EMNLP 2025 Papers — Page 16
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Supervised Attention Mechanism for Low-quality Multimodal Data
Sijie Mai (South China Normal University), Haifeng Hu (Sun Yat-sen University)
Representation LearningData-Centric LearningTransformerLarge Language ModelMultimodality
🎯 What it does: Proposed the SAM-LML framework, which jointly addresses missing and noisy data in multimodal data, and enhances fusion robustness through a supervised attention mechanism.
SURE: Safety Understanding and Reasoning Enhancement for Multimodal Large Language Models
Yuxin Gou (Hefei University of Technology), Wenbo Hu (Hefei University of Technology)
Safty and PrivacySupervised Fine-TuningMultimodalityChain-of-Thought
🎯 What it does: Propose the SURE framework, training MLLM to achieve safe intent recognition and rejection for multi-modal inputs through chain-of-thought reasoning.
Surge: On the Potential of Large Language Models as General-Purpose Surrogate Code Executors
Bohan Lyu (Tsinghua), Jiaming Zhang (Tsinghua)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Designed and constructed the SURGE benchmark to systematically evaluate the ability of large language models (LLMs) to predict code execution outcomes without executing the code; conducted comprehensive experiments on 21 open-source and closed-source LLMs to explore the impact of scale, data volume, and prompting strategies on prediction accuracy.
Surprise Calibration for Better In-Context Learning
Zhihang Tan (Wuhan University), Peng Zhu (Nanjing University of Science and Technology)
Computational EfficiencyRepresentation LearningMeta LearningRecurrent Neural NetworkLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a calibration method based on surprise (Surprise Calibration, SC), dynamically adjusting category priors in in-context learning (ICL) to enhance the adaptability of large language models.
SurveyGen: Quality-Aware Scientific Survey Generation with Large Language Models
Tong Bao (Nanjing University of Science and Technology), Chengzhi Zhang (Nanjing University of Science and Technology)
GenerationTransformerLarge Language ModelTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Constructed a large-scale scientific review dataset named SurveyGen and proposed a quality-aware retrieval framework QUAL-SG to evaluate and enhance the performance of LLMs in automatically generating scientific reviews.
SWAM: Adaptive Sliding Window and Memory-Augmented Attention Model for Rumor Detection
Mei Guo (Nankai University), Xiaojie Yuan (Nankai University)
ClassificationGraph Neural NetworkTransformerText
🎯 What it does: Proposed a new model called SWAM for detecting social media rumors using adaptive sliding windows and memory-enhanced attention.
SWAN: An Efficient and Scalable Approach for Long-Context Language Modeling
Krishna C Puvvada, Boris Ginsburg (NVIDIA)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a decoder-specific Transformer architecture named SWAN, which achieves robust inference for sequences with training lengths much longer than training durations by alternately using NoPE layers and sliding window RoPE layers, and applies dynamic logarithmic scaling of attention weights during inference to enhance long-context performance.
SwarmAgentic: Towards Fully Automated Agentic System Generation via Swarm Intelligence
Yao Zhang (LMU Munich), Volker Tresp (LMU Munich)
OptimizationLarge Language ModelAgentic AIText
🎯 What it does: Proposed the SwarmAgentic framework, utilizing language-based Particle Swarm Optimization (PSO) to autonomously generate, optimize, and collaborate multi-agent systems from scratch without any human intervention.
SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation
Aurick Qiao (Snowflake AI Research), Yuxiong He (Snowflake AI Research)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This study proposes SwiftKV, a technique that restructures LLMs to skip subsequent transformer layers during inference (only for prefix/prompt tokens) and directly generate KV caches for the remaining layers using the hidden states from the previous layer, while combining AcrossKV for cross-layer KV cache sharing and lightweight knowledge retention distillation, significantly reducing prefix computational costs and compressing KV caches;
Sycophancy Mitigation Through Reinforcement Learning with Uncertainty-Aware Adaptive Reasoning Trajectories
Mohammad Beigi (University of California Davis), Lifu Huang (University of California Davis)
OptimizationSafty and PrivacyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the SMART framework, which generates high-quality reasoning trajectories in two stages through uncertainty-aware adaptive Monte Carlo Tree Search (UA-MCTS), and employs dense progress rewards for reinforcement learning to alleviate the sycophancy behavior of LLMs.
SynC-LLM: Generation of Large-Scale Synthetic Circuit Code with Hierarchical Language Models
Shang Liu (Hong Kong University of Science and Technology), Zhiyao Xie (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextGraphRetrieval-Augmented Generation
🎯 What it does: Introduces SynC-LLM, a three-stage hierarchical framework for generating synthetic circuit code at the million-gate level, addressing the scarcity of public circuit data.
SYNC: A Synthetic Long-Context Understanding Benchmark for Controlled Comparisons of Model Capabilities
Shuyang Cao (University of Michigan), Lu Wang (University of Michigan)
RetrievalLarge Language ModelTextGraphBenchmarkChain-of-Thought
🎯 What it does: Designed and released the SYNC benchmark, which includes three types of tasks involving graph structures and synthetic language translation, unifying long-text contexts to evaluate the long-context understanding capabilities of large language models (LLMs).
