EMNLP 2025 Papers — Page 13
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Probing and Boosting Large Language Models Capabilities via Attention Heads
Dezhi Zhao (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Explore the correspondence between attention heads in large language models and five fundamental capabilities (mathematical reasoning, reading comprehension, common-sense reasoning, scientific reasoning, professional expertise), and utilize this correspondence for targeted instruction fine-tuning.
Probing for Arithmetic Errors in Language Models
Yucheng Sun (ETH Zürich), Mrinmaya Sachan (ETH Zürich)
Explainability and InterpretabilityTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper trains a lightweight probe to extract arithmetic operation results and correct answers from internal activations of large language models, and uses this to detect model errors;
Probing LLM World Models: Enhancing Guesstimation with Wisdom of Crowds Decoding
Yun-Shiuan Chuang (University of Wisconsin-Madison), Timothy T. Rogers (University of Wisconsin-Madison)
TransformerLarge Language ModelWorld ModelTextBenchmarkChain-of-Thought
🎯 What it does: Constructed three guesstimation datasets (MARBLES, FUTURE, ELECPRED) and evaluated the estimation capabilities of large language models on these tasks.
Probing Logical Reasoning of MLLMs in Scientific Diagrams
Yufei Wang (University of Pittsburgh), Adriana Kovashka (University of Pittsburgh)
Data SynthesisExplainability and InterpretabilityLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed a visual question-answering dataset based on food web/food chain images, and generated a large number of logical reasoning questions using seven logical templates to evaluate the logical reasoning ability of multimodal large language models on scientific charts.
Probing Narrative Morals: A New Character-Focused MFT Framework for Use with Large Language Models
Luca Mitran (McGill University), Andrew Piper (McGill University)
Large Language ModelPrompt EngineeringText
🎯 What it does: This paper constructs a role-based Moral Foundations Theory (MFT) assessment tool called MFCAQ and quantitatively analyzes character behaviors in 2,697 folktales from 55 countries using large language models (LLMs).
Procedural Environment Generation for Tool-Use Agents
Michael Sullivan (Saarland University), Alexander Koller (Saarland University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose a pipeline named RandomWorld for automatically generating synthetic environments containing interactive tools and nonlinear composite tasks to train LLM tool-use agents.
Process-Supervised Reinforcement Learning for Code Generation
Yufan Ye (Beijing Institute Of Technology), Hua Huang (Beijing Normal University)
AI Code AssistantReinforcement Learning from Human FeedbackReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes PRLCoder, which enhances code generation performance using reinforcement learning with process supervision, and for the first time directly applies DPO to code generation, automatically constructing a process supervision reward dataset through 'mutation/refactoring-execution validation'.
Process-Supervised Reward Models for Verifying Clinical Note Generation: A Scalable Approach Guided by Domain Expertise
Hanyin Wang (Mayo Clinic Health System), Jimeng Sun (Mayo Clinic Health System)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBiomedical DataElectronic Health Records
🎯 What it does: This paper designs and trains a process-supervised reward model (PRM) by step-by-step verification of LLM-generated clinical records and utilizes a large language model to automatically generate real error samples to construct large-scale training data.
ProcWorld: Benchmarking Large Model Planning in Reachability-Constrained Environments
Dong Wang (Tsinghua University), Huaping Liu (Tsinghua University)
Large Language ModelVision Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the ProcWORLD benchmark to evaluate the planning and spatial reasoning capabilities of large language models and vision-language models in multi-room environments with restricted reachability and partial observability.
Profiler: Black-box AI-generated Text Origin Detection via Context-aware Inference Pattern Analysis
Hanxi Guo (Purdue University), Xiangyu Zhang (Purdue University)
ClassificationLarge Language ModelText
🎯 What it does: Propose a black-box AI text source detection method called PROFILER, which identifies the original LLM that generated the text by leveraging context reasoning patterns.
Program of Thoughts for Financial Reasoning: Leveraging Dynamic In-Context Examples and Generative Retrieval
Subhendu Khatuya (Indian Institute of Technology Kharagpur), Niloy Ganguly (Indian Institute of Technology Kharagpur)
RetrievalTransformerSupervised Fine-TuningPrompt EngineeringTextTabularFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a two-step framework FINDER, which first uses an instruction-tuned generative retriever to extract relevant facts from text and tables, then employs dynamically selected context examples and Program-of-Thought (PoT) prompts to let GPT-4 generate executable Python programs and obtain answers.
ProLongVid: A Simple but Strong Baseline for Long-context Video Instruction Tuning
Rui Wang (Fudan University), Yu-Gang Jiang (Fudan University)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoText
🎯 What it does: Construct a large-scale long video instruction dataset and propose a progressive video instruction fine-tuning strategy to achieve efficient understanding of long videos by multimodal large models.
