ACL 2025 Papers — Page 11
Annual Meeting of the Association for Computational Linguistics · 1699 papers
Meaning Variation and Data Quality in the Corpus of Founding Era American English
Dallas Card (University of Michigan)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper automatically detects semantic change and variation in all terms of the 1787 US Constitution using the BERT masked language model (MLM), and systematically evaluates data quality in the COFEA corpus, including OCR accuracy and metadata reliability.
Measuring Data Diversity for Instruction Tuning: A Systematic Analysis and A Reliable Metric
Yuming Yang (Fudan University), Xuanjing Huang (Fudan University)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Proposed and evaluated a measurement method for data diversity in instruction tuning, introducing a new metric called NovelSum and designing a data selection strategy called NovelSelect based on it.
Measuring Social Biases in Masked Language Models by Proxy of Prediction Quality
Rahul Zalkikar (Independent Researcher), Kanchan Chandra (New York University)
TransformerTextBenchmark
🎯 What it does: Evaluated social bias in masked language models (MLM), proposing an attention-weighted prediction quality proxy function to measure model preferences between sentences from disadvantaged and advantaged groups, and tested the model's bias tendency under the masked language modeling objective through iterative mask experiments.
Measuring the Effect of Transcription Noise on Downstream Language Understanding Tasks
Ori Shapira (OriginAI), Amir David Nissan Cohen
ClassificationTransformerLarge Language ModelTextAudio
🎯 What it does: Propose a configurable framework (ENDOW) for systematically evaluating the impact of transcription noise on downstream SLU tasks, and conduct experimental analysis on various noise levels, noise types, and cleaning techniques.
MEDDxAgent: A Unified Modular Agent Framework for Explainable Automatic Differential Diagnosis
Daniel Philip Rose (University of California), Carolin Lawrence (NEC Laboratories Europe)
Explainability and InterpretabilityTransformerLarge Language ModelTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a modular and interpretable automated differential diagnosis framework called MEDDxAgent, which achieves iterative diagnosis through three modules: interactive history acquisition, knowledge retrieval, and diagnostic strategies;
Medical Graph RAG: Evidence-based Medical Large Language Model via Graph Retrieval-Augmented Generation
Junde Wu (University of Oxford), Vicente Grau (University of Oxford)
Graph Neural NetworkLarge Language ModelTextGraphElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Propose a medical retrieval-augmented generation framework called MedGraphRAG, based on triple graph construction and U-Retrieval, which can generate trustworthy, evidence-supported medical answers within LLMs.
MegaPairs: Massive Data Synthesis for Universal Multimodal Retrieval
Junjie Zhou (Beijing University of Posts and Telecommunications), Defu Lian (University of Science and Technology of China)
Data SynthesisRetrievalLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes the MegaPairs data synthesis method, which extracts diverse image pairs from open-domain images using multiple similarity models, and generates corresponding open-source text instructions via multimodal large language models, thereby producing a large number of high-quality multimodal retrieval training samples.
MemeQA: Holistic Evaluation for Meme Understanding
Khoi P. N. Nguyen (University of Texas at Dallas), Vincent Ng (University of Texas at Dallas)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed the MemeQA dataset, containing 9,031 multiple-choice questions for comprehensive evaluation of meme understanding.
MEMERAG: A Multilingual End-to-End Meta-Evaluation Benchmark for Retrieval Augmented Generation
María Andrea Cruz Blandón (Tampere University), Marcello Federico (Tampere University)
RetrievalLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a meta-evaluation benchmark for multilingual retrieval-augmented generation (RAG) systems called MEMERAG, and manually annotated the answers generated by the model.
Memorization Inheritance in Sequence-Level Knowledge Distillation for Neural Machine Translation
Verna Dankers (University of Edinburgh), Vikas Raunak (Microsoft)
Knowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: Investigated how sequence-level knowledge distillation (SeqKD) enables student models to inherit instance-level memory from teacher models, evaluated changes in memory and hallucination through experiments, and subsequently proposed Adaptive-SeqKD to reduce memory and hallucination via teacher fine-tuning.
Memorizing is Not Enough: Deep Knowledge Injection Through Reasoning
Ruoxi Xu (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences), Le Sun (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a four-layer knowledge injection framework (Recall, Extraction, Reasoning, Association) and constructed the DeepKnowledge benchmark to fine-grainedly evaluate the effectiveness of knowledge injection in LLMs;
MEraser: An Effective Fingerprint Erasure Approach for Large Language Models
Jingxuan Zhang (Indiana University), Meng Han (Zhejiang University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the MEraser method, which uses a two-stage fine-tuning approach (first using mismatched data to eliminate trigger associations, then using clean data to restore model performance) to remove backdoor fingerprints in LLMs.
