ACL 2026 Papers — Page 17
Annual Meeting of the Association for Computational Linguistics · 2296 papers
R1-RE: Cross-Domain Relation Extraction with RLVR
Runpeng Dai (University of Maryland), Hongtu Zhu (University of Maryland)
Domain AdaptationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextGraphChain-of-Thought
🎯 What it does: Re-defining the relation extraction task as a reasoning process based on annotation guidelines, and achieving cross-domain reasoning by training a small LLM with RLVR (Reinforcement Learning with Verifiable Rewards).
R2IF: Aligning Reasoning with Decisions via Composite Rewards for Interpretable LLM Function Calling
Aijia Cheng (East China Normal University), Yongxin Zhao (East China Normal University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose the R2IF framework, which enhances the interpretability and accuracy of LLMs through composite rewards that combine joint reasoning and function calls.
RADO: Reasoning Audit-Driven Optimization for Rigorous Reasoning in High-Stakes Domains
Zhijie Tan (Peking University), Tong Mo (Peking University)
OptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextTabularSequentialBiomedical DataBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the RADO framework, which enhances reasoning accuracy and logical rigor in high-risk domains through specialized auditing models and tool calls.
RAG-on-a-Diet: A Reinforcement Learning-Based Dynamic Resource Optimization Framework for RAG
Hongwen Ding (Shenzhen College of International Education), Yizheng Zhao (Nanjing University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a RAG routing framework based on reinforcement learning, which can dynamically select the most suitable and resource-efficient language model at each inference hop.
RAM-SD: Retrieval-Augmented Multi-agent framework for Sarcasm Detection
Ziyang Zhou (Xi'an Jiaotong-Liverpool University), Yangbin Chen (Xi'an Jiaotong-Liverpool University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the RAM-SD framework, which uses retrieval-enhanced multi-agent reasoning for irony detection.
Ranking Reasoning LLMs under Test-Time Scaling
Mohsen Hariri (Case Western Reserve University), Vipin Chaudhary (Case Western Reserve University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper studies the stability and convergence of ranking large language models (LLMs) under test-time scaling conditions, and systematically compares various statistical ranking methods.
RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora
Hanjun Cho (Allganize), Jay-Yoon Lee (Seoul National University)
Data SynthesisRetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Propose the RARE framework and the RedQA benchmark, constructing a multi-hop RAG evaluation dataset under high redundancy corpus by utilizing atomic fact splitting, redundancy tracking, and multi-criteria filtering.
RATE: Reviewer Profiling and Annotation-free Training for Expertise Ranking in Peer Review Systems
Weicong Liu (East China Normal University), Xiang Li (East China Normal University)
Recommendation SystemData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the LR-Bench benchmark and the RATE framework for unsupervised learning, used for reviewer assignment
Re^3: Relevance & Recency Retrieval for Mitigating Temporal Hallucination
Jiawei Cao (University Of Science And Technology Of China), Shijin Wang (University Of Science And Technology Of China)
GenerationData SynthesisRetrievalTransformerLarge Language ModelContrastive LearningTextTabularTime SeriesBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the Re3 framework, which reduces temporal illusion in retrieval-augmented generation through time-aware dual correlation encoding and conflict-aware temporal filtering.
ReActR: Reasoning through Error-Activated Reflection for LLM Post-Training
Lina Sun (Shanghai University of Engineering Science)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose the ReActR framework, which trains LLMs by reflecting on and correcting erroneous reasoning trajectories.
Read As Human: Compressing Context via Parallelizable Close Reading and Skimming
Jiwei Tang (Tsinghua University), Bo Zheng (Alibaba)
CompressionComputational EfficiencyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a context compression framework named RAM (Read As Human), which adopts an adaptive hybrid reading strategy to address the issues of low computational efficiency and redundant information in long context scenarios.
Ready Jurist One: Benchmarking Language Agents for Legal Intelligence in Dynamic Environments
Zheng Jia (Fudan University), Zhongyu Wei
TransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the first dynamic interactive legal environment J1-ENVS and a dual-indicator evaluation framework J1-EVAL to systematically evaluate LLM-driven legal intelligent agents.
REAL: REtrieval-reAsoning and Logic-constructed Attention Behaviors for Long-Context KV Cache Compression
Mengjie Li (Yonsei University), William J. Song (Yonsei University)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a KV cache compression method based on multi-behavior attention heads (REAL), which dynamically allocates KV cache budget by performing fine-grained analysis of four attention behaviors: retrieval inference, bias, and interference.