Synergizing Multimodal Temporal Knowledge Graphs and Large Language Models for Social Relation Recognition
Haorui Wang (Beijing University of Posts and Telecommunications), Bin Wu (Beijing University of Posts and Telecommunications)
RecognitionTransformerLarge Language ModelVideoMultimodalityRetrieval-Augmented Generation
🎯 What it does: A method combining multimodal temporal knowledge graphs with large language models to achieve social relationship recognition in videos.
Syntax-Aware Retrieval Augmentation for Neural Symbolic Regression
Canmiao Zhou (South China University of Technology), Han Huang (Guangzhou Zhisuan Linghang Technology Co., Ltd.)
RetrievalOptimizationRepresentation LearningData-Centric LearningTransformerTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a grammar-aware retrieval-enhanced neural symbolic regression method (SRASR), which improves the fitting accuracy of expressions by leveraging retrieved prior grammar structures during the autoregressive generation process of a pre-trained model.
Synth-SBDH: A Synthetic Dataset of Social and Behavioral Determinants of Health for Clinical Text
Avijit Mitra (University of Massachusetts Amherst), Hong Yu (U.S. Department of Veterans Affairs)
Data SynthesisExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Constructed Synth-SBDH, a publicly available synthetic Social and Behavioral Determinants of Health (SBDH) dataset covering 15 SBDH categories, with fine-grained annotations including presence, tense, and rationale.
Synthetic Socratic Debates: Examining Persona Effects on Moral Decision and Persuasion Dynamics
Jiarui Liu (Carnegie Mellon University), Maarten Sap (Carnegie Mellon University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically investigates the impact of personality traits on moral judgment and persuasion strategies by conducting single-round judgments and multi-round debates on 131 relational everyday moral dilemmas using six dimensions (age, gender, country, social class, political ideology, and the Big Five personality traits) across four large language models.
T-MAD: Target-driven Multimodal Alignment for Stance Detection
ZhaoDan Zhang, Hui Xu (Chinese Academy of Sciences)
ClassificationTransformerContrastive LearningMultimodality
🎯 What it does: Proposed a target-driven multi-modal alignment and dynamic weighting stance detection model (T-MAD), incorporating iterative reasoning to enhance performance.
T^2: An Adaptive Test-Time Scaling Strategy for Contextual Question Answering
Zhengyi Zhao (Chinese University of Hong Kong), Xian Wu (Tencent Jarvis Lab)
Computational EfficiencyTransformerMixture of ExpertsTextChain-of-Thought
🎯 What it does: Propose a framework named T2, which dynamically selects reasoning strategies to adapt to problem complexity by leveraging problem structure decomposition, similar example generation, and multi-criteria matching, thereby balancing accuracy and efficiency in context-aware question answering.
T2R-BENCH: A Benchmark for Real World Table-to-Report Task
Jie Zhang (Institute of Artificial Intelligence (TeleAI), China Telecom), Xuelong Li (Institute of Artificial Intelligence (TeleAI), China Telecom)
GenerationLarge Language ModelTextTabularBenchmark
🎯 What it does: Proposed the 'table-to-report' task and constructed a bilingual benchmark T2R-bench for industrial real-world tables (457 tables, 910 questions, and 4,320 report key points), while providing a three-dimensional evaluation system.
Table-LLM-Specialist: Language Model Specialists for Tables using Iterative Fine-tuning
Junjie Xing (University of Michigan), Surajit Chaudhuri (Microsoft Corporation)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTabular
🎯 What it does: Proposed a self-supervised fine-tuning framework called TABLE-SPECIALIST based on a generator-validator cycle, designed to train specialized language models for table tasks without relying on human annotations;
Table-R1: Inference-Time Scaling for Table Reasoning Tasks
Zheyuan Yang (Yale NLP Lab), Yilun Zhao (Yale NLP Lab)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTabularBenchmarkChain-of-Thought
🎯 What it does: This paper explores the application of inference time scaling in table reasoning tasks and proposes two post-training methods: distilling the inference trajectories of leading models (Table-R1-SFT) and reinforcement learning based on verifiable rewards (RLVR, Table-R1-Zero);
TableEval: A Real-World Benchmark for Complex, Multilingual, and Multi-Structured Table Question Answering
Junnan Zhu (Beijing Wenge Technology Co., Ltd.), Nan Xu (State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, CAS)
Large Language ModelTabularBenchmark
🎯 What it does: Constructed the TableEval benchmark, covering real-world table question-answering tasks with multi-structural, multi-lingual, and multi-domain characteristics, and proposed the SEAT evaluation framework.
TableRAG: A Retrieval Augmented Generation Framework for Heterogeneous Document Reasoning
Xiaohan Yu (Huawei Cloud BU), Chong Chen (Huawei Cloud BU)
GenerationRetrievalTransformerLarge Language ModelTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed TableRAG, an SQL-based retrieval-augmented generation framework specialized for multi-hop question answering over heterogeneous documents (text + tables), and constructed a new HeteQA benchmark dataset.