Promote, Suppress, Iterate: How Language Models Answer One-to-Many Factual Queries
Tianyi Lorena Yan (University of Southern California), Robin Jia (University of Southern California)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This study investigates how large language models (LLMs) answer one-to-many fact queries (1MKR), revealing an internal 'promote-then-suppress' mechanism, and further analyzes the specific roles of attention and MLP in knowledge recall and repetition avoidance;
PropRAG: Guiding Retrieval with Beam Search over Proposition Paths
Jingjin Wang (University of Illinois Urbana-Champaign), Jiawei Han (University of Illinois Urbana-Champaign)
GenerationRetrievalTextRetrieval-Augmented Generation
🎯 What it does: Developed PropRAG, a retrieval-augmented generation framework that combines high-fidelity propositional knowledge representation with LLM-free online beam search for multi-step reasoning question answering.
ProReason: Multi-Modal Proactive Reasoning with Decoupled Eyesight and Wisdom
Jingqi Zhou (University of Hong Kong), Chuan Wu (University of Hong Kong)
Knowledge DistillationLarge Language ModelMixture of ExpertsMultimodalityBenchmark
🎯 What it does: Proposes the PROREASON framework, decomposing visual reasoning into two stages: active visual perception and text reasoning, employing a multi-agent iterative active information acquisition approach.
Prototypical Human-AI Collaboration Behaviors from LLM-Assisted Writing in the Wild
Sheshera Mysore (Microsoft), Bahareh Sarrafzadeh
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper conducts a large-scale analysis of writing session logs from Microsoft Bing Copilot and WildChat, extracting and quantifying typical user behavior patterns (PATH) when collaborating with large language models (LLMs). These behaviors are associated with users' writing intentions to reveal collaborative characteristics in real-world environments and insights into alignment with LLMs.
ProtoVQA: An Adaptable Prototypical Framework for Explainable Fine-Grained Visual Question Answering
Xingjian Diao (Dartmouth College), Jiang Gui (Dartmouth College)
Explainability and InterpretabilityTransformerVision Language ModelMultimodality
🎯 What it does: Propose the ProtoVQA framework, achieving interpretable visual question answering through problem-aware prototype learning and spatially constrained greedy matching.
PruneCD: Contrasting Pruned Self Model to Improve Decoding Factuality
Byeongho Yu (Pohang University of Science and Technology), Eunhyeok Park (Pohang University of Science and Technology)
GenerationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose PruneCD, a method that improves the factuality of large language model generation by constructing an amateur model using layer pruning.
Pruning the Paradox: How CLIP’s Most Informative Heads Enhance Performance While Amplifying Bias
Avinash Madasu (Intel Labs), Phillip Howard (Thoughtworks)
ClassificationRetrievalExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelVision Language ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes the Concept Consistency Score (CCS) to measure the consistency of each attention head in the CLIP vision-language model with respect to specific concepts, and validates through soft pruning experiments the importance of high CCS heads for model performance, as well as their dual roles in reasoning, OOV detection, video retrieval, and social bias.
PSET: a Phonetics-Semantics Evaluation Testbed
Gianluca Sperduti (Istituto di Scienza e Tecnologie dell'Informazione A Faedo), Dong Nguyen (Utrecht University)
Representation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Designed and utilized the Phonetics-Semantics Evaluation Testbed (PSET) to evaluate English phoneme embeddings.
Pun Unintended: LLMs and the Illusion of Humor Understanding
Alessandro Zangari (Ca' Foscari University of Venice), Jose Camacho-Collados (Cardiff University)
RecognitionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The paper systematically evaluates the robustness and interpretability of large language models in identifying and explaining puns by constructing new pun evaluation datasets (PunnyPattern and PunBreak) and improving existing datasets.
PunMemeCN: A Benchmark to Explore Vision-Language Models’ Understanding of Chinese Pun Memes
Zhijun Xu (Fudan University), Deqing Yang (Fudan University)
ClassificationGenerationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented the PUNMEMECN benchmark for evaluating the identification, sentiment analysis, and appropriate responses in chat scenarios for Chinese pun memes.
Puzzled by Puzzles: When Vision-Language Models Can’t Take a Hint
Heekyung Lee (POSTECH), David M. Chan (University of California, Berkeley)
Large Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper constructs a dataset of 432 manually designed English rebus puzzles, systematically evaluates the performance of visual language models (VLMs) on puzzle-solving tasks, and experimentally analyzes the cognitive limitations of models through various prompting strategies, skill guidance, iterative refinement, and other methods.
PychoAgent: Psychology-driven LLM Agents for Explainable Panic Prediction on Social Media during Sudden Disaster Events
Mengzhu Liu (National University of Defense Technology), Quanjun Yin (National University of Defense Technology)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextChain-of-Thought
🎯 What it does: Propose the PsychoAgent framework, which utilizes psychology-driven LLM agents and multi-domain feature fusion to predict panic emotions in social media disaster events, and constructs a fine-grained panic emotion dataset called COPE.
