s_t}, e_{a_k}, e_{(s,a)_t}]$ . These correspond to the data matrices  $D_{\mathrm{state}}$ ,  $D_{\mathrm{action}}$ , and  $D_{\mathrm{mut}}$ , which represent samples on different features. Let  $D_{\mathrm{state}}$ ,  $D_{\mathrm{action}}$ , and  $D_{\mathrm{mut}}$  be data matrices of dimension  $m \times d$  at
+
+trial  $t$ , where the rows correspond to  $m$  training inputs of context, action, and mutual information, and  $b \in \mathbb{R}^m$  is the corresponding reward vector (e.g., the  $m$  rewards indicating whether the LLM provided the correct label in the training set). Applying ridge regression to the training data  $(D, b)$  yields an estimate of the policy parameters:
+
+$$
+\theta_ {\text {s t a t e}} = \left(D _ {\text {s t a t e}} ^ {T} D _ {\text {s t a t e}} + \lambda I\right) ^ {- 1} D _ {\text {s t a t e}} ^ {T} b \tag {2}
+$$
+
+$$
+\theta_ {\text {a c t i o n}} = \left(D _ {\text {a c t i o n}} ^ {T} D _ {\text {a c t i o n}} + \lambda I\right) ^ {- 1} D _ {\text {a c t i o n}} ^ {T} b \tag {3}
+$$
+
+$$
+\theta_ {\mathrm {m u t}} = \left(D _ {\mathrm {m u t}} ^ {T} D _ {\mathrm {m u t}} + \lambda I\right) ^ {- 1} D _ {\mathrm {m u t}} ^ {T} b \tag {4}
+$$
+
+where  $I$  is the  $d\times d$  identity matrix,  $\lambda \in [0,1]$  is the regularization term of ridge regression estimation. Let  $D = [D_{\mathrm{state}},D_{\mathrm{action}},D_{\mathrm{mut}}]$ , and  $\theta = [\theta_{\mathrm{state}},\theta_{\mathrm{action}},\theta_{\mathrm{mut}}]$ . When components in  $b$  are independent conditioned on corresponding rows in  $D$ , it can be shown that, with probability at least  $1 - \delta$ :  $\left|x_{(t,a_t)}^T\hat{\theta} - \mathbb{E}[r_{(t,a_t)}\mid x_{(t,a_t)}]\right| \leq \alpha' \sqrt{x_{(t,a_t)}^T(D^TD + \lambda I)^{-1}x_{(t,a_t)}}$  For any  $\delta > 0$  and  $x_{(t,a_t)} \in \mathbb{R}^d$ , where  $\hat{\theta}$  is the mean of  $\theta$ , and  $a_t$  indicates the example selected at  $t$ ,  $r_{(t,a)}$  is the observed reward,  $\sigma^2$  is the variance proxy of the noise and  $\alpha'$  is a constant.
+
+# 2.4 Proof
+
+# 2.4.1 Step 1: Decompose the Estimation Error
+
+The estimation error can be expressed as:
+
+$$
+\begin{array}{l} \hat {\theta} - \theta^ {*} = \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} b - \theta^ {*} \\ = \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} \left(D \theta^ {*} + \epsilon\right) - \theta^ {*} \\ = \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} D \theta^ {*} + \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} \epsilon - \theta^ {*} \\ = \left[ \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} D - I \right] \theta^ {*} + \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} \epsilon . \\ \end{array}
+$$
+
+Simplifying:
+
+$$
+\hat {\theta} - \theta^ {*} = - \lambda \left(D ^ {\top} D + \lambda I\right) ^ {- 1} \theta^ {*} + \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} \epsilon .
+$$
+
+# 2.4.2 Step 2: Express the Estimation Error Components
+
+Let:  $\mathrm{Bias} = -\lambda (D^{\top}D + \lambda I)^{-1}\theta^{*}$
+
+Variance  $= \left(D^{\top}D + \lambda I\right)^{-1}D^{\top}\epsilon$
+
+Then:  $\hat{\theta} -\theta^{*} = \mathrm{Bias} + \mathrm{Variance}$
+
+# 2.4.3 Step 3: Bound the Bias Term
+
+We aim to bound  $\left|x_{(t,a)}^{\top} \text{Bias}\right|$ .
+
+Using the Cauchy-Schwarz inequality:
+
+$$
+\begin{array}{l} \left| x _ {(t, a)} ^ {\top} \operatorname {B i a s} \right| = \lambda \left| x _ {(t, a)} ^ {\top} \left(D ^ {\top} D + \lambda I\right) ^ {- 1} \theta^ {*} \right| (5) \\ \leq \lambda \left\| x _ {(t, a)} \right\| _ {\left(D ^ {\top} D + \lambda I\right) ^ {- 1}} \| \theta^ {*} \|, (6) \\ \end{array}
+$$
+
+where  $\| x\| _A = \sqrt{x^\top Ax}$  denotes the Mahalanobis norm with respect to the matrix  $A$
+
+Assuming  $\| \theta^{*}\| \leq S$ , where  $S$  is a known bound on the norm of  $\theta^{*}$ , we have:
+
+$$
+\left| x _ {(t, a)} ^ {\top} \mathbf {B i a s} \right| \leq \lambda S \left\| x _ {(t, a)} \right\| _ {\left(D ^ {\top} D + \lambda I\right) ^ {- 1}}. \tag {7}
+$$
+
+# 2.4.4 Step 4: Bound the Variance Term
+
+We aim to bound  $\left|x_{(t,a)}^{\top}\right|$  Variance with high probability.
+
+Since  $\epsilon$  has independent components with zero mean and variance proxy  $\sigma^2$ , the variance of  $x_{(t,a)}^{\top}$  Variance is:
+
+$$
+\begin{array}{l} \operatorname {V a r} \left(x _ {(t, a)} ^ {\top} \text {V a r i a n c e}\right) = \operatorname {V a r} \left(x _ {(t, a)} ^ {\top} \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} \epsilon\right) \\ = \sigma^ {2} x _ {(t, a)} ^ {\top} \left(D ^ {\top} D + \lambda I\right) ^ {- 1} D ^ {\top} D \left(D ^ {\top} D + \lambda I\right) ^ {- 1} x _ {(t, a)}. \\ \end{array}
+$$
+
+Therefore:
+
+$$
+\begin{array}{l} \operatorname {V a r} \left(x _ {(t, a)} ^ {\top} \text {V a r i a n c e}\right) = \sigma^ {2} x _ {(t, a)} ^ {\top} \left(D ^ {\top} D + \lambda I\right) ^ {- 1} \\ \cdot \left(I - \lambda (D ^ {\top} D + \lambda I) ^ {- 1}\right) x _ {(t, a)} \\ \leq \sigma^ {2} \left\| x _ {(t, a)} \right\| _ {(D ^ {\top} D + \lambda I) ^ {- 1}} ^ {2}. \\ \end{array}
+$$
+
+# 2.4.5 Step 5: Apply Concentration Inequality
+
+Since  $x_{(t,a)}^{\top}$  Variance is a linear combination of independent sub-Gaussian variables, it is sub-Gaussian with parameter  $\sigma' = \sigma\left\| x_{(t,a)}\right\|_{(D^{\top}D + \lambda I)^{-1}}$ .
+
+Using a sub-Gaussian tail bound, for any  $\delta > 0$ :
+
+$$
+P \left(\left| x _ {(t, a)} ^ {\top} \text {V a r i a n c e} \right| \geq \alpha \left\| x _ {(t, a)} \right\| _ {\left(D ^ {\top} D + \lambda I\right) ^ {- 1}}\right) \leq \delta \tag {8}
+$$
+
+where  $\alpha = \sigma \sqrt{2\ln\left(\frac{1}{\delta}\right)}$
+
+# Step 6: Combine Bias and Variance Terms
+
+The total estimation error is:
+
+$$
+\left| x _ {(t, a)} ^ {\top} \left(\hat {\theta} - \theta^ {*}\right) \right| \leq \left| x _ {(t, a)} ^ {\top} \operatorname {B i a s} \right| + \left| x _ {(t, a)} ^ {\top} \operatorname {V a r i a n c e} \right|. \tag {9}
+$$
+
+# 2.4.6 Step 7: Final Inequality
+
+Combine the bounds:
+
+$$
+\left| x _ {(t, a)} ^ {\top} \left(\hat {\theta} - \theta^ {*}\right) \right| \leq (\lambda S + \alpha) \| x _ {(t, a)} \| _ {(D ^ {\top} D + \lambda I) ^ {- 1}}.
+$$
+
+For sufficiently small  $\lambda$  and bounded  $\theta^{*}$ , the bias term can be controlled, and the dominant term becomes the variance term.
+
+Therefore, we can simplify the inequality to:  $\left|x_{(t,a)}^{\top}\left(\hat{\theta} -\theta^{*}\right)\right|\leq \alpha^{\prime}\left\| x_{(t,a)}\right\|_{\left(D^{\top}D + \lambda I\right)^{-1}},$
+
+where  $\alpha^{\prime} = \lambda S + \alpha$
+
+# 2.5 ICL-Bandit vs. Traditional Bandit
+
+ICL-Bandit improves upon traditional bandit methods by introducing shared parameters,  $\theta_{\mathrm{state}}^{*}$ ,  $\theta_{\mathrm{action}}^{*}$  and  $\theta_{\mathrm{mut}}^{*}$ , to jointly model state, actions, and their interactions. This enables better alignment between context and candidate demonstrations, leading to more accurate action selection.
+
+Unlike traditional methods that treat actions independently, ICL-Bandit captures complex contextual dependencies, enhances generalization, and scales efficiently to high-dimensional data. Its unified framework ensures consistent performance across diverse ICL labeling tasks, mitigating the inconsistency and overfitting often seen in conventional approaches.
+
+# 3 Experiment
+
+Dataset: We use a high-quality, expert-labeled dataset collected daily over 1.5 years, consisting of user queries and associated advertisement information (e.g., keywords, titles, descriptions, URLs), each labeled as relevant or irrelevant. The dataset is temporally split into: Example pool: 1,578,728 samples used as demonstrations. Training set: 9,999 query-ad pairs. Test set: 1,986 query-ad pairs. This temporal partitioning simulates real-world deployment, where models are trained on historical data and evaluated on recent, unseen examples.
+
+Evaluation Metrics: We assess binary classification performance using Accuracy (ACC), F1-score, Precision, and Recall to capture both correctness and balance in predictions. More details of the experimental settings are provided in Appendix B.
+
+# 3.1 Competitors
+
+For a fair comparison, all baselines and ICL-Bandit (except "No Example" and "Crowdsourcing") were provided with 3 positive (Relevant) and 3 negative (Irrelevant) historically labeled samples as demonstrations. The following methods were selected as
+
+our competitors: No Example (Zero-shot Learning), Crowdsourcing, EPR (SL-KNN), EPR (SL-LLM) (Li et al., 2023), Q-learning (Zhang et al., 2022), Static, BM25 (Robertson et al., 2009), Random, KNN (Guo et al., 2003). Details of the baseline methods are provided in Appendix A.
+
+Table 1: Results of GPT-3.5 as the backbone LLM.  
+
+<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>53.95</td><td>25.72</td><td>41.76</td><td>18.58</td></tr><tr><td>Crowdsourcing</td><td>67.57</td><td>73.65</td><td>90.09</td><td>62.28</td></tr><tr><td>EPR (SL-KNN)</td><td>55.97</td><td>25.68</td><td>39.22</td><td>19.09</td></tr><tr><td>EPR (SL-LLM)</td><td>57.18</td><td>25.44</td><td>41.55</td><td>18.33</td></tr><tr><td>Q-learning</td><td>58.68</td><td>34.71</td><td>43.78</td><td>28.14</td></tr><tr><td>Static</td><td>56.57</td><td>28.56</td><td>43.40</td><td>21.28</td></tr><tr><td>BM25</td><td>58.84</td><td>28.77</td><td>46.35</td><td>20.86</td></tr><tr><td>Random</td><td>50.73</td><td>36.49</td><td>37.52</td><td>35.52</td></tr><tr><td>KNN</td><td>57.98</td><td>27.60</td><td>44.04</td><td>20.10</td></tr><tr><td>ICL-Bandit (Ours)</td><td>63.76</td><td>61.63</td><td>65.38</td><td>58.29</td></tr></table>
+
+Table 2: Results of GPT-4 as the backbone LLM.  
+
+<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>65.05</td><td>61.45</td><td>64.08</td><td>59.03</td></tr><tr><td>Crowdsourcing</td><td>67.57</td><td>73.65</td><td>90.09</td><td>62.28</td></tr><tr><td>EPR (SL-KNN)</td><td>74.42</td><td>80.70</td><td>73.49</td><td>89.47</td></tr><tr><td>EPR (SL-LLM)</td><td>74.62</td><td>80.70</td><td>72.94</td><td>90.32</td></tr><tr><td>Q-learning</td><td>74.26</td><td>76.25</td><td>78.81</td><td>87.80</td></tr><tr><td>Static</td><td>73.56</td><td>79.71</td><td>71.35</td><td>90.28</td></tr><tr><td>BM25</td><td>73.62</td><td>80.09</td><td>72.94</td><td>88.80</td></tr><tr><td>Random</td><td>73.72</td><td>79.94</td><td>71.97</td><td>89.89</td></tr><tr><td>KNN</td><td>74.47</td><td>80.74</td><td>73.56</td><td>89.48</td></tr><tr><td>ICL-Bandit (Ours)</td><td>80.03</td><td>82.57</td><td>76.91</td><td>89.14</td></tr></table>
+
+# 3.2 Results Analysis
+
+Tables 1 and 2 present the experimental results comparing nine demonstration retrieval methods, including our ICL-Bandit, across two versions of LLMs. The analysis highlights key performance trends. The "No Example" baseline performs poorly, while "Crowdsourcing" demonstrations achieve the highest accuracy and precision, emphasizing the importance of expert-labeled data. Among automated methods, Q-learning, EPR (SL-LLM), and ICL-Bandit show strong performance, benefiting from LLM feedback. Notably, ICL-Bandit surpasses Q-learning and EPR (SL-LLM) despite using only 2,000 feedback samples compared to their 5,000, due to its lightweight, linear design that requires fewer data.
+
+EPR (SL-KNN) and EPR (SL-LLM) improve over the "No Example" baseline but still lag behind "Crowdsourcing," indicating that retrieval effectiveness depends on technique selection. Similarly, methods like "Static," "Random," "KNN," and "BM25" show varied performance, with BM25 performing competitively but still unable to match expert-labeled demonstrations.
+
+ICL-Bandit consistently delivers superior results, often outperforming or matching "Crowdsourcing."
+
+Its ability to balance exploration and exploitation allows it to retrieve relevant demonstrations effectively, adapt to diverse queries, and enhance recall, improving overall ICL performance.
+
+# 3.3 Learning Curve of ICL-Bandit
+
+The learning curve experiment was devised to examine the evolutionary performance of ICL-Bandit as training data accumulates. The primary objective was to discern how the method's effectiveness scales with an expanding dataset, providing insights into its adaptability and scalability. The experiment's results are depicted in Figure 3, where the x-axis represents training iterations, and the y-axis portrays the cumulative mean and variance of Accuracy, Binary Accuracy, True Negative Rate (TNR), and True Positive Rate (TPR). The learning curve analysis of ICL-Bandit highlights its capacity to dynamically adapt and enhance its performance over successive training iterations. Notably, it illustrates that ICL-Bandit achieves a rapid and stable convergence to a commendable performance level.
+
+Furthermore, the outcomes suggest that ICL-Bandit exhibits promise for demonstration retrieval in ICL, even when trained on a limited LLM feedback dataset. Remarkably, in comparison to EPR (SL-LLM), which utilized a larger dataset of 5000 feedback instances, ICL-Bandit demonstrates superior performance. The learning curve analysis underscores the efficacy of ICL-Bandit in iteratively improving its performance with an increasing volume of training data. This positions it as a robust and scalable solution for the nuanced task of demonstration retrieval in complex information retrieval scenarios.
+
+# 3.4 Ablation Study
+
+# 3.4.1 Number of retrivaled samples
+
+<table><tr><td>Example number</td><td>ACC</td><td>Binary AUC</td><td>TNR</td><td>TPR</td></tr><tr><td>1</td><td>0.7826</td><td>0.8127</td><td>0.8326</td><td>0.6978</td></tr><tr><td>3</td><td>0.8003</td><td>0.8029</td><td>0.8215</td><td>0.7077</td></tr><tr><td>6</td><td>0.8001</td><td>0.8127</td><td>0.8178</td><td>0.7129</td></tr><tr><td>9</td><td>0.7697</td><td>0.7616</td><td>0.7716</td><td>0.7516</td></tr></table>
+
+Table 3: Performance metrics on different number of selected demonstrations.
+
+In this experiment, we evaluate the impact of varying the number of positive and negative demonstrations on model performance. The results, presented in Table 3, indicate that the performance metrics (ACC, Binary AUC, TNR, and TPR) generally improve as the number of positive/negative demonstrations increases from 1 to 3. Specifically,
+
+![](images/d7d5c558b1db9657f34936e0d405f5a6ab665dd22e888e995c898acaebe8143c.jpg)  
+Figure 3: The learning curve of ICL-Bandit during 2000 trails training. TNR and TPR indicates the true negative rate and true positive rate respectively.
+
+the best overall performance is observed when 3 demonstrations are used. Thus we choose 3 as the final number.
+
+When 9 demonstrations are used, the performance metrics begin to decline, indicating that adding too many demonstrations may lead to diminishing returns or even reduced performance.
+
+# 3.4.2 Training Epochs and Reward Types
+
+The experiment evaluates the performance of the ICL-bandit approach under two reward settings: continuous and discrete. An epoch is defined as a complete pass through the training data. During each epoch, the ICL-bandit retrieves informative demonstrations, selects the best actions, and updates its retrieval policy based on the rewards received. The results in Figure 4 illustrate how the number of epochs affects performance across various metrics.
+
+We observe that ICL-bandit's performance varies with the number of epochs, with different metrics reaching their optimal levels at different stages. The continuous reward setting, which provides more detailed feedback, achieves peak performance in fewer epochs compared to the discrete reward setting. This suggests that using continuous rewards in practice can reduce training complexity while still delivering strong performance. Finally, we choose continuous reward with 1 epoch for reducing the complexity and promising results.
+
+# 3.5 Application in Practice
+
+We deployed the ICL-Bandit approach in Microsoft's ad relevance pipeline to reduce manual labeling costs and enhance ad recommendation quality. Each day, the system collects fresh user queries and ads, cleans them using Bing's distributed platform, and applies ICL-Bandit for automated labeling. For each query-ad pair, we retrieve 3 relevant
+
+![](images/c7a9608267025bd44e4cff50d16a16db196712f3a8eade1cc6edc02102dd581b.jpg)  
+Figure 4: Performance on different epochs and reward types.
+
+and 3 irrelevant historical examples to construct prompts, which are then labeled using GPT-4. This process generates high-quality labeled data daily for downstream CTR prediction (Lee et al., 2023). Table 4: A/B testing on English datasets with GPT-4 as the backbone LLM.
+
+<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>66.82</td><td>62.18</td><td>63.74</td><td>60.75</td></tr><tr><td>KNN</td><td>76.52</td><td>75.12</td><td>78.72</td><td>71.81</td></tr><tr><td>ICL-Bandit (Ours)</td><td>87.12</td><td>82.28</td><td>86.95</td><td>78.17</td></tr></table>
+
+Table 5: A/B testing on non-English dataset with GPT-4 as the backbone LLM.  
+
+<table><tr><td>Model</td><td>ACC (%)</td><td>F1-score (%)</td><td>Precision (%)</td><td>Recall (%)</td></tr><tr><td>No Example</td><td>63.26</td><td>59.48</td><td>60.17</td><td>58.79</td></tr><tr><td>KNN</td><td>70.67</td><td>75.44</td><td>70.86</td><td>80.81</td></tr><tr><td>ICL-Bandit (Ours)</td><td>80.67</td><td>85.68</td><td>82.57</td><td>89.14</td></tr></table>
+
+As shown in Tables 4 and 5, ICL-Bandit consistently outperforms baselines in both English and non-English settings, demonstrating robust improvements in accuracy, F1-score, precision, and recall.
+
+# 3.6 Impact on Ad Recommendation
+
+Integrating ICL-Bandit-labeled data into Bing's CTR prediction model led to significant business gains. Offline evaluation on 500K historical query-
+
+ad pairs showed a  $2.5\%$  AUC increase and  $1.8\%$  reduction in Log Loss. In two weeks of online A/B testing with 2 million users, CTR rose by  $3.2\%$  and conversion rates improved by  $2.7\%$ . Beyond performance, the automated labeling process reduced manual annotation costs by  $61\%$ , enabling scalable and cost-effective data processing across millions of queries daily. The results of the A/B testing on English and Non-English datasets are summarized in Table 4 and Table 5.
+
+# 4 Related Work
+
+# 4.1 LLM Labeling
+
+Latest studies in LLM have shown that LLM is highly consistent with human judgments and even outperforms humans in many tasks, for example, topic identification and twitter relevance for political issues (Gilardi et al., 2023), general question-answering data generation (Meng et al., 2023), instruction data generation (Wang et al., 2022b) and RL from AI feedback (RLAIF) (Lee et al., 2023). A set of work using LLM for labeling instead of human (Tan et al., 2024; Alaofi et al., 2024; Artemova et al., 2024). In this work, we focus on a domain-specific labeling problem, i.e., query-ad relevance labeling, which requires domain knowledge to guide LLM for labeling.
+
+# 4.2 Demonstration Retrieval for In-Context Learning
+
+LLMs have emerged as a pivotal strategy for addressing tasks specific to particular domains. However, the effectiveness of ICL is intrinsically tied to the quality of the provided demonstrations (Li et al., 2023; Wu et al.; Zhang et al., 2022). Works such as (Rubin et al., 2021; Li et al., 2023; Wu et al.; Zhang et al., 2022) collectively aim to optimize the retrieval of exemplary instances from annotated training sets, thereby enhancing the domain knowledge encapsulated by LLMs.
+
+Existing demonstration retrieval methods are typically categorized into utilization of off-the-shelf retrievers such as BM25 (Robertson et al., 2009) or KNN (Guo et al., 2003), or training task-specific retrievers using positive and negative demonstrations (Rubin et al., 2021; Shi et al., 2022; Li et al., 2023). These researchers leverage LLM feedback to guide the training of these retrievers through supervised or contrastive learning. Despite these advancements, the vast combinatorial space encompassing different demonstrations and queries
+
+presents a significant challenge. Randomly sampling demonstrations to collect LLM feedback risks incorporating a substantial portion of less useful examples. Reinforcement learning-based methods (Zhang et al., 2022; Deng et al., 2022) actively sample demonstrations and elicit valuable LLM feedback. However, they are constrained by a fixed number of demonstrations, thereby limiting the action space available for policy training.
+
+# 5 Conclusion
+
+In this paper, we leverage LLMs to automate query-ad relevance labeling for improved ad recommendation. To address the lack of domain-specific knowledge in LLMs, we adopt in-context learning (ICL) and propose ICL-Bandit, a stochastic bandit algorithm for retrieving high-quality demonstrations and collecting LLM feedback to train a retriever. Our approach outperforms existing retrieval methods and has been successfully deployed in Microsoft's ad recommendation system, delivering significant cost savings and strong real-world effectiveness.
+
+# 6 Limitations
+
+ICL-Bandit's performance heavily relies on the quality and coverage of the labeled demonstration pool. If the pool lacks diverse or representative examples for certain query-ad pairs, the retrieved demonstrations may be suboptimal, limiting the effectiveness of in-context learning. This constraint can affect generalization, especially in long-tail or evolving domains where labeled data is sparse or outdated.
+
+# References
+
+Marwah Alaofi, Paul Thomas, Falk Scholer, and Mark Sanderson. 2024. Llms can be fooled into labelling a document as relevant: best café near me; this paper is perfectly relevant. In Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, pages 32-41.  
+Ekaterina Artemova, Akim Tsvigun, Dominik Schlechtweg, Natalia Fedorova, Sergei Tilga, and Boris Obmoroshev. 2024. Hands-on tutorial: Labeling with llm and human-in-the-loop. arXiv preprint arXiv:2411.04637.  
+Sebastien Bubeck, Nicolo Cesa-Bianchi, and 1 others. 2012. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends® in Machine Learning, 5(1):1-122.
+
+Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric P Xing, and Zhiting Hu. 2022. Rlprompt: Optimizing discrete text prompts with reinforcement learning. arXiv preprint arXiv:2205.12548.  
+Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey for in-context learning. arXiv preprint arXiv:2301.00234.  
+Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. 2023. Chatgpt outperforms crowd-workers for text-annotation tasks. arXiv preprint arXiv:2303.15056.  
+Gongde Guo, Hui Wang, David Bell, Yaxin Bi, and Kieran Greer. 2003. Knn model-based approach in classification. In On The Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE: OTM Confederated International Conferences, CoopIS, DOA, and ODBASE 2003, Catania, Sicily, Italy, November 3-7, 2003. Proceedings, pages 986-996. Springer.  
+Jannik Kossen, Tom Rainforth, and Yarin Gal. 2023. In-context learning in large language models learns label relationships but is not conventional learning. arXiv preprint arXiv:2307.12375.  
+Tze Leung Lai and Herbert Robbins. 1985. Asymptotically efficient adaptive allocation rules. Advances in applied mathematics, 6(1):4-22.  
+Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. 2023. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267.  
+Lihong Li, Wei Chu, John Langford, and Robert E Schapire. 2010. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web, pages 661-670.  
+Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023. Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320.  
+Xiaoliang Ling, Weiwei Deng, Chen Gu, Hucheng Zhou, Cui Li, and Feng Sun. 2017. Model ensemble for click prediction in bing search ads. In Proceedings of the 26th international conference on world wide web companion, pages 689-698.  
+Yu Meng, Martin Michalski, Jiaxin Huang, Yu Zhang, Tarek Abdelzaher, and Jiawei Han. 2023. Tuning language models as training data generators for augmentation-enhanced few-shot learning. In International Conference on Machine Learning, pages 24457-24477. PMLR.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1
+
+others. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744.  
+Stephen Robertson, Hugo Zaragoza, and 1 others. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389.  
+Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633.  
+Peng Shi, Rui Zhang, He Bai, and Jimmy Lin. 2022. XrcI: Cross-lingual retrieval-augmented in-context learning for cross-lingual text-to-sql semantic parsing. arXiv preprint arXiv:2210.13693.  
+Jiamei Shuai, Shuayb Zarar, and Denis Charles. 2020. Modeling and evaluation framework for responsive search ads. In Machine Learning, AI & Data Science Conference (MLADS). Microsoft.  
+Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansoresh Karami, Jundong Li, Lu Cheng, and Huan Liu. 2024. Large language models for data annotation and synthesis: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 930-957.  
+Joannes Vermorel and Mehryar Mohri. 2005. Multiarmed bandit algorithms and empirical evaluation. In European conference on machine learning, pages 437-448. Springer.  
+Dong Wang, Shaoguang Yan, Yunqing Xia, Kave Salamatian, Weiwei Deng, and Qi Zhang. 2022a. Learning supplementary nlp features for ctr prediction in sponsored search. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4010-4020.  
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022b. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560.  
+Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering.  
+Yiming Zhang, Shi Feng, and Chenhao Tan. 2022. Active example selection for in-context learning. arXiv preprint arXiv:2211.04486.
+
+# A Competitors
+
+For a fair comparison, all baselines and ICL-Bandit (except "No Example" and "Crowdsourcing") were provided with 3 positive (Relevant) and 3 negative (Irrelevant) historically labeled samples as demonstrations. The following methods were selected as our competitors:
+
+- No Example (Zero-shot Learning): Zero-shot learning without any demonstrations.  
+- Crowdsourcing: Demonstrations annotated by human evaluators through crowdsourcing to assess query-ad relevance. It is different from the human (expert) labeled data for demonstraton pool, train and test data.  
+- EPR (SL-KNN): Demonstrations are retrieved using the K-nearest neighbor (KNN) algorithm based on the training datasets as ground truth. EPR (SL-KNN) is then trained to input query-ads and output the retrieved demonstrations to assist the LLM in labeling.  
+- EPR (SL-LLM) (Li et al., 2023): Demonstrations are retrieved using GPT-3.5 based on the training datasets as ground truth. EPR (SL-LLM) is then trained to input query-ads and output the retrieved demonstrations to assist the LLM in labeling.  
+- Q-learning (Zhang et al., 2022): A demonstration candidate is predefined, and Q-learning is utilized to learn the retrieval policy. Demonstrations are clustered into 50 clusters to implement this algorithm.  
+- Static: Demonstrations are pre-defined and kept static.  
+- BM25 (Robertson et al., 2009): Demonstrations retrieved using the BM25 algorithm.  
+- Random: Demonstrations randomly sampled for each user query.  
+- KNN (Guo et al., 2003): Demonstrations are retrieved using the K-nearest neighbor (KNN) algorithm based on the user query. We use the same feature embedding as our method to retrieve the demonstrations with cosine similarity in KNN.
+
+# B Experimental Setup
+
+Dataset: In the experiments, we leveraged a meticulously curated dataset tailored specifically for assessing the efficacy of demonstration retrieval systems. This dataset is derived from high-quality human (expert)-labeled data collected daily over the recent 1.5-year period. Each sample in the dataset consists of a user query along with associated infor
+
+mation about recommended advertisements. This information includes query keywords, ad titles, ad descriptions, ad URLs, and other pertinent content, each labeled as either relevant or irrelevant.
+
+To facilitate a robust evaluation, we partitioned the dataset temporally into three distinct subsets: an example pool, a training set, and a test set. The example pool contains all 1,578,728 samples as demonstrations, ensuring a comprehensive range of instances. For the purpose of training the model, we selected a subset of 9,999 samples specifically for query-ads pair labeling. The evaluation phase was carried out on a test set, which included 1,986 samples also designated for query-ads pair labeling. This temporal division helps in mimicking real-world scenarios where models are trained on historical data and tested on recent, unseen data, thereby providing insights into the practical applicability and performance of the retrieval methods under study.
+
+Evaluation Metric: Our in-context learning method aims to enhance the labeling performance of large language models (LLMs). Given that the labeling task at hand is a binary classification problem, we evaluate the effectiveness of our approach using several key metrics. Specifically, we measure Accuracy (ACC), F1-Score, Precision, and Recall. These metrics collectively provide a comprehensive assessment of the model's performance in terms of both its ability to correctly label data and its balance between precision and recall.
+
+Computational Resource All experiments are performed on single Ubuntu 20.04 LTS system with Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz CPU, 112 Gigabyte memory and single NVIDIA Tesla P100 accelerator.

2025/ICL-Bandit_ Relevance Labeling in Advertisement Recommendation Systems via LLM/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f854ae584d2f72ca259f48586334593741fe9054f5db15f92eb253761d0912bb
+size 414886

2025/ICLER_ Intent CLassification with Enhanced Reasoning/80ea3d72-543b-4c68-9989-ef5d6936a0b2_content_list.json

@@ -0,0 +1,2153 @@
+[
+    {
+        "type": "text",
+        "text": "ICLER: Intent Classification with Enhanced Reasoning",
+        "text_level": 1,
+        "bbox": [
+            206,
+            90,
+            789,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Dezheng Gao<sup>1</sup>, Xiaozheng Dong<sup>2</sup>, SHuangtao Yang<sup>3</sup>, and Bo Fu<sup>4</sup>",
+        "bbox": [
+            231,
+            129,
+            763,
+            147
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Beijing University Of Technology, gaodz@mails.bjut.edu.cn",
+        "bbox": [
+            225,
+            159,
+            771,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Knowdee Intelligence, dongxz@knowdee.com",
+        "bbox": [
+            304,
+            177,
+            692,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$ Knowdee Intelligence, yangst@knowdee.com",
+        "bbox": [
+            304,
+            193,
+            692,
+            209
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{4}$ Knowdee Intelligence, fubo@knowdee.com",
+        "bbox": [
+            314,
+            210,
+            680,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In recent years, intent classification technology based on In-Context Learning (ICL) has made significant progress. However, when applied to enterprise vertical domains, existing methods are inadequate in identifying micrograined intentions. This study identifies two primary causes of errors in data analysis: (1) Incorrect instance retrieval, often due to embedding models' limitations in capturing subtle sentence-level information in business scenarios (such as entity-related or phenomenon-specific details) (2) Insufficient reasoning ability of Large Language Models (LLMs), which tend to rely on surface-level semantics while overlooking deeper semantic associations and business logic, leading to misclassification. To address these issues, we propose ICLER, an intent classification method with enhanced reasoning. This method first optimizes the embedding model by introducing a reasoning mechanism to enhance its ability to capture fine-grained sentence-level information. Then, this mechanism is incorporated into the ICL framework, maintaining computational efficiency while significantly enhancing intent recognition accuracy. Experimental results demonstrate that ICLER significantly outperforms the original ICL method in intent identification within vertical domains. Moreover, it yields accuracy improvements of  $0.04\\%$  to  $1.14\\%$  on general datasets and its fine-tuned embedding model achieves an average performance gain of  $5.56\\%$  on selected classification tasks in the MTEB benchmark. https://github.com/gaodz-111/ICLER.git",
+        "bbox": [
+            144,
+            288,
+            460,
+            785
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            799,
+            258,
+            814
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Natural Language Processing (NLP) has witnessed remarkable advancements through large-scale pretrained language models, which demonstrate exceptional performance across various tasks via In-Context Learning(ICL). ICL leverages analogical learning principles, enabling models to iden",
+        "bbox": [
+            112,
+            825,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tify patterns from provided examples and generate predictions for few-shot classification tasks. Unlike traditional supervised learning approaches, this paradigm eliminates the need for parameter updates, enabling direct implementation on pretrained LLMs while significantly reducing computational costs and training time. Despite these impressive achievements, the critical role of reasoning capabilities in intent classification tasks based on ICL remains underexplored.",
+        "bbox": [
+            507,
+            261,
+            884,
+            420
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "When confronted with highly complex scenarios in vertical domains, the ICL-based intent classification framework tends to reveal two critical shortcomings with greater prominence. As shown on the left side of Figure 1, the example illustrates that embedding models are difficult to recall valid examples because they lack domain-specific semantic adaptation optimization, and their representation space not being able to effectively distinguish fine-grained intent differences. Specifically, if the sample library has too few samples of the \"camera failure\" category or if they are mostly vague expressions (such as \"It can't take pictures anymore\"), the embedding model may overemphasize generic phrases like \"not working\" while overlooking critical domain-specific entities such as \"camera\". This limitation not only leads to misclassification but also highlights the embedding model's inability to capture domain-specific nuances.",
+        "bbox": [
+            507,
+            423,
+            884,
+            741
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "On the right side of Figure 1, LLMs exhibit a tendency to disproportionately focus on the action verb \"turn on\" and the generic noun \"computer\" when processing extended textual inputs. This observed pattern suggests that it is difficult for LLMs to achieve the mapping extension from surface semantics to deep semantics, for example, the causal relationship between \"Pressed the power button but there was no response\" and \"The computer won't turn on\", and thus cannot complete the reasoning of users' actual intent through ICL-based",
+        "bbox": [
+            507,
+            744,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "3031",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 3031-3044",
+        "bbox": [
+            216,
+            945,
+            779,
+            957
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            280,
+            959,
+            715,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/5eff723a45e86362f3ce755998a66cbb0a9d08d501a8e3aa514a4048e36ec395.jpg",
+        "image_caption": [
+            "Figure 1: The example on the left side of the figure shows that the examples retrieved based on similarity by the embedding model are not all valid. The example on the right side of the figure indicates that LLM may have issues recognizing multiple intentions when lacking reasoning ability, and it may also have recognition errors due to subtle differences in sentences."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            129,
+            85,
+            430,
+            354
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/e4520b181e6dd663c9393f782a4fb4832ff1c66815ec27483f56726261da494a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            450,
+            86,
+            867,
+            354
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "intent classification framework, ultimately leading to a significant reduction in the accuracy of intent recognition. Therefore, text classification requires not only the model to understand the surface semantics of the text but also to perform deep logical reasoning.",
+        "bbox": [
+            112,
+            457,
+            487,
+            554
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges in ICL-based intent classification, we propose ICLER. First, we implement reasoning-driven vector representation enhancement, employing domain knowledge integration to help embedding models capture complex semantic patterns in business scenarios, thereby improving instance retrieval quality. Second, we strengthen intent understanding through reasoning-augmented analysis, enhancing LLMs' capability to differentiate between semantically proximate intents. Experimental results demonstrate that ICLER significantly improves classification accuracy and robustness in complex business environments, effectively bridging the gap between surface semantics and deep intent reasoning.",
+        "bbox": [
+            112,
+            558,
+            489,
+            799
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our paper contributions are:",
+        "bbox": [
+            132,
+            803,
+            347,
+            819
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- In terms of vector representation, we use inference optimization techniques to enable embedding models to better capture complex semantic information in the business scenario, thus improving the quality of example recalls.",
+            "- We propose a new method to introduce the in"
+        ],
+        "bbox": [
+            134,
+            822,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ference process into the ICL-based text classification framework, which effectively combines domain knowledge and multi-step inference and enhances the reasoning ability of LLM.",
+        "bbox": [
+            544,
+            457,
+            884,
+            536
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "- We conduct comprehensive experiments on both vertical-domain and general-purpose datasets to validate the effectiveness and broad applicability of our method.",
+        "bbox": [
+            531,
+            539,
+            882,
+            604
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Chapter 2 will review the research work, Chapter 3 will detail the theoretical basis of the method, Chapter 4 will present the experimental results, Chapter 5 will summarize the conclusions, and the last chapter will analyze the limitations of the method.",
+        "bbox": [
+            507,
+            605,
+            882,
+            700
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            713,
+            663,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 In-Context learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            739,
+            712,
+            755
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As an emerging natural language processing paradigm (Brown et al., 2020), ICL has received wide attention in recent years. The core concept revolves around enabling LLMs to directly acquire task-specific patterns through minimal in-context instances, eliminating the requirement for explicit parameter updates. Recent studies explain the ICL mechanism from the perspectives of Bayesian inference (Xie et al., 2022) and gradient descent (Yang et al., 2023), and believe that LLMs realize",
+        "bbox": [
+            507,
+            760,
+            882,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "3032",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "context learning through implicit optimization or pattern matching, which provides theoretical support for the interpretability of this paradigm.",
+        "bbox": [
+            112,
+            84,
+            490,
+            131
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Meanwhile, ICL performance strongly depends on the prompt structure, including selection of instances, ordering of instances and formatting of instructions (Zhao et al., 2021; Lu et al., 2022). The k-nearest neighbor search KATE-based method (Liu et al., 2022) improved the effect of ICL in sentiment analysis, form to text generation and question and answer. Order sensitivity is mitigated with increasing complexity (Liu et al., 2024). In the research of improving the task processing ability of large language model, the researchers developed the task description automatic generation framework (Zhou et al., 2023), which optimized the task description quality through automatic generation mechanism, and the research team systematically constructed the thinking chain prompt mechanism (Wei et al., 2022). This innovative method significantly improves model performance in complex reasoning tasks by explicitly modeling reasoning paths.",
+        "bbox": [
+            115,
+            133,
+            489,
+            454
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Embedding Models",
+        "text_level": 1,
+        "bbox": [
+            112,
+            464,
+            314,
+            481
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The embedding model maps semantic information into a high-dimensional vector space via text embedding techniques. Significant progress has been made in this field, with models such as SBERT (Reimers and Gurevych, 2019), which enhances sentence embeddings through siamese networks; SimCSE (Gao et al., 2021), which improves contrastive learning for sentence embeddings; Contriever (Izacard et al., 2022), which leverages contrastive learning for unsupervised retrieval; and Multilingual E5 (Wang et al., 2024), which demonstrates strong performance in multilingual and multi-task scenarios.",
+        "bbox": [
+            112,
+            486,
+            489,
+            694
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The development of embedding models has been further accelerated by the large-scale text embedding benchmark (MTEB) (Muennighoff et al., 2023), which provides a comprehensive evaluation framework for comparing and improving embedding techniques. Building on these foundations, recent studies have introduced innovative approaches to address specific challenges—particularly the effectiveness of instruction-based embedding frameworks in cross-task adaptation. Notably, Asai's team (Asai et al., 2022) established foundational evidence that semantic instruction constraints can effectively align query intent with document functionality. Building on this theoretical basis, Su's",
+        "bbox": [
+            112,
+            696,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "team (Su et al., 2023) subsequently developed the INSTRUCTOR framework, which operationalizes this principle by encoding both task descriptions and domain features as natural language instructions. This implementation achieves task-adaptive embeddings while eliminating the need for resource-intensive fine-tuning processes.",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Additionally, the GTE model (Li et al., 2023) employs a multi-stage contrastive learning framework, demonstrating exceptional generalization capabilities, particularly in unidirectional quantity embedding tasks. More recently, the M3-Embedding model (Chen et al., 2024) has achieved breakthroughs in multi-language retrieval, input granularity processing, and retrieval function unification. By leveraging self-knowledge distillation, efficient batch processing algorithms, and high-quality multimodal datasets, this model significantly improves the semantic representation quality of embedded vectors.",
+        "bbox": [
+            507,
+            198,
+            885,
+            407
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Reasoning ability of LLMs",
+        "text_level": 1,
+        "bbox": [
+            507,
+            420,
+            764,
+            436
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The reasoning ability of LLMs has been significantly enhanced by techniques that simulate the human problem-solving process. A major breakthrough in this field is the Chain-of-Thought (CoT) (Wei et al., 2023), which prompts LLMs to solve complex problems by breaking them down into multiple reasoning steps.",
+        "bbox": [
+            507,
+            443,
+            882,
+            556
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Recently, LLMs with complex reasoning capabilities, such as OpenAI's GPT-4/4o (OpenAI et al., 2024), DeepSeek-R1 (DeepSeek-AI et al., 2025), and Google's Gemini 2.0, have significantly validated the effectiveness of systematic reasoning methods in practical applications. These advanced models employ a human-like \"slow thinking\" mechanism when dealing with higher-order cognitive tasks: they first generate a systematic reasoning process through a multi-stage cognitive processing path, and then deduce the final solution. This hierarchical problem-solving paradigm enables LLMs to perform close to human experts in complex areas such as program code generation and mathematical theorem proof.",
+        "bbox": [
+            507,
+            558,
+            885,
+            797
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            507,
+            813,
+            611,
+            828
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 illustrates the processing pipeline of ICLER. This method processes query by retrieving relevant examples from the Instance Database  $\\mathcal{D}$  and logical analysis results from the Reasoning Database  $\\mathcal{R}$  through a dual-channel retrieval mech",
+        "bbox": [
+            507,
+            841,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "3033",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/d06008fb8ffc47ce34ad46f1c8aa58eb6f6574d917bc77ad733ce0ed5d5ee2fd.jpg",
+        "image_caption": [
+            "Figure 2: Processing pipeline of ICLER: Embedding Model is a model jointly trained based on reasoning and embedding tasks, and Reasoning database is an error analysis generated based on strong negative cases."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            121,
+            85,
+            880,
+            350
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "anism. The retrieved demonstration samples are integrated with a predefined prompt template to construct a structured input sequence, which is then fed into the LLM for intent recognition analysis.",
+        "bbox": [
+            112,
+            443,
+            489,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To further improve the performance of this process, we employ an embedding model that is optimized through a multi-task learning framework. Not only does this framework maintain basic semantic representation capabilities, but it also introduces joint training of reasoning-generation tasks and vector-optimization tasks. Specifically, the reasoning-generation task enhances the model's ability to infer logical relationships, while the vector-optimization task improves the quality of its semantic representations. As a result, ICLER significantly boosts the model's capacity to capture fine-grained features in vertical domains, thereby leading to higher classification accuracy in complex semantic scenarios.",
+        "bbox": [
+            112,
+            513,
+            489,
+            753
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To support this framework, the following technical components are developed: (1) Build Retrieval Database and Fine-tuning Dataset: a Reasoning Database that stores structured logical analysis results for retrieval and an Instance Database to retrieve related instances. Then, build a Fine-tuning Dataset with the corresponding database; (2) Multi-task Training Pipeline: a multi-task training pipeline that optimizes embedding quality through joint learning of reason",
+        "bbox": [
+            112,
+            760,
+            490,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ing and representation tasks; (3) Results Obtain: an intent classification module that integrates improved embeddings and logical reasoning for accurate intent recognition.",
+        "bbox": [
+            507,
+            443,
+            884,
+            506
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Build Retrieval Database and Fine-tuning Dataset",
+        "text_level": 1,
+        "bbox": [
+            507,
+            519,
+            882,
+            550
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1.1 Build Retrieval Database",
+        "text_level": 1,
+        "bbox": [
+            507,
+            558,
+            766,
+            571
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We designed a concise and efficient data synthesis process that use the reasoning power of LLM to generate Reasoning Database. First, we trained an embedding model based on the SetFit framework (Tunstall et al., 2022). The training set is trained by five-fold cross-validation, and the divided test set is also identified to generate a strong negative case data set  $W = \\{w_{i} = \\text{(case, true_label, false_label)}\\}$ , where each sample contains a case, a correct label and an error label. This data is then combined with the corresponding prompt:",
+        "bbox": [
+            505,
+            577,
+            884,
+            771
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} P _ {R i} = \\operatorname {T a k} \\left\\{B _ {-} \\operatorname {I n s} \\right\\}; \\\\ \\text {E x a m p l e s} \\{B _ {-} E \\}; \\\\ Q u e r y \\{w _ {i} \\}, \\\\ r _ {i} = L L M _ {r e a s o n i n g} (P _ {R i}), \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            613,
+            783,
+            791,
+            860
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Among them,  $B\\_Ins$  is used to clearly explain the requirements of the thinking task. Example  $B\\_E$  is used to show the form of the desired out",
+        "bbox": [
+            507,
+            873,
+            884,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "3034",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "put to guide the model to generate compliant responses.  $w_{i}$  represents the input specific case for model generation reasoning. The designed  $P_{R_i}$  (A) is input into LLM to generate the reasoning results. These results will be used to construct the Reasoning Database  $\\mathcal{R}$  and provide high-quality reasoning data support for subsequent tasks.",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training corpora are processed into fixed-dimensional vectors with semantic representations through a embedding model, and the normalized text embedding vectors along with their supervised labels are subsequently stored in an Instance Database  $\\mathcal{D}$ .",
+        "bbox": [
+            112,
+            197,
+            489,
+            292
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.1.2 Build Fine-tuning Dataset",
+        "text_level": 1,
+        "bbox": [
+            112,
+            300,
+            378,
+            317
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The model is trained to respond to instructions by generating answers. We need to combine reasoning and its corresponding  $P_R$  construction to generate the task dataset  $T_1$ .",
+        "bbox": [
+            112,
+            319,
+            489,
+            384
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT _ {I} = \\left\\{\\left\\{t e x t: \\left[ P _ {R i}, r _ {i} \\right] \\right\\}, \\dots \\right\\}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            395,
+            487,
+            413
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Where  $r_i \\in \\mathcal{R}$ , is the prompt to generate in 3.1.1. And at the beginning of the training, the data set  $T_1$  will be task marked, the specific marker form is as follows: generate the task tag:  $<s><|user|>\\{...$ $|\\text{assistant}|>\\{\\} < /s>$ .",
+        "bbox": [
+            112,
+            420,
+            487,
+            501
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, in order to meet the representational instruction tuning, we need to combine the embedded instructions with the training data to build the data set  $T_{2}$ .",
+        "bbox": [
+            112,
+            502,
+            489,
+            565
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nT _ {2} = \\left\\{ \\begin{array}{c} \\left\\{\\text {q u e r y}: [ Q \\_ I n s, q _ {\\text {t r a i n} i} ], \\right. \\\\ \\quad \\text {p o s}: \\left[ \\left[ P \\_ I n s, p _ {i 1} \\right], \\left[ P \\_ I n s, p _ {i 2} \\right], \\dots \\right], \\\\ \\quad \\text {n e g}: \\left[ \\left[ N \\_ I n s, n _ {i 1} \\right], \\left[ N \\_ I n s, n _ {i 2} \\right], \\dots \\right] \\\\ \\}, \\dots \\end{array} \\right\\} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            573,
+            487,
+            663
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "$q_{train_i} \\in \\mathcal{D}$  and  $\\{p_{ij}\\}$  is a set of sentences with the same composition as  $q_{train_i}$ 's label, and  $\\{n_{ik}\\}$  is a set of sentences with different intents from  $q_{train_i}$ 's label (to reduce the scale of  $T_2$ , we optimized the construction process. Same label: extract a sentence in order to form a positive sample set; Different labels: extract one sentence from each label to form a negative sample set).  $Q_{-}\\text{Ins}, P_{-}\\text{Ins},$  and  $N_{-}\\text{Ins}$  are the embedding instructions corresponding to the query  $q_{train_i}$ , the positive instance set  $\\{p_{ij}\\}$ , and the negative instance set  $\\{n_{ik}\\}$ , respectively. At the same time, study has confirmed that the above three embedding instructions can be the same one, So we use emb_Instruction = \"Represent the example for the following task: Given a question about computer-related",
+        "bbox": [
+            112,
+            664,
+            490,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "issues or casual conversation, it is necessary to determine the intent of the given question in order to better address the user's issue.\" as the instruction corresponding to all instances.",
+        "bbox": [
+            507,
+            84,
+            884,
+            148
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "At the beginning of the training, the dataset will be task labeled, the specific marker form is as follows: embedding task tag:  $<s><|user|>\\{emb_Instruction\\}<|embed|\\} \\{q_{traini}$  or  $p_{ij}$  or  $n_{ik}\\}$",
+        "bbox": [
+            507,
+            149,
+            884,
+            215
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2 Multi-task Training Pipeline",
+        "text_level": 1,
+        "bbox": [
+            507,
+            223,
+            781,
+            239
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We used the GRIT architecture (Muennighoff et al., 2025) to train with the contrast loss function and the generation loss function, and used the ability of the generation task optimization embedding model to represent fine-grained features. Taking a similar approach, we let the embedding model fine-tune the instructions of the corresponding task according to the different task markers.",
+        "bbox": [
+            507,
+            243,
+            884,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = - \\frac {\\lambda_ {e m b}}{M} \\sum_ {i = 1} ^ {M} \\log \\frac {\\sum_ {j = 1} ^ {A} \\exp (\\tau \\sigma (f _ {\\theta} (q _ {i}) , f _ {\\theta} (d _ {i j})))}{\\sum_ {k = 1} ^ {A + B} \\exp (\\tau \\sigma (f _ {\\theta} (q _ {i}) , f _ {\\theta} (d _ {i k})))} \\\\ - \\frac {\\lambda_ {g e n}}{N} \\sum_ {i = 1} ^ {N} \\log P \\left(f _ {\\theta , \\eta} \\left(x _ {i}\\right) \\mid f _ {\\theta , \\eta} \\left(x _ {(<   i)}\\right)\\right), \\tag {4} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            509,
+            376,
+            889,
+            512
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Here, M indicates the embedding task batch and N indicates the generation task batch.  $f_{\\theta}$  denotes an embedding model parameterized by  $\\theta$  while  $f_{\\theta,\\eta}$  refers to the architecture augmented with an  $\\eta$ -parameterized language modeling head for generation tasks. The temperature hyperparameter is denoted as  $\\tau$ , and  $\\sigma$  signifies the process of applying a pooling operation to each output embedding followed by cosine similarity computation. Given a query  $q_{i}$ , we define its positive sample set of size A and its negative sample set of size B,  $d_{ij}$  represents the relevant positive examples, and  $d_{ik}$  represents the relevant positive and negative examples. For generation tasks,  $x_{i}$  indicates the i-th token in the sequence, with  $x_{(<i)}$  encompassing all preceding tokens. The training objective is to minimize the negative log-likelihood loss exclusively over the tokens ( $\\{\\text{response}\\} < /s>$ ). Based on experimental results identifying  $(\\lambda_{gen}, \\lambda_{emb}) = (2,1)$  as optimal, we initialize the loss function accordingly.",
+        "bbox": [
+            507,
+            518,
+            884,
+            841
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Intent Classification",
+        "text_level": 1,
+        "bbox": [
+            507,
+            848,
+            714,
+            864
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 Information Retrieval",
+        "text_level": 1,
+        "bbox": [
+            507,
+            870,
+            741,
+            884
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Incorporating reasoning support and a domain-adaptive retrieval embedding model, we clas",
+        "bbox": [
+            507,
+            889,
+            884,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "3035",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "sify intents via the designed pipeline. We prompt the LLM with the instruction tuple  $I\\_Ins, r_i^*, d_i^*, q_{testi}$  (where  $q_{testi}$  is drawn from the test set), and retrieve the associated reasoning  $r_i^*$  and instances  $d_i^*$ . Retrieval then uses embedding similarity between the user query and each  $d_i^*$  to assess relevance. Given a query  $q_{testi}$ , Instance Database  $\\mathcal{D}$ , and Reasoning Database  $\\mathcal{R}$ , the embedding model  $M$  encodes them as  $v_{q_i}, V_D, V_R$ :",
+        "bbox": [
+            112,
+            84,
+            489,
+            231
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nv _ {q _ {i}} \\leftarrow M \\left(q _ {\\text {t e s t} i}\\right), V _ {D} \\leftarrow M (\\mathcal {D}), V _ {R} \\leftarrow M (\\mathcal {R}), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            121,
+            265,
+            487,
+            298
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To retrieve related instances  $d_{i}^{*}$  and related reasoning  $r_i^*$  from a large-scale dataset, we used the graph-based approximate nearest neighbor search algorithm Hierarchical Navigable Small Worlds(HNSW)(Malkov and Yashunin, 2020). The relevant instances need to meet the following nearest-neighbor criteria:",
+        "bbox": [
+            112,
+            332,
+            489,
+            447
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} d _ {i} ^ {*} = \\arg \\min  _ {d _ {i} ^ {*} \\in \\mathcal {D}} \\left\\{d i s (v _ {q _ {i}}, V _ {D}) \\right\\}, \\\\ r _ {i} ^ {*} = \\arg \\min  _ {r _ {i} ^ {*} \\in \\mathcal {R}} \\left\\{d i s \\left(v _ {q _ {i}}, V _ {R}\\right) \\right\\}, \\tag {6} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            479,
+            485,
+            533
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Here, the function  $dis()$  denotes the cosine distance computed by the pre-trained embedding model  $M$ . First, pre-process the Reasoning Dataset  $\\mathcal{R}$  and the Instance dataset  $\\mathcal{D}$  in the reasoning library, plus the embedded instruction tag:  $<|user|>\\{\\} <|embed|>\\{\\mathcal{D},\\mathcal{R}\\}$ . (Note that  $\\mathcal{D}$  and  $\\mathcal{R}$  are already included in the training data, so no extra instructions are needed for embedding.). Next, for each user query  $q_{test_i}$ , we add a task tag:  $<|user|>\\{I_Ins\\}<|embed|>\\{q_{test_i}\\}$ . ( $I_Ins:Given$  a sentence, return sentences that are semantically similar to it.) The processed data is divided into the embedding model M to obtain the last hidden state of the model, and the vector representation is obtained using mean pooling aggregation.",
+        "bbox": [
+            112,
+            576,
+            489,
+            819
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Finally, the similarity between the embedding vectors is obtained by applying the HNSW search algorithm, quickly retrieve and returning the most relevant TOP-N instances  $d_{i}^{*}$ . Meanwhile, we select the TOP-1 reason  $r_{i}^{*}$  that best matched the query for intent judgment.",
+        "bbox": [
+            112,
+            824,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 Results Obtain",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            687,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Construct a  $P_{C_i}$  for classification and input it into an LLM for classification:",
+        "bbox": [
+            507,
+            105,
+            882,
+            136
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP _ {C i} = \\operatorname {T a s k} \\left\\{C \\_ I n s \\right\\};\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            600,
+            149,
+            764,
+            167
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {R e a s o n} \\left\\{r _ {i} ^ {*} \\right\\};\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            652,
+            168,
+            754,
+            186
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\text {E x a m p l e s} \\left\\{d _ {i} ^ {*} \\right\\}; \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            652,
+            189,
+            882,
+            206
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nQ u e y \\left\\{q _ {t e s t i} \\right\\},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            652,
+            209,
+            759,
+            225
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {r e s u l t} _ {i} = L L M _ {\\text {c l a s s i f i c a t i o n}} \\left(P _ {C _ {i}}\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            229,
+            811,
+            244
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Where Ins represents the input intent classification task instruction, reasoning data  $r_i^*$  is used to provide the analysis process, instance data  $d_i^*$  provides the relevant instances, and  $q_{test_i}$  is the user's query. Then, the corresponding  $P_{Ci}$  (B) is input into the LLM used for classification, and the classification result  $result_i$  is obtained. In this mechanism, the LLM strictly selects the corresponding labels from the examples, and does not generate new labels, ensuring the consistency and predictability of the process.",
+        "bbox": [
+            507,
+            259,
+            884,
+            436
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            451,
+            648,
+            468
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Dataset and Evaluation Method",
+        "text_level": 1,
+        "bbox": [
+            507,
+            479,
+            805,
+            494
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the performance of ICLER in general scenarios, we used multiple datasets: the widely adopted multilingual Amazon Review corpus (MARC) including its en-US, zh-CN, and ja-JP subsets (Keung et al., 2020), which contain cross-lingual user reviews to support the model's multilingual text processing; the bankingIntent dataset (focused on bank-related intent identification, critical for assessing performance in the financial domain); the mistral-intent-data1816 dataset (covering diverse scenarios to comprehensively test intent recognition capabilities); and a self-constructed PC domain dataset (with over 500 labels and industrial-domain intent classification characteristics). Comprehensive evaluations across these datasets confirm ICLER's high accuracy and reliability in varied scenarios, with Top-1 Accuracy adopted as the primary evaluation metric.",
+        "bbox": [
+            505,
+            501,
+            884,
+            804
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In the recall task, Top-1 Accuracy represents the proportion of the first example intention of the embedding model recall consistent with the user query; in the classification task, Top-1 Accuracy represents the proportion of the results generated by the reasoning model consistent with the user query.",
+        "bbox": [
+            507,
+            809,
+            882,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "3036",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/ea78ad8aeeaf1c93a3cbbba8217a5d2c46e52d8859ab76f115c72e0bad3a838d.jpg",
+        "table_caption": [
+            "Table 1: Evaluation results of employing Multilingual-E5-Base and Qwen1.5-1.8B for embedding, with Qwen2.5-7B handling reasoning tasks. Evaluation uses Top-1 Accuracy(%) as the primary metric."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td rowspan=\"2\">Method</td><td colspan=\"6\">Top-1 Accuracy (%)</td></tr><tr><td>pc</td><td>banking</td><td>en</td><td>ja</td><td>zh</td><td>mistral</td></tr><tr><td rowspan=\"3\">Multilingual-E5</td><td>Base</td><td>33.72</td><td>66.23</td><td>67.99</td><td>59.72</td><td>65.27</td><td>67.05</td></tr><tr><td>SetFit</td><td>56.08</td><td>83.18</td><td>81.98</td><td>77.34</td><td>76.70</td><td>73.86</td></tr><tr><td>Emb+Reason</td><td>61.15</td><td>90.68</td><td>80.70</td><td>78.28</td><td>78.41</td><td>75.00</td></tr><tr><td rowspan=\"3\">Qwen1.5-1.8B</td><td>Base</td><td>28.59</td><td>63.28</td><td>67.18</td><td>66.95</td><td>68.70</td><td>56.00</td></tr><tr><td>Emb-only</td><td>66.76</td><td>89.84</td><td>85.54</td><td>81.24</td><td>83.15</td><td>79.55</td></tr><tr><td>Emb+Reason</td><td>67.81</td><td>92.05</td><td>86.08</td><td>83.96</td><td>84.53</td><td>81.82</td></tr><tr><td rowspan=\"2\">Qwen2.5-7B</td><td>ICL</td><td>81.44</td><td>92.21</td><td>87.79</td><td>84.36</td><td>85.14</td><td>84.09</td></tr><tr><td>ICLER</td><td>86.21</td><td>93.00</td><td>87.83</td><td>85.04</td><td>85.94</td><td>85.23</td></tr></table>",
+        "bbox": [
+            115,
+            141,
+            884,
+            296
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            319,
+            225,
+            336
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Given strict time constraints, we constructed a recall system based on an embedding model, selecting two base models—Multilingual-E5-Base and Qwen1.5-1.8B-Chat—as embedding models. For data preparation, we utilized Qwen2.5-72B-instruct to generate high-quality reasoning content; for intent recognition, we adopted Qwen2.5-7B-Instruct. Training was conducted on a high-performance A100 GPU, with a batch size of 4 to balance data throughput and computational efficiency, a learning rate of 1e-5, and a fixed temperature parameter of 0.02 (the latter maintains high certainty in text generation to reduce stochastic noise). We also set maximum lengths of 256 for both queries and paragraphs to optimize long-text processing and avoid computational waste from overly long inputs. For classification, we reference the Top-5 instances and Top-1 reasoning.",
+        "bbox": [
+            112,
+            340,
+            490,
+            630
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Main Result",
+        "text_level": 1,
+        "bbox": [
+            112,
+            640,
+            258,
+            653
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3.1 Validation Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            661,
+            317,
+            674
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 3, we addressed the issue of insufficient LLM reasoning ability proposed in the Introduction section. By introducing the reasoning mechanism, ICLER achieves a significant performance improvement in the intention classification task.",
+        "bbox": [
+            112,
+            678,
+            487,
+            774
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Simultaneously, Figure 4 reveals the key pathways to improving the embedding model's fine-grained intent recognition capabilities: Conventional training strategies yield nearly identical attention weights for domain-critical entities (camera) and generic descriptors (not working), resulting in confused semantic representations. However, our joint training paradigm integrating reasoning generation and vector optimization tasks re",
+        "bbox": [
+            112,
+            776,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/e42afce435356ac0f2aca8dbc33f6dcb49ef44c06f85599c78ef3284b315b14f.jpg",
+        "image_caption": [
+            "Figure 3: Results of the ICLER. The picture shows that through the given analysis process related to the user query."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            318,
+            882,
+            577
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "calibrates this attention distribution, amplifying focus on technical entities (camera: 0.73 attention weight) while suppressing generic terms. This optimized attention mechanism enables precise mapping of user queries to specialized technical scenarios.",
+        "bbox": [
+            507,
+            652,
+            884,
+            747
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3.2 Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            757,
+            741,
+            772
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In the evaluation, the performance of the embedding models is summarized in Table 1.",
+        "bbox": [
+            507,
+            776,
+            882,
+            808
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "All experimental results in Table 1 are \"domain evaluations\"—specifically, the model is fine-tuned on one dataset (e.g., PC, Banking, Amazon Reviews) and then evaluated on the test set of the same dataset. As shown, multi-task fine-tuning of the embedding model yields absolute accuracy improvements of 2.72 points on ja-JP and 1.38",
+        "bbox": [
+            507,
+            809,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "3037",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/f8f867a1a67b5381889de814dd1eb2282fa5a26c746eef2228f0c8225be39270.jpg",
+        "image_caption": [
+            "Figure 4: Attention-heatmap. The Qwen1.5-1.8B-Chat model is trained based on the pc dataset, and the attention mechanism visualization is studied for the user query \"camera is not working.\""
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            83,
+            487,
+            149
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/a76d2c30d5652da7e8c10e84b5c710be8bd220ee57591f50a38a6fd92ad380e8.jpg",
+        "table_caption": [
+            "Table 2: Top1-Accuracy(‰) of GTE-Qwen2-1.5B-Instruct Model on MTEB Classification Tasks after Multi-task Fine-tuning"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Classification Task</td><td>Base</td><td>Emb+ Reason</td></tr><tr><td>ToxicConversationsClassification</td><td>82.30</td><td>82.04↓</td></tr><tr><td>MTOPDomainClassification</td><td>97.54</td><td>98.39↑</td></tr><tr><td>MassiveIntentClassification</td><td>82.41</td><td>84.05↑</td></tr><tr><td>MTOPIntentClassification</td><td>88.71</td><td>87.44↓</td></tr><tr><td>AmazonPolarityClassification</td><td>96.60</td><td>96.33↓</td></tr><tr><td>TweetSentimentExtractionClassification</td><td>72.55</td><td>71.55↓</td></tr><tr><td>MassiveScenarioClassification</td><td>84.38</td><td>89.09↑</td></tr><tr><td>Banking77Classification</td><td>87.33</td><td>88.44↑</td></tr><tr><td>AmazonCounterfactualClassification</td><td>84.04</td><td>90.12↑</td></tr><tr><td>AmazonReviewsClassification</td><td>55.24</td><td>54.70↓</td></tr><tr><td>ImdbClassification</td><td>95.71</td><td>95.22↓</td></tr><tr><td>EmotionClassification</td><td>61.05</td><td>80.43↑</td></tr></table>",
+        "bbox": [
+            115,
+            319,
+            497,
+            501
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "points on zh-CN over single-task embedding-only fine-tuning. Among the tested models, Qwen1.5-1.8B achieves the highest Top-1 recall accuracy under joint optimization of embedding and generation tasks, and is selected as the backbone embedding model. These gains reflect the effectiveness of multi-task supervision in enhancing embedding representations.",
+        "bbox": [
+            112,
+            530,
+            489,
+            657
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In addition, we evaluate our proposed ICLER framework, which integrates reasoning into the ICL-based intent classification process. Compared to the best-performing embedding-only baseline, ICLER achieves an absolute accuracy gain of 4.77 points on PC-related domain datasets. For general-domain evaluations, the maximum absolute accuracy improvement is 1.14 points, and this gain only occurs on the Mistral dataset (the improvement range of other general datasets is  $0.04\\% - 0.89\\%$ ). It is important to note that all \"performance improvements\" reported in this study refer to absolute accuracy differences, not relative percentage changes.",
+        "bbox": [
+            112,
+            661,
+            489,
+            869
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To further assess the effectiveness of our approach, we select six classification tasks from the MTEB benchmark: MTOPDomainClassification,",
+        "bbox": [
+            112,
+            873,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "MassiveIntentClassification, MassiveScenarioClassification, Banking77Classification, AmazonCounterfactualClassification, and EmotionClassification. These tasks are chosen because they involve short-text classification, which aligns well with the nature of user queries in intent recognition scenarios. For each task, corresponding reasoning traces are generated to support multi-task fine-tuning of the embedding model. We then fine-tune the GteQwen2-1.5B-instruct model with both generation and embedding objectives.",
+        "bbox": [
+            507,
+            84,
+            884,
+            261
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table2, our approach achieves consistent improvements on the target tasks. Although a slight performance drop is observed on out-of-domain tasks not included in the fine-tuning data (average absolute decrease of 0.63 points), this is outweighed by substantial gains on the target tasks (average absolute increase of 5.56 points). Overall, the multi-task fine-tuned embedding model achieves an average improvement of 2.50 points in accuracy across all evaluated MTEB classification tasks compared to the baseline.",
+        "bbox": [
+            507,
+            262,
+            884,
+            438
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.3.3 Ablation Studies",
+        "text_level": 1,
+        "bbox": [
+            507,
+            449,
+            699,
+            463
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To verify the impact of key parameters and design choices on model performance, we conducted ablation experiments focusing on three core aspects: the number of retrieved examples, the relevance of reasoning samples, and the decoupling strategy of reasoning and instances. The results are as follows:",
+        "bbox": [
+            507,
+            469,
+            882,
+            565
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(1) Number of Retrieved Examples: When retrieving the Top-5 instances, over  $97\\%$  of test samples included the correct intent in the retrieval results. Continuing to increase the number of retrieved examples (e.g., Top-10) only improved the accuracy by  $0.12\\%$ , while increasing the computational overhead of the retrieval module by  $23\\%$ . Considering the balance between performance and efficiency, we fixed the number of retrieved examples at  $\\mathrm{N} = 5$  in the final framework.",
+        "bbox": [
+            507,
+            567,
+            884,
+            726
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "(2)Relevance of Reasoning Samples: We tested the effect of adding 2-3 weakly relevant reasoning samples (i.e., reasoning content not strongly associated with the user query intent) to the prompt. The results showed that the model accuracy decreased by  $0.3\\% - 0.5\\%$  compared to using only strongly relevant samples. We speculate this is due to the \"hallucination\" phenomenon of LLMs—weakly relevant reasoning content interferes with the model's judgment of core intent. Additionally, the increase in reasoning samples extended the input sequence length by  $40\\%$ , leading to a  $15\\%$  rise in inference",
+        "bbox": [
+            507,
+            728,
+            884,
+            920
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "3038",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "time. Therefore, we ultimately chose to use only the Top-1 strongly relevant reasoning sample for intent classification.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Reason for Decoupling Reasoning and Instances: In real business scenarios, the instance retrieval module (responsible for recalling similar cases) and the reasoning generation module (responsible for generating logical analysis) usually belong to different system components, and their data sources and update cycles are independent. If the two modules are strongly bound (e.g., generating reasoning content specifically for each retrieved instance), the system design complexity increases by  $60\\%$ , and the iteration cycle of the framework is extended by 1.8 times. The decoupling design (generating reasoning data independently and combining it with retrieved instances during classification) reduces the system deployment complexity by  $30\\%$  while ensuring no significant loss in classification accuracy.",
+        "bbox": [
+            115,
+            133,
+            489,
+            405
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "These ablation results provide experimental basis for the parameter setting and framework design of ICLER, confirming that the selected key parameters and decoupling strategy are optimal for balancing performance, efficiency, and practical deployability.",
+        "bbox": [
+            112,
+            407,
+            489,
+            502
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            516,
+            245,
+            532
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, ICLER successfully improves the performance of ICL in the intention classification task. In the evaluation, our method achieves better results on all the datasets tested, demonstrating its validity and utility. We look forward to continuing to explore and optimize these methods to make a greater contribution to the development of the field of natural language processing.",
+        "bbox": [
+            112,
+            543,
+            487,
+            671
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            684,
+            220,
+            700
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Although ICLER shows significant advantages in each dataset, there are still three limitations. First, the domain generalization ability of this method has not been fully validated, the current experiments mainly focus on PC related fields and general datasets, and the migration effect in other professional fields still needs to be further explored. Secondly, the joint optimization process requires the simultaneous maintenance of generative tasks and embedding tasks, leading to increased memory usage during training. Finally, the model performance improvement still depends on the scale effect of the base model, and how to achieve a sim",
+        "bbox": [
+            112,
+            712,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "ilar optimization path in the model with a smaller number of parameters will be an important future research direction.",
+        "bbox": [
+            507,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            162,
+            606,
+            177
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Akari Asai, Timo Schick, Patrick Lewis, Xilun Chen, Gautier Izacard, Sebastian Riedel, Hannaneh Hajishirzi, and Wen tau Yih. 2022. Task-aware retrieval with instructions. Preprint, arXiv:2211.09260.",
+            "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, and 12 others. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.",
+            "Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. Preprint, arXiv:2402.03216.",
+            "DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.",
+            "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6894-6910, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised dense information retrieval with contrastive learning. Preprint, arXiv:2112.09118.",
+            "Phillip Keung, Yichao Lu, György Szarvas, and Noah A. Smith. 2020. The multilingual Amazon reviews corpus. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4563-4568, Online. Association for Computational Linguistics.",
+            "Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. Preprint, arXiv:2308.03281."
+        ],
+        "bbox": [
+            510,
+            186,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "3039",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pages 100–114, Dublin, Ireland and Online. Association for Computational Linguistics.",
+            "Sheng Liu, Haotian Ye, Lei Xing, and James Zou. 2024. In-context vectors: Making in context learning more effective and controllable through latent space steering. Preprint, arXiv:2311.06668.",
+            "Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics.",
+            "Yu A. Malkov and D. A. Yashunin. 2020. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):824-836.",
+            "Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2025. Generative representational instruction tuning. Preprint, arXiv:2402.09906.",
+            "Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. 2023. Mteb: Massive text embedding benchmark. Preprint, arXiv:2210.07316.",
+            "OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, and 262 others. 2024. Gpt-4 technical report. Preprint, arXiv:2303.08774.",
+            "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.",
+            "Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2023. One embedder, any task: Instruction-finetuned text embeddings. In Findings of the Association for Computational Linguistics: ACL 2023, pages 1102-1121, Toronto, Canada. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Lewis Tunstall, Nils Reimers, Unso Eun Seo Jo, Luke Bates, Daniel Korat, Moshe Wasserblat, and Oren Pereg. 2022. Efficient few-shot learning without prompts. Preprint, arXiv:2209.11055.",
+            "Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. Preprint, arXiv:2402.05672.",
+            "Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent abilities of large language models. Transactions on Machine Learning Research. Survey Certification.",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models. Preprint, arXiv:2201.11903.",
+            "Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2022. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations.",
+            "Zhe Yang, Damai Dai, Peiyi Wang, and Zhifang Sui. 2023. Not all demonstration examples are equally beneficial: Reweighting demonstration examples for in-context learning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13209-13221, Singapore. Association for Computational Linguistics.",
+            "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12697-12706. PMLR.",
+            "Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large language models are human-level prompt engineers. Preprint, arXiv:2211.01910."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            720
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "3040",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "System Prompt",
+        "text_level": 1,
+        "bbox": [
+            147,
+            124,
+            278,
+            141
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+        "bbox": [
+            152,
+            147,
+            665,
+            164
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "User Prompt",
+        "text_level": 1,
+        "bbox": [
+            152,
+            170,
+            258,
+            186
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Task Specification",
+        "text_level": 1,
+        "bbox": [
+            152,
+            193,
+            309,
+            208
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Given a specific case and two labels (a correct label and a wrong label), analyze the reason why the case is a correct label and why it is not a false label.",
+        "bbox": [
+            152,
+            209,
+            843,
+            240
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Requirements",
+        "text_level": 1,
+        "bbox": [
+            152,
+            241,
+            275,
+            256
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The reason for analysis is given and must be marked with  $\\langle \\text{reason} \\rangle$",
+        "bbox": [
+            152,
+            256,
+            737,
+            272
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Examples",
+        "text_level": 1,
+        "bbox": [
+            152,
+            274,
+            243,
+            288
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "case: \"I am still waiting on my card?\"",
+        "bbox": [
+            152,
+            288,
+            435,
+            305
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "True_label: card Arrival",
+        "bbox": [
+            152,
+            306,
+            337,
+            319
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "False_label: carddelivery Estimate",
+        "bbox": [
+            152,
+            322,
+            421,
+            337
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Output Examples",
+        "text_level": 1,
+        "bbox": [
+            152,
+            338,
+            305,
+            354
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "<reason>The case \"I am still waiting on my card?\" belongs to \"card Arrival\" because the speaker expresses they're still waiting for the card, indicating a concern about the card's arrival. It's not \"card Delivery Estimate\" as there's no mention of estimating delivery time.</reason>",
+        "bbox": [
+            152,
+            354,
+            845,
+            401
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Candidate Intent And The Corresponding Examples",
+        "text_level": 1,
+        "bbox": [
+            152,
+            403,
+            576,
+            418
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "case: {query}",
+        "bbox": [
+            152,
+            419,
+            258,
+            435
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "True_label: {true_label}",
+        "bbox": [
+            152,
+            435,
+            337,
+            450
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "False_label: {false_label}",
+        "bbox": [
+            152,
+            451,
+            346,
+            466
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Analyze The Reasons",
+        "bbox": [
+            152,
+            467,
+            334,
+            482
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "header",
+        "text": "A  $P_{R}$",
+        "bbox": [
+            114,
+            83,
+            179,
+            99
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "3041",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "System Prompt",
+        "text_level": 1,
+        "bbox": [
+            151,
+            124,
+            278,
+            141
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+        "bbox": [
+            152,
+            147,
+            665,
+            164
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "User Prompt",
+        "text_level": 1,
+        "bbox": [
+            152,
+            170,
+            258,
+            186
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Based on the given intents and one of its examples, you need to determine which intent the user input Content belongs to. Additional, Context is the historical user input.",
+        "bbox": [
+            151,
+            193,
+            845,
+            224
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Examples of analysis for other cases",
+        "text_level": 1,
+        "bbox": [
+            151,
+            225,
+            447,
+            240
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The case \"wake me up at this time\" belongs to \"alarm_set\" because the speaker is requesting to set an alarm for a specific time. It's not \"alarm_query\" as there's no indication of the speaker asking about the current alarm settings or status.",
+        "bbox": [
+            151,
+            241,
+            843,
+            288
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Requirements",
+        "text_level": 1,
+        "bbox": [
+            152,
+            290,
+            275,
+            305
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- DO NOT create new intent on your own, you must strictly use the intents in the examples.",
+            "- Give you with some other cases for the analysis process of determining intent, but you do not need to offer analysis in the round of intent.",
+            "- If no examples match the targeted Content, Output -1.",
+            "- Only output all index numbers of the matched intent for the targeted Content.",
+            "- IF there are multiple intents that match the targeted Content, output these index and use ’, to separate them.",
+            "- Consider the context from previous messages if the targeted Content is unclear."
+        ],
+        "bbox": [
+            151,
+            306,
+            843,
+            434
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Examples",
+        "text_level": 1,
+        "bbox": [
+            152,
+            435,
+            243,
+            450
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[case] wake me at daybreak [label] alarm_set [Index] 1",
+        "bbox": [
+            188,
+            470,
+            601,
+            487
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[case] tell me when the next alarms are for [label] alarm_query [Index] 2",
+        "bbox": [
+            189,
+            497,
+            734,
+            514
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[case] i need to set an alarm how many do i have set [label] alarm_query [Index] 3",
+        "bbox": [
+            188,
+            524,
+            801,
+            541
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[case] can you set my alarm [label] alarm_set [Index] 4",
+        "bbox": [
+            189,
+            551,
+            605,
+            567
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "[case] then tap ok [label] alarm_query [Index] 5",
+        "bbox": [
+            189,
+            577,
+            552,
+            594
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "#Content",
+        "text_level": 1,
+        "bbox": [
+            152,
+            606,
+            230,
+            619
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "alarm settings",
+        "bbox": [
+            152,
+            623,
+            258,
+            638
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Output",
+        "text_level": 1,
+        "bbox": [
+            152,
+            639,
+            223,
+            653
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "2",
+        "bbox": [
+            154,
+            656,
+            166,
+            668
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "header",
+        "text": "B  $P_{C}$",
+        "bbox": [
+            114,
+            83,
+            179,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "3042",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1 Build Reasoning Database"
+        ],
+        "code_body": "Query dataset  $Q_{train} = \\{q_{traini}\\}$   \nDocument dataset  $D = \\{d_i\\}$   \nModel for generating reasonings  $M_1$   \nModel for embedding  $M_2$   \nModel for classification  $M_3$   \nPrompt template:  $P_R, P_C$   \nSimilarity score function of Retrieval:  $dis()$   \n1. Initialize:  $\\mathcal{R} \\gets \\emptyset, V_D \\gets M_2(D)$   \n2. for each  $q_{traini} \\in Q_{train}$ :  \n3.  $V_q^i \\gets M_2(q_{traini})$   \n4.  $\\mathcal{D}_i^* = \\{d_j^*\\}_i \\gets \\arg \\min_{d_j \\in \\mathcal{D}} \\text{dis}(V_q^i, V_D^j)$   \n5.  $P_{Ci} \\gets (q_{traini}, \\mathcal{D}_i^*, \\mathcal{R}_i^* = \\emptyset)$   \n6. labelresulti  $\\gets M_3(P_{Ci})$   \n7. If labelresulti == true_label_qi:  \n8. Continue  \n9. else:  \n10.  $P_{Ri} \\gets (qi, \\text{true_label}_i, \\text{labelresult})$   \n11.  $r_i \\gets M_1(P_{Ri})$   \n12.  $\\mathcal{R} \\gets r_i$   \n13. end for  \n14. return  $\\mathcal{R}$",
+        "bbox": [
+            129,
+            141,
+            477,
+            494
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "header",
+        "text": "C Algorithm",
+        "bbox": [
+            115,
+            83,
+            243,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "3043",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Algorithm 2 Results Obtain",
+        "text_level": 1,
+        "bbox": [
+            115,
+            83,
+            326,
+            98
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Query dataset  $Q_{test} = \\{q_{testj}\\}$",
+        "bbox": [
+            134,
+            104,
+            374,
+            120
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Document dataset  $\\mathcal{D} = \\{d_k\\}$",
+        "bbox": [
+            134,
+            120,
+            356,
+            135
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Reasoning dataset  $\\mathcal{R} = \\{r_l\\}$",
+        "bbox": [
+            136,
+            136,
+            352,
+            152
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Model for embedding  $M_{1}$",
+        "bbox": [
+            134,
+            153,
+            334,
+            167
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Model for classification  $M_2$",
+        "bbox": [
+            134,
+            168,
+            347,
+            183
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Prompt template:  $P_{C}$",
+        "bbox": [
+            134,
+            185,
+            302,
+            199
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Similarity score function of Retrieval:  $dis()$",
+        "bbox": [
+            134,
+            200,
+            475,
+            216
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Initialize: Result  $\\leftarrow \\emptyset$ ,  $V_{D} \\gets M_{1}(D)$ ,  $V_{R} \\gets M_{1}(R)$",
+            "2. for each  $q_{test_i} \\in Q_{test}$ :",
+            "3.  $V_{q}^{i}\\gets M_{1}(q_{test_{i}})$",
+            "4.  $\\mathcal{D}_i^* = \\{d_j^*\\}_{i}\\gets \\underset {d_j^*\\in \\mathcal{D}}{\\arg \\min}dis(V_q^i,V_D^j)$",
+            "5.  $r_i^*\\gets \\arg \\min_{r_i^*\\in \\mathcal{R}}dis(V_q^i,V_R)$",
+            "6.  $P_{C_i} \\gets (q_{test_i}, \\mathcal{D}_i^*, r_i^*)$",
+            "7. resulti  $\\leftarrow M_2(P_{C_i})$",
+            "8. Result  $\\leftarrow$  resulti",
+            "9. end for",
+            "10. return Result"
+        ],
+        "bbox": [
+            134,
+            217,
+            559,
+            395
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "3044",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2025/ICLER_ Intent CLassification with Enhanced Reasoning/80ea3d72-543b-4c68-9989-ef5d6936a0b2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88124c23dff0fb67c14c63f5b4002a8a71bd9c51610376b5edfcb87baf82eb12
+size 930554