TACO: Enhancing Multimodal In-context Learning via Task Mapping-Guided Sequence Configuration
Yanshu Li (Brown University), Ruixiang Tang (University of Bristol)
ClassificationTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a multimodal ICL analysis framework based on task mapping, and design a lightweight Transformer model TACO, which dynamically configures ICL sequences through task-aware attention to enhance LVLM inference performance.
TactfulToM: Do LLMs have the Theory of Mind ability to understand White Lies?
Yiwei Liu (EPFL), Saku Sugawara (National Institute of Informatics)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Designed and constructed the TactfulToM benchmark to evaluate large language models' understanding and reasoning abilities regarding white lies in real-world dialogues.
Tailoring Table Retrieval from a Field-aware Hybrid Matching Perspective
Da Li (Chinese Academy Of Sciences), Xueqi Cheng (University Of Chinese Academy Of Sciences)
RetrievalTransformerLarge Language ModelMixture of ExpertsContrastive LearningTabularBenchmark
🎯 What it does: Propose THYME table retriever, which utilizes field-aware hybrid dense-sparse matching to learn different matching preferences for table titles, headers, and cells.
Taking Notes Brings Focus? Towards Multi-Turn Multimodal Dialogue Learning
Jiazheng Liu (Peking University), Zongqing Lu (Peking University)
TransformerLarge Language ModelVision Language ModelMultimodalityTabularBenchmarkChain-of-Thought
🎯 What it does: Proposed a multi-turn multi-modal dialogue dataset MMDiag and its extended version MMDiag-E, and designed the DiagNote model, enhancing the multi-turn dialogue and visual localization capabilities of MLLMs through interactive reasoning and visual focusing mechanisms.
TALON: A Multi-Agent Framework for Long-Table Exploration and Question Answering
Ruochun Jin, Meng Zhang (National University Of Defense Technology)
Agentic AITabularChain-of-Thought
🎯 What it does: Designed a multi-agent framework called TALON to address question-answering tasks involving long tables (50~10,000+ rows).
Taming Text-to-Image Synthesis for Novices: User-centric Prompt Generation via Multi-turn Guidance
Yilun Liu (Huawei), Osamu Yoshie (Huawei)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: Propose a text-to-image (prompt) generation framework called DialPrompt based on multi-turn dialogue, helping novice users gradually optimize prompts and generate high-quality images through conversations.
TAPS: Tool-Augmented Personalisation via Structured Tagging
Ekaterina Taktasheva (University of Edinburgh), Jeff Dalton (University of Edinburgh)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied the personalization of tool-enhanced LLMs, finding that existing models often suffer from semantic errors, missing parameters, and hallucinations when integrating user preferences. The proposed TAPS method improves personalized tool usage through structured tagging and uncertainty-based tool detection.
Targeted Distillation for Sentiment Analysis
Yice Zhang (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
ClassificationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a two-stage knowledge distillation framework (KNOWDIST+ICLDIST) for sentiment analysis, and constructed a comprehensive sentiment analysis benchmark called SENTIBENCH containing 12 datasets.
Task-aware Contrastive Mixture of Experts for Quadruple Extraction in Conversations with Code-like Replies and Non-opinion Detection
Chenyuan He (Zhengzhou University), Min Peng (Wuhan University)
TransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: Propose a multi-task framework called TaCoMoE based on large language models for extracting quadruples (goal, aspect, opinion, sentiment) from multi-turn dialogues and detecting non-opinion utterances;
Task-Aware Resolution Optimization for Visual Large Language Models
Weiqing Luo (University of North Carolina at Chapel Hill), Tianlong Chen (University of North Carolina at Chapel Hill)
OptimizationLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper addresses the varying resolution requirements of visual large language models across different tasks and proposes a task-aware resolution optimization method.
TASO: Task-Aligned Sparse Optimization for Parameter-Efficient Model Adaptation
Daiye Miao (East China Normal University), Yuanbin Wu (East China Normal University)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the TASO method, which sparsifies LoRA on task-specific core regions to significantly reduce trainable parameters while maintaining performance.
TaxoAlign: Scholarly Taxonomy Generation Using Language Models
Avishek Lahiri (Indian Association for the Cultivation of Science), Debarshi Kumar Sanyal (Indian Association for the Cultivation of Science)
ClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a three-stage pipeline (TAXOALIGN) based on large language models for automatically generating hierarchical taxonomies in academic survey papers;
TCP: a Benchmark for Temporal Constraint-Based Planning
Zifeng Ding (University of Cambridge), Andreas Vlachos (University of Cambridge)
Large Language ModelTextBenchmark
🎯 What it does: Proposed the TCP benchmark to evaluate the capabilities of large language models in multi-constraint temporal planning tasks.
TCPO: Thought-Centric Preference Optimization for Effective Embodied Decision-making
Kechen Jiao (Tsinghua University), Xiu Li (Northeastern University)
OptimizationRobotic IntelligenceReinforcement LearningTextSequentialChain-of-Thought
🎯 What it does: Proposes a thought-directed preference optimization (TCPO) framework for embodied artificial intelligence, leveraging step-by-step preference learning and consistency constraints in a chained reasoning process to enhance decision efficiency and robustness in dynamic environments.