RealChart2Code: Bridging the Gap in Real-World Chart-to-Code Generation via Multi-Task Evaluation
Jiajun Zhang (University of Science and Technology of China), Liang Wang (University of Chinese Academy of Sciences)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the RealChart2Code benchmark, which focuses on generating Matplotlib code from real, multi-panel visualizations;
RealTalk-CN: A Realistic Chinese Speech Task-Oriented Dialogue Benchmark with Cross-Modal Analysis
Enzhi Wang (Nankai University), Yong Qin (Nankai University)
ClassificationRecognitionAnomaly DetectionData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: Proposed RealTalk-CN, a Chinese task-oriented dialogue benchmark containing 5.4k turns, 150 hours, and approximately 60k speech-text interactions, covering 58 multi-domain scenarios, fine-grained dialogue states, intents, slot values, and 10% spontaneous oral distortions and speaker feature annotations.
REaR : Retrieve, Expand and Refine for Effective Multitable Retrieval
Rishita Agarwal (Indian Institute of Technology Guwahati), Vivek Gupta (Arizona State University)
RetrievalComputational EfficiencyTransformerPrompt EngineeringContrastive LearningTextTabularRetrieval-Augmented Generation
🎯 What it does: Proposed a three-stage, LLM-free multi-table retrieval framework called REAR (Retrieve, Expand, Refine), aimed at improving the quality of multi-table retrieval for natural language queries in relational databases.
ReasMark: A Robust Watermark for Attributing LLM Reasoning Under Knowledge Distillation Attacks
Peizhuo Lv (Nanyang Technological University), Yuling Liu (Chinese Academy of Sciences)
Safty and PrivacyExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a robust watermarking method called REASMARK, which enables attribution verification of the reasoning behavior of large language models (LLMs) even after knowledge distillation attacks.
ReasonAny: Incorporating Reasoning Capability to Any Model via Simple and Effective Model Merging
Junyao Yang (Shanghai Artificial Intelligence Laboratory), Dongrui Liu (Shanghai Artificial Intelligence Laboratory)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerMixture of ExpertsContrastive LearningTextTabularSequentialBenchmarkFinance RelatedChain-of-Thought
🎯 What it does: Designed an untrained model merging framework called ReasonAny, which combines the long-chain reasoning capability with domain-specific models to obtain a single Reasoning+X model.
ReasonEmbed: Enhanced Text Embeddings for Reasoning-Intensive Document Retrieval
Jianlyu Chen (University of Science and Technology of China), Zheng Liu (Beijing Academy of Artificial Intelligence)
RetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose ReasonEmbed, a text embedding model designed for reasoning-intensive document retrieval;
Reasoning Fails Where Step Flow Breaks
Xiaoyu Xu (Shanghai Jiao Tong University), Xiaofeng Zhang (Shanghai Jiao Tong University)
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the StepSaliency diagnostic method, which aggregates token-level attention gradients to the step level, revealing two information loss patterns: shallow locking and deep attenuation. Based on this, the paper designs StepFlow intervention during testing to repair information loss in large models during long-chain reasoning.
Reasoning Gets Harder for LLMs Inside A Dialogue
Ivan Kartáč (Charles University), Ondrej Dusek (Charles University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a dynamic benchmark called BOULDER to evaluate the reasoning capabilities of large language models (LLMs) in task-oriented dialogue (TOD) environments, and compared the performance of reasoning tasks conducted in isolation versus those embedded within dialogues.
Reasoning Hijacking: The Fragility of Reasoning Alignment in Large Language Models
Yuansen Liu (National University of Singapore), Anthony Kum Hoe Tung (National University of Singapore)
ClassificationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the 'reasoning hijacking' attack, which deceives the LLM's reasoning process by injecting fake decision criteria without altering the task objective.
Reasoning Is Not All You Need: Examining LLMs for Multi-Turn Mental Health Conversations
Mohit Chandra (Georgia Institute of Technology), Munmun De Choudhury (Georgia Institute of Technology)
GenerationData SynthesisRecommendation SystemExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBiomedical DataElectronic Health RecordsReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the MedAgent framework to generate multi-turn mental health conversations, construct the MHSD dataset, and design the MultiSenseEval evaluation framework to systematically assess the performance of LLMs in multi-turn mental health dialogues.
Reasoning over Precedents Alongside Statutes: Case-Augmented Deliberative Alignment for LLM Safety
Can Jin (Rutgers University), Dimitris N. Metaxas (Rutgers University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented a case-based incremental reasoning method for safety alignment called CADA, which guides large language models to make safe judgments during the reasoning phase using minimized safety specifications and self-generated cases, and evaluated on multiple safety and practicality benchmarks.
Reasoning Over Space: Enabling Geographic Reasoning for LLM-Based Generative Next POI Recommendation
Dongyi Lv (School of Software Engineering, Xi'an Jiaotong University), Mu Xu (Amap, Alibaba Group)
Recommendation SystemTransformerLarge Language ModelReinforcement LearningTextTabularChain-of-Thought
🎯 What it does: Proposed the ROS framework, which treats geographic information as a decision variable, and utilizes hierarchical spatial semantic ID and a three-stage mobile chain-of-thought to achieve generative next POI recommendation.