2025/ICLER_ Intent CLassification with Enhanced Reasoning/full.md

@@ -0,0 +1,415 @@
+# ICLER: Intent Classification with Enhanced Reasoning
+
+Dezheng Gao<sup>1</sup>, Xiaozheng Dong<sup>2</sup>, SHuangtao Yang<sup>3</sup>, and Bo Fu<sup>4</sup>
+
+<sup>1</sup>Beijing University Of Technology, gaodz@mails.bjut.edu.cn
+
+$^{2}$ Knowdee Intelligence, dongxz@knowdee.com
+
+$^{3}$ Knowdee Intelligence, yangst@knowdee.com
+
+$^{4}$ Knowdee Intelligence, fubo@knowdee.com
+
+# Abstract
+
+In recent years, intent classification technology based on In-Context Learning (ICL) has made significant progress. However, when applied to enterprise vertical domains, existing methods are inadequate in identifying micrograined intentions. This study identifies two primary causes of errors in data analysis: (1) Incorrect instance retrieval, often due to embedding models' limitations in capturing subtle sentence-level information in business scenarios (such as entity-related or phenomenon-specific details) (2) Insufficient reasoning ability of Large Language Models (LLMs), which tend to rely on surface-level semantics while overlooking deeper semantic associations and business logic, leading to misclassification. To address these issues, we propose ICLER, an intent classification method with enhanced reasoning. This method first optimizes the embedding model by introducing a reasoning mechanism to enhance its ability to capture fine-grained sentence-level information. Then, this mechanism is incorporated into the ICL framework, maintaining computational efficiency while significantly enhancing intent recognition accuracy. Experimental results demonstrate that ICLER significantly outperforms the original ICL method in intent identification within vertical domains. Moreover, it yields accuracy improvements of  $0.04\%$  to  $1.14\%$  on general datasets and its fine-tuned embedding model achieves an average performance gain of  $5.56\%$  on selected classification tasks in the MTEB benchmark. https://github.com/gaodz-111/ICLER.git
+
+# 1 Introduction
+
+Natural Language Processing (NLP) has witnessed remarkable advancements through large-scale pretrained language models, which demonstrate exceptional performance across various tasks via In-Context Learning(ICL). ICL leverages analogical learning principles, enabling models to iden
+
+tify patterns from provided examples and generate predictions for few-shot classification tasks. Unlike traditional supervised learning approaches, this paradigm eliminates the need for parameter updates, enabling direct implementation on pretrained LLMs while significantly reducing computational costs and training time. Despite these impressive achievements, the critical role of reasoning capabilities in intent classification tasks based on ICL remains underexplored.
+
+When confronted with highly complex scenarios in vertical domains, the ICL-based intent classification framework tends to reveal two critical shortcomings with greater prominence. As shown on the left side of Figure 1, the example illustrates that embedding models are difficult to recall valid examples because they lack domain-specific semantic adaptation optimization, and their representation space not being able to effectively distinguish fine-grained intent differences. Specifically, if the sample library has too few samples of the "camera failure" category or if they are mostly vague expressions (such as "It can't take pictures anymore"), the embedding model may overemphasize generic phrases like "not working" while overlooking critical domain-specific entities such as "camera". This limitation not only leads to misclassification but also highlights the embedding model's inability to capture domain-specific nuances.
+
+On the right side of Figure 1, LLMs exhibit a tendency to disproportionately focus on the action verb "turn on" and the generic noun "computer" when processing extended textual inputs. This observed pattern suggests that it is difficult for LLMs to achieve the mapping extension from surface semantics to deep semantics, for example, the causal relationship between "Pressed the power button but there was no response" and "The computer won't turn on", and thus cannot complete the reasoning of users' actual intent through ICL-based
+
+![](images/5eff723a45e86362f3ce755998a66cbb0a9d08d501a8e3aa514a4048e36ec395.jpg)  
+Figure 1: The example on the left side of the figure shows that the examples retrieved based on similarity by the embedding model are not all valid. The example on the right side of the figure indicates that LLM may have issues recognizing multiple intentions when lacking reasoning ability, and it may also have recognition errors due to subtle differences in sentences.
+
+![](images/e4520b181e6dd663c9393f782a4fb4832ff1c66815ec27483f56726261da494a.jpg)
+
+intent classification framework, ultimately leading to a significant reduction in the accuracy of intent recognition. Therefore, text classification requires not only the model to understand the surface semantics of the text but also to perform deep logical reasoning.
+
+To address these challenges in ICL-based intent classification, we propose ICLER. First, we implement reasoning-driven vector representation enhancement, employing domain knowledge integration to help embedding models capture complex semantic patterns in business scenarios, thereby improving instance retrieval quality. Second, we strengthen intent understanding through reasoning-augmented analysis, enhancing LLMs' capability to differentiate between semantically proximate intents. Experimental results demonstrate that ICLER significantly improves classification accuracy and robustness in complex business environments, effectively bridging the gap between surface semantics and deep intent reasoning.
+
+Our paper contributions are:
+
+- In terms of vector representation, we use inference optimization techniques to enable embedding models to better capture complex semantic information in the business scenario, thus improving the quality of example recalls.  
+- We propose a new method to introduce the in
+
+ference process into the ICL-based text classification framework, which effectively combines domain knowledge and multi-step inference and enhances the reasoning ability of LLM.
+
+- We conduct comprehensive experiments on both vertical-domain and general-purpose datasets to validate the effectiveness and broad applicability of our method.
+
+Chapter 2 will review the research work, Chapter 3 will detail the theoretical basis of the method, Chapter 4 will present the experimental results, Chapter 5 will summarize the conclusions, and the last chapter will analyze the limitations of the method.
+
+# 2 Related Work
+
+# 2.1 In-Context learning
+
+As an emerging natural language processing paradigm (Brown et al., 2020), ICL has received wide attention in recent years. The core concept revolves around enabling LLMs to directly acquire task-specific patterns through minimal in-context instances, eliminating the requirement for explicit parameter updates. Recent studies explain the ICL mechanism from the perspectives of Bayesian inference (Xie et al., 2022) and gradient descent (Yang et al., 2023), and believe that LLMs realize
+
+context learning through implicit optimization or pattern matching, which provides theoretical support for the interpretability of this paradigm.
+
+Meanwhile, ICL performance strongly depends on the prompt structure, including selection of instances, ordering of instances and formatting of instructions (Zhao et al., 2021; Lu et al., 2022). The k-nearest neighbor search KATE-based method (Liu et al., 2022) improved the effect of ICL in sentiment analysis, form to text generation and question and answer. Order sensitivity is mitigated with increasing complexity (Liu et al., 2024). In the research of improving the task processing ability of large language model, the researchers developed the task description automatic generation framework (Zhou et al., 2023), which optimized the task description quality through automatic generation mechanism, and the research team systematically constructed the thinking chain prompt mechanism (Wei et al., 2022). This innovative method significantly improves model performance in complex reasoning tasks by explicitly modeling reasoning paths.
+
+# 2.2 Embedding Models
+
+The embedding model maps semantic information into a high-dimensional vector space via text embedding techniques. Significant progress has been made in this field, with models such as SBERT (Reimers and Gurevych, 2019), which enhances sentence embeddings through siamese networks; SimCSE (Gao et al., 2021), which improves contrastive learning for sentence embeddings; Contriever (Izacard et al., 2022), which leverages contrastive learning for unsupervised retrieval; and Multilingual E5 (Wang et al., 2024), which demonstrates strong performance in multilingual and multi-task scenarios.
+
+The development of embedding models has been further accelerated by the large-scale text embedding benchmark (MTEB) (Muennighoff et al., 2023), which provides a comprehensive evaluation framework for comparing and improving embedding techniques. Building on these foundations, recent studies have introduced innovative approaches to address specific challenges—particularly the effectiveness of instruction-based embedding frameworks in cross-task adaptation. Notably, Asai's team (Asai et al., 2022) established foundational evidence that semantic instruction constraints can effectively align query intent with document functionality. Building on this theoretical basis, Su's
+
+team (Su et al., 2023) subsequently developed the INSTRUCTOR framework, which operationalizes this principle by encoding both task descriptions and domain features as natural language instructions. This implementation achieves task-adaptive embeddings while eliminating the need for resource-intensive fine-tuning processes.
+
+Additionally, the GTE model (Li et al., 2023) employs a multi-stage contrastive learning framework, demonstrating exceptional generalization capabilities, particularly in unidirectional quantity embedding tasks. More recently, the M3-Embedding model (Chen et al., 2024) has achieved breakthroughs in multi-language retrieval, input granularity processing, and retrieval function unification. By leveraging self-knowledge distillation, efficient batch processing algorithms, and high-quality multimodal datasets, this model significantly improves the semantic representation quality of embedded vectors.
+
+# 2.3 Reasoning ability of LLMs
+
+The reasoning ability of LLMs has been significantly enhanced by techniques that simulate the human problem-solving process. A major breakthrough in this field is the Chain-of-Thought (CoT) (Wei et al., 2023), which prompts LLMs to solve complex problems by breaking them down into multiple reasoning steps.
+
+Recently, LLMs with complex reasoning capabilities, such as OpenAI's GPT-4/4o (OpenAI et al., 2024), DeepSeek-R1 (DeepSeek-AI et al., 2025), and Google's Gemini 2.0, have significantly validated the effectiveness of systematic reasoning methods in practical applications. These advanced models employ a human-like "slow thinking" mechanism when dealing with higher-order cognitive tasks: they first generate a systematic reasoning process through a multi-stage cognitive processing path, and then deduce the final solution. This hierarchical problem-solving paradigm enables LLMs to perform close to human experts in complex areas such as program code generation and mathematical theorem proof.
+
+# 3 Method
+
+Figure 2 illustrates the processing pipeline of ICLER. This method processes query by retrieving relevant examples from the Instance Database  $\mathcal{D}$  and logical analysis results from the Reasoning Database  $\mathcal{R}$  through a dual-channel retrieval mech
+
+![](images/d06008fb8ffc47ce34ad46f1c8aa58eb6f6574d917bc77ad733ce0ed5d5ee2fd.jpg)  
+Figure 2: Processing pipeline of ICLER: Embedding Model is a model jointly trained based on reasoning and embedding tasks, and Reasoning database is an error analysis generated based on strong negative cases.
+
+anism. The retrieved demonstration samples are integrated with a predefined prompt template to construct a structured input sequence, which is then fed into the LLM for intent recognition analysis.
+
+To further improve the performance of this process, we employ an embedding model that is optimized through a multi-task learning framework. Not only does this framework maintain basic semantic representation capabilities, but it also introduces joint training of reasoning-generation tasks and vector-optimization tasks. Specifically, the reasoning-generation task enhances the model's ability to infer logical relationships, while the vector-optimization task improves the quality of its semantic representations. As a result, ICLER significantly boosts the model's capacity to capture fine-grained features in vertical domains, thereby leading to higher classification accuracy in complex semantic scenarios.
+
+To support this framework, the following technical components are developed: (1) Build Retrieval Database and Fine-tuning Dataset: a Reasoning Database that stores structured logical analysis results for retrieval and an Instance Database to retrieve related instances. Then, build a Fine-tuning Dataset with the corresponding database; (2) Multi-task Training Pipeline: a multi-task training pipeline that optimizes embedding quality through joint learning of reason
+
+ing and representation tasks; (3) Results Obtain: an intent classification module that integrates improved embeddings and logical reasoning for accurate intent recognition.
+
+# 3.1 Build Retrieval Database and Fine-tuning Dataset
+
+# 3.1.1 Build Retrieval Database
+
+We designed a concise and efficient data synthesis process that use the reasoning power of LLM to generate Reasoning Database. First, we trained an embedding model based on the SetFit framework (Tunstall et al., 2022). The training set is trained by five-fold cross-validation, and the divided test set is also identified to generate a strong negative case data set  $W = \{w_{i} = \text{(case, true_label, false_label)}\}$ , where each sample contains a case, a correct label and an error label. This data is then combined with the corresponding prompt:
+
+$$
+\begin{array}{l} P _ {R i} = \operatorname {T a k} \left\{B _ {-} \operatorname {I n s} \right\}; \\ \text {E x a m p l e s} \{B _ {-} E \}; \\ Q u e r y \{w _ {i} \}, \\ r _ {i} = L L M _ {r e a s o n i n g} (P _ {R i}), \\ \end{array}
+$$
+
+Among them,  $B\_Ins$  is used to clearly explain the requirements of the thinking task. Example  $B\_E$  is used to show the form of the desired out
+
+put to guide the model to generate compliant responses.  $w_{i}$  represents the input specific case for model generation reasoning. The designed  $P_{R_i}$  (A) is input into LLM to generate the reasoning results. These results will be used to construct the Reasoning Database  $\mathcal{R}$  and provide high-quality reasoning data support for subsequent tasks.
+
+Training corpora are processed into fixed-dimensional vectors with semantic representations through a embedding model, and the normalized text embedding vectors along with their supervised labels are subsequently stored in an Instance Database  $\mathcal{D}$ .
+
+# 3.1.2 Build Fine-tuning Dataset
+
+The model is trained to respond to instructions by generating answers. We need to combine reasoning and its corresponding  $P_R$  construction to generate the task dataset  $T_1$ .
+
+$$
+T _ {I} = \left\{\left\{t e x t: \left[ P _ {R i}, r _ {i} \right] \right\}, \dots \right\}, \tag {2}
+$$
+
+Where  $r_i \in \mathcal{R}$ , is the prompt to generate in 3.1.1. And at the beginning of the training, the data set  $T_1$  will be task marked, the specific marker form is as follows: generate the task tag:  $<s><|user|>\{...$ $|\text{assistant}|>\{\} < /s>$ .
+
+Furthermore, in order to meet the representational instruction tuning, we need to combine the embedded instructions with the training data to build the data set  $T_{2}$ .
+
+$$
+T _ {2} = \left\{ \begin{array}{c} \left\{\text {q u e r y}: [ Q \_ I n s, q _ {\text {t r a i n} i} ], \right. \\ \quad \text {p o s}: \left[ \left[ P \_ I n s, p _ {i 1} \right], \left[ P \_ I n s, p _ {i 2} \right], \dots \right], \\ \quad \text {n e g}: \left[ \left[ N \_ I n s, n _ {i 1} \right], \left[ N \_ I n s, n _ {i 2} \right], \dots \right] \\ \}, \dots \end{array} \right\} \tag {3}
+$$
+
+$q_{train_i} \in \mathcal{D}$  and  $\{p_{ij}\}$  is a set of sentences with the same composition as  $q_{train_i}$ 's label, and  $\{n_{ik}\}$  is a set of sentences with different intents from  $q_{train_i}$ 's label (to reduce the scale of  $T_2$ , we optimized the construction process. Same label: extract a sentence in order to form a positive sample set; Different labels: extract one sentence from each label to form a negative sample set).  $Q_{-}\text{Ins}, P_{-}\text{Ins},$  and  $N_{-}\text{Ins}$  are the embedding instructions corresponding to the query  $q_{train_i}$ , the positive instance set  $\{p_{ij}\}$ , and the negative instance set  $\{n_{ik}\}$ , respectively. At the same time, study has confirmed that the above three embedding instructions can be the same one, So we use emb_Instruction = "Represent the example for the following task: Given a question about computer-related
+
+issues or casual conversation, it is necessary to determine the intent of the given question in order to better address the user's issue." as the instruction corresponding to all instances.
+
+At the beginning of the training, the dataset will be task labeled, the specific marker form is as follows: embedding task tag:  $<s><|user|>\{emb_Instruction\}<|embed|\} \{q_{traini}$  or  $p_{ij}$  or  $n_{ik}\}$
+
+# 3.2 Multi-task Training Pipeline
+
+We used the GRIT architecture (Muennighoff et al., 2025) to train with the contrast loss function and the generation loss function, and used the ability of the generation task optimization embedding model to represent fine-grained features. Taking a similar approach, we let the embedding model fine-tune the instructions of the corresponding task according to the different task markers.
+
+$$
+\begin{array}{l} \mathcal {L} = - \frac {\lambda_ {e m b}}{M} \sum_ {i = 1} ^ {M} \log \frac {\sum_ {j = 1} ^ {A} \exp (\tau \sigma (f _ {\theta} (q _ {i}) , f _ {\theta} (d _ {i j})))}{\sum_ {k = 1} ^ {A + B} \exp (\tau \sigma (f _ {\theta} (q _ {i}) , f _ {\theta} (d _ {i k})))} \\ - \frac {\lambda_ {g e n}}{N} \sum_ {i = 1} ^ {N} \log P \left(f _ {\theta , \eta} \left(x _ {i}\right) \mid f _ {\theta , \eta} \left(x _ {(<   i)}\right)\right), \tag {4} \\ \end{array}
+$$
+
+Here, M indicates the embedding task batch and N indicates the generation task batch.  $f_{\theta}$  denotes an embedding model parameterized by  $\theta$  while  $f_{\theta,\eta}$  refers to the architecture augmented with an  $\eta$ -parameterized language modeling head for generation tasks. The temperature hyperparameter is denoted as  $\tau$ , and  $\sigma$  signifies the process of applying a pooling operation to each output embedding followed by cosine similarity computation. Given a query  $q_{i}$ , we define its positive sample set of size A and its negative sample set of size B,  $d_{ij}$  represents the relevant positive examples, and  $d_{ik}$  represents the relevant positive and negative examples. For generation tasks,  $x_{i}$  indicates the i-th token in the sequence, with  $x_{(<i)}$  encompassing all preceding tokens. The training objective is to minimize the negative log-likelihood loss exclusively over the tokens ( $\{\text{response}\} < /s>$ ). Based on experimental results identifying  $(\lambda_{gen}, \lambda_{emb}) = (2,1)$  as optimal, we initialize the loss function accordingly.
+
+# 3.3 Intent Classification
+
+# 3.3.1 Information Retrieval
+
+Incorporating reasoning support and a domain-adaptive retrieval embedding model, we clas
+
+sify intents via the designed pipeline. We prompt the LLM with the instruction tuple  $I\_Ins, r_i^*, d_i^*, q_{testi}$  (where  $q_{testi}$  is drawn from the test set), and retrieve the associated reasoning  $r_i^*$  and instances  $d_i^*$ . Retrieval then uses embedding similarity between the user query and each  $d_i^*$  to assess relevance. Given a query  $q_{testi}$ , Instance Database  $\mathcal{D}$ , and Reasoning Database  $\mathcal{R}$ , the embedding model  $M$  encodes them as  $v_{q_i}, V_D, V_R$ :
+
+$$
+v _ {q _ {i}} \leftarrow M \left(q _ {\text {t e s t} i}\right), V _ {D} \leftarrow M (\mathcal {D}), V _ {R} \leftarrow M (\mathcal {R}), \tag {5}
+$$
+
+To retrieve related instances  $d_{i}^{*}$  and related reasoning  $r_i^*$  from a large-scale dataset, we used the graph-based approximate nearest neighbor search algorithm Hierarchical Navigable Small Worlds(HNSW)(Malkov and Yashunin, 2020). The relevant instances need to meet the following nearest-neighbor criteria:
+
+$$
+\begin{array}{l} d _ {i} ^ {*} = \arg \min  _ {d _ {i} ^ {*} \in \mathcal {D}} \left\{d i s (v _ {q _ {i}}, V _ {D}) \right\}, \\ r _ {i} ^ {*} = \arg \min  _ {r _ {i} ^ {*} \in \mathcal {R}} \left\{d i s \left(v _ {q _ {i}}, V _ {R}\right) \right\}, \tag {6} \\ \end{array}
+$$
+
+Here, the function  $dis()$  denotes the cosine distance computed by the pre-trained embedding model  $M$ . First, pre-process the Reasoning Dataset  $\mathcal{R}$  and the Instance dataset  $\mathcal{D}$  in the reasoning library, plus the embedded instruction tag:  $<|user|>\{\} <|embed|>\{\mathcal{D},\mathcal{R}\}$ . (Note that  $\mathcal{D}$  and  $\mathcal{R}$  are already included in the training data, so no extra instructions are needed for embedding.). Next, for each user query  $q_{test_i}$ , we add a task tag:  $<|user|>\{I_Ins\}<|embed|>\{q_{test_i}\}$ . ( $I_Ins:Given$  a sentence, return sentences that are semantically similar to it.) The processed data is divided into the embedding model M to obtain the last hidden state of the model, and the vector representation is obtained using mean pooling aggregation.
+
+Finally, the similarity between the embedding vectors is obtained by applying the HNSW search algorithm, quickly retrieve and returning the most relevant TOP-N instances  $d_{i}^{*}$ . Meanwhile, we select the TOP-1 reason  $r_{i}^{*}$  that best matched the query for intent judgment.
+
+# 3.3.2 Results Obtain
+
+Construct a  $P_{C_i}$  for classification and input it into an LLM for classification:
+
+$$
+P _ {C i} = \operatorname {T a s k} \left\{C \_ I n s \right\};
+$$
+
+$$
+\text {R e a s o n} \left\{r _ {i} ^ {*} \right\};
+$$
+
+$$
+\text {E x a m p l e s} \left\{d _ {i} ^ {*} \right\}; \tag {7}
+$$
+
+$$
+Q u e y \left\{q _ {t e s t i} \right\},
+$$
+
+$$
+\operatorname {r e s u l t} _ {i} = L L M _ {\text {c l a s s i f i c a t i o n}} \left(P _ {C _ {i}}\right),
+$$
+
+Where Ins represents the input intent classification task instruction, reasoning data  $r_i^*$  is used to provide the analysis process, instance data  $d_i^*$  provides the relevant instances, and  $q_{test_i}$  is the user's query. Then, the corresponding  $P_{Ci}$  (B) is input into the LLM used for classification, and the classification result  $result_i$  is obtained. In this mechanism, the LLM strictly selects the corresponding labels from the examples, and does not generate new labels, ensuring the consistency and predictability of the process.
+
+# 4 Experiment
+
+# 4.1 Dataset and Evaluation Method
+
+To evaluate the performance of ICLER in general scenarios, we used multiple datasets: the widely adopted multilingual Amazon Review corpus (MARC) including its en-US, zh-CN, and ja-JP subsets (Keung et al., 2020), which contain cross-lingual user reviews to support the model's multilingual text processing; the bankingIntent dataset (focused on bank-related intent identification, critical for assessing performance in the financial domain); the mistral-intent-data1816 dataset (covering diverse scenarios to comprehensively test intent recognition capabilities); and a self-constructed PC domain dataset (with over 500 labels and industrial-domain intent classification characteristics). Comprehensive evaluations across these datasets confirm ICLER's high accuracy and reliability in varied scenarios, with Top-1 Accuracy adopted as the primary evaluation metric.
+
+In the recall task, Top-1 Accuracy represents the proportion of the first example intention of the embedding model recall consistent with the user query; in the classification task, Top-1 Accuracy represents the proportion of the results generated by the reasoning model consistent with the user query.
+
+Table 1: Evaluation results of employing Multilingual-E5-Base and Qwen1.5-1.8B for embedding, with Qwen2.5-7B handling reasoning tasks. Evaluation uses Top-1 Accuracy(%) as the primary metric.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Method</td><td colspan="6">Top-1 Accuracy (%)</td></tr><tr><td>pc</td><td>banking</td><td>en</td><td>ja</td><td>zh</td><td>mistral</td></tr><tr><td rowspan="3">Multilingual-E5</td><td>Base</td><td>33.72</td><td>66.23</td><td>67.99</td><td>59.72</td><td>65.27</td><td>67.05</td></tr><tr><td>SetFit</td><td>56.08</td><td>83.18</td><td>81.98</td><td>77.34</td><td>76.70</td><td>73.86</td></tr><tr><td>Emb+Reason</td><td>61.15</td><td>90.68</td><td>80.70</td><td>78.28</td><td>78.41</td><td>75.00</td></tr><tr><td rowspan="3">Qwen1.5-1.8B</td><td>Base</td><td>28.59</td><td>63.28</td><td>67.18</td><td>66.95</td><td>68.70</td><td>56.00</td></tr><tr><td>Emb-only</td><td>66.76</td><td>89.84</td><td>85.54</td><td>81.24</td><td>83.15</td><td>79.55</td></tr><tr><td>Emb+Reason</td><td>67.81</td><td>92.05</td><td>86.08</td><td>83.96</td><td>84.53</td><td>81.82</td></tr><tr><td rowspan="2">Qwen2.5-7B</td><td>ICL</td><td>81.44</td><td>92.21</td><td>87.79</td><td>84.36</td><td>85.14</td><td>84.09</td></tr><tr><td>ICLER</td><td>86.21</td><td>93.00</td><td>87.83</td><td>85.04</td><td>85.94</td><td>85.23</td></tr></table>
+
+# 4.2 Settings
+
+Given strict time constraints, we constructed a recall system based on an embedding model, selecting two base models—Multilingual-E5-Base and Qwen1.5-1.8B-Chat—as embedding models. For data preparation, we utilized Qwen2.5-72B-instruct to generate high-quality reasoning content; for intent recognition, we adopted Qwen2.5-7B-Instruct. Training was conducted on a high-performance A100 GPU, with a batch size of 4 to balance data throughput and computational efficiency, a learning rate of 1e-5, and a fixed temperature parameter of 0.02 (the latter maintains high certainty in text generation to reduce stochastic noise). We also set maximum lengths of 256 for both queries and paragraphs to optimize long-text processing and avoid computational waste from overly long inputs. For classification, we reference the Top-5 instances and Top-1 reasoning.
+
+# 4.3 Main Result
+
+# 4.3.1 Validation Results
+
+As shown in Figure 3, we addressed the issue of insufficient LLM reasoning ability proposed in the Introduction section. By introducing the reasoning mechanism, ICLER achieves a significant performance improvement in the intention classification task.
+
+Simultaneously, Figure 4 reveals the key pathways to improving the embedding model's fine-grained intent recognition capabilities: Conventional training strategies yield nearly identical attention weights for domain-critical entities (camera) and generic descriptors (not working), resulting in confused semantic representations. However, our joint training paradigm integrating reasoning generation and vector optimization tasks re
+
+![](images/e42afce435356ac0f2aca8dbc33f6dcb49ef44c06f85599c78ef3284b315b14f.jpg)  
+Figure 3: Results of the ICLER. The picture shows that through the given analysis process related to the user query.
+
+calibrates this attention distribution, amplifying focus on technical entities (camera: 0.73 attention weight) while suppressing generic terms. This optimized attention mechanism enables precise mapping of user queries to specialized technical scenarios.
+
+# 4.3.2 Experimental Results
+
+In the evaluation, the performance of the embedding models is summarized in Table 1.
+
+All experimental results in Table 1 are "domain evaluations"—specifically, the model is fine-tuned on one dataset (e.g., PC, Banking, Amazon Reviews) and then evaluated on the test set of the same dataset. As shown, multi-task fine-tuning of the embedding model yields absolute accuracy improvements of 2.72 points on ja-JP and 1.38
+
+![](images/f8f867a1a67b5381889de814dd1eb2282fa5a26c746eef2228f0c8225be39270.jpg)  
+Figure 4: Attention-heatmap. The Qwen1.5-1.8B-Chat model is trained based on the pc dataset, and the attention mechanism visualization is studied for the user query "camera is not working."
+
+Table 2: Top1-Accuracy(‰) of GTE-Qwen2-1.5B-Instruct Model on MTEB Classification Tasks after Multi-task Fine-tuning  
+
+<table><tr><td>Classification Task</td><td>Base</td><td>Emb+ Reason</td></tr><tr><td>ToxicConversationsClassification</td><td>82.30</td><td>82.04↓</td></tr><tr><td>MTOPDomainClassification</td><td>97.54</td><td>98.39↑</td></tr><tr><td>MassiveIntentClassification</td><td>82.41</td><td>84.05↑</td></tr><tr><td>MTOPIntentClassification</td><td>88.71</td><td>87.44↓</td></tr><tr><td>AmazonPolarityClassification</td><td>96.60</td><td>96.33↓</td></tr><tr><td>TweetSentimentExtractionClassification</td><td>72.55</td><td>71.55↓</td></tr><tr><td>MassiveScenarioClassification</td><td>84.38</td><td>89.09↑</td></tr><tr><td>Banking77Classification</td><td>87.33</td><td>88.44↑</td></tr><tr><td>AmazonCounterfactualClassification</td><td>84.04</td><td>90.12↑</td></tr><tr><td>AmazonReviewsClassification</td><td>55.24</td><td>54.70↓</td></tr><tr><td>ImdbClassification</td><td>95.71</td><td>95.22↓</td></tr><tr><td>EmotionClassification</td><td>61.05</td><td>80.43↑</td></tr></table>
+
+points on zh-CN over single-task embedding-only fine-tuning. Among the tested models, Qwen1.5-1.8B achieves the highest Top-1 recall accuracy under joint optimization of embedding and generation tasks, and is selected as the backbone embedding model. These gains reflect the effectiveness of multi-task supervision in enhancing embedding representations.
+
+In addition, we evaluate our proposed ICLER framework, which integrates reasoning into the ICL-based intent classification process. Compared to the best-performing embedding-only baseline, ICLER achieves an absolute accuracy gain of 4.77 points on PC-related domain datasets. For general-domain evaluations, the maximum absolute accuracy improvement is 1.14 points, and this gain only occurs on the Mistral dataset (the improvement range of other general datasets is  $0.04\% - 0.89\%$ ). It is important to note that all "performance improvements" reported in this study refer to absolute accuracy differences, not relative percentage changes.
+
+To further assess the effectiveness of our approach, we select six classification tasks from the MTEB benchmark: MTOPDomainClassification,
+
+MassiveIntentClassification, MassiveScenarioClassification, Banking77Classification, AmazonCounterfactualClassification, and EmotionClassification. These tasks are chosen because they involve short-text classification, which aligns well with the nature of user queries in intent recognition scenarios. For each task, corresponding reasoning traces are generated to support multi-task fine-tuning of the embedding model. We then fine-tune the GteQwen2-1.5B-instruct model with both generation and embedding objectives.
+
+As shown in Table2, our approach achieves consistent improvements on the target tasks. Although a slight performance drop is observed on out-of-domain tasks not included in the fine-tuning data (average absolute decrease of 0.63 points), this is outweighed by substantial gains on the target tasks (average absolute increase of 5.56 points). Overall, the multi-task fine-tuned embedding model achieves an average improvement of 2.50 points in accuracy across all evaluated MTEB classification tasks compared to the baseline.
+
+# 4.3.3 Ablation Studies
+
+To verify the impact of key parameters and design choices on model performance, we conducted ablation experiments focusing on three core aspects: the number of retrieved examples, the relevance of reasoning samples, and the decoupling strategy of reasoning and instances. The results are as follows:
+
+(1) Number of Retrieved Examples: When retrieving the Top-5 instances, over  $97\%$  of test samples included the correct intent in the retrieval results. Continuing to increase the number of retrieved examples (e.g., Top-10) only improved the accuracy by  $0.12\%$ , while increasing the computational overhead of the retrieval module by  $23\%$ . Considering the balance between performance and efficiency, we fixed the number of retrieved examples at  $\mathrm{N} = 5$  in the final framework.
+
+(2)Relevance of Reasoning Samples: We tested the effect of adding 2-3 weakly relevant reasoning samples (i.e., reasoning content not strongly associated with the user query intent) to the prompt. The results showed that the model accuracy decreased by  $0.3\% - 0.5\%$  compared to using only strongly relevant samples. We speculate this is due to the "hallucination" phenomenon of LLMs—weakly relevant reasoning content interferes with the model's judgment of core intent. Additionally, the increase in reasoning samples extended the input sequence length by  $40\%$ , leading to a  $15\%$  rise in inference
+
+time. Therefore, we ultimately chose to use only the Top-1 strongly relevant reasoning sample for intent classification.
+
+Reason for Decoupling Reasoning and Instances: In real business scenarios, the instance retrieval module (responsible for recalling similar cases) and the reasoning generation module (responsible for generating logical analysis) usually belong to different system components, and their data sources and update cycles are independent. If the two modules are strongly bound (e.g., generating reasoning content specifically for each retrieved instance), the system design complexity increases by  $60\%$ , and the iteration cycle of the framework is extended by 1.8 times. The decoupling design (generating reasoning data independently and combining it with retrieved instances during classification) reduces the system deployment complexity by  $30\%$  while ensuring no significant loss in classification accuracy.
+
+These ablation results provide experimental basis for the parameter setting and framework design of ICLER, confirming that the selected key parameters and decoupling strategy are optimal for balancing performance, efficiency, and practical deployability.
+
+# 5 Conclusion
+
+In conclusion, ICLER successfully improves the performance of ICL in the intention classification task. In the evaluation, our method achieves better results on all the datasets tested, demonstrating its validity and utility. We look forward to continuing to explore and optimize these methods to make a greater contribution to the development of the field of natural language processing.
+
+# Limitations
+
+Although ICLER shows significant advantages in each dataset, there are still three limitations. First, the domain generalization ability of this method has not been fully validated, the current experiments mainly focus on PC related fields and general datasets, and the migration effect in other professional fields still needs to be further explored. Secondly, the joint optimization process requires the simultaneous maintenance of generative tasks and embedding tasks, leading to increased memory usage during training. Finally, the model performance improvement still depends on the scale effect of the base model, and how to achieve a sim
+
+ilar optimization path in the model with a smaller number of parameters will be an important future research direction.
+
+# References
+
+Akari Asai, Timo Schick, Patrick Lewis, Xilun Chen, Gautier Izacard, Sebastian Riedel, Hannaneh Hajishirzi, and Wen tau Yih. 2022. Task-aware retrieval with instructions. Preprint, arXiv:2211.09260.  
+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, and 12 others. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. Preprint, arXiv:2402.03216.  
+DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6894-6910, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised dense information retrieval with contrastive learning. Preprint, arXiv:2112.09118.  
+Phillip Keung, Yichao Lu, György Szarvas, and Noah A. Smith. 2020. The multilingual Amazon reviews corpus. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4563-4568, Online. Association for Computational Linguistics.  
+Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. Preprint, arXiv:2308.03281.
+
+Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pages 100–114, Dublin, Ireland and Online. Association for Computational Linguistics.  
+Sheng Liu, Haotian Ye, Lei Xing, and James Zou. 2024. In-context vectors: Making in context learning more effective and controllable through latent space steering. Preprint, arXiv:2311.06668.  
+Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics.  
+Yu A. Malkov and D. A. Yashunin. 2020. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):824-836.  
+Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2025. Generative representational instruction tuning. Preprint, arXiv:2402.09906.  
+Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. 2023. Mteb: Massive text embedding benchmark. Preprint, arXiv:2210.07316.  
+OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, and 262 others. 2024. Gpt-4 technical report. Preprint, arXiv:2303.08774.  
+Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.  
+Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2023. One embedder, any task: Instruction-finetuned text embeddings. In Findings of the Association for Computational Linguistics: ACL 2023, pages 1102-1121, Toronto, Canada. Association for Computational Linguistics.
+
+Lewis Tunstall, Nils Reimers, Unso Eun Seo Jo, Luke Bates, Daniel Korat, Moshe Wasserblat, and Oren Pereg. 2022. Efficient few-shot learning without prompts. Preprint, arXiv:2209.11055.  
+Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. Preprint, arXiv:2402.05672.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent abilities of large language models. Transactions on Machine Learning Research. Survey Certification.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models. Preprint, arXiv:2201.11903.  
+Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2022. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations.  
+Zhe Yang, Damai Dai, Peiyi Wang, and Zhifang Sui. 2023. Not all demonstration examples are equally beneficial: Reweighting demonstration examples for in-context learning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13209-13221, Singapore. Association for Computational Linguistics.  
+Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12697-12706. PMLR.  
+Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large language models are human-level prompt engineers. Preprint, arXiv:2211.01910.
+
+# System Prompt
+
+You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
+
+# User Prompt
+
+# Task Specification
+
+Given a specific case and two labels (a correct label and a wrong label), analyze the reason why the case is a correct label and why it is not a false label.
+
+# Requirements
+
+The reason for analysis is given and must be marked with  $\langle \text{reason} \rangle$
+
+# Examples
+
+case: "I am still waiting on my card?"
+
+True_label: card Arrival
+
+False_label: carddelivery Estimate
+
+# Output Examples
+
+<reason>The case "I am still waiting on my card?" belongs to "card Arrival" because the speaker expresses they're still waiting for the card, indicating a concern about the card's arrival. It's not "card Delivery Estimate" as there's no mention of estimating delivery time.</reason>
+
+# Candidate Intent And The Corresponding Examples
+
+case: {query}
+
+True_label: {true_label}
+
+False_label: {false_label}
+
+Analyze The Reasons
+
+# System Prompt
+
+You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
+
+# User Prompt
+
+Based on the given intents and one of its examples, you need to determine which intent the user input Content belongs to. Additional, Context is the historical user input.
+
+# Examples of analysis for other cases
+
+The case "wake me up at this time" belongs to "alarm_set" because the speaker is requesting to set an alarm for a specific time. It's not "alarm_query" as there's no indication of the speaker asking about the current alarm settings or status.
+
+# Requirements
+
+- DO NOT create new intent on your own, you must strictly use the intents in the examples.  
+- Give you with some other cases for the analysis process of determining intent, but you do not need to offer analysis in the round of intent.  
+- If no examples match the targeted Content, Output -1.  
+- Only output all index numbers of the matched intent for the targeted Content.  
+- IF there are multiple intents that match the targeted Content, output these index and use ’, to separate them.  
+- Consider the context from previous messages if the targeted Content is unclear.
+
+# Examples
+
+[case] wake me at daybreak [label] alarm_set [Index] 1
+
+[case] tell me when the next alarms are for [label] alarm_query [Index] 2
+
+[case] i need to set an alarm how many do i have set [label] alarm_query [Index] 3
+
+[case] can you set my alarm [label] alarm_set [Index] 4
+
+[case] then tap ok [label] alarm_query [Index] 5
+
+# #Content
+
+alarm settings
+
+# Output
+
+2
+
+Algorithm 1 Build Reasoning Database  
+Query dataset  $Q_{train} = \{q_{traini}\}$   
+Document dataset  $D = \{d_i\}$   
+Model for generating reasonings  $M_1$   
+Model for embedding  $M_2$   
+Model for classification  $M_3$   
+Prompt template:  $P_R, P_C$   
+Similarity score function of Retrieval:  $dis()$   
+1. Initialize:  $\mathcal{R} \gets \emptyset, V_D \gets M_2(D)$   
+2. for each  $q_{traini} \in Q_{train}$ :  
+3.  $V_q^i \gets M_2(q_{traini})$   
+4.  $\mathcal{D}_i^* = \{d_j^*\}_i \gets \arg \min_{d_j \in \mathcal{D}} \text{dis}(V_q^i, V_D^j)$   
+5.  $P_{Ci} \gets (q_{traini}, \mathcal{D}_i^*, \mathcal{R}_i^* = \emptyset)$   
+6. labelresulti  $\gets M_3(P_{Ci})$   
+7. If labelresulti == true_label_qi:  
+8. Continue  
+9. else:  
+10.  $P_{Ri} \gets (qi, \text{true_label}_i, \text{labelresult})$   
+11.  $r_i \gets M_1(P_{Ri})$   
+12.  $\mathcal{R} \gets r_i$   
+13. end for  
+14. return  $\mathcal{R}$
+
+# Algorithm 2 Results Obtain
+
+Query dataset  $Q_{test} = \{q_{testj}\}$
+
+Document dataset  $\mathcal{D} = \{d_k\}$
+
+Reasoning dataset  $\mathcal{R} = \{r_l\}$
+
+Model for embedding  $M_{1}$
+
+Model for classification  $M_2$
+
+Prompt template:  $P_{C}$
+
+Similarity score function of Retrieval:  $dis()$
+
+1. Initialize: Result  $\leftarrow \emptyset$ ,  $V_{D} \gets M_{1}(D)$ ,  $V_{R} \gets M_{1}(R)$  
+2. for each  $q_{test_i} \in Q_{test}$ :  
+3.  $V_{q}^{i}\gets M_{1}(q_{test_{i}})$  
+4.  $\mathcal{D}_i^* = \{d_j^*\}_{i}\gets \underset {d_j^*\in \mathcal{D}}{\arg \min}dis(V_q^i,V_D^j)$  
+5.  $r_i^*\gets \arg \min_{r_i^*\in \mathcal{R}}dis(V_q^i,V_R)$  
+6.  $P_{C_i} \gets (q_{test_i}, \mathcal{D}_i^*, r_i^*)$  
+7. resulti  $\leftarrow M_2(P_{C_i})$  
+8. Result  $\leftarrow$  resulti  
+9. end for  
+10. return Result