Teach Small Models to Reason by Curriculum Distillation
Wangyi Jiang (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
Knowledge DistillationLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented a two-phase curriculum distillation framework to transfer implicit reasoning and explicit chain-of-thought reasoning from large inference models to small models.
Teaching Your Models to Understand Code via Focal Preference Alignment
Jie Wu, Scarlett Li (Tsinghua University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningText
🎯 What it does: Studied a fine-grained preference alignment framework called Target-DPO based on iterative debugging, aiming to enable code generation models to more accurately learn error correction strategies.
Temporal Referential Consistency: Do LLMs Favor Sequences Over Absolute Time References?
Ashutosh Bajpai (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Proposes a new multilingual benchmark, TEMP-ReCon, for evaluating time reference consistency of large language models under different temporal references (absolute time vs. event sequence), and designs the UnTRaP method to enhance consistency by aligning event-oriented and time-oriented reasoning paths.
Temporal Scaling Law for Large Language Models
Yizhe Xiong (Tsinghua University), Guiguang Ding (Tsinghua University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Studied and proposed the temporal scaling law of test loss evolution during the pre-training of large language models, based on dynamic hyperbolic modeling according to token positions to achieve future loss prediction;
TempParaphraser: “Heating Up” Text to Evade AI-Text Detection through Paraphrasing
Junjie Huang (Xiamen University), Yidong Chen (Xiamen University)
Data SynthesisAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose the TempParaphraser framework, which uses multiple normal-temperature sampling to simulate high-temperature sampling to evade AI text detection.
Text Detoxification: Data Efficiency, Semantic Preservation and Model Generalization
Jing Yu (East China Normal University), Xiang Li (Shanghai EastWonder Info-tech Co., Ltd)
GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a two-stage training framework named GEM. It first performs supervised fine-tuning using a small amount of high-quality filtered parallel aligned data, and then employs unannotated toxic text and a custom reward function through GRPO (Group Relative Policy Optimization) for reinforcement learning, achieving text detoxification while maintaining semantic consistency.
Text Meets Topology: Rethinking Out-of-distribution Detection in Text-Rich Networks
Danny Wang (University of Queensland), Zi Huang (University of Queensland)
Anomaly DetectionGraph Neural NetworkTransformerContrastive LearningTextGraphBenchmark
🎯 What it does: This paper proposes a benchmark framework for out-of-distribution (OOD) detection in text-network (TrN) scenarios, named TextTopoOOD, and designs a novel detection model called TNT‑OOD based on this framework.
Text Takes Over: A Study of Modality Bias in Multimodal Intent Detection
Ankan Mullick (IIT Kharagpur), Pawan Goyal (IIT Kharagpur)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodalityAudio
🎯 What it does: Study the multimodal intent detection task, investigate the impact of text bias on model performance, and propose a debiasing method.
Text2Vis: A Challenging and Diverse Benchmark for Generating Multimodal Visualizations from Text
Mizanur Rahman (York University), Enamul Hoque (York University)
GenerationTransformerLarge Language ModelReinforcement LearningAgentic AIMultimodalityTabularBenchmarkFinance Related
🎯 What it does: Proposed the Text2Vis benchmark dataset to evaluate the performance of text-to-visualization models and introduced a cross-modal actor-critic improvement framework
Textual Aesthetics in Large Language Models
Lingjie Jiang (Microsoft Research), Furu Wei (Microsoft Research)
GenerationOptimizationLarge Language ModelSupervised Fine-TuningReinforcement LearningImageText
🎯 What it does: This paper constructs the first LLM text aesthetics dataset, TEXAES, and proposes the TAPO fine-tuning method based on this dataset, significantly improving the readability, layout, consistency, and coherence of LLM-generated text.
TFDP: Token-Efficient Disparity Audits for Autoregressive LLMs via Single-Token Masked Evaluation
Inderjeet Singh (Fujitsu Research of Europe), Hisashi Kojima (Fujitsu Limited)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose Token-Focused Disparity Probing (TFDP), a word-level disparity auditing method for autoregressive large language models.
The Arabic Generality Score: Another Dimension of Modeling Arabic Dialectness
Sanad Sha’ban, Nizar Habash (Mohamed bin Zayed University of Artificial Intelligence)
Representation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the Arabic Generalization Score (AGS) and constructed a complete process for annotating, calculating, and predicting word-level AGS from parallel corpora, subsequently extending it to the sentence level.
The discordance between embedded ethics and cultural inference in large language models
Aida Ramezani (University of Toronto), Yang Xu (University of Toronto)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper investigates the differences between ethical representations embedded in large language models (LLMs) and their inference of cultural norms in cross-cultural contexts, proposing and verifying the 'ethical-cultural inconsistency' hypothesis.
The Emperor’s New Reasoning: Format Imitation Overshadows Genuine Mathematical Understanding in SFT
Linyao Yang (Zhejiang Lab), Guirong Xue (Zhejiang Lab)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: By performing 10-sample supervised fine-tuning (SFT) on a small-scale LLM, the study systematically evaluates the relationship between format alignment and the enhancement of genuine reasoning capabilities, verifying that significant performance improvements mainly stem from imitating the 'step-by-step reasoning format.'