QCRD: Quality-guided Contrastive Rationale Distillation for Large Language Models
Wei Wang (University of Science and Technology of China), Li Xiao (University of Science and Technology of China)
Knowledge DistillationTransformerContrastive LearningTextChain-of-Thought
🎯 What it does: This paper proposes a quality-guided contrastive rationalization distillation framework QCRD, which extracts positive and negative rationalization knowledge from large language models and distills it to small models.
QFrCoLA: a Quebec-French Corpus of Linguistic Acceptability Judgments
David Beauchemin (Laval University), Richard Khoury (Laval University)
ClassificationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed and released QFrCoLA, a binary dataset for French acceptability judgment containing 25,153 in-domain examples and 2,675 out-of-domain examples, and conducted fine-tuning and zero-shot evaluation on various monolingual and cross-lingual Transformers using this dataset.
QG-CoC: Question-Guided Chain-of-Captions for Large Multimodal Models
Kuei-Chun Kao (University of California), Cho-Jui Hsieh (University of California)
Large Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a zero-shot prompting method called Question-Guided Chain-of-Captions (QG-CoC), which first decomposes the question into sub-questions, then generates targeted caption texts for each sub-question, and finally combines sub-questions and sub-answers to answer the original question;
QSpec: Speculative Decoding with Complementary Quantization Schemes
Juntao Zhao (University of Hong Kong), Chuan Wu (University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a reasoning framework QSPEC that combines low-precision activation joint quantization (W4A4) for drafting with high-precision weight quantization (W4A16) for verification, achieving zero additional memory cost through shared weights and KV cache;
QualBench: Benchmarking Chinese LLMs with Localized Professional Qualifications for Vertical Domain Evaluation
Mengze Hong (Hong Kong Polytechnic University), Di Jiang (Hong Kong Polytechnic University)
Federated LearningTransformerPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Constructed a multi-domain Chinese professional qualification exam QA benchmark (QualBench), containing 17,316 questions from 24 qualification exams across six vertical fields, and evaluated the professional knowledge of Chinese LLMs.
Quantifying Language Disparities in Multilingual Large Language Models
Songbo Hu (University of Cambridge), Anna Korhonen (University of Cambridge)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a multilingual large language model (LLM) evaluation framework that disentangles confounding factors (language, task, model) to quantify performance differences across languages, and validates its effectiveness through case studies.
Quantifying Logical Consistency in Transformers via Query-Key Alignment
Eduard Tulchinskii (Skolkovo Institute of Science and Technology), Serguei Barannikov (Skolkovo Institute of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a lightweight metric (QK-score) based on the internal query-key (QK) alignment in Transformers, which quantifies the logical consistency of large language models (LLMs) in a single forward pass, and evaluates logical reasoning by identifying high-scoring heads.
Quantized but Deceptive? A Multi-Dimensional Truthfulness Evaluation of Quantized LLMs
Yao Fu (Case Western Reserve University), Pan Li (Case Western Reserve University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed the TruthfulnessEval framework, systematically evaluating the authenticity of quantized large language models across three dimensions: logical reasoning, common sense judgment, and imitative false answers, while investigating prompt sensitivity and internal representations.
Query-Focused Retrieval Heads Improve Long-Context Reasoning and Re-ranking
Wuwei Zhang (Princeton University), Xi Ye (Princeton University)
RetrievalTransformerTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a new query-focused retrieval head QRHEAD and built a retriever QRRETRIEVER based on it for long-context reasoning and paragraph re-ranking.
QUIDS: Query Intent Description for Exploratory Search via Dual Space Modeling
Yumeng Wang (Leiden University), Suzan Verberne (Leiden University)
RetrievalTransformerContrastive LearningText
🎯 What it does: Propose a dual-space contrastive learning model called QUIDS to generate natural language query intent descriptions in exploratory search, helping users understand the inferences made by search engines.
QuZO: Quantized Zeroth-Order Fine-Tuning for Large Language Models
Jiajun Zhou (University of California Santa Barbara), Zheng Zhang (University of California Santa Barbara)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A quantization zeroth-order fine-tuning framework called QuZO, which does not require backpropagation, is proposed for low-precision (INT4/INT8) inference hardware. It directly utilizes forward inference for parameter updates.
R-Bind: Unified Enhancement of Attribute and Relation Binding in Text-to-Image Diffusion Models
Huixuan Zhang (Peking University), Xiaojun Wan (Peking University)
GenerationConvolutional Neural NetworkTransformerDiffusion modelImageTextBenchmark
🎯 What it does: Developed a training-free method called R-Bind, which jointly enhances entity-attribute binding and entity-relation-entity binding by performing gradient optimization on the attention maps of diffusion models during inference, achieving semantic consistency between text and images.
R-BPE: Improving BPE-Tokenizers with Token Reuse
Nancy Hamdan (Arab Center for Research and Policy Studies), Fadi A. Zaraket (Arab Center for Research and Policy Studies)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Proposes R-BPE, a lightweight vocabulary reuse framework that enhances existing BPE tokenizers to better support specified target languages (e.g., Arabic) without altering model size or introducing additional parameters.