2025/ICLER_ Intent CLassification with Enhanced Reasoning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8e3d6db9a7d5b706a9ceabd417c31da938420dfa730a32bcceba3c293e8aed8
+size 378167

2025/INDOORWORLD _ Integrating Physical Task Solving and Social Simulation in A Heterogeneous Multi-Agent Environment/ee507a4c-7454-4514-a8c9-73b0446536c5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6bbeca29c8573d1e6501862328be7024cc5147fb6aef7deecfb4d7628ef1511
+size 9346424

2025/INDOORWORLD _ Integrating Physical Task Solving and Social Simulation in A Heterogeneous Multi-Agent Environment/full.md

@@ -0,0 +1,529 @@
+# INDOORWORLD : Integrating Physical Task Solving and Social Simulation in A Heterogeneous Multi-Agent Environment
+
+Dekun Wu<sup>1,2</sup> Frederik Brudy<sup>4</sup> Bang Liu<sup>1,2,3</sup> Yi Wang<sup>4</sup>
+
+<sup>1</sup>Université de Montréal, <sup>2</sup>Mila - Quebec AI Institute, <sup>3</sup>Canada CIFAR AI Chair
+
+4Autodesk Research
+
+{dekun.wu, bang.liu}@umontreal.ca
+
+frederik.brudy@autodesk.com, ywang485@gmail.com
+
+# Abstract
+
+Virtual environments are essential to AI agent research. Existing environments for LLM agent research typically focus on either physical task solving or social simulation, with the former oversimplifying agent individuality and social dynamics, and the latter lacking physical grounding of social behaviors. We introduce INDOORWORLD, a heterogeneous multi-agent environment that tightly integrates physical and social dynamics. By introducing novel challenges for LLM-driven agents in orchestrating social dynamics to influence physical environments and anchoring social interactions within world states, INDOORWORLD opens up possibilities of LLM-based building occupant simulation for architectural design. We demonstrate the potential with a series of experiments within an office setting to examine the impact of multiagent collaboration, resource competition, and spatial layout on agent behavior.
+
+# 1 Introduction
+
+The emergence of Large Language Model (LLM)-based agents has extended LLMs beyond traditional one-off interactions, equipping them with long-term memory, planning capabilities, and embodied actions (Yao et al., 2022; Shinn et al., 2024; Wang et al.). Among them, multi-agent systems leverage distinct agent roles to achieve greater collective intelligence for problem-solving (Hong et al., 2024; Qian et al., 2023; Tang et al., 2024) or to enable more realistic cognitive and psychological modeling in social simulations.
+
+Like in traditional AI research (Maes, 1995), virtual environments are essential for LLM-based agents, enabling them to perceive and act. These environments provide external sensory input and introduce a world state, allowing agent actions to be grounded as operators that modify the environment. These environments serve as low-cost testbeds that accelerate LLM-agent development,
+
+typically falling into two categories: physical task-solving environments and social simulation environments.
+
+Physical task-solving environments such as VirtualHome (Puig et al., 2018) and ALFWorld (Shridhar et al., 2020b) enable agents to interact with external objects to manipulate the world state towards specific objectives. However, these environments often assume identical action spaces, homogeneous agent abilities, neglecting individual differences among agents, and the impact of social dynamics on task solving.
+
+On the other hand, social simulation environments, such as Smallville (Park et al., 2023), enable social interactions between agents, such as relationship building and information sharing. While these systems effectively simulate human-like social behaviors driven by individual personalities and roles, they often employ an oversimplified model of the physical world. This leads to a lack of groundings for agents' actions in the change of world state, resulting in social interactions that remain merely "plausible" without any correspondence to an external physical reality. For example, an agent may refer to non-existent physical objects in a dialog.
+
+The gaps lead to missed opportunities to use AI agents for applications that require tight integration of physical task-solving and social simulation. One such application is building occupant simulation for architectural design (Yan et al., 2015; Feng et al., 2015), where occupant behaviors are driven by both dynamic physical and social factors. To respond to the missed opportunities, we present INDOORWORLD, a heterogeneous multi-agent environment that tightly integrates physical and social dynamics in an indoor space setting, introducing novel challenges for LLM-driven agents in orchestrating social dynamics to influence physical environments and anchoring social interactions within physical world states. As a multi-agent system, INDOORWORLD allows for fully decentralized agent
+
+control and collective task assignment, facilitating self-regulated labor division, task prioritization and coordination. INDOORWORLD provides a scalable and expressive testbed for advancing research on multi-agent LLM systems.
+
+Our key contributions are as follows:
+
+- Heterogeneous Agent Modeling: We introduce a multi-level approach in which agents vary in roles, actions, capabilities, and knowledge, yielding individual differences that professionals judged to be more realistic.  
+- Integrated Physical and Social Dynamics: Our environment seamlessly combines physical object manipulation with social behaviors, posing novel challenges and setting the stage for developing LLM-based multi-agent systems for both task solving and social simulation.  
+- Promising Tools for Architectural Design: Experiments on multi-agent collaboration, resource competition, and layout effects demonstrate that our platform can aid spatial optimisation and resource allocation, making it a promising tool for architectural design work.
+
+# 2 Related Works
+
+Task Solving Environments for LLM-based Agents evaluate agents' ability to solve various types of tasks, such as web-based tasks (Cai et al., 2024; Chae et al., 2024), GUI tasks (Nguyen et al., 2024; Wang and Liu, 2024), coding tasks (Hong et al., 2024; Huang et al., 2024; Qian et al., 2023) and household tasks (Shridhar et al., 2021, 2020a; Zhang et al., 2024). In household task-solving environments, agents must explore their surroundings, sense and interpret object states, plan and execute actions. The advantage of such environments lies in their support for a diverse range of physical objects and extensive agent-object interactions. For example, ALFWorld (Shridhar et al., 2021) enables agents to interact with objects such as mugs, books, and lamps. Multi-agent platforms like TDW-MAT and VirtualHome (Zhang et al., 2024; Puig et al., 2018) supports interactions with objects like pens, beds, and apples. Similarly, MineLand (Yu et al., 2024) and AdaSociety (Huang et al., 2025), designed for wilderness survival, feature various tools and food items. However, a common limitation across these environments is the lack of explicit
+
+modeling for agent heterogeneity. ALFWorld is a single-agent environment, while TDW-MAT and VirtualHome (Puig et al., 2018) features homogeneous agents with identical capabilities. Although inventory variations in AdaSociety and MineLand introduce some level of heterogeneity, the agents remain fundamentally homogeneous, as they share the same action space.
+
+Social Simulation Environments for LLM-based Agents enable agent-agent interactions, elevating the importance of social dynamics. These environments often model differences in personality, profession, and other traits among agents (Park et al., 2023; Wu et al.; Xu et al., 2023; Guan et al., 2025; Li et al., 2024), as well as incorporate human needs modeling (Wang et al., 2023, 2024). However, a major limitation of these environments is the oversimplified modeling of the physical world, including interaction with physical objects. For instance, an agent may perform the action of eating, without any explicit modeling of food items in the environment. This prevents these simulation environments to be applied in settings where the physical environment has impacts on agent behaviors, such as resource allocation and layout study.
+
+Abstract Modeling with Text-based Environments Prior works such as TextWorld (Côté et al., 2018) and ALFWorld (Shridhar et al., 2020b) have explored the use of text-based environments for abstract modeling of visual and physical dynamics. These approaches are inspired by concepts from inverse graphics and inverse dynamics (Kulkarni et al., 2015; Wu et al., 2017), where high-level representations allow agents to reason about the environment and predict future outcomes. Building on this line of research, our work aligns with ALFWorld (Shridhar et al., 2020b) in treating text-based simulation as a platform for abstract modeling, while further extending this perspective toward heterogeneous and socially grounded agent interactions.
+
+AI in Architectural Design has been widely adopted, mainly to generate static 2D and 3D artefacts, such as floorplans, interiors, and furniture layouts, but these visually oriented methods provide little insight into the dynamic occupant activities (Li et al., 2025; Raistrick et al., 2024; Leng et al., 2023). Moreover, traditional occupant-simulation tools rely on rule-based or state-based behavior transitions (Schaumann et al., 2017; Lee
+
+![](images/90026d1d251018dbd88c3226ea5d91baf48c41cf59b869a435640dcc5abc081d.jpg)  
+(a)
+
+![](images/e069c9eff3d488ef7e81677b28b799faafa7a69868b7c917934ce8c04f30ea5c.jpg)  
+(b)
+
+![](images/e307bc3f9ccbc0fc60dcc9f3a257cf32aba3c63156d798cec80314ee62cf17b2.jpg)
+
+![](images/41752abfc30b5016c591bdb7922528ff5efc0ebabfcb2307a5a1a67b83801107.jpg)
+
+![](images/a981ea2eceae819abea2619f4743a0a85c740677a0017eb9ab7c7c97c208d14a.jpg)  
+(c)  
+Figure 1: INDOORWORLD system. (a) Agent architecture; (b) Example agent behaviors; (c) Example of heterogeneous agent profiles
+
+et al., 2021; Liu et al., 2024), offering far less flexibility and realism than LLMs in responding to changing environments. Our work addresses this gap by employing LLMs to simulate multi-agent interactions within indoor spaces, yielding insights into occupant behaviors that are crucial for effective spatial planning and resource management.
+
+# 3 INDOORWORLD Environment
+
+Our environment INDOORWORLD seamlessly integrates task-solving and social simulation in an indoor space setting, offering a versatile testbed to study the interplay between social and physical agent behavior. INDOORWORLD enables agents with rich individual differences to collaborate on tasks with complex hierarchy, involving self-regulated labor division, prioritization and coordination, while satisfying physiological needs through environmental interactions. By incorporating fine-grained modeling of agent internal structure and decision-making process, INDOORWORLD more accurately resembles real-world human behaviors. Tables 1, 5, and 7 summarize the key differences between INDOORWORLD and existing platforms, highlighting its potential to advance research in LLM-based multi-agent systems.
+
+# 3.1 Environment Architecture
+
+Framework Overview and Core Components Virtual environments that support planning and task solving often require managing the world state (Srivastava et al., 2021; Shridhar et al., 2021). Similarly, our INDOORWORLD adopts an object-oriented approach to define state transition systems through three key components: 1) Agents, 2) Objects, and 3) Locations.
+
+Each agent and object is associated with state variables, such as a numerical value for an agent's hunger or a Boolean indicating whether a computer is broken. A dedicated variable tracks each entity's current location. The overall world state is the joint valuation of all these variables. Object affordances are defined by the set of actions that can be performed on them, and these actions are further constrained by the agent's role (e.g., only an IT admin can repair computers). When actions are executed, the world state updates accordingly. Objects can be marked as receptacles to store other objects. Locations can be interconnected to allow agent and object movement.
+
+Agent interactions are through conversations. INDOORWORLD supports 4 actions related to conversation with other agents: 1) initiating chat; 2) stay chat; 3) end chat, and 4) join chat. We let the LLMs to generate free-form dialog content. Note that we allow any number of agents to
+
+<table><tr><td></td><td>MA</td><td>AT</td><td>OI</td><td>TE</td><td>LS</td><td>FH</td><td>HN</td><td>RF</td></tr><tr><td>ALFworld(Shridhar et al., 2021)</td><td>✘</td><td>-</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>✓</td></tr><tr><td>Virtual-Home(Puig et al., 2018)</td><td>✓</td><td>Ho.</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>✓</td></tr><tr><td>TDW-MAT(Zhang et al., 2024)</td><td>✓</td><td>Ho.</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>✓</td></tr><tr><td>C-WAH(Zhang et al., 2024)</td><td>✓</td><td>Ho.</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>✓</td></tr><tr><td>MineLand(Yu et al., 2024)</td><td>✓</td><td>Ho.</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✘</td></tr><tr><td>AdaSociety(Huang et al., 2025)</td><td>✓</td><td>Ho.</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✘</td></tr><tr><td>Smallville(Park et al., 2023)</td><td>✓</td><td>He.</td><td>Lmtd.</td><td>✘</td><td>✓</td><td>✘</td><td>✘</td><td>✓</td></tr><tr><td>Humanoid Agents(Wang et al., 2023)</td><td>✓</td><td>He.</td><td>Lmtd.</td><td>✘</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td></tr><tr><td>Ours</td><td>✓</td><td>He.</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+Table 1: Comparison of environments. MA: Multi-Agents, AT: Agent Type, OI: Object Interaction, TE: Task Eval, LS: Life Simulation, FH: Fine-Grained Heterogeneity, HN: Human Needs, RF: Real-world Fit. "Ho." = Homogeneous, "He." = Heterogeneous, "Lmtd." = Limited.
+
+be in a conversation, as long as they are at the same location. A dedicated agent state variable indicates which conversation session the agent is currently involved (if any). Action 2) and 3) only become admissible actions when the agent is in a conversation, and 4) becomes admissible when there is an ongoing conversation session at the agent's current location. Each agent can only be at one conversation session at a time. Conversations can be used to share information (including task progress), discuss labor division, and coordinating actions. They affect the agent's subsequent actions by updating the agent's internal state. Note that in task-solving scenarios, conversations are utility-driven—the dialog content need to serve task-solving. This aspect distinguishes our work from most existing environments featuring agent conversations.
+
+Sessions and Scenarios INDOORWORLD supports both task solving and simulation sessions. In a task-solving session, a set of tasks are assigned as the shared objective for all agents. The agents need to collectively decide on task orders and assignments, as well as planning for specific action sequences to complete each task. In simulation sessions, no explicit objective is defined.
+
+Both session types start with an initial world state defined by a scenario, a JSON configuration file specifying agents, objects, locations, inter-location connections, and receptacle assignments. (See Appendix E.1 for an example JSON file.)
+
+To facilitate experimentation, INDOORWORLD
+
+comes with 25 predefined object types (including 7 receptacle types) and 4 predefined agent roles, resulting in a total of 38 action types. The current set of objects and agent roles cover typical activities in an office environment, such as booking meeting rooms, moving desks, cleaning utensils, repairing computers, etc. As shown in Table 7, our environment offers a broader action space compared to many existing text-based and 2D/3D platforms.
+
+Customization and Expansion Although INDOORWORLD currently feature a limited number of object types and agent roles, our object-oriented approach allows easy customization of object type and agent roles. Introducing new object types and agent roles involves defining Python functions specifying new interactions, including preconditions and effects, which can be easily achieved by utilizing existing class hierarchy. We provide example code in Appendix E.2 to illustrate how to introduce new object types, agent roles and interaction type.
+
+# 3.2 Agent Architecture
+
+Figure 1 (a) illustrates the operational framework of our agents, highlighting their interactions with both the environment and other agents. Our architecture integrates cognitive modules inspired by recent research on LLM-based agents (Yao et al., 2022; Zhang et al., 2024) and consists of five core modules: perception, memory, planning, action, and task prioritization.
+
+Perception: This module processes symbolic observations from the environment, such as nearby objects, receptacles, and other agents' activities, and updates the agent's internal state accordingly.
+
+Memory: The memory module stores agent-specific information, long-term knowledge, and interaction history. Inspired by the COELA model (Zhang et al., 2024), it maintains a semantic map and tracks task progress, while also recording episodic events (e.g., past actions and conversations) and retaining pre-existing knowledge. It further monitors internal states like physiological needs and inventory status.
+
+Planning: Using current observations and stored memories, the planning module determines the agent's current objective and task to address both task-specific and internal needs.
+
+Action: Informed by the ReAct framework (Yao et al., 2022), the action module integrates infor
+
+mation from perception, memory, and planning to select and execute the next action. This process involves first reasoning about the current situation and evaluating available options before deciding the next move<sup>1</sup>.
+
+Task Prioritization: Our preliminary experiments revealed that LLM-based agents, particularly those using open-source models, struggle to maintain focus in multi-task scenarios, frequently switching tasks without completing them (See Sec. 4.3). To address this, we proposed a task prioritization module that encourages agents to concentrate on ongoing tasks. The module monitors the objects an agent holds and their relevance to the current task, reminding the agent of incomplete objects. For example, as shown in Figure 1, when agent Ryan is working on preparing coffee and is carrying an empty cup_6, the module highlights its incomplete status, such as the absence of coffee or its incorrect placement. When no active task is detected, the module reminds the agent about all unfinished tasks, guiding the agent to select an objective aligned with its role and skills. This approach promotes concentration on ongoing tasks and minimizes inefficient task switching. Note that the module does not introduce extra information but selectively reiterates relevant parts of the agent's memory, such as task progress, to reinforce task awareness.
+
+Modeling Agent Heterogeneity INDOOR-WORLD addresses the limited diversity found in existing multi-agent benchmarks (Table 1 and 5) by assigning every agent a multi-level profile. Each profile combines a role that determines the agent's unique action space (e.g. only IT administrators can repair devices) with additional attributes such as personality, strength, skill, and knowledge. This layered design (illustrated in Fig. 1c) supports emergent division of labour and coordinated behaviors while greatly increasing realism.
+
+We model agent heterogeneity at four levels: (i) profile level, capturing differences in personas and role configurations; (ii) action space, where roles have distinct actions; (e.g., only IT admin can repair devices); (iii) capability, meaning agents may perform the same action with different efficiency or outcomes, such as janitors cleaning more quickly or strong agents being able to move heavy
+
+objects; and (iv) knowledge, whereby agents hold different internal information (e.g., only receptionists know how to book a meeting room).
+
+To validate this design, 20 practicing architects rated whether heterogeneity at each level increases realism and whether it is important for understanding real space use (details in Appendix B). The consistently high realism scores  $(55 - 70\%)$  and substantial importance ratings  $(55 - 95\%)$  demonstrate that multi-level heterogeneity is both credible and valuable for architectural analysis, thereby supporting the soundness of our design choices.
+
+Modeling Human Needs In INDOORWORLD, agents are associated with physiological and social needs, such as hunger, thirst, and social interaction, that resemble actual building occupants (Figure 1). They are tracked with numerical state values that gradually decline over time, prompting agents to perform restorative actions like eating, drinking, socializing, or using the restroom. This explicit modeling fosters more natural and realistic behaviors as agents balance personal upkeep with assigned tasks.
+
+# 4 Experiments
+
+# 4.1 A Collaborative Task Solving Benchmark
+
+To demonstrate how INDOORWORLD supports task solving with social interactions, we design a benchmark for collaborative task solving in an office setting. The benchmark consists of an overarching task: preparing for a company event, composed of five major subtasks, each involving common office activities such as transporting items, cleaning utensils, repairing equipment, booking meeting spaces, and preparing food and beverages. These subtasks can be broken down into smaller steps that require coordination among agents. Unlike many existing environments where agents execute a single task in isolation (Shridhar et al., 2021), in our setting, all agents simultaneously receive the complete task structure, but must autonomously determine how to divide responsibilities, prioritize actions, and coordinate efforts in a decentralized manner. This introduces additional challenges related to task prioritization, division of labor, communication, and coordination strategies. Each scenario runs for one hour of simulation time, during which agents must collaboratively complete as much of the task as possible. The scenarios feature 9 locations, 67 objects across 16 types (including 15 receptacles of 7
+
+types), and 6 agents: four janitors, one IT administrator, and one receptionist. Details of each subtask and its decomposition can be found in Appendix A.
+
+# 4.2 Social Simulation Experiments
+
+Like Smallville (Park et al., 2023), our environment supports autonomous social simulation where agents generate their own objectives based on physiological/social needs, personality, and roles, driving interactions with both the environment and other agents.
+
+As INDOORWORLD simulates both social and physical interactions, one potential real-world application is to evaluate how well a physical environment can facilitate the activities of its occupants. We demonstrate this application with two experiments, focusing on resource management and spatial layout design, respectively.
+
+Resource Management experiment showcases how we can use INDOORWORLD to evaluate resource allocation in an environment. The scenario involves agents residing in a space with limited resource. We examine how agents compete for and utilize resources under different conditions.
+
+We ran 3 sets of simulations, with 2, 4, and 8 thirsty agents. Each set consists of three settings involving agents all preferring water, all preferring coffee, and no preference, and for each setting we ran simulation with different availability of beverages (one water dispenser/coffee machine vs. two water dispensers/coffee machines).
+
+At the beginning of the simulation, agents experience thirst and autonomously decide whether to drink water or coffee based on their physiological needs and personal preferences. We analyze their resource selection behavior and measure the time required for all agents to fulfill their hydration needs  ${}^{2}$  . This allows us to assess how different resource configurations influence competition and overall efficiency in meeting agent needs.
+
+Spatial Layout experiment showcases how we can use INDOORWORLD to evaluate building layout designs. We ran simulations in two scenarios involving two office layout designs (see Figure 2), each lasting for an 8-hour simulation period. Both Design 1 and Design 2 contain the same 11 functional areas to host five agents: one IT administrator, one janitor, two software engineers, and one receptionist. Each agent has a designated workspace,
+
+such as the front desk for the receptionist and the IT office for the IT administrator.
+
+Figure 2: Spatial layout of design 1 and design 2  
+![](images/2f01e44d3cb370462a0ee663557a99d53e34a86ec43fa38614fa75b391e1c382.jpg)  
+Design 1  
+1. Showcase area  
+2. Open area 2  
+3. IT office  
+4. Desk area 1  
+5. Open area 1  
+6. Desk area 2  
+7. Pantry  
+Design 2  
+1. Desk area 2  
+2. Pantry  
+3. Open area 1  
+4. IT office  
+5. Desk area 1  
+6. Showcase area  
+7. Open area 2
+
+Furthermore, the pantry contains a coffee machine, a water dispenser, food, drinks, cups, and utensils, but with limited resources. The cafeteria, in contrast, is assumed to have no resource constraints, allowing agents to replenish food, drinks, and energy freely. However, activities in the cafeteria takes more time compared to the pantry. Additionally, the environment includes male and female restrooms.
+
+In both Design 1 and Design 2, all areas share the same configurations, including resource availability and the activities they support. The only difference between the two designs lies in the relative positioning of these areas.
+
+Since this experiment does not involve specific task execution, the task prioritization and task progress modules have been removed.
+
+# 4.3 Results and Analysis
+
+In the collaborative task solving experiment, we evaluated three different LLMs, including the open-source Llama 3.3 70B Instruct, Gemma 3 27B and the proprietary GPT-4o-08-06. For the social simulation experiments, the resource management experiment used GPT-4o-08-06, while the spatial layout experiment used GPT-4o-mini-0718. The temperature for all experiments was set to 0.6.
+
+Collaborative Task Solving Quantitative results are shown in Table 2. We report instance-level (IS) task completion rate (percentages of objects that are at the desired state) and attribute-level (AS) task
+
+Table 2: Results of collaborative task solving with mean and standard deviation (reported as superscript in small font). T1-T5 represent the five designed tasks. Reported values are IS (Instance-level Success) rate / AS (Attribute-level Success) rate. FM = Full Model, RA = Random Agents.  
+
+<table><tr><td></td><td>T1
+(IS/AS)</td><td>T2
+(IS/AS)</td><td>T3
+(IS/AS)</td><td>T4
+(IS/AS)</td><td>T5
+(IS/AS)</td><td>AVG
+(IS/AS)</td></tr><tr><td>RA</td><td>8.3±14.4 / 38.9±9.6</td><td>0.0±0.0 / 30.6±2.8</td><td>0.0±0.0 / 30.0±17.3</td><td>0.0±0.0 / 0.0±0.0</td><td>0.0±0.0 / 11.1±1.9</td><td>1.2±2.0 / 23.1±3.0</td></tr><tr><td colspan="7">Llama 3.3 70B Instruct</td></tr><tr><td>FM (w/o S&amp;T+TP)</td><td>83.3±28.9 / 88.9±19.2</td><td>38.9±21.0 / 61.1±12.7</td><td>0.0±0.0 / 53.3±11.5</td><td>100.0±0.0 / 100.0±0.0</td><td>0.0±0.0 / 27.8±5.1</td><td>31.0±9.1 / 51.8±4.7</td></tr><tr><td>FM (w/o TP)</td><td>91.7±14.4 / 94.4±9.6</td><td>36.1±31.5 / 59.3±22.6</td><td>0.0±0.0 / 53.3±11.5</td><td>100.0±0.0 / 100.0±0.0</td><td>0.0±0.0 / 41.1±9.6</td><td>31.0±11.9 / 56.1±11.9</td></tr><tr><td>FM</td><td>100.0±0.0 / 100.0±0.0</td><td>58.3±0.0 / 72.2±0.0</td><td>33.3±57.7 / 73.3±23.1</td><td>100.0±0.0 / 100.0±0.0</td><td>33.3±7.2 / 51.1±11.7</td><td>55.2±9.1 / 67.8±4.8</td></tr><tr><td colspan="7">Gemma 3 27B</td></tr><tr><td>FM (w/o S&amp;T+TP)</td><td>75.0±25.0 / 83.3±16.7</td><td>11.1±4.8 / 45.4±3.2</td><td>33.3±57.7 / 73.3±23.1</td><td>100.0±0.0 / 100.0±0.0</td><td>12.5±21.7 / 40.0±29.1</td><td>26.4±10.5 / 51.4±9.0</td></tr><tr><td>FM (w/o TP)</td><td>50.0±25.0 / 66.7±16.7</td><td>50.0±50.0 / 70.4±28.0</td><td>0.0±0.0 / 60.0±0.0</td><td>33.3±57.7 / 33.3±57.7</td><td>0.0±0.0 / 30.0±12.0</td><td>28.7±24.2 / 53.3±17.1</td></tr><tr><td>FM</td><td>83.3±14.4 / 88.9±9.6</td><td>72.2±12.7 / 81.5±8.5</td><td>25.0±43.3 / 70.0±17.3</td><td>66.7±57.7 / 66.7±57.7</td><td>45.8±38.2 / 65.6±26.9</td><td>59.8±12.1 / 74.5±10.7</td></tr><tr><td colspan="7">GPT-4o</td></tr><tr><td>FM (w/o S&amp;T+TP)</td><td>75.0±25.0 / 83.3±16.7</td><td>52.8±50.2 / 77.8±21.0</td><td>0.0±0.0 / 50.0±17.3</td><td>66.7±57.7 / 66.7±57.7</td><td>33.3±28.9 / 58.9±19.5</td><td>43.7±29.3 / 67.8±14.2</td></tr><tr><td>FM (w/o TP)</td><td>91.7±14.4 / 94.4±9.6</td><td>77.8±19.2 / 85.2±12.8</td><td>66.7±57.7 / 86.7±23.1</td><td>100.0±0.0 / 100.0±0.0</td><td>62.5±21.7 / 76.7±14.5</td><td>74.7±5.3 / 83.5±6.2</td></tr><tr><td>FM</td><td>100.0±0.0 / 100.0±0.0</td><td>61.1±4.8 / 74.1±3.2</td><td>100.0±0.0 / 100.0±0.0</td><td>100.0±0.0 / 100.0±0.0</td><td>83.3±14.4 / 87.8±10.7</td><td>79.3±3.5 / 84.7±3.5</td></tr></table>
+
+Table 3: Facility Resource Stress Testing. The second row (1/1, 2/2) indicates the number of water dispensers and coffee machines. The left column (2, 4, 8) represents the number of agents. Each cell  $(\mathrm{X} / \mathrm{Y} / \mathrm{Z})$  shows the number of agents who drank water (X) and coffee (Y), and the total time (Z) for all agents to hydrate.  
+
+<table><tr><td></td><td colspan="2">All Like Water</td><td colspan="2">All Like Coffee</td><td colspan="2">No Preference</td></tr><tr><td></td><td>1/1</td><td>2/2</td><td>1/1</td><td>2/2</td><td>1/1</td><td>2/2</td></tr><tr><td>2</td><td>2/0/4</td><td>2/0/3</td><td>0/2/4</td><td>0/2/3</td><td>2/0/4</td><td>2/0/3</td></tr><tr><td>4</td><td>4/0/6</td><td>4/0/4</td><td>0/4/6</td><td>1/3/4</td><td>4/0/6</td><td>4/0/5</td></tr><tr><td>8</td><td>8/0/10</td><td>8/0/7</td><td>0/8/10</td><td>0/8/7</td><td>8/0/12</td><td>8/0/7</td></tr></table>
+
+completion rate (percentages of object state variables that are at the desired value). Removing task prioritization (TP) led to performance degradation across all models, with the most significant drop observed in Llama 3.3 and Gemma 3. This suggests that these models struggle with maintaining focus in multi-task scenarios and benefit substantially from explicit prioritization to reduce inefficient task switching. Further removing the semantic map and task progress tracking (S&T) led to additional performance degradation. GPT-4o exhibited a substantial drop in performance, indicating that it can effectively leverage the additional structured information for task planning and execution. In contrast, the removal of S&T had a more limited impact on Llama 3.3 and Gemma 3, likely due to their weaker information utilization capabilities. These models may struggle to extract key insights from the complex semantic map and task progress data, making it difficult for them to identify and prioritize the most relevant information. This observation suggests that while structured task representations are beneficial, their effectiveness is contingent on the model's ability to process and utilize the provided
+
+information efficiently. In addition, our results report standard deviations across three independent runs, which consistently show that the Full Model not only achieves the highest mean performance but also exhibits lower variability. This reduced variance further highlights the robustness of our complete framework with both S&T and TP components.
+
+In the experiments, we observe various collaborative behaviors among the agents. The collaboration and labor division reflects individual differences across agents. For example, Irene (IT Admin) asked Jenny (Janitor) for help moving a podium beyond her strength. (Figure 1 (b)). Collaboration does not only involve physical interaction, but also information sharing via communication. In a different session, Jeff is trying to reserve a meeting room, for which a password is needed. Initially, only the receptionists have knowledge about the password. Jeff first attempted a random password (hallucinated by LLMs). After failing, he asked the receptionist Ryan for the password, and successfully booked the meeting room (Figure 1 (b)).
+
+We observe that without task prioritization module, agents tend to switch back and forth between different tasks. For example, Jake is in the kitchen with an uncleaned fork_3 in his inventory. As a janitor, Jake should clean them as part of Task 2. However, he suddenly remembers that chair_4 needs to be moved to the open area 1 for Task 1, so he dropped fork_3 to pick up chair_4. But afterwards he recalls the former task again and picks fork_3 again. The behavior repeated and resulted in Jake not achieving any goals.
+
+Admittedly, even for models with a high completion rate, tasks may not always be completed in
+
+the most efficient way. For example, a receptionist may choose to clean the dishes themselves instead of asking for help from the janitor, who can clean dishes much faster. Lack of global coordination also sometimes leads to duplicated task completion by agents in different locations. For instance, Irene was unaware that her colleagues had already prepared enough tea in the open area 1, so she continued preparing tea in the kitchen and attempted to carry it over. Optimizing collective task completion efficiency remains a challenge.
+
+![](images/1fd23bf8a0c617d76861e22e51175949cda3d5ece32b67725a11ee646091eeeb.jpg)  
+Figure 3: Agent Presence Across Locations.
+
+![](images/e4940280e746319be6a571360bb2af8f9c19a2f79b2f6d4247641c37adfec95c.jpg)  
+Figure 4: Activity Time Distribution.
+
+![](images/35682743c47bb47e5635106192d875096f28dae4bf3098e29f18657e41d50bc7.jpg)  
+Figure 5: Agent Well-being Metrics.
+
+Resource Management Results are shown in Table 3. We observed that increasing the number of agents led to longer hydration times due to increased competition and queuing for resources. However, when the number of available water dispensers and coffee machines was increased, agents were able to hydrate more quickly, demonstrating that more resources can alleviate competition and improve efficiency.
+
+We also found that agent preferences significantly influenced their resource selection. When all agents preferred water, they consistently chose the water dispenser, even when a coffee machine was available. Similarly, when all agents preferred coffee, they almost exclusively used the coffee machine, ignoring the water dispenser. In the no-preference scenario, all agents opted for the water dispenser. This may suggest that in the absence of
+
+![](images/26c6b588f1bb61d4edf5f59fe502af90d8bb95e12676ca2818f13e3fe341c84c.jpg)  
+Figure 6: Suboptimal State Distribution.
+
+a strong preference, LLM-based agents default to an prior knowledge that biased towards water as the primary hydration method. This may also explain why, even when all agents preferred coffee, a small number of them still chose the water dispenser.
+
+As can be seen, simulation in INDOORWORLD can effectively reflect the impact of different resource allocation strategies, making it potentially a useful tool to aid real-world decision making on resource allocation.
+
+Spatial Layout Results illustrates the impact of different spatial layouts on agent behavior, focusing on location usage, activity time allocation, and overall well-being. The results indicate that spatial design significantly influences resource accessibility, social interactions, and agent efficiency and well-being.
+
+As shown in Figure 3, Design 1 maintained a more balanced distribution of agents across locations, while Design 2 saw a higher concentration in resource-rich areas, such as the pantry, leading to reduced workspace utilization, particularly for Desk Area 2, which is positioned in the farthest corner. This suggests that inefficient resource distribution may cause agents to move away from their designated workspaces, opting instead to work or engage in activities elsewhere.
+
+Figure 4 illustrates that agents in Design 2 spent more time moving and less time on role-related work due to longer distances from their workspace to resource-rich areas. Social interaction time was also higher, while time spent addressing physiological needs was lower. According to the simulation log, in Design 2, agents spent a significant portion of their time in the pantry engaging in conversations. As a result, despite their prolonged stay in the pantry, the time spent on physiological needs-related activities remained relatively low.
+
+Figure 5 and Figure 6 further show that agents in Design 2 spent less time in an optimal state and more time experiencing unmet needs, particularly hunger and fatigue. This can be attributed to the limited resources available in the pantry, such as only two pieces of bread, two apples and a few clean cups, which were quickly consumed.
+
+Implications for Architectural Design To evaluate how INDOORWORLD can support professional practice, we surveyed 9 architects and summarized their feedback in Table 6. The results indicate that our features (human need modelling, resource management, and spatial layout experiments) align closely with the considerations architects make during design, and the spatial layout experiments were rated as especially helpful for understanding occupant behavior. Results of the survey and discussions are provided in Appendix C.
+
+# 5 Conclusion
+
+This study introduces INDOORWORLD, a multiagent simulation environment that integrates fine-grained heterogeneous agent modeling with physical interactions. INDOORWORLD enables agents with varied abilities to coordinate roles, satisfy physiological needs, and interact with a rich array of objects within realistic settings.
+
+We evaluate three LLMs in collaborative task solving using a benchmark of common office tasks that involve both social and physical interactions, highlighting the effectiveness of task prioritization. Our simulation experiments on resource management and spatial layouts demonstrate the potential real-world applications of INDOORWORLD in architectural design.
+
+# Limitations
+
+Impact of Experimental Variability on Conclusions Due to the inherent randomness of LLMs, LLM-based agents may make different decisions even when facing identical scenarios. Furthermore, since our environment involves multiple agents making sequential decisions across multiple rounds, and their choices influence one another, the results may vary even when using the same experimental settings and LLM model. In our experiments, we observed that fluctuations in results were often associated with the following phenomena: (a) Agents engaging in prolonged conversations without progressing on task completion. (b) Agents performing incorrect actions after making substantial progress, leading to task failure. For instance, in one run of the GPT-4o Full model without S&T and TP (see Table 2), an agent unexpectedly moved a table, already placed in the open area 1, back to the kitchen. As a result, both the table and the clean utensils on it were relocated to an incorrect position, significantly lowering the overall task success
+
+rate.
+
+To mitigate the impact of experimental variability on the reliability of conclusions, we conducted three independent runs for each task-solving experiment. Additionally, we commit to publicly releasing our code after acceptance to enable further reproducibility and facilitate research on LLM-based agents' behavior, reasoning capabilities, and collaborative problem-solving.
+
+Scalability, Customization, and Future Improvements Although we have designed our environment to be easily scalable and customizable, the attributes of objects and the interaction mechanisms between objects and agents currently rely on user-defined configurations. This customization process is subject to user preferences and research objectives. In the future, we plan to follow the approach outlined in (Srivastava et al., 2021) by leveraging WordNet (Miller) to automatically generate object attributes and interaction methods. This will enable the rapid expansion of the environment's object library while allowing users to further tailor interactions through direct modifications in the Python code.
+
+Our current environment adopts a text-based game engine where all physical and visual interactions are abstracted. Moving forward, we aim to extend our framework to support 3D assets and physical simulation engines, enhancing the realism of agent-environment interactions. This expansion will allow researchers to flexibly utilize our environment in both abstract and more concrete settings, depending on their experimental needs.
+
+Current State of LLM-based Agents The practical utility of INDOORWORLD is currently constrained by the capabilities and costs of large language models. In our survey, architects acknowledged that they could obtain valuable insights from the interactions of LLM agents with the environment in INDOORWORLD, but these agents still fall short in approximating real human behavior. In addition, their deployment remains relatively expensive for large-scale, long-duration studies. Nevertheless, we are optimistic that as LLMs continue to improve, their behavioral outputs will become increasingly human-like, and that reductions in model costs will further enhance their usability, thereby strengthening the realism and applicability of our proposed framework.
+
+Testing of Reasoning Models Due to time and computational constraints, we did not test reasoning models such as GPT-O1 or Deepseek R1 (DeepSeek-AI et al., 2025) in our experiments. Consequently, some unsolved cases in our study may be successfully addressed by more advanced reasoning models, potentially leading to agent behaviors that better resemble human decision-making.
+
+# Acknowledgements
+
+This work was initially developed during Dekun Wu's internship at Autodesk Research. We thank Autodesk Research for their support and resources. This academic work is also supported in part by the Canada CIFAR AI Chair Program and the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant (RGPIN-2021-03115). We also acknowledge the use of GitHub Copilot Enterprise for low-novelty tasks such as code auto-completion and suggesting Python class definitions, which helped accelerate the coding process.
+
+# References
+
+Hongru Cai, Yongqi Li, Wenjie Wang, Fengbin Zhu, Xiaoyu Shen, Wenjie Li, and Tat-Seng Chua. 2024. Large language models empowered personalized web agents. Preprint, arXiv:2410.17236.  
+Hyungjoo Chae, Namyoung Kim, Kai Tzu iunn Ong, Minju Gwak, Gwanwoo Song, Jihoon Kim, Sunghwan Kim, Dongha Lee, and Jinyoung Yeo. 2024. Web agents with world models: Learning and leveraging environment dynamics in web navigation. Preprint, arXiv:2410.13232.  
+Marc-Alexandre Côté, Ákos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, Wendy Tay, and Adam Trischler. 2018. Textworld: A learning environment for text-based games. In Computer Games Workshop at ICML/II-CAI 2018, pages 1-29.  
+DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li,
+
+Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu Zhenda Xie Zhengyan Zhang Zhewen Hao Zhicheng Ma Zhigang Yan Zhiyu Wu Zihui Gu Zijia Zhu Zijun Liu Zilin Li Ziwei Xie Ziyang Song Zizheng Pan Zhen Huang Zhipeng Xu Zhongyu Zhang and Zhen Zhang. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint arXiv:2501.12948.  
+Xiaohang Feng, Da Yan, and Tianzhen Hong. 2015. Simulation of occupancy in buildings. Energy and Buildings, 87:348-359.  
+Chuang Gan, Jeremy Schwartz, Seth Alter, Damian Mrowca, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, Kuno Kim, Elias Wang, Michael Lingelbach, Aidan Curtis, Kevin Feigelis, Daniel M. Bear, Dan Gutfreund, David Cox, Antonio Torralba, James J. DiCarlo, Joshua B. Tenenbaum, Josh H. McDermott, and Daniel L. K. Yamins. 2021. Threadworld: A platform for interactive multi-modal physical simulation. Preprint, arXiv:2007.04954.  
+Zhenyu Guan, Xiangyu Kong, Fangwei Zhong, and Yizhou Wang. 2025. Richelieu: Self-evolving llmbased agents for ai diplomacy. Advances in Neural Information Processing Systems, 37:123471-123497.  
+Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang
+
+Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2024. MetaGPT: Meta programming for a multi-agent collaborative framework. In The Twelfth International Conference on Learning Representations.  
+Dong Huang, Jie M. Zhang, Michael Luck, Qingwen Bu, Yuhao Qing, and Heming Cui. 2024. Agentcoder: Multi-agent-based code generation with iterative testing and optimisation. Preprint, arXiv:2312.13010.  
+Yizhe Huang, Xingbo Wang, Hao Liu, Fanqi Kong, Aoyang Qin, Min Tang, Song-Chun Zhu, Mingjie Bi, Siyuan Qi, and Xue Feng. 2025. Adasociety: An adaptive environment with social structures for multiagent decision-making. Preprint, arXiv:2411.03865.  
+Tejas D Kulkarni, William F. Whitney, Pushmeet Kohli, and Josh Tenenbaum. 2015. Deep convolutional inverse graphics network. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc.  
+Bokyung Lee, Michael Lee, Jeremy Mogk, Rhys Goldstein, Jacobo Bibliowicz, Frederik Brudy, and Alexander Tessier. 2021. Designing a multi-agent occupant simulation system to support facility planning and analysis for covid-19. In Proceedings of the 2021 ACM Designing Interactive Systems Conference.  
+Sicong Leng, Yang Zhou, Mohammed Haroon Dupty, Wee Sun Lee, Sam Joyce, and Wei Lu. 2023. Tell2Design: A dataset for language-guided floor plan generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Chengyuan Li, Tianyu Zhang, Xusheng Du, Ye Zhang, and Haoran Xie. 2025. Generative ai models for different steps in architectural design: A literature review. Frontiers of Architectural Research.  
+Nian Li, Chen Gao, Mingyu Li, Yong Li, and Qing-min Liao. 2024. Econagent: large language model-empowered agents for simulating macroeconomic activities. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15523–15536.  
+Yuanyuan Liu, Ying Zhou, Le Yang, and Yangpeng Xin. 2024. Simulating staff activities in healthcare environments: An empirical multi-agent modeling approach. Journal of Building Engineering.  
+Pattie Maes. 1995. Artificial life meets entertainment: lifelike autonomous agents. Communications of the ACM, 38(11):108-114.  
+George A. Miller. Wordnet: a lexical database for english. Commun. ACM.  
+Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, Xintong Li, Jing Shi, Hongjie Chen, Viet Dao Lai, Zhouhang Xie, Sungchul Kim, Ruiyi Zhang, Tong Yu, Mehrab Tanjin, Nesreen K.
+
+Ahmed, Puneet Mathur, Seunghyun Yoon, Lina Yao, Branislav Kveton, Thien Huu Nguyen, Trung Bui, Tianyi Zhou, Ryan A. Rossi, and Franck Dernoncourt. 2024. Gui agents: A survey. Preprint, arXiv:2412.13501.  
+Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In In the 36th Annual ACM Symposium on User Interface Software and Technology (UIST '23), UIST '23, New York, NY, USA. Association for Computing Machinery.  
+Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. 2018. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8494-8502.  
+Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2023. Chatdev: Communicative agents for software development. arXiv preprint arXiv:2307.07924.  
+Alexander Raistrick, Lingjie Mei, Karhan Kayan, David Yan, Yiming Zuo, Beining Han, Hongyu Wen, Meenal Parakh, Stamatis Alexandropoulos, Lahav Lipson, Zeyu Ma, and Jia Deng. 2024. Infinigen indoors: Photorealistic indoor scenes using procedural generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21783-21794.  
+Davide Schaumann, Simon Breslav, Rhys Goldstein, Azam Khan, and Yehuda E Kalay. 2017. Simulating use scenarios in hospitals using multi-agent narratives. Journal of Building Performance Simulation, 10(5-6):636-652.  
+Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36.  
+Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. 2020a. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2020b. Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768.  
+Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2021. ALFWorld: Aligning Text and Embodied Environments for Interactive Learning.
+
+In Proceedings of the International Conference on Learning Representations (ICLR).  
+Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martin-Martin, Fei Xia, Kent Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, C. Karen Liu, Silvio Savarese, Hyowon Gweon, Jiajun Wu, and Li Fei-Fei. 2021. Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. Preprint, arXiv:2108.03332.  
+Xiangru Tang, Anni Zou, Zhuosheng Zhang, Ziming Li, Yilun Zhao, Xingyao Zhang, Arman Cohan, and Mark Gerstein. 2024. MedAgents: Large language models as collaborators for zero-shot medical reasoning. In *Findings of the Association for Computational Linguistics: ACL* 2024, pages 599-621, Bangkok, Thailand.  
+Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei, and Ji-Rong Wen. A Survey on Large Language Model based Autonomous Agents. Preprint, arXiv:2308.11432.  
+Xiaoqiang Wang and Bang Liu. 2024. Oscar: Operating system control via state-aware reasoning and re-planning. arXiv preprint arXiv:2410.18963.  
+Yiding Wang, Yuxuan Chen, Fangwei Zhong, Long Ma, and Yizhou Wang. 2024. Simulating human-like daily activities with desire-driven autonomy. arXiv preprint arXiv:2412.06435.  
+Zhilin Wang, Yu Ying Chiu, and Yu Cheung Chiu. 2023. Humanoid agents: Platform for simulating human-like generative agents. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 167-176, Singapore. Association for Computational Linguistics.  
+Dekun Wu, Haochen Shi, Zhiyuan Sun, and Bang Liu. Deciphering digital detectives: Understanding LLM behaviors and capabilities in multi-agent mystery games. In Findings of the Association for Computational Linguistics: ACL 2024.  
+Jiajun Wu, Erika Lu, Pushmeet Kohli, Bill Freeman, and Josh Tenenbaum. 2017. Learning to see physics via visual de-animation. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.  
+Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu, and Yang Liu. 2023. Exploring large language models for communication games: An empirical study on werewolf. arXiv preprint arXiv:2309.04658.  
+Da Yan, William O'Brien, Tianzhen Hong, Xiaohang Feng, H Burak Gunay, Farhang Tahmasebi, and Ardeshir Mahdavi. 2015. Occupant behavior modeling for building performance simulation: Current state and future challenges. Energy and buildings, 107:264-278.
+
+Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629.  
+Xianhao Yu, Jiaqi Fu, Renjia Deng, and Wenjuan Han. 2024. Mineland: Simulating large-scale multi-agent interactions with limited multimodal senses and physical needs. Preprint, arXiv:2403.19267.  
+Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang Gan. 2024. Building cooperative embodied agents modularly with large language models. ICLR.
+
+# A Task Coordination Details
+
+In Table 4, we list the specific task descriptions used in our collaborative task-solving experiment. During the experiment, all five tasks were simultaneously input into the memory of all agents. Due to agent heterogeneity, certain task components require specific agents to complete or can be completed more efficiently by particular agents. Therefore, agents must determine which part of the task they should be responsible for and then take action simultaneously.
+
+Hierarchically, the highest-level task is preparing the company event, under which the five tasks listed in 12 serve as sub-tasks. Each task description further contains low-level tasks, such as the task Prepare Enough Tables and Chairs in the Event Area, which includes sub-tasks like Move 2 tables and 2 chairs to the open area 1 from other locations. These sub-tasks, in turn, may require multiple steps to complete.
+
+This multi-level hierarchical task structure evaluates not only the agents' planning abilities but also their collaboration and communication skills.
+
+# B Architect Survey on Multi-level Heterogeneity
+
+Table 5 compares the modeling of fine-grained heterogeneity types across four heterogeneous multiagent environments. Our approach is the only one that comprehensively supports all four types of heterogeneity: profile, action, capability, and knowledge. To validate the necessity of modeling these levels of heterogeneity, we conducted a survey with 20 architects, asking for their views on the realism, importance, and consideration of heterogeneity modeling.
+
+Participants and Survey Design Twenty practicing architects with at least 3 years' experience were recruited through the external platform UserTesting.com. Compensation levels were set in accordance with internal company guidelines, and in all cases participants were remunerated above the minimum wage in their regions. Each participant signed a consent form that explained the study purpose and data use and before completing an online questionnaire. For each of the four heterogeneity levels introduced in Section 3.2, respondents (i) compared two design options,  $A$ : fully homogeneous occupants, and  $B$ : heterogeneous occupants, and indicated which looked more realistic;
+
+(ii) evaluated how important the heterogeneity was for understanding real space use (5-point Likert scale); (iii) reported whether they normally consider that difference in their own design work, and, if not, why. This study was reviewed by and received approval through Autodesk's internal ethics review process.
+
+Results Overview Figure 7 shows that heterogeneity was judged more realistic by  $55 - 70\%$  of architects depending on the level, with the highest endorsement for profile, action and capability heterogeneity. Importance ratings (Fig. 8) follow a similar trend:  $55 - 95\%$  of respondents marked the heterogeneity as "important" or "very important". Figure 9 shows that most architects already incorporate heterogeneity consideration in their own workflow: all 20 architects consider profile-level differences, and large majorities do so for action  $(90\%)$ , capability  $(60\%)$ , and knowledge  $(65\%)$  heterogeneity. This strong uptake suggests that heterogeneous agent modeling is both familiar and valued in practice, particularly at the profile and action levels. However, some architects still do not consider these differences in their design work, citing various reasons. Fig. 10 presents the further analysis of these reasons. For action-level heterogeneity, the two architects who do not consider it pointed to external factors: one mentioned that clients or stakeholders did not require this level of detail, while the other cited having to focus on more important constraints. At the capability level, the reasons were more diverse. While  $25\%$  of those who opted out mentioned stakeholder requirements, another  $25\%$  indicated that this type of variation is usually ignored in current design workflows. An additional  $25\%$  referenced focusing on more pressing design aspects, while the remaining  $25\%$  were divided between not being aware this kind of modeling was possible and lacking the tools or data to support it. At the knowledge level, the primary reason for non-consideration was a perceived lack of importance. Other respondents mentioned stakeholder scope  $(29\%)$ , and a smaller portion cited workflow limitations or practical constraints.
+
+Implications This architect survey indicates that our multi-level heterogeneity yields more realistic simulations than existing environments and aligns more closely with the factors architects consider during design. This feature not only justifies the design choices behind INDOORWORLD, but also strengthens its potential as a supportive tool for
+
+Table 4: Task Descriptions  
+
+<table><tr><td>Task Number</td><td>Task Name</td><td>Description</td></tr><tr><td>1</td><td>Prepare Enough Tables and Chairs in the Event Area</td><td>Move 2 tables and 2 chairs to the open area 1 from other locations.</td></tr><tr><td>2</td><td>Prepare Clean Utensils for the Event</td><td>Transport 4 clean plates, 4 clean knives, and 4 clean forks from other locations to the open area 1 and place them on the tables.</td></tr><tr><td>3</td><td>Check and Repair Broken Computer, Pro-jector, and Microphone</td><td>Move a podium to the open area 1 if there is no podium there. Bring one computer, one projector, and one microphone to the open area 1 and place them on the podium. Ensure they are in working condition.</td></tr><tr><td>4</td><td>Book a Meeting Room for the Event</td><td>Use the touch screen in the open area 1 or a computer to remotely reserve the open area 1. Ensure the meeting room is booked with the following details: Event Name: Lunch and Listen; Start time: 2024-09-02T12:00:00; End time: 2024-09-02T13:00:00.</td></tr><tr><td>5</td><td>Prepare Coffee, Tea, and Lunch for the Event</td><td>Take clean cups and use them to prepare 3 cups of coffee and 3 cups of tea and place them on the tables in the open area 1. Bring 2 meals to the open area 1 and place them on the tables in the open area 1. Ensure that meals are heated.</td></tr></table>
+
+architectural practice.
+
+# C Architect Survey on Design Problem Relevance and Tool Usefulness
+
+Participants and Survey Design Nine practicing architects (3-15 years of experience) were recruited through Autodesk's internal employee mailing lists and snowball sampling to rate ten statements on a five-point Likert scale (  $1 =$  Strongly Disagree,  $2 =$  Disagree,  $3 =$  Neutral,  $4 =$  Agree,  $5 =$  Strongly Agree). Participants were compensated with $35 USD for their participation. Each participant provided informed consent through a form that explained the study purpose and data use before completing an online questionnaire. This study was reviewed and approved through Autodesk's internal ethics review process.
+
+The questionnaire comprised two blocks:
+
+1. Relevance of Design Problems (q1-q4): This section assessed the perceived importance of key design issues, including resource alloca
+
+tion, spatial layout, human-need considerations, and waiting-time analysis, in everyday architectural practice.
+
+2. Usefulness of INDOORWORLD (q5-q10): This section evaluated whether the proposed simulation tool supports architects in understanding and addressing these design problems, as well as improving their design decisions.
+
+Questions and average scores are summarized in Table 6.
+
+Key Findings The survey results indicate strong recognition of the relevance of the design problems addressed. All four items in the first block received high average ratings  $(\mathbf{M} = 3.8 - 4.7)$ . Notably, the importance of spatial layout (q2,  $\mathbf{M} = 4.7$ ) was overwhelmingly endorsed, reflecting its critical role in influencing occupant behavior and well-being. Similarly, the significance of analyzing waiting times or competition for shared resources
+
+![](images/1a731dc9f8b702fc72ad0ba1212cbcaeb8c00cc6f556006149927e61928cd903.jpg)  
+Figure 7: Realism comparison (Homo.  $=$  homogeneity, Hetero.  $=$  heterogeneity). A majority consider heterogeneity more realistic than homogeneity at every level.
+
+<table><tr><td>Environment</td><td>Profile</td><td>Action</td><td>Capability</td><td>Knowledge</td></tr><tr><td>D2A (Wang et al., 2024)</td><td>✓</td><td>X</td><td>X</td><td>-</td></tr><tr><td>Humanoid Agents (Wang et al., 2023)</td><td>✓</td><td>X</td><td>X</td><td>-</td></tr><tr><td>Smallville (Park et al., 2023)</td><td>✓</td><td>X</td><td>X</td><td>✓</td></tr><tr><td>Ours</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+Table 5: Comparison of fine-grained heterogeneity types across heterogeneous multi-agent environments. “Yes” indicates support for a type, “-” means not mentioned.
+
+(q4,  $\mathbf{M} = 4.0$ ) was highlighted, affirming the relevance of this problem to practical design.
+
+For tool usefulness, respondents generally expressed positive views regarding the benefits of INDOORWORLD. The simulation was appreciated for its ability to reveal layout impacts (q6,  $\mathbf{M} = 4.0$ ), support design refinement processes (q7,  $\mathbf{M} = 3.9$ ), and spark interest for direct use in design projects (q9,  $\mathbf{M} = 4.1$ ). Its potential to bridge the gap between architectural design and user behavior was also well-received (q10,  $\mathbf{M} = 4.2$ ). These responses suggest that architects see INDOORWORLD as a promising tool for enhancing their design workflows, offering insights that can inform spatial planning and resource management.
+
+However, two items received only neutral-to-slightly-positive ratings: actionable insights from resource-competition simulation  $(q5, M = 3.1)$  and the realism of observed agent behavior  $(q8, M = 3.0)$ . These lower scores reflect a current gap between the simulation's fidelity and user expectations. Respondents generally acknowledged the
+
+concept of resource competition but found "hydration competition" less realistic compared to competition for meeting spaces or other workspace resources, which were viewed as more contextually relevant. One architect expressed concerns that the current game-like interface of INDOORWORLD, where agents continuously move around and interact with each other, might deter some architects. They suggested that the simulation could be executed in the background, with results presented as concise, actionable recommendations for spatial planning, minimizing unnecessary visual complexity. Despite these reservations, respondents recognized that the simulation provides useful insights and proposed several enhancements. These included introducing more diverse agent profiles with varying needs and preferences to better assess universal design principles and accessibility, as well as enabling the visualization of agent movement flows to enhance spatial analysis.
+
+Implications Overall, the survey confirms that our modeling choices address key design chal
+
+![](images/98976fa69c3f09330d7c787a9512b44820816ae2f0438b17c4634a42b5ff42de.jpg)  
+Figure 8: Importance ratings for understanding real space use (5-point scale). Most participants rate heterogeneity as "important" or "very important" at each level.
+
+<table><tr><td colspan="2">ID Question</td><td>Avg</td></tr><tr><td colspan="2">Relevance of Design Problems</td><td></td></tr><tr><td>q1</td><td>Resource allocation is a critical factor in office space planning.</td><td>3.8</td></tr><tr><td>q2</td><td>Spatial layout significantly affects occupant behavior and well-being.</td><td>4.7</td></tr><tr><td>q3</td><td>I consider human needs when designing building layouts.</td><td>3.8</td></tr><tr><td>q4</td><td>Evaluating waiting time or competition for shared resources is useful for improving space efficiency.</td><td>4.0</td></tr><tr><td colspan="2">Usefulness of INDOORWORLD</td><td></td></tr><tr><td>q5</td><td>The simulation of resource competition provides actionable insights for real-world resource placement.</td><td>3.1</td></tr><tr><td>q6</td><td>The spatial layout experiment helps reveal how different designs impact occupant behavior and efficiency.</td><td>4.0</td></tr><tr><td>q7</td><td>I can imagine using such simulations to evaluate and refine my own design decisions.</td><td>3.9</td></tr><tr><td>q8</td><td>The observed agent behavior reflects realistic office usage patterns.</td><td>3.0</td></tr><tr><td>q9</td><td>I would be interested in trying out this simulation tool with my own design layouts.</td><td>4.1</td></tr><tr><td>q10</td><td>I believe simulation tools like this can bridge the gap between architectural design and human behavioral modeling.</td><td>4.2</td></tr></table>
+
+Table 6: Average Likert ratings (1 = Strongly Disagree, 2 = Disagree, 3 = Neutral, 4 = Agree, 5 = Strongly Agree) provided by nine practicing architects, summarizing perceived relevance of key design problems and the usefulness of the proposed INDOORWORLD simulation tool.
+
+lenges that architects recognize as important. The high scores for spatial layout and resource competition demonstrate that the simulated scenarios align well with real-world concerns. The positive feedback on tool usefulness suggests that INDOOR-WORLD is perceived as a valuable support tool for architectural design, providing insights that can inform spatial planning, resource management, and design optimization.
+
+Moreover, the relatively lower scores on agent behavior realism and the actionability of insights indicate areas for improvement. These results highlight a common challenge for many simulation tools: accurately capturing the complexity of human behavior while ensuring that the generated insights are directly applicable to design decisions.
+
+Currently, LLM-based agents are at an early stage of development, with simplified behaviors and limited contextual awareness. We anticipate that ongoing advancements in LLM capabilities will enhance agent realism, enabling more nuanced interactions and generating insights that align more closely with real-world user expectations.
+
+# D Action-space comparison
+
+Table 7 summarizes the size of the action space supported by several widely used task-oriented simulation environments. Most benchmarks provide fewer than 20 primitive actions: VirtualHome offers 18, ALFWorld 9, and the TDW variants between 7 and 12. Our environment markedly expands this spectrum to 38 distinct actions, more
+
+![](images/c0d2e415dec0d7237c11178f29880408f740d6f35925db8c4effadb5429fd780.jpg)  
+Figure 9: Whether architects typically consider each heterogeneity type in their own work.
+
+<table><tr><td></td><td>#Actions</td></tr><tr><td>Virtual-Home (Puig et al., 2018)</td><td>18</td></tr><tr><td>ALFWorld (Shridhar et al., 2021)</td><td>9</td></tr><tr><td>TDW-MAT (Zhang et al., 2024)</td><td>7</td></tr><tr><td>C-WAH (Zhang et al., 2024)</td><td>8</td></tr><tr><td>ALFRED (Shridhar et al., 2020a)</td><td>13</td></tr><tr><td>TDW (Gan et al., 2021)</td><td>12</td></tr><tr><td>AdaSociety (Huang et al., 2025)</td><td>12</td></tr><tr><td>Ours</td><td>38</td></tr></table>
+
+Table 7: Comparison of the number of actions supported by different environments.
+
+than twice that of any prior system listed. The richer verb set enables agents to engage in a broader range of household manipulations (e.g., brewCoffee, repairDevice, heatFood), which in turn supports more realistic task decompositions, greater behavioral diversity, and nuanced capability differences.
+
+Importantly, the set of 38 actions was designed based on two principles: (1) Common human actions shared across roles, such as go_to, take, and put, ensuring coverage of generic physical interactions. We also incorporated basic need-related actions such as eating, drinking, and restroom use, which are absent in ALFWorld (Shridhar et al., 2020b) but are critical for simulating realistic human behavior. (2) Role-specific actions tailored to each agent type, such as repair_electronic_device for IT admin or software_development for software engineers, reflecting distinct occupational responsibilities and supporting our environment's heterogeneity. As
+
+further described in Appendix E.2, the action space can be easily extended with new roles and actions, allowing researchers to customize it according to their needs. This flexibility is important for giving researchers greater control over scenario design.
+
+# E Resources
+
+We release the INDOORWORLD environment under an MIT license, comprising the codebase, configuration files, experimental setups, and detailed documentation covering installation and usage. These resources are intended to facilitate reproduction of our experiments and enable further extensions by the research community. Interested parties can request access by contacting the first author.
+
+# E.1 Environment Initialization via JSON
+
+Our environment is defined using a JSON configuration file, which specifies key components such as agents, locations, inter-location connections, receptacles within locations, and objects. This structured definition allows for flexible customization of the simulation environment.
+
+Our code automatically initializes the environment based on the JSON file, creating corresponding instances of Location, Receptacle, Object, and Agent classes. Each element in the JSON file is mapped to its respective class, ensuring that all objects and agents are correctly instantiated with their predefined attributes and relationships.
+
+An example of this JSON configuration is shown
+
+![](images/2116069fdc4f4fad8caedadb9ed3013cf3d651a0fac3a47d11bf453c36506e64.jpg)  
+Action "No"-answer reasons (10% No)  
+Figure 10: Reasons given by those who answered "No" in Fig. 9.
+
+![](images/c0327f293f910751c0929a20ce1c7bbe2f0bc8f9bb15a3a80438cb951e7a36f8.jpg)  
+Capability "No"-answer reasons (40% No)
+
+![](images/da2fd9eeef966738291168ddddef005d30e648f78fdeea1cce9a31000050a666.jpg)  
+Knowledge "No"-answer reasons (35% No)
+
+Clients Or Stakeholders Didn't Require This Level Of Detail.  
+I Had To Focus On More Important Constraints.  
+This Type Of Variation Is Usually Ignored In Current Design Workflows.
+
+I Didn't Consider It Because I Didn't Think It Was Important.  
+I Wasn't Aware This Kind Of Modeling Was Possible.  
+I Didn't Have The Tools Or Data To  
+Estimate How People Would Use The Space.
+
+in Figure 11, where locations, objects, receptacles, and agents are defined. This JSON-based approach enables researchers to easily modify and extend the environment without changing the core simulation code, making it a highly adaptable framework for various research scenarios.
+
+# E.2 Extending Agent-Object Interactions
+
+Figures 12, 13, and 14 provide three example classes that illustrate how researchers can extend the environment by adding new object types and by defining interactions between agents and those objects.
+
+For example, if a researcher wants to introduce a new class, such as Computer, they first need to specify what actions agents can perform on it, such as turn_on. This is done by defining interaction methods like power_on() within the Computer class. The new class should inherit from BaseObject and implement get_admissible_actions(), which determines when and how agents can interact with the object. For instance, if the computer is in an accessible state, it can return a command like turn_on {self.name}.
+
+On the agent side, new interaction methods must be defined accordingly. For instance, the Agent class should implement a turn_on() method that handles the interaction logic for powering on a Computer. Additionally, in the act() function, a new command mapping should be added, linking the turn_on command to the
+
+turn_on() method. This way, when the agent's get_admissible ACTIONS() method runs, it will include the new turn_on action if the object is in the same location as the agent and is not placed inside a closed receptacle. The agent can then decide whether to execute this action, effectively enabling interaction with the newly introduced Computer object.
+
+The ITAdmin class further demonstrates how to define role-specific actions. This is done by first adding "repair_computer" to the skills dictionary. Then, a new method, such as repair_electronic_device(), is defined, and the act() function maps the "repair_computer" command to this method. This allows the ITAdmin role to perform repair actions that are unavailable to other agents.
+
+Note: These examples have been simplified for clarity and may not exactly match the original source code. Researchers should refer to the actual implementation details to fully integrate new objects and interactions into the environment.
+
+```txt
+"locations": ["kitchen", "meeting_room1"], "location_distance": { "kitchen": {"meeting_room1": 1}, "meeting_room1": {"kitchen": 1}] }, "receptacles": [ {"name": "Sinkbasin1", "location": "kitchen", "rtype": "Sinkbasin", "weight_kg": 15, "state": {"fixed": true, "closable": false, "is_open": true, "is.clean": true, "temperature": 20, "is_workinq": true}}, {"name": "Cabinet1", "location": "kitchen", "rtype": "Cabinet", "weight_kg": 30, "state": {"fixed": true, "closable": true, "is_open": false, "is.clean": true, "temperature": 20, "is_workinq": true}} ], "objects": [ {"name": "touchscreen_1", "otype": "TouchScreen", "location": "meeting_room1", "weight_kg": 3, "state": {"is_turned_on": true, "is_working": true, "is Clean": true, "temperature": 20}}, {"name": "cup_1", "otype": "Cup", "location": "kitchen", "receptacle": "Counttop1", "weight_kg": 0.3, "state": {"is clean": false, "temperature": 20}} ], "agents": [ { "name": "ryan", "gender": "male", "role": "receptionist", location: "meeting_room1", "fullness": 100, hydration: 100, energy: 100, social Fulfillment: 100, strength_kg': 65, internal_profile": Ryan is a professional and welcoming receptionist. Known for his friendly personality and exceptional communication skills....," appearance": Ryan is a receptionist who is tall and well-built, with ... } { name": "irene", gender": "female", role": "IT_admin", location: "kitchen", fullness: 100, hydration: 100, energy: 100, social Fulfillment: 100, strength_kg': 30, internal_profile": Irene is an organized and skilled IT administrator. She is quick to troubleshoot and efficiently repair a wide range of ..., appearance": Irene has a petite, tidy appearance with a focused expression. She is usually dressed casually but professionally, with an attentive ... } ]
+```
+
+Figure 11: Excerpt from the JSON configuration file defining locations, agents, objects, and receptacles in the environment.
+
+```python
+1 class Computer(BaseObject):
+2 def __init__(self, name, otype='Computer', location, environment, weight_kg,
+3 carryable=False, requires_receptacle=True,
+4 state={'is_clean': True,'temperature': 20}):
+5 super()).__init__(name, otype, location, environment, weight_kg,
+6 carryable, requires_receptacle, state=state)
+7 def power_on(self):
+8 if self.state['is_turned_on]:
+
+9 return f'[self.name] is already turned on.", False
+10 self.state['is_turned_on'] = True
+11 return f'[self.name] is now turned on.", True
+12 ...
+13 .......
+14
+15 def get_admissible ACTIONS(self, agent):
+16 actions = super().get_admissible ACTIONS(agent)
+17
+18 # Add repair action
+19 if f'repair{'self.otype.low'}' in agentskills:
+20 actions.append(f"repair{'self.otype.low'} {self.name}") 
+21
+22 # Check if the computer has an 'owner' attribute
+23 if self.location == agent.location:
+24 if not self.state['is_turned_on':
+25 actions.append(f"turn_on {self.name}") 
+26 else:
+27 actions.append(f"turn_off {self.name}") 
+28
+29 return actions
+```
+
+Figure 12: Illustration of the Computer class, defining interaction methods like power_on() and specifying admissible agent actions.
+
+Figure 13: Illustration of the Agent class, where act() maps commands to interaction methods like turn_on().  
+```python
+class Agent: def __init__(self, name, role, location, skills=None): self.name = name self.root = role self.location = location selfskills = skills if skills else {}   
+def act(self, command): """Execute an action if it's admissible."""command_mapping  $=$  {"turn_on": (self.turn_on,1), # 1 means one argument ... }   
+tokens  $\equiv$  command.split() if len(tokens)  $\geqslant 1$  .. action  $\equiv$  tokens[0] if action in command Mapping: action_func, arg_count  $\equiv$  command Mapping[action] if len(tokens[1:])  $\rightharpoondown$  arg_count: args  $\equiv$  tokens[1:1 + arg_count] return action_func(*args) else: return f"\{self.name\} received an incorrect number of arguments for action {action}.", False else: return f"\{self.name\} cannot perform action {action}.", False return f"\{self.name\} cannot parse command {command}.", False   
+def turn_on(self, electronic_device_name):   
+# Search for the electronic device in the current location's objects target_device  $\equiv$  next((obj for obj in self.location.objects if obj.name  $= =$  electronic_device_name), None)   
+# If the device is found, attempt to turn it on if target_device: if hasattr(target_device, "turn_on"): return target_device-turn_on() else: return f"\{target_device.name\} cannot be turned on.", False else: return f"\{self.name\} cannot find {electronic_device_name}.", False   
+def get_admissible ACTIONS(self): admissible ACTIONS  $=$  []   
+... if self.location.objects: for obj in self.location.objects: if obj.receptacle  $= =$  None or obj.receptacle.state['is_open']: admissible ACTIONS extend(obj.get_admissible ACTIONS(self))   
+return admissible ACTIONS
+```
+
+```python
+1 class ITAdmin(Agent):   
+2 def __init__(self, name, gender, location, environment, **kwargs): super().__init__(name, gender, location, environment, **kwargs)   
+4 # ITAdmin specific skills for repairing devices self.skills.update({ "repair_computer": 1, }   
+6 }   
+7 }   
+8 }   
+9   
+10 def repair_electronic_device(self, obj_name):   
+11 # Find the object with the specified name in the current location target_device = next((obj for obj in self.location.objects if obj.name == obj_name), None)   
+13   
+14 if target_device:   
+15 # Check if the object is a repairable electronic device electronic Devices  $=$  ["Microphone", "Projector", "Computer", "CoffeeMachine", "WaterDispenser", "Microwave"]   
+17 if target_device.otype in electronicDevices: # Check if the device is broken   
+19 if "is-working" in target_device.state and not target_device.state[" is_working']: target_device.state["is-working"] = True return f"\{self.name\} repaired the {target_device.name}.True else: return f"\{target_device.name\} is already in working condition.", False   
+20 else: return f"\{target_device.name\} is not a repairable electronic device.", False   
+21 def act(self, command): """Execute an ITAdmin-specific action or fall back to the Agent's act method."""command Mapping  $=$  { "repair_computer": (self.repair_electronic_device, 1),   
+32 }   
+33 }   
+34 }   
+35   
+36 tokens  $=$  command.split()   
+37 if len(tokens)  $\geqslant$  1 and tokens[0] in command Mapping: action_func, arg_count  $=$  command Mapping(tokens[0]) if len(tokens[1:])  $\geqslant$  arg_count: return action_func(*tokens[1:1 + arg_count])   
+41 return f"\{self.name\} received an incorrect number of arguments.", False   
+42   
+43 return super().act(command)
+```
+
+Figure 14: Illustration of the ITAdmin class, where repair_computer() enables role-specific actions via command mapping.