The Enemy from Within: A Study of Political Delegitimization Discourse in Israeli Political Speech
Naama Rivlin-Angert (Hebrew University of Jerusalem), Guy Mor-Lan (Hebrew University of Jerusalem)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied political legitimacy weakening discourse (PDD), constructed and manually annotated 10,410 Hebrew sentences (sources include speeches from the Israeli parliament, Facebook official account posts, and news media), and proposed a two-stage detection pipeline (first identifying PDD, then predicting its intensity, aggressiveness, target type, etc.), followed by using the model for cross-platform and cross-temporal quantitative analysis.
The Good, the Bad and the Constructive: Automatically Measuring Peer Review’s Utility for Authors
Abdelrahman Sadallah (Mohamed Bin Zayed University of Artificial Intelligence), Ted Briscoe (Mohamed Bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a framework for automatically evaluating the usefulness of comments in the 'weaknesses' section of peer review, implementing four dimensions of scoring (operability, text-based localization and detail, verifiability, and overall usefulness), and building the RevUtil dataset based on this.
The Good, the Bad, and the Debatable: A Survey on the Impacts of Data for In-Context Learning
Stephanie Schoch (University of Virginia), Yangfeng Ji (University of Virginia)
Data-Centric LearningTransformerLarge Language ModelTextReview/Survey Paper
🎯 What it does: This paper classifies and analyzes the impact of demonstration data in in-context learning (ICL) for large language models (LLMs) through a systematic review approach, proposing a data-centric evaluation perspective;
The Hidden Strength of Disagreement: Unraveling the Consensus-Diversity Tradeoff in Adaptive Multi-Agent Systems
Zengqing Wu (Kyoto University), Takayuki Ito (Kyoto University)
TransformerLarge Language ModelPrompt EngineeringSequential
🎯 What it does: Propose and verify the use of implicit consensus (agents maintain independent decisions after discussion) instead of traditional explicit consensus (enforced unified decisions) in large language model-driven multi-agent systems, while introducing controllable diversity through role diversification.
The Illusion of Progress: Re-evaluating Hallucination Detection in LLMs
Denis Janiak (Wroclaw University of Science and Technology), Tomasz Jan Kajdanowicz
Large Language ModelTextBenchmark
🎯 What it does: This paper re-evaluates LLM hallucination detection methods through large-scale human evaluations and multiple assessment metrics, revealing that traditional ROUGE evaluation methods severely overestimate their performance, and demonstrating that LLM-as-Judge aligns more closely with human judgment. It also discovers that answer length itself is a powerful indicator of hallucination.
The Impact of Language Mixing on Bilingual LLM Reasoning
Yihao Li (University of Pennsylvania), Lyle Ungar (University of Pennsylvania)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper investigates the impact of English-Chinese code mixing on bilingual LLM reasoning, analyzes the causal effects of the training phase, code mixing modes, and proposes a lightweight probe-guided decoding method.
The Impact of Negated Text on Hallucination with Large Language Models
Jaehyung Seo (Korea University), Heuiseok Lim (Korea University)
Explainability and InterpretabilityTransformerPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Investigate the impact of negated text on hallucination detection in large language models, construct the NegHalu dataset, and conduct experiments on multiple models.
The LLM Already Knows: Estimating LLM-Perceived Question Difficulty via Hidden Representations
Yubo Zhu (State Key Laboratory for Novel Software Technology, Nanjing University), Jing Shao (State Key Laboratory for Novel Software Technology, Nanjing University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningTextMultimodality
🎯 What it does: Leverages the hidden representations of large language models (LLMs) by modeling the generation process as a Markov chain and defining a value function to estimate the model's perceived difficulty for input questions; subsequently, adaptively selects repeated sampling strategies such as Self-Consistency, Best-of-N, or Self-Refine based on the difficulty estimation to improve reasoning efficiency.
The Medium Is Not the Message: Deconfounding Document Embeddings via Linear Concept Erasure
Yu Fan (ETH Zurich), Alexander Hoyle (ETH Zurich)
RetrievalRepresentation LearningTransformerTextBenchmark
🎯 What it does: This paper proposes to improve the performance of similarity, clustering, and retrieval tasks by removing observable confounding factors (e.g., text source or language) from document embeddings through linear concept elimination (LEACE).
The Missing Parts: Augmenting Fact Verification with Half Truth Detection
Yixuan Tang (National University of Singapore), Anthony Kum Hoe Tung
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a semi-true detection task and construct the POLITIFACT-HIDDEN dataset, aligning evidence and annotating intents for political statements.
The Practical Impacts of Theoretical Constructs on Empathy Modeling
Allison Lahnala (McMaster University), Lucie Flek (University of Bonn)
Explainability and InterpretabilityRepresentation LearningTransformerSupervised Fine-TuningText
🎯 What it does: This paper conducts mediator task transfer experiments on 18 NLP empathy tasks to evaluate the impact of theoretical constructs on transfer effectiveness.