Merge Hijacking: Backdoor Attacks to Model Merging of Large Language Models
Zenghui Yuan (Huazhong University of Science and Technology), Lichao Sun (Lehigh University)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: Proposed a backdoor attack method called Merge Hijacking targeting LLM model fusion
MergePrint: Merge-Resistant Fingerprints for Robust Black-box Ownership Verification of Large Language Models
Shojiro Yamabe (Institute of Science Tokyo), Koki Wataoka (SB Intuitions)
OptimizationSafty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: The paper proposes MERGEPRINT, a black-box fingerprint embedding method targeting merge attacks on large language models (LLMs);
Meta-Learning Neural Mechanisms rather than Bayesian Priors
Michael Eric Goodale (PSL University), Yair Lakretz (PSL University)
Meta LearningRecurrent Neural NetworkSequential
🎯 What it does: Investigated the role of meta-learning in formal language learning, and compared two hypotheses: 'simplicity prior' and 'mechanism complexity'.
Meta-rater: A Multi-dimensional Data Selection Method for Pre-training Language Models
Xinlin Zhuang (Shanghai Artificial Intelligence Laboratory), Conghui He (Shanghai Artificial Intelligence Laboratory)
Representation LearningData-Centric LearningTransformerText
🎯 What it does: Propose Meta-rater, a multi-dimensional data selection framework that screens high-quality pre-training data through the weighted fusion of PRRC four dimensions (professionalism, readability, reasoning, neatness) and existing quality metrics.
Meta-Reflection: A Feedback-Free Reflection Learning Framework
Yaoke Wang (Zhejiang University), Yueting Zhuang (Zhejiang University)
Meta LearningAI Code AssistantTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposed the Meta-Reflection framework, achieving an LLM improvement scheme that enables reflection without external feedback and through a single inference step.
Meta-Tool: Unleash Open-World Function Calling Capabilities of General-Purpose Large Language Models
Shengqian Qin (Shanghai Jiao Tong University), Xiaofan Zhang (Shanghai Jiao Tong University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Meta-Tool proposes a pluggable tool retrieval framework that enables LLMs to autonomously retrieve and invoke tools in an open world.
METAL: A Multi-Agent Framework for Chart Generation with Test-Time Scaling
Bingxuan Li (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
GenerationLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposes METAL, a multi-agent VLM framework for chart code generation, iteratively improving chart reproducibility quality through four dedicated agents: generation, visual critique, code critique, and revision.
MEXMA: Token-level objectives improve sentence representations
João Maria Janeiro (Meta AI), Loic Barrault
Representation LearningTransformerContrastive LearningText
🎯 What it does: Propose the MEXMA model, which trains cross-lingual sentence encoders by simultaneously using sentence-level and word-level objectives;
Micro-Act: Mitigate Knowledge Conflict in Question Answering via Actionable Self-Reasoning
Nan Huo (University of Hong Kong), Reynold Cheng (University of Hong Kong)
Large Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: In retrieval-augmented generation (RAG) QA, the MICRO-ACT framework is proposed, achieving adaptive fine-grained comparison through a hierarchical action space to alleviate conflicts between retrieved information and model internal knowledge.
Middle-Layer Representation Alignment for Cross-Lingual Transfer in Fine-Tuned LLMs
Danni Liu (Karlsruhe Institute of Technology), Jan Niehues (Karlsruhe Institute of Technology)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Analyze the internal representations of large language models (LLMs), finding that intermediate layers have the greatest potential for cross-lingual alignment. Subsequently, introduce an alignment objective during task-specific fine-tuning, using alignment loss to align sentence representations across languages in the intermediate layers. Validate the effectiveness on multiple tasks (slot filling, machine translation, structured text generation).
Mimicking the Familiar: Dynamic Command Generation for Information Theft Attacks in LLM Tool-Learning System
Ziyou Jiang (State Key Laboratory of Intelligent Game), Qing Wang (State Key Laboratory of Intelligent Game)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose AUTOCMD, dynamically generating commands to implement information stealing attacks in LLM tool learning systems.