R-CHAR: A Metacognition-Driven Framework for Role-Playing in Large Language Models
Haiming Qin (Shenzhen University), Rui Mao (Shenzhen University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose the R-CHAR framework, which leverages metacognitive-driven hierarchical adaptive reasoning to enhance cognitive consistency and reasoning quality of large language models in role-playing tasks.
R-PRM: Reasoning-Driven Process Reward Modeling
Shuaijie She (Nanjing University), Shujian Huang (Nanjing University)
Reinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a reasoning-based process reward model, R-PRM, to evaluate the mathematical reasoning steps generated by large language models.
R-TOFU: Unlearning in Large Reasoning Models
Sangyeon Yoon (Yonsei University), Albert No (Yonsei University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the R-TOFU benchmark, specifically designed to evaluate the simultaneous removal of reasoning chains (CoT) and answers by large reasoning models (LRMs) when forgetting specific training data.
R2I-Bench: Benchmarking Reasoning-Driven Text-to-Image Generation
Kaijie Chen (University of California Davis), Lifu Huang (University of California Davis)
GenerationLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the R2I-Bench benchmark and R2I-Score evaluation metric to assess the reasoning capabilities of text-to-image models.
RACCooN: Versatile Instructional Video Editing with Auto-Generated Narratives
Jaehong Yoon (UNC Chapel Hill), Mohit Bansal (UNC Chapel Hill)
GenerationLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelAuto EncoderVideoTextMultimodality
🎯 What it does: Proposes a two-stage video-to-paragraph-to-video editing framework named RACCOON, which automatically generates structured narratives and supports adding, deleting, and modifying video content.
RAcQUEt: Unveiling the Dangers of Overlooked Referential Ambiguity in Visual LLMs
Alberto Testoni (Amsterdam UMC), Raquel Fernández (University of Amsterdam)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes a new dataset called RACQUET (including RACQUET-GENERAL and RACQUET-BIAS) focused on referential ambiguity in visual question answering, and constructs an evaluation framework based on three types of responses (Explicit, Implicit, High Risk).
RaDeR: Reasoning-aware Dense Retrieval Models
Debrup Das (University of Massachusetts Amherst), Razieh Rahimi (University of Massachusetts Amherst)
RetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Built and trained a reasoning-oriented dense retriever called RaDeR, leveraging retrieval trajectories generated by MCTS to provide the retrieval model with diverse and high-quality training data.
RAED: Retrieval-Augmented Entity Description Generation for Emerging Entity Linking and Disambiguation
Karim Ghonim (Sapienza University of Rome), Roberto Navigli (Sapienza University of Rome)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: RAED addresses the problem of new entity linking and disambiguation by retrieving external knowledge and generating entity titles and definitions.
RAG-Instruct: Boosting LLMs with Diverse Retrieval-Augmented Instructions
Wanlong Liu (University of Electronic Science and Technology of China), Benyou Wang (Chinese University of Hong Kong Shenzhen)
RetrievalTransformerPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the RAG-Instruct method, which enhances the RAG capabilities of large language models by synthesizing diverse, high-quality retrieval-augmented instruction data, and constructs a 40K instruction set based on Wikipedia.
RAG-Zeval: Enhancing RAG Responses Evaluator through End-to-End Reasoning and Ranking-Based Reinforcement Learning
Kun Li (Chinese University of Hong Kong), Helen M. Meng
RetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Built a rule-guided end-to-end evaluator, RAG-Zeval, for reliable and interpretable assessment of answers generated by retrieval-augmented generation (RAG) systems, and trained a compact LLM using reinforcement learning to achieve automated evaluation without human annotations.
RAG+: Enhancing Retrieval-Augmented Generation with Application-Aware Reasoning
Yu Wang (Xi'an Jiaotong University), Ting Liu (Xi'an Jiaotong University)
RetrievalTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Introduce RAG+ in the retrieval-augmented generation (RAG) framework by constructing a bilingual corpus of knowledge and corresponding application examples, and simultaneously retrieving both during the inference phase to provide structured, goal-oriented reasoning prompts, thereby enhancing the performance of large language models (LLMs) in complex reasoning tasks.
RAGferee: Building Contextual Reward Models for Retrieval-Augmented Generation
Andrei C. Coman (Idiap Research Institute), Adrià de Gispert (Amazon AGI)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the RAGferee method, which retransforms QA datasets into RAG-specific preference pairs for training contextual reward models.
RALS: Resources and Baselines for Romanian Automatic Lexical Simplification
Fabian Anghel, Sergiu Nisioi (University of Bucharest)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Created a dataset for predicting and simplifying lexical complexity in Romanian, and proposed a hybrid simplification system based on DexFlex.