The Psychology of Falsehood: A Human-Centric Survey of Misinformation Detection
Arghodeep Nandi (Indian Institute of Technology Delhi), Tanmoy Chakraborty (University of Minnesota Twin Cities)
Reinforcement Learning from Human FeedbackGraph Neural NetworkLarge Language ModelTextMultimodalityBiomedical DataReview/Survey PaperChain-of-Thought
🎯 What it does: This paper reviews current research on misinformation detection, shifting from traditional fact-based detection methods to a human-centered cognitive psychology perspective, emphasizing the impact of emotion, narrative, and social dynamics on the spread of misinformation.
The Pursuit of Empathy: Evaluating Small Language Models for PTSD Dialogue Support
Suhas Bn, Saeed Abdullah (Penn State University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study uses small-scale language models with 0.5-5B parameters to generate empathetic dialogues for PTSD patients and created and utilized a clinically reviewed synthetic dialogue dataset named TIDE.
The Ranking Blind Spot: Decision Hijacking in LLM-based Text Ranking
Yaoyao Qian (Northeastern University), Huazheng Wang (Oregon State University)
RetrievalAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigated the 'Ranking Blind Spot' of LLMs in multi-document comparison tasks, and proposed two decision hijacking attacks (Decision Objective Hijacking and Decision Criteria Hijacking) to manipulate LLM text retrieval rankings.
The Role of Outgoing Connection Heterogeneity in Feedforward Layers of Large Language Models
Felix Stahlberg (Google Research), Shankar Kumar (Google Research)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper investigates the importance of output connection heterogeneity in feedforward layers of large language models (LLMs) and proposes improving model performance by reducing the entropy of neuron output weights;
The Sound of Syntax: Finetuning and Comprehensive Evaluation of Language Models for Speech Pathology
Fagun Patel (Stanford University), Nick Haber (Sound Speech and Hearing Clinic)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBiomedical DataBenchmarkChain-of-ThoughtAudio
🎯 What it does: This paper constructs the first benchmark containing 1,000 manually annotated data points for five core speech pathology tasks, systematically evaluates the performance of 15 multimodal language models in children's speech diagnosis, transcription, symptom and disease type classification, and enhances model performance through fine-tuning and integration.
The Staircase of Ethics: Probing LLM Value Priorities through Multi-Step Induction to Complex Moral Dilemmas
Ya Wu (Institute of Computing Technology Chinese Academy of Sciences), Juan Cao (Institute of Computing Technology Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Constructed the multi-step moral dilemmas (MMDs) dataset and dynamically evaluated the value preferences of large language models (LLMs) across five progressive stages.
The State of Multilingual LLM Safety Research: From Measuring The Language Gap To Mitigating It
Zheng Xin Yong, Julia Kreutzer (Cohere Labs)
Safty and PrivacyLarge Language ModelTextReview/Survey Paper
🎯 What it does: This paper systematically evaluates nearly 300 papers on LLM security from the ∗ACL conferences and workshops between 2020-2024, revealing that the field is highly English-centric, with scarce research on non-English languages, which are often incorporated into large-scale multilingual assessments but lack depth.
The Stepwise Deception: Simulating the Evolution from True News to Fake News with LLM Agents
Yuhan Liu (Renmin University of China), Rui Yan (Renmin University of China)
TransformerLarge Language ModelAgentic AITextSequential
🎯 What it does: The FUSE framework was constructed using LLM-driven agents to simulate the gradual evolution of true news into fake news on social networks. The framework includes four role agents (spreaders, commentators, fact-checkers, bystanders), a hierarchical memory mechanism, different network topologies, and official interventions based on bias thresholds.
The Strawberry Problem: Emergence of Character-level Understanding in Tokenized Language Models
Adrian Cosma (National University of Science and Technology POLITEHNICA), Mihai Dascalu (National University of Science and Technology POLITEHNICA)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: This study investigates the limitations of tokenization on character-level tasks in large language models, proposes 19 synthetic tasks to measure the emergence of character understanding, and subsequently designs a lightweight character-aware module to enhance character-level reasoning capabilities.
The Transfer Neurons Hypothesis: An Underlying Mechanism for Language Latent Space Transitions in Multilingual LLMs
Hinata Tezuka (Japan Advanced Institute of Science and Technology), Naoya Inoue (Japan Advanced Institute of Science and Technology)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper systematically investigates the spatial transfer of internal representations in multilingual large language models through hierarchical visualization and intervention experiments, and proposes the 'transmission neuron hypothesis,' which posits that specific MLP neurons are responsible for transferring representations between language-specific spaces and shared semantic spaces.
The Validation Gap: A Mechanistic Analysis of How Language Models Compute Arithmetic but Fail to Validate It
Leonardo Bertolazzi (University of Trento), Raffaella Bernardi (Free University of Bozen-Bolzano)
Explainability and InterpretabilityTransformerText
🎯 What it does: This study systematically analyzes the mechanisms by which small LLMs detect errors in simple arithmetic problems, revealing how models identify and correct arithmetic mistakes.
