ACL 2025 Papers — Page 13
Annual Meeting of the Association for Computational Linguistics · 1699 papers
Predicting Turn-Taking and Backchannel in Human-Machine Conversations Using Linguistic, Acoustic, and Visual Signals
Yuxin Lin (Xiamen University), Wangzheng Shi (Xiamen University)
ClassificationTransformerLarge Language ModelVideoTextMultimodalityAudio
🎯 What it does: Constructed the MM-F2F human-machine dialogue multimodal dataset and proposed an end-to-end multimodal framework to predict alternation and background speech in conversations.
Prediction Hubs are Context-Informed Frequent Tokens in LLMs
Beatrix Miranda Ginn Nielsen, Marco Baroni
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper investigates the 'hubness' phenomenon in autoregressive large language models (LLMs) within high-dimensional representation spaces, proving that the probability distance (softmax dot product) used by LLMs for generating the next word is not affected by meaningless hubs caused by distance concentration, yet still exhibits high hubness, revealing that these hubs correspond to context-regulated high-frequency words; additionally, experiments on other commonly used Euclidean/cosine distances show they produce meaningless hubs, and suggest using hubness mitigation techniques in such comparisons.
PreP-OCR: A Complete Pipeline for Document Image Restoration and Enhanced OCR Accuracy
Shuhao Guan (University College Dublin), Derek Greene (Trinity College Dublin)
RecognitionRestorationData SynthesisTransformerLarge Language ModelImageText
🎯 What it does: Built a two-stage complete pipeline, first training an image restoration model using synthetic data to enhance the quality of historical document images, then using a ByT5-based semantic correction module to correct OCR errors.
Pretraining Context Compressor for Large Language Models with Embedding-Based Memory
Yuhong Dai (Shenzhen University), Hao Liao (Microsoft Gaming)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderText
🎯 What it does: Propose a decoupled pre-trained context compressor (PCC), which achieves efficient inference for LLMs on long contexts by compressing long texts into dense embedding slots;
Principled Content Selection to Generate Diverse and Personalized Multi-Document Summaries
Vishakh Padmakumar (New York University), Jennifer Healey (Adobe Research)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Achieve multi-document summarization through a three-step process: ① Use LLM to decompose each document into atomic key points; ② Employ Determinantal Point Process (DPP) to select a subset of key points that are both diverse and relevant to user intent; ③ Reuse LLM to rewrite the selected key points into a coherent summary; and evaluate on the DIVERSESUMM benchmark.
Principled Understanding of Generalization for Generative Transformer Models in Arithmetic Reasoning Tasks
Xingcheng Xu (Shanghai Artificial Intelligence Laboratory), Yanqing Yang (Fudan University)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes a unified theoretical framework to explain the length generalization (upstream/downstream OOD) behavior of Transformers in arithmetic tasks (addition, multiplication, modular addition, modular multiplication).
PRISM: A Framework for Producing Interpretable Political Bias Embeddings with Political-Aware Cross-Encoder
Yiqun Sun (National University of Singapore), Jun Yu (Harbin Institute of Technology)
RetrievalExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed the PRISM framework for automatically generating interpretable political bias embeddings;
PrivaCI-Bench: Evaluating Privacy with Contextual Integrity and Legal Compliance
Haoran Li (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
Safty and PrivacyLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Establish the PrivaCI-Bench evaluation benchmark, covering real court cases, privacy policies, and synthetic EU AI Act scenarios, to assess LLMs' legal compliance and context understanding capabilities.
PrivacyRestore: Privacy-Preserving Inference in Large Language Models via Privacy Removal and Restoration
Ziqian Zeng (South China University of Technology), Cen Chen (South China University of Technology)
Safty and PrivacyTransformerLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: Propose the PrivacyRestore method, which first actively removes privacy spans from user inputs on the client side, and then restores complete information through the server-side activation-guided recovery mechanism and meta-vector reconstruction, achieving privacy protection during online LLM inference.
Private Memorization Editing: Turning Memorization into a Defense to Strengthen Data Privacy in Large Language Models
Elena Sofia Ruzzetti (University of Rome Tor Vergata), Fabio Massimo Zanzotto (University of Rome Tor Vergata)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Developed a model editing method called Private Memorization Editing (PME), which defends against privacy leakage by identifying and modifying personally identifiable information (PII) already memorized in large language models (LLMs).
PRMBench: A Fine-grained and Challenging Benchmark for Process-Level Reward Models
Mingyang Song (Fudan University), Yu Cheng (Chinese University of Hong Kong)
Reinforcement Learning from Human FeedbackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the PRMBENCH benchmark for fine-grained evaluation of process-level reward models' ability to detect erroneous steps
Probing LLMs for Multilingual Discourse Generalization Through a Unified Label Set
Florian Eichin (LMU Munich), Michael A. Hedderich (LMU Munich)
ClassificationRecognitionExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a cross-framework and cross-lingual unified discourse relation label set, and uses this label set to conduct probe experiments on large language models (LLMs) on the multilingual, multi-framework DISRPT dataset to evaluate their generalization ability in discourse relation extraction.