Randomized Smoothing Meets Vision-Language Models
Emmanouil Seferis (National Technical University of Athens), Chih-Hong Cheng (Carl von Ossietzky University of Oldenburg)
Safty and PrivacyComputational EfficiencyAdversarial AttackLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Provide provable robustness certificates for vision-language models (VLMs) by migrating randomized smoothing (RS) from classification tasks to generative models, and complete robustness verification by introducing an 'oracle' classifier to cluster generated text, determine harm/non-harm, or recognize discrete actions.
Randomly Removing 50% of Dimensions in Text Embeddings has Minimal Impact on Retrieval and Classification Tasks
Sotaro Takeshita (University of Mannheim), Simone Paolo Ponzetto (University of Mannheim)
ClassificationRetrievalTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Investigate the performance impact of text embeddings after randomly truncating 50% of dimensions on retrieval and classification tasks, and reveal through dimension attribution analysis that many dimensions in the embeddings negatively affect performance.
Rank-Awareness and Angular Constraints: A New Perspective on Learning Sentence Embeddings from NLI Data
Zicheng Zhou (University of Chinese Academy of Sciences), Qinghai Miao (University of Chinese Academy of Sciences)
Representation LearningData-Centric LearningTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: This work proposes the RAOE framework, which leverages complete NLI data (Entailment, Neutral, Contradiction) and precomputes continuous similarity scores for each sentence pair to learn higher-quality sentence embeddings.
Rapid Word Learning Through Meta In-Context Learning
Wentao Wang (New York University), Brenden Lake
Meta LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the Minnow method, leveraging meta-learning and placeholder tokens to enable language models to rapidly learn new words and generate novel usages with few examples;
RareSyn: Health Record Synthesis for Rare Disease Diagnosis
Huimin Wang (Tencent Jarvis Lab), Xian Wu (Tencent Jarvis Lab)
Data SynthesisTransformerLarge Language ModelContrastive LearningBiomedical DataElectronic Health Records
🎯 What it does: Designed the RareSyn framework by integrating LLM with medical knowledge graphs, generating synthetic electronic health records (EHR) for rare disease diagnosis through layered recall, imap structuring, KG-weighted entity sampling, and LLM rewriting.
RAV: Retrieval-Augmented Voting for Tactile Descriptions Without Training
Jinlin Wang, Hongyu Yang (Sichuan University)
GenerationRetrievalVision Language ModelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed a parameter-free cross-modal perception method called RAV, which retrieves CLIP feature vectors from visual and tactile modalities and generates tactile descriptions using a voting mechanism.
RAVEN: Query-Guided Representation Alignment for Question Answering over Audio, Video, Embedded Sensors, and Natural Language
Subrata Biswas (Worcester Polytechnic Institute), Bashima Islam (Worcester Polytechnic Institute)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose the RAVEN framework, which adaptively fuses video, audio, sensor, and language modalities through query-guided cross-modal gating (QuART) to achieve multi-modal question answering.
RBPtool: A Deep Language Model Framework for Multi-Resolution RBP-RNA Binding Prediction and RNA Molecule Design
Jiyue Jiang (Chinese University of Hong Kong), Yu Li (Chinese University of Hong Kong)
GenerationGraph Neural NetworkTransformerLarge Language ModelBiomedical Data
🎯 What it does: Developed a unified deep language model framework RBPtool for multi-resolution RBP-RNA binding prediction and RNA molecule design.
RCScore: Quantifying Response Consistency in Large Language Models
Dongjun Jang (Seoul National University), Hyopil Shin (Seoul National University)
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: Proposed and studied the RCScore evaluation framework, which quantifies the response consistency of large language models to different instruction styles, and examines the association between consistency and task accuracy through Cross-Response Similarity (CRS).
RD-MCSA: A Multi-Class Sentiment Analysis Approach Integrating In-Context Classification Rationales and Demonstrations
Haihua Xie (Beijing Institute of Mathematical Sciences and Applications), Mingming Sun (Beijing Institute of Mathematical Sciences and Applications)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Studied multi-class sentiment analysis, proposing the RD-MCSA framework that combines classification reasoning with adaptive example selection, leveraging ICL to enhance performance under limited supervision.
Re-Align: Aligning Vision Language Models via Retrieval-Augmented Direct Preference Optimization
Shuo Xing (Texas A&M University), Zhengzhong Tu (Texas A&M University)
RetrievalOptimizationReinforcement Learning from Human FeedbackTransformerVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study generates aligned data through image retrieval and aligns vision-language models by combining direct preference optimization (DPO), significantly reducing model hallucinations and improving visual question answering performance.
REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing
Haitian Zhong (Institute of Automation, Chinese Academy of Sciences), Tieniu Tan (Institute of Automation, Chinese Academy of Sciences)
Representation LearningLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the two-stage REACT framework, achieving precise editing of knowledge in large language models by extracting directional vectors of factual representations and applying controlled perturbations in hidden layers.