Mind the Gap: Static and Interactive Evaluations of Large Audio Models
Minzhi Li (Georgia Institute of Technology), Diyi Yang (Institute for Infocomm Research)
RecognitionTransformerLarge Language ModelBenchmarkAudio
🎯 What it does: Collected 7,500 voice interaction data through interactive evaluation, compared six large audio models (including closed-source and open-source LAM as well as ASR+LLM combinations), and analyzed task types and model advantages via user preference analysis.
Mind the Gesture: Evaluating AI Sensitivity to Culturally Offensive Non-Verbal Gestures
Akhila Yerukola (Carnegie Mellon University), Maarten Sap (Carnegie Mellon University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: Constructed the MC-SIGNS dataset and evaluated the detection and interpretation capabilities of cross-cultural gestures using text-to-image models, language models, and vision-language models.
MIND: A Multi-agent Framework for Zero-shot Harmful Meme Detection
Ziyan Liu (Beijing University of Posts and Telecommunications), Kaiwei Deng (Beijing University of Posts and Telecommunications)
ClassificationRetrievalAnomaly DetectionTransformerLarge Language ModelAgentic AIVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes MIND, a multi-agent framework for zero-shot harmful meme detection that completely does not rely on labeled data.
MindRef: Mimicking Human Memory for Hierarchical Reference Retrieval with Fine-Grained Location Awareness
Ye Wang, Zhiming Ding (Chinese Academy of Sciences)
RetrievalTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the MindRef framework, which uses a large language model to achieve variable-position retrieval without pre-chunking through two-stage prompting (first recalling document titles, then locating fine-grained paragraphs), and further improves speed via short prefix localization.
MiniLongBench: The Low-cost Long Context Understanding Benchmark for Large Language Models
Zhongzhan Huang (Sun Yat Sen University), Liang Lin (Sun Yat Sen University)
Large Language ModelTextBenchmark
🎯 What it does: Compress LongBench to construct MiniLongBench, retaining only 237 test samples to achieve efficient long-text evaluation.
Minimal Pair-Based Evaluation of Code-Switching
Igor Sterner (University of Cambridge), Simone Teufel (University of Cambridge)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: By constructing an evaluation benchmark based on minimal pairs, systematically assess the acceptance of large language models (LLMs) in code-switching (CS) texts, verifying whether LLMs handle CS in a manner consistent with bilingual speakers.
Mining Complex Patterns of Argumentative Reasoning in Natural Language Dialogue
Ramon Ruiz-Dolz (University of Dundee), John Lawrence (University of Passau)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Constructed the QT-SCHEMES natural dialogue argumentation scheme corpus, and trained and evaluated Transformer and LLM models on this corpus for identifying argumentation schemes in natural language dialogues.
Mining the uncertainty patterns of humans and models in the annotation of moral foundations and human values
Neele Falk (University of Stuttgart), Gabriella Lapesa (GESIS - Leibniz Institute for the Social Sciences)
TransformerLarge Language ModelText
🎯 What it does: Investigated the relationship between human labeler variance (HLV) and model uncertainty, and explained their common patterns through linguistic features (complexity, polarity, vocabulary choice, pragmatic phenomena).
MIR: Methodology Inspiration Retrieval for Scientific Research Problems
Aniketh Garikaparthi (TCS Research), Arman Cohan (Yale University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextGraph
🎯 What it does: Proposes the 'Methodology Inspiration Retrieval (MIR)' task aimed at retrieving papers from literature that can provide methodological inspiration for research questions, and constructs a specialized dataset and retrieval method.
MIRAGE: Exploring How Large Language Models Perform in Complex Social Interactive Environments
Yin Cai (Institute of Big Data, Fudan University), Ping Chen (Institute of Big Data, Fudan University)
TransformerLarge Language ModelAgentic AITextGraphBenchmark
🎯 What it does: Proposed the MIRAGE framework, evaluating large language models' role-playing and reasoning abilities in complex social interaction environments through eight murder mystery game scripts.
Mis-prompt: Benchmarking Large Language Models for Proactive Error Handling
Jiayi Zeng (East China Normal University), Aimin Zhou (East China Normal University)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the 'Mis-prompt' benchmark, which includes four proactive error-handling subtasks (detection, localization, correction, and guidance) and an error classification system, and conducted systematic evaluations of multiple large language models on this benchmark.