Probing Relative Interaction and Dynamic Calibration in Multi-modal Entity Alignment
Chenxiao Li (Northeastern University), Cairui Wang (Northeastern University)
Representation LearningGraph Neural NetworkTransformerMultimodality
🎯 What it does: Proposed the RICEA framework, utilizing relative interaction and dynamic calibration techniques to achieve multi-modal entity alignment.
ProcessBench: Identifying Process Errors in Mathematical Reasoning
Chujie Zheng (Qwen Team, Alibaba Inc.), Junyang Lin (Qwen Team, Alibaba Inc.)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed and released the PROCESSBENCH dataset to evaluate the ability of language models to identify erroneous steps in mathematical reasoning processes.
ProgCo: Program Helps Self-Correction of Large Language Models
Xiaoshuai Song (Taobao & Tmall Group of Alibaba), Bo Zheng (Taobao & Tmall Group of Alibaba)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Designed a program-driven self-correction framework named ProgCo, which first allows LLM to self-generate and self-execute pseudo-programs for verification (ProgVe), and then performs dual reflection and program refinement based on feedback to improve answers (ProgRe)
Program Synthesis Benchmark for Visual Programming in XLogoOnline Environment
Chao Wen (Max Planck Institute for Software Systems), Adish Singla (Max Planck Institute for Software Systems)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Created the XLOGOMINIPROG visual programming program synthesis benchmark to evaluate the performance of large models on tasks requiring multiple skills (spatial planning, logical reasoning, arithmetic, loops, variables, etc.), and proposed a fine-tuning pipeline based on synthetic data;
Programming by Example meets Historical Linguistics: A Large Language Model Based Approach to Sound Law Induction
Atharva Naik (Carnegie Mellon University), David R. Mortensen
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Convert sound change induction tasks into Programming by Example (PBE) problems, construct a low-resource evaluation benchmark, design multiple synthetic data generation methods with adjustable structural/substantial bias, and train the PySLICoder model using LLM fine-tuning.
Progressive Multimodal Reasoning via Active Retrieval
Guanting Dong (Renmin University of China), Ji-Rong Wen (Renmin University of China)
TransformerLarge Language ModelReinforcement LearningMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a unified framework AR-MCTS, which dynamically acquires high-quality reasoning prompts at each step of Monte Carlo Tree Search (MCTS) through active retrieval, and gradually trains the process reward model (PRM) using step-level annotations generated during the process, thereby achieving multi-step reasoning verification and improvement for multimodal models.
ProMALex: Progressive Modular Adapters for Multi-Jurisdictional Legal Language Modeling
Santosh T.y.s.s, Mohamed Hesham Elganayni (Technical University of Munich)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: In multi-jurisdictional legal corpora, a hierarchical adapter framework (ProMALex) was constructed through parameter-efficient LoRA adapters, achieving fine-grained adaptation for multi-jurisdictional legal language models by progressively sharing and specializing layers.
Prompt Candidates, then Distill: A Teacher-Student Framework for LLM-driven Data Annotation
Mingxuan Xia (Zhejiang University), Runze Wu (NetEase Fuxi AI Lab)
Knowledge DistillationData-Centric LearningTransformerPrompt EngineeringText
🎯 What it does: Propose the CanDist framework, which first uses a large language model (LLM) to generate candidate labels, then employs a small language model (SLM) for knowledge distillation to address mislabeling caused by single-label assignments.
Prompt-Guided Internal States for Hallucination Detection of Large Language Models
Fujie Zhang (Nankai University), Zheli Liu (Nankai University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Construct the PRISM framework by prompting the internal states of LLMs to enhance cross-domain hallucination detection performance.
PROPER: A Progressive Learning Framework for Personalized Large Language Models with Group-Level Adaptation
Linhai Zhang (King's College London), Yulan He (King's College London)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Propose the PROPER framework, achieving LLM personalization through hierarchical fine-tuning using LoRA and MoE at multiple levels (population-level, group-level, user-level);
ProtoLens: Advancing Prototype Learning for Fine-Grained Interpretability in Text Classification
Bowen Wei (George Mason University), Ziwei Zhu (George Mason University)
ClassificationExplainability and InterpretabilityTransformerTextBenchmark
🎯 What it does: Propose the ProtoLens model to achieve fine-grained clause-level interpretable text classification.
ProvBench: A Benchmark of Legal Provision Recommendation for Contract Auto-Reviewing
Xiuxuan Shen (Xidian University), Philip S. Yu (University of Illinois Chicago)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose PROVBENCH as a benchmark for legal clause recommendation and conflict detection in automated contract review, and construct the PROVDATA dataset
ProxAnn: Use-Oriented Evaluations of Topic Models and Document Clustering
Alexander Hoyle (ETH Zürich), Philip Resnik (University of Maryland)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Designed a human-assessment-based topic model and document clustering evaluation protocol, and proposed an LLM proxy named PROXANN that can replace humans.