Read to Hear: A Zero-Shot Pronunciation Assessment Using Textual Descriptions and LLMs
Yu-Wen Chen (Columbia University), Julia Hirschberg (Columbia University)
RecognitionTransformerLarge Language ModelTextAudio
🎯 What it does: Propose a zero-shot pronunciation assessment method called TextPA, which evaluates text-based speech descriptions (transcripts, IPA, CMU phoneme sequences) using a large language model and provides interpretable feedback.
Reading Between the Prompts: How Stereotypes Shape LLM’s Implicit Personalization
Vera Neplenbroek (University of Amsterdam), Raquel Fernández (University of Groningen)
Data SynthesisExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper constructs controllable synthetic dialogues to systematically evaluate how large language models implicitly infer and memorize user identities during multi-round interactions, and quantitatively analyzes this through two perspectives: model internal representations and generated responses.
ReAgent: Reversible Multi-Agent Reasoning for Knowledge-Enhanced Multi-Hop QA
Zhao Xinjie (University of Tokyo), Irene Li (University of Tokyo)
Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed a reversible multi-agent reasoning framework called ReAgent, specifically designed for multi-hop question answering, and implemented local and global backtracking mechanisms to correct reasoning errors.
Real-time Ad Retrieval via LLM-generative Commercial Intention for Sponsored Search Advertising
Tongtong Liu (Tencent Inc.), Peng Shu (Tencent Inc.)
RetrievalRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a real-time ad retrieval framework RARE based on LLM, which directly matches user queries with ads by utilizing commercial intent (CI) text generated by LLM.
Realistic Training Data Generation and Rule Enhanced Decoding in LLM for NameGuess
Yikuan Xia (Peking University), Jun Gao (Peking University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextTabular
🎯 What it does: Improve training data by generating subsequence abbreviations using a model and collecting non-subsequence abbreviation tables; during inference, use an automaton to constrain Beam search, ensuring outputs comply with abbreviation rules.
REALM: Recursive Relevance Modeling for LLM-based Document Re-Ranking
Pinhuan Wang (Rutgers State University of New Jersey), Hang Liu (Rutgers State University of New Jersey)
RetrievalComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a document re-ranking framework REALM based on large language models (LLMs), which models document relevance using a Gaussian distribution and iteratively refines the ranking through recursive Bayesian updates.
REARANK: Reasoning Re-ranking Agent via Reinforcement Learning
Le Zhang (Mila - Quebec AI Institute), Aishwarya Agrawal (Mila - Quebec AI Institute)
RetrievalExplainability and InterpretabilityLarge Language ModelReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: Propose a list-based reasoning re-ranking agent REARANK based on a large language model, which is trained using reinforcement learning to generate interpretable reasoning chains and provide final ranking results during the re-ranking process.
Reason to Rote: Rethinking Memorization in Reasoning
Yupei Du (Utrecht University), Barbara Plank (LMU Munich)
Explainability and InterpretabilityRepresentation LearningTransformerText
🎯 What it does: Studies how language models, when trained with label noise, can simultaneously memorize incorrect labels while maintaining reasoning ability on clean samples, and quantitatively and qualitatively analyze their internal mechanisms.
Reasoning Model Unlearning: Forgetting Traces, Not Just Answers, While Preserving Reasoning Skills
Changsheng Wang (Michigan State University), Sijia Liu (Michigan State University)
Safty and PrivacyTransformerTextChain-of-Thought
🎯 What it does: Studied the machine forgetting problem in Large Retrieval Models (LRM), proposing a new method R2MU that can completely remove sensitive information (including reasoning trajectories and final answers) without compromising inference capabilities.
Reasoning under Uncertainty: Efficient LLM Inference via Unsupervised Confidence Dilution and Convergent Adaptive Sampling
Zhenning Shi (Tsinghua University), Qing Li (Peng Cheng Laboratory)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a LLM inference framework based on self-supervised confidence calibration and adaptive sampling, which can significantly reduce inference costs while maintaining accuracy.
Reasoning-to-Defend: Safety-Aware Reasoning Can Defend Large Language Models from Jailbreaking
Junda Zhu (Beihang University), Lei Sha (Beihang University)
Safty and PrivacyKnowledge DistillationTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper proposes a training paradigm called Reasoning-to-Defend (R2D), enabling large language models to identify and block jailbreak attacks during generation through safety-aware reasoning, avoiding direct hard rejections;
ReasonMed: A 370K Multi-Agent Generated Dataset for Advancing Medical Reasoning
Yu Sun (Alibaba DAMO Academy), Yu Rong (Alibaba DAMO Academy)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAgentic AIBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Constructed the largest medical reasoning dataset, ReasonMed, containing 370,000 high-quality multi-step Chain-of-Thought (CoT) examples;
Recall with Reasoning: Chain-of-Thought Distillation for Mamba’s Long-Context Memory and Extrapolation
Jun-Yu Ma (Tencent AI Lab), Dong Yu (Tencent AI Lab)
Computational EfficiencyKnowledge DistillationRepresentation LearningSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the Recall with Reasoning (RwR) method, enhancing the Mamba model's performance in long context memory and reasoning through Chain-of-Thought (CoT) transfer.