MISP-Meeting: A Real-World Dataset with Multimodal Cues for Long-form Meeting Transcription and Summarization
HangChen HangChen, Jun Du (NERC-SLIP University of Science and Technology of China)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Introduces the MISP-Meeting dataset and benchmarks automatic meeting transcription and summarization on this dataset, exploring the performance improvements brought by multimodal cues.
Mitigating Confounding in Speech-Based Dementia Detection through Weight Masking
Zhecheng Sheng (University of Minnesota), Serguei V. S. Pakhomov
ClassificationTransformerSupervised Fine-TuningTextAlzheimer's Disease
🎯 What it does: Researchers propose two bias mitigation methods based on weight masking (Extended Confounding Filter and Dual Filter) to eliminate gender-related spurious correlations in speech text, thereby improving fairness in detecting cognitive decline (Alzheimer's disease).
Mitigating Lost-in-Retrieval Problems in Retrieval Augmented Multi-Hop Question Answering
Rongzhi Zhu (Nanjing University), Wei Hu (Nanjing University)
RetrievalLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a progressive retrieval framework called ChainRAG, addressing retrieval errors in multi-hop QA by using sub-question rewriting and sentence graphs to complete missing entities, thereby improving retrieval quality.
Mitigating Non-Representative Prototypes and Representation Bias in Few-Shot Continual Relation Extraction
Thanh Duc Pham (FPT Software AI Center), Thien Huu Nguyen (University of Oregon)
ClassificationRepresentation LearningTransformerContrastive LearningText
🎯 What it does: Propose the Minion framework, addressing the issues of non-representative prototypes and representation bias in few-shot continual relation extraction (FCRE), by constructing dynamically updatable class prototypes and introducing contrastive learning assisted by label descriptions.
Mitigating Posterior Salience Attenuation in Long-Context LLMs with Positional Contrastive Decoding
Zikai Xiao (Zhejiang University), Zuozhu Liu (Zhejiang University)
TransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper discovered the phenomenon of Posterior Salience Attenuation (PSA) when large language models (LLMs) process long texts, and proposed an untrained decoding strategy—Positional Contrastive Decoding (PCD)—which enhances long-text reasoning by comparing standard RoPE attention with over-rotated local attention.
Mitigating Selection Bias with Node Pruning and Auxiliary Options
Hyeong Kyu Choi (University of Wisconsin-Madison), Chandan K. Reddy (Amazon)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: This paper analyzes the internal representations of LLMs and proposes two methods, Bias Node Pruning and Auxiliary Option Injection, to reduce selection bias in multiple-choice answers.
Mitigating Shortcut Learning with InterpoLated Learning
Michalis Korakakis (University of Cambridge), Adrian Weller (University of Cambridge)
Domain AdaptationRepresentation LearningTransformerText
🎯 What it does: Propose the InterpoLated Learning (InterpoLL) method, which interpolates representations of minority class samples with those of majority class samples to weaken the model's reliance on dominant shortcuts in the training set, thereby enhancing minority class and out-of-distribution (OOD) generalization capabilities.
Mitigating Visual Forgetting via Take-along Visual Conditioning for Multi-modal Long CoT Reasoning
Hai-Long Sun (Nanjing University), Han-Jia Ye (Nanjing University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Studies the problem of visual information decay (visual forgetting) in multimodal long-chain reasoning, and proposes the Take-along Visual Conditioning (TVC) method, which dynamically restarts visual input and compresses visual tokens during critical reasoning stages to maintain the model's sustained attention to images;
Mixture of insighTful Experts (MoTE): The Synergy of Reasoning Chains and Expert Mixtures in Self-Alignment
Zhili Liu (Hong Kong University of Science and Technology), James Kwok (Huawei Noah's Ark Lab)
Safty and PrivacyMixture of ExpertsTextChain-of-Thought
🎯 What it does: Propose the Mixture of insighTful Experts (MoTE) framework, combining multi-step reasoning chains with expert mixing techniques for LLM self-alignment;
Mixture of Ordered Scoring Experts for Cross-prompt Essay Trait Scoring
Po-Kai Chen (National Central University), Yi-Ting Huang (National Taiwan University of Science and Technology)
ClassificationTransformerMixture of ExpertsText
🎯 What it does: Propose the MOOSE framework to improve cross-prompt article feature scoring by simulating the human expert scoring process
Mixture of Small and Large Models for Chinese Spelling Check
Ziheng Qiao (Soochow University), Zhenghua Li (Soochow University)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a Chinese spelling correction method that dynamically hybridizes a small model (BERT) with a large language model (LLM) during the beam search phase, leveraging the high precision of the small model and the fluency of the LLM for complementary enhancement.