Proxy-Driven Robust Multimodal Sentiment Analysis with Incomplete Data
Aoqiang Zhu (Hefei University of Technology), Ning An (Hefei University of Technology)
ClassificationData-Centric LearningTransformerAuto EncoderMultimodality
🎯 What it does: Propose a robust multimodal sentiment analysis method for data centers, P-RMF, achieving higher robustness and accuracy under scenarios with randomly missing multimodal data.
PsyAdvisor: A Plug-and-Play Strategy Advice Planner with Proactive Questioning in Psychological Conversations
Yuxin Hu (Southeast University), Yan Liu (Nanjing Derong Wisdom Information Technology Co., Ltd.)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposed a plug-and-play plugin called PsyAdvisor, enabling psychological LLMs to proactively determine the timing of questions and provide strategic advice during conversations, thereby enhancing the depth and quality of psychological counseling dialogues.
PsyDial: A Large-scale Long-term Conversational Dataset for Mental Health Support
Huachuan Qiu (Zhejiang University), Zhenzhong Lan (Westlake University)
Data SynthesisSafty and PrivacyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Developed a privacy-preserving dialogue reconstruction method RMRR for generating the long-term psychological counseling dialogue dataset PsyDial.
PsyDT: Using LLMs to Construct the Digital Twin of Psychological Counselor with Personalized Counseling Style for Psychological Counseling
Haojie Xie (South China University of Technology), Xiangmin Xu (South China University of Technology)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Built a framework based on LLM (PsyDT) that can quickly generate personalized counseling-style digital twin models of psychological counselors.
PTQ1.61: Push the Real Limit of Extremely Low-Bit Post-Training Quantization Methods for Large Language Models
Jiaqi Zhao (Harbin Institute of Technology (Shenzhen)), Min Zhang (Harbin Institute of Technology (Shenzhen))
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed an extremely low-bit post-training quantization method, PTQ 1.61, which can quantize LLM weights to 1.61 bits while maintaining excellent performance.
PunchBench: Benchmarking MLLMs in Multimodal Punchline Comprehension
Kun Ouyang (Peking University), Xu Sun (Peking University)
Large Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and constructed a multimodal punchline understanding benchmark called PunchBench, and designed a simple-to-complex chained questioning (SC-CoQ) method tailored to this benchmark to enhance the performance of MLLMs.
PVP: An Image Dataset for Personalized Visual Persuasion with Persuasion Strategies, Viewer Characteristics, and Persuasiveness Ratings
Junseo Kim (Seoul National University), Yohan Jo (Seoul National University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the PVP (Personalized Visual Persuasion) dataset, containing 28,454 visual persuasion images based on 596 informational messages (e.g., 'Don't smoke'), covering 9 persuasion strategies; simultaneously collected 2,521 raters' persuasiveness scores, demographic characteristics, and psychological traits (Big Five personality, values, moral foundations). On this basis, two tasks were proposed and implemented: ① Using an estimator to predict image persuasiveness scores; ② Generating personalized persuasive images tailored to individual psychological traits.
PwnGPT: Automatic Exploit Generation Based on Large Language Models
Wanzong Peng (Harbin Institute of Technology), Chen Zhang (China Mobile Group Design Institute Company Ltd)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed PwnGPT, an LLM-based automated exploit framework that improves the efficiency of automatically exploiting binary vulnerabilities (CTF challenges) through a modular workflow (analysis, generation, verification).
QAEncoder: Towards Aligned Representation Learning in Question Answering Systems
Zhengren Wang (Peking University), Wentao Zhang (Peking University)
Representation LearningLarge Language ModelTextBenchmark
🎯 What it does: Proposed QAEncoder, which eliminates the semantic gap between documents and user queries by generating diverse queries and taking their expected value in the vector space as the document representation.
QAEval: Mixture of Evaluators for Question-Answering Task Evaluation
Tan Yue (Peking University), Dongyan Zhao (Peking University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark
🎯 What it does: Propose the QAEval hybrid evaluation framework, combining answer extraction, rule-based rapid scoring, and MOE evaluation to efficiently and accurately assess QA tasks.
QDTSynth: Quality-Driven Formal Theorem Synthesis for Enhancing Proving Performance of LLMs
Lei Wang (East China Normal University), Zhengfeng Yang (Henan University)
Data SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Synthesize a high-quality Lean4 formal theorem dataset through adaptive MCTS, cluster diversity screening, and self-evaluation mechanisms, and use this dataset to supervise the fine-tuning of open-source LLMs, thereby improving their proof success rate on the miniF2F benchmark.
QG-SMS: Enhancing Test Item Analysis via Student Modeling and Simulation
Bang Nguyen (University of Notre Dame), Meng Jiang (University of Notre Dame)
TransformerLarge Language ModelText
🎯 What it does: Constructing test question pairs of varying quality, utilizing LLM for student modeling and simulation to assess the educational value of generated questions.