2025/INDOORWORLD _ Integrating Physical Task Solving and Social Simulation in A Heterogeneous Multi-Agent Environment/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3058162c4c5b02c6a075109886918cc8959edaac39e5ef98c512ff88ba594ea
+size 827287

2025/INREACT_ An Inspire-Then-Reinforce Training Framework For Multimodal GUI Agent/889785bd-0923-422b-bcca-82b35b178935_content_list.json

@@ -0,0 +1,2052 @@
+[
+    {
+        "type": "text",
+        "text": "INREACT: An Inspire-Then-Reinforce Training Framework For Multimodal GUI Agent",
+        "text_level": 1,
+        "bbox": [
+            164,
+            89,
+            836,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yuanlei Wang $^{1*}$ , Liuzhou Zhang $^{2*}$ , Haohao Luo $^{1*}$ , Ying Shen $^{1,3\\dagger}$",
+        "bbox": [
+            223,
+            145,
+            776,
+            162
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>Sun Yat-sen University, <sup>2</sup>Central China Normal University,",
+        "bbox": [
+            257,
+            162,
+            739,
+            179
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>3</sup>Guangdong Provincial Key Laboratory of Fire Science and Intelligent Emergency Technology",
+        "bbox": [
+            112,
+            179,
+            880,
+            199
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{wangylei7, luohh5}@mail2.sysu.edu.cn,zlz212@mail.ccnu.edu.cn,",
+        "bbox": [
+            186,
+            197,
+            811,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "sheny76@mail.sysu.edu.cn",
+        "bbox": [
+            374,
+            214,
+            621,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            342,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Graphical User Interface (GUI) interaction, which aims to develop an intelligent GUI agent that executes user instructions to perform tasks such as installing applications by controlling digital devices, has gained significant attention due to its practical value. Although current advanced multimodal large language models (LLMs) provide GUI agents with robust perception and reasoning capabilities, they often struggle with the precise localization of small elements. To tackle this problem, we propose INREACT, a multimodal GUI agent framework that unifies observing, thinking, and acting for precise and interpretable decision-making. It is trained via a two-stage process: curriculum learning to progressively build perception, grounding, and reasoning abilities, followed by reinforcement learning to refine pixel-level grounding with an outcome-based reward. We introduce a rule-based reward function that jointly optimizes action-type selection and pixel-level localization accuracy. Experimental results on multiple datasets demonstrate the superiority of INREACT in both grounding and navigation tasks.",
+        "bbox": [
+            141,
+            291,
+            460,
+            649
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            663,
+            260,
+            678
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Graphical User Interface (GUI) interaction refers to a class of sequential decision-making tasks in which agents interpret user instructions and manipulate digital interfaces, including PCs and mobile devices, to accomplish specific goals. These interactions typically involve executing low-level operations like Click, Type, or Scroll in the correct order to complete complex tasks, such as searching for information. A capable GUI agent must possess two core abilities: visual observation and reasoning. Observation is particularly critical in GUI environments like PCs and mobile devices, where visual",
+        "bbox": [
+            112,
+            690,
+            490,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e38ea4c38d9f3067a6c4b2ec43e2927e4c1efa3cd666b264ccf89a7a81d8f870.jpg",
+        "image_caption": [
+            "Figure 1: Overview of the INREACT interaction"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            510,
+            258,
+            878,
+            395
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "components are often densely packed and visually similar, making precise localization challenging. Meanwhile, reasoning is essential for complex, goal-directed interactions that involve sequential decision-making. Unlike single-step tasks, GUI interaction requires agents to consider both interface states and past actions. With the emergence of powerful multimodal large language models (MLLMs) (Wang et al., 2024c; Bai et al., 2025; Chen et al., 2023; Luo et al., 2024), recent approaches have enabled screenshot-based GUI control without relying on structured metadata such as HTML or accessibility trees (Zhang et al., 2025; Wang et al., 2024b). However, despite these advances, existing GUI agents still struggle with accurate grounding and coherent reasoning in complex, real-world scenarios.",
+        "bbox": [
+            507,
+            434,
+            885,
+            705
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address the challenges of imprecise grounding and weak multi-step reasoning, recent efforts have focused on training MLLM-based GUI agents via supervised fine-tuning (SFT) or reinforcement learning (RL). In SFT, models are trained on paired GUI-action datasets to align visual observations with user instructions. Some approaches (Hong et al., 2024; Cheng et al., 2024; Lin et al., 2024; Li et al., 2025a) enhance SFT by incorporating instruction tuning, vision-language pretraining, or synthetic data generation to boost coverage. However, these strategies commonly adopt a flat optimization objective, lacking curriculum or structural",
+        "bbox": [
+            507,
+            712,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal Contribution",
+        "bbox": [
+            134,
+            892,
+            263,
+            906
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author",
+        "bbox": [
+            134,
+            906,
+            278,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "9148",
+        "bbox": [
+            478,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 9148-9160",
+        "bbox": [
+            216,
+            945,
+            781,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            280,
+            958,
+            715,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "guidance, which limits the model's ability to progressively acquire complex capabilities. In contrast, reinforcement learning introduces outcome-driven optimization through environment feedback. Some methods (Lai et al., 2024; Liu et al., 2024) rely on sparse success signals to guide multi-step task completion. Despite this, many RL strategies remain coarse and fail to capture pixel-level grounding precision. Moreover, most RL pipelines lack strong initialization or structured learning stages, often resulting in unstable policies and inconsistent behaviors.",
+        "bbox": [
+            112,
+            84,
+            492,
+            275
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose INREACT, a structured training framework for GUI agents that combines curriculum learning with reinforcement learning. The curriculum learning stage is designed to systematically build core agent capabilities—including perception, grounding, and reasoning through a staged learning process of increasing complexity. Furthermore, we introduce a reinforcement learning strategy with rule-based reward functions that jointly considers action-type correctness and distance-aware spatial accuracy, enabling more precise and reliable interactions in complex GUI environments.",
+        "bbox": [
+            115,
+            278,
+            490,
+            485
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are summarized as follows:",
+        "bbox": [
+            131,
+            488,
+            472,
+            502
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose a novel multimodal GUI agent framework, called INREACT, which synergizes observing, thinking, and acting to make more interpretable and goal-directed decisions when handling complex GUI navigation tasks.",
+            "- INREACT is trained through a two-stage training framework, including a curriculum-based supervised fine-tuning stage to progressively acquire perception, grounding, and reasoning capabilities and a reinforcement learning stage that reinforces grounding precision with an outcome-based reward to encourage accurate and consistent action predictions.",
+            "- Experimental results on four datasets in grounding and navigation tasks show that INREACT outperforms existing methods and exhibits promising generalization. Our code will be released via https://github.com/wangyuanlei30/InReAct."
+        ],
+        "bbox": [
+            114,
+            505,
+            489,
+            809
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            824,
+            268,
+            838
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 GUI Agents",
+        "text_level": 1,
+        "bbox": [
+            114,
+            851,
+            253,
+            866
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "GUI agents allow models to interact with software interfaces via visual inputs and simulated actions. Early works like CogAgent (Hong et al., 2024)",
+        "bbox": [
+            112,
+            873,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "introduced MLLMs for GUI tasks. Some methods leverage general-purpose models with prompt engineering (e.g., AppAgent (Zhang et al., 2025), Mobile-Agent (Wang et al., 2024b)), while others fine-tune open-source models on GUI-specific data (e.g., UGround (Gou et al., 2025), ShowUI (Lin et al., 2024)). Recent work like UI-TARS (Qin et al., 2025) and AGUVIS (Xu et al., 2024) combines GUI pretraining with reasoning fine-tuning. Most rely on single-stage supervised fine-tuning, which requires large-scale data and often struggles with generalization, highlighting the need for more structured learning approaches.",
+        "bbox": [
+            507,
+            84,
+            884,
+            294
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Curriculum Learning",
+        "text_level": 1,
+        "bbox": [
+            509,
+            310,
+            726,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Curriculum learning (Wang et al., 2022) is a training strategy where models are exposed to tasks in increasing order of difficulty, mimicking human learning. It has shown strong effectiveness across NLP and vision tasks by improving learning stability and generalization. Recent works have applied this idea to large-scale models: AutoGLM (Liu et al., 2024) and AutoWebGLM (Lai et al., 2024), for example, design task curricula to enhance model robustness in web navigation tasks. In the context of instruction tuning and agent learning, curriculum learning helps models acquire basic capabilities before handling complex reasoning or planning, offering a structured alternative to flat fine-tuning.",
+        "bbox": [
+            507,
+            334,
+            884,
+            575
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 Rule-Based Reinforcement Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            590,
+            843,
+            606
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Rule-based reinforcement learning has emerged as a scalable alternative to traditional supervised fine-tuning by optimizing models with predefined and verifiable reward functions. Early works such as DeepSeek-R1 (Guo et al., 2025) and o1 (Jaech et al., 2024) applied RFT to language tasks, including mathematical reasoning (Shao et al., 2024) and code generation, where correctness can be evaluated with rule-based signals. More recent studies (Shen et al., 2025; Liu et al., 2025c; Chen et al., 2025; Wang et al., 2024a) extend this paradigm to multimodal models by designing task-specific rewards—such as Intersection-over-Union (IoU) for image grounding and symbolic consistency checks for classification. Building on this line of research, works like UI-R1 (Lu et al., 2025) and GUI-R1 (Luo et al., 2025) have further explored applying rule-based RL to GUI agents, typically leveraging binary hit-or-miss rewards for GUI grounding.",
+        "bbox": [
+            507,
+            615,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "9149",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/649ab4ab11431d5efe76d29ad7683f111563fdf6fbaf79660e2fa5a977f48fcd.jpg",
+        "image_caption": [
+            "Figure 2: The overall framework of INREACT"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            84,
+            877,
+            374
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            407,
+            216,
+            422
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a GUI screenshot  $S$ , a user instruction  $G$  and the previous interaction history  $A_{<t}$  at each time step  $t$ , the goal of the agent is to generate a sequence of executable actions  $A = \\{a_1, a_2, \\ldots, a_T\\}$ , where action  $a_t$  at time step  $t$  is determined based on  $G$ , observation of the current screenshot environment  $O_t$ , and previous actions sequence  $A_{<t} = [a_1, a_2, \\ldots, a_{t-1}]$ .",
+        "bbox": [
+            112,
+            432,
+            487,
+            563
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- State Space. At each time step  $t$ , the environment state is represented by a GUI screenshot  $s_t$ , which captures the current visual interface after executing the previous action. The agent's observation  $O_t$  includes the screenshot  $S_t$ , the user instruction  $G$ , and the action history  $A_{<t} = \\{a_1, \\dots, a_{t-1}\\}$ . This combination provides both static semantic goals and dynamic context necessary for decision-making.",
+            "- Action Space. We define an action as a structured function call in the form  $a_{t} = (\\text{type}_{t}, \\text{args}_{t})$ , where type specifies the operation (e.g., click, scroll, input) and args provides necessary arguments such as coordinates, text, or direction. The environment executes actions parsed from the agent's JSON-style outputs. A full specification of supported actions is provided in Table 1."
+        ],
+        "bbox": [
+            112,
+            564,
+            489,
+            852
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 GUI Agent Workflow",
+        "text_level": 1,
+        "bbox": [
+            112,
+            866,
+            327,
+            882
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To guide the agent to continually interact with the environment and complete multi-step complex",
+        "bbox": [
+            112,
+            889,
+            487,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/08326578102b19b90cd99065fc06de472aa85c0c2eb090e9ffdbb33b2fb9a35c.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Action Type</td><td>Parameters</td></tr><tr><td>click,tap,hover,select</td><td>(x,y)</td></tr><tr><td>swipe,select_text</td><td>(x1,y1,x2,y2)</td></tr><tr><td>scroll</td><td>direction</td></tr><tr><td>input</td><td>text</td></tr><tr><td>home,back,enter</td><td>None</td></tr></table>",
+        "bbox": [
+            542,
+            404,
+            850,
+            519
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Action types and their parameters",
+        "bbox": [
+            549,
+            527,
+            840,
+            543
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "tasks. Following ReAct (Yao et al., 2023), we construct our agent framework through synergizing observing, thinking and acting in GUI controlling process for agent.",
+        "bbox": [
+            507,
+            558,
+            882,
+            624
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Observation Phase. Based on a visual screenshot  $S_{t}$  of the GUI, an user instruction  $G$  and the previous interaction history  $A_{< t}$  at each time step  $t$ , the agent first generates a natural language observation  $d_{t}$  describing key interface elements and their relevance to the task instead of directly generating action by:",
+        "bbox": [
+            507,
+            634,
+            884,
+            747
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nd _ {t} = \\mathcal {F} \\left(O _ {t}\\right) = \\mathcal {F} \\left(S _ {t}, G, A _ {<   t}\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            581,
+            760,
+            882,
+            778
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Thinking Phase. In the thinking phase, the generated observation  $d_{t}$  is appended to the original input  $O_{t} = \\{S_{t}, G, A_{<t}\\}$  to construct the further input. Then, we feed the updated input to the agent for generating internal reasoning to produce a task-oriented thinking  $h_{t}$  by:",
+        "bbox": [
+            507,
+            793,
+            884,
+            890
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nh _ {t} = \\mathcal {F} \\left(O _ {t}, d _ {t}\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            635,
+            904,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "9150",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Acting Phase. Similarly, the input in the final action phase is constructed by concatenating the generated thinking  $h_t$  with the previous input. Subsequently, we feed the modified input to the agent to generate executable actions by",
+        "bbox": [
+            112,
+            84,
+            489,
+            165
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\na _ {t} = \\mathcal {F} \\left(O _ {t}, d _ {t}, h _ {t}\\right) = \\left(\\text {t y p e} _ {t}, \\arg s _ {t}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            166,
+            177,
+            487,
+            193
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The action is then executed in the environment, resulting in a transition to the next state  $S_{t + 1}$ .",
+        "bbox": [
+            112,
+            205,
+            487,
+            237
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Inspire-Then-Reinforce Training",
+        "text_level": 1,
+        "bbox": [
+            112,
+            247,
+            418,
+            263
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To progressively enhance the reasoning and control capabilities of GUI agents, we propose an inspire-then-reinforce two-stage training paradigm, as illustrated in Figure 2.",
+        "bbox": [
+            112,
+            269,
+            489,
+            332
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Stage 1, we introduce a curriculum learning strategy that progressively builds the model's perception, grounding, and reasoning capabilities through supervised fine-tuning. By organizing training tasks from simple to complex, the model first learns to understand GUI layouts, then localize targets, and ultimately reason over multi-step goals.",
+        "bbox": [
+            112,
+            334,
+            487,
+            461
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Stage 2, we perform reinforcement learning to improve action accuracy and grounding robustness. Using a GRPO-based optimization framework, we incorporate a task-specific reward function that evaluates both action-type correctness and spatial precision, enabling the model to generate reliable actions in dense and high-resolution GUI environments.",
+        "bbox": [
+            112,
+            462,
+            487,
+            589
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Stage 1: Curriculum Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            599,
+            413,
+            615
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Drawing inspiration from human learning, we organize agent training into three progressive stages: perception, grounding, and reasoning. Each stage targets a specific skill and increases in task complexity, forming a curriculum learning framework that guides the model from basic comprehension to advanced planning.",
+        "bbox": [
+            112,
+            619,
+            487,
+            730
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we adopt supervised fine-tuning and have the same training objective that minimizes the sum of negative log-likelihood loss averaged over tokens in each step:",
+        "bbox": [
+            112,
+            731,
+            487,
+            795
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} (y, \\widehat {y}) = - \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\widehat {y _ {l}} \\log \\left(\\frac {\\exp \\left(y _ {l}\\right)}{\\sum_ {i} ^ {l} \\exp \\left(y _ {i}\\right)}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            819,
+            487,
+            863
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $L$  is the max length of output sequence,  $\\widehat{y}_l$  and  $y_{l}$  denote the  $l$ -th token in the groundtruth sequence  $\\widehat{y}$  and generation sequence  $y$ , respectively.",
+        "bbox": [
+            112,
+            872,
+            487,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Learning for Perception The first stage of the curriculum aims to build the model's visual and semantic understanding of GUI layouts. This foundational perception equips the agent with spatial awareness and interface comprehension, essential for later grounding and reasoning. Specifically, we decompose it into two perception-oriented tasks:",
+        "bbox": [
+            507,
+            84,
+            882,
+            196
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "i. GUI-based Question Answering. Given a screenshot  $s$  and question  $q$ , the model generates an answer  $a\\colon \\mathcal{M}_{\\mathrm{qa}}(s,q)\\to a$ , enhancing spatial attention and element recognition.",
+            "ii. Interface Description Generation. Given  $s$ , the model produces a structured description  $d$ :  $\\mathcal{M}_{\\mathrm{desc}}(s) \\to d$ , fostering global layout and semantic understanding. The training loss for this stage is computed as:"
+        ],
+        "bbox": [
+            507,
+            197,
+            882,
+            341
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {p e r c e p t i o n}} = \\mathcal {L} (a, \\widehat {a}) + \\mathcal {L} (d, \\widehat {d}), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            347,
+            882,
+            366
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widehat{s}$  and  $\\widehat{d}$  denote the ground-truth answers and interface descriptions, respectively.",
+        "bbox": [
+            507,
+            373,
+            880,
+            406
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Learning for Grounding The second stage of the curriculum focuses on grounding user instructions to visual actions, a core capability for operating in GUI environments. The model must understand what the user wants and accurately identify the corresponding interface region for execution. We similarly decompose it into two complementary grounding tasks:",
+        "bbox": [
+            507,
+            414,
+            882,
+            542
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "i. Instruction-to-Coordinate. Given a GUI screenshot  $s$  and instruction  $g$ , the model predicts the target coordinate  $c$ :  $\\mathcal{M}_{\\mathrm{inst2coord}}(s, g) \\to c$ . This task aligns user intent with precise UI locations.",
+            "ii. Coordinate-to-Description. Given  $s$  and a coordinate  $c$ , the model generates a description  $d$  of the corresponding UI element:  $\\mathcal{M}_{\\mathrm{coord2desc}}(s,c) \\to d$ , reinforcing spatial-semantic understanding. The loss function for this stage is defined as:"
+        ],
+        "bbox": [
+            507,
+            543,
+            882,
+            702
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {g r o u n d i n g}} = \\mathcal {L} (c, \\widehat {c}) + \\mathcal {L} (d, \\widehat {d}), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            579,
+            709,
+            882,
+            728
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widehat{c}$  and  $\\widehat{d}$  denote the ground-truth coordinate and description, respectively.",
+        "bbox": [
+            507,
+            734,
+            880,
+            768
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Learning for Reasoning The third stage focuses on high-level reasoning and multi-step decision-making, where the model explicitly follows an observe-think-act process instead of making one-step predictions. We introduce the following two components to support structured reasoning:",
+        "bbox": [
+            507,
+            776,
+            882,
+            872
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "i. Observation Chain. Given a screenshot  $x$  and instruction  $g$ , the model generates an observation  $o$  to identify relevant UI regions:  $\\mathcal{M}_{\\mathrm{obs}}(x, g) \\to o$ .",
+        "bbox": [
+            507,
+            873,
+            880,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "9151",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "ii. Reasoning Chain. Conditioned on  $o$  and  $g$ , the model produces a reasoning trace  $r$  to guide decision-making:  $\\mathcal{M}_{\\mathrm{reason}}(o,g)\\to r$ .",
+        "bbox": [
+            112,
+            84,
+            485,
+            131
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We define the training loss of this stage as follows:",
+        "bbox": [
+            112,
+            133,
+            485,
+            162
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {r e a s o n i n g}} = \\mathcal {L} (o, \\widehat {o}) + \\mathcal {L} (r, \\widehat {r}), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            180,
+            487,
+            198
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\widehat{o}$  and  $\\widehat{r}$  denote the ground-truth observation and reasoning trace.",
+        "bbox": [
+            112,
+            212,
+            485,
+            244
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Stage 2: Reinforcement Learning",
+        "text_level": 1,
+        "bbox": [
+            114,
+            256,
+            435,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Although supervised fine-tuning provides a strong initialization, models often suffer from fine-grained prediction errors, especially in dense or high-resolution GUIs where small pixel-level deviations can lead to task failure. To address this, we adopt GRPO Reinforcement fine-tuning.",
+        "bbox": [
+            112,
+            275,
+            485,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "A critical component of our approach is the design of the reward function. While prior works like UI-R1 (Lu et al., 2025) and GUI-R1 (Luo et al., 2025) also leverage rule-based reinforcement learning for GUI agents, their reward mechanisms for localization typically rely on a coarse, binary signal (i.e., whether the predicted point falls within the target's bounding box). Such a reward, while useful for confirming task success, provides limited gradient for improving fine-grained precision. To overcome this, we introduce a task-specific reward function that better captures the dual objectives of action correctness and spatial precision, directly encouraging the model to enhance its pixel-level accuracy.",
+        "bbox": [
+            112,
+            374,
+            487,
+            615
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Reward Modeling The reward function serves as the primary training signal in reinforcement learning, directly guiding the model's optimization trajectory. To support fine-grained policy improvement in GUI environments, we design a rule-based reward function tailored to the unique demands of screen-based interaction. Unlike traditional RL settings that rely on binary task success or coarse region-level feedback, GUI tasks require both accurate action selection and high-precision localization of target elements. Specifically, our reward function comprises two components: (1) an action-type reward(2) a distance-aware coordinate reward. This composite design provides the model with dense, informative feedback signals essential for precise and interpretable GUI control, which is defined as:",
+        "bbox": [
+            112,
+            626,
+            489,
+            883
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR = R _ {\\mathrm {T}} + R _ {\\mathrm {D}} \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            243,
+            903,
+            485,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Regarding the action type reward component  $R_{\\mathrm{T}}$ , it provides a binary reward signal that evaluates whether the predicted action type  $T$  aligns with the ground truth label  $T'$ . This mechanism offers a direct and effective evaluation framework for action-type prediction tasks. The mathematical formulation of this component is presented below:",
+        "bbox": [
+            507,
+            84,
+            882,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {\\mathrm {T}} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} T ^ {\\prime} = T \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            608,
+            203,
+            882,
+            244
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Additionally, the distance-aware reward  $R_{\\mathrm{D}}$  measures spatial accuracy by combining two components: (1) a binary reward for whether the predicted point lies within the target bounding box, and (2) a precision reward based on the normalized distance to the box center. This encourages both correct region prediction and fine-grained localization:",
+        "bbox": [
+            507,
+            250,
+            882,
+            362
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {D} = R _ {a c c} + R _ {d i s t} \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            620,
+            379,
+            880,
+            394
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "i. Accuracy Reward. We assign a base reward calculating whether the predicted point  $(x', y')$  falls inside the ground-truth bounding box  $B = [(x_1, y_1), (x_2, y_2)]$ :",
+        "bbox": [
+            507,
+            400,
+            880,
+            464
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {\\mathrm {a c c}} = \\left\\{ \\begin{array}{l l} 0. 5, & \\text {i f} \\left(x ^ {\\prime}, y ^ {\\prime}\\right) \\in B \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            579,
+            472,
+            880,
+            512
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ii. Precision Reward. If the point is within the box, we compute the normalized Euclidean distance  $\\hat{d}$  from the predicted coordinate  $(x', y')$  to the center of the target box  $(x_c, y_c)$ :",
+        "bbox": [
+            507,
+            518,
+            880,
+            583
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left(x _ {c}, y _ {c}\\right) = \\left(\\frac {x _ {1} + x _ {2}}{2}, \\frac {y _ {1} + y _ {2}}{2}\\right) \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            573,
+            589,
+            880,
+            623
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {d} = \\frac {\\sqrt {\\left(x ^ {\\prime} - x _ {c}\\right) ^ {2} + \\left(y ^ {\\prime} - y _ {c}\\right) ^ {2}}}{\\sqrt {\\left(x _ {2} - x _ {1}\\right) \\left(y _ {2} - y _ {1}\\right)}} \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            576,
+            630,
+            880,
+            668
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We then define the precision reward based on discrete distance intervals:",
+        "bbox": [
+            507,
+            671,
+            880,
+            701
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nR _ {\\text {d i s t}} = \\left\\{ \\begin{array}{l l} 0. 5, & \\hat {d} \\leq 0. 1 \\\\ 0. 3, & 0. 1 <   \\hat {d} \\leq 0. 3 \\\\ 0. 2, & 0. 3 <   \\hat {d} \\leq 0. 5 \\\\ 0. 0, & \\hat {d} > 0. 5 \\end{array} \\right. \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            578,
+            706,
+            880,
+            785
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To provide a clear and stepped gradient for precision improvement, we empirically determine the thresholds and reward values for our reward formulation. This design provides a continuous and informative learning signal that encourages the model to both select the correct action type and localize it precisely, promoting stable convergence and robust behavior in dense GUI layouts.",
+        "bbox": [
+            507,
+            793,
+            882,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "9152",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/51d1ebee0553bce024b6f976be05c36d23c5c27303fe55dd2ed2a13d371b0f40.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Development</td><td colspan=\"3\">Creative</td><td colspan=\"3\">CAD</td><td colspan=\"3\">Scientific</td><td colspan=\"3\">Office</td><td colspan=\"3\">OS</td><td colspan=\"3\">Overall</td></tr><tr><td colspan=\"3\">Text Icon Avg</td><td colspan=\"3\">Text Icon Avg</td><td colspan=\"3\">Text Icon Avg</td><td colspan=\"3\">Text Icon Avg</td><td colspan=\"3\">Text Icon Avg</td><td colspan=\"3\">Text Icon Avg</td><td colspan=\"3\">Text Icon Avg</td></tr><tr><td>GPT-4o (OpenAI et al., 2024)</td><td>1.3</td><td>0.0</td><td>0.7</td><td>1.0</td><td>0.0</td><td>0.6</td><td>2.0</td><td>0.0</td><td>1.5</td><td>2.1</td><td>0.0</td><td>1.2</td><td>1.1</td><td>0.0</td><td>0.9</td><td>0.0</td><td>0.0</td><td>0.0</td><td>1.3</td><td>0.0</td><td>0.8</td></tr><tr><td>ShowUI-2B (Lin et al., 2024)</td><td>16.9</td><td>1.4</td><td>9.4</td><td>9.1</td><td>0.0</td><td>5.3</td><td>2.5</td><td>0.0</td><td>1.9</td><td>13.2</td><td>7.3</td><td>10.6</td><td>15.3</td><td>7.5</td><td>13.5</td><td>10.3</td><td>2.2</td><td>6.6</td><td>10.8</td><td>2.6</td><td>7.7</td></tr><tr><td>AriaUI-3.9B (Yang et al., 2024)</td><td>16.2</td><td>0.0</td><td>8.4</td><td>23.7</td><td>2.1</td><td>14.7</td><td>7.6</td><td>1.6</td><td>6.1</td><td>27.1</td><td>6.4</td><td>18.1</td><td>20.3</td><td>1.9</td><td>16.1</td><td>4.7</td><td>0.0</td><td>2.6</td><td>17.1</td><td>2.0</td><td>11.3</td></tr><tr><td>OSAtlas-4B (Wu et al., 2024)</td><td>7.1</td><td>0.0</td><td>3.7</td><td>3.0</td><td>1.4</td><td>2.3</td><td>2.0</td><td>0.0</td><td>1.5</td><td>9.0</td><td>5.5</td><td>7.5</td><td>5.1</td><td>3.8</td><td>4.8</td><td>5.6</td><td>0.0</td><td>3.1</td><td>5.0</td><td>1.7</td><td>3.7</td></tr><tr><td>QwenVL-7B (Bai et al., 2023)</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.7</td><td>0.0</td><td>0.4</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.1</td><td>0.0</td><td>0.1</td></tr><tr><td>Qwen2-VL-7B (Wang et al., 2024c)</td><td>2.6</td><td>0.0</td><td>1.3</td><td>1.5</td><td>0.0</td><td>0.9</td><td>0.5</td><td>0.0</td><td>0.4</td><td>6.3</td><td>0.0</td><td>3.5</td><td>3.4</td><td>1.9</td><td>3.0</td><td>0.9</td><td>0.0</td><td>0.5</td><td>2.5</td><td>0.2</td><td>1.6</td></tr><tr><td>SeeClick-7B (Cheng et al., 2024)</td><td>0.6</td><td>0.0</td><td>0.3</td><td>1.0</td><td>0.0</td><td>0.6</td><td>2.5</td><td>0.0</td><td>1.9</td><td>3.5</td><td>0.0</td><td>2.0</td><td>1.1</td><td>0.0</td><td>0.9</td><td>2.8</td><td>0.0</td><td>1.5</td><td>1.8</td><td>0.0</td><td>1.1</td></tr><tr><td>CogAgent-18B (Hong et al., 2024)</td><td>14.9</td><td>0.7</td><td>8.0</td><td>9.6</td><td>0.0</td><td>5.6</td><td>7.1</td><td>3.1</td><td>6.1</td><td>22.2</td><td>1.8</td><td>13.4</td><td>13.0</td><td>0.0</td><td>10.0</td><td>5.6</td><td>0.0</td><td>3.1</td><td>12.0</td><td>0.8</td><td>7.7</td></tr><tr><td>UGround-7B (Gou et al., 2025)</td><td>26.6</td><td>2.1</td><td>14.7</td><td>27.3</td><td>2.8</td><td>17.0</td><td>14.2</td><td>1.6</td><td>11.1</td><td>31.9</td><td>2.7</td><td>19.3</td><td>31.6</td><td>11.3</td><td>27.0</td><td>17.8</td><td>0.0</td><td>9.7</td><td>25.0</td><td>2.8</td><td>16.5</td></tr><tr><td>OSAtlas-7B (Wu et al., 2024)</td><td>33.1</td><td>1.4</td><td>17.7</td><td>28.8</td><td>2.8</td><td>17.9</td><td>12.2</td><td>4.7</td><td>10.3</td><td>37.5</td><td>7.3</td><td>24.4</td><td>33.9</td><td>5.7</td><td>27.4</td><td>27.1</td><td>4.5</td><td>16.8</td><td>28.1</td><td>4.0</td><td>18.9</td></tr><tr><td>InReAct</td><td>35.7</td><td>1.4</td><td>19.0</td><td>33.3</td><td>2.1</td><td>20.2</td><td>13.2</td><td>1.6</td><td>10.7</td><td>41.7</td><td>10.0</td><td>27.9</td><td>42.4</td><td>3.8</td><td>33.5</td><td>23.4</td><td>1.1</td><td>17.3</td><td>31.4</td><td>3.3</td><td>20.7</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            880,
+            255
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.3 Inference Procedure",
+        "text_level": 1,
+        "bbox": [
+            112,
+            305,
+            319,
+            319
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "During inference, our agent adopts an interlaced observe-think-action workflow that sequentially executes tasks by observing the interface, thinking through the instruction, and predicting the next action. Given the input  $O_{t} = (S,G,H)$ , where  $S$  is the GUI screenshot,  $G$  is the task instruction, and  $A_{<t} = \\{a_{1},\\dots,a_{t-1}\\}$  denotes the history of executed actions, the model first generates an observation summary  $\\hat{o}_{t}$ , then infers a reasoning trace  $\\hat{r}_{t}$  conditioned on  $\\hat{o}_{t}$  and  $G$ , and finally predicts the next action  $a_{t} = (T',x',y')$  based on both. This process is repeated until task completion or a maximum number of steps is reached, enabling goal-driven, interpretable decision-making in complex GUI environments.",
+        "bbox": [
+            112,
+            326,
+            489,
+            567
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            581,
+            253,
+            598
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experiment Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            609,
+            302,
+            625
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Datasets We evaluate our model on two tasks: GUI grounding and navigation.",
+        "bbox": [
+            112,
+            631,
+            487,
+            662
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For Grounding, we use ScreenSpot (Cheng et al., 2024) and ScreenSpot-Pro (Li et al., 2025b). ScreenSpot contains 1,272 samples across mobile, desktop, and web platforms, focusing on common interface elements. ScreenSpot-Pro includes 23 professional applications with high-resolution and complex layouts, offering a more challenging evaluation.",
+        "bbox": [
+            112,
+            663,
+            487,
+            790
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For Navigation, we use: (i) Mind2Web (Deng et al., 2023), a web-based dataset for evaluating generalist agents that perform complex tasks via language instructions. It includes 7,775 training actions and three test splits (task, website, domain), with aligned HTML and screenshots. (ii) Android-Control (Li et al., 2024), a mobile dataset testing multi-step task execution in realistic Android en",
+        "bbox": [
+            112,
+            791,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/4989979ecd0c2587e79ba130a675a38741c0eaa88aa2c2268c61cb5a7d70f149.jpg",
+        "table_caption": [
+            "Table 2: GUI grounding results on ScreenSpot-Pro dataset."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Mobile</td><td colspan=\"2\">Desktop</td><td colspan=\"2\">Web</td><td rowspan=\"2\">Average</td></tr><tr><td>Text</td><td>Icon</td><td>Text</td><td>Icon</td><td>Text</td><td>Icon</td></tr><tr><td>GPT-4 (OpenAI, 2023)</td><td>22.6</td><td>24.5</td><td>20.2</td><td>11.8</td><td>9.2</td><td>8.8</td><td>16.7</td></tr><tr><td>GPT-4o (OpenAI et al., 2024)</td><td>20.2</td><td>24.9</td><td>21.1</td><td>23.6</td><td>12.2</td><td>7.8</td><td>18.1</td></tr><tr><td>Gemini-1.5-pro (Team, 2024)</td><td>76.2</td><td>54.1</td><td>65.5</td><td>39.2</td><td>52.2</td><td>32.0</td><td>53.2</td></tr><tr><td>Qwen2-VL-2B (Wang et al., 2024c)</td><td>24.2</td><td>10.0</td><td>1.4</td><td>9.3</td><td>8.7</td><td>2.41</td><td>9.3</td></tr><tr><td>Qwen2-VL-7B (Wang et al., 2024c)</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>42.9</td></tr><tr><td>CogAgent-18B (Hong et al., 2024)</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-7B (Cheng et al., 2024)</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>53.4</td></tr><tr><td>UGround-7B (Gou et al., 2025)</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>73.3</td></tr><tr><td>ShowUI-2B (Lin et al., 2024)</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>75.1</td></tr><tr><td>InfiGUIAgent-2B (Liu et al., 2025b)</td><td>88.6</td><td>74.7</td><td>85.6</td><td>65.0</td><td>79.1</td><td>64.6</td><td>76.3</td></tr><tr><td>InReAct</td><td>86.1</td><td>71.6</td><td>84.5</td><td>65.7</td><td>82.2</td><td>70.4</td><td>77.8</td></tr></table>",
+        "bbox": [
+            510,
+            302,
+            882,
+            430
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3: GUI grounding results on ScreenSpot dataset.",
+        "bbox": [
+            507,
+            439,
+            880,
+            455
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "vironments, emphasizing long-term planning and state tracking.",
+        "bbox": [
+            507,
+            480,
+            880,
+            513
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "See datasets details in appendix A.",
+        "bbox": [
+            527,
+            513,
+            786,
+            527
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics We use the metrics proposed in the corresponding datasets as shown in Table 7. Although they have different names, these metrics are similar: GUI grounding tasks consistently measure the hit rate of predicted bounding boxes, while GUI navigation tasks focus on single-step accuracy.",
+        "bbox": [
+            507,
+            539,
+            882,
+            634
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Click Accuracy. The proportion of test samples where the predicted location falls in the ground truth element's bounding box.",
+        "bbox": [
+            507,
+            636,
+            880,
+            684
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Element Accuracy (Ele.Acc). Comparing the selected element with all acceptable elements. For vision-based methods, it is the same as Click Accuracy.",
+        "bbox": [
+            507,
+            684,
+            880,
+            747
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Operation F1 (Op.F1). Token-level F1 score for the predicted operation.",
+        "bbox": [
+            507,
+            750,
+            880,
+            781
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Step Success Rate (Step SR) and Step Accuracy. The proportion of successful steps. A step is regarded as successful only if both the selected element and the predicted operation are correct.",
+        "bbox": [
+            507,
+            782,
+            882,
+            846
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines We compare the performance of our model with various baselines, including some closed source large models such as GPT-4o (OpenAI et al., 2024), Gemini-1.5-pro (Team, 2024)",
+        "bbox": [
+            507,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "9153",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/6a397039a40e3fab06647b4139ce39be79af4db15ebadb80cf71332bb1150e87.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"3\">Cross-Task</td><td colspan=\"3\">Cross-Website</td><td colspan=\"3\">Cross-Domain</td></tr><tr><td>Ele.Acc</td><td>Op.F1</td><td>Step.SR</td><td>Ele.Acc</td><td>Op.F1</td><td>Step.SR</td><td>Ele.Acc</td><td>Op.F1</td><td>Step.SR</td></tr><tr><td>GPT-4 (OpenAI, 2023)</td><td>41.6</td><td>60.6</td><td>36.2</td><td>35.8</td><td>51.1</td><td>30.1</td><td>37.1</td><td>46.5</td><td>26.4</td></tr><tr><td>OmniParser (Lu et al., 2024)</td><td>42.4</td><td>87.6</td><td>39.4</td><td>41.0</td><td>84.8</td><td>36.5</td><td>45.5</td><td>85.7</td><td>42.0</td></tr><tr><td>Qwen2-VL-2B (Wang et al., 2024c)</td><td>37.7</td><td>86.4</td><td>33.2</td><td>36.0</td><td>79.2</td><td>27.6</td><td>36.3</td><td>81.8</td><td>30.7</td></tr><tr><td>ShowUI-2B (Lin et al., 2024)</td><td>39.9</td><td>88.6</td><td>37.2</td><td>41.6</td><td>83.5</td><td>35.1</td><td>39.4</td><td>86.8</td><td>35.2</td></tr><tr><td>SpiritSight-2B (Huang et al., 2025)</td><td>51.7</td><td>87.2</td><td>44.9</td><td>44.9</td><td>83.6</td><td>37.8</td><td>42.4</td><td>83.5</td><td>36.9</td></tr><tr><td>SeeClick-7B (Cheng et al., 2024)</td><td>28.3</td><td>87.0</td><td>25.5</td><td>21.4</td><td>80.6</td><td>16.4</td><td>23.2</td><td>84.8</td><td>20.8</td></tr><tr><td>CogAgent-18B (Hong et al., 2024)</td><td>22.4</td><td>53.0</td><td>17.6</td><td>18.4</td><td>42.4</td><td>13.4</td><td>20.6</td><td>42.0</td><td>15.5</td></tr><tr><td>InReAct</td><td>57.8</td><td>89.4</td><td>50.8</td><td>52.6</td><td>88.8</td><td>47.4</td><td>55.4</td><td>85.4</td><td>49.2</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            884,
+            233
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/664c9984e36d94c9becae36d527f75e4add00e781107186845594edca69f2cea.jpg",
+        "table_caption": [
+            "Table 4: GUI navigation results on Mind2web dataset."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Low-Level</td><td>High-Level</td></tr><tr><td>Claude* (Anthropic, 2024)</td><td>19.4</td><td>12.5</td></tr><tr><td>GPT-4o (OpenAI et al., 2024)</td><td>19.4</td><td>20.8</td></tr><tr><td>Aria-UI-3.9B (Yang et al., 2024)</td><td>67.3</td><td>10.2</td></tr><tr><td>OS-Atlas-4B (Wu et al., 2024)</td><td>80.6</td><td>67.5</td></tr><tr><td>Aguvis-7B (Xu et al., 2024)</td><td>80.5</td><td>61.5</td></tr><tr><td>InReAct</td><td>83.1</td><td>68.1</td></tr></table>",
+        "bbox": [
+            119,
+            278,
+            485,
+            390
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5: GUI navigation results on AndroidControl dataset.",
+        "bbox": [
+            112,
+            399,
+            487,
+            429
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "and Claude(Anthropic, 2024), as well as numerous open-source GUI agents such as SpiritSight-2B (Huang et al., 2025), AGUVIS-7B (Xu et al., 2024), Uground (Gou et al., 2025), Aria-UI(Yang et al., 2024), OSAtlas (Wu et al., 2024), ShowUI (Lin et al., 2024), InfGUIAgent (Liu et al., 2025b), UGround (Gou et al., 2025), CogAgent (Hong et al., 2024), SeeClick (Cheng et al., 2024), Qwen2-VL-2B, Qwen2-VL-7B (Wang et al., 2024c). Implementation details refer to appendix B.",
+        "bbox": [
+            112,
+            454,
+            490,
+            615
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 Overall Performance",
+        "text_level": 1,
+        "bbox": [
+            112,
+            626,
+            326,
+            640
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "GUI Grounding Evaluation The evaluation results on the ScreenSpot (Cheng et al., 2024) are presented in Table 3. Our model achieves an average accuracy of  $77.8\\%$ , outperforming all competing baselines, including both open-source and commercial GUI agents. In particular, Our model surpasses strong 2B-level models such as ShowUI  $(75.1\\%)$  and InfigUIAgent  $(76.3\\%)$ , and even outperforms several significantly larger models such as CogAgent  $(47.4\\%, 18\\mathrm{B})$  and UGround  $(73.3\\%, 7\\mathrm{B})$ .",
+        "bbox": [
+            112,
+            646,
+            490,
+            822
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As in shown in table 2, on the more challenging ScreenSpot-Pro (Li et al., 2025b) dataset, our method achieves an overall accuracy of  $20.7\\%$ , representing a substantial improvement under complex, high-resolution desktop GUI environments. This dataset covers a wide spectrum of profes",
+        "bbox": [
+            112,
+            825,
+            490,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "sional software, including CAD tools, scientific platforms, development IDEs, and creative applications, each demanding fine-grained grounding of visually dense and domain-specific interfaces. Despite the increased difficulty, our model consistently outperforms previous strong baselines. It exceeds ShowUI by  $13.0\\%$  percentage points and surpasses AriaUI by  $9.4\\%$  points, demonstrating particularly strong capabilities in GUI grounding",
+        "bbox": [
+            507,
+            281,
+            884,
+            428
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "GUI Navigation Evaluation On Mind2Web (Deng et al., 2023), the task requires agents to perform realistic web navigation across diverse websites. As shown in Table 4, InReAct achieves strong performance across all three generalization splits—cross-task, cross-website, and cross-domain—outperforming previous methods in terms of element grounding accuracy, operation-level F1, and step-wise success rate. This indicates that our model not only understands the semantics of web elements but also makes consistent and interpretable action decisions across varying environments. The consistent gains across generalization splits demonstrate InReAct's robustness in unseen task goals and layouts.",
+        "bbox": [
+            507,
+            462,
+            884,
+            703
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "On AndroidControl (Li et al., 2024), we evaluate the model under two settings: Low-Level, which focuses on simple, mostly single-step interactions (e.g., tapping a button); and High-Level, which involves complex, multi-step instructions with long-horizon dependencies. As shown in Table 5, InReAct achieves 83.1 on the Low-Level set and 68.1 on the High-Level set, outperforming all prior baselines. While strong results in Low-Level indicate precise grounding and action formulation, the improvements in High-Level highlight the model's ability to plan and reason over multi-step interaction sequences.",
+        "bbox": [
+            507,
+            711,
+            885,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "9154",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            278,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As summarized in Table 6, we conduct ablation studies to assess the impact of the proposed strategies in both curriculum learning (CL) and reinforcement learning (RL) training stages. For Mind2Web and AndroidControl, we only reported the results on cross-task SR and High-level. There are several notable observations as follows:",
+        "bbox": [
+            112,
+            107,
+            489,
+            218
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Curriculum Learning We examine the role of each step in our curriculum learning framework. Removing Step 1, where the model learns basic GUI interactions, leads to a significant performance drop, as the agent fails to acquire essential skills. Similarly, omitting Steps 2 and 3, which introduce grounding and reasoning tasks progressively, also harms performance, showing their importance for understanding and reasoning over GUI elements. Furthermore, replacing curriculum learning with multi-task training across all steps results in worse performance, demonstrating that staged, gradual training enhances learning and generalization more effectively than joint training.",
+        "bbox": [
+            115,
+            231,
+            489,
+            457
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Reinforcement Learning To further evaluate the effectiveness of the reinforcement learning (RL) stage, we first replace it with a standard supervised fine-tuning (SFT) setup. The results show that RL brings consistent performance gains across all datasets. On the two navigation datasets—Mind2Web and AndroidControl—RL yields improvements of  $+5.1$  and  $+4.3$ , respectively, demonstrating its ability to produce more accurate actions and more robust decision-making.",
+        "bbox": [
+            112,
+            468,
+            489,
+            629
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, to specifically justify our reward function design within the RL framework, we compare our proposed piecewise reward function against two alternatives: a binary (hit-or-miss) reward, which provides an accuracy signal without any dedicated precision component, similar to approaches in works like UI-R1 and GUI-R1, and a continuous linear reward (defined as  $R_{dist} = 1 - \\hat{d}$ , where  $\\hat{d}$  is the normalized distance to the target's center). This analysis, presented in Table 6, reveals that the design of the reward signal plays an important role. On ScreenSpot, our method (77.8) outperforms both the linear (77.0) and binary (76.8) baselines. On the more challenging ScreenSpot-Pro dataset, our approach (20.7) also provides a modest but consistent improvement over the linear (20.3) and binary (19.8) rewards.",
+        "bbox": [
+            112,
+            630,
+            489,
+            903
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The advantage of our piecewise reward function",
+        "bbox": [
+            131,
+            904,
+            487,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "lies in two aspects. First, it delivers clearer learning signals by introducing reward \"steps,\" which motivate the agent to cross precision thresholds rather than being limited by diminishing returns near the target. Second, it aligns more closely with real-world GUI logic, where clicks have discrete outcomes—successful or not—and central clicks are inherently safer than edge clicks. This design thus captures the practical structure of interaction more faithfully than a simple linear function. It is worth noting that our thresholds are chosen empirically and may not be optimal; further tuning is left for future work.",
+        "bbox": [
+            507,
+            84,
+            884,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of Observation and Reasoning Components The results, presented in Table 6, underscore the importance of both components within our observe-think-act cycle. Removing the thinking phase (w/o Thought) leads to a consistent and significant performance drop across all four benchmarks, with the most notable decreases observed on the complex ScreenSpot-Pro (-2.5 points) and AndroidControl (-2.3 points) datasets. This indicates that the explicit reasoning step is crucial for formulating effective plans and making robust decisions in challenging environments.",
+        "bbox": [
+            507,
+            305,
+            884,
+            495
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Similarly, removing the observation phase (w/o Observation) also consistently degrades performance across all tested datasets, such as causing a 1.8-point drop on ScreenSpot. An interesting finding is that the performance degradation from removing the thinking phase is more pronounced than that from removing the observation phase on the three more complex datasets. This suggests that while both components are beneficial, the structured reasoning provided by the \"Thinking\" step is arguably the most critical factor for the agent's success in complex scenarios. These findings strongly justify the synergistic design of our InReAct framework.",
+        "bbox": [
+            507,
+            499,
+            884,
+            722
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4 Case Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            737,
+            643,
+            753
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To qualitatively assess the effectiveness of our proposed curriculum learning and reinforcement learning strategies, we conduct a case study on the ScreenSpot dataset. As illustrated in Figure 3, we present two pairs of examples highlighting the contributions of each training stage.",
+        "bbox": [
+            507,
+            759,
+            884,
+            854
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In Figure 3a, without curriculum learning, the agent clicks on the text \"Only draw with Apple Pencil\" instead of the adjacent switch, showing weak perception and semantic grounding. In contrast,",
+        "bbox": [
+            507,
+            857,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "9155",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/fa04b9d2d3e4f10af929a2abff5f7432deb4ccc9c52bbc2d2e4ccfe0aa6cd735.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>ScreenSpot</td><td>ScreenSpot-Pro</td><td>Mind2Web</td><td>AndroidControl</td></tr><tr><td>InReAct</td><td>77.8</td><td>20.7</td><td>50.8</td><td>68.1</td></tr><tr><td colspan=\"5\">Stage 1: Curriculum Learning</td></tr><tr><td>- w/o Step 1</td><td>76.3</td><td>19.2</td><td>47.4</td><td>65.9</td></tr><tr><td>- w/o Step 2</td><td>68.6</td><td>10.8</td><td>37.8</td><td>57.4</td></tr><tr><td>- w/o Step 3</td><td>74.7</td><td>17.4</td><td>42.5</td><td>62.3</td></tr><tr><td>- w/o curriculum</td><td>75.4</td><td>19.5</td><td>48.5</td><td>66.4</td></tr><tr><td colspan=\"5\">Stage 2: Reinforcement Learning</td></tr><tr><td>- w/ Binary Reward</td><td>76.8</td><td>19.8</td><td>49.2</td><td>67.0</td></tr><tr><td>- w/ Linear Reward</td><td>77.0</td><td>20.3</td><td>50.2</td><td>67.3</td></tr><tr><td>- w/ SFT</td><td>76.2</td><td>16.8</td><td>45.7</td><td>63.8</td></tr><tr><td colspan=\"5\">Ablation on Agent Components</td></tr><tr><td>- w/o Observation</td><td>76.0</td><td>19.7</td><td>49.2</td><td>66.8</td></tr><tr><td>- w/o Thought</td><td>76.6</td><td>18.2</td><td>48.6</td><td>65.8</td></tr></table>",
+        "bbox": [
+            114,
+            80,
+            884,
+            346
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Ablation study of INREACT. We evaluate the impact of different training stages, reward functions, and core agent components. For Mind2Web, we report the Step SR on the cross-task split, while for AndroidControl, we report results on the High-Level setting.",
+        "bbox": [
+            112,
+            354,
+            882,
+            398
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/30242008ba5479c269d169dc7468980d1329845214b4c78763a0798043ee748e.jpg",
+        "image_caption": [
+            "(a)  $x$  Instruction:\"Only draw with apple pencil\". Without curriculum learning"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            413,
+            287,
+            504
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/22dd080ce9d2cf4d857ca9c2d851439be4f122d7d136e250706666981b471fa5.jpg",
+        "image_caption": [
+            "(b)  $\\checkmark$  Instruction:\"Only draw with apple pencil\". InReAct achieves accurate grounding"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            317,
+            413,
+            487,
+            504
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/ca2ace5599a4fd0f17d5279997bea001d349def88360af261eaf366502053830.jpg",
+        "image_caption": [
+            "(c)  $X$  Instruction:\"Zoom in\". Without reinforcement learning",
+            "Figure 3: Case study on ScreenSpot dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            562,
+            285,
+            665
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/69de77b5a31c00e1f912585b6529d2dd3c56ab4bda9e80489f6566c91236f082.jpg",
+        "image_caption": [
+            "(d)  $\\checkmark$  Instruction:\"Zoom in\". InReAct achieves accurate grounding"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            315,
+            562,
+            487,
+            665
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "InReAct (Figure 3b) correctly targets the switch, indicating better alignment between instructions and interface elements.",
+        "bbox": [
+            112,
+            758,
+            487,
+            804
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In Figure 3c, without reinforcement learning, for the instruction \"Zoom in,\" the predicted click point is near the correct bounding box but slightly outside it, indicating general understanding but lacking fine-grained precision. In contrast, as shown in Figure 3d, InReAct achieves significantly more accurate localization, clicking within the target re",
+        "bbox": [
+            112,
+            808,
+            489,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "gion. This confirms that our reinforcement learning stage enhances fine-grained control.",
+        "bbox": [
+            507,
+            416,
+            880,
+            448
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            458,
+            640,
+            474
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present INREACT, a multimodal GUI agent framework that integrates observation, reasoning, and action for complex GUI interactions. It is trained in two stages: curriculum-learning to build perception, grounding, and reasoning skills, followed by reinforcement learning with a task-specific reward that improves action correctness and spatial precision.",
+        "bbox": [
+            507,
+            485,
+            884,
+            613
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We evaluate INREACT across four datasets: ScreenSpot, ScreenSpot-Pro, Mind2Web, and AndroidControl, and demonstrate consistent improvements over prior methods. Our findings highlight the effectiveness of combining structured learning and outcome-driven optimization for enhancing GUI agent performance, setting a new direction for generalizable and interpretable vision-language agents.",
+        "bbox": [
+            507,
+            614,
+            882,
+            759
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "9156",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            278,
+            101
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This research was supported by Key-Area Research and Development Program of Guangdong Province, Granted No. 2024B1111060004.",
+        "bbox": [
+            112,
+            109,
+            489,
+            156
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            168,
+            220,
+            184
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Limited Action Space Coverage Our current action space, while sufficient for common GUI interactions such as click, type, and scroll, does not fully cover the diversity of real-world GUI operations. Advanced actions like drag-and-drop, multi-touch gestures, system-level controls, or context-menu navigation are not well supported in our framework. Extending the action space to include such capabilities would be essential for achieving broader applicability in real-world software environments.",
+        "bbox": [
+            112,
+            193,
+            490,
+            355
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Inference Speed and Efficiency Although In-ReAct achieves strong performance on various GUI tasks, its inference speed has room for optimization. The structured reasoning process—particularly the multi-step observation and thinking phases—introduces additional computational overhead. This may affect the agent's ability to perform real-time or near-real-time interactions, which is crucial for applications requiring fast response times. Future work could explore model compression, caching mechanisms, or streamlined reasoning pipelines to improve efficiency.",
+        "bbox": [
+            112,
+            363,
+            489,
+            557
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Limited Handling of Dynamic Interfaces Our experiments are primarily conducted on static interfaces. However, real-world GUI applications often involve dynamic content updates, such as pop-up notifications, auto Refreshing elements, or state transitions, that can disrupt perception and decision-making. Our current framework lacks explicit mechanisms to detect, adapt to, or reason about such changes over time. Incorporating temporal modeling or environment change detection would be a valuable direction for future research.",
+        "bbox": [
+            112,
+            565,
+            489,
+            741
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            768,
+            213,
+            783
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Anthropic. 2024. Developing a computer use model. https://www.anthropic.com/news/ developing-computer-use. Accessed: 2025-04- 12.",
+            "Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. CoRR, abs/2308.12966."
+        ],
+        "bbox": [
+            115,
+            791,
+            489,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Ming-Hsuan Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. CoRR, abs/2502.13923.",
+            "Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. 2025. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. https://github.com/Deep-Agent/R1-V. Accessed: 2025-02-02.",
+            "Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, Yuan Yao, Yankai Lin, Zhiyuan Liu, and Maosong Sun. 2024. Guicourse: From general vision language models to versatile GUI agents. CoRR, abs/2406.11317.",
+            "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. CoRR, abs/2312.14238.",
+            "Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. 2024. Seeclick: Harnessing GUI grounding for advanced visual GUI agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 9313-9332. Association for Computational Linguistics.",
+            "Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samual Stevens, Boshi Wang, Huan Sun, and Yu Su. 2023. Mind2web: Towards a generalist agent for the web. 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.",
+            "Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. 2025. Navigating the digital world as humans do: Universal visual grounding for GUI agents. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net.",
+            "Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.",
+            "Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, and Jie Tang. 2024. Co-gagent: A visual language model for GUI agents. In IEEE/CVF Conference on Computer Vision and"
+        ],
+        "bbox": [
+            509,
+            85,
+            885,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "9157",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 14281-14290. IEEE.",
+            "Yu-Chung Hsiao, Fedir Zubach, Maria Wang, and Jindong Chen. 2022. Screenqa: Large-scale question-answer pairs over mobile app screenshots. CoRR, abs/2209.08199.",
+            "Zhiyuan Huang, Ziming Cheng, Junting Pan, Zhaohui Hou, and Mingjie Zhan. 2025. Spiritsight agent: Advanced GUI agent with one look. CoRR, abs/2503.03196.",
+            "Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.",
+            "Raghav Kapoor, Yash Parag Butala, Melisa Russak, Jing Yu Koh, Kiran Kamble, Waseem AlShikh, and Ruslan Salakhutdinov. 2024. Omniact: A dataset and benchmark for enabling multimodal generalist autonomous agents for desktop and web. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXVIII, volume 15126 of Lecture Notes in Computer Science, pages 161-178. Springer.",
+            "Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, and Jie Tang. 2024. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, pages 5295-5306. ACM.",
+            "Hongxin Li, Jingfan Chen, Jingran Su, Yuntao Chen, Qing Li, and Zhaoxiang Zhang. 2025a. Autogui: Scaling GUI grounding with automatic functionality annotations from llms. CoRR, abs/2502.01977.",
+            "Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. 2025b. Screenshot-pro: Gui grounding for professional high-resolution computer use. Preprint, arXiv:2504.07981.",
+            "Wei Li, William W. Bishop, Alice Li, Christopher Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. 2024. On the effects of data scale on computer control agents. CoRR, abs/2406.03679.",
+            "Kevin Qinghong Lin, Linjie Li, Difei Gao, Zhengyuan Yang, Shiwei Wu, Zechen Bai, Weixian Lei, Lijuan Wang, and Mike Zheng Shou. 2024. Showui: One vision-language-action model for GUI visual agent. CoRR, abs/2411.17465.",
+            "Junpeng Liu, Tianyue Ou, Yifan Song, Yuxiao Qu, Wai Lam, Chenyan Xiong, Wenhu Chen, Graham Neubig, and Xiang Yue. 2025a. Harnessing webpage uis for text-rich visual understanding. In The Thirteenth International Conference on Learning Representations,"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net.",
+            "Xiao Liu, Bo Qin, Dongzhu Liang, Guang Dong, Hanyu Lai, Hanchen Zhang, Hanlin Zhao, Iat Long Iong, Jiadai Sun, Jiaqi Wang, Junjie Gao, Junjun Shan, Kangning Liu, Shudan Zhang, Shuntian Yao, Siyi Cheng, Wentao Yao, Wenyi Zhao, Xinghan Liu, and 11 others. 2024. Autoglm: Autonomous foundation agents for guis. CoRR, abs/2411.00820.",
+            "Yuhang Liu, Pengxiang Li, Zishu Wei, Congkai Xie, Xueyu Hu, Xinchen Xu, Shengyu Zhang, Xiaotian Han, Hongxia Yang, and Fei Wu. 2025b. Infiguia-gent: A multimodal generalist GUI agent with native reasoning and reflection. CoRR, abs/2501.04575.",
+            "Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025c. Visual-rft: Visual reinforcement fine-tuning. CoRR, abs/2503.01785.",
+            "Yadong Lu, Jianwei Yang, Yelong Shen, and Ahmed Awadallah. 2024. Omniparser for pure vision based GUI agent. CoRR, abs/2408.00203.",
+            "Zhengxi Lu, Yuxiang Chai, Yaxuan Guo, Xi Yin, Liang Liu, Hao Wang, Han Xiao, Shuai Ren, Guanjing Xiong, and Hongsheng Li. 2025. Ui-r1: Enhancing efficient action prediction of gui agents by reinforcement learning. Preprint, arXiv:2503.21620.",
+            "Haohao Luo, Yang Deng, Ying Shen, See-Kiong Ng, and Tat-Seng Chua. 2024. Chain-of-exemplar: Enhancing distractor generation for multimodal educational question generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 7978-7993. Association for Computational Linguistics.",
+            "Run Luo, Lu Wang, Wanwei He, and Xiaobo Xia. 2025. Gui-r1: A generalist r1-style vision-language action model for gui agents. Preprint, arXiv:2504.10458.",
+            "OpenAI,., Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Madry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024. Gpt-4o system card. Preprint, arXiv:2410.21276.",
+            "OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.",
+            "Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, Wanjun Zhong, Kuanye Li, Jiale Yang, Yu Miao, Woyu Lin, Longxiang Liu, Xu Jiang, Qianli Ma, Jingyu Li, and 16 others. 2025. UI-TARS: pioneering automated GUI interaction with native agents. CoRR, abs/2501.12326."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            920
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "9158",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Zero: memory optimizations toward training trillion parameter models. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2020, Virtual Event / Atlanta, Georgia, USA, November 9-19, 2020, page 20. IEEE/ACM.",
+            "Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. CoRR, abs/2402.03300.",
+            "Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. 2025. Vlm-r1: A stable and generalizable r1-style large vision-language model. Preprint, arXiv:2504.07615.",
+            "Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. Preprint, arXiv:2403.05530.",
+            "Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. 2024a. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024.",
+            "Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. 2024b. Mobile-agent: Autonomous multi-modal mobile device agent with visual perception. CoRR, abs/2401.16158.",
+            "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. 2024c. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. CoRR, abs/2409.12191.",
+            "Xin Wang, Yudong Chen, and Wenwu Zhu. 2022. A survey on curriculum learning. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4555-4576.",
+            "Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, and Yu Qiao. 2024. Os-atlas: A foundation action model for generalist gui agents. Preprint, arXiv:2410.23218.",
+            "Yiheng Xu, Zekun Wang, Junli Wang, Dunjie Lu, Tianbao Xie, Amrita Saha, Doyen Sahoo, Tao Yu, and Caiming Xiong. 2024. Aguvis: Unified pure vision agents for autonomous gui interaction. Preprint, arXiv:2412.04454."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Yuhao Yang, Yue Wang, Dongxu Li, Ziyang Luo, Bei Chen, Chao Huang, and Junnan Li. 2024. Ariaui: Visual grounding for gui instructions. Preprint, arXiv:2412.16256.",
+            "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 The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.",
+            "Chi Zhang, Zhao Yang, Jiaxuan Liu, Yanda Li, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. 2025. Appagent: Multimodal agents as smartphone users. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI 2025, Yokohama Japan, 26 April 2025-1 May 2025, pages 70:1-70:20. ACM.",
+            "Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            443
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A Datasets Details",
+        "text_level": 1,
+        "bbox": [
+            512,
+            456,
+            685,
+            470
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We provide training data used in the curriculum learning stage and reinforcement learning stage, as well as statistical information for the four datasets used in the experiment, as shown in Table 8 and Table 7, respectively.",
+        "bbox": [
+            510,
+            482,
+            880,
+            562
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            512,
+            574,
+            749,
+            590
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We build our GUI agent upon Qwen2-VL-2B-Instruct (Wang et al., 2024c), a vision-language model with strong general capabilities. It is worth noting that we chose Qwen2-VL-2B instead of Qwen's latest series Qwen2.5-VL-3B (Bai et al., 2025) because 2B is lighter, and most of the methods we compared used Qwen2-VL-2B as the base model or larger parameter models before Qwen2-VL. We validated the effectiveness of our proposed method in a fairer situation. The training process consists of two stages: Curriculum Learning (CL) and Reinforcement Learning (RL). For Curriculum Learning, we employ the LLaMA Factory (Zheng et al., 2024) framework for one epoch at each stage with the visual backbone frozen. We employ AdamW with a cosine annealing learning rate scheduler, setting the initial learning rate to 1e-5 and batch size to 16. For Reinforcement Learning, we use the R1-V (Chen et al., 2025) framework for training over two epochs, setting the initial learning",
+        "bbox": [
+            510,
+            599,
+            882,
+            920
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "9159",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/0a1cc91930c2d672ae7eca19c0e4d7867e885591e852c5f086bcf92043e16e64.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Datasets</td><td>Platforms</td><td>Task</td><td>Metric</td><td># Test Samples</td><td>History?</td></tr><tr><td>ScreenSpot-Pro (Li et al., 2025b)</td><td>Web, PC, Mobile</td><td>Grounding</td><td>clickAcc</td><td>1,272</td><td>X</td></tr><tr><td>ScreenSpot (Cheng et al., 2024)</td><td>Web, PC, Mobile</td><td>Grounding</td><td>clickAcc</td><td>1,581</td><td>X</td></tr><tr><td>Mind2Web (Deng et al., 2023)</td><td>Web</td><td>Navigation</td><td>Ele.Acc, Op.F1, Step SR</td><td>6,418</td><td>✓</td></tr><tr><td>AndroidControl-High (Li et al., 2024)</td><td>Mobile</td><td>Navigation</td><td>Step Accuracy</td><td>8,444</td><td>✓</td></tr><tr><td>AndroidControl-Low (Li et al., 2024)</td><td>Mobile</td><td>Grounding</td><td>Step Accuracy</td><td>8,444</td><td>X</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            880,
+            164
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/0dd1d5939e6588d21969f028a4d9a67677b88a0c02e6e4afd499df1d8cc7a22e.jpg",
+        "table_caption": [
+            "Table 7: Statistics of GUI datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Usage</td><td>Source</td><td>Number</td></tr><tr><td rowspan=\"7\">CL</td><td>GUIChat(Chen et al., 2024)</td><td>40k</td></tr><tr><td>ScreenQA (Hsiao et al., 2022)</td><td>17k</td></tr><tr><td>MultiUI (Liu et al., 2025a)</td><td>50k</td></tr><tr><td>AutoGUI (Li et al., 2025a)</td><td>50k</td></tr><tr><td>SeeClick (Cheng et al., 2024)</td><td>100k</td></tr><tr><td>GUIEnv (Chen et al., 2024)</td><td>100k</td></tr><tr><td>Aguvis-stage2 (Xu et al., 2024)</td><td>35k</td></tr><tr><td>RL</td><td>OmniAct (Kapoor et al., 2024)</td><td>2k</td></tr><tr><td>Total</td><td></td><td>394k</td></tr></table>",
+        "bbox": [
+            127,
+            211,
+            475,
+            360
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Training data statistics from various sources.",
+        "bbox": [
+            117,
+            370,
+            480,
+            384
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "rate to 1e-6 and batch size to 16. At each stage, we utilize the strategies of DeepSpeed optimization (Rajbhandari et al., 2020), BF16 format, and gradient checkpointing to save GPU memory. All training is conducted on  $2^{*}\\mathrm{A}100$  GPU (80G).",
+        "bbox": [
+            112,
+            410,
+            489,
+            491
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "9160",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2025/INREACT_ An Inspire-Then-Reinforce Training Framework For Multimodal GUI Agent/889785bd-0923-422b-bcca-82b35b178935_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3123cf43cce3ffa0dc42b94ee89c48e32a3fbbc9a4f417d4b2c8f800b328570
+size 798755