Mixtures of In-Context Learners
Giwon Hong (University of Edinburgh), Pasquale Minervini (University of Edinburgh)
ClassificationComputational EfficiencyLarge Language ModelMixture of ExpertsText
🎯 What it does: This study proposes Mixtures of In-Context Learners (MOICL), which splits the demonstration set into several subsets, treating each as an expert for ICL, and then learns a trainable weighted function to gradient-optimizedly combine the next-token distributions from these experts;
MLAS-LoRA: Language-Aware Parameters Detection and LoRA-Based Knowledge Transfer for Multilingual Machine Translation
Tianyu Dong (Tianjin University), Deyi Xiong (Tianjin University)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the MLAS-LoRA framework, which identifies language-aware neurons and separates them into language-general and language-specific knowledge, leveraging multi-language LoRA modules to achieve efficient knowledge transfer from teacher models to student models, thereby improving multilingual machine translation performance.
MM-Verify: Enhancing Multimodal Reasoning with Chain-of-Thought Verification
Linzhuang Sun (University of Chinese Academy of Sciences), Wentao Zhang (Peking University)
Supervised Fine-TuningMultimodalityChain-of-Thought
🎯 What it does: Propose MM-Verifier and MM-Reasoner, leveraging long Chain-of-Thought (COT) verification and reasoning, with a two-step synthetic data method for model training.
MMBoundary: Advancing MLLM Knowledge Boundary Awareness through Reasoning Step Confidence Calibration
Zhitao He (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose the MMBoundary framework for multi-step reasoning chains in multi-modal large language models (MLLMs), enabling the model to output self-confidence expressions after each reasoning step, thereby enhancing the self-correction capability of the reasoning chain.
MMDEND: Dendrite-Inspired Multi-Branch Multi-Compartment Parallel Spiking Neuron for Sequence Modeling
Kexin Wang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Guoqi Li (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
Computational EfficiencySpiking Neural NetworkImageTextSequentialAudio
🎯 What it does: Proposed the multi-branch, multi-compartment parallel spiking neuron MMDEND based on dendritic structure, and addressed the long-tailed membrane potential distribution and binarization errors through the Scaling-Shifting Integer Firing (SSF) mechanism.
MMLU-CF: A Contamination-free Multi-task Language Understanding Benchmark
Qihao Zhao (Microsoft Research), Furu Wei (Microsoft Research)
Data-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed a contamination-free multi-task language understanding benchmark called MMLU-CF, re-evaluating the world knowledge comprehension capabilities of large language models.
MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark
Xiang Yue, Graham Neubig
Large Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Designed and released a more rigorous multi-disciplinary, multi-modal evaluation benchmark called MMMU-Pro, removing text-solvable problems, increasing the number of options, and introducing a view-only input setting to force models to genuinely integrate visual and textual information.
MMRC: A Large-Scale Benchmark for Understanding Multimodal Large Language Model in Real-World Conversation
Haochen Xue (Xi'an Jiaotong-Liverpool University), Yu Qiao (Shanghai Artificial Intelligence Laboratory)
Large Language ModelMultimodalityBenchmark
🎯 What it does: Built MMRC — an open-source evaluation benchmark containing 5,120 real multimodal dialogues and 28,720 manually annotated question-answer pairs;
MobiLoRA: Accelerating LoRA-based LLM Inference on Mobile Devices via Context-aware KV Cache Optimization
Borui Li (Southeast University), Shuai Wang (Southeast University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Implemented efficient inference of large language models (LLM) based on LoRA on mobile devices, primarily through optimizing KV cache.
MoC: Mixtures of Text Chunking Learners for Retrieval-Augmented Generation System
Jihao Zhao (Renmin University of China), Zhiyu Li (Institute for Advanced Algorithms Research)
RetrievalTransformerLarge Language ModelMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: Proposed two new metrics for evaluating text chunking quality—Boundary Clarity (BC) and Chunk Stickiness (CS)—and designed a multi-granularity hybrid chunking framework (MoC) based on these metrics to enhance the performance of retrieval-augmented generation (RAG) systems.