Reasoning Structure Matters for Safety Alignment of Reasoning Models
Yeonjun In (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose ALTTRAIN by rewriting the inference structure of large reasoning models to achieve safe alignment
Reasoning Traces Shape Outputs but Models Won’t Say So
Yijie Hao (Emory University), Joyce C. Ho
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The study investigates whether the reasoning trajectories of large reasoning models (LRMs) genuinely affect model outputs and evaluates whether models will truthfully report the impact after being injected with reasoning. It proposes the THOUGHT INJECTION intervention method.
Reasoning While Asking: Transforming Reasoning Large Language Models from Passive Solvers to Proactive Inquirers
Xin Chen (Nanjing University), Shujian Huang (Nanjing University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Proactive Interactive Reasoning (PIR) framework, enabling large language models to actively ask users questions during the reasoning process and integrate feedback, overcoming the problem of blind self-thinking.
Reasoning with OmniThought: A Large CoT Dataset with Verbosity and Cognitive Difficulty Annotations
Wenrui Cai (Shanghai Jiao Tong University), Xiangzhong Fang (Alibaba Group)
Data-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Created a large-scale Chain-of-Thought (CoT) dataset called OmniThought, and trained and released a new large-scale reasoning model based on this dataset.
Reasoning with Ontology Graph: Toward Type-Constrained Knowledge Graph Question Answering
Yongxue Shan (National University of Defense Technology), Xiaodong Wang (National University of Defense Technology)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a relation-centric ontology graph based on Freebase, and implemented the Planner-Judge-Generative Backoff framework on this graph for multi-hop knowledge graph question answering.
ReasoningGuard: Safeguarding Large Reasoning Models with Inference-time Safety Aha Moments
Yuquan Wang (Fudan University), Min Yang (Shanghai University of Electric Power)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes ReasoningGuard, an inference-time safeguard mechanism based on internal attention monitoring during the reasoning process and the injection of secure Aha moments, aiming to defend large reasoning models (LRM) against jailbreak attacks.
ReasonRank: Empowering Passage Ranking with Strong Reasoning Ability
Wenhan Liu (Renmin University of China), Zhicheng Dou (Renmin University of China)
Data SynthesisRetrievalRecommendation SystemComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose ReasonRank, a list-wise retrieval result re-ranking model with strong reasoning capabilities, and improve its performance through automated data synthesis and two-stage training.
ReCode: Reinforcing Code Generation with Reasoning-Process Rewards
Lishui Fan (Zhejiang University), Zhongxin Liu (Zhejiang University)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the ReCode framework, which combines reinforcement learning to supervise and optimize the reasoning quality during the code generation process.
ReContraster: Making Your Posters Stand Out with Regional Contrast
Peixuan Zhang (Beijing University of Posts and Telecommunications), Boxin Shi (Peking University)
Image TranslationImage HarmonizationGenerationOptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose an untrained multi-agent framework, ReContraster, for generating posters with regional contrast, layout harmony, and smooth boundaries.
ReCoQA: A Benchmark for Tool-Augmented and Multi-Step Reasoning in Real Estate Question and Answering
Yindong Zhang (Hong Kong Baptist University), Weijia Jia (Beijing Normal-Hong Kong Baptist University)
TransformerLarge Language ModelAgentic AITextTabularBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the ReCoQA dataset and the HIRE-Agent baseline for evaluating tool-enhanced, multi-step reasoning QA in real estate decision-making scenarios;
ReCreate: Reasoning and Creating Domain Agents Driven by Experience
Zhezheng Hao (Zhejiang University), Jiawei Chen (Zhejiang University)
Autonomous DrivingOptimizationDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes an experience-driven framework called ReCreate, which can automatically build and optimize domain-specific LLM agents from scratch, using the interaction history, execution logs, and evaluation results of the agent to update the agent scaffold;
Rectifying the Emotional Flow: Aligning Priors and Dynamic Guidance for High-Arousal Text-to-Speech
Fangming Feng (Zhejiang University), Tao Jin (Zhejiang University)
GenerationTransformerDiffusion modelFlow-based ModelRectified FlowTextAudio
🎯 What it does: A model-free retraining framework is proposed by introducing an emotion correction noise prior (ERNP) and likelihood inverse guided (LIG) dynamic guidance during the inference stage, aiming to address the stability and expressiveness challenges in high-intensity emotion text-to-speech.
Red Teaming Large Reasoning Models
Jiawei Chen (East China Normal University), Zhaoxia Yin (East China Normal University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposes a unified Rt-LRM benchmark for evaluating the accuracy, security, and inference efficiency of large reasoning models (LRMs), covering 30 carefully designed tasks.
Red-Bandit: Test-Time Adaptation for LLM Red-Teaming via Bandit-Guided LoRA Experts
Christos Ziakas (Imperial College London), Alessandra Russo (Imperial College London)
Safty and PrivacyAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose a black-box red teaming method called Red-Bandit, which dynamically selects multiple attack-style LoRA experts using a multi-armed bandit during testing to generate prompts that can cause large language models to produce unsafe outputs.