QQSUM: A Novel Task and Model of Quantitative Query-Focused Summarization for Review-based Product Question Answering
An Quang Tang (RMIT University), Zhuang Li (RMIT University)
GenerationRetrievalTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose the QQSUM task and develop the QQSUM-RAG model to generate multi-perspective quantitative summaries based on product reviews, enhancing the diversity and accuracy of question-answering.
Quaff: Quantized Parameter-Efficient Fine-Tuning under Outlier Spatial Stability Hypothesis
Hong Huang (City University of Hong Kong), Dapeng Wu (City University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a Quantized Parameter-efficient Fine-tuning framework called Quaff, based on the outlier spatial stability hypothesis, which enables efficient fine-tuning with low-bit quantization on Large Language Models (LLMs).
QualiSpeech: A Speech Quality Assessment Dataset with Natural Language Reasoning and Descriptions
Siyin Wang (Tsinghua University), Chao Zhang (Tsinghua University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkAudio
🎯 What it does: Proposed the QualiSpeech dataset and benchmark, assessing speech quality using natural language descriptions;
Quantification of Large Language Model Distillation
Sunbowen Lee (Shenzhen Institutes of Advanced Technology), Shiwen Ni (Shenzhen Institutes of Advanced Technology)
Explainability and InterpretabilityKnowledge DistillationLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a framework to quantify the distillation degree of large language models (LLMs), primarily through two metrics: Identity Consistency Evaluation (ICE) and Response Similarity Evaluation (RSE);
Quantifying Lexical Semantic Shift via Unbalanced Optimal Transport
Ryo Kishino (Kyoto University), Hidetoshi Shimodaira (Kyoto University)
Explainability and InterpretabilityRepresentation LearningText
🎯 What it does: This paper utilizes Unbalanced Optimal Transport (UOT) to match context embedding sets of the same word in corpora from different periods, and proposes the Sense Usage Shift (SUS) metric to quantify the frequency changes of each usage instance, thereby enabling semantic change detection at both the sense and instance levels.
Quantifying Misattribution Unfairness in Authorship Attribution
Pegah Alipoormolabashi (Stony Brook University), Niranjan Balasubramanian (Stony Brook University)
ClassificationRetrievalTransformerText
🎯 What it does: Proposed and utilized the Misattribution Unfairness Index (MAUI) in author attribution tasks to quantify the unfairness of models when incorrectly attributing non-author text to specific authors;
Quantifying Semantic Emergence in Language Models
Hang Chen (Xi'an Jiaotong University), Wenya Wang (Nanyang Technological University)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed the Informational Emergence (IE) metric to quantitatively measure the ability of large language models to extract semantics from token-level during autoregressive processes;
Quantized Can Still Be Calibrated: A Unified Framework to Calibration in Quantized Large Language Models
Mingyu Zhong (University Of Houston), Na Zou (University Of Houston)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Research on uncertainty calibration for quantized large language models (LLMs), proposing the UBCE (Upper Bound Calibration Error) evaluation metric, and restoring calibration performance after quantization through soft-prompt tuning.
QuASAR: A Question-Driven Structure-Aware Approach for Table-to-Text Generation
WeiJie Liu, Fang Kong (Soochow University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularRetrieval-Augmented Generation
🎯 What it does: Propose the QuASAR framework, achieving table-to-text generation through three pre-training tasks: self-supervised structural-related questions, sentence reconstruction, and numerical summarization.
Query-driven Document-level Scientific Evidence Extraction from Biomedical Studies
Massimiliano Pronesti (IBM Research Europe), Yufang Hou (IT:U Interdisciplinary Transformation University)
RetrievalData-Centric LearningTransformerLarge Language ModelBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a document-level scientific evidence extraction task for clinical questions and releases the COCHRANEFOREST dataset based on Cochrane systematic reviews;
Qwen2.5-xCoder: Multi-Agent Collaboration for Multilingual Code Instruction Tuning
Jian Yang (Beihang University), Junyang Lin (Alibaba Group)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: Proposed a multilingual multi-agent collaborative framework to generate multilingual code instruction data for instruction fine-tuning of Qwen2.5-xCoder, significantly enhancing cross-language code generation capabilities.
R-Fairness: Assessing Fairness of Ranking in Subjective Data
Lorenzo Balzotti (Sapienza Università di Roma), Sihem Amer-Yahia (CNRS, University Grenoble Alpes)
Recommendation SystemText
🎯 What it does: Designed and evaluated a framework (R-Fairness) to assess ranking fairness across different reviewer groups in collaborative rating platforms, with experiments conducted on real-world Yelp and Amazon datasets.
R2-MultiOmnia: Leading Multilingual Multimodal Reasoning via Self-Training
Leonardo Ranaldi (University of Edinburgh), Giulia Pucci (University of Aberdeen)
Representation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes the R2-MultiOmnia framework, combining multilingual modal bridging and language-agnostic reasoning alignment to enhance multimodal multilingual reasoning capabilities through self-supervised demonstration generation and incremental reinforcement learning.