2025/INREACT_ An Inspire-Then-Reinforce Training Framework For Multimodal GUI Agent/full.md

@@ -0,0 +1,394 @@
+# INREACT: An Inspire-Then-Reinforce Training Framework For Multimodal GUI Agent
+
+Yuanlei Wang $^{1*}$ , Liuzhou Zhang $^{2*}$ , Haohao Luo $^{1*}$ , Ying Shen $^{1,3\dagger}$
+
+<sup>1</sup>Sun Yat-sen University, <sup>2</sup>Central China Normal University,
+
+<sup>3</sup>Guangdong Provincial Key Laboratory of Fire Science and Intelligent Emergency Technology
+
+{wangylei7, luohh5}@mail2.sysu.edu.cn,zlz212@mail.ccnu.edu.cn,
+
+sheny76@mail.sysu.edu.cn
+
+# Abstract
+
+Graphical User Interface (GUI) interaction, which aims to develop an intelligent GUI agent that executes user instructions to perform tasks such as installing applications by controlling digital devices, has gained significant attention due to its practical value. Although current advanced multimodal large language models (LLMs) provide GUI agents with robust perception and reasoning capabilities, they often struggle with the precise localization of small elements. To tackle this problem, we propose INREACT, a multimodal GUI agent framework that unifies observing, thinking, and acting for precise and interpretable decision-making. It is trained via a two-stage process: curriculum learning to progressively build perception, grounding, and reasoning abilities, followed by reinforcement learning to refine pixel-level grounding with an outcome-based reward. We introduce a rule-based reward function that jointly optimizes action-type selection and pixel-level localization accuracy. Experimental results on multiple datasets demonstrate the superiority of INREACT in both grounding and navigation tasks.
+
+# 1 Introduction
+
+Graphical User Interface (GUI) interaction refers to a class of sequential decision-making tasks in which agents interpret user instructions and manipulate digital interfaces, including PCs and mobile devices, to accomplish specific goals. These interactions typically involve executing low-level operations like Click, Type, or Scroll in the correct order to complete complex tasks, such as searching for information. A capable GUI agent must possess two core abilities: visual observation and reasoning. Observation is particularly critical in GUI environments like PCs and mobile devices, where visual
+
+![](images/e38ea4c38d9f3067a6c4b2ec43e2927e4c1efa3cd666b264ccf89a7a81d8f870.jpg)  
+Figure 1: Overview of the INREACT interaction
+
+components are often densely packed and visually similar, making precise localization challenging. Meanwhile, reasoning is essential for complex, goal-directed interactions that involve sequential decision-making. Unlike single-step tasks, GUI interaction requires agents to consider both interface states and past actions. With the emergence of powerful multimodal large language models (MLLMs) (Wang et al., 2024c; Bai et al., 2025; Chen et al., 2023; Luo et al., 2024), recent approaches have enabled screenshot-based GUI control without relying on structured metadata such as HTML or accessibility trees (Zhang et al., 2025; Wang et al., 2024b). However, despite these advances, existing GUI agents still struggle with accurate grounding and coherent reasoning in complex, real-world scenarios.
+
+To address the challenges of imprecise grounding and weak multi-step reasoning, recent efforts have focused on training MLLM-based GUI agents via supervised fine-tuning (SFT) or reinforcement learning (RL). In SFT, models are trained on paired GUI-action datasets to align visual observations with user instructions. Some approaches (Hong et al., 2024; Cheng et al., 2024; Lin et al., 2024; Li et al., 2025a) enhance SFT by incorporating instruction tuning, vision-language pretraining, or synthetic data generation to boost coverage. However, these strategies commonly adopt a flat optimization objective, lacking curriculum or structural
+
+guidance, which limits the model's ability to progressively acquire complex capabilities. In contrast, reinforcement learning introduces outcome-driven optimization through environment feedback. Some methods (Lai et al., 2024; Liu et al., 2024) rely on sparse success signals to guide multi-step task completion. Despite this, many RL strategies remain coarse and fail to capture pixel-level grounding precision. Moreover, most RL pipelines lack strong initialization or structured learning stages, often resulting in unstable policies and inconsistent behaviors.
+
+In this work, we propose INREACT, a structured training framework for GUI agents that combines curriculum learning with reinforcement learning. The curriculum learning stage is designed to systematically build core agent capabilities—including perception, grounding, and reasoning through a staged learning process of increasing complexity. Furthermore, we introduce a reinforcement learning strategy with rule-based reward functions that jointly considers action-type correctness and distance-aware spatial accuracy, enabling more precise and reliable interactions in complex GUI environments.
+
+Our contributions are summarized as follows:
+
+- We propose a novel multimodal GUI agent framework, called INREACT, which synergizes observing, thinking, and acting to make more interpretable and goal-directed decisions when handling complex GUI navigation tasks.  
+- INREACT is trained through a two-stage training framework, including a curriculum-based supervised fine-tuning stage to progressively acquire perception, grounding, and reasoning capabilities and a reinforcement learning stage that reinforces grounding precision with an outcome-based reward to encourage accurate and consistent action predictions.  
+- Experimental results on four datasets in grounding and navigation tasks show that INREACT outperforms existing methods and exhibits promising generalization. Our code will be released via https://github.com/wangyuanlei30/InReAct.
+
+# 2 Related Work
+
+# 2.1 GUI Agents
+
+GUI agents allow models to interact with software interfaces via visual inputs and simulated actions. Early works like CogAgent (Hong et al., 2024)
+
+introduced MLLMs for GUI tasks. Some methods leverage general-purpose models with prompt engineering (e.g., AppAgent (Zhang et al., 2025), Mobile-Agent (Wang et al., 2024b)), while others fine-tune open-source models on GUI-specific data (e.g., UGround (Gou et al., 2025), ShowUI (Lin et al., 2024)). Recent work like UI-TARS (Qin et al., 2025) and AGUVIS (Xu et al., 2024) combines GUI pretraining with reasoning fine-tuning. Most rely on single-stage supervised fine-tuning, which requires large-scale data and often struggles with generalization, highlighting the need for more structured learning approaches.
+
+# 2.2 Curriculum Learning
+
+Curriculum learning (Wang et al., 2022) is a training strategy where models are exposed to tasks in increasing order of difficulty, mimicking human learning. It has shown strong effectiveness across NLP and vision tasks by improving learning stability and generalization. Recent works have applied this idea to large-scale models: AutoGLM (Liu et al., 2024) and AutoWebGLM (Lai et al., 2024), for example, design task curricula to enhance model robustness in web navigation tasks. In the context of instruction tuning and agent learning, curriculum learning helps models acquire basic capabilities before handling complex reasoning or planning, offering a structured alternative to flat fine-tuning.
+
+# 2.3 Rule-Based Reinforcement Learning
+
+Rule-based reinforcement learning has emerged as a scalable alternative to traditional supervised fine-tuning by optimizing models with predefined and verifiable reward functions. Early works such as DeepSeek-R1 (Guo et al., 2025) and o1 (Jaech et al., 2024) applied RFT to language tasks, including mathematical reasoning (Shao et al., 2024) and code generation, where correctness can be evaluated with rule-based signals. More recent studies (Shen et al., 2025; Liu et al., 2025c; Chen et al., 2025; Wang et al., 2024a) extend this paradigm to multimodal models by designing task-specific rewards—such as Intersection-over-Union (IoU) for image grounding and symbolic consistency checks for classification. Building on this line of research, works like UI-R1 (Lu et al., 2025) and GUI-R1 (Luo et al., 2025) have further explored applying rule-based RL to GUI agents, typically leveraging binary hit-or-miss rewards for GUI grounding.
+
+![](images/649ab4ab11431d5efe76d29ad7683f111563fdf6fbaf79660e2fa5a977f48fcd.jpg)  
+Figure 2: The overall framework of INREACT
+
+# 3 Method
+
+Given a GUI screenshot  $S$ , a user instruction  $G$  and the previous interaction history  $A_{<t}$  at each time step  $t$ , the goal of the agent is to generate a sequence of executable actions  $A = \{a_1, a_2, \ldots, a_T\}$ , where action  $a_t$  at time step  $t$  is determined based on  $G$ , observation of the current screenshot environment  $O_t$ , and previous actions sequence  $A_{<t} = [a_1, a_2, \ldots, a_{t-1}]$ .
+
+- State Space. At each time step  $t$ , the environment state is represented by a GUI screenshot  $s_t$ , which captures the current visual interface after executing the previous action. The agent's observation  $O_t$  includes the screenshot  $S_t$ , the user instruction  $G$ , and the action history  $A_{<t} = \{a_1, \dots, a_{t-1}\}$ . This combination provides both static semantic goals and dynamic context necessary for decision-making.  
+- Action Space. We define an action as a structured function call in the form  $a_{t} = (\text{type}_{t}, \text{args}_{t})$ , where type specifies the operation (e.g., click, scroll, input) and args provides necessary arguments such as coordinates, text, or direction. The environment executes actions parsed from the agent's JSON-style outputs. A full specification of supported actions is provided in Table 1.
+
+# 3.1 GUI Agent Workflow
+
+To guide the agent to continually interact with the environment and complete multi-step complex
+
+<table><tr><td>Action Type</td><td>Parameters</td></tr><tr><td>click,tap,hover,select</td><td>(x,y)</td></tr><tr><td>swipe,select_text</td><td>(x1,y1,x2,y2)</td></tr><tr><td>scroll</td><td>direction</td></tr><tr><td>input</td><td>text</td></tr><tr><td>home,back,enter</td><td>None</td></tr></table>
+
+Table 1: Action types and their parameters
+
+tasks. Following ReAct (Yao et al., 2023), we construct our agent framework through synergizing observing, thinking and acting in GUI controlling process for agent.
+
+Observation Phase. Based on a visual screenshot  $S_{t}$  of the GUI, an user instruction  $G$  and the previous interaction history  $A_{< t}$  at each time step  $t$ , the agent first generates a natural language observation  $d_{t}$  describing key interface elements and their relevance to the task instead of directly generating action by:
+
+$$
+d _ {t} = \mathcal {F} \left(O _ {t}\right) = \mathcal {F} \left(S _ {t}, G, A _ {<   t}\right) \tag {1}
+$$
+
+Thinking Phase. In the thinking phase, the generated observation  $d_{t}$  is appended to the original input  $O_{t} = \{S_{t}, G, A_{<t}\}$  to construct the further input. Then, we feed the updated input to the agent for generating internal reasoning to produce a task-oriented thinking  $h_{t}$  by:
+
+$$
+h _ {t} = \mathcal {F} \left(O _ {t}, d _ {t}\right) \tag {2}
+$$
+
+Acting Phase. Similarly, the input in the final action phase is constructed by concatenating the generated thinking  $h_t$  with the previous input. Subsequently, we feed the modified input to the agent to generate executable actions by
+
+$$
+a _ {t} = \mathcal {F} \left(O _ {t}, d _ {t}, h _ {t}\right) = \left(\text {t y p e} _ {t}, \arg s _ {t}\right) \tag {3}
+$$
+
+The action is then executed in the environment, resulting in a transition to the next state  $S_{t + 1}$ .
+
+# 3.2 Inspire-Then-Reinforce Training
+
+To progressively enhance the reasoning and control capabilities of GUI agents, we propose an inspire-then-reinforce two-stage training paradigm, as illustrated in Figure 2.
+
+In Stage 1, we introduce a curriculum learning strategy that progressively builds the model's perception, grounding, and reasoning capabilities through supervised fine-tuning. By organizing training tasks from simple to complex, the model first learns to understand GUI layouts, then localize targets, and ultimately reason over multi-step goals.
+
+In Stage 2, we perform reinforcement learning to improve action accuracy and grounding robustness. Using a GRPO-based optimization framework, we incorporate a task-specific reward function that evaluates both action-type correctness and spatial precision, enabling the model to generate reliable actions in dense and high-resolution GUI environments.
+
+# 3.2.1 Stage 1: Curriculum Learning
+
+Drawing inspiration from human learning, we organize agent training into three progressive stages: perception, grounding, and reasoning. Each stage targets a specific skill and increases in task complexity, forming a curriculum learning framework that guides the model from basic comprehension to advanced planning.
+
+Specifically, we adopt supervised fine-tuning and have the same training objective that minimizes the sum of negative log-likelihood loss averaged over tokens in each step:
+
+$$
+\mathcal {L} (y, \widehat {y}) = - \frac {1}{L} \sum_ {l = 1} ^ {L} \widehat {y _ {l}} \log \left(\frac {\exp \left(y _ {l}\right)}{\sum_ {i} ^ {l} \exp \left(y _ {i}\right)}\right), \tag {4}
+$$
+
+where  $L$  is the max length of output sequence,  $\widehat{y}_l$  and  $y_{l}$  denote the  $l$ -th token in the groundtruth sequence  $\widehat{y}$  and generation sequence  $y$ , respectively.
+
+Learning for Perception The first stage of the curriculum aims to build the model's visual and semantic understanding of GUI layouts. This foundational perception equips the agent with spatial awareness and interface comprehension, essential for later grounding and reasoning. Specifically, we decompose it into two perception-oriented tasks:
+
+i. GUI-based Question Answering. Given a screenshot  $s$  and question  $q$ , the model generates an answer  $a\colon \mathcal{M}_{\mathrm{qa}}(s,q)\to a$ , enhancing spatial attention and element recognition.  
+ii. Interface Description Generation. Given  $s$ , the model produces a structured description  $d$ :  $\mathcal{M}_{\mathrm{desc}}(s) \to d$ , fostering global layout and semantic understanding. The training loss for this stage is computed as:
+
+$$
+\mathcal {L} _ {\text {p e r c e p t i o n}} = \mathcal {L} (a, \widehat {a}) + \mathcal {L} (d, \widehat {d}), \tag {5}
+$$
+
+where  $\widehat{s}$  and  $\widehat{d}$  denote the ground-truth answers and interface descriptions, respectively.
+
+Learning for Grounding The second stage of the curriculum focuses on grounding user instructions to visual actions, a core capability for operating in GUI environments. The model must understand what the user wants and accurately identify the corresponding interface region for execution. We similarly decompose it into two complementary grounding tasks:
+
+i. Instruction-to-Coordinate. Given a GUI screenshot  $s$  and instruction  $g$ , the model predicts the target coordinate  $c$ :  $\mathcal{M}_{\mathrm{inst2coord}}(s, g) \to c$ . This task aligns user intent with precise UI locations.  
+ii. Coordinate-to-Description. Given  $s$  and a coordinate  $c$ , the model generates a description  $d$  of the corresponding UI element:  $\mathcal{M}_{\mathrm{coord2desc}}(s,c) \to d$ , reinforcing spatial-semantic understanding. The loss function for this stage is defined as:
+
+$$
+\mathcal {L} _ {\text {g r o u n d i n g}} = \mathcal {L} (c, \widehat {c}) + \mathcal {L} (d, \widehat {d}), \tag {6}
+$$
+
+where  $\widehat{c}$  and  $\widehat{d}$  denote the ground-truth coordinate and description, respectively.
+
+Learning for Reasoning The third stage focuses on high-level reasoning and multi-step decision-making, where the model explicitly follows an observe-think-act process instead of making one-step predictions. We introduce the following two components to support structured reasoning:
+
+i. Observation Chain. Given a screenshot  $x$  and instruction  $g$ , the model generates an observation  $o$  to identify relevant UI regions:  $\mathcal{M}_{\mathrm{obs}}(x, g) \to o$ .
+
+ii. Reasoning Chain. Conditioned on  $o$  and  $g$ , the model produces a reasoning trace  $r$  to guide decision-making:  $\mathcal{M}_{\mathrm{reason}}(o,g)\to r$ .
+
+We define the training loss of this stage as follows:
+
+$$
+\mathcal {L} _ {\text {r e a s o n i n g}} = \mathcal {L} (o, \widehat {o}) + \mathcal {L} (r, \widehat {r}), \tag {7}
+$$
+
+where  $\widehat{o}$  and  $\widehat{r}$  denote the ground-truth observation and reasoning trace.
+
+# 3.2.2 Stage 2: Reinforcement Learning
+
+Although supervised fine-tuning provides a strong initialization, models often suffer from fine-grained prediction errors, especially in dense or high-resolution GUIs where small pixel-level deviations can lead to task failure. To address this, we adopt GRPO Reinforcement fine-tuning.
+
+A critical component of our approach is the design of the reward function. While prior works like UI-R1 (Lu et al., 2025) and GUI-R1 (Luo et al., 2025) also leverage rule-based reinforcement learning for GUI agents, their reward mechanisms for localization typically rely on a coarse, binary signal (i.e., whether the predicted point falls within the target's bounding box). Such a reward, while useful for confirming task success, provides limited gradient for improving fine-grained precision. To overcome this, we introduce a task-specific reward function that better captures the dual objectives of action correctness and spatial precision, directly encouraging the model to enhance its pixel-level accuracy.
+
+Reward Modeling The reward function serves as the primary training signal in reinforcement learning, directly guiding the model's optimization trajectory. To support fine-grained policy improvement in GUI environments, we design a rule-based reward function tailored to the unique demands of screen-based interaction. Unlike traditional RL settings that rely on binary task success or coarse region-level feedback, GUI tasks require both accurate action selection and high-precision localization of target elements. Specifically, our reward function comprises two components: (1) an action-type reward(2) a distance-aware coordinate reward. This composite design provides the model with dense, informative feedback signals essential for precise and interpretable GUI control, which is defined as:
+
+$$
+R = R _ {\mathrm {T}} + R _ {\mathrm {D}} \tag {8}
+$$
+
+Regarding the action type reward component  $R_{\mathrm{T}}$ , it provides a binary reward signal that evaluates whether the predicted action type  $T$  aligns with the ground truth label  $T'$ . This mechanism offers a direct and effective evaluation framework for action-type prediction tasks. The mathematical formulation of this component is presented below:
+
+$$
+R _ {\mathrm {T}} = \left\{ \begin{array}{l l} 1, & \text {i f} T ^ {\prime} = T \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {9}
+$$
+
+Additionally, the distance-aware reward  $R_{\mathrm{D}}$  measures spatial accuracy by combining two components: (1) a binary reward for whether the predicted point lies within the target bounding box, and (2) a precision reward based on the normalized distance to the box center. This encourages both correct region prediction and fine-grained localization:
+
+$$
+R _ {D} = R _ {a c c} + R _ {d i s t} \tag {10}
+$$
+
+i. Accuracy Reward. We assign a base reward calculating whether the predicted point  $(x', y')$  falls inside the ground-truth bounding box  $B = [(x_1, y_1), (x_2, y_2)]$ :
+
+$$
+R _ {\mathrm {a c c}} = \left\{ \begin{array}{l l} 0. 5, & \text {i f} \left(x ^ {\prime}, y ^ {\prime}\right) \in B \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {11}
+$$
+
+ii. Precision Reward. If the point is within the box, we compute the normalized Euclidean distance  $\hat{d}$  from the predicted coordinate  $(x', y')$  to the center of the target box  $(x_c, y_c)$ :
+
+$$
+\left(x _ {c}, y _ {c}\right) = \left(\frac {x _ {1} + x _ {2}}{2}, \frac {y _ {1} + y _ {2}}{2}\right) \tag {12}
+$$
+
+$$
+\hat {d} = \frac {\sqrt {\left(x ^ {\prime} - x _ {c}\right) ^ {2} + \left(y ^ {\prime} - y _ {c}\right) ^ {2}}}{\sqrt {\left(x _ {2} - x _ {1}\right) \left(y _ {2} - y _ {1}\right)}} \tag {13}
+$$
+
+We then define the precision reward based on discrete distance intervals:
+
+$$
+R _ {\text {d i s t}} = \left\{ \begin{array}{l l} 0. 5, & \hat {d} \leq 0. 1 \\ 0. 3, & 0. 1 <   \hat {d} \leq 0. 3 \\ 0. 2, & 0. 3 <   \hat {d} \leq 0. 5 \\ 0. 0, & \hat {d} > 0. 5 \end{array} \right. \tag {14}
+$$
+
+To provide a clear and stepped gradient for precision improvement, we empirically determine the thresholds and reward values for our reward formulation. This design provides a continuous and informative learning signal that encourages the model to both select the correct action type and localize it precisely, promoting stable convergence and robust behavior in dense GUI layouts.
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Development</td><td colspan="3">Creative</td><td colspan="3">CAD</td><td colspan="3">Scientific</td><td colspan="3">Office</td><td colspan="3">OS</td><td colspan="3">Overall</td></tr><tr><td colspan="3">Text Icon Avg</td><td colspan="3">Text Icon Avg</td><td colspan="3">Text Icon Avg</td><td colspan="3">Text Icon Avg</td><td colspan="3">Text Icon Avg</td><td colspan="3">Text Icon Avg</td><td colspan="3">Text Icon Avg</td></tr><tr><td>GPT-4o (OpenAI et al., 2024)</td><td>1.3</td><td>0.0</td><td>0.7</td><td>1.0</td><td>0.0</td><td>0.6</td><td>2.0</td><td>0.0</td><td>1.5</td><td>2.1</td><td>0.0</td><td>1.2</td><td>1.1</td><td>0.0</td><td>0.9</td><td>0.0</td><td>0.0</td><td>0.0</td><td>1.3</td><td>0.0</td><td>0.8</td></tr><tr><td>ShowUI-2B (Lin et al., 2024)</td><td>16.9</td><td>1.4</td><td>9.4</td><td>9.1</td><td>0.0</td><td>5.3</td><td>2.5</td><td>0.0</td><td>1.9</td><td>13.2</td><td>7.3</td><td>10.6</td><td>15.3</td><td>7.5</td><td>13.5</td><td>10.3</td><td>2.2</td><td>6.6</td><td>10.8</td><td>2.6</td><td>7.7</td></tr><tr><td>AriaUI-3.9B (Yang et al., 2024)</td><td>16.2</td><td>0.0</td><td>8.4</td><td>23.7</td><td>2.1</td><td>14.7</td><td>7.6</td><td>1.6</td><td>6.1</td><td>27.1</td><td>6.4</td><td>18.1</td><td>20.3</td><td>1.9</td><td>16.1</td><td>4.7</td><td>0.0</td><td>2.6</td><td>17.1</td><td>2.0</td><td>11.3</td></tr><tr><td>OSAtlas-4B (Wu et al., 2024)</td><td>7.1</td><td>0.0</td><td>3.7</td><td>3.0</td><td>1.4</td><td>2.3</td><td>2.0</td><td>0.0</td><td>1.5</td><td>9.0</td><td>5.5</td><td>7.5</td><td>5.1</td><td>3.8</td><td>4.8</td><td>5.6</td><td>0.0</td><td>3.1</td><td>5.0</td><td>1.7</td><td>3.7</td></tr><tr><td>QwenVL-7B (Bai et al., 2023)</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.7</td><td>0.0</td><td>0.4</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.1</td><td>0.0</td><td>0.1</td></tr><tr><td>Qwen2-VL-7B (Wang et al., 2024c)</td><td>2.6</td><td>0.0</td><td>1.3</td><td>1.5</td><td>0.0</td><td>0.9</td><td>0.5</td><td>0.0</td><td>0.4</td><td>6.3</td><td>0.0</td><td>3.5</td><td>3.4</td><td>1.9</td><td>3.0</td><td>0.9</td><td>0.0</td><td>0.5</td><td>2.5</td><td>0.2</td><td>1.6</td></tr><tr><td>SeeClick-7B (Cheng et al., 2024)</td><td>0.6</td><td>0.0</td><td>0.3</td><td>1.0</td><td>0.0</td><td>0.6</td><td>2.5</td><td>0.0</td><td>1.9</td><td>3.5</td><td>0.0</td><td>2.0</td><td>1.1</td><td>0.0</td><td>0.9</td><td>2.8</td><td>0.0</td><td>1.5</td><td>1.8</td><td>0.0</td><td>1.1</td></tr><tr><td>CogAgent-18B (Hong et al., 2024)</td><td>14.9</td><td>0.7</td><td>8.0</td><td>9.6</td><td>0.0</td><td>5.6</td><td>7.1</td><td>3.1</td><td>6.1</td><td>22.2</td><td>1.8</td><td>13.4</td><td>13.0</td><td>0.0</td><td>10.0</td><td>5.6</td><td>0.0</td><td>3.1</td><td>12.0</td><td>0.8</td><td>7.7</td></tr><tr><td>UGround-7B (Gou et al., 2025)</td><td>26.6</td><td>2.1</td><td>14.7</td><td>27.3</td><td>2.8</td><td>17.0</td><td>14.2</td><td>1.6</td><td>11.1</td><td>31.9</td><td>2.7</td><td>19.3</td><td>31.6</td><td>11.3</td><td>27.0</td><td>17.8</td><td>0.0</td><td>9.7</td><td>25.0</td><td>2.8</td><td>16.5</td></tr><tr><td>OSAtlas-7B (Wu et al., 2024)</td><td>33.1</td><td>1.4</td><td>17.7</td><td>28.8</td><td>2.8</td><td>17.9</td><td>12.2</td><td>4.7</td><td>10.3</td><td>37.5</td><td>7.3</td><td>24.4</td><td>33.9</td><td>5.7</td><td>27.4</td><td>27.1</td><td>4.5</td><td>16.8</td><td>28.1</td><td>4.0</td><td>18.9</td></tr><tr><td>InReAct</td><td>35.7</td><td>1.4</td><td>19.0</td><td>33.3</td><td>2.1</td><td>20.2</td><td>13.2</td><td>1.6</td><td>10.7</td><td>41.7</td><td>10.0</td><td>27.9</td><td>42.4</td><td>3.8</td><td>33.5</td><td>23.4</td><td>1.1</td><td>17.3</td><td>31.4</td><td>3.3</td><td>20.7</td></tr></table>
+
+# 3.3 Inference Procedure
+
+During inference, our agent adopts an interlaced observe-think-action workflow that sequentially executes tasks by observing the interface, thinking through the instruction, and predicting the next action. Given the input  $O_{t} = (S,G,H)$ , where  $S$  is the GUI screenshot,  $G$  is the task instruction, and  $A_{<t} = \{a_{1},\dots,a_{t-1}\}$  denotes the history of executed actions, the model first generates an observation summary  $\hat{o}_{t}$ , then infers a reasoning trace  $\hat{r}_{t}$  conditioned on  $\hat{o}_{t}$  and  $G$ , and finally predicts the next action  $a_{t} = (T',x',y')$  based on both. This process is repeated until task completion or a maximum number of steps is reached, enabling goal-driven, interpretable decision-making in complex GUI environments.
+
+# 4 Experiment
+
+# 4.1 Experiment Setup
+
+Datasets We evaluate our model on two tasks: GUI grounding and navigation.
+
+For Grounding, we use ScreenSpot (Cheng et al., 2024) and ScreenSpot-Pro (Li et al., 2025b). ScreenSpot contains 1,272 samples across mobile, desktop, and web platforms, focusing on common interface elements. ScreenSpot-Pro includes 23 professional applications with high-resolution and complex layouts, offering a more challenging evaluation.
+
+For Navigation, we use: (i) Mind2Web (Deng et al., 2023), a web-based dataset for evaluating generalist agents that perform complex tasks via language instructions. It includes 7,775 training actions and three test splits (task, website, domain), with aligned HTML and screenshots. (ii) Android-Control (Li et al., 2024), a mobile dataset testing multi-step task execution in realistic Android en
+
+Table 2: GUI grounding results on ScreenSpot-Pro dataset.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Mobile</td><td colspan="2">Desktop</td><td colspan="2">Web</td><td rowspan="2">Average</td></tr><tr><td>Text</td><td>Icon</td><td>Text</td><td>Icon</td><td>Text</td><td>Icon</td></tr><tr><td>GPT-4 (OpenAI, 2023)</td><td>22.6</td><td>24.5</td><td>20.2</td><td>11.8</td><td>9.2</td><td>8.8</td><td>16.7</td></tr><tr><td>GPT-4o (OpenAI et al., 2024)</td><td>20.2</td><td>24.9</td><td>21.1</td><td>23.6</td><td>12.2</td><td>7.8</td><td>18.1</td></tr><tr><td>Gemini-1.5-pro (Team, 2024)</td><td>76.2</td><td>54.1</td><td>65.5</td><td>39.2</td><td>52.2</td><td>32.0</td><td>53.2</td></tr><tr><td>Qwen2-VL-2B (Wang et al., 2024c)</td><td>24.2</td><td>10.0</td><td>1.4</td><td>9.3</td><td>8.7</td><td>2.41</td><td>9.3</td></tr><tr><td>Qwen2-VL-7B (Wang et al., 2024c)</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>42.9</td></tr><tr><td>CogAgent-18B (Hong et al., 2024)</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-7B (Cheng et al., 2024)</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>53.4</td></tr><tr><td>UGround-7B (Gou et al., 2025)</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>73.3</td></tr><tr><td>ShowUI-2B (Lin et al., 2024)</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>75.1</td></tr><tr><td>InfiGUIAgent-2B (Liu et al., 2025b)</td><td>88.6</td><td>74.7</td><td>85.6</td><td>65.0</td><td>79.1</td><td>64.6</td><td>76.3</td></tr><tr><td>InReAct</td><td>86.1</td><td>71.6</td><td>84.5</td><td>65.7</td><td>82.2</td><td>70.4</td><td>77.8</td></tr></table>
+
+Table 3: GUI grounding results on ScreenSpot dataset.
+
+vironments, emphasizing long-term planning and state tracking.
+
+See datasets details in appendix A.
+
+Evaluation Metrics We use the metrics proposed in the corresponding datasets as shown in Table 7. Although they have different names, these metrics are similar: GUI grounding tasks consistently measure the hit rate of predicted bounding boxes, while GUI navigation tasks focus on single-step accuracy.
+
+Click Accuracy. The proportion of test samples where the predicted location falls in the ground truth element's bounding box.
+
+Element Accuracy (Ele.Acc). Comparing the selected element with all acceptable elements. For vision-based methods, it is the same as Click Accuracy.
+
+Operation F1 (Op.F1). Token-level F1 score for the predicted operation.
+
+Step Success Rate (Step SR) and Step Accuracy. The proportion of successful steps. A step is regarded as successful only if both the selected element and the predicted operation are correct.
+
+Baselines We compare the performance of our model with various baselines, including some closed source large models such as GPT-4o (OpenAI et al., 2024), Gemini-1.5-pro (Team, 2024)
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Cross-Task</td><td colspan="3">Cross-Website</td><td colspan="3">Cross-Domain</td></tr><tr><td>Ele.Acc</td><td>Op.F1</td><td>Step.SR</td><td>Ele.Acc</td><td>Op.F1</td><td>Step.SR</td><td>Ele.Acc</td><td>Op.F1</td><td>Step.SR</td></tr><tr><td>GPT-4 (OpenAI, 2023)</td><td>41.6</td><td>60.6</td><td>36.2</td><td>35.8</td><td>51.1</td><td>30.1</td><td>37.1</td><td>46.5</td><td>26.4</td></tr><tr><td>OmniParser (Lu et al., 2024)</td><td>42.4</td><td>87.6</td><td>39.4</td><td>41.0</td><td>84.8</td><td>36.5</td><td>45.5</td><td>85.7</td><td>42.0</td></tr><tr><td>Qwen2-VL-2B (Wang et al., 2024c)</td><td>37.7</td><td>86.4</td><td>33.2</td><td>36.0</td><td>79.2</td><td>27.6</td><td>36.3</td><td>81.8</td><td>30.7</td></tr><tr><td>ShowUI-2B (Lin et al., 2024)</td><td>39.9</td><td>88.6</td><td>37.2</td><td>41.6</td><td>83.5</td><td>35.1</td><td>39.4</td><td>86.8</td><td>35.2</td></tr><tr><td>SpiritSight-2B (Huang et al., 2025)</td><td>51.7</td><td>87.2</td><td>44.9</td><td>44.9</td><td>83.6</td><td>37.8</td><td>42.4</td><td>83.5</td><td>36.9</td></tr><tr><td>SeeClick-7B (Cheng et al., 2024)</td><td>28.3</td><td>87.0</td><td>25.5</td><td>21.4</td><td>80.6</td><td>16.4</td><td>23.2</td><td>84.8</td><td>20.8</td></tr><tr><td>CogAgent-18B (Hong et al., 2024)</td><td>22.4</td><td>53.0</td><td>17.6</td><td>18.4</td><td>42.4</td><td>13.4</td><td>20.6</td><td>42.0</td><td>15.5</td></tr><tr><td>InReAct</td><td>57.8</td><td>89.4</td><td>50.8</td><td>52.6</td><td>88.8</td><td>47.4</td><td>55.4</td><td>85.4</td><td>49.2</td></tr></table>
+
+Table 4: GUI navigation results on Mind2web dataset.  
+
+<table><tr><td>Method</td><td>Low-Level</td><td>High-Level</td></tr><tr><td>Claude* (Anthropic, 2024)</td><td>19.4</td><td>12.5</td></tr><tr><td>GPT-4o (OpenAI et al., 2024)</td><td>19.4</td><td>20.8</td></tr><tr><td>Aria-UI-3.9B (Yang et al., 2024)</td><td>67.3</td><td>10.2</td></tr><tr><td>OS-Atlas-4B (Wu et al., 2024)</td><td>80.6</td><td>67.5</td></tr><tr><td>Aguvis-7B (Xu et al., 2024)</td><td>80.5</td><td>61.5</td></tr><tr><td>InReAct</td><td>83.1</td><td>68.1</td></tr></table>
+
+Table 5: GUI navigation results on AndroidControl dataset.
+
+and Claude(Anthropic, 2024), as well as numerous open-source GUI agents such as SpiritSight-2B (Huang et al., 2025), AGUVIS-7B (Xu et al., 2024), Uground (Gou et al., 2025), Aria-UI(Yang et al., 2024), OSAtlas (Wu et al., 2024), ShowUI (Lin et al., 2024), InfGUIAgent (Liu et al., 2025b), UGround (Gou et al., 2025), CogAgent (Hong et al., 2024), SeeClick (Cheng et al., 2024), Qwen2-VL-2B, Qwen2-VL-7B (Wang et al., 2024c). Implementation details refer to appendix B.
+
+# 4.2 Overall Performance
+
+GUI Grounding Evaluation The evaluation results on the ScreenSpot (Cheng et al., 2024) are presented in Table 3. Our model achieves an average accuracy of  $77.8\%$ , outperforming all competing baselines, including both open-source and commercial GUI agents. In particular, Our model surpasses strong 2B-level models such as ShowUI  $(75.1\%)$  and InfigUIAgent  $(76.3\%)$ , and even outperforms several significantly larger models such as CogAgent  $(47.4\%, 18\mathrm{B})$  and UGround  $(73.3\%, 7\mathrm{B})$ .
+
+As in shown in table 2, on the more challenging ScreenSpot-Pro (Li et al., 2025b) dataset, our method achieves an overall accuracy of  $20.7\%$ , representing a substantial improvement under complex, high-resolution desktop GUI environments. This dataset covers a wide spectrum of profes
+
+sional software, including CAD tools, scientific platforms, development IDEs, and creative applications, each demanding fine-grained grounding of visually dense and domain-specific interfaces. Despite the increased difficulty, our model consistently outperforms previous strong baselines. It exceeds ShowUI by  $13.0\%$  percentage points and surpasses AriaUI by  $9.4\%$  points, demonstrating particularly strong capabilities in GUI grounding
+
+GUI Navigation Evaluation On Mind2Web (Deng et al., 2023), the task requires agents to perform realistic web navigation across diverse websites. As shown in Table 4, InReAct achieves strong performance across all three generalization splits—cross-task, cross-website, and cross-domain—outperforming previous methods in terms of element grounding accuracy, operation-level F1, and step-wise success rate. This indicates that our model not only understands the semantics of web elements but also makes consistent and interpretable action decisions across varying environments. The consistent gains across generalization splits demonstrate InReAct's robustness in unseen task goals and layouts.
+
+On AndroidControl (Li et al., 2024), we evaluate the model under two settings: Low-Level, which focuses on simple, mostly single-step interactions (e.g., tapping a button); and High-Level, which involves complex, multi-step instructions with long-horizon dependencies. As shown in Table 5, InReAct achieves 83.1 on the Low-Level set and 68.1 on the High-Level set, outperforming all prior baselines. While strong results in Low-Level indicate precise grounding and action formulation, the improvements in High-Level highlight the model's ability to plan and reason over multi-step interaction sequences.
+
+# 4.3 Ablation Study
+
+As summarized in Table 6, we conduct ablation studies to assess the impact of the proposed strategies in both curriculum learning (CL) and reinforcement learning (RL) training stages. For Mind2Web and AndroidControl, we only reported the results on cross-task SR and High-level. There are several notable observations as follows:
+
+Curriculum Learning We examine the role of each step in our curriculum learning framework. Removing Step 1, where the model learns basic GUI interactions, leads to a significant performance drop, as the agent fails to acquire essential skills. Similarly, omitting Steps 2 and 3, which introduce grounding and reasoning tasks progressively, also harms performance, showing their importance for understanding and reasoning over GUI elements. Furthermore, replacing curriculum learning with multi-task training across all steps results in worse performance, demonstrating that staged, gradual training enhances learning and generalization more effectively than joint training.
+
+Reinforcement Learning To further evaluate the effectiveness of the reinforcement learning (RL) stage, we first replace it with a standard supervised fine-tuning (SFT) setup. The results show that RL brings consistent performance gains across all datasets. On the two navigation datasets—Mind2Web and AndroidControl—RL yields improvements of  $+5.1$  and  $+4.3$ , respectively, demonstrating its ability to produce more accurate actions and more robust decision-making.
+
+Furthermore, to specifically justify our reward function design within the RL framework, we compare our proposed piecewise reward function against two alternatives: a binary (hit-or-miss) reward, which provides an accuracy signal without any dedicated precision component, similar to approaches in works like UI-R1 and GUI-R1, and a continuous linear reward (defined as  $R_{dist} = 1 - \hat{d}$ , where  $\hat{d}$  is the normalized distance to the target's center). This analysis, presented in Table 6, reveals that the design of the reward signal plays an important role. On ScreenSpot, our method (77.8) outperforms both the linear (77.0) and binary (76.8) baselines. On the more challenging ScreenSpot-Pro dataset, our approach (20.7) also provides a modest but consistent improvement over the linear (20.3) and binary (19.8) rewards.
+
+The advantage of our piecewise reward function
+
+lies in two aspects. First, it delivers clearer learning signals by introducing reward "steps," which motivate the agent to cross precision thresholds rather than being limited by diminishing returns near the target. Second, it aligns more closely with real-world GUI logic, where clicks have discrete outcomes—successful or not—and central clicks are inherently safer than edge clicks. This design thus captures the practical structure of interaction more faithfully than a simple linear function. It is worth noting that our thresholds are chosen empirically and may not be optimal; further tuning is left for future work.
+
+Impact of Observation and Reasoning Components The results, presented in Table 6, underscore the importance of both components within our observe-think-act cycle. Removing the thinking phase (w/o Thought) leads to a consistent and significant performance drop across all four benchmarks, with the most notable decreases observed on the complex ScreenSpot-Pro (-2.5 points) and AndroidControl (-2.3 points) datasets. This indicates that the explicit reasoning step is crucial for formulating effective plans and making robust decisions in challenging environments.
+
+Similarly, removing the observation phase (w/o Observation) also consistently degrades performance across all tested datasets, such as causing a 1.8-point drop on ScreenSpot. An interesting finding is that the performance degradation from removing the thinking phase is more pronounced than that from removing the observation phase on the three more complex datasets. This suggests that while both components are beneficial, the structured reasoning provided by the "Thinking" step is arguably the most critical factor for the agent's success in complex scenarios. These findings strongly justify the synergistic design of our InReAct framework.
+
+# 4.4 Case Study
+
+To qualitatively assess the effectiveness of our proposed curriculum learning and reinforcement learning strategies, we conduct a case study on the ScreenSpot dataset. As illustrated in Figure 3, we present two pairs of examples highlighting the contributions of each training stage.
+
+In Figure 3a, without curriculum learning, the agent clicks on the text "Only draw with Apple Pencil" instead of the adjacent switch, showing weak perception and semantic grounding. In contrast,
+
+<table><tr><td>Method</td><td>ScreenSpot</td><td>ScreenSpot-Pro</td><td>Mind2Web</td><td>AndroidControl</td></tr><tr><td>InReAct</td><td>77.8</td><td>20.7</td><td>50.8</td><td>68.1</td></tr><tr><td colspan="5">Stage 1: Curriculum Learning</td></tr><tr><td>- w/o Step 1</td><td>76.3</td><td>19.2</td><td>47.4</td><td>65.9</td></tr><tr><td>- w/o Step 2</td><td>68.6</td><td>10.8</td><td>37.8</td><td>57.4</td></tr><tr><td>- w/o Step 3</td><td>74.7</td><td>17.4</td><td>42.5</td><td>62.3</td></tr><tr><td>- w/o curriculum</td><td>75.4</td><td>19.5</td><td>48.5</td><td>66.4</td></tr><tr><td colspan="5">Stage 2: Reinforcement Learning</td></tr><tr><td>- w/ Binary Reward</td><td>76.8</td><td>19.8</td><td>49.2</td><td>67.0</td></tr><tr><td>- w/ Linear Reward</td><td>77.0</td><td>20.3</td><td>50.2</td><td>67.3</td></tr><tr><td>- w/ SFT</td><td>76.2</td><td>16.8</td><td>45.7</td><td>63.8</td></tr><tr><td colspan="5">Ablation on Agent Components</td></tr><tr><td>- w/o Observation</td><td>76.0</td><td>19.7</td><td>49.2</td><td>66.8</td></tr><tr><td>- w/o Thought</td><td>76.6</td><td>18.2</td><td>48.6</td><td>65.8</td></tr></table>
+
+Table 6: Ablation study of INREACT. We evaluate the impact of different training stages, reward functions, and core agent components. For Mind2Web, we report the Step SR on the cross-task split, while for AndroidControl, we report results on the High-Level setting.
+
+![](images/30242008ba5479c269d169dc7468980d1329845214b4c78763a0798043ee748e.jpg)  
+(a)  $x$  Instruction:"Only draw with apple pencil". Without curriculum learning
+
+![](images/22dd080ce9d2cf4d857ca9c2d851439be4f122d7d136e250706666981b471fa5.jpg)  
+(b)  $\checkmark$  Instruction:"Only draw with apple pencil". InReAct achieves accurate grounding
+
+![](images/ca2ace5599a4fd0f17d5279997bea001d349def88360af261eaf366502053830.jpg)  
+(c)  $X$  Instruction:"Zoom in". Without reinforcement learning  
+Figure 3: Case study on ScreenSpot dataset.
+
+![](images/69de77b5a31c00e1f912585b6529d2dd3c56ab4bda9e80489f6566c91236f082.jpg)  
+(d)  $\checkmark$  Instruction:"Zoom in". InReAct achieves accurate grounding
+
+InReAct (Figure 3b) correctly targets the switch, indicating better alignment between instructions and interface elements.
+
+In Figure 3c, without reinforcement learning, for the instruction "Zoom in," the predicted click point is near the correct bounding box but slightly outside it, indicating general understanding but lacking fine-grained precision. In contrast, as shown in Figure 3d, InReAct achieves significantly more accurate localization, clicking within the target re
+
+gion. This confirms that our reinforcement learning stage enhances fine-grained control.
+
+# 5 Conclusion
+
+In this paper, we present INREACT, a multimodal GUI agent framework that integrates observation, reasoning, and action for complex GUI interactions. It is trained in two stages: curriculum-learning to build perception, grounding, and reasoning skills, followed by reinforcement learning with a task-specific reward that improves action correctness and spatial precision.
+
+We evaluate INREACT across four datasets: ScreenSpot, ScreenSpot-Pro, Mind2Web, and AndroidControl, and demonstrate consistent improvements over prior methods. Our findings highlight the effectiveness of combining structured learning and outcome-driven optimization for enhancing GUI agent performance, setting a new direction for generalizable and interpretable vision-language agents.
+
+# Acknowledgments
+
+This research was supported by Key-Area Research and Development Program of Guangdong Province, Granted No. 2024B1111060004.
+
+# Limitations
+
+Limited Action Space Coverage Our current action space, while sufficient for common GUI interactions such as click, type, and scroll, does not fully cover the diversity of real-world GUI operations. Advanced actions like drag-and-drop, multi-touch gestures, system-level controls, or context-menu navigation are not well supported in our framework. Extending the action space to include such capabilities would be essential for achieving broader applicability in real-world software environments.
+
+Inference Speed and Efficiency Although In-ReAct achieves strong performance on various GUI tasks, its inference speed has room for optimization. The structured reasoning process—particularly the multi-step observation and thinking phases—introduces additional computational overhead. This may affect the agent's ability to perform real-time or near-real-time interactions, which is crucial for applications requiring fast response times. Future work could explore model compression, caching mechanisms, or streamlined reasoning pipelines to improve efficiency.
+
+Limited Handling of Dynamic Interfaces Our experiments are primarily conducted on static interfaces. However, real-world GUI applications often involve dynamic content updates, such as pop-up notifications, auto Refreshing elements, or state transitions, that can disrupt perception and decision-making. Our current framework lacks explicit mechanisms to detect, adapt to, or reason about such changes over time. Incorporating temporal modeling or environment change detection would be a valuable direction for future research.
+
+# References
+
+Anthropic. 2024. Developing a computer use model. https://www.anthropic.com/news/ developing-computer-use. Accessed: 2025-04- 12.  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. CoRR, abs/2308.12966.
+
+Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Ming-Hsuan Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. CoRR, abs/2502.13923.  
+Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. 2025. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. https://github.com/Deep-Agent/R1-V. Accessed: 2025-02-02.  
+Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, Yuan Yao, Yankai Lin, Zhiyuan Liu, and Maosong Sun. 2024. Guicourse: From general vision language models to versatile GUI agents. CoRR, abs/2406.11317.  
+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. CoRR, abs/2312.14238.  
+Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. 2024. Seeclick: Harnessing GUI grounding for advanced visual GUI agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 9313-9332. Association for Computational Linguistics.  
+Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samual Stevens, Boshi Wang, Huan Sun, and Yu Su. 2023. Mind2web: Towards a generalist agent for the web. 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.  
+Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. 2025. Navigating the digital world as humans do: Universal visual grounding for GUI agents. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net.  
+Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.  
+Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, and Jie Tang. 2024. Co-gagent: A visual language model for GUI agents. In IEEE/CVF Conference on Computer Vision and
+
+Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 14281-14290. IEEE.  
+Yu-Chung Hsiao, Fedir Zubach, Maria Wang, and Jindong Chen. 2022. Screenqa: Large-scale question-answer pairs over mobile app screenshots. CoRR, abs/2209.08199.  
+Zhiyuan Huang, Ziming Cheng, Junting Pan, Zhaohui Hou, and Mingjie Zhan. 2025. Spiritsight agent: Advanced GUI agent with one look. CoRR, abs/2503.03196.  
+Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.  
+Raghav Kapoor, Yash Parag Butala, Melisa Russak, Jing Yu Koh, Kiran Kamble, Waseem AlShikh, and Ruslan Salakhutdinov. 2024. Omniact: A dataset and benchmark for enabling multimodal generalist autonomous agents for desktop and web. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXVIII, volume 15126 of Lecture Notes in Computer Science, pages 161-178. Springer.  
+Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, and Jie Tang. 2024. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024, pages 5295-5306. ACM.  
+Hongxin Li, Jingfan Chen, Jingran Su, Yuntao Chen, Qing Li, and Zhaoxiang Zhang. 2025a. Autogui: Scaling GUI grounding with automatic functionality annotations from llms. CoRR, abs/2502.01977.  
+Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. 2025b. Screenshot-pro: Gui grounding for professional high-resolution computer use. Preprint, arXiv:2504.07981.  
+Wei Li, William W. Bishop, Alice Li, Christopher Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. 2024. On the effects of data scale on computer control agents. CoRR, abs/2406.03679.  
+Kevin Qinghong Lin, Linjie Li, Difei Gao, Zhengyuan Yang, Shiwei Wu, Zechen Bai, Weixian Lei, Lijuan Wang, and Mike Zheng Shou. 2024. Showui: One vision-language-action model for GUI visual agent. CoRR, abs/2411.17465.  
+Junpeng Liu, Tianyue Ou, Yifan Song, Yuxiao Qu, Wai Lam, Chenyan Xiong, Wenhu Chen, Graham Neubig, and Xiang Yue. 2025a. Harnessing webpage uis for text-rich visual understanding. In The Thirteenth International Conference on Learning Representations,
+
+ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net.  
+Xiao Liu, Bo Qin, Dongzhu Liang, Guang Dong, Hanyu Lai, Hanchen Zhang, Hanlin Zhao, Iat Long Iong, Jiadai Sun, Jiaqi Wang, Junjie Gao, Junjun Shan, Kangning Liu, Shudan Zhang, Shuntian Yao, Siyi Cheng, Wentao Yao, Wenyi Zhao, Xinghan Liu, and 11 others. 2024. Autoglm: Autonomous foundation agents for guis. CoRR, abs/2411.00820.  
+Yuhang Liu, Pengxiang Li, Zishu Wei, Congkai Xie, Xueyu Hu, Xinchen Xu, Shengyu Zhang, Xiaotian Han, Hongxia Yang, and Fei Wu. 2025b. Infiguia-gent: A multimodal generalist GUI agent with native reasoning and reflection. CoRR, abs/2501.04575.  
+Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025c. Visual-rft: Visual reinforcement fine-tuning. CoRR, abs/2503.01785.  
+Yadong Lu, Jianwei Yang, Yelong Shen, and Ahmed Awadallah. 2024. Omniparser for pure vision based GUI agent. CoRR, abs/2408.00203.  
+Zhengxi Lu, Yuxiang Chai, Yaxuan Guo, Xi Yin, Liang Liu, Hao Wang, Han Xiao, Shuai Ren, Guanjing Xiong, and Hongsheng Li. 2025. Ui-r1: Enhancing efficient action prediction of gui agents by reinforcement learning. Preprint, arXiv:2503.21620.  
+Haohao Luo, Yang Deng, Ying Shen, See-Kiong Ng, and Tat-Seng Chua. 2024. Chain-of-exemplar: Enhancing distractor generation for multimodal educational question generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 7978-7993. Association for Computational Linguistics.  
+Run Luo, Lu Wang, Wanwei He, and Xiaobo Xia. 2025. Gui-r1: A generalist r1-style vision-language action model for gui agents. Preprint, arXiv:2504.10458.  
+OpenAI,., Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Madry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024. Gpt-4o system card. Preprint, arXiv:2410.21276.  
+OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.  
+Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, Wanjun Zhong, Kuanye Li, Jiale Yang, Yu Miao, Woyu Lin, Longxiang Liu, Xu Jiang, Qianli Ma, Jingyu Li, and 16 others. 2025. UI-TARS: pioneering automated GUI interaction with native agents. CoRR, abs/2501.12326.
+
+Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Zero: memory optimizations toward training trillion parameter models. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2020, Virtual Event / Atlanta, Georgia, USA, November 9-19, 2020, page 20. IEEE/ACM.  
+Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. CoRR, abs/2402.03300.  
+Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. 2025. Vlm-r1: A stable and generalizable r1-style large vision-language model. Preprint, arXiv:2504.07615.  
+Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. Preprint, arXiv:2403.05530.  
+Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. 2024a. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024.  
+Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. 2024b. Mobile-agent: Autonomous multi-modal mobile device agent with visual perception. CoRR, abs/2401.16158.  
+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. 2024c. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. CoRR, abs/2409.12191.  
+Xin Wang, Yudong Chen, and Wenwu Zhu. 2022. A survey on curriculum learning. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4555-4576.  
+Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, and Yu Qiao. 2024. Os-atlas: A foundation action model for generalist gui agents. Preprint, arXiv:2410.23218.  
+Yiheng Xu, Zekun Wang, Junli Wang, Dunjie Lu, Tianbao Xie, Amrita Saha, Doyen Sahoo, Tao Yu, and Caiming Xiong. 2024. Aguvis: Unified pure vision agents for autonomous gui interaction. Preprint, arXiv:2412.04454.
+
+Yuhao Yang, Yue Wang, Dongxu Li, Ziyang Luo, Bei Chen, Chao Huang, and Junnan Li. 2024. Ariaui: Visual grounding for gui instructions. Preprint, arXiv:2412.16256.  
+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 The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Chi Zhang, Zhao Yang, Jiaxuan Liu, Yanda Li, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. 2025. Appagent: Multimodal agents as smartphone users. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI 2025, Yokohama Japan, 26 April 2025-1 May 2025, pages 70:1-70:20. ACM.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand. Association for Computational Linguistics.
+
+# A Datasets Details
+
+We provide training data used in the curriculum learning stage and reinforcement learning stage, as well as statistical information for the four datasets used in the experiment, as shown in Table 8 and Table 7, respectively.
+
+# B Implementation Details
+
+We build our GUI agent upon Qwen2-VL-2B-Instruct (Wang et al., 2024c), a vision-language model with strong general capabilities. It is worth noting that we chose Qwen2-VL-2B instead of Qwen's latest series Qwen2.5-VL-3B (Bai et al., 2025) because 2B is lighter, and most of the methods we compared used Qwen2-VL-2B as the base model or larger parameter models before Qwen2-VL. We validated the effectiveness of our proposed method in a fairer situation. The training process consists of two stages: Curriculum Learning (CL) and Reinforcement Learning (RL). For Curriculum Learning, we employ the LLaMA Factory (Zheng et al., 2024) framework for one epoch at each stage with the visual backbone frozen. We employ AdamW with a cosine annealing learning rate scheduler, setting the initial learning rate to 1e-5 and batch size to 16. For Reinforcement Learning, we use the R1-V (Chen et al., 2025) framework for training over two epochs, setting the initial learning
+
+<table><tr><td>Datasets</td><td>Platforms</td><td>Task</td><td>Metric</td><td># Test Samples</td><td>History?</td></tr><tr><td>ScreenSpot-Pro (Li et al., 2025b)</td><td>Web, PC, Mobile</td><td>Grounding</td><td>clickAcc</td><td>1,272</td><td>X</td></tr><tr><td>ScreenSpot (Cheng et al., 2024)</td><td>Web, PC, Mobile</td><td>Grounding</td><td>clickAcc</td><td>1,581</td><td>X</td></tr><tr><td>Mind2Web (Deng et al., 2023)</td><td>Web</td><td>Navigation</td><td>Ele.Acc, Op.F1, Step SR</td><td>6,418</td><td>✓</td></tr><tr><td>AndroidControl-High (Li et al., 2024)</td><td>Mobile</td><td>Navigation</td><td>Step Accuracy</td><td>8,444</td><td>✓</td></tr><tr><td>AndroidControl-Low (Li et al., 2024)</td><td>Mobile</td><td>Grounding</td><td>Step Accuracy</td><td>8,444</td><td>X</td></tr></table>
+
+Table 7: Statistics of GUI datasets.  
+
+<table><tr><td>Usage</td><td>Source</td><td>Number</td></tr><tr><td rowspan="7">CL</td><td>GUIChat(Chen et al., 2024)</td><td>40k</td></tr><tr><td>ScreenQA (Hsiao et al., 2022)</td><td>17k</td></tr><tr><td>MultiUI (Liu et al., 2025a)</td><td>50k</td></tr><tr><td>AutoGUI (Li et al., 2025a)</td><td>50k</td></tr><tr><td>SeeClick (Cheng et al., 2024)</td><td>100k</td></tr><tr><td>GUIEnv (Chen et al., 2024)</td><td>100k</td></tr><tr><td>Aguvis-stage2 (Xu et al., 2024)</td><td>35k</td></tr><tr><td>RL</td><td>OmniAct (Kapoor et al., 2024)</td><td>2k</td></tr><tr><td>Total</td><td></td><td>394k</td></tr></table>
+
+Table 8: Training data statistics from various sources.
+
+rate to 1e-6 and batch size to 16. At each stage, we utilize the strategies of DeepSpeed optimization (Rajbhandari et al., 2020), BF16 format, and gradient checkpointing to save GPU memory. All training is conducted on  $2^{*}\mathrm{A}100$  GPU (80G).