RedCoder: Automated Multi-Turn Red Teaming for Code LLMs
Wenjie Jacky Mo (University of California Davis), Muhao Chen (University of California Davis)
Safty and PrivacyAdversarial AttackAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextRetrieval-Augmented Generation
🎯 What it does: This paper proposes and implements an automatic multi-round red team agent called REDCODER, which can interactively converse with code generation large language models (Code LLMs) to induce them to generate code with security vulnerabilities.
Reducing Token Redundancy in LVLMs: A Systematic Review of Token Pruning Methods
Hanzhang Yuan (University of Virginia), Sheng Li (University of Virginia)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageVideoTextMultimodalityReview/Survey Paper
🎯 What it does: Systematically summarizes the visual token pruning techniques in large vision-language models (LVLM), proposes three major paradigms for pruning positions (vision-side, LLM-side, hybrid), and classifies and comprehensively evaluates the judgment criteria and strategies of each method.
ReEfBench: Quantifying the Reasoning Efficiency of LLMs
Zhizhang Fu (Westlake University), Yue Zhang (Westlake University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Evaluated the reasoning efficiency of LLMs and constructed a neuro-symbolic evaluation framework based on First-Order Logic called ReEfBench.
ReEx-SQL: Reasoning with Execution-Aware Reinforcement Learning for Text-to-SQL
Yaxun Dai (Soochow University), Wenhao Jiang (Guangdong Laboratory of Artificial Intelligence and Digital Economy)
AI Code AssistantTransformerReinforcement LearningPrompt EngineeringTextTabularChain-of-Thought
🎯 What it does: Propose a Text-to-SQL framework called ReEx-SQL that can dynamically correct the query statements during the generation process by real-time acquisition and utilization of SQL execution feedback;
ReFEree: Reference-Free and Fine-Grained Method for Evaluating Factual Consistency in Real-World Code Summarization
Suyoung Bae (Sungkyunkwan University), Jee-Hyong Lee (Sungkyunkwan University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a no-reference, fine-grained evaluation method called ReFEree to measure the factual consistency of code summaries generated by LLMs in real-world code.
Reference Attack: A New Cross-Modal Jailbreaking Attack against Multimodal Large Language Models
Yulong Wang (Beijing University of Posts and Telecommunications), Jiayi Gao (Beijing University of Posts and Telecommunications)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelFlow-based ModelRectified FlowGenerative Adversarial NetworkImageTextTabular
🎯 What it does: Propose a novel cross-modal jailbreaking method called Reference Attack, which embeds malicious text into non-text modalities (images, spreadsheets) using recursive symbolic references, inducing multi-modal large language models (MLLMs) to progressively parse and generate违规 content.
Refining and Reusing Annotation Guidelines for LLM Annotation
Kon Woo Kim (Graduate University for Advanced Studies SOKENDAI), Akiko Aizawa (Graduate University for Advanced Studies SOKENDAI)
RecognitionTransformerLarge Language ModelPrompt EngineeringTextBiomedical Data
🎯 What it does: This paper proposes an iterative review framework to reuse and refine human-made annotation guidelines to guide large language models (LLMs) in text annotation.
ReFL: Reflective Feedback Learning for Hallucination Detection of Large Language Models
Cunhang Fan (Anhui University), Zhengqi Wen (Tsinghua University)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Propose the ReFL framework, which utilizes corrective in-context learning to enable self-feedback in LLMs, dynamically reshaping internal representations to enhance hallucination detection capabilities.
ReflectRM: Boosting Generative Reward Models via Self-Reflection within a Unified Judgment Framework
Kai Qin (Tsinghua University), Daiting Shi (Baidu Inc)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a unified generative reward model called ReflectRM, which significantly improves the accuracy and consistency of preference judgment through self-reflection evaluation of the analytical process and utilizes high-confidence anchors and self-reflection voting in two-stage reasoning.
REFLEX: Self-Refining Explainable Fact-Checking via Verdict-Anchored Style Control
Chuyi Kong (Hong Kong Baptist University), Yuxi Sun (Hong Kong Baptist University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a self-improving explanatory fact-checking framework called REFLEX, which enhances the credibility and accuracy of explanations by anchoring the explanatory style to the judgment results.
Reframing Responsibility: Framing-Aware Event Causality Identification
Jin Zhao (Brandeis University), Nianwen Xue (Brandeis University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed and implemented the Frame-based Event Causality Identification (FRECI) task, constructed a multilingual framed causality dataset, and designed a joint neural model to extract information such as responsibility, source, and modality.
REG: Retrieval via Emotion Similarity for Guiding Empathetic Dialogue Generation
Xu Wang (Tianjin University), Yuexian Hou (China Mobile Communication Group Tianjin Co., Ltd.)
GenerationRetrievalKnowledge DistillationTransformerContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the REG framework, which achieves more empathetic dialogue generation by guiding retrieval with emotional attributes and combining it with a generator.