R2D2: Remembering, Replaying and Dynamic Decision Making with a Reflective Agentic Memory
Tenghao Huang (University of Southern California), Muhao Chen (University of California Davis)
TransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the R2D2 framework, combining two paradigms: Remember (replay buffer + A* search) and Reflect (error analysis and reflective memory), to enhance the navigation and execution capabilities of web agents.
RADAR: Enhancing Radiology Report Generation with Supplementary Knowledge Injection
Wenjun Hou (Hong Kong Polytechnic University), Jiang Liu (Hong Kong Polytechnic University)
GenerationTransformerLarge Language ModelImageTextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Propose the RADAR framework, which first generates reports using LLM, identifies reliable internal knowledge with an expert model, and then retrieves and fuses external supplementary knowledge to produce more accurate radiology reports.
RAEmoLLM: Retrieval Augmented LLMs for Cross-Domain Misinformation Detection Using In-Context Learning Based on Emotional Information
Zhiwei Liu (University of Manchester), Eduard Hovy (University of Melbourne)
ClassificationDomain AdaptationAnomaly DetectionTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the RAEmoLLM framework, which utilizes retrieval-augmented LLM (RAG) for cross-domain misinformation detection by using sentiment information as the retrieval basis, constructing unsupervised few-shot examples, and completing judgment through in-context learning;
RAG-Critic: Leveraging Automated Critic-Guided Agentic Workflow for Retrieval Augmented Generation
Guanting Dong (Renmin University of China), Ji-Rong Wen (Renmin University of China)
GenerationLarge Language ModelReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the RAG-Critic framework, which achieves automated error correction and improvement in RAG systems through an automated critic and agent-based workflow.
RAGEval: Scenario Specific RAG Evaluation Dataset Generation Framework
Kunlun Zhu (Tsinghua University), Maosong Sun (Tsinghua University)
GenerationData SynthesisRetrievalLarge Language ModelTextBiomedical DataBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: This paper proposes the RAGEval framework, which can automatically generate scenario-based RAG evaluation datasets, including structured documents, question-answer pairs, and reference snippets, and evaluates answers through key point extraction.
RankCoT: Refining Knowledge for Retrieval-Augmented Generation through Ranking Chain-of-Thoughts
Mingyan Wu, Ge Yu (Northeastern University)
GenerationRetrievalTransformerPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the RankCoT method, integrating reranking signals into Chain-of-Thought (CoT) generation to improve the knowledge refinement process of Retrieval-Augmented Generation (RAG).
Ranking Unraveled: Recipes for LLM Rankings in Head-to-Head AI Combat
Roland Daynauth (University of Michigan), Jason Mars (University of Michigan)
Large Language ModelTextBenchmark
🎯 What it does: This paper systematically evaluates and compares the performance of four traditional ranking algorithms (Elo, Bradley-Terry, Glicko, Markov Chain) in head-to-head evaluations of large language models (LLMs), and proposes best practices for selecting appropriate algorithms.
RARE: Retrieval-Augmented Reasoning Enhancement for Large Language Models
Hieu Tran (University of Massachusetts, Amherst), Hong Yu (University of Massachusetts, Amherst)
TransformerLarge Language ModelTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the RARE framework, integrating retrieval-augmented actions (A6, A7) and retrieval-augmented factual scorer (RAFS) into MCTS search to improve the accuracy and factual reliability of LLMs on medical and common-sense reasoning tasks.
RATIONALYST: Pre-training Process-Supervision for Improving Reasoning
Dongwei Jiang (Johns Hopkins University), Daniel Khashabi (Johns Hopkins University)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the RATIONALYST model, which utilizes implicitly self-supervised extracted reasoning steps from unlabelled text to guide LLM reasoning
Re-identification of De-identified Documents with Autoregressive Infilling
Lucas Georges Gabriel Charpentier (University of Oslo), Pierre Lison (Norwegian Computing Center)
RetrievalSafty and PrivacyTransformerLarge Language ModelTextElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-enhanced autoregressive cloze framework to reverse recover masked personally identifiable information (PII)
Re-ranking Using Large Language Models for Mitigating Exposure to Harmful Content on Social Media Platforms
Rajvardhan Oak (University of California, Davis), Anshuman Chhabra (University of South Florida)
Recommendation SystemSafty and PrivacyLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a re-ranking method based on a large language model (LLM) to reduce the probability of users being exposed to harmful content in social media recommendation sequences.
Re^{3}Syn: A Dependency-Based Data Synthesis Framework for Long-Context Post-training
Zhiyang Zhang (Central South University), De Wen Soh (Singapore University of Technology and Design)
Data SynthesisTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose a long-text synthesis framework RE SYN 3 based on semantic similarity retrieval, dependency identification, and re-ranking for generating high-quality long-context training data.