RECALL: REpresentation-aligned Catastrophic-forgetting ALLeviation via Hierarchical Model Merging
Bowen Wang (Tsinghua University), Sheng Zhang (Tsinghua University)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a hierarchical model fusion framework called RECALL based on internal representation similarity, aiming to achieve continual learning and catastrophic forgetting suppression for LLMs without relying on historical data or task labels.
RecBase: Generative Foundation Model Pretraining for Zero-Shot Recommendation
Sashuai Zhou (Zhejiang University), Zhou Zhao (Zhejiang University)
Domain AdaptationRecommendation SystemTransformerAuto EncoderTextSequential
🎯 What it does: Construct and pretrain a foundational model named RecBase, leveraging unified text representations and hierarchical concept IDs for autoregressive training to achieve zero-shot cross-domain recommendation.
RecGPT: A Foundation Model for Sequential Recommendation
Yangqin Jiang (University of Hong Kong), Chao Huang (University of Hong Kong)
Recommendation SystemTransformerLarge Language ModelTextSequential
🎯 What it does: Propose a foundational model for sequence recommendation called RecGPT, which utilizes text-driven unified tokenization and autoregressive modeling to achieve zero-shot cross-domain recommendation;
Recontextualizing Revitalization: A Mixed Media Approach to Reviving the Nüshu Language
Ivory Yang (Dartmouth College), Soroush Vosoughi (Dartmouth College)
RecognitionData SynthesisTransformerSupervised Fine-TuningImageMultimodalitySequential
🎯 What it does: Proposed a multimodal framework for the Nüshu language, achieving digitization and machine recognition of this extremely endangered female-specific writing system through OCR images and stroke sequences;
Recursive Training Loops in LLMs: How training data properties modulate distribution shift in generated data?
Grgur Kovač (INRIA), Pierre-Yves Oudeyer (INRIA)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: This paper systematically investigates the impact of human data attributes (such as lexical diversity, semantic diversity, quality, positivity/negativity, etc.) on distribution shift, quality and diversity degradation, and changes in political bias caused by recursive training, through training LLMs in an iterative chain.
RED: Unleashing Token-Level Rewards from Holistic Feedback via Reward Redistribution
Jiahui Li (Zhejiang University), Cheng Yang (Ant Group)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Achieve fine-grained rewards by redistributing the overall reward in a time-difference manner to each generated token during the RLHF training phase.
ReDepress: A Cognitive Framework for Detecting Depression Relapse from Social Media
Aakash Kumar Agarwal (Indian Institute of Technology Bombay), Pushpak Bhattacharyya (Indian Institute of Technology Bombay)
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the first clinically validated social media depression relapse dataset ReDepress, and annotated and modeled relapse status using cognitive theory;
RedHerring Attack: Testing the Reliability of Attack Detection
Jonathan Rusert (Purdue University)
Adversarial AttackTransformerText
🎯 What it does: Proposed and tested a new adversarial attack called RedHerring, designed to trick detection models into generating false positives while maintaining the accuracy of text classifiers, thereby undermining human trust in detection models.
ReEvalMed: Rethinking Medical Report Evaluation by Aligning Metrics with Real-World Clinical Judgment
Ruochen Li (Technical University of Munich), Youxiang Zhu (University of Massachusetts Boston)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelTextBiomedical DataElectronic Health Records
🎯 What it does: This paper rethinks medical report evaluation methods, proposing a Meta-Evaluation framework based on clinical needs and constructing a GT-ME report pair dataset verified by experts for fine-grained assessment of existing evaluation metrics.
Refining Attention for Explainable and Noise-Robust Fact-Checking with Transformers
Jean-Flavien Bussotti (Megagon Labs), Paolo Papotti (EURECOM)
ClassificationExplainability and InterpretabilityTransformerText
🎯 What it does: Propose the ATTUN model, which enhances model interpretability and robustness to noise by directly adjusting attention weights within the Transformer architecture.
Refining Text Generation for Realistic Conversational Recommendation via Direct Preference Optimization
Manato Tajiri (University of Electro-Communications), Michimasa Inaba (University of Electro-Communications)
GenerationRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Utilize DPO to fine-tune LLM for generating conversation summaries and item recommendations to enhance the quality of real-world conversational recommendation systems.
ReflAct: World-Grounded Decision Making in LLM Agents via Goal-State Reflection
Jeonghye Kim (KAIST), Kyomin Jung (Seoul National University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the ReflAct framework, replacing ReAct's 'next action' thinking with continuous reflection on state and goals in LLM agents to achieve more robust decision-making.