ReGATE: Learning Faster and Better with Fewer Tokens in MLLMs
Chaoyu Li (Arizona State University), Pooyan Fazli (Arizona State University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelVision Language ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes an adaptive token pruning framework called REGATE, which dynamically removes low-importance tokens during the training of multi-modal large language models;
Region-Grounded Report Generation for 3D Medical Imaging: A Fine-Grained Dataset and Graph-Enhanced Framework
Cong Huy Nguyen (AI4LIFE Hanoi University of Science and Technology), Noel Crespi (SAMOVAR Télécom SudParis Institut Polytechnique de Paris)
GenerationGraph Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextBiomedical DataComputed TomographyPositron Emission Tomography
🎯 What it does: This paper addresses the automatic report generation for 3D PET/CT medical images, constructing a fine-grained RoI annotated dataset called VietPET-RoI, and proposes the HiRRA framework that simulates the diagnostic process of radiologists.
Reinforced Efficient Reasoning via Semantically Diverse Exploration
Ziqi Zhao (Shandong University), Xin Xin (Shandong University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the ROSE framework, which improves the accuracy and reasoning efficiency of LLMs in mathematical reasoning tasks through semantic diversification exploration and length-aware advantage estimation.
Reinforced Informativeness Optimization for Long-Form Retrieval-Augmented Generation
Yuhao Wang (Renmin University of China), Haifeng Wang (Baidu Inc)
RetrievalOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented RioRAG, a reinforcement learning-based long-text retrieval-augmented generation framework that generates fact-rich long answers by reinforcing information coverage.
Reinforcement Learning for Diffusion LLMs via Energy-Based Gibbs Alignment
Yijia Fan (Sun Yat-sen University), Jusheng Zhang (Sun Yat-sen University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelContrastive LearningText
🎯 What it does: This paper proposes a new Diffusion-Gibbs Alignment (DGA) method, which reformulates the reinforcement learning alignment problem of diffusion-based large language models as an energy-based variational distribution matching, thereby avoiding the dependence on intractable marginal likelihoods.
Reinforcement Learning for Self-Improving Agent with Skill Library
Jiongxiao Wang (University of Wisconsin Madison), Lin Lee Cheong (Aws Agentic Ai)
Autonomous DrivingOptimizationAI Code AssistantTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a self-improving agent framework called SAGE based on reinforcement learning, which utilizes a skill library to achieve knowledge transfer and reuse;
Reinforcement Learning on Pre-Training Data
Siheng Li (Tencent), Di Wang (Tencent)
Representation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a method called RLPT that directly applies reinforcement learning on pre-training data. The model predicts the next semantic segment given a context and receives a reward based on the similarity between the predicted and actual segments, achieving self-supervised reinforcement learning without manual annotations.
Reinforcement Learning–Guided Adaptive Tuning for Out-of-Distribution Harmful Text Detection
Mengyu Xiang (State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences), Daniel Dajun Zeng (State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences)
ClassificationDomain AdaptationReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes an adaptive tuning framework based on reinforcement learning, RLAT, which dynamically adjusts model parameters during the testing phase to improve the performance of harmful text detection across platforms and time points.
Reinforcing Agentic Search Via Reward Density Optimization
Kun Luo (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Kang Liu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the InfoFlow framework, combining a dual-agent architecture, subgoal reward shaping, and path prompting to enhance reward density in reinforcement learning, achieving efficient LLM agent search and reasoning.
Reliable Evaluation Protocol for Low-Precision Retrieval
Kisu Yang (VAIV Company), Heuiseok Lim (Korea University)
RetrievalScore-based ModelContrastive LearningTextMultimodalityBenchmark
🎯 What it does: This paper proposes a reliable low-precision retrieval evaluation protocol, combining high-precision scoring (HPS) and association evaluation metrics (TRM), to eliminate false ties and provide measurements of expected scores, ranges, and biases.
Reliable Use of Lemmas via Eligibility Reasoning and Section-Aware Reinforcement Learning
Zhikun Xu (Arizona State University), Zicheng Liu (Advanced Micro Devices, Inc)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Studied the phenomenon of large models misusing lemmas in mathematical reasoning, and proposed the RULES framework, which enhances the model's ability to judge the applicability of lemmas through two-stage preconditions and conclusion checks, as well as paragraph-aware reinforcement learning.
ReLook: Vision-Grounded RL with a Multimodal LLM Critic for Agentic Web Coding
Yuhang Li (Tencent), Bo Zhou (Tencent)
OptimizationKnowledge DistillationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes the ReLook framework, which enables frontend code generation to achieve a generation-diagnosis-improvement cycle through a multi-modal LLM as a visual reviewer, enhancing visual consistency and interactive correctness.