Read it in Two Steps: Translating Extremely Low-Resource Languages with Code-Augmented Grammar Books
Chen Zhang (Wangxuan Institute of Computer Technology, Peking University), Yansong Feng (Wangxuan Institute of Computer Technology, Peking University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Explored how to utilize grammar books in machine translation for extremely low-resource languages by splitting the process into two steps: rule retrieval and rule application, and proposed using code-based rules and a per-item retrieval strategy to improve translation effectiveness.
REAL-MM-RAG: A Real-World Multi-Modal Retrieval Benchmark
Navve Wasserman (IBM Research Israel), Leonid Karlinsky (IBM Research Israel)
RetrievalLarge Language ModelVision Language ModelImageTextMultimodalityTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose REAL-MM-RAG-Bench, a retrieval benchmark designed for real-world multimodal RAG systems, and evaluate the semantic robustness of models through multi-layer query rewriting;
Real-time Factuality Assessment from Adversarial Feedback
Sanxing Chen (Duke University), Bhuwan Dhingra (Duke University)
GenerationData SynthesisAdversarial AttackTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed a news rewriting generation pipeline based on adversarial iteration, utilizing feedback from retrieval-enhanced detectors to generate real-time fake news capable of misleading powerful LLM detectors.
Recent Advances in Speech Language Models: A Survey
Wenqian Cui (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
RecognitionGenerationTransformerLarge Language ModelFlow-based ModelGenerative Adversarial NetworkContrastive LearningMultimodalityReview/Survey PaperAudio
🎯 What it does: This paper reviews the latest research progress on SpeechLM, systematically organizing its three core components (speech tokenizer, language model, vocoder), training process (pre-training, instruction fine-tuning, post-alignment), and multi-dimensional evaluation methods, and proposes a complete technical roadmap;
RecLM: Recommendation Instruction Tuning
Yangqin Jiang (University of Hong Kong), Chao Huang (University of Hong Kong)
Recommendation SystemKnowledge DistillationGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a model-agnostic instruction tuning framework called RecLM, which combines large language models (LLMs) with collaborative filtering. It leverages multi-round dialogues and reinforcement learning to generate high-quality user/item features, achieving significant improvements in cold start and sparse data scenarios.
Reconsidering LLM Uncertainty Estimation Methods in the Wild
Yavuz Faruk Bakman (University of Southern California), Sai Praneeth Karimireddy (University of Southern California)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Systematically evaluated 19 LLM uncertainty estimation methods across four key challenges in real-world environments: threshold selection sensitivity, input transformation robustness, long-text applicability, and multi-method integration.
Recurrent Knowledge Identification and Fusion for Language Model Continual Learning
Yujie Feng (Hong Kong Polytechnic University), Xiao-Ming Wu (Hong Kong Polytechnic University)
OptimizationSafty and PrivacyRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the Recurrent-KIF framework to achieve dynamic parameter importance estimation for continual learning and multi-round knowledge fusion
Redundancy Principles for MLLMs Benchmarks
Zicheng Zhang (Shanghai AI Laboratory), Guangtao Zhai (Shanghai AI Laboratory)
Vision Language ModelBenchmark
🎯 What it does: This paper proposes a framework for evaluating redundancy in multimodal large language model (MLLM) benchmarks and systematically analyzes dimensional, instance, and cross-benchmark redundancies.
Ref-Long: Benchmarking the Long-context Referencing Capability of Long-context Language Models
Junjie Wu (Hong Kong University Of Science And Technology), Arman Cohan (Yale University)
TransformerSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Introduce the Ref-Long benchmark to assess the ability of LCLMs to locate document indices containing specific keywords in long contexts.
Refining Salience-Aware Sparse Fine-Tuning Strategies for Language Models
Xinxin Liu (Southern University of Science and Technology), Xitong Gao (Shenzhen Institutes of Advanced Technology, CAS)
Computational EfficiencyRepresentation LearningNeural Architecture SearchTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Studied a sparse parameter efficient fine-tuning (SPEFT) method, systematically evaluated multiple importance metrics to construct sparse masks, and compared the effects of static and dynamic masks.
ReflectDiffu: Reflect between Emotion-intent Contagion and Mimicry for Empathetic Response Generation via a RL-Diffusion Framework
Jiahao Yuan (University of Shanghai for Science and Technology), Usman Naseem (Macquarie University)
GenerationLarge Language ModelReinforcement LearningDiffusion modelAuto EncoderText
🎯 What it does: Propose the ReflectDiffu framework, combining emotion contagion encoding with an intent dual cycle (Explore-Sampling-Correct) mechanism integrated with reinforcement learning diffusion models to generate empathetic dialogue responses.
ReflectionCoder: Learning from Reflection Sequence for Enhanced One-off Code Generation
Houxing Ren (CUHK MMLab), Hongsheng Li (CUHK MMLab)
GenerationKnowledge DistillationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By leveraging reflection sequences generated through compiler feedback, the code language model is fine-tuned, significantly improving the accuracy of one-time code generation.