Theorem-Validated Reverse Chain-of-Thought Problem Generation for Geometric Reasoning
Deng Linger (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Data SynthesisLarge Language ModelPrompt EngineeringImageTextChain-of-Thought
🎯 What it does: Propose the TR-CoT two-stage framework to generate images, descriptions, and attributes based on geometric theorems, and automatically generate and verify logically consistent question-answer data through reverse chain reasoning.
Think and Recall: Layer-Level Prompting for Lifelong Model Editing
Jinke Wang (University of Science and Technology of China), Zhi Zheng (University of Science and Technology of China)
TransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the Layer-Level Prompting (LLP) method, achieving lifelong model editing by retrieving prompts in the early layers of LLMs and injecting them in the later layers.
Think Globally, Group Locally: Evaluating LLMs Using Multi-Lingual Word Grouping Games
César Guerra-Solano (University of Pittsburgh), Xiang Lorraine Li (University of Pittsburgh)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper introduces GLOBALGROUP, a multilingual abstract reasoning game benchmark, to evaluate the abstract reasoning and cross-lingual transfer capabilities of LLMs across different linguistic environments.
Think in Safety: Unveiling and Mitigating Safety Alignment Collapse in Multimodal Large Reasoning Model
Xinyue Lou (Beijing Jiaotong University), Kaiyu Huang (Beijing Jiaotong University)
Safty and PrivacyTransformerSupervised Fine-TuningMultimodalityChain-of-Thought
🎯 What it does: Systematically evaluated the safety performance of 13 multi-modal large reasoning models (MLRM), revealing different manifestations between safety performance degradation and safety awareness assessment.
Think Wider, Detect Sharper: Reinforced Reference Coverage for Document-Level Self-Contradiction Detection
Yuhao Chen (University of Science and Technology of China), Tong Xu (University of Science and Technology of China)
ClassificationKnowledge DistillationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a two-stage training framework, first obtaining high-quality Chain-of-Thought training data through teacher model distillation for supervised fine-tuning (SFT), and then further enhancing the model's reasoning completeness and consistency in the document-level self-contradiction detection (DSCD) task using GRPO reinforcement learning with custom accuracy, coverage, and format rewards.
Think, Verbalize, then Speak: Bridging Complex Thoughts and Comprehensible Speech
Tony Woo (Seoul National University), Gunhee Kim (Seoul National University)
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposed the THINK-VERBALIZE-SPEAK framework, decoupling reasoning (THINK) from verbalization (VERBALIZE), and introduced the REVERT model to achieve incremental verbalization, enhancing the naturalness of speech output.
ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models
Chung-En Sun (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: The study finds that large language models suffer performance degradation in chain-of-thought reasoning due to excessively short reasoning chains, and proposes ThinkEdit, a method to alleviate this issue by modifying the output projection weights of a small number of attention heads.
Thinking Out Loud: Do Reasoning Models Know When They’re Right?
Qingcheng Zeng (Northwestern University), Rob Voigt (Westlake University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Through a comparative analysis of three training methods—instruction tuning, reasoning SFT, and reasoning RL—the paper systematically evaluates the calibration performance (accuracy and verbalized confidence) of large reasoning models on mathematical, scientific reasoning, factual question-answering, and general reasoning tasks.
ThinkSLM: Towards Reasoning in Small Language Models
Gaurav Srivastava (Virginia Tech), Xuan Wang (Virginia Tech)
Computational EfficiencyKnowledge DistillationTransformerPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the THINKSLM benchmark, systematically evaluated the performance of 72 small language models across 17 reasoning tasks, and analyzed the impact of factors such as model scale, training, compression, and prompting on reasoning.
ThinkTuning: Instilling Cognitive Reflections without Distillation
Aswin Rrv (Arizona State University), Ben Zhou (Arizona State University)
Reinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringText
🎯 What it does: Designed and implemented an interactive training framework called THINKTUNING, which enhances the student model's thinking and self-reflection capabilities in RL (GRPO) by leveraging feedback from a teacher model.
Thought calibration: Efficient and confident test-time scaling
Menghua Wu (Massachusetts Institute of Technology), Tommi Jaakkola (Massachusetts Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a 'thinking calibration' method that dynamically decides when to stop the reasoning process of large language models, thereby shortening the thinking length and reducing computational costs during testing.
ThoughtProbe: Classifier-Guided LLM Thought Space Exploration via Probing Representations
Zijian Wang (University of Sydney), Chang Xu (University of Sydney)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes ThoughtProbe, a reasoning framework that utilizes LLM hidden layer representations through a classifier for discrimination, thereby guiding tree-structured idea exploration and aggregating answers;
Thread: A Logic-Based Data Organization Paradigm for How-To Question Answering with Retrieval Augmented Generation
Kaikai An (Peking University), Baobao Chang (Microsoft)
RetrievalLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose THREAD: a data organization paradigm based on logical units (LU), combined with retrieval-augmented generation (RAG) systems, for efficiently answering how-to questions.