REMIND: Memorization and Unlearning in LLMs Through the Lens of Input Loss Landscapes
Liran Cohen (Technion), Avi Mendelson (Technion)
Federated LearningSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This study investigates the knowledge residue phenomenon of large language models (LLMs) during the unlearning process, and proposes a geometric diagnostic framework called REMIND based on the input loss landscape (ILL), which is used to distinguish retained, forgotten, and unseen samples.
Render-of-Thought: Rendering Textual Chain-of-Thought as Images for Visual Latent Reasoning
Yifan Wang (Tencent), Yang Tang (Tencent)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Render the textual steps in the Chain-of-Thought (CoT) reasoning process into images and perform reasoning in the visual space, thereby achieving compression and visualization of the thought chain.
Rendering Data Unlearnable by Exploiting LLM Alignment Mechanisms
Ruihan Zhang (Singapore Management University), Jun Sun (Singapore Management University)
Safty and PrivacyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the Disclaimer Injection method, which injects alignment triggers (disclaimers) into the training data to make text unlearnable during the fine-tuning process.
Repeated Sequences Reveal Gaps between Large Language Models and Natural Language
Kumiko Tanaka-Ishii (Waseda University)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Propose a long-range structural evaluation framework based on the distribution of repeated subsequences, associating it with higher-order Rényi entropy to quantify the entropy growth of text at finite lengths;
RepoGenesis: Benchmarking End-to-End Microservice Generation from Readme to Repository
Zhiyuan Peng (Shanghai Jiao Tong University), Dongmei Zhang (Microsoft)
GenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented RepoGenesis, a multilingual end-to-end microservice generation benchmark that generates fully deployable code repositories from natural language requirements;
ReportLogic: Evaluating Logical Quality in Deep Research Reports
Jujia Zhao (Leiden University), Zhaochun Ren (Leiden University)
Knowledge DistillationTransformerReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the ReportLogic benchmark to evaluate the logical quality of deep research reports and construct human-annotated data
Representation Interventions Enable Lifelong Knowledge Memory Control in LLMs
Xuyuan Liu (Dartmouth College), Zhengzhang Chen (NEC Laboratories America)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Achieving lifelong knowledge updating and control in large language models through controllable interventions in the model's hidden layer space.
ReProbe: Efficient Test-Time Scaling of Multi-Step Reasoning by Probing Internal States of Large Language Models
Jingwei Ni (ETH Zürich), Mrinmaya Sachan (ETH Zürich)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Designed and implemented ReProbe, a lightweight step-level verifier based on the internal states of large language models, used to detect the correctness of each step in real-time during inference, and supports Test-Time Scaling (such as Best-of-N and Beam Search).
ReproEvalCard: A Reporting Standard for Reproducible Evaluation of LLM Pipelines
Priyaranjan Pattnayak (Oracle America Inc), Apoorv Bhatia (Oracle Canada)
TransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically evaluates the reproducibility of 55 large language model (LLM) multi-stage pipeline papers published between 2022 and 2025, identifying missing key Artifacts such as random control and execution traces. It proposes a ReproEvalCard reporting standard and checklist based on evidence, aiming to improve the reproducibility and comparability of evaluation execution.
ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning
Jiani Huang (Hong Kong Polytechnic University), Li Qing (Hong Kong Polytechnic University)
Recommendation SystemReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextGraphTabularRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a ReRec LLM recommendation assistant based on reinforcement learning fine-tuning, which can perform multi-step reasoning and provide recommendation results in single-round complex queries;
Resolving the Security-Auditability Dilemma with Auditable Latent Chain-of-Thought Alignment
Guan Wang (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering, Chinese Academy of Sciences)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: In large model safety alignment, the Auditable Latent CoT Alignment (ALCA) framework is proposed, transferring secure reasoning from visible Chain-of-Thought (CoT) text to a continuous latent space, and ensuring auditability through a self-decoding mechanism; a detector determines when to enter the latent reasoning mode.
Resonating with RoPE: Spectral Quantization for High-Fidelity Key Cache Compression
Xuefei Wang (Shandong University), Weili Guan (Harbin Institute Of Technology)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringAuto EncoderTextBenchmark
🎯 What it does: Propose SpectrumQuant, which compresses the Key cache generated by RoPE's periodic oscillations in the frequency domain using DCT, to address the KV cache quantization bottleneck.
RespiraMFM: A Multimodal Foundation Model with Contrastive Audio-Language Alignment for Respiratory Disease Identification
Shakhrul Iman Siam (Ohio State University), Mi Zhang (Ohio State University)
ClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityElectronic Health RecordsBenchmarkAudio
🎯 What it does: Proposes RespiraMFM, a multi-modal foundation model that integrates respiratory sounds with patient symptoms/history, achieving semantic alignment between audio and text through contrastive learning, thereby improving diagnostic accuracy for respiratory diseases (COVID-19, TB, COPD, asthma, pneumonia).