ReflecTool: Towards Reflection-Aware Tool-Augmented Clinical Agents
Yusheng Liao (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
TransformerLarge Language ModelAgentic AIMultimodalityBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the Clinical Agent Benchmark (CAB) (18 tasks, 5 dimensions) and the reflective tool-enhanced framework REFLECTOOL to improve LLM performance on clinical multimodal, multidimensional tasks.
RefreshKV: Updating Small KV Cache During Long-form Generation
Fangyuan Xu (New York University), Eunsol Choi (New York University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the RefreshKV inference method, which utilizes dynamic refreshing of small KV caches and alternates between full KV attention and small KV attention in long text generation to improve inference speed while reducing performance loss.
Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training
Youliang Yuan (School of Data Science, Chinese University of Hong Kong), Zhaopeng Tu (School of Data Science, Chinese University of Hong Kong)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Improve the safety of large language models by introducing Disentangled Rejection Training (DeRTa), enabling the model to identify and reject unsafe content at any position during the generation process.
Registering Source Tokens to Target Language Spaces in Multilingual Neural Machine Translation
Zhi Qu (Nara Institute of Science and Technology), Taro Watanabe (National Institute of Information and Communications Technology)
GenerationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied a method called Registering to address the off-target problem in multilingual neural machine translation, and trained the MITRE series of models based on this method.
Reinforced IR: A Self-Boosting Framework For Domain-Adapted Information Retrieval
Chaofan Li (Beijing University of Posts and Telecommunications), Zheng Liu (Beijing Academy of Artificial Intelligence)
RetrievalDomain AdaptationKnowledge DistillationTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes Reinforced-IR, a cross-domain retrieval framework that jointly adapts a retriever and a generator through a self-boosting mechanism.
RelationalCoder: Rethinking Complex Tables via Programmatic Relational Transformation
Haoyu Dong (Institute of Information Engineering, Chinese Academy of Sciences), Yanan Cao (Institute of Information Engineering, Chinese Academy of Sciences)
Data-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTabularFinance RelatedChain-of-Thought
🎯 What it does: Propose the RELATIONALCODER framework, which unifies complex semi-structured tables into SQL-compatible relational tables directly usable by SQL; simultaneously introduce Loop Reference Decoding (LRD) to achieve large table compression and avoid cell-level hallucinations.
ReLearn: Unlearning via Learning for Large Language Models
Haoming Xu (Zhejiang University), Ningyu Zhang (National University of Singapore)
Data SynthesisOptimizationSafty and PrivacyComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose ReLearn, a forgetting framework for LLMs based on positive optimization, achieving targeted knowledge forgetting through data augmentation and fine-tuning while maintaining model fluency and relevance.
Reliably Bounding False Positives: A Zero-Shot Machine-Generated Text Detection Framework via Multiscaled Conformal Prediction
Xiaowei Zhu (Institute Of Information Engineering Chinese Academy Of Sciences), Yangxi Li (National Computer Network Emergency Response Technical Team)
Anomaly DetectionTextBenchmark
🎯 What it does: Proposes a zero-shot machine-generated text detection framework based on multi-scale consistency prediction (MCP), and constructs a large-scale dataset called RealDet covering 15 domains and 22 large language models (LLMs).
Removal of Hallucination on Hallucination: Debate-Augmented RAG
Wentao Hu (Hong Kong Polytechnic University), Li Qing
GenerationRetrievalAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a training-free retrieval-augmented generation framework called DRAG, which eliminates 'retrieval-induced misinformation' and model hallucinations through a multi-agent debate mechanism in two stages: retrieval and generation.
RePanda: Pandas-powered Tabular Verification and Reasoning
Atoosa Chegini, Soheil Feizi (University of Maryland)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTabularBenchmark
🎯 What it does: Propose RePanda, a structured reasoning framework that converts natural language statements into executable pandas queries for table fact verification and question answering.
Representation Bending for Large Language Model Safety
Ashkan Yousefpour (Seoul National University), Jonghyun Choi (Seoul National University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the REPBEND method, which enhances the safety of large language models by 'bending' the model's internal representations during training, bringing safe representations closer and dangerous representations further apart.
Representations of Fact, Fiction and Forecast in Large Language Models: Epistemics and Attitudes
Meng Li (University of Potsdam), David Schlangen (University of Potsdam)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Evaluate the cognitive modal knowledge of open-weight LLMs in expressing facts, imagination, and prediction through experiments on controlled stories, revealing their limitations in generating expressions of uncertainty.
ReSCORE: Label-free Iterative Retriever Training for Multi-hop Question Answering with Relevance-Consistency Supervision
Dosung Lee (Korea University), Paul Hongsuck Seo (Korea University)
RetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose an iterative retrieval training method called ReSCORE for unlabeled documents, combined with the multi-hop question answering framework IQATR to improve retrieval and answering quality.