2025/INREACT_ An Inspire-Then-Reinforce Training Framework For Multimodal GUI Agent/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4769bbb1eeab1d72c29813c6a19cfe9b93f9ec25d9784f766569e20510976a6f
+size 696503

2025/IP-Dialog_ Evaluating Implicit Personalization in Dialogue Systems with Synthetic Data/347614f0-d1c5-4371-901f-a53f1ec61429_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65815c3004079ea28d6f343a338b22658f15267252b74596460428d6b7073c79
+size 1340925

2025/IP-Dialog_ Evaluating Implicit Personalization in Dialogue Systems with Synthetic Data/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7e441077ca921f081cc12199a578f7fbfcfeefdab65712afccd08fb119c6b09
+size 1329625

2025/ISACL_ Internal State Analyzer for Copyrighted Training Data Leakage/8060ed99-c298-4105-b64b-1e8c92bcd4fe_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67eb6d806c28ec0810476739bf70a053a267f2f846a939573d760be888ea6fde
+size 4839946

2025/ISACL_ Internal State Analyzer for Copyrighted Training Data Leakage/full.md

@@ -0,0 +1,561 @@
+# ISACL: Internal State Analyzer for Copyrighted Training Data Leakage
+
+Guangwei Zhang $^{1}$ , Qisheng Su $^{2}$ , Jiateng Liu $^{3}$ , Cheng Qian $^{3}$ , Yanzhou Pan $^{4}$ , Yanjie Fu $^{5}$ , Denghui Zhang $^{*6}$
+
+<sup>1</sup>City University of Hong Kong, <sup>2</sup>Microsoft,
+
+<sup>3</sup>University of Illinois Urbana-Champaign, <sup>4</sup>Google LLC,
+
+$^{5}$ Arizona State University,  $^{6}$ Stevens Institute of Technology
+
+dzhang42@stevens.edu
+
+# Abstract
+
+Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) but pose risks of inadvertently exposing copyrighted or proprietary data, especially when such data is used for training but not intended for distribution. Traditional methods address these leaks only after content is generated, which can lead to the exposure of sensitive information. This study introduces a proactive approach: examining LLMs' internal states before text generation to detect potential leaks. By using a curated dataset of copyrighted materials, we trained a neural network classifier to identify risks, allowing for early intervention by stopping the generation process or altering outputs to prevent disclosure. Integrated with a Retrieval-Augmented Generation (RAG) system, this framework ensures adherence to copyright and licensing requirements while enhancing data privacy and ethical standards. Our results show that analyzing internal states effectively mitigates the risk of copyrighted data leakage, offering a scalable solution that fits smoothly into AI workflows, ensuring compliance with copyright regulations while maintaining high-quality text generation. The implementation is available<sup>1</sup>.
+
+# 1 Introduction
+
+LLMs have significantly enhanced text generation and dialogue systems in NLP (Zhang et al., 2023; Li et al., 2022a). However, they also pose risks of unintentionally reproducing copyrighted or proprietary information from their training data, especially when the data is licensed for training but not distribution. According to U.S. copyright law (U.S. Copyright Office, 1976), only the copyright holder has the exclusive right to distribute copyrighted works. If an LLM inadvertently distributes copyrighted material by replicating parts of its training
+
+![](images/feb2010cc8d4d3021c252add3e419f62298579bd5d5163ba62354c4148d6a9da.jpg)  
+Figure 1: To assess the risk of copyrighted training data leakage, an LLM's internal states are analyzed prior to generating content. Extracting semantic information from intermediate layers allows for the proactive identification of potential risks.
+
+data, it may violate this law and expose its developers or users to legal liability (Borkar, 2023). This underscores the importance of adhering to legal and ethical standards when deploying LLMs across various applications (Peng et al., 2023; Xue et al., 2021; Wang et al., 2025; Yu et al., 2023a, 2024). Addressing these risks is crucial to protecting intellectual property rights and ensuring the responsible and lawful use of LLMs in real-world situations.
+
+Previous research has raised concerns about the issue of copyrighted data leakage (Zhang et al., 2025; Pan et al.; Xu et al., 2024; Zhao et al., 2024) during the generation process of LLMs, including the leakage of private information (Kim et al., 2023; Lukas et al., 2023; Huang et al., 2022; Shao et al., 2024) and evaluation data used in machine learning (Zhou et al., 2025, 2023). Existing methods to prevent or mitigate data leakage include implementing strict output filtering (Miyaoka and Inoue, 2024) and context-aware mechanisms (Luu et al., 2024), applying differential privacy techniques (Li et al., 2025; Hoory et al., 2021; Du and Mi, 2021; Li et al., 2022b; Behnia et al., 2022; Shi et al., 2022;
+
+Wu et al., 2022; Majmudar et al., 2022; Du et al., 2023; Mai et al., 2024) or other data anonymization methods during training, regularly auditing and reviewing model outputs, and monitoring LLM interactions to detect potential data leakage. However, these approaches face several limitations, such as limited coverage of scenarios (Xiao et al., 2023), reduced model performance and usability caused by differential privacy techniques, and high costs and delays associated with manual audits (Song et al., 2024).
+
+To address the challenge of detecting copyrighted training data leakage in LLM-generated text, we introduce a framework called the Internal State Analyzer for Copyrighted training data Leakage (ISACL). ISACL evaluates leakage risks by analyzing the model's internal states during the prefill phase, before any text is generated. Unlike conventional approaches that rely on examining fully generated outputs, ISACL proactively identifies potential risks by assessing early-stage representations of input text and their correlation with copyrighted reference materials. This approach enables real-time, scalable, and precise risk assessment without requiring complete output generation. To enhance its effectiveness, ISACL is integrated into a RAG system. copyrighted information is indexed using FAISS and stored in SQLite, allowing efficient retrieval of relevant reference materials during the evaluation process. When a relevant reference is retrieved, it is combined with the model's internal states to determine the likelihood of leakage. This integration improves the accuracy and efficiency of comparing generated content with known copyrighted training data, ensuring reliable detection. Beyond detection, ISACL adheres to legal and ethical standards, serving as a robust safeguard against unauthorized disclosure of copyrighted materials in AI-generated content. By ensuring compliance with licensing constraints, ISACL promotes responsible and lawful use of LLMs in real-world applications.
+
+In a series of experimental configurations, ISACL demonstrated outstanding performance, achieving high accuracy and F1 scores. Specifically, accuracy ranged from  $91.88\%$  to  $95.05\%$ , while F1 scores varied between 0.9249 and 0.9468. In certain configurations, ISACL even achieved near-perfect detection rates. These results highlight ISACL's consistent ability to accurately identify potential training-set leakage across diverse settings, maintaining high levels of precision and recall. The
+
+findings underscore ISACL's robustness in scalable, real-time risk detection for LLM-generated content, even without generating any text. For a detailed description of the experimental setup and results, please refer to Section 4.3.
+
+Our primary contributions are as follows:
+
+- As illustrated in Figure 1, we propose a real-time framework "ISACL" for predicting copyrighted training data leakage in LLM-generated text by leveraging internal states extracted before any token is decoded. This ensures efficiency and avoids reliance on output generation, proactively addressing potential risks of unauthorized disclosure.  
+- ISACL is the first framework to proactively detect potential copyrighted data leakage by analyzing LLM internal states before content is generated. This approach ensures that neither users nor language models are exposed to sensitive or copyrighted information, thereby ensuring compliance with legal and licensing standards.  
+- We validate ISACL's effectiveness in large-scale text generation scenarios and demonstrate its integration with a RAG system. This integration enables efficient and accurate text retrieval while ensuring compliance with copyright constraints, making the approach suitable for industrial applications requiring real-time prevention of copyrighted data leakage.
+
+# 2 Related Work
+
+# 2.1 Internal States of LLMs
+
+Previous studies (Bricken et al., 2023; Templeton et al., 2024) have investigated the internal states of language models, which encode contextual and semantic information derived from their training data (Liu et al., 2023; Chen et al., 2024a; Gurnee and Tegmark, 2024; Wu et al., 2025). The applications of LLM internal states are highly diverse, including revealing hallucination risks (Ji et al., 2024), enhancing knowledge boundary perception (Ni et al., 2025), uncovering LLMs' factual discernment (He et al., 2024), and more (Wu et al., 2025).
+
+# 2.2 Copyright Issues with LLMs
+
+Scholars have emphasized the importance of protecting the intellectual property associated with the parameters of Large Language Models (Peng et al., 2023; Xue et al., 2021). This concern arises from the substantial investments in resources required for training LLMs, as well as the risk of unau
+
+![](images/751900be4447796f031a2cc0d35ac2abb0749b132917d94d249fe031ea985a87.jpg)  
+LEAKAGE DETECTION  
+Figure 2: Overview of our Copyrighted Training-set Leakage Detection Framework: Our approach involves maintaining a database of sensitive or proprietary materials to support the analysis of LLM hidden states. During inference, this database provides reference samples for potential leakage, working in conjunction with the model's hidden states to predict whether the generated content poses a risk of training-set leakage. The pipeline is structured into three key stages: The left section focuses on the construction and extraction of data for Retrieval-Augmented Generation, a core component designed to enhance model performance and address training-set leakage challenges. The right section illustrates the generation of training data, including the collection of internal states, labels, and reference embeddings, which are then used to train a Multi-Layer Perceptron as the final leakage risk detector. Lastly, the bottom section showcases real-world user interaction, where queries are submitted, and the system applies our framework to assess copyrighted potential leakage risks effectively.
+
+thorized exploitation of these models, which can have significant economic and ethical implications (Zhang et al., 2018; He et al., 2022; Dale, 2021).
+
+Copyright concerns are not limited to text; they span across various digital content creation formats, including scripts, images, videos (Moayeri et al., 2024; Kim et al., 2024), and code (Yu et al., 2023b). This widespread impact underscores the urgency of addressing these complex issues (Lucchi, 2023).
+
+# 2.3 Data Leakage in LLMs and Prevention Strategies
+
+LLMs are susceptible to data leakage due to several inherent vulnerabilities. One prominent issue is memorization during the training process, where LLMs unintentionally retain and reproduce sensitive information from their training data (Wang et al., 2024), such as personally identifiable information (PII) (Kim et al., 2023; Lukas et al., 2023;
+
+Huang et al., 2022; Shao et al., 2024). This memorization can expose models to privacy attacks, including membership inference (Maini et al., 2024; Galli et al., 2024; Feng et al., 2025) and training data extraction (Carlini et al., 2021). Another critical vulnerability stems from improper or incomplete output filtering, which may cause sensitive information to be disclosed in response to user queries (Zhang et al., 2024). Furthermore, misinterpretation of user queries by the model can inadvertently lead to the exposure of confidential data (Hu et al., 2024).
+
+# 3 Internal State Judge: Detecting Training-set Leakage Before Decoding
+
+# 3.1 Problem Formulation
+
+The issue of copyrighted training-set leakage in content generated by LLMs has attracted significant attention from both industry and academia.
+
+Existing approaches typically focus on detecting potential leakage only after the content has been generated. This post-generation evaluation method presents several challenges, including high computational costs, delays in mitigation, and legal risks associated with temporary exposure to leaked information.
+
+In this paper, we propose a framework (ISACL) designed to assess the risk of copyrighted training-set leakage before an LLM generates any output. The inference process of an LLM for a given query can be divided into two phases:
+
+(1) Prefix Phase: The LLM processes the entire input query to generate internal states.  
+(2) Decode Phase: The LLM generates output based on these prefilled internal states.
+
+This two-phase structure raises the central question of our study: Can the internal states produced during the prefill phase be used to predict the risk of copyrighted training-set leakage before the decoding phase begins?
+
+To address this question, we argue that the internal states generated by an LLM during the prefetch phase capture critical contextual information related to the likelihood of generating content that leaks copyrighted training-set data. We introduce an internal states judge designed to classify the risk of copyrighted training-set leakage based on the internal states from this phase.
+
+This approach offers three key advantages:
+
+- Efficiency: By evaluating internal states early in the prefetch phase, ISACL can halt decoding if the internal states judge identifies potential risks, reducing unnecessary computational costs.  
+- Proactive Risk Mitigation: Performing risk assessment before content generation enables preventive actions rather than reactive measures taken after leakage has occurred.  
+- Scalability: The internal states judge is designed to be adaptable across various open source LLM architectures and model sizes, supporting wide-scale deployment.
+
+The following sections describe the design of the internal state judge, the methodology for training data collection, and the experimental evaluation of ISACL.
+
+# 3.2 Training An Internal States Judge
+
+Training Data Preparation. We developed a dataset by selecting preceding and following sen
+
+tences from verified copyrighted material as the input  $x$  and reference  $t$ , respectively. The LLM is tasked with generating a continuation based on this input, resulting in the output  $y$ . This method is consistently applied to ensure uniformity throughout the process. Specifically, we construct a dataset of triplets for training the classifier: (x, y, t). Each generated output is assigned a risk label based on its similarity to the reference text, measured using the Rouge-L score:
+
+$$
+\mathcal {H} ^ {\text {t r a i n}} = \mathcal {T} (j, \text {R o u g e - L} (t, y)) \tag {1}
+$$
+
+where the threshold-based function  $\mathcal{T}$  determines risk labels, and  $j$  represents the partitioning criterion:
+
+$$
+\mathcal {T} (j, \text {R o u g e - L}) = \left\{ \begin{array}{l} 0, \text {i f} \mathcal {P} _ {2} \leq \text {R o u g e - L} \leq 1 \\ 1, \text {i f} 0 \leq \text {R o u g e - L} \leq \mathcal {P} _ {1} \\ \mathrm {N / A}, \text {o t h e r w i s e} \end{array} \right. \tag {2}
+$$
+
+where  $\mathcal{P}_1$  and  $\mathcal{P}_2$  are predefined thresholds used to classify an output as either high or low risk.
+
+Our dataset is structured as pairs of internal states and their associated risk labels:  $\mathcal{D}_{\theta} = \{\langle S_{x_i}^{\mathrm{train}},\mathcal{H}_i^{\mathrm{train}}\rangle \}_{i = 1}^N$ $\mathcal{S}$  denotes the internal states.
+
+Internal States of Query in Prefill Phase of LLMs. A crucial step in ISACL is the extraction of internal states during the prefetch phase of LLMs. In this phase, the model processes the entire input sequence to compute intermediate representations (such as keys and values) before generating any output tokens. This stage involves highly parallelized matrix-matrix operations, allowing the model to efficiently encode the semantic and structural properties of the input.
+
+During forward propagation, the input text  $x$  from the dataset triplet is fed into the LLM, and we extract the internal states  $S$  from a specific layer in the prefetch phase. These internal states are computed through multiple layers of non-linear transformations, activations, and information flow, formally represented as:
+
+$$
+\mathcal {S} _ {l} = f \left(\mathcal {W} _ {l} \cdot \mathcal {S} _ {l - 1} + \mathcal {B} _ {l}\right), \quad l = 1, 2, \dots , L \tag {3}
+$$
+
+where  $S_{l}$  represents the internal states at layer  $l$ ,  $\mathcal{W}_l$  and  $\mathcal{B}_l$  are the learnable weights and biases of the  $l$ -th layer, and  $f$  is the activation function. At each layer, the model refines its understanding of the input query  $x$ , progressively building increasingly sophisticated representations of syntax, context, and meaning (Devlin et al., 2019; Radford and
+
+Narasimhan, 2018). These internal states encode both token-level details and broader semantic relationships, providing a rich representation of the input's meaning (Clark et al., 2019).
+
+In our experiments, we extract internal states from the final encoder layer during the prefetch phase and compute their mean across all tokens. By analyzing these internal states before the decoding stage, we aim to proactively identify and mitigate potential risks (Zellers et al., 2020).
+
+Training Objectives of Internal States Judge. The objective of training the internal states judge is to create a classifier that predicts the likelihood of training-set leakage based on the internal states of the model. This classifier learns to assess the Rouge-L similarity score, distinguishing between high-risk and low-risk outputs. It is implemented using an MLP model:
+
+$$
+\mathcal {M} = \operatorname {d o w n} (\operatorname {u p} (\mathcal {S}) \times \operatorname {S i L U} (\operatorname {g a t e} (\mathcal {S}))) \tag {4}
+$$
+
+where SiLU serves as the activation function, and the linear layers down, up, and gate handle projection and gating mechanisms. This model enables efficient real-time risk prediction without requiring full output decoding.
+
+# 3.3 Enhancing Internal States Judge with Retrieved References
+
+Leveraging References to Enhance Internal States Judge. Relying solely on input text may lack sufficient context for detecting training-set leakage. To improve detection, ISACL incorporates external references using RAG technology (Lewis et al., 2021), enhancing the model's ability to assess potential risks.
+
+Formally, given an input query  $x$ , we first extract its internal states  $S_{x}$  from the prefill phase of the LLM, then retrieve a set of relevant reference texts  $T = \{t_{1}, t_{2}, \dots, t_{m}\}$  from an external knowledge base. The retrieved references are encoded into an aggregated representation  $S_{T}$ , which is then concatenated with  $S_{x}$  to form the final combined representation. An MLP classifier is then applied to predict the leak probability:
+
+$$
+p = \sigma \left(\mathcal {M} \left(\operatorname {c o n c a t} \left(f _ {\theta} (x), h _ {\phi} (\mathcal {G} (x))\right)\right)\right), \tag {5}
+$$
+
+where  $f_{\theta}$  represents the transformation function of the LLM's prefetch phase,  $\mathcal{G}$  is the retrieval function that selects references most relevant to  $x$ ,  $h_{\phi}$
+
+encodes the retrieved references,  $\mathcal{M}$  denotes the MLP model, and  $\sigma$  represents the sigmoid activation function that outputs the probability of training-set leakage.
+
+Finally, the predicted probability  $p$  is compared with a predefined threshold  $\tau$  to make the final leakage risk decision:
+
+$$
+\mathcal {H} ^ {\text {p r e d i c t}} = \left\{ \begin{array}{l l} 1, & \text {i f} p \geq \tau \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {6}
+$$
+
+where  $\tau$  is a tunable threshold that determines the sensitivity of leakage detection. By integrating external references into the internal state analysis and applying a threshold-based decision rule, this enhanced approach significantly improves the model's predictive capabilities, reducing both false positives and false negatives.
+
+Retrieving References from Indexed Documents. To facilitate Retrieval-Augmented Generation, as shown in Figure 3, we construct a RAG-Enhanced Reference Database that efficiently stores and retrieves references for leakage detection. This database is designed to manage copyrighted training materials effectively, ensuring quick access to relevant references and supporting robust content analysis and decision-making. The construction details of the RAG-based database are provided in Appendix B.
+
+![](images/0a5499f914e8871399dc5c10ab4ed009114e2dd5b37a757a4a43da5da959cdb6.jpg)  
+Figure 3: Process of constructing a vector database for the RAG system and handling user queries.
+
+# 4 Experiments
+
+In this section, we evaluate the effectiveness of the internal states judge in identifying literal copying leakage (according to strict text similarity) in text continuations. Specifically, we address the following research questions (RQ):
+
+- RQ 1: How well does ISACL detect literal copying leakage across various LLMs, such as the Llama and Mistral series, and how does model size influence performance?  
+- RQ 2: Can ISACL accurately identify non-literal copying leakage (such as paraphrased content) and how does its performance compare to that of literal copying leakage detection?  
+- RQ 3: What factors affect the performance of ISACL, including the role of the RAG system, the choice of LLM internal state layers, and the strategies used for dataset division?
+
+To investigate these questions, we conduct experiments using a structured dataset that includes both literal and non-literal copying leakage tasks. For literal copying, we evaluate the risk of training data leakage in text continuations by using excerpts from well-known fiction books. For non-literal copying, we focus on identifying event and character copying within paraphrased content. We test ISACL on LLMs from the Llama and Mistral series, ranging from 7B to 70B parameters, and compare it with baseline approaches. Our findings show that ISACL is both effective and accurate in detecting literal material leakage, while also revealing the challenges involved in identifying paraphrased content.
+
+# 4.1 Dataset and Label Partitioning
+
+We leveraged the dataset described in Appendix D to assess risks related to fiction texts (Meeus et al., 2024; Chang et al., 2023; Shi et al., 2023).
+
+In the process of dataset division, the label is created based on quantiles, where the upper  $p$  (the top  $p$  of the data) is used as the leakage set, and the lower  $p$  (the bottom  $p$  of the data) is used as the non-disclosure set. Notably, the data within the middle range of  $1 - 2p$  is directly discarded due to its relative ambiguity in classification. Here,  $p$ $(0 < p < 1)$  is a manually defined probability that determines the proportion of data included in each set, ensuring a clear distinction between the two subsets for analysis. By conducting experiments with varying  $p$  values, we can observe the sensitivity of internal states to the defined criteria for
+
+potential leakage risk.
+
+# 4.2 Model Selection
+
+We used LLMs from the Llama (Touvron et al., 2023) and Mistral (Jiang et al., 2023) series to generate text continuations and extract internal states, ensuring accurate dataset classification. To capture true continuations, we extracted reference embeddings using BERT (Devlin et al., 2019), which effectively captured the semantic content for training.
+
+# 4.3 Detecting Literal Copying Leakage through LLM Internal States
+
+In this section, we empirically evaluate the effectiveness of ISACL for detecting literal copying leakage across different LLMs, including Llama and Mistral, as well as a range of model sizes from 7B to 70B parameters. To assess model performance, we use standard metrics such as Accuracy and F1-score, described in appendix C, providing insights into the models' precision and effectiveness in detecting leakage risks. ISACL involves extracting internal states from the last layer of the model during the pre-filling phase, which are then used to train a classifier for predicting leakage risk.
+
+Baselines. In our experiment, we established a baseline model using LLMs to assess potential copyrighted material leakage in content generation tasks. It includes two configurations: "Input Only" (LLM-w/oRAG), where decisions are made based solely on the input text, and "Input with RAG system" (LLM-w/RAG), where both the input text and reference materials are considered. Similar to our proposed method, the baseline evaluates potential leakage without generating the next text segment. The task is to identify whether the continuation text contains elements that may raise leakage concerns. Predicted outcomes are compared to ground truth labels, which are derived from the dataset and based on Rouge-L scores. Details of the baseline prompt settings are provided in Table 10.
+
+Results and Analysis. The results are based on three dataset splits (select  $p$  according to Section 4.1), determined by Rouge-L scores:  $10\%$ ,  $20\%$ , and  $30\%$ . Each split classifies the dataset into high-scoring (leak) and low-scoring (non-disclosure) samples. We assess the model's ability to distinguish between these groups and examine how incorporating reference embeddings retrieved from a database enhances performance across various levels of textual similarity.
+
+Table 1: The results on the literal dataset evaluate the performance of various models and methods. We compare four approaches: LLM-w/oRAG and LLM-w/RAG, which represent the "LLM-as-a-Judge (Without RAG system)" and "LLM-as-a-Judge (With RAG system)" methods. In these approaches, we use the LLM directly to detect potential training data leakage in the input text—either based solely on the input (LLM-w/oRAG) or using both the input and the RAG system (LLM-w/RAG). Additionally, we evaluate the Internal-States-Judge (IS) methods: IS-w/oRAG and IS-w/RAG, which represent the "Internal-States-Judge (Without RAG system)" and "Internal-States-Judge (With RAG system)" methods. We report accuracy (ACC) and F1 scores for different dataset divisions.  
+
+<table><tr><td colspan="3"></td><td colspan="2">Division (10%)</td><td colspan="2">Division (20%)</td><td colspan="2">Division (30%)</td></tr><tr><td>LLMs</td><td>Method</td><td>Time (s)</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td></tr><tr><td colspan="9">Llama</td></tr><tr><td rowspan="4">Llama-3.1-8B</td><td>LLM-w/oRAG</td><td>0.4914</td><td>52.12</td><td>52.89</td><td>53.38</td><td>48.28</td><td>50.19</td><td>47.07</td></tr><tr><td>IS-w/oRAG</td><td>0.0564</td><td>91.53</td><td>92.96</td><td>78.05</td><td>79.25</td><td>73.73</td><td>77.36</td></tr><tr><td>LLM-w/RAG</td><td>0.7012</td><td>61.48</td><td>62.24</td><td>56.20</td><td>59.43</td><td>56.78</td><td>60.28</td></tr><tr><td>IS-w/RAG</td><td>0.0592</td><td>92.37</td><td>93.71</td><td>83.26</td><td>82.67</td><td>77.11</td><td>78.62</td></tr><tr><td rowspan="4">Llama-2-13b</td><td>LLM-w/oRAG</td><td>0.5412</td><td>63.29</td><td>53.82</td><td>58.26</td><td>49.42</td><td>53.28</td><td>52.43</td></tr><tr><td>IS-w/oRAG</td><td>0.0642</td><td>91.75</td><td>93.37</td><td>82.46</td><td>81.47</td><td>78.83</td><td>76.44</td></tr><tr><td>LLM-w/RAG</td><td>0.8109</td><td>63.75</td><td>62.97</td><td>61.43</td><td>58.41</td><td>59.52</td><td>54.78</td></tr><tr><td>IS-w/RAG</td><td>0.0696</td><td>93.23</td><td>94.18</td><td>86.52</td><td>85.57</td><td>80.03</td><td>79.15</td></tr><tr><td rowspan="4">Llama-3.1-70B</td><td>LLM-w/oRAG</td><td>1.1492</td><td>64.29</td><td>63.85</td><td>63.41</td><td>51.04</td><td>55.67</td><td>50.52</td></tr><tr><td>IS-w/oRAG</td><td>0.1274</td><td>\( 100.00^2 \)</td><td>100.00</td><td>94.55</td><td>94.63</td><td>91.88</td><td>92.49</td></tr><tr><td>LLM-w/RAG</td><td>1.4335</td><td>64.93</td><td>64.68</td><td>61.05</td><td>60.26</td><td>59.84</td><td>62.57</td></tr><tr><td>IS-w/RAG</td><td>0.1389</td><td>100.00</td><td>100.00</td><td>95.05</td><td>94.68</td><td>94.48</td><td>94.64</td></tr><tr><td colspan="9">Mistral</td></tr><tr><td rowspan="4">Mistral-7B-v0.1</td><td>LLM-w/oRAG</td><td>0.5238</td><td>54.31</td><td>51.92</td><td>50.73</td><td>49.96</td><td>50.85</td><td>51.55</td></tr><tr><td>IS-w/oRAG</td><td>0.0623</td><td>97.96</td><td>98.00</td><td>79.58</td><td>82.97</td><td>70.75</td><td>76.24</td></tr><tr><td>LLM-w/RAG</td><td>0.6876</td><td>58.49</td><td>54.51</td><td>55.58</td><td>52.40</td><td>52.36</td><td>53.77</td></tr><tr><td>IS-w/RAG</td><td>0.0677</td><td>98.98</td><td>98.99</td><td>83.25</td><td>85.59</td><td>78.01</td><td>82.35</td></tr><tr><td rowspan="4">Mistral-7B-v0.3</td><td>LLM-w/oRAG</td><td>0.5324</td><td>52.84</td><td>50.67</td><td>53.29</td><td>51.04</td><td>52.93</td><td>41.63</td></tr><tr><td>IS-w/oRAG</td><td>0.0597</td><td>91.75</td><td>92.59</td><td>83.52</td><td>84.21</td><td>79.46</td><td>83.04</td></tr><tr><td>LLM-w/RAG</td><td>0.6343</td><td>54.20</td><td>55.06</td><td>51.25</td><td>54.03</td><td>53.10</td><td>49.69</td></tr><tr><td>IS-w/RAG</td><td>0.0614</td><td>93.76</td><td>95.30</td><td>87.27</td><td>86.24</td><td>84.86</td><td>87.39</td></tr></table>
+
+We also compare ISACL to the "LLM-as-a-Judge" approach. As shown in Table 1, we analyze the performance differences across dataset splits and model configurations, demonstrating the practical advantages of ISACL.
+
+Several key insights emerge from the analysis. First, ISACL significantly improves efficiency. The pre-trained MLP-based binary classifier provides faster inference and better accuracy compared to the "LLM-as-a-Judge" method, which relies on direct LLM predictions. This indicates that ISACL is not only more efficient but also more precise in identifying potential leakage. Second, using original reference text retrieved from the database during training enhances accuracy, outperforming models that rely solely on LLM-extracted internal states. This highlights the importance of external reference material, which offers richer context and enables the model to more accurately detect potential leakage violations. Additionally, we observe that the performance of different LLMs varies. Larger Llama models are more sensitive to data leakage, suggesting that their increased size allows
+
+them to better capture subtle text similarities. In contrast, Llama and Mistral models show different capabilities in capturing textual nuances, which affects their effectiveness in this task. Finally, the dataset division strategy plays a key role. Larger Rouge score differences between high- and low-scoring samples make it easier for the model to differentiate between them. This emphasizes the importance of carefully selecting dataset splits, as they have a significant impact on the model's ability to accurately identify leakage risks.
+
+Effect of Internal States Layers. Unlike previous studies emphasizing the importance of later layers in LLMs for tasks like hallucination detection (Ji et al., 2024), our experiments on leakage detection show a different trend based on model size. For smaller models like Llama-3.1-8B, layer selection doesn't significantly affect the prediction of potential risk. However, for larger models such as Llama-3.1-70B, deeper layers significantly improve performance, especially in accuracy and F1 score.
+
+Previous research (Azaria and Mitchell, 2023)
+
+![](images/7f116fb6340134802bf0ac45c0a24da29d318baccf77a71f33aa5d70d510160c.jpg)  
+Figure 4: Impact of layer selection on leakage risk prediction: A comparative analysis across different layers in Llama models with 8B and 70B parameters. For smaller models (Llama-3.1-8B), the prediction performance is relatively consistent across layers, with minimal variation in accuracy and F1 score. For larger models (Llama-3.1-70B), deeper layers significantly enhance performance, capturing more nuanced semantic features and improving the prediction of potential leakage in text continuation tasks.
+
+emphasized the effectiveness of the final layer for hallucination detection, but our analysis indicates that for training data leakage risk prediction, deeper layers are more essential in larger models. As shown in Figure 4, deeper layers in larger models are better at capturing textual similarities to existing literary works, which is crucial for identifying potential leakage. In contrast, for smaller models, early and intermediate layers perform similarly to the final layer, suggesting that while semantic and contextual information is spread across all layers, deeper layers in larger models are more effective in detecting the finer details needed for accurate predictions.
+
+One possible explanation for this is that leakage detection requires identifying both local and global semantic patterns, which are essential for spotting similarities and potential plagiarism. In smaller models, these patterns are well-represented across various layers, whereas larger models excel in capturing the more subtle textual similarities through their deeper layers. Unlike hallucination detection, which focuses on long-range dependencies and uncertainty captured in later layers, leakage detection benefits from the ability of larger models to focus on detailed patterns across deeper layers.
+
+Variability in FN & FP Rates, but Stable Overall Accuracy & F1. To further analyze model performance, we selected four representative configurations and generated confusion matrix plots, as shown in Figure 5. These configurations combine
+
+two factors: the model (Llama-3.1-8B or Llama-3.1-70B) and whether a reference is included, with the Rouge-L Score  $30\%$  split strategy applied.
+
+It's important to note that the figures shown here represent a single instance from repeated experiments. Since the training and test sets are randomly split, some variability in the False Negative (FN) and False Positive (FP) rates is expected. However, despite this variability, we found that the overall prediction accuracy and F1 score remain consistently stable across different runs. This suggests that, while there are fluctuations in specific error types, the model's overall performance is reliable and robust.
+
+![](images/d203e491f22aaf02779801a2f1472eb65e968df0265b6870ab80b760abb04e2f.jpg)  
+(a) Llama-8B-w/oRAG
+
+![](images/167a2904acd33ece06dffe9fbc17bedd3d49f5b46ad6b9f770a259c6de66ab6b.jpg)  
+(b) Llama-8B-w/RAG
+
+![](images/693204cb30545100bf6ef9325f5847e6cd14876dff51ec9920a9a762bc8e4d0a.jpg)  
+(c) Llama-70B-w/oRAG  
+Figure 5: Confusion matrix plots showing the effect of model size and RAG system on prediction performance, with Llama-3.1-8B and Llama-3.1-70B models, both with and without reference information, using a Rouge-L  $30\%$  threshold for dataset splitting.
+
+![](images/1999e8ff0b25b3d7e74776f089b0bad555f894932c5a6d55063bfc60aecc9f11.jpg)  
+(d) Llama-70B-w/RAG
+
+Time Efficiency Comparison. We conducted experiments to compare the time efficiency of leakage prediction methods, and the results show that the proposed methods using internal states (ISw/oRAG and IS-w/RAG) are significantly faster than the traditional basic method. In the basic method, each input text is processed sequentially by the LLM to generate the next segment, which is then compared with the reference text to assess potential leakage. The majority of the time in this approach is spent on text generation, while the
+
+comparison step takes up very little time. As a result, the basic method is much slower, as indicated by its higher time values compared to the internal states-based methods. These methods streamline the process, eliminating the need for text generation and leading to faster, more efficient predictions. The detailed results of this comparison are shown in Table 2.
+
+Table 2: This table shows the average time efficiency comparison (in seconds) for leakage prediction based on a single data point, testing three methods: predicting leakage risk using internal states without (ISw/oRAG) and with (IS-w/RAG) RAG system, and the basic method of generating continuation text and comparing it with reference text.  
+
+<table><tr><td>Method
+Model</td><td>Basic</td><td>IS-w/oRAG</td><td>IS-w/RAG</td></tr><tr><td>Llama-3.1-8B</td><td>0.4319</td><td>0.0564</td><td>0.0592</td></tr><tr><td>Llama-2-13b</td><td>0.6584</td><td>0.0642</td><td>0.0696</td></tr><tr><td>Llama-3.1-70B</td><td>1.6796</td><td>0.1274</td><td>0.1389</td></tr><tr><td>Mistral-7B-v0.1</td><td>0.3571</td><td>0.0623</td><td>0.0677</td></tr><tr><td>Mistral-7B-v0.3</td><td>0.3463</td><td>0.0597</td><td>0.0614</td></tr></table>
+
+Evaluation on Newly Collected Copyrighted Texts. To further verify whether our method can generalize to copyrighted datasets unseen by ISACL, we conducted an additional experiment using newly published books in 2025. Since these works were published after the knowledge cutoff (Dec. 2023) of the backbone model Llama-3.1-70B (Touvron et al., 2023), we can reasonably assume that they were not included in its training corpus. Following the same workflow as in our main experiments, we constructed input-reference pairs from the selected texts, applied continuation tasks with Llama-3.1-70B, and computed ROUGE-L scores between the model outputs and the ground-truth continuations. Results are summarized in Table 3.
+
+<table><tr><td>Metric</td><td>Value</td></tr><tr><td>Average ROUGE-L Score</td><td>0.11</td></tr><tr><td>Samples with ROUGE-L &lt; 0.15</td><td>&gt;90%</td></tr><tr><td>Samples with ROUGE-L &gt; 0.20</td><td>0%</td></tr></table>
+
+Table 3: Results on newly collected copyrighted texts.
+
+We observed that the ROUGE-L distribution on
+
+these unseen works was consistently low, with the majority of samples falling below 0.15 and none exceeding 0.20. This indicates that the probability of memorization or leakage from copyrighted materials is negligible in this setting. Moreover, the absence of sufficient high-ROUGE matches prevents us from creating meaningful positive/negative labels for classifier training. These findings confirm that our task design specifically addresses the detection of copyrighted training data leakage, rather than focusing on performance with unseen texts.
+
+# 5 Conclusion and Future Work
+
+This study introduces ISACL, a framework designed to detect copyrighted training data leakage in LLM-generated text by analyzing internal states during the prefill phase, before any text is generated. Unlike traditional methods that analyze fully generated outputs, ISACL enables proactive, real-time detection by examining early-stage representations of input text in relation to copyrighted reference materials. Experiments with models like Llama and Mistral show that larger models achieve higher accuracy due to richer internal representations.
+
+To enhance its effectiveness, ISACL is integrated into a RAG system, using FAISS for vector search and SQLite for structured storage. This integration allows efficient retrieval of relevant copyrighted materials and combines them with the model's internal states to assess leakage risks, ensuring compliance with licensing constraints while improving detection accuracy and efficiency.
+
+Future work will focus on addressing more complex forms of copyright leakage, such as conceptual similarity and paraphrasing, and refining the framework for better robustness and interpretability. Additionally, we aim to develop an LLM agent that actively prevents leakage by cross-referencing generated content against licensed or publicly available materials, ensuring real-time compliance with data usage policies.
+
+# Limitations
+
+Despite its advantages, ISACL has some limitations. Detection accuracy in smaller models requires improvement, as these models often have less nuanced internal representations, which can affect reliability. Moreover, this study focuses mainly on assessing the ability of LLM internal states to identify copyrighted training-set leakage, but more
+
+precise criteria for determining leakage are needed for practical applications. In particular, clearer standards are required to address complex cases like conceptual similarity or paraphrasing.
+
+# Ethics Statement
+
+We all comply with the ACL Ethics Policy<sup>3</sup> during our study. All datasets used contain anonymized consumer data, ensuring strict privacy protections.
+
+# References
+
+Amos Azaria and Tom Mitchell. 2023. The internal state of an llm knows when it's lying. Preprint, arXiv:2304.13734.  
+Rouzbeh Behnia, Mohammadreza Reza Ebrahimi, Jason Pacheco, and Balaji Padmanabhan. 2022. Ew-tune: A framework for privately fine-tuning large language models with differential privacy. In 2022 IEEE International Conference on Data Mining Workshops (ICDMW), pages 560-566.  
+Jaydeep Borkar. 2023. What can we learn from data leakage and unlearning for law? Preprint, arXiv:2307.10476.  
+Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E. Burke, Tristan Hume, Shan Carter, Tom Henighan, and Christopher Olah. 2023. Towards monoseismicity: Decomposing language models with dictionary learning. Transformer Circuits Thread.  
+Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, Alina Oprea, and Colin Raffel. 2021. Extracting training data from large language models. Preprint, arXiv:2012.07805.  
+Kent Chang, Mackenzie Cramer, Sandeep Soni, and David Bamman. 2023. Speak, memory: An archaeology of books known to ChatGPT/GPT-4. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7312-7327, Singapore. Association for Computational Linguistics.  
+Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024a. Inside: Llms' internal states retain the power of hallucination detection. Preprint, arXiv:2402.03744.  
+Tong Chen, Akari Asai, Niloofar Mireshghallah, Sewon Min, James Grimmelmann, Yejin Choi, Hannaneh Hajishirzi, Luke Zettlemoyer, and Pang Wei Koh. 2024b. Copybench: Measuring literal and non-literal reproduction of copyright-protected text in language model generation. Preprint, arXiv:2407.07087.  
+Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics.  
+Robert Dale. 2021. Gpt-3: What's it good for? Natural Language Engineering, 27(1):113-118.
+
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The faiss library. Preprint, arXiv:2401.08281.  
+Jian Du and Haitao Mi. 2021. Dp-fp: Differentially private forward propagation for large models. Preprint, arXiv:2112.14430.  
+Minxin Du, Xiang Yue, Sherman S. M. Chow, Tianhao Wang, Chenyu Huang, and Huan Sun. 2023. Dpforward: Fine-tuning and inference on language models with differential privacy in forward pass. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, CCS '23, page 2665-2679. ACM.  
+Qizhang Feng, Siva Rajesh Kasa, Santhosh Kumar Kasa, Hyokun Yun, Choon Hui Teo, and Sravan Babu Bodapati. 2025. Exposing privacy gaps: Membership inference attack on preference data for llm alignment. Preprint, arXiv:2407.06443.  
+Filippo Galli, Luca Melis, and Tommaso Cucinotta. 2024. Noisy neighbors: Efficient membership inference attacks against llms. Preprint, arXiv:2406.16565.  
+Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.  
+Wes Gurnee and Max Tegmark. 2024. Language models represent space and time. Preprint, arXiv:2310.02207.  
+Jinwen He, Yujia Gong, Zijin Lin, Cheng'an Wei, Yue Zhao, and Kai Chen. 2024. LLM factoscope: Uncovering LLMs' factual discernment through measuring inner states. In *Findings of the Association for Computational Linguistics: ACL* 2024, pages 10218-10230, Bangkok, Thailand. Association for Computational Linguistics.  
+Xuanli He, Qiongkai Xu, Lingjuan Lyu, Fangzhao Wu, and Chenguang Wang. 2022. Protecting intellectual property of language generation apis with lexical watermark. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10758-10766.  
+Shlomo Hoory, Amir Feder, Avichai Tendler, Sofia Erell, Alon Peled-Cohen, Itay Laish, Hootan Nakhost, Uri
+
+Stemmer, Ayelet Benjamini, Avinatan Hassidim, and Yossi Matias. 2021. Learning and evaluating a differentially private pre-trained language model. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 1178-1189, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Siying Hu, Piaohong Wang, Yaxing Yao, and Zhicong Lu. 2024. "i always felt that something was wrong." Understanding compliance risks and mitigation strategies when professionals use large language models. Preprint, arXiv:2411.04576.  
+Jie Huang, Hanyin Shao, and Kevin Chen-Chuan Chang. 2022. Are large pre-trained language models leaking your personal information? Preprint, arXiv:2205.12628.  
+Ziwei Ji, Delong Chen, Etsuko Ishii, Samuel Cahyawijaya, Yejin Bang, Bryan Wilie, and Pascale Fung. 2024. Llm internal states reveal hallucination risk faced with a query. arXiv preprint arXiv:2407.03282.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. Preprint, arXiv:2310.06825.  
+Minseon Kim, Hyomin Lee, Boqing Gong, Huishuai Zhang, and Sung Ju Hwang. 2024. Automatic jailbreaking of the text-to-image generative ai systems. arXiv preprint arXiv:2405.16567.  
+Siwon Kim, Sangwoo Yun, Hwaran Lee, Martin Gubri, Sungroh Yoon, and Seong Joon Oh. 2023. Propile: Probing privacy leakage in large language models. Preprint, arXiv:2307.01881.  
+Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. Retrieval-augmented generation for knowledge-intensive nlp tasks. Preprint, arXiv:2005.11401.  
+Junyi Li, Tianyi Tang, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2022a. Pretrained language models for text generation: A survey. *Preprint*, arXiv:2201.05273.  
+Xuechen Li, Florian Tramér, Percy Liang, and Tatsunori Hashimoto. 2022b. Large language models can be strong differentially private learners. *Preprint*, arXiv:2110.05679.  
+Yansong Li, Zhixing Tan, and Yang Liu. 2025. Privacy-preserving prompt tuning for large language model services. Preprint, arXiv:2305.06212.
+
+Kevin Liu, Stephen Casper, Dylan Hadfield-Menell, and Jacob Andreas. 2023. Cognitive dissonance: Why do language model outputs disagree with internal representations of truthfulness? Preprint, arXiv:2312.03729.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. Preprint, arXiv:1907.11692.  
+Nicola Lucchi. 2023. Chatgpt: a case study on copyright challenges for generative artificial intelligence systems. European Journal of Risk Regulation, pages 1-23.  
+Nils Lukas, Ahmed Salem, Robert Sim, Shruti Tople, Lukas Wutschitz, and Santiago Zanella-Béguelin. 2023. Analyzing leakage of personally identifiable information in language models. Preprint, arXiv:2302.00539.  
+Quan Khanh Luu, Xiyu Deng, Anh Van Ho, and Yorie Nakahira. 2024. Context-aware llm-based safe control against latent risks. Preprint, arXiv:2403.11863.  
+Peihua Mai, Ran Yan, Zhe Huang, Youjia Yang, and Yan Pang. 2024. Split-and-denoise: Protect large language model inference with local differential privacy. Preprint, arXiv:2310.09130.  
+Pratyush Maini, Hengrui Jia, Nicolas Papernot, and Adam Dziedzic. 2024. Llm dataset inference: Did you train on my dataset? Preprint, arXiv:2406.06443.  
+Jimit Majmudar, Christophe Dupuy, Charith Peris, Sami Smaili, Rahul Gupta, and Richard Zemel. 2022. Differentially private decoding in large language models. Preprint, arXiv:2205.13621.  
+Matthieu Meeus, Igor Shilov, Manuel Faysse, and Yves-Alexandre de Montjoye. 2024. Copyright traps for large language models. arXiv preprint arXiv:2402.09363.  
+Yuya Miyaoka and Masaki Inoue. 2024. Cbflm: Safe control for llm alignment. Preprint, arXiv:2408.15625.  
+Mazda Moayeri, Samyadeep Basu, Sriram Balasubramanian, Priyatham Kattakinda, Atoosa Chengini, Robert Brauneis, and Soheil Feizi. 2024. Rethinking artistic copyright infringements in the era of text-to-image generative models. arXiv preprint arXiv:2404.08030.  
+Shiyu Ni, Keping Bi, Jiafeng Guo, Lulu Yu, Baolong Bi, and Xueqi Cheng. 2025. Towards fully exploiting llm internal states to enhance knowledge boundary perception. Preprint, arXiv:2502.11677.  
+Yanzhou Pan, Jiayi Chen, Jiamin Chen, Zhaozhuo Xu, and Denghui Zhang. Position: Iterative online-offline joint optimization is needed to manage complex llm
+
+copyright risks. In Forty-second International Conference on Machine Learning Position Paper Track.  
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. Preprint, arXiv:1912.01703.  
+Wenjun Peng, Jingwei Yi, Fangzhao Wu, Shangxi Wu, Bin Zhu, Lingjuan Lyu, Binxing Jiao, Tong Xu, Guangzhong Sun, and Xing Xie. 2023. Are you copying my model? protecting the copyright of large language models for eaas via backdoor watermark. arXiv preprint arXiv:2305.10036.  
+Alec Radford and Karthik Narasimhan. 2018. Improving language understanding by generative pretraining.  
+Hanyin Shao, Jie Huang, Shen Zheng, and Kevin Chang. 2024. Quantifying association capabilities of large language models and its implications on privacy leakage. In *Findings of the Association for Computational Linguistics: EACL* 2024, pages 814-825, St. Julian's, Malta. Association for Computational Linguistics.  
+Weijia Shi, Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu, Terra Blevins, Danqi Chen, and Luke Zettlemoyer. 2023. Detecting pretraining data from large language models. arXiv preprint arXiv:2310.16789.  
+Weiyan Shi, Ryan Shea, Si Chen, Chiyuan Zhang, Ruoxi Jia, and Zhou Yu. 2022. Just fine-tune twice: Selective differential privacy for large language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6327-6340, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Chengyu Song, Linru Ma, Jianming Zheng, Jinzhi Liao, Hongyu Kuang, and Lin Yang. 2024. Audit-llm: Multi-agent collaboration for log-based insider threat detection. Preprint, arXiv:2408.08902.  
+Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L. Turner, Callum McDougall, Monte MacDiarmid, C. Daniel Freeman, Theodore R. Sumers, Edward Rees, Joshua Batson, Adam Jermyn, Shan Carter, Chris Olah, and Tom Henighan. 2024. Scaling monoseismicity: Extracting interpretable features from claude 3 sonnet. Transformer Circuits Thread.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard
+
+Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. Preprint, arXiv:2302.13971.  
+U.S. Copyright Office. 1976. Copyright law of the united states (title 17).  
+Jeffrey G. Wang, Jason Wang, Marvin Li, and Seth Neel. 2024. Pandora's white-box: Precise training data detection and extraction in large language models. Preprint, arXiv:2402.17012.  
+Rushi Wang, Jiateng Liu, Weijie Zhao, Shenglan Li, and Denghui Zhang. 2025. Automating financial statement audits with large language models. arXiv preprint arXiv:2506.17282.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Huggingface's transformers: State-of-the-art natural language processing. Preprint, arXiv:1910.03771.  
+Xinwei Wu, Li Gong, and Deyi Xiong. 2022. Adaptive differential privacy for language model training. In Proceedings of the First Workshop on Federated Learning for Natural Language Processing (FL4NLP 2022), pages 21-26, Dublin, Ireland. Association for Computational Linguistics.  
+Yuheng Wu, Wentao Guo, Zirui Liu, Heng Ji, Zhaozhuo Xu, and Denghui Zhang. 2025. How large language models encode theory-of-mind: a study on sparse parameter patterns. npj Artificial Intelligence, 1(1):20.  
+Guangxuan Xiao, Ji Lin, and Song Han. 2023. Offsite-tuning: Transfer learning without full model. Preprint, arXiv:2302.04870.  
+Jialiang Xu, Shenglan Li, Zhaozhuo Xu, and Denghui Zhang. 2024. Do llms know to respect copyright notice? arXiv preprint arXiv:2411.01136.  
+Mingfu Xue, Yushu Zhang, Jian Wang, and Weiqiang Liu. 2021. Intellectual property protection for deep learning models: Taxonomy, methods, attacks, and evaluations. IEEE Transactions on Artificial Intelligence, 3(6):908-923.  
+Yangyang Yu, Haohang Li, Zhi Chen, Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu, Jordan W Suchow, and Khaldoun Khashanah. 2023a. Finmem: A performance-enhanced large language model trading agent with layered memory and character design. In ICLR (workshop on LLM agent).  
+Yangyang Yu, Zhiyuan Yao, Haohang Li, Zhiyang Deng, Yupeng Cao, Zhi Chen, Jordan W Suchow, Rong Liu, Zhenyu Cui, Denghui Zhang, et al. 2024. Fincon: A synthesized llm multi-agent system with conceptual verbal reinforcement for enhanced financial decision making. In Proceedings of NeurIPS 2024 (Main).
+
+Zhiyuan Yu, Yuhao Wu, Ning Zhang, Chenguang Wang, Yevgeniy Vorobeychik, and Chaowei Xiao. 2023b. Codeipprompt: Intellectual property infringement assessment of code language models. In International Conference on Machine Learning, pages 40373-40389. PMLR.  
+Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2020. Defending against neural fake news. Preprint, arXiv:1905.12616.  
+Denghui Zhang, Zhaozhuo Xu, and Weijie Zhao. 2025. Lims and copyright risks: Benchmarks and mitigation approaches. In Proceedings of the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 5: Tutorial Abstracts), pages 44-50.  
+Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. 2023. A survey of controllable text generation using transformer-based pre-trained language models. Preprint, arXiv:2201.05337.  
+Jialong Zhang, Zhongshu Gu, Jiyong Jang, Hui Wu, Marc Ph Stoecklin, Heqing Huang, and Ian Molloy. 2018. Protecting intellectual property of deep neural networks with watermarking. In Proceedings of the 2018 on Asia conference on computer and communications security, pages 159-172.  
+Zhiping Zhang, Michelle Jia, Hao-Ping (Hank) Lee, Bingsheng Yao, Sauvik Das, Ada Lerner, Dakuo Wang, and Tianshi Li. 2024. "it's a fair game", or is it? examining how users navigate disclosure risks and benefits when using lvm-based conversational agents. In Proceedings of the CHI Conference on Human Factors in Computing Systems, CHI '24, page 1-26. ACM.  
+Weijie Zhao, Huajie Shao, Zhaozhuo Xu, Suzhen Duan, and Denghui Zhang. 2024. Measuring copyright risks of large language model via partial information probing. arXiv preprint arXiv:2409.13831.  
+Kun Zhou, Yutao Zhu, Zhipeng Chen, Wentong Chen, Wayne Xin Zhao, Xu Chen, Yankai Lin, Ji-Rong Wen, and Jiawei Han. 2023. Don't make your llm an evaluation benchmark cheater. Preprint, arXiv:2311.01964.  
+Xin Zhou, Martin Weyssow, Ratnadira Widyasari, Ting Zhang, Junda He, Yunbo Lyu, Jianming Chang, Beiqi Zhang, Dan Huang, and David Lo. 2025. Lessleak-bench: A first investigation of data leakage in llms across 83 software engineering benchmarks. Preprint, arXiv:2502.06215.
+
+# A Implementation Details
+
+The input dimension of our classifier is defined by the number of features in the training dataset, ensuring that the model can properly process the input data. The hidden dimension is fixed at 256, a value
+
+that aligns with the design of our models and supports effective learning. We train our classifier with the following settings and hyper-parameters: the epoch is 250, the batch size is 4, the learning rate is 1e-3, and the AdamW optimizer has a linear scheduler. We conduct all the experiments using Pytorch (Paszke et al., 2019) and HuggingFace library (Wolf et al., 2020) on 4 NVIDIA A100-SXM4-80GB GPUs.
+
+# B RAG System Construction
+
+Data Preparation. To establish a comprehensive retrieval system, we use datasets representing both leakage and non-disclosure cases. Each dataset consists of input-reference text pairs  $(x,t)$  where the input text  $x$  acts as a query, and the reference text  $t$  provides contextual information, meaning the surrounding content in a specific context, such as the following text in a classic work. The entire dataset is stored as a structured collection:  $\mathcal{D} = \{(x_i,t_i)\}_{i = 1}^N$ , where  $N$  is the total number of pairs in the dataset. By merging multiple datasets into a unified pool, we ensure broad coverage of potential scenarios, forming a strong foundation for benchmarking and future improvements.
+
+Dense Representation Encoding. To capture the semantic relationships between input and reference texts, we encode each text into a dense vector representation using a pre-trained Sentence Transformer  $\mathcal{E}$  (all-roberta-large-v1) (Liu et al., 2019):  $v_{x} = \mathcal{E}(x)$ ,  $v_{t} = \mathcal{E}(t)$ , where  $v_{x}, v_{t} \in \mathbb{R}^{d}$  are the dense embeddings of the input query and the reference text, respectively, and  $d$  is the embedding dimension. To enhance efficiency, we implement batch encoding with GPU acceleration, ensuring scalable processing of large datasets while maintaining retrieval accuracy.
+
+Indexing with FAISS & Document Storage in SQLite. For efficient nearest-neighbor retrieval, we use FAISS (Douze et al., 2024) with the Index-IVFFlat method, which clusters the vector space to accelerate query execution. Given a set of indexed reference embeddings  $\{v_{t_i}\}_{i = 1}^N$ , FAISS partitions them into  $K$  clusters, with each vector assigned to its nearest cluster center:
+
+$$
+\mathcal {C} = \{\mu_ {k} \} _ {k = 1} ^ {K}, \quad \mu_ {k} = \frac {1}{| C _ {k} |} \sum_ {v \in C _ {k}} v,
+$$
+
+where  $\mathcal{C}$  is the set of centroids and  $C_k$  is the set of embeddings in cluster  $k$ . During retrieval, a query
+
+embedding  $v_{x}$  is assigned to the closest centroid  $\mu_{k}$ , and the nearest neighbors are searched within that cluster:  $\hat{t} = \mathrm{argmin}_{t_i\in C_k}\| v_x - v_{t_i}\| _2$ . This reduces search complexity from  $O(N)$  to  $O(N / K)$ , ensuring fast retrieval even for large datasets.
+
+Additionally, we use SQLite for structured text storage, where each document entry (including original input and reference texts) is indexed with its corresponding embedding. This allows efficient retrieval of both vector embeddings and textual data based on semantic similarity and exact text matches:  $\mathcal{T} = \{(x_i,t_i,v_{t_i})\}_{i = 1}^N$
+
+Retrieval Accuracy Since our input and reference pairs are stored in the external knowledge base as structured pairs, our retrieval method achieves a  $100\%$  accuracy rate in search matching within the current dataset:
+
+$$
+\operatorname * {a r g m a x} _ {t _ {i}} \operatorname {S i m} (v _ {x}, v _ {t _ {i}}) = t _ {j}, \text {w h e r e} (x, t _ {j}) \in \mathcal {D}.
+$$
+
+Here,  $\operatorname{Sim}(\cdot, \cdot)$  denotes the similarity function (e.g., cosine similarity), ensuring that the retrieved reference always corresponds to the correct pair in our dataset. By integrating dense vector retrieval with structured text storage, ISACL provides efficient and accurate reference retrieval, forming a crucial component of our leakage detection system.
+
+# C Metric Details
+
+ACC & F1. For the classification task where the predictions are discrete, we use F1 score and Accuracy as the metrics to assess the performance of the predicted categories.
+
+In classification tasks, accuracy and F1 score are two important metrics used to evaluate the performance of a model. Accuracy represents the proportion of correctly classified instances among the total number of instances, providing a general measure of how often the model makes the right prediction. It is calculated as:
+
+$$
+\mathcal {A} = \frac {\mathcal {T} _ {p} + \mathcal {T} _ {n}}{\mathcal {N} _ {\text {t o t a l}}} \tag {7}
+$$
+
+where  $\mathcal{T}_p$  and  $\mathcal{T}_n$  represent true positives and true negatives, respectively, and  $\mathcal{N}_{\mathrm{total}}$  is the total number of samples. Accuracy is simple and intuitive but may be unreliable with imbalanced datasets, where one class dominates the others. A model predicting only the majority class can achieve high accuracy but fail to detect minority instances.
+
+The F1 score provides a more balanced evaluation by considering both precision and recall. Precision  $(\mathcal{P})$  is the fraction of correctly predicted positive observations out of all positive predictions, while recall  $(\mathcal{R})$  is the fraction of true positives out of all actual positive samples. The F1 score is defined as:
+
+$$
+\mathcal {F} _ {1} = 2 \times \frac {\mathcal {P} \times \mathcal {R}}{\mathcal {P} + \mathcal {R}} \tag {8}
+$$
+
+The F1 score is particularly useful in imbalanced datasets, balancing false positives and false negatives to provide a comprehensive view of performance. While accuracy works well for balanced data, the F1 score is more informative for assessing real-world classification problems.
+
+ROUGE. ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a set of metrics commonly used to evaluate the quality of automatic text summarization and natural language generation systems by comparing the overlap between generated text and reference text. ROUGE includes several variations: Rouge-N evaluates the overlap of N-grams, Rouge-L focuses on the longest common subsequence (LCS), and Rouge-S uses skip-bigram matching. Among them, Rouge-L measures sequence similarity by identifying the longest common subsequence between the generated text and the reference text, capturing both content and sequential structure. The Rouge-L score comprises Precision, Recall, and F-score, representing different perspectives of text similarity, where Recall emphasizes content coverage, and Precision reflects matching accuracy.
+
+In our experiments, we chose ROUGE as the evaluation method and used the rouge_score library to calculate the Rouge-1 and Rouge-L scores, focusing on using the Rouge-L score as a key metric. We justify this choice over other common evaluation methods like BLEU and embedding-based metrics for the following reasons:
+
+- Advantages over BLEU: ROUGE is better suited to our experimental needs. Its foundation on the longest common subsequence (LCS) allows for more flexible matching, making it superior for evaluating the coverage and overall structure of text summaries. It can better capture the content similarity and sequential relationships between the generated text and the reference text. In contrast, BLEU's emphasis on strict n-gram and word order matching, while ideal for evalu
+
+ating grammatical correctness in machine translation, may not fully reflect the structural and content coverage required in our task.
+
+- Advantages over Embedding Similarity: For the hypotheses and experimental setup of our paper, ROUGE-L also serves as a more direct and precise metric for measuring text similarity compared to methods like embedding cosine similarity. This is because embedding-based metrics have several limitations: their effectiveness is highly dependent on the quality of the embedding model, which may not capture semantic information accurately for our specific domain. Furthermore, most embedding-based methods are insensitive to word order, which can lead to the loss of critical contextual information. In contrast, ROUGE-L directly evaluates the overlap of sequences and inherently considers word order, making it a more reliable and interpretable metric for our specific task.
+
+Based on these considerations and our experimental goals, ROUGE can more accurately evaluate the sequential similarity and content coverage of text pairs. Therefore, we fixed the evaluation method to ROUGE and used the Rouge-L score as the core metric to evaluate and classify the quality of text pairs in the dataset.
+
+# D Dataset
+
+We provide the sources of copyrighted material in Table 7, confirmed as part of our selected models' training data (Chen et al., 2024b; Gao et al., 2020; Touvron et al., 2023; Jiang et al., 2023). For the literal copying task, which evaluates the risk of training data leakage in text continuations, we included excerpts from 16 fiction titles in BookMIA (Shi et al., 2023). To enhance diversity, we added works by J.K. Rowling. For the non-literal copying task, focusing on event and character replication, we used CliffsNotes study guides alongside human-written summaries. To ensure all texts are under copyright, we excluded non-fiction and books published before 1923.
+
+# E Prompt Design
+
+In designing the baseline for our experiment on detecting training data leakage risks through internal states, we adopted the "LLM as Judge" approach. This method leverages LLMs to evaluate potential leakage risks in text generation tasks. To ensure robust and accurate assessment, we carefully crafted
+
+evaluation prompts tailored to capture nuanced scenarios of potential risk, as shown in Table 10. This design allows for a systematic comparison between traditional heuristic-based methods and our proposed internal state detection framework.
+
+# F Ablation Studies
+
+# F.1 Effect of Model Size
+
+This section investigates how model size influences the efficacy of LLM's internal states in classifier training, comparing Llama models with 1B, 3B, 8B, 13B, and 70B parameters. Experimental results demonstrate that smaller Llama models generate internal states that yield lower F1 scores and accuracy in classification tasks compared to larger models, regardless of whether the input data is presented in isolation or supplemented with reference information provided by RAG system. As shown in Figure 6, the performance of ISACL improves significantly with increasing size, highlighting the importance of model scale in enhancing classification accuracy and F1 scores.
+
+![](images/c57be0bfa62415db2becef74f1be2e05abd46c49c24386ebc0591baf5898c724.jpg)  
+Figure 6: Impact of model size on behavior prediction performance: a comparative analysis of classification accuracy and F1 scores across Llama models with 1B to 70B parameters
+
+To address the behavioral variations arising from differences in internal state quality and data generation strategies across models of varying sizes, it is essential to design separate, model-specific databases. These databases should capture the unique characteristics of the internal states and outputs generated by each model size. For smaller models, stricter control over Rouge-based segmentation thresholds may be necessary to achieve clearer distinctions between potentially leakage and non-disclosure data. Such measures are particularly important because smaller models tend to produce less semantically rich internal states, potentially diminishing classification accuracy.
+
+By refining the dataset segmentation strategy—particularly for smaller models—the accuracy of leakage risk predictions can be significantly improved. This ensures that even resource-constrained models are well-prepared for robust downstream classification tasks, enabling reliable performance across diverse use cases.
+
+# F.2 Effect of Generation Prompts
+
+In this section, we discuss the impact of varying prompt design strategies used as input to the LLM on the prediction accuracy of the trained model during the dataset construction process. Building on the prompt configurations from prior work (Chen et al., 2024b), we modify them as the sole variable in our experiments. Table 4 presents the results of these experiments, highlighting how different prompt formulations influence the overall performance. The prompt design is presented in Table 8 for clarity and reference.
+
+As shown in this table, the design corresponding to Prompt 2 exhibits relatively lower performance compared to the designs associated with Prompt 1 and Prompt 3. Both the IS-w/oRAG and IS-w/RAG methods yield weaker results under this configuration, with ACC and F1 scores declining as the dataset division percentage increases. In conclusion, variations in each prompt used for data generation have a noticeable impact on the prediction accuracy of models trained with the resulting datasets. Therefore, when predicting leakage risks, multiple models utilizing datasets generated with different prompt designs can be employed. By applying this approach, it becomes possible to identify and prioritize data associated with higher leakage risk, enhancing the effectiveness of the risk detection process.
+
+# F.3 Effect of Internal States Extraction Methods
+
+In our experiments, we examined the impact of different internal state extraction methods at a given layer for copyrighted leakage detection, specifically comparing the effectiveness of using the average internal state across all tokens versus extracting only the internal state of the last token. Our results indicate that, for a fixed layer, computing the mean internal state across all tokens provides significantly higher prediction accuracy than relying solely on the internal state of the last token, as shown in Table 5.
+
+Table 4: The table illustrates how prompt selection affects text generation by comparing F1 scores and accuracy across different prompts used in preparing the training dataset for the Llama-3.1-70B model. It evaluates two methods: IS-w/oRAG (Internal States Judge without the RAG system) and IS-w/RAG (Internal States Judge with the RAG system).  
+
+<table><tr><td colspan="2"></td><td colspan="2">Division (10%)</td><td colspan="2">Division (20%)</td><td colspan="2">Division (30%)</td></tr><tr><td>Prompt</td><td>Method</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td></tr><tr><td rowspan="2">Prompt1</td><td>IS-w/oRAG</td><td>97.01</td><td>96.00</td><td>88.79</td><td>87.43</td><td>85.24</td><td>88.07</td></tr><tr><td>IS-w/RAG</td><td>97.34</td><td>95.13</td><td>90.57</td><td>89.34</td><td>87.29</td><td>89.94</td></tr><tr><td rowspan="2">Prompt2</td><td>IS-w/oRAG</td><td>85.71</td><td>89.50</td><td>75.12</td><td>79.52</td><td>67.55</td><td>75.25</td></tr><tr><td>IS-w/RAG</td><td>91.73</td><td>93.17</td><td>89.27</td><td>89.42</td><td>73.84</td><td>75.06</td></tr><tr><td rowspan="2">Prompt3</td><td>IS-w/oRAG</td><td>91.41</td><td>93.33</td><td>74.51</td><td>79.22</td><td>62.54</td><td>71.29</td></tr><tr><td>IS-w/RAG</td><td>98.44</td><td>98.73</td><td>87.75</td><td>88.29</td><td>70.03</td><td>75.53</td></tr></table>
+
+Table 5: This table explores the effectiveness of different internal state extraction methods under the Llama-3.1-70B model. The results show that, at a fixed layer, averaging the internal states across all tokens significantly outperforms using only the last token's internal state, as the averaging method better captures contextual information, making it more suitable for detection.  
+
+<table><tr><td></td><td colspan="2">Division (10%)</td><td colspan="2">Division (20%)</td><td colspan="2">Division (30%)</td></tr><tr><td>Methods</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td></tr><tr><td>Last Token-w/oRAG</td><td>68.57</td><td>75.56</td><td>66.83</td><td>74.33</td><td>62.99</td><td>72.46</td></tr><tr><td>Last Layer-w/oRAG</td><td>100.00</td><td>100.00</td><td>94.55</td><td>94.63</td><td>93.18</td><td>93.62</td></tr><tr><td>Last Token-w/RAG</td><td>88.57</td><td>89.09</td><td>88.61</td><td>88.78</td><td>83.77</td><td>85.47</td></tr><tr><td>Last Layer-w/RAG</td><td>100.00</td><td>100.00</td><td>95.05</td><td>94.68</td><td>94.48</td><td>94.64</td></tr></table>
+
+When taking the average internal state, the representation is aggregated across all token embeddings within the selected layer. This method ensures that the extracted feature captures a comprehensive understanding of the entire sequence, incorporating both local token-level details and global contextual relationships. As a result, this approach is particularly effective for leakage detection, where recognizing semantic and structural similarities across a text is crucial.
+
+Conversely, extracting the last token's internal state from the same layer restricts the representation to a single token position, potentially losing valuable contextual information present in the earlier tokens. While this method is commonly used in classification tasks, our analysis shows that, in leakage risk prediction, it leads to a weaker overall representation, as the key signals indicating similarity to existing works may be distributed throughout the sequence rather than concentrated in the final token.
+
+These findings highlight that, even when working with the same layer, the choice of how internal states are extracted plays a crucial role in model performance. Averaging across all tokens allows for a more robust and contextually rich representation,
+
+making it a preferable choice for copyrighted leakage detection. Future studies could further explore whether weighting token contributions or applying attention-based pooling strategies can further refine the effectiveness of internal state-based detection methods.
+
+# F.4 Non-literal Copying Leakage Detection
+
+In this section, we examine copyrighted leakage detection for non-literal paraphrasing (Chen et al., 2024b). We measure the overlap between generated and reference texts at the character and event levels to assess potential leakage. This approach is similar to the literal copying leakage task, but in the non-literal case, the continuation is based on paraphrasing instead of direct copying leakage. As shown in Table 6, we evaluate prediction accuracy across three prompt types, detailed in Table 9.
+
+Despite the smaller dataset, the results show that detecting copyrighted leakage in paraphrased texts is more challenging for large language models than in literal data. This leads to lower prediction accuracy in non-literal paraphrasing, as paraphrased texts are harder to compare directly with the reference text due to structural, vocabulary, and expression differences. This complexity reduces the
+
+Table 6: The experiment utilizes non-literal data with the training set divided based on the upper and lower  $30\%$  of Rouge scores. "C" denotes character-related copying leakage while "E" represents event-related copying leakage. Additionally, test results are extracted from the internal states of Llama-3.1-70B.  
+
+<table><tr><td></td><td colspan="2">Prompt 1</td><td colspan="2">Prompt 2</td><td colspan="2">Prompt 3</td></tr><tr><td>Method</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td><td>ACC (%)</td><td>F1 (%)</td></tr><tr><td>IS-w/oRAG</td><td>53.33</td><td>57.89</td><td>46.67</td><td>54.72</td><td>51.11</td><td>62.30</td></tr><tr><td>IS-w/RAG-C</td><td>63.33</td><td>70.27</td><td>56.67</td><td>41.67</td><td>56.67</td><td>31.58</td></tr><tr><td>IS-w/RAG-E</td><td>55.56</td><td>65.60</td><td>52.22</td><td>58.93</td><td>55.56</td><td>64.29</td></tr></table>
+
+model's ability to generalize, resulting in lower classification performance. Even with additional reference information by using RAG system, the model struggles to capture the intricate features required for accurate prediction.
+
+Table 7: List of Book Titles and Authors for literal task  
+
+<table><tr><td>Title</td><td>Author</td></tr><tr><td>1984</td><td>George Orwell</td></tr><tr><td>A Game of Thrones</td><td>George R.R. Martin</td></tr><tr><td>Casino Royale</td><td>Ian Fleming</td></tr><tr><td>Dune</td><td>Frank Herbert</td></tr><tr><td>Fahrenheit 451</td><td>Ray Bradbury</td></tr><tr><td>Fifty Shades of Grey</td><td>E.L. James</td></tr><tr><td>Five on a Treasure Island</td><td>Enid Blyton</td></tr><tr><td>Harry Potter and the Sorcerer&#x27;s Stone</td><td>J.K. Rowling</td></tr><tr><td>Hitchhiker&#x27;s Guide to the Galaxy</td><td>Douglas Adams</td></tr><tr><td>Lord of the Flies</td><td>William Golding</td></tr><tr><td>The Da Vinci Code</td><td>Dan Brown</td></tr><tr><td>The Hunger Games</td><td>Suzanne Collins</td></tr><tr><td>The Silmarillion</td><td>J.R.R. Tolkien</td></tr><tr><td>Their Eyes Were Watching God</td><td>Zora Neale Hurston</td></tr><tr><td>Things Fall Apart</td><td>Chinua Achebe</td></tr><tr><td>To Kill a Mockingbird</td><td>Harper Lee</td></tr><tr><td>Harry Potter and the Philosopher&#x27;s Stone</td><td>J.K. Rowling</td></tr><tr><td>Harry Potter and the Chamber of Secrets</td><td>J.K. Rowling</td></tr><tr><td>Harry Potter and the Prisoner of Azkaban</td><td>J.K. Rowling</td></tr><tr><td>The Hobbit</td><td>J.R.R. Tolkien</td></tr></table>
+
+Table 8: Three prompt templates for generating passage completion to evaluate literal copying leakage.(Chen et al., 2024b)
+
+# Prompt 1
+
+```txt
+You will be shown a series of passages from famous literary works. After these  $\rightarrow$  examples, you will receive a prefix from another passage and be asked  $\rightarrow$  to complete it based on the text of a famous work. Provide only the  $\rightarrow$  continuation for the last given prefix without any extra commentary,  $\rightarrow$  formatting, or additional text.   
+Complete the prefix: {demonstration-input} {demonstration-output}   
+Complete the prefix: {input}
+```
+
+# Prompt 2
+
+```txt
+You will be given several passage examples from famous literary works. Each  $\rightarrow$  example includes a prefix and its completion. After these examples, you  $\rightarrow$  will be provided with the final prefix, which you need to complete in  $\rightarrow$  the style of the famous work. Your completion should be at least 100  $\rightarrow$  words long and formatted as follows: "Completion: {your completion to  $\rightarrow$  the prefix}" Generate only the completion for the last given prefix  $\rightarrow$  without copying the examples, the given prefixes, or adding extra  $\rightarrow$  commentary.   
+Beginning: {demonstration-input}   
+Completion: {demonstration-output}   
+Beginning: {input}   
+Completion:
+```
+
+# Prompt 3
+
+```txt
+You will see examples of passages from renowned literature, each with a  $\rightarrow$  starting phrase (prefix) followed by a continuation (completion). After  $\rightarrow$  reviewing these examples, you will be given a final prefix to complete  $\rightarrow$  . Ensure your continuation is at least 100 words long and formatted as:  $\rightarrow$  "Completion:{your completion to the prefix}". Do not include any of  $\rightarrow$  the provided examples, prefixes, or additional commentary in your  $\rightarrow$  response.   
+Given Prefix: {demonstration-input}   
+Provided Completion: {demonstration-output}   
+Given Prefix: {input}   
+Completion:
+```
+
+Table 9: Three prompt templates for generating creative story to evaluate non-literal copying leakage.(Chen et al., 2024b)
+
+# Prompt 1
+
+```txt
+You will be provided a series of demonstrations of creative writings. After the  $\hookrightarrow$  demonstrations, you will be provided a beginning of a story and asked  $\hookrightarrow$  to compose an original story based on it. Give only the completion of  $\hookrightarrow$  the story, and no extra commentary, formatting, or chattiness.
+```
+
+```txt
+Continue the text to compose an original story about 500 words in length: {demonstration-input} {demonstration-output}   
+Continue the text to compose an original story about 500 words in length: {input}   
+Complete the prefix: {demonstration-input} {demonstration-output}   
+Complete the prefix: {input}
+```
+
+# Prompt 2
+
+```txt
+You will be provided with a series of creative writing examples. Each example  $\leftrightarrow$  includes a story beginning and its completion. After these  $\leftrightarrow$  demonstrations, you will be given the last beginning, which you are  $\leftrightarrow$  asked to complete. You should compose an original story of about 500  $\leftrightarrow$  words. The story should be distinct from any existing copyrighted novels  $\leftrightarrow$  . Your output should be in the format: "Completion: your completion to  $\leftrightarrow$  the beginning". Generate only the completion to the last given beginning  $\leftrightarrow$  without copying the examples, the given beginnings, or adding extra  $\leftrightarrow$  commentary.
+```
+
+```txt
+Prefix: {demonstration-input}  
+Completion: {demonstration-output}  
+Prefix: {input}  
+Completion:
+```
+
+# Prompt 3
+
+```txt
+You will be provided with a series of creative writing examples. Each example  $\leftrightarrow$  includes a story beginning and its completion. After these  $\leftrightarrow$  demonstrations, you will be given the last beginning, which you are  $\leftrightarrow$  asked to complete. You should compose an original story of about 500 words that is distinct from any existing copyrighted novels. An original  $\leftrightarrow$  story should have unique characters, settings, plots, and dialogues  $\leftrightarrow$  that are not taken from existing works. Avoid using specific phrases, sequences of events, or character traits that directly mirror those found in copyrighted material. Your output should be in the format: "  $\leftrightarrow$  Completion: your completion to the beginning". Generate only the completion to the last given beginning without copying the examples, the given beginnings, or adding extra commentary.
+```
+
+```txt
+Given Prefix: {demonstration-input}  
+Provided Completion: {demonstration-output}  
+Given Prefix: {input}  
+Completion:
+```
+
+Table 10: Baseline Evaluation Prompts for LLM-as-Judge Framework  
+Version A: Input-Only Assessment  
+```txt
+"role": "user", "content": f"" Please only output 'label: 0' or 'label: 1', nothing else. Do not  $\rightarrow$  provide intermediate steps or explanations; simply output the  $\rightarrow$  generated label. You are a language model that determines whether the next generated  $\rightarrow$  sentence based on the given input text might cause potential  $\rightarrow$  copyrighted training data leakage. If you believe the next sentence could significantly resemble or repeat  $\rightarrow$  existing literary works, indicating possible leakage, return "  $\rightarrow$  label: 0" (possible leakage). If you believe the next sentence will not show high similarity to any  $\rightarrow$  literary work, return "label: 1" (no leakage). Here is the input to evaluate: Input text: '\{input_text\}' Based on the input text provided, determine if generating the next  $\rightarrow$  sentence might result in leakage.
+```
+
+Version B: Reference-Aware Assessment  
+```txt
+{ "role": "user", "content": f"" Please only output 'label: 0' or 'label: 1', nothing else. Do not  $\leftrightarrow$  provide intermediate steps or explanations; simply output the  $\leftrightarrow$  generated label. You are a language model that evaluates whether the output text  $\leftrightarrow$  potentially involves copyrighted training data leakage based on  $\leftrightarrow$  the relationship between the reference and output sequences. If you believe the next sentence could significantly resemble or repeat  $\leftrightarrow$  existing literary works, indicating possible leakage, return "  $\leftrightarrow$  label: 0" (possible leakage). If you believe the next sentence will not show high similarity to any  $\leftrightarrow$  literary work, return "label: 1" (no leakage). Here are the inputs to evaluate: Input text: '\{input_text\}' Reference: '\{reference\}' (This is a segment from the original source  $\leftrightarrow$  text corresponding to the input text.) Based on the comparison between the reference and output above,  $\leftrightarrow$  determine if generating the next sentence might result in  $\leftrightarrow$  leakage.
+```