MockConf: A Student Interpretation Dataset: Analysis, Word- and Span-level Alignment and Baselines
Dávid Javorský (Charles University), François Yvon (Sorbonne Université)
Data-Centric LearningTransformerLarge Language ModelTextMultimodalityBenchmarkAudio
🎯 What it does: This paper constructs and releases the MockConf dataset, collecting students' simultaneous interpretation records in simulated conferences, and provides word-level and segment-level alignment annotations; meanwhile, it introduces a web tool called InterAlign for this alignment, and presents a baseline alignment system along with evaluation metrics;
Modality-Aware Neuron Pruning for Unlearning in Multimodal Large Language Models
Zheyuan Liu (University Of Notre Dame), Meng Jiang (University Of Notre Dame)
Safty and PrivacyTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Studied the unlearning problem in multimodal large language models, proposing a modality-aware neuron pruning framework called MANU, which selectively removes sensitive knowledge under multimodal inputs while maintaining model performance.
Model Extrapolation Expedites Alignment
Chujie Zheng (Tsinghua University), Nanyun Peng (University of California, Los Angeles)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes the EXPO (Model Extrapolation) method, which enhances the alignment performance of large language models (LLMs) without training overhead by amplifying parameter changes from SFT (Supervised Fine-Tuning) to alignment models.
Modeling Complex Semantics Relation with Contrastively Fine-Tuned Relational Encoders
Naïm Es-sebbani (CRIL, University of Artois and CNRS), Zied Bouraoui (CRIL, University of Artois and CNRS)
Representation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsContrastive LearningText
🎯 What it does: Propose and experiment with a series of contrastive learning-based relation encoders, learning relationship embeddings between entity pairs through diverse prompts and multi-model fusion.
Modeling the Evolution of English Noun Compounds with Feature-Rich Diachronic Compositionality Prediction
Filip Miletić (University of Stuttgart), Sabine Schulte im Walde (University of Stuttgart)
ClassificationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Analyzes the compositionality of English noun compounds over time, constructs various time-based feature vectors, and uses them to perform binary classification prediction on the composability of contemporary compounds.
Modeling Uncertainty in Composed Image Retrieval via Probabilistic Embeddings
Haomiao Tang (Tsinghua University), Shu-Tao Xia (Tsinghua University)
RetrievalRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Propose a compositional probabilistic embedding framework named COPE, which embeds queries and targets as combinations of Gaussian distributions to address uncertainty issues in compositional image retrieval.
Modular Sentence Encoders: Separating Language Specialization from Cross-Lingual Alignment
Yongxin Huang (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
Domain AdaptationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Construct a modular multilingual sentence encoder: first train language-specific tokenizers, word embeddings, and language adapters for each language, then train monolingual sentence adapters using LoRA based on these; subsequently train a cross-lingual alignment adapter (CLA) using cross-lingual paraphrase pairs and parallel aligned corpora to map non-English sentence embeddings into the English semantic space; during inference, only activate the corresponding language module.
MolRAG: Unlocking the Power of Large Language Models for Molecular Property Prediction
Ziting Xian (Sun Yat-sen University), Shangsong Liang (Sun Yat-sen University)
Explainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the MolRAG framework, utilizing retrieval-augmented generation (RAG) and chain-of-thought (CoT) to achieve molecular property prediction without additional training.
MoQAE: Mixed-Precision Quantization for Long-Context LLM Inference via Mixture of Quantization-Aware Experts
Wei Tao (Huazhong University of Science and Technology), Jianzong Wang (Ping An Technology Co., Ltd.)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: This paper addresses the problem of high memory consumption in KV cache during long-text inference in large language models by proposing the MoQAE mixed-precision quantization scheme.
More is not always better? Enhancing Many-Shot In-Context Learning with Differentiated and Reweighting Objectives
Xiaoqing Zhang (Renmin University of China), Rui Yan (Renmin University of China)
OptimizationMeta LearningTextBenchmark
🎯 What it does: Proposed a multi-example context learning optimization framework named DrICL, addressing performance degradation caused by improper NLL objectives and demonstration noise.
MorphMark: Flexible Adaptive Watermarking for Large Language Models
Zongqi Wang (Tsinghua University), Yujiu Yang (Tsinghua University)
GenerationTransformerLarge Language ModelText
🎯 What it does: Propose MorphMark, an adaptive watermarking method based on green-red lists for text generation in large language models, addressing the trade-off between watermark effectiveness and text quality.