Research Borderlands: Analysing Writing Across Research Cultures
Shaily Bhatt (Carnegie Mellon University), Maria Antoniak (University of Copenhagen)
TransformerLarge Language ModelText
🎯 What it does: This paper collects the experiences of interdisciplinary researchers on writing papers in different research cultures through interviews and surveys, constructs a writing standards framework, quantifies it using computational metrics, and subsequently evaluates the cultural adaptability of LLMs in cross-cultural writing.
Response Wide Shut? Surprising Observations in Basic Vision Language Model Capabilities
Shivam Chandhok (University of British Columbia), Leonid Sigal (University of British Columbia)
ClassificationRecognitionVision Language ModelImageTextMultimodality
🎯 What it does: Systematically evaluate the capabilities of multiple vision-language models (VLMs) on basic visual tasks (coarse-to-fine grain classification, counting, spatial relationships), and construct a three-layer subspace (visual, visual-language projection, response) to detect specific locations of information loss.
Rethinking Evaluation Metrics for Grammatical Error Correction: Why Use a Different Evaluation Process than Human?
Takumi Goto (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
Data-Centric LearningText
🎯 What it does: Researchers propose a new automatic evaluation process that converts sentence-level scores of existing evaluation metrics into pairwise comparisons, and then uses the same TrueSkill algorithm as human assessments to generate system rankings, thereby narrowing the gap between automatic evaluation and human evaluation.
Rethinking KenLM: Good and Bad Model Ensembles for Efficient Text Quality Filtering in Large Web Corpora
Yungi Kim (Liner), Chanjun Park (Korea University)
Computational EfficiencyData-Centric LearningMixture of ExpertsText
🎯 What it does: Proposed a new integrated method that utilizes two contrasting KenLM models (Good KenLM and Bad KenLM) to effectively filter low-quality text from large web corpora while preserving high-quality text.
Rethinking Repetition Problems of LLMs in Code Generation
Yihong Dong (Peking University), Ge Li (Peking University)
GenerationAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Study and address the problem of structural repetition in large language models during code generation, proposing a grammar-based decoding method (RPG) to detect and penalize repetitive fragments.
Rethinking Reward Model Evaluation Through the Lens of Reward Overoptimization
Sunghwan Kim (Yonsei University), Jinyoung Yeo (Yonsei University)
OptimizationReinforcement LearningTextBenchmark
🎯 What it does: Investigated the relationship between reward model evaluation and reward over-optimization, and proposed constructing a more reliable benchmark by measuring over-optimization.
Rethinking Semantic Parsing for Large Language Models: Enhancing LLM Performance with Semantic Hints
Kaikai An (Peking University), Baobao Chang (Peking University)
Data-Centric LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Explores the role of semantic parsing on large language models (LLMs) and proposes a novel prompting method called SENSE, which enhances LLM performance in understanding and generation tasks through semantic prompts.
Rethinking the Role of Prompting Strategies in LLM Test-Time Scaling: A Perspective of Probability Theory
Yexiang Liu (MAIS, Institute of Automation, Chinese Academy of Sciences), Tieniu Tan (MAIS, Institute of Automation, Chinese Academy of Sciences)
TransformerPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Investigate the performance of different prompting strategies under computational amplification (majority voting) during LLM testing, propose a method for quickly predicting amplification performance, and further provide improvement strategies.
Retrieval-Augmented Fine-Tuning With Preference Optimization For Visual Program Generation
Deokhyung Kang (POSTECH), Gary Lee
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextGraphRetrieval-Augmented Generation
🎯 What it does: A two-stage training method combining retrieval-augmented fine-tuning and graph-editing-based preference optimization is proposed for generating industrial visual programming languages (Ladder Diagram).
Retrieve to Explain: Evidence-driven Predictions for Explainable Drug Target Identification
Ravi Patel (BenevolentAI), Dane S. Corneil (BenevolentAI)
Explainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a retrieval-driven model named R2E (Retrieve to Explain), which scores and ranks each candidate answer (e.g., genes) based on retrieved evidence in high-risk scientific problems such as drug target identification, and quantitatively explains the contribution of each piece of evidence to the answer score using Shapley values.
Retrofitting Large Language Models with Dynamic Tokenization
Darius Feher (University of Cambridge), Benjamin Minixhofer (University of Cambridge)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Developed a dynamic tokenization method based on BPE merging and pre-trained super networks, which can generate token embeddings in real-time during inference, replacing traditional static subword tokenizers to improve the efficiency and fairness of multilingual models.
RetroLLM: Empowering Large Language Models to Retrieve Fine-grained Evidence within Generation
Xiaoxi Li (Renmin University of China), Zhicheng Dou
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose RetroLLM, a unified autoregressive framework that directly performs retrieval and generation within the LLM through FM-Index constraints;
Retrospective Learning from Interactions
Zizhao Chen (Cornell University), Yoav Artzi (Cornell University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringImageTextMultimodality
🎯 What it does: Propose a framework based on retrospection learning (RESPECT), enabling large language models to self-learn and improve from implicit feedback (e.g., frustration, confirmation, restatement) during multi-round interactions with humans.