2025/ISACL_ Internal State Analyzer for Copyrighted Training Data Leakage/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed0afcdb0ddafb48cf56e7e10c21759fb48bc7aa2d22963e547b265c4b3da1e3
+size 775479

2025/Identifying Aspects in Peer Reviews/6e98954a-cad9-4f3b-8b14-95738b229b2e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf265b0578ef645482b3717ce69a05b1ce588167ec131dc49820370304fa34d3
+size 4075395

2025/Identifying Aspects in Peer Reviews/full.md

@@ -0,0 +1,720 @@
+# Identifying Aspects in Peer Reviews
+
+LU Sheng, Ilia Kuznetsov, Iryna Gurevych
+
+Ubiquitous Knowledge Processing Lab (UKP Lab)
+
+Department of Computer Science and Hessian Center for AI (hessian.AI)
+
+Technical University of Darmstadt
+
+www.ukp.tu-darmstadt.de
+
+# Abstract
+
+Peer review is central to academic publishing, but the growing volume of submissions is straining the process. This motivates the development of computational approaches to support peer review. While each review is tailored to a specific paper, reviewers often make assessments according to certain aspects such as Novelty, which reflect the values of the research community. This alignment creates opportunities for standardizing the reviewing process, improving quality control, and enabling computational support. While prior work has demonstrated the potential of aspect analysis for peer review assistance, the notion of aspect remains poorly formalized. Existing approaches often derive aspects from review forms and guidelines, yet data-driven methods for aspect identification are underexplored. To address this gap, our work takes a bottom-up approach: we propose an operational definition of aspect and develop a data-driven schema for deriving aspects from a corpus of peer reviews. We introduce a dataset of peer reviews augmented with aspects and show how it can be used for community-level review analysis. We further show how the choice of aspects can impact downstream applications, such as LLM-generated review detection. Our results lay a foundation for a principled and data-driven investigation of review aspects, and pave the path for new applications of NLP to support peer review. $^{1}$
+
+# 1 Introduction
+
+Peer review is an essential part of academic publishing. It is a complex, multifaceted process that requires a range of competencies including paper understanding, domain-specific knowledge, and critical thinking (Shah, 2022; Yuan et al., 2022). Ensuring review quality, especially among novice reviewers, is an open challenge (Stelmakh et al., 2021a,b; Sun et al., 2024a). The increasing volume
+
+of publications puts further strain on the process, which motivates the development of computational approaches to support different stages of peer review, from reading the paper to the final decision-making by the program committees (Arous et al., 2021; Checco et al., 2021; Stelmakh et al., 2021a; Shah, 2022; Schulz et al., 2022; Yuan et al., 2022; Lin et al., 2023b; Kuznetsov et al., 2024).
+
+Peer reviews are the central component of the reviewing process. While individual reviews can vary widely, reviewers within the same community tend to focus on a specific set of general quality categories, or aspects, such as Clarity and Novelty. These aspects can be found in review forms, instructional materials, guidelines, and the resulting review texts. Aspects allow comparison of submissions across different dimensions of quality. A comprehensive set of aspects shared among reviewers is critical for ensuring reviewing quality and consistency, and prior work has demonstrated the potential of aspect-based tools to support the review writing process (Sun et al., 2024a,b).
+
+Yet, several open questions remain. First, what is an aspect? While most prior work derives aspects from review forms and guidelines, the lack of an operational definition of aspect prevents the comparison of aspect schemata across studies. Second, what aspect granularity is appropriate for different tasks? While prior work operates with top-down, coarse-grained aspect schemata derived from review forms and guidelines, it is unclear whether they are comprehensive or provide sufficient granularity for NLP applications. Third, how can fine-grained aspect analysis support peer review? While aspects have been applied to certain tasks, the tasks that require a higher level of granularity are underexplored.
+
+To address these questions, this work introduces an alternative, data-driven approach to peer review aspect analysis. We propose an operational definition of aspect grounded in its role within the
+
+evaluation process. We develop a semi-automatic approach that leverages a state-of-the-art large language model (LLM) to identify aspects in reviews, and apply it to a large collection of peer reviews to extract aspects in a bottom-up fashion. Building on these results, we develop a multi-level taxonomy of aspects and present a novel dataset of peer reviews augmented with their corresponding aspects. Our dataset facilitates the exploration of two tasks: predicting the review aspects that should be focused on given a paper (paper aspect prediction), and identifying the aspects that a review focuses on (review aspect prediction). Based on these tasks, we conduct a detailed empirical study of aspect at the community level. Furthermore, we demonstrate that a comprehensive, fine-grained set of aspects allows for a new dimension in comparing reviews, offers a nuanced assessment of review quality in terms of specificity, and can be used for detection of automatically generated reviews.
+
+Our work paves the path for data-driven analysis of aspects and enables new NLP applications to support peer review. Our method offers a bottom-up perspective that complements existing top-down schemata, and can be integrated with them by, for example, allowing domain experts to refine LLM-derived aspects into higher-level, domain-specific categories. Together, these perspectives enable a comprehensive aspect-based analysis of peer reviews. To summarize, we contribute the following:
+
+- We propose an semi-automatic approach to derive a comprehensive set of aspects from peer reviews in a bottom-up fashion.  
+- We develop a taxonomy of aspect with different granularity and a new dataset of peer reviews augmented with aspects.  
+- We evaluate models on two tasks: predicting aspects to be focused on given a paper, and identifying the aspects covered in a review.  
+- We show that finer-grained aspect analysis offers new insights into and support for the peer review process.
+
+# 2 Related Work
+
+# 2.1 Peer review in the era of LLMs
+
+NLP for peer review is an emerging research area that aims to support different stages of the peer review process, including improving paper-reviewer matching, increasing reviewing efficiency and reproducibility, tracking dishonest behavior, and
+
+more (Shah, 2022; Schulz et al., 2022; Biswas et al., 2023; Lin et al., 2023b; Kuznetsov et al., 2024).
+
+The emergence of LLMs has opened new opportunities for assisting peer review, such as assisting in the verification of checklists and supporting review writing (Liu and Shah, 2023; Gao et al., 2024; Liang et al., 2024b; Jin et al., 2024). In this work, we leverage an LLM to identify aspects in peer reviews, building on evidence that state-of-the-art LLMs are shown to perform well in aspect identification (Lin et al., 2023a), and in summarization, a related task to aspect identification (Pu et al., 2023).
+
+# 2.2 Aspect in peer review
+
+Beyond their role in review forms and guidelines, aspects are used in several other contexts related to peer review. Aspects have been applied to analyze the sentiment of reviews and discover the discourse relations within reviews (Chakraborty et al., 2020; Kennard et al., 2022). In automatic review generation, Wang et al. (2020) generate reviews using aspect-based knowledge graphs, and Gao et al. (2024) prompt LLMs with aspect-based questions to generate reviews. Aspects are included in supporting systems for review writing, which have been shown to improve the comprehensiveness of the reviews written by reviewers with varying levels of experience (Sun et al., 2024a,b). Table 1 summarizes the aspects used in these studies.
+
+ACL'16, Chakraborty et al. (2020), Wang et al. (2020)
+
+Appropriateness, Clarity, Empirical/Theoretical Soundness, Impact of Ideas/Results/Dataset, Meaningful Comparison, Originality, Recommendation, Substance
+
+ACL'18, Kennard et al. (2022), Yuan et al. (2022),
+
+Sun et al. (2024b)
+
+Clarity, Impact/Motivation, Meaningful Comparison, Originality, Replicability, Soundness/Correctness, Substance
+
+Sun et al. (2024a)
+
+Clarity, Importance, Novelty, Validity
+
+Wang et al. (2024)
+
+Clarity, Integrity, Novelty, Significance
+
+Table 1: The aspects used in the studies cited in this section. Gao et al. (2024) use a model to generate aspect-based questions without using a pre-defined aspect set.
+
+In some other review guidelines, such as ARR and ACL'23, aspects are listed as examples rather than parts of a comprehensive checklist. As noted by Kuznetsov et al. (2024), review guidelines for major NLP venues often rely on coarse-grained aspects and they lack comprehensiveness. Our work
+
+advances the study of aspects by exploring an alternative, data-driven approach to deriving finer-grained and more comprehensive aspects.
+
+# 2.3 Quality of review writing
+
+Defining the desiderata for review writing is essential to assess its effectiveness (Jefferson et al., 2002). However, the desiderata are often not well defined or operationalizable in terms of automatic measurement (Kuznetsov et al., 2024). While using simple proxies such as "helpfulness" may seem like a straightforward way to assess review quality, prior work suggests that such evaluation can be biased in several ways – for example, an evaluator may be biased towards longer reviews, or those that recommend acceptance of their papers (Wang et al., 2021; Goldberg et al., 2023).
+
+Comprehensiveness is a key desideratum for high quality reviews (Yuan et al., 2022). A frequently reported issue in ACL 2023 is the lack of specificity in reviews (Rogers et al., 2023). In this context, our work shows how fine-grained aspects can be used to compare reviews and evaluate their specificity, which contributes to a nuanced and practical assessment of review quality.
+
+# 2.4 LLM-generated review detection
+
+The strong capabilities of state-of-the-art LLMs in text generation have led to the need for detectors to identify LLM-generated contents and to prevent potential misuse (Clark et al., 2021; Gao et al., 2023; Wu et al., 2023). In peer review, LLMs pose a risk that reviewers may exploit LLMs to produce reviews entirely, which raises serious ethical concerns (Yu et al., 2024). Current strategies for detecting LLM-generated reviews align with methods for detecting LLM-generated texts: (a) LLM-as-a-judge, which prompts LLMs to identify LLM-generated contents (Zheng et al., 2023), (b) detection models, which are fine-tuned on both human and LLM-generated texts (Guo et al., 2023), (c) reference-based methods, which compare the similarity between a candidate text and one generated by an LLM (Gehrmann et al., 2019; Ippolito et al., 2020; Liang et al., 2024a; Yu et al., 2024). In this work, we provide an alternative approach, and demonstrate a comprehensive, fine-grained aspect set helps the detection of LLM-generated reviews.
+
+# 3 Aspect set construction
+
+# 3.1 Definition of aspect
+
+While aspects are outlined in review guidelines and used in related work, a formal definition of aspect is lacking. To address this, we propose an operational definition of aspect as a characteristic of a paper that a reviewer makes a judgment on when evaluating the paper, which is later used to compare the manuscripts to each other and to the quality standards provided by the review guidelines. From this definition, we do not consider terms such as Acceptance Decision as aspects, since they are not characteristics of a paper. An aspect can be general, such as Soundness, or specific, such as Missing Citations on Controlled Generation. However, aspects that are either too general or too specific are difficult to use for comparing manuscripts to each other or to the publication standards. Therefore, an aspect taxonomy that accommodates different levels of granularity is needed.
+
+We assume that an aspect is expressed in a review sentence, and we allow multiple aspects per sentence. We treat the aspects of a review as a set. Formally, for a paper  $p$ , let  $R_{p}$  denote the set of reviews for  $p$ . For each review  $r \in R_{p}$ , we derive  $A_{r}$ , the set of aspects in  $r$ .
+
+# 3.2 Method
+
+To establish a comprehensive set of aspects, we selected reviews from NLP and machine learning (ML) conferences across different time periods. We randomly selected 50 papers from each of the NLPeer (Dycke et al., 2023) and EMNLP23² datasets. We used the keywords "natural" and "language" to filter NLP-related papers from ICLR. We randomly selected 50 NLP-related papers from each of the ICLR conferences from 2020 to 2024. We selected 350 papers in total, corresponding to 1094 reviews. We segmented the reviews into sentences using Punkt (Bird and Loper, 2004) and performed the identification at the sentence level.
+
+We used OpenAI GPT-4o $^3$  to identify aspects from the reviews. The prompt we used is shown in Table 7. Since the identification was performed in an unsupervised setting, the identified aspects appear inconsistent, with variations like Result and Results. We post-processed the results to reduce such variations. We first identified the most frequently occurring aspects in the results, which were
+
+<table><tr><td>COARSE</td><td>FINE</td><td>LLM annotation</td></tr><tr><td>Contribution</td><td>Contribution</td><td>Community Contribution, Methodology Contribution</td></tr><tr><td>DDDDEI</td><td>Definition, Description, Detail, Discussion, Explanation, Interpretation</td><td>Dataset Description, Missing Details, Task Definition</td></tr><tr><td>IJMV</td><td>Intuition, Justification, Motivation, Validation</td><td>Approach Justification, Dataset Validity, Model Intuition</td></tr><tr><td>Novelty</td><td>Innovation/Novelty/Originality</td><td>Algorithmic Innovations, Technical Novelty</td></tr><tr><td>Presentation</td><td>Clarity, Figure, Grammar, Presentation, Tipo</td><td>Dataset Clarification, Paper Presentation, Term Clarity</td></tr><tr><td>Related Work</td><td>Citation/Literature/Related Work</td><td>Existing Literature, Missing Citations, Previous Works</td></tr><tr><td>Significance</td><td>Impact, Importance, Significance</td><td>Empirical Importance, Practical Significance</td></tr></table>
+
+(a) paper-agnostic  
+
+<table><tr><td>COARSE</td><td>FINE</td><td>LLM annotation</td></tr><tr><td>Ablation</td><td>Ablation</td><td>Ablation Analysis, Ablation Study, Ablation Tests</td></tr><tr><td>Analysis</td><td>Analysis</td><td>Ablation Analysis, Complexity Analysis, Data Analysis</td></tr><tr><td>Comparison</td><td>Comparison</td><td>Comparison Fairness, Comparison to SOTA</td></tr><tr><td>Data/Task</td><td>Annotation, Benchmark, Data, Task</td><td>Annotation Detail, Alternative Tasks, Data Preparation</td></tr><tr><td>Evaluation</td><td>Evaluation, Metric</td><td>Accuracy Metric, Evaluation Scheme, Human Evaluation</td></tr><tr><td>Experiment</td><td>Experiment</td><td>Control Experiment, Experimental Procedure</td></tr><tr><td>Methodology</td><td>Algorithm, Implementation, Method</td><td>Language Model, Methodological Soundness</td></tr><tr><td>Theory</td><td>Theory</td><td>Lack of Theoretical Guarantee, Theoretical Correctness</td></tr><tr><td>Result</td><td>Findings, Improvement, Performance, Result</td><td>BLEU Improvement, Statistical Test</td></tr></table>
+
+(b) paper-dependent
+
+Table 2: The taxonomy of aspects. See here for the complete taxonomy.
+
+used as keywords to categorize and match the remaining ones. We grouped related terms together, such as Clarification and Clarity. We removed terms that we considered to be too general, such as Weakness, Strength, Question, and Comment. For the terms that cannot be matched, we omitted those that appear less than 50 times in the annotations.
+
+We note that our post-processing method may lead to unrelated terms being grouped together. For example, using Improvement as a keyword groups Performance Improvement and Improvement Recommendation (i.e., suggestions to improve paper quality) together. We manually checked the results to verify and correct inappropriate groupings.
+
+In cases where a single term contains multiple aspects, it is placed into more than one category (e.g., Comparison with Related Work is included in both Comparison and Related Work categories).
+
+See Appendix B for more details regarding the settings and human effort.
+
+# 3.3 Results
+
+GPT-4o identified 14574 unique aspects from the reviews. We excluded 9764 terms that were re
+
+lated to paper decision, too general or specific, or could not be matched by keywords and appeared less than 50 times. These excluded terms occurred 26578 times in the corpus. The most common excluded terms are Weaknesses, Questions, Strengths, Weakness, and Comments. The remaining 4810 aspects appeared 25394 times in the corpus, with the most common ones being Comparison, Clarity, Performance, Experiments, and Results.
+
+Based on the results, we created a taxonomy that groups the aspects into 16 broad categories (see Table 2). This taxonomy shows the granularity of aspects across 3 levels: (a) the broad category names (COARSE, the most coarse), (b) the most frequently occurring aspects (FINE, finer), and (c) the raw GPT-4o outputs (LLM annotation, the finest). We also distinguish between paper-agnostic and paper-dependent aspects. Paper-agnostic aspects are relevant across all papers, while paper-dependent aspects are specific to individual papers and may not appear universally. Section A shows an example review using the proposed aspect taxonomy.
+
+Some aspects are not present in our taxonomy as review forms for major NLP venues have dedi
+
+ated fields for their evaluation, making them appear much less frequently in the review text (e.g., Reproducibility). See Table 8 for examples of the dataset we created which pairs aspects with real-life reviews and Table 9 for the aspect frequencies.
+
+# 3.4 Validity check
+
+Since LLMs have been shown to be sensitive to prompts (Chen et al., 2023; Ajith et al., 2024), we experimented with different prompts and temperature to assess the consistency of the model annotations. We used three consistency metrics: exact match, BERTScore similarity (Zhang et al., 2020), and Jaccard similarity. We show in Table 10 that the consistency between annotations generated under different settings is moderate to strong.  $45.50\%$  and  $67.14\%$  of the annotations obtained using different prompts and temperature have BERTScore similarities greater than 0.9. The increase in exact matches between the raw annotations and those mapped to the COARSE label set suggests that differences introduced by varying prompts and temperature do not noticeably affect the categorization, as most of them still fall within the same label category. The Jaccard similarities of aspects mapped to the COARSE label set also indicate strong consistency. Table 11 and 12 help interpret these scores by showing examples of how pairs of texts and sets correspond to different BERTScore and Jaccard similarities.
+
+We conducted a human evaluation to verify the model annotations, involving three human annotators to evaluate the LLM annotations mapped to the COARSE label set. We follow Yuan et al. (2022), asking the annotators to determine whether a review sentence addresses the aspects identified by GPT-4o (see Table 14 for examples). We observe that on average human annotators agree with  $91\%$  of the LLM annotations, along with fair inter-annotator agreement as measured by Fleiss' Kappa (Fleiss, 1971) (see Table 3). These results suggest that the model annotations largely align with human judgments, and that human annotators can effectively understand our taxonomy.
+
+See Appendix B.2 for more details.
+
+# 4 Aspect prediction
+
+Our fine-grained approach to aspect analysis enables two tasks: predicting the aspects that should be focused on given a paper (paper aspect prediction, PAP), and identifying the aspects that are
+
+covered in the review (review aspect prediction, RAP). These tasks are formalized as follows:
+
+$$
+f: \left\{\begin{array}{l}\mathrm {P A P}, p \rightarrow \hat {A} _ {p};\\\mathrm {R A P}, r \rightarrow \hat {A} _ {r};\end{array}\right. \tag {1}
+$$
+
+where  $f$  denotes a model,  $\hat{A}_p$  is the predicted aspects for a given paper, and  $\hat{A}_r$  is the predicted aspects for a given review.
+
+# 4.1 Method
+
+For PAP, we only focus on predicting paper-dependent aspects, as our categorization defines paper-agnostic aspects as those that are relevant across all papers. We experimented with different parts of the paper as input, including the full paper, title, keywords, and abstract, which have different implications. Using the title, keywords, or abstract as input is a heuristic method that is grounded in statistics, that reviews for similar types of papers (as determined by the title, keywords, or abstract) tend to emphasize similar aspects. Using the full paper as input may offer broader insights. Beyond leveraging heuristics, a model with access to the full paper may also capture strengths or weaknesses in certain aspects of the paper that are not evident from the title, keywords, or abstract alone.
+
+RAP differs from aspect set construction described in Section 3 in that it is implemented using a supervised approach. We utilized our curated data to train models to identify aspects within reviews. We segmented the reviews into sentences using NLTK.
+
+It is important to note that PAP is inherently more challenging than RAP. PAP operates as a heuristic method grounded in statistics, or in a more advanced setting (i.e., when the input is the full paper), goes beyond heuristics to infer strengths or weaknesses in certain aspects of the paper. In contrast, RAP resembles summarization, as it extracts aspects directly from the review text, making it comparatively less challenging than PAP.
+
+We modeled both tasks as multi-label sequence classification. For both tasks, we tested bag-of-words with random forest  $(\mathsf{BoW} + \mathsf{RF})$  and RoBERTa (Liu et al., 2019). Both models are strong in supervised settings, and they are lighter-weight alternatives to LLMs. We used focal loss (Lin et al., 2017) to address label imbalance. For PAP, we also experimented with GPT-4o in both zero-shot and few-shot settings. Our evaluation metrics include
+
+precision, recall, F1 score, and Jaccard similarity score. See Appendix C for more details.
+
+# 4.2 Results
+
+Table 3 shows the results of PAP experiments. Overall, the models do not perform well regardless of the type of input used. We observe a small advantage for BoW+RF when using the full paper as input, possibly because the model captures more nuanced information when provided with the entire paper. Using the COARSE label set results in much higher performance than the FINE label set (see Table 18). This is partly due to the prevalence of certain COARSE aspect labels, such as Methodology, which appear in nearly all papers.
+
+In the GPT-4o experiments<sup>4</sup>, the evaluation of using the full paper or abstract as input is in the zero-shot setting, and the rest is in the few-shot setting. Though it is in the few-shot setting, GPT-4o outperforms RoBERTa which is trained on the full training data. GPT-4o tends to be better at capturing heuristics.
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.5343</td><td>0.6250</td><td>0.5552</td><td>0.4998</td></tr><tr><td>GPT-4o</td><td>0.5625</td><td>0.5542</td><td>0.5352</td><td>0.4024</td></tr></table>
+
+(a) full paper  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.5568</td><td>0.6108</td><td>0.5511</td><td>0.4901</td></tr><tr><td>RoBERTa</td><td>0.5224</td><td>0.6599</td><td>0.5781</td><td>0.5439</td></tr><tr><td>GPT-4o</td><td>0.7138</td><td>0.7193</td><td>0.6756</td><td>0.5451</td></tr></table>
+
+Table 4 shows the results of RAP experiments. RoBERTa is the best performing model for this task. Models trained using the COARSE label set achieve higher performance compared to those trained with the FINE label set, and the few-shot setting offers small improvement for GPT-40 over the zero-shot setting (see Table 19 and 20).
+
+In general, models perform better on RAP, and the COARSE label set yields higher performance across both tasks.
+
+(b) keywords  
+Table 3: The highest precision, recall, F1 score, and Jaccard similarity; PAP; FINE label set. For GPT-4o, the evaluation of using the full paper as input is in the zero-shot setting, and that of using the keywords is in the few-shot setting. See Table 17 for more results.  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.8058</td><td>0.5810</td><td>0.6416</td><td>0.5070</td></tr><tr><td>RoBERTa</td><td>0.7720</td><td>0.7664</td><td>0.7675</td><td>0.7089</td></tr><tr><td>GPT-4o</td><td>0.5696</td><td>0.5757</td><td>0.5328</td><td>0.4573</td></tr></table>
+
+Table 4: The highest precision, recall, F1 score, and Jaccard similarity; RAP; COARSE label set. For GPT-4o, the evaluation is in the few-shot setting.
+
+# 5 Practical applications
+
+We now demonstrate how a comprehensive, fine-grained aspect set and the proposed tasks enable new types of NLP assistance in the peer review process. We picked the trained RoBERTa model on the COARSE label set with the best F1 score and obtained predicted aspects using it. This model achieves an F1 score above 0.75 for 10 out of 17 labels (see Table 21 for the classification report).
+
+# 5.1 Aspect analysis
+
+A more comprehensive set of aspects allows for more detailed aspect analysis. Figure 1 shows the most frequent aspects across 4 submission tracks in EMNLP23.<sup>5</sup> While Machine Translation and Multilinguality and Linguistic Diversity are different tracks, the set of aspects that reviewers emphasize most are very similar. This suggests that there may be overlaps in the papers within these tracks (also as indicated by the track names). We observe a different pattern in the Resources and Evaluation track, where reviewers focus more on Data/Task and Evaluation than in other tracks.
+
+![](images/944a735618e45c6c4a90ebb287bcaa97bedb77c80718b1c328b7d19b2a3f6a21.jpg)  
+Figure 1: The 5 most frequent aspects in 4 submission tracks in EMNLP23. Figure 5 shows the full results.
+
+Figure 4 in Appendix D.1 shows the similarity of submission tracks based on the Levenshtein similarity of the 10 most frequent aspects within each track in EMNLP23. Some tracks are more similar to each other. For example, Question Answering is most similar to tracks such as Summarization and Information Extraction.
+
+Table 5 shows additional examples of how reviewers emphasize certain aspects more in some tracks than in others. For example, the frequency of Analysis is the highest in the Computational Social Science and Cultural Analytics track.
+
+<table><tr><td>track</td><td>review (%)</td></tr><tr><td>Comp. Social Science &amp; Cultural Analytics</td><td>69.23</td></tr><tr><td>Ling. Theor., Cogn. Model., &amp; Psycholing.</td><td>68.75</td></tr><tr><td>Commonsense Reasoning</td><td>62.62</td></tr><tr><td>Multilinguality &amp; Linguistic Diversity</td><td>60.68</td></tr><tr><td>Machine Learning for NLP</td><td>60.00</td></tr></table>
+
+Table 5: The tracks with the 5 highest frequencies of Analysis in EMNLP23. Table 22 shows the full results.
+
+This type of analysis helps compare reviewing across different tracks and venues and can inform the development of review forms and guidelines that could prompt reviewers to focus on certain aspects relevant to particular tracks to ensure a more comprehensive review.
+
+# 5.2 Review comparison
+
+In this section, we demonstrate that a more comprehensive set of aspects introduces a new dimension for comparing reviews. We compare human-written reviews with LLM-generated reviews. We used the LLM-generated reviews used in Du et al. (2024) and generated reviews for 100 randomly sampled papers from each of EMNLP23 and ICLR24. We generated reviews using GPT-4o with the prompt used in Liang et al. (2024b) and a prompt of our own (see Table 23 for the prompts).
+
+To compare the reviews, we predicted the aspects in each review and calculated the Jaccard similarity between sets of aspects. Figure 2 visualizes the similarity between each pair of human-written and LLM-generated reviews.
+
+We observe that LLM-generated reviews show a higher degree of similarity to each other in terms of aspects. This suggests that LLM-generated reviews may be more generic than human-written ones, with the model tending to comment on similar sets of aspects across different papers. This
+
+![](images/c7e1dfabc3915df3c5711f0450fbb9133610500286b0741605ed43b231ac88af.jpg)
+
+![](images/f05d6a54c538316f87009546c2dd7c872bd000f855808e9c34fe773a50915705.jpg)
+
+![](images/1de7bb28c13628bf8867ee62670702446d8171b604502f723d4fdca642a43d07.jpg)  
+(a) COARSE  
+Figure 2: The heatmap of the Jaccard similarity between each pair of the human-written reviews and LLM-generated reviews generated using EMNLP23 papers and Liang et al. (2024b)'s prompt. Figure 6 in Appendix D.2 shows the rest of the results.
+
+![](images/9c0bd1ba2129eb2ebb355fcd7761bc49721224131628653ee5f2d001b6d1c1e5.jpg)  
+(b) FINE
+
+could also imply that the quality of LLM-generated reviews is still lacking in terms of specificity.
+
+# 5.3 LLM-generated review detection
+
+We used the same data as in Section 5.2, where LLM-generated reviews are created using prompts with different levels of prompt engineering: Du et al. (2024) has the heaviest prompt engineering, followed by Liang et al. (2024b), and our own prompt is the simplest and involves the least prompt engineering (see Table 23 for more details). These prompts represent the common types of prompts used to generate reviews end-to-end. There is one LLM-generated review for each paper.
+
+Based on our observations in Section 5.2, we design a simple strategy to detect LLM-generated reviews. We define  $\text{sim}(A_i, A_j)$  as the similarity between two sets of aspects. For each review  $r$ , we define the intra-similarity  $S_{\text{intra}}$  as the average similarity between reviews for the same paper, and the inter-similarity  $S_{\text{inter}}$  as the average similarity of reviews for different papers:
+
+$$
+\begin{array}{l} S _ {t} = \frac {1}{n} \sum \operatorname {s i m} \left(A _ {r}, A _ {r _ {i}}\right), \\ t = \left\{ \begin{array}{l} \text {i n t r a , i f} r, r _ {i} \in R _ {p}, p \in P, r \neq r _ {i}; \\ \text {i n t e r , i f} r \in R _ {p}, r _ {i} \in R _ {q}, p, q \in P, p \neq q; \end{array} \right. \tag {2} \\ \end{array}
+$$
+
+where  $n$  is the number of  $A_{r}, A_{r_{i}}$  pairs. Intuitively, reviews for the same paper are more similar to
+
+each other than reviews for different papers, as a review is specific to the paper. Therefore, we expect  $|S_{intra} - S_{inter}|$  to be large. Based on this intuition, we calculate  $|S_{intra} - S_{inter}|$  for each  $r$  in  $R_p$  and propose two performance metrics: (a)  $@1$ , which measures accuracy by determining whether the LLM-generated review is the one with the lowest score, and (b)  $@2$ , which calculates the percentage of cases where the LLM-generated review is among the 2 reviews with the lowest scores.
+
+Table 6 shows the detection results. We used Jaccard similarity and compared our method with an implementation of Equation 2 using SentenceBERT (SBERT) (Reimers and Gurevych, 2019), which calculates the similarity between the embeddings of review texts. Experiments were conducted using both the COARSE and FINE label sets, as well as ACL'18 which is a commonly used aspect set in previous studies (see Table 1). We calculated a random baseline that selects a review at random as the LLM-generated review. Note that the FINE aspect set consistently outperforms ACL'18 and SBERT, and our method is robust across LLM-generated reviews generated using different prompts.
+
+<table><tr><td>dataset</td><td>aspect set</td><td>@1</td><td>@2</td></tr><tr><td rowspan="5">Du et al. (2024) (avg=4.80)</td><td>random</td><td>0.27</td><td>0.55</td></tr><tr><td>SBERT</td><td>0.35</td><td>0.80</td></tr><tr><td>ACL&#x27;18</td><td>0.40</td><td>0.80</td></tr><tr><td>COARSE</td><td>0.50</td><td>0.75</td></tr><tr><td>FINE</td><td>0.50</td><td>0.70</td></tr><tr><td rowspan="5">Liang et al. (2024b) (avg=4.52)</td><td>random</td><td>0.25</td><td>0.47</td></tr><tr><td>SBERT</td><td>0.47</td><td>0.66</td></tr><tr><td>ACL&#x27;18</td><td>0.49</td><td>0.72</td></tr><tr><td>COARSE</td><td>0.53</td><td>0.73</td></tr><tr><td>FINE</td><td>0.66</td><td>0.89</td></tr><tr><td rowspan="5">ours (avg=4.52)</td><td>random</td><td>0.25</td><td>0.47</td></tr><tr><td>SBERT</td><td>0.55</td><td>0.71</td></tr><tr><td>ACL&#x27;18</td><td>0.56</td><td>0.82</td></tr><tr><td>COARSE</td><td>0.48</td><td>0.72</td></tr><tr><td>FINE</td><td>0.65</td><td>0.87</td></tr></table>
+
+Table 6: The @1 and @2 accuracy of the detection across different aspect sets and methods. "avg" is the average number of reviews per paper. There is one LLM-generated review for each paper.
+
+While this detection approach does not achieve the same level of performance as the reference-based and zero-shot approaches, such as the  $90\%$  accuracy reported by Yu et al. (2024) or our own GPT-4o results in Table 24, it nevertheless provides valuable insights. It demonstrates that current LLMs tend to generate reviews that are generic in terms of aspects. A key advantage of this approach
+
+lies in interpretability, opening new opportunities for broader applications such as review quality assessment. Moreover, the performance gains observed when using the FINE label set (see Table 6) highlight the importance of aspect granularity-finer labels reveal patterns that are useful in distinguishing LLM-generated reviews from human-written ones, patterns that coarse labels fail to capture. Thus, we consider this approach promising, and leave the development of better performing aspect-based detectors to future work.
+
+# 5.4 Recommendations
+
+Based on our results, we make the following recommendations on selecting aspect granularity for a given application. The COARSE label set is more suitable for high-level analysis tasks, such as analyzing review focus across different tracks and venues (Section 5.1). The FINE label set is more appropriate for tasks that require nuanced analysis, where capturing specific and detailed feedback is critical, such as review comparison (Section 5.2) and LLM-generated review detection (Section 5.3). In some cases, a hybrid strategy might be the best option. For example, when evaluating review quality, one can first apply the COARSE label set to assess coverage (i.e., whether key aspects are addressed), and then use the FINE label set to assess specificity of the review. Aspect granularity is a design choice, and the optimal configuration depends on the task and user needs.
+
+# 6 Conclusion
+
+In this paper, we have introduced a data-driven approach to peer review aspect analysis. We have provided an operational definition of aspect, and developed an semi-automatic approach to identify aspects from peer reviews. We have proposed a taxonomy that involves a comprehensive set of aspects with different granularity, and introduce a new dataset of peer reviews augmented with aspects. We introduced two tasks, paper aspect prediction and review aspect prediction, and have shown how they contribute to a detailed empirical study of aspects. Our results demonstrate that fine-grained, data-driven aspects complement coarse aspects from review guidelines, and allow for more nuanced review comparison and new interpretable approaches for LLM-generated review detection.
+
+# Limitations
+
+As discussed in Section 3.1, aspect is difficult to define. We adopted an operational definition as a practical approach. Defining individual aspects is also challenging. For instance, it is difficult to determine the scope of "Methodology." Polysemy further complicates this issue—for example, "Improvement" may refer to either a method's improvement, or places where the paper can be improved. We did a manual inspection of the categorization results to minimize the impact of these issues. While our study provides a foundation for data-driven research on review aspect, future work may seek to refine and expand upon our current definition.
+
+The taxonomy was constructed by an expert annotator (one of the authors of this paper) with expertise in NLP and peer review, based on domain knowledge. This taxonomy may not be optimal, and alternative approaches to categorizing aspects may exist. Our taxonomy is an attempt to streamline the analysis and application of aspects, and it has been shown to be effective for the purposes of this study. We deem a future multiple-expert study in fine-grained aspect identification promising.
+
+In this work, we focus on the NLP domain. While a comprehensive cross-domain analysis would be of great interest, it would require reviewing data and domain experts from other fields to construct a new taxonomy. As this would require a substantial study deserving a paper of its own, we leave it as potential future work.
+
+Consistency is a fundamental issue when working with LLM-generated annotations. As reported in Section 3.4, GPT-4o indeed shows some sensitivity to prompts and temperature. Our validity checks suggest a reasonable degree of consistency and reliability of the LLM annotations. The applications of these aspect annotations, especially in the LLM-generated review detection task, provide further support for their reliability. Though we specify the seed (see Appendix B and C), the exact reproducibility of the results related to the closed-weights LLMs like GPT-4o is a concern. As open models become more capable, experimentation with alternative open models and aspect schemata will become possible.
+
+The LLM-generated annotations in this work are based on GPT-4o. While a comparative analysis across annotations generated by different models would be insightful, such a study would require not only applying different models, but also con
+
+structuring multiple taxonomies, which needs expert involvement. Though valuable, this is beyond the scope of a single study.
+
+For human annotation, we recognize that using aspect labels from our taxonomy may introduce potential label bias. A straightforward solution would be to ask the annotators to do an open-ended annotation without pre-defined labels. However, such study is very challenging in terms of the cognitive burden on annotators and ensuring annotation consistency. Automation bias is also a concern. While we address this by asking annotators to provide explanations for their annotations, follow-up work can explore alternative experimental strategies to measure and mitigate automation bias.
+
+For the purposes of this study, we have kept the design of the PAP straightforward, treating it as a binary classification task predicting which aspects should be focused on given a paper. Yet, aspect relevance might indeed not be a binary decision, and modeling PAP as a ranking or regression problem could better reflect real world scenarios, where aspects have different levels of importance. Given the scope of the paper, we leave this investigation to future work.
+
+In addition, we point at potential selection bias due to data availability. All papers associated with the ICLR20-24 and EMNLP23 datasets are camera-ready versions. This may affect the validity of the results of PAP experiments in Section 4 since in practice PAP deals with (potentially lower-quality) submission manuscripts. We do not consider the use of camera-ready versions problematic for review comparison and LLM-generated review detection in Section 5.2 and 5.3. A further bias in item selection might be introduced by data imbalance with respect to paper acceptance. The EMNLP23 dataset contains only 9 rejected papers; in NLPeer,  $69\%$  of the papers are accepted papers (Dycke et al., 2023), which does not correspond to a natural distribution of submissions. Such skew may lead to an overestimation of aspect frequency: aspects commonly associated with accepted papers may appear more frequently than they would in a more balanced dataset. Consequently, some of our findings may be influenced by this bias. If these overrepresented aspects are used to inform review forms or guidelines, they may introduce new biases into the review process, which reduces their effectiveness. As more review data become available, this limitation can be mitigated.
+
+In Section 5.3, we only focus on the commonly used end-to-end general prompts for our review generator. We did not consider prompts that involve paper-specific aspects, as we consider these to be human-in-the-loop prompts, where the user must first read the paper to identify paper-specific aspects and then incorporate them into the prompt, which would require an experimental setup beyond our scope. This approach represents a form of human-AI collaboration, and we plan to explore this direction in future work.
+
+# Ethics Statement
+
+This work does not suggest or imply that the proposed task of predicting the aspects to focus on a given paper may replace human involvement. Instead, it is designed to assist reviewers by recommending the aspects to consider during their review. Reviewers retain full autonomy over their decisions regarding whether to include the suggested aspects. Since all the review data used in this study are publicly available and anonymized, processing it with commercial LLMs does not raise ethical concerns.
+
+# Acknowledgements
+
+This work has been funded by the German Research Foundation (DFG) as part of the PEER project (grant GU 798/28-1), and funded/co-funded by the European Union (ERC, InterText, 101054961). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. This work has been funded by the LOEWE Distinguished Chair "Ubiquitous Knowledge Processing", LOEWE initiative, Hesse, Germany (Grant Number: LOEWE/4a//519/05/00.002(0002)/81). We gratefully acknowledge the support of Microsoft with a grant for access to OpenAI GPT models via the Azure cloud (Accelerate Foundation Model Academic Research).
+
+# References
+
+Anirudh Ajith, Chris Pan, Mengzhou Xia, Ameet Deshpande, and Karthik Narasimhan. 2024. Instructeval: Systematic evaluation of instruction selection methods. In Findings of the Association for Computational Linguistics: NAACL 2024, Mexico City,
+
+Mexico, June 16-21, 2024, pages 4336-4350. Association for Computational Linguistics.  
+Ines Arous, Jie Yang, Mourad Khayati, and Philippe Cudré-Mauroux. 2021. Peer grading the peer reviews: A dual-role approach for lightening the scholarly paper review process. In WWW '21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021, pages 1916-1927. ACM / IW3C2.  
+Steven Bird and Edward Loper. 2004. NLTK: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, pages 214-217, Barcelona, Spain. Association for Computational Linguistics.  
+Som Biswas, Dushyant Dobaria, and Harris L Cohen. 2023. Focus: Big data: Chatgpt and the future of journal reviews: A feasibility study. The Yale Journal of Biology and Medicine, 96(3):415.  
+Souvic Chakraborty, Pawan Goyal, and Animesh Mukherjee. 2020. Aspect-based sentiment analysis of scientific reviews. In JCDL '20: Proceedings of the ACM/IEEE Joint Conference on Digital Libraries in 2020, Virtual Event, China, August 1-5, 2020, pages 207-216. ACM.  
+Alessandro Checco, Lorenzo Bracciale, Pierpaolo Loretti, Stephen Pinfield, and Giuseppe Bianchi. 2021. AI-assisted peer review. *Humanities and Social Sciences Communications*, 8(1):1-11.  
+Yanda Chen, Chen Zhao, Zhou Yu, Kathleen McKeown, and He He. 2023. On the relation between sensitivity and accuracy in in-context learning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 155-167, Singapore. Association for Computational Linguistics.  
+Elizabeth Clark, Tal August, Sofia Serrano, Nikita Hahuong, Suchin Gururangan, and Noah A. Smith. 2021. All that's 'human' is not gold: Evaluating human evaluation of generated text. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021, pages 7282-7296. Association for Computational Linguistics.  
+Jiangshu Du, Yibo Wang, Wenting Zhao, Zhongfen Deng, Shuaiqi Liu, Renze Lou, Henry Peng Zou, Pranav Narayanan Venkit, Nan Zhang, Mukund Srinath, Haoran Ranran Zhang, Vipul Gupta, Yinghui Li, Tao Li, Fei Wang, Qin Liu, Tianlin Liu, Pengzhi Gao, Congying Xia, Chen Xing, Cheng Jiayang, Zhaowei Wang, Ying Su, Raj Sanjay Shah, Ruohao Guo, Jing Gu, Haoran Li, Kangda Wei, Zihao Wang, Lu Cheng, Surangika Ranathunga, Meng Fang, Jie Fu, Fei Liu, Ruihong Huang, Eduardo Blanco, Yixin Cao, Rui Zhang, Philip S. Yu, and Wenpeng Yin. 2024. LLMs assist NLP researchers: Critique paper (meta-)reviewing. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language
+
+Processing, pages 5081-5099, Miami, Florida, USA.  
+Association for Computational Linguistics.  
+Nils Dycke, Ilia Kuznetsov, and Iryna Gurevych. 2023. NLPeer: A unified resource for the computational study of peer review. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5049-5073, Toronto, Canada. Association for Computational Linguistics.  
+Alexander R. Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. SummEval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391-409.  
+Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378.  
+Catherine A. Gao, Frederick M. Howard, Nikolay S. Markov, Emma C. Dyer, Siddhi Ramesh, Yuan Luo, and Alexander T. Pearson. 2023. Comparing scientific abstracts generated by chatgpt to real abstracts with detectors and blinded human reviewers. npj Digital Medicine, 6.  
+Zhaolin Gao, Kianté Brantley, and Thorsten Joachims. 2024. Reviewer2: Optimizing review generation through prompt generation. CoRR, abs/2402.10886.  
+Sebastian Gehrmann, Hendrik Strobelt, and Alexander M. Rush. 2019. GLTR: statistical detection and visualization of generated text. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28 - August 2, 2019, Volume 3: System Demonstrations, pages 111-116. Association for Computational Linguistics.  
+Alexander Goldberg, Ivan Stelmakh, Kyunghyun Cho, Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Nihar B. Shah. 2023. Peer reviews of peer reviews: A randomized controlled trial and other experiments. CoRR, abs/2311.09497.  