Response-G1: Explicit Scene Graph Modeling for Proactive Streaming Video Understanding
Ke Ma (Northwestern Polytechnical University), Yunhao Liu (Tsinghua University)
RetrievalExplainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringVision Language ModelVideoTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Response-G1 framework to achieve active streaming video understanding by explicitly modeling and retrieving scene graphs to determine when to answer.
Responsible Evaluation of AI for Mental Health
Hiba Arnaout (Technische Universität Darmstadt), Iryna Gurevych (Technische Universität Darmstadt)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTextBiomedical DataElectronic Health RecordsReview/Survey PaperBenchmarkAudio
🎯 What it does: Propose an interdisciplinary framework and taxonomy for responsibly evaluating AI tools in the field of mental health.
REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once
Zhuoshi Pan (Tsinghua University), Lijun Wu (Shanghai Artificial Intelligence Laboratory)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the REST framework, which conducts parallel multi-question stress testing on large-scale reasoning models (LRM) by concatenating multiple questions into a single prompt;
Rethinking Composed Image Retrieval Evaluation: A Fine-Grained Benchmark from Image Editing
Tingyu Song (CASIA), Shu Wu (CASIA)
Data SynthesisRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposes EDIR, a compositional image retrieval (CIR) benchmark that synthesizes fine-grained queries through image editing, containing 5,000 high-quality queries and 178,645 images.
Rethinking Data Mixing from the Perspective of Large Language Models
Yuanjian Xu (Hong Kong University of Science and Technology (Guangzhou)), Guang Zhang (Hong Kong University of Science and Technology (Guangzhou))
Domain AdaptationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposes DoGraph, a dynamic domain reweighting framework based on graph constraints, which dynamically adjusts domain sampling weights during the pre-training of large-scale language models according to the model's own gradients;
Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning
Zhenguang Wang (Northeastern University), Osmar Zaiane (University of Alberta)
ClassificationExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkTransformerSupervised Fine-TuningContrastive LearningTextBiomedical DataElectronic Health Records
🎯 What it does: Propose the MTSP (Multi-Task Subscore Prediction) framework, which uses multi-task learning to simultaneously predict subscale scores of depression scales from clinical interview texts, achieving fine-grained depression assessment.
Rethinking Entropy Interventions in RLVR: An Entropy Change Perspective
Zhezheng Hao (Zhejiang University), Jiawei Chen (Zhejiang University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Analyze and solve the entropy collapse problem in RLVR training, proposing an adaptive weighting method based on token entropy changes called STEER;
Rethinking Evaluation for LLM Hallucination Detection: A Desiderata, A New RAG-based Benchmark, New Insights
Wenbo Chen (Amazon), Leman Akoglu (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose seven requirements for LLM hallucination detection benchmarks, construct the first TRIVIA+ benchmark with long context, natural hallucinations, multiple models, multiple domains, manual verification, and multiple noisy labels, and experiment on existing detection methods.
Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning
Bowen Ding (Zhejiang University), Tao Lin (Westlake University)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Proposes the Plasticity-Ceiling framework, systematically decomposing and analyzing the utilization of expert trajectories in post-training of LLMs; under this framework, it compares various paradigms such as synchronous and sequential SFT-RL, pure RL, etc., determining sequential SFT-then-RL as the optimal strategy, and provides insights into the impact of the best SFT transition timing (Stable phase) and data attributes (scale, difficulty) on the final performance ceiling.
Rethinking Jailbreak Detection of Large Vision Language Models with Representational Contrastive Scoring
Peichun Hua (Washington University in St. Louis), Ning Zhang (Washington University in St. Louis)
Anomaly DetectionSafty and PrivacyComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Proposes a contrastive scoring framework (Representational Contrastive Scoring, RCS) that leverages internal representations of large vision-language models to detect multi-modal 'jailbreak' attacks, balancing generalization and efficiency;
Rethinking Meeting Effectiveness: A Benchmark and Framework for Temporal Fine-grained Automatic Meeting Effectiveness Evaluation
Yihang Li (Kyoto University), Chenhui Chu (Kyoto University)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Redefine meeting effectiveness by quantifying it as the ratio of goal achievement to time cost, propose a fine-grained segment evaluation based on topics, and construct the AMI-ME meta-evaluation dataset and an LLM judgment framework.
Rethinking Reasoning-Intensive Retrieval: Evaluating and Advancing Retrievers in Agentic Search Systems
Yilun Zhao (Yale NLP Lab), Arman Cohan (Yale NLP Lab)
Data SynthesisRetrievalOptimizationComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningAgentic AIPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the BRIGHT-PRO evaluation framework and RTriever-Synth synthetic corpus, and trained the RTriever-4B retriever, evaluating its performance in agentic search.
Rethinking Sample Polarity in Reinforcement Learning with Verifiable Rewards
Xinyu Tang (Renmin University of China), Jun Zhou (Ant Group)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: This paper systematically analyzes the roles of positive and negative samples in RLVR (Verifiable Reinforcement Learning with Rewards), and based on this, proposes an adaptive asymmetric token-level advantage shaping method called A3PO, significantly improving the reasoning performance of large language models.