Reflective Agreement: Combining Self-Mixture of Agents with a Sequence Tagger for Robust Event Extraction
Fatemeh Haji (Secure AI and Autonomy Lab), Peyman Najafirad (Secure AI and Autonomy Lab)
Large Language ModelSupervised Fine-TuningAgentic AIMixture of ExpertsTextBenchmark
🎯 What it does: Proposed the ARIS framework, integrating Self-Mixture of Agents with the sequence annotator (TagPrime) to achieve robust event extraction.
Reframe Your Life Story: Interactive Narrative Therapist and Innovative Moment Assessment with Large Language Models
Yi Feng (Beijing Jiaotong University), Jian Yu (Beijing Jiaotong University)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes an interactive narrative therapist framework INT (Interactive Narrative Therapist) and an innovative moment assessment method IMA (Innovative Moment Assessment), based on large language models, to simulate professional narrative therapists and quantify therapeutic progress.
Refusal-Aware Red Teaming: Exposing Inconsistency in Safety Evaluations
Yongkang Chen (Academy of Military Sciences), Xiaohui Kuang (Academy of Military Sciences)
Safty and PrivacyLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a rejection-aware red teaming framework that automatically generates test cases via reinforcement learning to expose the gap between internal rejection behaviors in large language models and external security assessments.
Reimagining Safety Alignment with An Image
Yifan Xia (Wuhan University), Jindong Gu (University of Oxford)
OptimizationSafty and PrivacyLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Proposes a method that simultaneously alleviates over-rejection and security issues against jailbreaking attacks in multi-modal large language models (MLLMs) by optimizing visual inputs (Magic Image).
Reinforced Query Reasoners for Reasoning-intensive Retrieval Tasks
Xubo Qin (State Key Laboratory of General Artificial Intelligence), Zilong Zheng (State Key Laboratory of General Artificial Intelligence)
RetrievalTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose and train a small-scale language model (TongSearch QR 7B/1.5B) for query reasoning and rewriting, using reinforcement learning to enhance performance in reasoning retrieval tasks.
Reinforcement Learning for Large Language Models via Group Preference Reward Shaping
Huaisheng Zhu (Pennsylvania State University), Vasant G. Honavar (Pennsylvania State University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a reinforcement learning method called GPRS without a value function, which generates multiple responses under the same prompt and compares their merits and demerits using a reward model, forming a reward based on preference (win-rate);
Reliable and Cost-Effective Exploratory Data Analysis via Graph-Guided RAG
Mossad Helali (Concordia University), Essam Mansour (Concordia University)
Data-Centric LearningAI Code AssistantLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelGraphTabularRetrieval-Augmented Generation
🎯 What it does: Automatically construct knowledge graphs and utilize retrieval-augmented generation to create executable exploratory data analysis (EDA) notebooks, combining LLM with self-correcting agents to achieve high reliability and low-cost code generation.
Reliable Evaluation and Benchmarks for Statement Autoformalization
Auguste Poiroux (EPFL), Antoine Bosselut (EPFL)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper studies the automatic translation of natural language mathematical statements into Lean 4 formal language (statement autoformalization), proposing new evaluation metrics and benchmarks to enhance the measurability and reliability of this task.
ReMedy: Learning Machine Translation Evaluation from Human Preferences with Reward Modeling
Shaomu Tan (University of Amsterdam), Christof Monz (University of Amsterdam)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose ReMedy, a reward modeling-based MT evaluation framework that trains models using contrastive human preference data to assess translation quality.
Representation Potentials of Foundation Models for Multimodal Alignment: A Survey
Jianglin Lu (Northeastern University), Yun Fu (Northeastern University)
Representation LearningTransformerImageTextMultimodalityBiomedical DataReview/Survey PaperAudio
🎯 What it does: This review systematically organizes and evaluates the 'representation potential' of foundation models in unimodal and cross-modal alignment, including visual, language, speech, multimodal, and alignment phenomena with neuroscience.
Rescorla-Wagner Steering of LLMs for Undesired Behaviors over Disproportionate Inappropriate Context
Rushi Wang (University of Illinois Urbana-Champaign), Denghui Zhang (Stevens Institute of Technology)
OptimizationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a novel robust regulation method called RW-Steering, aiming to enhance the answer quality and safety of large language models in mixed contexts (containing both compliant and non-compliant information).
ReSeeding Latent States for Sequential Language Understanding
Stéphane Aroca-Ouellette (University of Colorado Boulder), Alessandro Roncone (University of Colorado Boulder)
Representation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningTextSequential
🎯 What it does: Build and train a framework called RESEED, enabling large language models to learn from environmental states during the training phase and re-inject these states' latent representations during the inference phase to achieve better multi-step sequential reasoning.
RESF: Regularized-Entropy-Sensitive Fingerprinting for Black-Box Tamper Detection of Large Language Models
Pingyi Hu (Huazhong University of Science and Technology), Bin Benjamin Zhu (Microsoft Corporation)
Anomaly DetectionLarge Language ModelText
🎯 What it does: This paper proposes a black-box large language model (LLM) tampering detection method called RESF, which uses trainable fingerprint sequences to detect whether model parameters have been tampered with.