Threading the Needle: Reweaving Chain-of-Thought Reasoning to Explain Human Label Variation
Beiduo Chen (LMU Munich), Barbara Plank (LMU Munich)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes a method to automatically extract explanation-label (EL) pairs by leveraging chain-of-thought (CoT) generated by large language models, and enhances explanation quality through discourse segmentation to better capture human annotation differences;
Through the Valley: Path to Effective Long CoT Training for Small Language Models
Renjie Luo (Singapore University of Technology and Design), Wei Lu (Nanyang Technological University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper systematically studies the training effects of using long-chain reasoning (CoT) data on small language models (SLM), identifies and defines the 'Long CoT Degradation' phenomenon, where after a short period of receiving a small amount of long CoT supervision, the model's performance significantly decreases;
TIDES: Technical Information Discovery and Extraction System
Jihee Kim (Yonsei University), Kyungwoo Song (Yonsei University)
RetrievalComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Developed a four-stage technical question-answering system called TIDES based on TF-IDF and prompt-based LLM.
Tiny Budgets, Big Gains: Parameter Placement Strategy in Parameter Super-Efficient Fine-Tuning
Jinman Zhao (University of Toronto), Gerald Penn (University of Toronto)
Computational EfficiencyRepresentation LearningTransformerImageText
🎯 What it does: Proposes the FoRA-UA method, which utilizes sparse Fourier frequency domain low-rank adaptation combined with intermediate representations and split projections, achieving parameter-efficient fine-tuning with only 1-5% of LoRA parameters.
TinySQL: A Progressive Text-to-SQL Dataset for Mechanistic Interpretability Research
Abir Harrasse (Martian), Amir Abdullah (Thoughtworks)
Data SynthesisExplainability and InterpretabilityTransformerAuto EncoderTextTabularBenchmark
🎯 What it does: Propose the TinySQL dataset and multi-scale models to explore the internal mechanisms of text-to-SQL generation
TLUE: A Tibetan Language Understanding Evaluation Benchmark
Fan Gao (University of Electronic Science and Technology of China), Hao Wang (University of Electronic Science and Technology of China)
Large Language ModelTextBenchmark
🎯 What it does: This paper constructs and releases the TLUE benchmark, covering multi-domain knowledge tests and safety evaluations to assess the understanding ability of large language models in Tibetan.
To Mask or to Mirror: Human-AI Alignment in Collective Reasoning
Crystal Qian (Google DeepMind), Lucas Dixon (Google DeepMind)
TransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: Designed and conducted an online experiment with 748 participants using the 'Lost at Sea' collective decision-making task to investigate the impact of visible identity information (such as names, gender, etc.) on leader selection; subsequently, matched simulations were performed using four publicly available LLMs (Gemini 2.5, GPT-4.1, Claude Haiku 3.5, and Gemma 3) to evaluate alignment and optimality between humans and AI in collective reasoning.
To See a World in a Spark of Neuron: Disentangling Multi-Task Interference for Training-Free Model Merging
Zitao Fang (Xiamen University Malaysia), Sim Kuan Goh (Xiamen University Malaysia)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningImageText
🎯 What it does: Propose a model merging method called NeuroMerging based on neuronal mechanisms, which decomposes the fine-tuned task vectors into parallel and orthogonal subspaces, performing untrained merging within these subspaces to alleviate multi-task interference;
ToDi: Token-wise Distillation via Fine-Grained Divergence Control
Seongryong Jung (Chung-Ang University), Hwanhee Lee (Chung-Ang University)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose Token-wise Distillation (ToDi), dynamically balancing forward KL and reverse KL at the token level in knowledge distillation to achieve fine-grained alignment of output distributions between large and small models.
Token-Aware Editing of Internal Activations for Large Language Model Alignment
Tianbo Wang (Beihang University), Xianglong Liu (Beihang University)
Explainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the Token-Aware Editing (TAE) framework, achieving inference-time alignment for large language models through fine-tuning of internal activations;
Token-level Proximal Policy Optimization for Query Generation
Yichen Ouyang (Zhejiang University), Feng Sun (Microsoft)
RetrievalOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the Token-level Proximal Policy Optimization (TPPO) framework to improve the performance of large language models (LLMs) in search query generation.
Tokenization and Representation Biases in Multilingual Models on Dialectal NLP Tasks
Vani Kanjirangat (SUPSI), Fabio Rinaldi (SUPSI)
ClassificationTransformerLarge Language ModelText
🎯 What it does: Investigate representation bias in multilingual large language models for dialect NLP tasks, and analyze the correlation between Tokenization Parity (TP) and Information Parity (IP) with downstream task performance.
TokenSelect: Efficient Long-Context Inference and Length Extrapolation for LLMs via Dynamic Token-Level KV Cache Selection
Wei Wu (University of Science and Technology of China), Hui Xiong (Hong Kong University of Science and Technology (Guangzhou))
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes TokenSelect, a training-free dynamic token-level KV cache selection method for efficient long-text inference and length extrapolation.
TokenSkip: Controllable Chain-of-Thought Compression in LLMs
Heming Xia (Hong Kong Polytechnic University), Wenjie Li (University of Science and Technology of China)
CompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose TokenSkip, a method for controllable chain-of-thought compression by skipping low-importance tokens