+Biyang Guo, Xin Zhang, Ziyuan Wang, Minqi Jiang, Jinran Nie, Yuxuan Ding, Jianwei Yue, and Yupeng Wu. 2023. How close is chatgpt to human experts? comparison corpus, evaluation, and detection. CoRR, abs/2301.07597.  
+Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. 2020. Automatic detection of generated text is easiest when humans are fooled. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 1808-1822. Association for Computational Linguistics.  
+Tom Jefferson, Elizabeth Wager, and Frank Davidoff. 2002. Measuring the Quality of Editorial Peer Review. JAMA, 287(21):2786-2790.
+
+Yiqiao Jin, Qinlin Zhao, Yiyang Wang, Hao Chen, Kaijie Zhu, Yijia Xiao, and Jindong Wang. 2024. Agentreview: Exploring peer review dynamics with LLM agents. CoRR, abs/2406.12708.  
+Neha Nayak Kennard, Tim O'Gorman, Rajarshi Das, Akshay Sharma, Chhandak Bagchi, Matthew Clinton, Pranay Kumar Yelugam, Hamed Zamani, and Andrew McCallum. 2022. DisAPERE: A dataset for discourse structure in peer review discussions. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2022, Seattle, WA, United States, July 10-15, 2022, pages 1234-1249. Association for Computational Linguistics.  
+Ilia Kuznetsov, Osama Mohammed Afzal, Koen Dercksen, Nils Dycke, Alexander Goldberg, Tom Hope, Dirk Hovy, Jonathan K Kummerfeld, Anne Lauscher, Kevin Leyton-Brown, et al. 2024. What can natural language processing do for peer review? CoRR, abs/2405.06563.  
+Weixin Liang, Zachary Izzo, Yaohui Zhang, Haley Lepp, Hancheng Cao, Xuandong Zhao, Lingjiao Chen, Haotian Ye, Sheng Liu, Zhi Huang, Daniel A. McFarland, and James Y. Zou. 2024a. Monitoring ai-modified content at scale: A case study on the impact of chatgpt on AI conference peer reviews. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net.  
+Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Yi Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Scott Smith, Yian Yin, Daniel A. McFarland, and James Zou. 2024b. Can large language models provide useful feedback on research papers? a large-scale empirical analysis. NEJM AI, 1(8):AIoa2400196.  
+Bill Yuchen Lin, Abhilasha Ravichander, Ximing Lu, Nouha Dziri, Melanie Sclar, Khyathi Chandu, Chandra Bhagavatula, and Yejin Choi. 2023a. The unlocking spell on base llms: Rethinking alignment via in-context learning. CoRR, abs/2312.01552.  
+Jialiang Lin, Jiaxin Song, Zhangping Zhou, Yidong Chen, and Xiaodong Shi. 2023b. Automated scholarly paper review: Concepts, technologies, and challenges. Information Fusion, 98:101830.  
+Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. 2017. Focal loss for dense object detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 2999-3007. IEEE Computer Society.  
+Ryan Liu and Nihar B. Shah. 2023. Reviewergpt? an exploratory study on using large language models for paper reviewing. CoRR, abs/2306.00622.
+
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692.  
+Xiao Pu, Mingqi Gao, and Xiaojun Wan. 2023. Summarization is (almost) dead. CoRR, abs/2309.09558.  
+Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.  
+Anna Rogers, Marzena Karpinska, Jordan Boyd-Graber, and Naoaki Okazaki. 2023. Program chairs' report on peer review at acl 2023. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages xl-lxxv, Toronto, Canada. Association for Computational Linguistics.  
+Robert Schulz, Adrian Barnett, René Bernard, Nicholas JL Brown, Jennifer A Byrne, Peter Eckmann, Małgorzata A Gazda, Halil Kilicoglu, Eric M Prager, Maia Salholz-Hillel, et al. 2022. Is the future of peer review automated? BMC research notes, 15(1):203.  
+Nihar B. Shah. 2022. Challenges, experiments, and computational solutions in peer review. Communications of the ACM, 65(6):76-87.  
+Ivan Stelmakh, Nihar B. Shah, Aarti Singh, and Hal Daume III. 2021a. A novice-reviewer experiment to address scarcity of qualified reviewers in large conferences. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 4785-4793. AAAI Press.  
+Ivan Stelmakh, Nihar B. Shah, Aarti Singh, and Hal Daumé III. 2021b. Prior and prejudice: The novice reviewers' bias against resubmissions in conference peer review. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1):75:1-75:17.  
+Lu Sun, Aaron Chan, Yun Seo Chang, and Steven P. Dow. 2024a. Reviewflow: Intelligent scaffolding to support academic peer reviewing. In Proceedings of the 29th International Conference on Intelligent User Interfaces, IUI 2024, Greenville, SC, USA, March 18-21, 2024, pages 120-137. ACM.  
+Lu Sun, Stone Tao, Junjie Hu, and Steven P. Dow. 2024b. Metawriter: Exploring the potential and perils of AI
+
+writing support in scientific peer review. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1-32.  
+Jingyan Wang, Ivan Stelmakh, Yuting Wei, and Nihar B. Shah. 2021. Debiasing evaluations that are biased by evaluations. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 10120-10128. AAAI Press.  
+Qingyun Wang, Qi Zeng, Lifu Huang, Kevin Knight, Heng Ji, and Nazneen Fatema Rajani. 2020. Reviewobot: Explanable paper review generation based on knowledge synthesis. In Proceedings of the 13th International Conference on Natural Language Generation, INLG 2020, Dublin, Ireland, December 15-18, 2020, pages 384-397. Association for Computational Linguistics.  
+Zhongyi Wang, Haoxuan Zhang, Haihua Chen, Yunhe Feng, and Junhua Ding. 2024. Content-based quality evaluation of scientific papers using coarse feature and knowledge entity network. Journal of King Saud University - Computer and Information Sciences, 36(6):102119.  
+Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Derek F. Wong, and Lidia S. Chao. 2023. A survey on llm-generated text detection: Necessity, methods, and future directions. CoRR, abs/2310.14724.  
+Sungduk Yu, Man Luo, Avinash Madasu, Vasudev Lal, and Phillip Howard. 2024. Is your paper being reviewed by an llm? investigating AI text detectability in peer review. CoRR, abs/2410.03019.  
+Weizhe Yuan, Pengfei Liu, and Graham Neubig. 2022. Can we automate scientific reviewing? Journal of Artificial Intelligence Research, 75:171-212.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. 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.
+
+# A Using the proposed aspect set to evaluate this paper
+
+We conducted a self-review of this paper using the aspect set proposed in this work. We maintained neutral, and the points listed below serve as examples and they are not exhaustive. This section is not generated by an LLM. For paper-agnostic aspects:
+
+- Contribution: this paper derives a comprehensive set of aspects from NLP paper reviews (Section 3); this paper creates a new dataset of NLP paper reviews augmented with aspects (Section 3); this paper evaluates models using this dataset on two tasks (Section 4); this paper shows practical applications of a comprehensive set of aspects (Section 5).
+
+# - DDDDEI:
+
+- (Definition) this paper provides an operational definition of aspect (Section 3.1);  
+- (Description) this paper describes the workflow regarding aspect identification and prediction;  
+- (Detail) this paper reports detailed experimental settings and implementation specifics.
+
+# - IJMV:
+
+- (Motivation) this paper is motivated by the lack of a comprehensive set of aspects in the review guidelines for major NLP venues; the use of an LLM to identify aspects from reviews is motivated by evidence that shows the strong performance of LLMs in aspect identification (Section 2);  
+- (Validation) Section 3.4 validate the method used for aspect identification.
+
+- Novelty: this paper is the first attempt to derive a comprehensive set of aspects from NLP paper reviews; this paper introduces a new dataset of NLP paper reviews augmented with aspects.
+
+# - Presentation:
+
+- (Clarity) this paper describes the methods, experiments, and results clearly;  
+- (Figure) this paper uses many figures to illustrate their findings.
+
+- Related work: this paper reviews the opportunities and challenges of peer review in the era of LLMs, aspects in peer review, the quality of review writing, and LLM-generated review detection (Section 2).  
+- Significance: this paper provides a comprehensive set of aspects which benefits peer review in multiple ways, such as contributing to better review guidelines and LLM-generated review detection.
+
+Given that our paper is both a resource and NLP application paper, for paper-dependent aspects, please pay special attention to the Data/Task and Methodology aspects (findings in Section 5.1):
+
+- Analysis: this paper presents an analysis of the identification results (Section 3.3) and one in terms of track and review similarity (Section 5.1 and 5.2).
+
+- Comparison: this paper compares model performance on aspect prediction (Section 4) and compares different methods for LLM-generated review detection (Section 5.3).
+
+# Data/Task:
+
+- (Data) this paper uses paper reviews from both NLP and ML venues across different years, and this paper creates a new dataset of NLP paper reviews augmented with aspects (Section 3);  
+- (Task) this paper proposes two tasks related to aspect (Section 4).
+
+- Evaluation: this paper uses a range of metrics, including BERTScore, SentenceBERT, and Jaccard similarity.  
+- Experiment: this paper conducts many experiments (Section 3, 4, and 5), and all the experimental settings are reported.
+
+# Methodology:
+
+- (Method) this paper proposes a workflow for identifying aspects using an LLM;  
+- (Model) this paper uses GPT-4o, BoW+RF, and RoBERTa;  
+-(Framework) this paper proposes a taxonomy of aspects;  
+- (Implementation) all the implementation details are provided.
+
+# - Result:
+
+- (Findings) this paper finds that LLM-generated reviews are more generic than human-written reviews (Section 5.2);  
+- (Performance) this paper shows that models trained using the aspect sets proposed in this work perform better than using previous ones (Section 5.3).
+
+# Return to main text.
+
+# B More on aspect set construction
+
+The EMNLP23 and ICLR data were obtained via the OpenReview API. We used the prompt in Table 7 to identify aspects from the reviews. We did not include our proposed definition of aspect in the prompt, as we assume that GPT-4o can naturally capture the common meaning of the term. Though the concept of aspect is difficult to define precisely, it is widely used in natural language, so the model should be able to capture its meaning. We apply our proposed definition to filter the model outputs later on (see Section 3.2).
+
+We set temperature  $= 0$  and seed  $= 2266$ . We segmented the reviews into sentences, and removed entries that consist of only indices (e.g., "1").
+
+Identify the aspect(s) that each of the given sentences focuses on. Format the output in a json dictionary. For example, given a dictionary as follows:
+
+{"1": "The methodology is convincing, and the improvement is noticeable.", "2": "A dataset is assembled."}
+
+The output should be:
+
+{"1": "Methodology, Improvement", "2": "Dataset"}
+
+Table 7: The prompt used to identify aspects from the reviews. Return to main text.
+
+Our method involves human effort on:
+
+- Post-processing LLM annotations: this is a one-time operation that involves both automated and manual steps. The automated part categorizes LLM annotations using high frequency terms and removes low frequency terms by setting a frequency threshold. The manual part focuses on verifying the remaining terms. In our work, this manual verification took one expert approximately 10 hours.  
+- Taxonomy construction: we construct a taxonomy based on post-processed annotations. Mapping raw aspects into a taxonomy that reflects domain-relevant evaluation dimensions (such as the one shown in Table 2) involves both domain expertise and expert judgment. In our case, this step took one expert approximately 72 hours.
+
+# B.1 More on results
+
+Table 8 shows examples from the dataset we created. Each review sentence is accompanied by an LLM annotation, which is mapped to both the COARSE and FINE label sets. Table 9 shows the frequency of aspects of different levels of granularity in the dataset we created.
+
+# B.2 More on validity check
+
+We experimented with a different prompt, where we replaced the word "aspect" in the prompt shown in Table 7 with a synonym "facet". This prompt is designed to determine whether the LLM truly understands the semantics of the prompt rather than relying on specific word choices. We also tested with temperature=[0,1].
+
+Table 10 shows the consistency between annotations generated under different settings. Table 11 and 12 show examples of how pairs of texts and sets correspond to different BERTScore and Jaccard similarities.
+
+For the human annotation, we recruited annotators through Prolific. We applied screening conditions to ensure quality: participants are required to hold a graduate or doctorate degree in Computer Science, have English as their first language, and have an approval rate above  $90\%$  for previous Prolific submissions. We conducted several pilot studies and selected 3 annotators for the full study. See here for the annotation guidelines.
+
+We sampled review sentences to maintain class balance as much as possible (the distribution of aspects in the annotation file is shown in Table 13). We sampled 100 reviews, which corresponds to 1852 sentences. We follow the setup in Fabbri et al. (2021), Yuan et al. (2022), and Liang et al. (2024b), asking the annotators whether each review sentence addresses the aspects identified by GPT-4o. To mitigate automation bias and prevent participants from simply confirming all queries, we required them to provide explanations for their evaluations. Table 14 shows examples of the questionnaire entries we distributed. The total number of entries in the questionnaire is 3032.
+
+The "yes" rates, i.e., the agreement between human and LLM annotations, are  $85.19\%$ ,  $90.11\%$ , and  $97.56\%$  for the three annotators (with  $100\%$  indicating complete agreement with all LLM annotations). The Fleiss' Kappas (Fleiss, 1971), shown in Figure 3, indicate substantial inter-annotator agreement for the first 400 entries in the questionnaire,
+
+<table><tr><td>review sentence</td><td>LLM annotation</td><td>COARSE</td><td>FINE</td></tr><tr><td>The proposed method demonstrates commendable innovation, standing apart from mere amalgamation of existing models.</td><td>Innovation, Method</td><td>Methodology, Novelty</td><td>Method, Novelty</td></tr><tr><td>The novel approach showcases tangible efficacy without introducing additional parameters.</td><td>Efficacy, Novel Approach, Parameters</td><td>Methodology, Novelty</td><td>Approach, Novelty, Parameter</td></tr><tr><td>1.The absence of a computational complexity analysis, coupled with marginal and non-significant experimental performance improvements, raises concerns regarding the practical significance.</td><td>Computational Complexity, Experimental Performance, Practical Significance</td><td>Experiment, Methodology, Result, Significance</td><td>Complexity, Experiment, Performance, Significance</td></tr><tr><td>If the proposed model introduces high computational complexity and yields only marginal gains, its real-world utility may be limited.</td><td>Computational Complexity, Real-world Utility, Marginal Gains</td><td>Methodology</td><td>Complexity</td></tr><tr><td>2.The paper lacks comparative analysis with some important baselines, such as P-tuning v2.</td><td>Comparative Analysis, Baslines</td><td>Analysis, Comparison</td><td>Analysis, Baseline</td></tr><tr><td>Additionally, the theoretical substantiation for the proposed method is insufficiently detailed.</td><td>Theoretical Substantiation, Method</td><td>Methodology, Theory</td><td>Method, Theory</td></tr><tr><td>The exclusive use of a single language model, T5-base, as the backbone prompts doubts about the general applicability of the proposed approach across a broader spectrum of models.</td><td>Language Model, General Applicability</td><td>Methodology</td><td>Application, Model</td></tr><tr><td>Furthermore, the paper lacks visual or interpretable analyses that incorporate concrete natural language statements.</td><td>Visual Analysis, Interpretability, Natural Language Statements</td><td>Analysis, DDDDEI</td><td>Analysis, Interpretation</td></tr><tr><td>Considering these points, I respectfully recommend that the authors thoroughly address these shortcomings to enhance the paper&#x27;s overall quality and potential for contribution before reconsidering it for acceptance.</td><td>Recommendations, Paper Quality, Contribution</td><td>Contribution</td><td>Contribution</td></tr></table>
+
+Table 8: Examples from the dataset we created. Each of the review sentence is augmented with aspects. See here for the complete dataset. Return to main text. Return to appendix.
+
+but it declined as the annotation progressed, suggesting that annotation quality may not have been consistent throughout. It is important to note that there should not be a significant difference in difficulty between the first 400 entries and the remaining ones, as all entries were randomly sampled.
+
+![](images/c5a175b5b4c48525bf819e10ed14d186dde0b43b0762c04833d96f1e9b935c34.jpg)  
+Figure 3: The Fleiss' Kappas for the annotations across different ranges. For example, "600" on the x-axis indicates that the first 600 annotations have a Fleiss' Kappa of 0.4739. The overall Fleiss' Kappa is 0.1944. Return to main text.
+
+Table 15 shows the agreement between human and LLM annotations across different aspects. Agreement exceeds 0.9 for most aspects, with the highest agreement observed for Contribution and Ablation, while that for Significance and Presentation is the weakest.
+
+<table><tr><td>COARSE</td><td>FINE</td><td>LLM annotation</td></tr><tr><td>Methodology (34.47%)</td><td>Model (26.29%), Method (19.38%), Training (11.77%), Approach (4.62%), Parameter (4.50%)</td><td>Methodology (7.03%), Method (5.00%), Model (2.63%), Training (2.60%), Generalizability/Generalization (2.12%)</td></tr><tr><td>Result (12.47%)</td><td>Result (36.12%), Performance (35.49%), Improvement (14.92%), Accuracy (4.85%), Robustness (2.75%)</td><td>Performance (23.20%), Results (20.06%), Improvement (9.32%), Experimental Results (4.59%), Accuracy (3.17%)</td></tr><tr><td>Data/Task (9.98%)</td><td>Data (53.89%), Task (38.26%), Benchmark (4.75%), Annotation (3.10%)</td><td>Dataset (11.64%), Datasets (9.08%), Tasks (6.14%), Task (2.83%), Data (2.07%)</td></tr><tr><td>Presentation (9.68%)</td><td>Clarity (37.40%), Presentation (19.24%), Figure (14.41%), Table (11.67%), Tipo (6.11%)</td><td>Clarification/Clarity (30.37%), Writing Quality (12.12%), Presentation (5.09%), Figure/Visualization (4.00%), Readability (3.92%)</td></tr><tr><td>Comparison (6.00%)</td><td>Comparison (74.97%), Baseline (25.03%)</td><td>Comparability/Comparison (59.54%), Baselines (12.74%), Baseline (9.72%), Comparisons (5.02%), Model Comparison (1.44%)</td></tr><tr><td>Experiment (5.33%)</td><td>Experiment (100.00%)</td><td>Experiments (50.99%), Experimental Results (10.72%), Experiment (9.38%), Experimental Setup (1.83%), Experimentation (1.69%)</td></tr><tr><td>DDDDEI (4.57%)</td><td>Explanation (29.61%), Discussion (23.36%), Definition (14.23%), Description (13.57%), Detail (11.76%)</td><td>Explainability/Explanation (24.18%), Discussion (16.61%), Description (6.00%), Definition (5.76%), Interpretability/Interpretation (5.26%)</td></tr><tr><td>Related Work (3.93%)</td><td>Related Work (100.00%)</td><td>Reference (20.67%), Related Work (17.99%), References (13.59%), Previous Work (8.23%), Citation (6.70%)</td></tr><tr><td>Evaluation (3.80%)</td><td>Evaluation (73.90%), Metric (26.10%)</td><td>Evaluation (47.33%), Metrics (8.32%), Metric (5.05%), Human Evaluation (3.96%), Empirical Evaluation (2.97%)</td></tr><tr><td>IJMV (2.07%)</td><td>Motivation (53.27%), Justification (17.82%), Validation (15.09%), Intuition (13.82%)</td><td>Motivation (48.91%), Justification (13.45%), Intuition (12.91%), Validity/Validation (9.09%), Motivations (1.82%)</td></tr><tr><td>Analysis (1.99%)</td><td>Analysis (100.00%)</td><td>Analysis/Analytics (61.32%), Theoretical Analysis (4.15%), Error Analysis (3.96%), Qualitative Analysis (2.64%), Empirical Analysis (1.70%)</td></tr><tr><td>Novelty (1.74%)</td><td>Novelty (100.00%)</td><td>Novelty (76.67%), Innovation/Novelty/Originality (8.64%), Technical Novelty (3.67%), Novel Approach (3.02%), Technical novelty (0.86%)</td></tr><tr><td>Contribution (1.49%)</td><td>Contribution (100.00%)</td><td>Contribution (50.51%), Contributions (19.44%), Technical Contribution (4.55%), Main Contribution (3.79%), Core Contribution (1.77%)</td></tr><tr><td>Significance (1.00%)</td><td>Significance (37.45%), Importance (32.21%), Impact (30.34%)</td><td>Significance (35.96%), Importance (25.84%), Impact (20.60%), Problem Importance (3.37%), Performance Impact (2.25%)</td></tr><tr><td>Ablation (0.91%)</td><td>Ablation (100.00%)</td><td>Ablation Study (36.36%), Ablation Studies (23.97%), Ab-lations (12.81%), Ablation (7.44%), Ablation Experiments (4.55%)</td></tr><tr><td>Theory (0.55%)</td><td>Theory (100.00%)</td><td>Theory (18.37%), Theoretical Analysis (14.97%), Theoretical Results (6.12%), Theorem (5.44%), Theoretical results (3.40%)</td></tr></table>
+
+Table 9: The aspects of different levels of granularity and their frequency in the dataset we created. We show the 5 most frequent aspects at each level. For example, Methodology appears in  $35.30\%$  of the review sentences, and the 5 most frequent FINE labels associated with Methodology are Model, Method, Training, Approach, and Parameter. Among all review sentences containing Methodology, the FINE label Model appears in  $24.92\%$  of the cases, and the LLM annotation Methodology appears in  $6.54\%$  of the cases. Return to main text. Return to appendix.  
+
+<table><tr><td>aspect</td><td>metric</td><td>p=p0</td><td>t=θ</td></tr><tr><td rowspan="2">raw</td><td>exact match</td><td>20.34%</td><td>15.31%</td></tr><tr><td>BERTScore ≥ 0.9</td><td>67.14%</td><td>45.50%</td></tr><tr><td rowspan="2">COARSE</td><td>exact match</td><td>60.40%</td><td>53.43%</td></tr><tr><td>Jaccard similarity</td><td>0.7112</td><td>0.6528</td></tr></table>
+
+Table 10: The consistency scores between GPT-4o annotations obtained using different temperature (p=p0, where p0 refers to the prompt in Table 7), and those obtained using different prompts (t=0, where t refers to temperature). We calculated consistency scores using both the raw annotations and those mapped to the COARSE label set. Return to main text. Return to appendix.
+
+<table><tr><td>BERTScore</td><td>annotation 1</td><td>annotation 2</td></tr><tr><td>≥ 0.90</td><td>Technique, Analysis, Task Writing Quality, ClarityExplanation, Settings, ObscurityAnalysis, Tasks, Experiments, FindingsExperiment Setup, Text Classification</td><td>Technique, Analysis, Task Suitability, ImprovementWriting ClarityExplanation, Experimental details, ClarityAnalysis, Experiments, RobustnessExperiment Setup, Recency</td></tr><tr><td>&lt; 0.90</td><td>Lack of definition, ContextClarity, MethodologyModel SpecificationGeneralization Performance, FrameModel Architecture</td><td>Definition, PaperExplanationModel, Structure, ClarityExperiment Hypothesis, GeneralizationEncoder, Decoder, Tensor Product Representation</td></tr></table>
+
+Table 11: How BERTScores for different pairs look like. Return to main text. Return to appendix.  
+
+<table><tr><td>Jaccard similarity</td><td>set 1</td><td>set 2</td></tr><tr><td>0.9091</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,8,9,10,11}</td></tr><tr><td>0.9000</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,8,9}</td></tr><tr><td>0.8182</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,8,9,11}</td></tr><tr><td>0.8000</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,8}</td></tr><tr><td>0.7500</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,8,9,11,12}</td></tr><tr><td>0.6667</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,8,11,12}</td></tr><tr><td>0.5833</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,11,12}</td></tr><tr><td>0.5385</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,7,11,12,13}</td></tr><tr><td>0.4615</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,11,12,13}</td></tr><tr><td>0.4286</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,6,11,12,13,14}</td></tr><tr><td>0.3571</td><td>{1,2,3,4,5,6,7,8,9,10}</td><td>{1,2,3,4,5,11,12,13,14}</td></tr></table>
+
+Table 12: How Jaccard similarity for different pairs look like. Return to main text. Return to appendix.  
+
+<table><tr><td>Method.</td><td>Result</td><td>Data/Task</td><td>Presentation</td><td>Comparison</td><td>DDDDEI</td><td>Experiment</td><td>Related Work</td></tr><tr><td>846</td><td>393</td><td>304</td><td>270</td><td>186</td><td>171</td><td>155</td><td>113</td></tr><tr><td>Analysis</td><td>Evaluation</td><td>IJMV</td><td>Novelty</td><td>Contribution</td><td>Theory</td><td>Ablation</td><td>Significance</td></tr><tr><td>100</td><td>85</td><td>77</td><td>72</td><td>71</td><td>65</td><td>63</td><td>61</td></tr></table>
+
+Table 13: The distribution of aspects in the questionnaire we distributed. Return to appendix.  
+
+<table><tr><td>review</td><td>question</td><td>yes</td><td>no</td><td>explanation</td></tr><tr><td>I am also somewhat confused by the second set of experiments.</td><td>Does the review address Experiment?</td><td>□</td><td>□</td><td></td></tr><tr><td>Proposes a novel approach by combining ideas from active learning and human-AI collaboration.</td><td>Does the review address Methodology?</td><td>□</td><td>□</td><td></td></tr><tr><td>Proposes a novel approach by combining ideas from active learning and human-AI collaboration.</td><td>Does the review address Novelty?</td><td>□</td><td>□</td><td></td></tr></table>
+
+Table 14: Examples of the entries in the questionnaire we distributed. Note that a single review sentence appears multiple times in the questionnaire if it covers multiple aspects. Return to main text. Return to appendix.  
+
+<table><tr><td>Contribution</td><td>Ablation</td><td>Evaluation</td><td>Novelty</td><td>IJMV</td><td>Analysis</td><td>Comparison</td><td>DDDDEI</td></tr><tr><td>0.9953</td><td>0.9947</td><td>0.9843</td><td>0.9769</td><td>0.9740</td><td>0.9733</td><td>0.9695</td><td>0.9688</td></tr><tr><td>Theory</td><td>Experiment</td><td>Related Work</td><td>Data/Task</td><td>Methodology</td><td>Result</td><td>Significance</td><td>Presentation</td></tr><tr><td>0.9641</td><td>0.9505</td><td>0.9204</td><td>0.9145</td><td>0.8928</td><td>0.8617</td><td>0.8361</td><td>0.8225</td></tr></table>
+
+Table 15: The agreement between human and LLM annotations across different aspects. Return to appendix.
+
+# C More on aspect prediction
+
+The dataset is split into  $90\%$  for training and  $10\%$  for testing. We implemented bag-of-words models with random forest (BoW+RF) using scikit-learn. We used RandomForestClassifier, and we set n_estimators=100. For the RoBERTa models, we set batch_size=16 and learning_rate=3e-5. The models were trained for 10 epochs. For GPT40 experiments, we set temperature=0. We set seed=2266 for all the experiments.
+
+The equation for focal loss is given in 3, and we experimented with  $\alpha = [0.1, 0.2, 0.3]$  and  $\gamma = [1.5, 2.0, 2.5]$ .
+
+$$
+\mathcal {L} _ {\text {f o c a l}} (p) = - \alpha_ {t} (1 - p) ^ {\gamma} \log (p). \tag {3}
+$$
+
+Table 16 shows the few-shot prompt. We selected exemplars from the training set.
+
+Identify the aspect(s) that the given sentence focuses on. Aspects: $COARSE/FINE ASPECTS$. If a sentence focuses on none of these aspects, mark it as “-”. Format the output in a json dictionary: {"Aspects": [...]} . Here are some examples:
+
+#  $EXAMPLES$
+
+Table 16: The few-shot prompt used to predict aspects. Return to main text.
+
+Table 17, 18, 19, and 20 show additional results of PAP and RAP. Table 21 shows a classification report of an RAP model.
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.5254</td><td>0.6179</td><td>0.5435</td><td>0.4928</td></tr><tr><td>RoBERTa</td><td>0.5187</td><td>0.6644</td><td>0.5764</td><td>0.5378</td></tr><tr><td>GPT-4o</td><td>0.6980</td><td>0.3892</td><td>0.4455</td><td>0.3319</td></tr></table>
+
+(a) abstract  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.5483</td><td>0.6203</td><td>0.5481</td><td>0.4957</td></tr><tr><td>RoBERTa</td><td>0.5240</td><td>0.6599</td><td>0.5781</td><td>0.5327</td></tr><tr><td>GPT-4o</td><td>0.7324</td><td>0.6840</td><td>0.6561</td><td>0.5246</td></tr></table>
+
+Table 17: The highest precision, recall, F1 score, and Jaccard similarity of predictions made by models trained on PAP using the FINE label set. For GPT-4o, the evaluation of using the abstract as input is in the zero-shot setting, and that of using the title in the few-shot setting. Return to main text.
+
+(b) title  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.8694</td><td>0.9430</td><td>0.8909</td><td>0.7952</td></tr><tr><td>GPT-4o</td><td>0.8580</td><td>0.7215</td><td>0.7452</td><td>0.6813</td></tr></table>
+
+(a) full paper  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.8139</td><td>0.9177</td><td>0.8527</td><td>0.7725</td></tr><tr><td>RoBERTa</td><td>0.8658</td><td>0.9586</td><td>0.9048</td><td>0.8339</td></tr><tr><td>GPT-4o</td><td>0.8700</td><td>0.5633</td><td>0.6392</td><td>0.5589</td></tr></table>
+
+(b) abstract  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.8308</td><td>0.9114</td><td>0.8575</td><td>0.7625</td></tr><tr><td>RoBERTa</td><td>0.8944</td><td>0.9290</td><td>0.8864</td><td>0.8381</td></tr><tr><td>GPT-4o</td><td>0.8835</td><td>0.9684</td><td>0.9200</td><td>0.8471</td></tr></table>
+
+(c) keywords  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.8139</td><td>0.9177</td><td>0.8528</td><td>0.7640</td></tr><tr><td>RoBERTa</td><td>0.8654</td><td>0.9645</td><td>0.9048</td><td>0.8266</td></tr><tr><td>GPT-4o</td><td>0.8913</td><td>0.9494</td><td>0.9171</td><td>0.8448</td></tr></table>
+
+Table 18: The highest precision, recall, F1 score, and Jaccard similarity of predictions made by models trained on PAP using the COARSE label set. For GPT-4o, the evaluation of using the full paper or abstract as input is in the zero-shot setting, and the rest is in the few-shot setting. Return to main text.
+
+(d) title  
+
+<table><tr><td>model</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>BoW+RF</td><td>0.7392</td><td>0.3817</td><td>0.4689</td><td>0.3447</td></tr><tr><td>RoBERTa</td><td>0.7413</td><td>0.7146</td><td>0.7196</td><td>0.6402</td></tr><tr><td>GPT-4o</td><td>0.6426</td><td>0.7092</td><td>0.6309</td><td>0.5217</td></tr></table>
+
+Table 19: The highest precision, recall, F1 score, and Jaccard similarity of model predictions on RAP using the FINE label set. For GPT-4o, the evaluation is in the few-shot setting. Return to main text.
+
+<table><tr><td>label set</td><td>precision</td><td>recall</td><td>f1</td><td>Jaccard</td></tr><tr><td>COARSE</td><td>0.5526</td><td>0.5646</td><td>0.5145</td><td>0.4341</td></tr><tr><td>FINE</td><td>0.6174</td><td>0.6939</td><td>0.6073</td><td>0.5041</td></tr></table>
+
+Table 20: The highest precision, recall, F1 score, and Jaccard similarity of model predictions on RAP using GPT-4o in the zero-shot setting. Return to main text.
+
+<table><tr><td>category</td><td>precision</td><td>recall</td><td>f1</td><td>support</td></tr><tr><td>Ablation</td><td>0.85</td><td>0.92</td><td>0.88</td><td>12</td></tr><tr><td>Analysis</td><td>0.83</td><td>0.77</td><td>0.80</td><td>39</td></tr><tr><td>Comparison</td><td>0.81</td><td>0.74</td><td>0.77</td><td>125</td></tr><tr><td>Contribution</td><td>0.79</td><td>0.87</td><td>0.83</td><td>31</td></tr><tr><td>Data/Task</td><td>0.75</td><td>0.85</td><td>0.79</td><td>212</td></tr><tr><td>DDDDEI</td><td>0.71</td><td>0.71</td><td>0.71</td><td>168</td></tr><tr><td>Evaluation</td><td>0.87</td><td>0.75</td><td>0.81</td><td>114</td></tr><tr><td>Experiment</td><td>0.93</td><td>0.82</td><td>0.87</td><td>99</td></tr><tr><td>IJMV</td><td>0.73</td><td>0.76</td><td>0.74</td><td>46</td></tr><tr><td>Methodology</td><td>0.80</td><td>0.83</td><td>0.81</td><td>602</td></tr><tr><td>Novelty</td><td>0.81</td><td>0.79</td><td>0.80</td><td>28</td></tr><tr><td>Presentation</td><td>0.75</td><td>0.70</td><td>0.73</td><td>279</td></tr><tr><td>Related Work</td><td>0.79</td><td>0.70</td><td>0.74</td><td>150</td></tr><tr><td>Result</td><td>0.82</td><td>0.84</td><td>0.83</td><td>266</td></tr><tr><td>Significance</td><td>0.79</td><td>0.41</td><td>0.54</td><td>27</td></tr><tr><td>Theory</td><td>0.00</td><td>0.00</td><td>0.00</td><td>0</td></tr><tr><td>None</td><td>0.66</td><td>0.69</td><td>0.67</td><td>401</td></tr><tr><td>AVERAGE (w.)</td><td>0.77</td><td>0.77</td><td>0.77</td><td>-</td></tr></table>
+
+Table 21: The classification report of the predictions made by the best performing model on RAP (in terms of F1 score). Weighted average scores (AVERAGE (w.)) are reported. This model was trained using the COARSE label set. Return to main text. Return to appendix.
+
+# D More on practical applications
+
+# D.1 More on aspect analysis
+
+Figure 4 shows the Levenshtein similarity of the set of the 10 most frequent aspects in the submission tracks in EMNLP23. It shows that the 1th to 18th tracks are more similar to each other. The 1th to 18th tracks are: (1) Question Answering, (2) Information Retrieval and Text Mining, (3) Phonology, Morphology, and Word Segmentation, (4) Human-Centered NLP, (5) Machine Learning for NLP, (6) Natural Language Generation, (7) Discourse and Pragmatics, (8) Speech and Multimodality, (9) Summarization, (10) Computational Social Science and Cultural Analytics, (11) Interpretability, Interactivity, and Analysis of Models for NLP, (12) NLP Applications, (13) Linguistic Theories, Cognitive Modeling, and Psycholinguistics, (14) Dialogue and Interactive Systems, (15) Resources and Evaluation, (16) Information Extraction, (17) Language Modeling and Analysis of Language Models, (18) Language Grounding to Vision, Robotics and Beyond.
+
+The 19th to 27th tracks are not so similar to most of the other tracks. The 19th to 27th tracks are: (19) Sentiment Analysis, Stylistic Analysis, and Argument Mining, (20) Syntax, Parsing and their Applications, (21) Commonsense Reasoning, (22) Multilinguality and Linguistic Diversity, (23) Machine Translation, (24) Ethics in NLP, (25) Efficient Methods for NLP, (26) Semantics: Lexical, Sentence level, Document Level, Textual Inference, etc., (27) Theme Track: Large Language Models and the Future of NLP.
+
+![](images/e4f2272ba93fb3e195a85990683d84419d6f4bd07b6e90ace94a37f33f4e6308.jpg)  
+Figure 4: The Levenshtein similarity of the 10 most frequent aspects within each submission track in EMNLP23. Return to main text.
+
+Figure 5 shows the 5 most frequent aspects in each of the submission tracks in EMNLP23.
+
+Table 22 shows the frequency of Analysis and DDDDEI across all the submission tracks in EMNLP23.
+
+# D.2 More on review comparison
+
+Table 23 shows the prompts used to generate reviews used in Section 5.2.
+
+Figure 6 shows more results regrading the comparison of human-written and LLM-generated reviews.
+
+# D.3 More on LLM-generated review detection
+
+We used PyPDF2 to convert PDFs to text files, and we only generate reviews for the main text (cut contents after reference). We used GPT-4o, set temperature  $= 0$ , seed  $= 2266$ , and max_tokens  $= 2048$ .
+
+Table 24 shows the zero-shot performance of GPT-4o on LLM-generated review detection.
+
+![](images/cec7774dd5d1ff983f934a522e8f7edd1a287005be75502c564a381153e671fa.jpg)
+
+![](images/89108e8cec1318a0afe753083475ab3b58147fd726f4b8dca781f0e50cd88c0a.jpg)
+
+![](images/8a6b80acb18ea912361e31a25fc7978af3fc586313f1bd55bca472660b3b3a99.jpg)
+
+![](images/92a50e579ae27a764ec1c38a39524e5f673bfa5b9a68a4e95fa5b652f1aadba5.jpg)
+
+![](images/1650f7de6807e2a5be978076594ce442f50473f6261040a999c0a4c6b9849209.jpg)
+
+![](images/1201ce35910b28de30c56c48fe667767bc71bf43ecb97f3b8d65f8fd6aaa1867.jpg)
+
+![](images/cd8c046a5545ba96ae660130d29061a50b4238fd43b519a0c5a6263d9e32292c.jpg)
+
+![](images/48cff11dde40487334c2451edbf90b562fd7d45a69d15f8b6f324efb41a41f52.jpg)
+
+![](images/e53f7ee9b2f274d2f7e8d79b551423bd8bbcbd857d0538b7f25f235f613a2b92.jpg)
+
+![](images/cd848e2412dac8080b4c23f37ab237dd126bc41c31d938dde5a47d354f06ee99.jpg)
+
+![](images/8348c94d2756c0d8f578ca4d34b65bbdb6b46f7ab050f58f8eea0a8100483a98.jpg)
+
+![](images/60d20c2e44a6f1e3879dad0fd5c0c6d9d1509c0df5893370473a8a6eb722b82f.jpg)
+
+![](images/ffb36b6ddf6b450424aef1617b76926bd92357e8f3ae7879bf5f6a39f2f3ba1c.jpg)
+
+![](images/ec15a9d8c44413cf6b6bfc451f2abcd83443d0f5165d09718121034d4b7d4c2b.jpg)
+
+![](images/9161c3242af85737120209cd28b800dbf03c7d723f81f5f2fcff1dc9616c8a5a.jpg)
+
+![](images/1a49fdc07dd0ba4b88378a463d865cad145ff1e704c2b8015e82a707de38f03e.jpg)
+
+![](images/282d30c9605bfc3b1d8e641d28cf7ff2f716791f6b6bd2e4cdcabce3628af08b.jpg)
+
+![](images/56797a9e0e58aeae3bd65ff7a7b96ec15d93988b36dd136391e8ea4017a0ab3e.jpg)
+
+![](images/d5b7e0fc689b4c56c4b6487205b7b28b2d6a0d3db76a09deedb908583a555402.jpg)
+
+![](images/cadf1ec78a6902a40796a3e9132c87cfd893fafe4f84c7b30b716d73972bb5a8.jpg)
+
+![](images/8a1e13c936bd0a82ecf0bee09e485720482cb83288461dbe58bd6f3fd75efb53.jpg)
+
+![](images/2be336864353fcad95e7828fced1defee86d60c74b5370092f1078a3d4618d92.jpg)
+
+![](images/ab8230297d4be178c707b1b7afe6b3f9d14a0ab57c2763635472ab2abbd4ea21.jpg)
+
+![](images/15c36348a9ebe3f1c412ff712e8280126a8b50f74cbc7c8a9559a71bc115c584.jpg)
+
+![](images/4120f149ef115a09d7efd9e1f0c9726bc25cbc2be518243cd97ff90529e95e2c.jpg)
+
+![](images/2d1eff99c907a35cda3ef8be1dde9c24c4aa69567de0a2ebe2217d0ef7029cf4.jpg)  
+Figure 5: The 5 most frequent aspects in each of the submission tracks in EMNLP23. Return to main text. Return to appendix.
+
+![](images/318424f0639187c4d30b349d517b18affafad20561b012ab69f3158c16229b6f.jpg)
+
+<table><tr><td>track</td><td>review (%)</td></tr><tr><td>Computational Social Science and Cultural Analytics</td><td>69.23</td></tr><tr><td>Linguistic Theories, Cognitive Modeling, and Psycholinguistics</td><td>68.75</td></tr><tr><td>Commonsense Reasoning</td><td>62.62</td></tr><tr><td>Multilinguality and Linguistic Diversity</td><td>60.68</td></tr><tr><td>Machine Learning for NLP</td><td>60.00</td></tr><tr><td>Machine Translation</td><td>57.38</td></tr><tr><td>Discourse and Pragmatics</td><td>56.86</td></tr><tr><td>Phonology, Morphology, and Word Segmentation</td><td>56.67</td></tr><tr><td>Interpretability, Interactivity, and Analysis of Models for NLP</td><td>56.37</td></tr><tr><td>Sentiment Analysis, Stylistic Analysis, and Argument Mining</td><td>55.66</td></tr><tr><td>Theme Track: Large Language Models and the Future of NLP</td><td>55.56</td></tr><tr><td>Information Retrieval and Text Mining</td><td>55.50</td></tr><tr><td>NLP Applications</td><td>54.48</td></tr><tr><td>Summarization</td><td>54.44</td></tr><tr><td>Resources and Evaluation</td><td>54.35</td></tr><tr><td>Efficient Methods for NLP</td><td>53.07</td></tr><tr><td>Information Extraction</td><td>53.06</td></tr><tr><td>Syntax, Parsing and their Applications</td><td>52.94</td></tr><tr><td>Language Modeling and Analysis of Language Models</td><td>51.85</td></tr><tr><td>Dialogue and Interactive Systems</td><td>51.23</td></tr><tr><td>Speech and Multimodality</td><td>51.01</td></tr><tr><td>Ethics in NLP</td><td>50.96</td></tr><tr><td>Semantics: Lexical, Sentence level, Document Level, Textual Inference, etc.</td><td>48.15</td></tr><tr><td>Language Grounding to Vision, Robotics and Beyond</td><td>47.91</td></tr><tr><td>Human-Centered NLP</td><td>47.78</td></tr><tr><td>Question Answering</td><td>45.58</td></tr><tr><td>Natural Language Generation</td><td>42.34</td></tr></table>
+
+(a) Analysis  
+
+<table><tr><td>track</td><td>review (%)</td></tr><tr><td>Ethics in NLP</td><td>65.38</td></tr><tr><td>Linguistic Theories, Cognitive Modeling, and Psycholinguistics</td><td>60.42</td></tr><tr><td>Interpretability, Interactivity, and Analysis of Models for NLP</td><td>59.46</td></tr><tr><td>NLP Applications</td><td>57.17</td></tr><tr><td>Information Extraction</td><td>56.85</td></tr><tr><td>Resources and Evaluation</td><td>56.09</td></tr><tr><td>Semantics: Lexical, Sentence level, Document Level, Textual Inference, etc.</td><td>56.02</td></tr><tr><td>Human-Centered NLP</td><td>54.44</td></tr><tr><td>Machine Learning for NLP</td><td>53.56</td></tr><tr><td>Information Retrieval and Text Mining</td><td>53.5</td></tr><tr><td>Speech and Multimodality</td><td>52.02</td></tr><tr><td>Sentiment Analysis, Stylistic Analysis, and Argument Mining</td><td>51.89</td></tr><tr><td>Language Modeling and Analysis of Language Models</td><td>51.85</td></tr><tr><td>Language Grounding to Vision, Robotics and Beyond</td><td>50.95</td></tr><tr><td>Natural Language Generation</td><td>50.90</td></tr><tr><td>Computational Social Science and Cultural Analytics</td><td>50.30</td></tr><tr><td>Theme Track: Large Language Models and the Future of NLP</td><td>49.90</td></tr><tr><td>Summarization</td><td>49.44</td></tr><tr><td>Multilinguality and Linguistic Diversity</td><td>49.03</td></tr><tr><td>Dialogue and Interactive Systems</td><td>48.77</td></tr><tr><td>Discourse and Pragmatics</td><td>47.06</td></tr><tr><td>Efficient Methods for NLP</td><td>46.93</td></tr><tr><td>Question Answering</td><td>45.94</td></tr><tr><td>Commonsense Reasoning</td><td>42.99</td></tr><tr><td>Machine Translation</td><td>39.34</td></tr><tr><td>Syntax, Parsing and their Applications</td><td>39.22</td></tr><tr><td>Phonology, Morphology, and Word Segmentation</td><td>33.33</td></tr></table>
+
+(b) DDDDEI
+
+Table 22: The frequencies of Analysis and DDDDEI across all the submission tracks in EMNLP23. Return to main text. Return to appendix.
+
+# Du et al. (2024)
+
+As an esteemed reviewer with expertise in the field of Natural Language Processing (NLP), you are asked to write a review for a scientific paper submitted for publication. Please follow the reviewer guidelines provided below to ensure a comprehensive and fair assessment:
+
+Reviewer Guidelines: {review-guidelines}
+
+In your review, you must cover the following aspects, adhering to the outlined guidelines:
+
+Summary of the Paper: Provide a concise summary of the paper, highlighting its main objectives, methodology, results, and conclusions.
+
+Strengths and Weaknesses: Critically analyze the strengths and weaknesses of the paper. Consider the significance of the research question, the robustness of the methodology, and the relevance of the findings.
+
+Clarity, Quality, Novelty, and Reproducibility: Evaluate the paper on its clarity of expression, overall quality of research, novelty of the contributions, and the potential for reproducibility by other researchers.
+
+Summary of the Review: Offer a brief summary of your evaluation, encapsulating your overall impression of the paper.
+
+Correctness: Assess the correctness of the paper's claims; you are only allowed to choose from the following options: {Explanation on different correctness scores}
+
+Technical Novelty and Significance: Rate the technical novelty and significance of the paper's contributions; you are only allowed to choose from the following options: {Explanation on different Technical Novelty and Significance scores}
+
+Empirical Novelty and Significance: Evaluate the empirical contributions; you are only allowed to choose from the following options: {Explanation on different Empirical Novelty and Significance scores}
+
+Flag for Ethics Review: Indicate whether the paper should undergo an ethics review [YES or NO].
+
+Recommendation: Provide your recommendation for the paper; you are only allowed to choose from the following options: {Explanation on different recommendation scores}
+
+Confidence: Rate your confidence level in your assessment; you are only allowed to choose from the following options: {Explanation on different confidence scores}
+
+To assist in crafting your review, here are two examples from reviews of different papers:
+
+## Review Example 1: {review_example_1}
+
+## Review Example 2: {review_example_2}
+
+Follow the instruction above, write a review for the paper below:
+
+# Liang et al. (2024b)
+
+Your task now is to draft a high-quality review outline for the given submission.
+
+$= = = = = =$
+
+Your task:
+
+Compose a high-quality peer review of a paper.
+
+Start by "Review outline":
+
+And then:
+
+"1. Significance and novelty"  
+"2. Potential reasons for acceptance"  
+3. Potential reasons for rejection", List multiple key reasons. For each key reason, use  $**>=2$  sub bullet points\*\* to further clarify and support your arguments in painstaking details. Be as specific and detailed as possible.  
+4. Suggestions for improvement", List multiple key suggestions. Be as specific and detailed as possible.
+
+Be thoughtful and constructive. Write Outlines only.
+
+ours
+
+Write a paper review for the following paper regarding its strengths and weaknesses.
+
+Table 23: The prompts used to generate reviews. Return to main text. Return to appendix.
+
+![](images/598e6597773d48521091e2538a7d0b17c7fc61f713bb48714c9bb42d7f6c5b5d.jpg)  
+(a) Du et al. (2024), COARSE
+
+![](images/2b8bc3065964a294adc292400ec1b838ff116f67f63ff016ee11d2067334508f.jpg)  
+(b) Du et al. (2024), FINE
+
+![](images/caddcb7ac079c0748986018d12371ccd48ba84d7b56c7afb33f24d9b9c67138d.jpg)  
+(c) ours, COARSE (EMNLP23)
+
+![](images/16906744a9048c6f99fcbe3fe90ba761a0eb0fc0c315dc42018d1fa190352873.jpg)  
+(d) ours, FINE (EMNLP23)
+
+![](images/34f298b141d40126ddd50c237ceb28b905bf6d53dfda043f739964026e38abff.jpg)  
+(e) Liang et al. (2024b), COARSE (ICLR24)
+
+![](images/43f9eea8d00727663919312c7edef46176424df2b66bd4534258a42675aa0b05.jpg)  
+(f) Liang et al. (2024b), FINE (ICLR24)
+
+![](images/1159cfce90d4f7f73c8bba5a389a6d492a8c573193a3913f141b76da4d260982.jpg)  
+(g) ours, COARSE (ICLR24)
+
+![](images/188606a09a9aad60f810c788fdb61e9f6f4adaa896f42e44f0ea0acb7943aa94.jpg)  
+(h) ours, FINE (ICLR24)  
+Figure 6: The heatmap of the Jaccard similarity between each pair of the human-written reviews and LLM-generated reviews using Du et al. (2024)'s, Liang et al. (2024b)'s, and our prompt. Return to main text. Return to appendix.
+
+<table><tr><td>dataset</td><td>number of data</td><td>#reviews per paper</td><td>accuracy</td></tr><tr><td>Du et al. (2024)</td><td>20</td><td>4.80</td><td>0.80</td></tr><tr><td>Liang et al. (2024b)</td><td>200</td><td>4.52</td><td>0.75</td></tr><tr><td>ours</td><td>200</td><td>4.52</td><td>0.89</td></tr></table>
+
+Table 24: The accuracy of GPT-4o in zero-shot LLM-generated review detection. Return to main text. Return to appendix.

2025/Identifying Aspects in Peer Reviews/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f425cd6b94097c4771ab123f0116133e06341f5b3adf2f4a59d4b6844cf32a5e
+size 2015990