Rethinking Table Pruning in TableQA: From Sequential Revisions to Gold Trajectory-Supervised Parallel Search
Yu Guo (University of Science and Technology of China), Xiaobin Tan (University of Science and Technology of China)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularBenchmark
🎯 What it does: Propose the TabTrim framework, achieving a transition from sequential revision to parallel search under the supervision of gold-standard trajectory for table trimming;
Rethinking the Idiomaticity Decomposability Hypothesis: Evidence from Distributional Learning
Maggie Mi (University of Sheffield), Nafise Sadat Moosavi (University of Sheffield)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Investigated the decomposability of idioms, re-examining the IDH hypothesis from a distributed learning perspective using contextualized language models.
Rethinking the Role of Entropy in Optimizing Tool-Use Behaviors for Large Language Model Agents
Zeping Li, Zhenfei Yin (Oxford University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Study entropy reduction as an intrinsic signal for tool call quality in large language model agents, and propose two reward designs to optimize tool usage behavior based on this.
ReTraceQA: Evaluating Reasoning Traces of Small Language Models in Commonsense Question Answering
Francesco Maria Molfese (Sapienza University of Rome), Roberto Navigli (Sapienza University of Rome)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the RETRACEQA benchmark, manually annotate the reasoning trajectories and error locations of small language models on commonsense reasoning tasks to evaluate the effectiveness of their reasoning process;
ReTRE: Benchmarking LLM Transfer Robustness with Structure-Preserving Variants
ZhongDong Li (Soochow University), Jiajie Xu (Soochow University)
Domain AdaptationKnowledge DistillationRepresentation LearningRobotic IntelligenceMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingImageTextMultimodalityTabularBenchmarkRetrieval-Augmented GenerationStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose the ReTRE benchmark, evaluating the robustness of LLMs in near and far transfer across knowledge domains and modal contexts through structure-preserving task rewriting;
Retrieval as Generation: A Unified Framework with Self-Triggered Information Planning
Bo Li (Peking University), Wei Ye (Hebei University of Technology)
GenerationRetrievalRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose GRIP, a Retrieval-as-Generation framework that integrates retrieval behavior into the generation process, achieving retrieval triggering, query reconstruction, and termination decision-making through token-level control tokens, supporting multi-step reasoning.
Retrieval Heads are Dynamic
Yuping Lin (Michigan State University), Jiliang Tang (Michigan State University)
RetrievalExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Study retrieval heads in large language models (LLMs) from a dynamic perspective, systematically analyzing their behavior, and verifying the dynamics, irreplaceability, and correlation with hidden states of retrieval heads on two tasks: Needle-in-a-Haystack (NIAH) and HotpotQA; subsequently, apply dynamic retrieval heads to a dynamic retrieval-augmented generation (Dynamic RAG) framework, evaluating their impact on question-answering performance.
Retrieval-Augmented Defense: Adaptive and Controllable Jailbreak Prevention for Large Language Models
Guangyu Yang (University of Cambridge), Bill Byrne (University of Cambridge)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-enhanced defense framework called RAD, which uses known jailbreak examples for retrieval, re-ranking, extraction, classification, and voting to detect and block jailbreak attacks on LLMs in real time.
Retrievals Can Be Detrimental: Unveiling the Backdoor Vulnerability of Retrieval-Augmented Diffusion Models
Hao Fang (Tsinghua University), Shu-Tao Xia (Tsinghua University)
GenerationRetrievalAdversarial AttackTransformerDiffusion modelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes BadRDM, a backdoor attack framework for retrieval-augmented diffusion models (RDM), which uses minimal contamination of the retrieval database and malicious contrastive learning on the retriever to make the model generate images specified by the attacker when triggered, while maintaining high-quality generation under normal inputs.
Retrieving to Recover: Towards Incomplete Audio-Visual Question Answering via Semantic-consistent Purification
Jiayu Zhang (Great Bay University), Zitong YU
RestorationRetrievalRepresentation LearningTransformerPrompt EngineeringMixture of ExpertsContrastive LearningVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: A retrieval-based missing modality recovery framework, R ScP, is proposed. It retrieves the real features of the missing modality in a unified semantic space and removes noise through context-aware adaptive purification (CAP), ultimately achieving robust reasoning for audio-visual question answering (AVQA) in scenarios with missing modalities.
Reusable Experiences: Latent Routing and Modular Composition in LLMs
Shuai Ling (Harbin Institute of Technology (Shenzhen)), Weili Guan (Harbin Institute of Technology (Shenzhen))
Computational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsAuto EncoderTextRetrieval-Augmented Generation
🎯 What it does: Proposes an experience-centralized LLM adaptation framework called ReX, which dynamically combines skill vectors from a shared experience bank through potential experience codes to generate input-specific LoRA adapters, enabling cross-task knowledge sharing and flexible adaptation without relying on task identifiers.