Movie101v2: Improved Movie Narration Benchmark
Zihao Yue (Renmin University of China), Qin Jin (Renmin University of China)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Proposed the Movie101v2 movie narration benchmark dataset, a staged task design (L1 visual fact description, L2 narrative plot, L3 deployable AD text), and LLM-based L1-Score and L2-Score evaluation metrics, while benchmarking multiple large vision-language models.
mPLUG-DocOwl2: High-resolution Compressing for OCR-free Multi-page Document Understanding
Anwen Hu (Alibaba Group), Jingren Zhou (Alibaba Group)
CompressionRepresentation LearningTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: Designed and implemented a high-resolution document compression module (High-resolution DocCompressor) and a three-stage training framework, constructing a multimodal large language model DocOwl2 capable of processing multi-page documents without OCR.
MPO: Multilingual Safety Alignment via Reward Gap Optimization
Weixiang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
OptimizationSafty and PrivacyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a multilingual reward gap optimization (MPO) method that leverages the reward gap of the main language to guide safe alignment in low-resource languages.
MPVStance: Mitigating Hallucinations in Stance Detection with Multi-Perspective Verification
ZhaoDan Zhang, Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the MPVStance framework, which systematically mitigates hallucinations in large language models (LLMs) for stance detection by leveraging multi-perspective verification (MPV) and retrieval-augmented generation (RAG), while enhancing stance detection accuracy through a five-step process (baseline generation, validation planning, validation execution, cross-checking and revision, final stance output).
MT-RAIG: Novel Benchmark and Evaluation Framework for Retrieval-Augmented Insight Generation over Multiple Tables
Kwangwook Seo (Yonsei University), Dongha Lee (Yonsei University)
GenerationRetrievalTransformerPrompt EngineeringMixture of ExpertsTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the MT-RAIG Benchmark and MT-RAIG EVAL to evaluate systems for multi-table retrieval-augmented insight generation.
MTSA: Multi-turn Safety Alignment for LLMs through Multi-round Red-teaming
Weiyang Guo (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the MTSA (Multi-Turn Safety Alignment) framework, combining thinking-guided multi-turn attack learning with multi-turn reinforcement learning to form a red-blue adversarial iterative training process, enhancing the safety and attack robustness of LLMs in multi-turn dialogues.
Multi-Attribute Steering of Language Models via Targeted Intervention
Duy Nguyen (UNC Chapel Hill), Mohit Bansal (UNC Chapel Hill)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Designed and implemented a reasoning-time intervention framework called MAT-STEER for multi-attribute target derivation, utilizing a gating mechanism to perform sparse and orthogonal interventions at the token level in large language models, thereby achieving a balance among multiple attributes (e.g., sincerity, harmlessness, and unbiasedness).
Multi-document Summarization through Multi-document Event Relation Graph Reasoning in LLMs: a case study in Framing Bias Mitigation
Yuanyuan Lei (Texas A&M University), Ruihong Huang (Texas A&M University)
GenerationGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraph
🎯 What it does: By constructing a multi-document event relationship graph (including events, event relationships, and moral attributes), and converting it into hard or soft prompts to guide large language models in generating neutralized summaries, thereby reducing media framing bias.
Multi-Facet Blending for Faceted Query-by-Example Retrieval
Heejin Do (POSTECH), Gary Lee (POSTECH)
RetrievalTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Proposed a multi-facet fusion (FaBle) augmentation method that decomposes and recombines documents using LLM to generate facet-specific positive and negative samples, enabling unlabeled facet-oriented QBE training;
Multi-level Association Refinement Network for Dialogue Aspect-based Sentiment Quadruple Analysis
Zeliang Tong, Xingyu Yan (Shenzhen Yishi Huolala Technology Limited)
RecognitionData-Centric LearningGraph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose a Multi-layer Associative Refinement Network (MARN) for complete extraction of dialogue sentiment quadruples (target, aspect, opinion, emotion).
Multi-Level Explanations for Generative Language Models
Lucas Monteiro Paes (Harvard University), Soumya Ghosh (IBM Research)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes the MExGen framework, which generates importance scores for input text to explain model outputs in context-driven generative language models (e.g., summarization, QA), through multi-level chunking, linear-complexity variants of LIME/SHAP, and text-to-real scalarizers.
Multi-level Relevance Document Identifier Learning for Generative Retrieval
Fuwei Zhang (Beihang University), Zhao Zhang (Beihang University)
RetrievalTransformerAuto EncoderContrastive LearningText
🎯 What it does: Propose the MERGE method, which utilizes multi-level query-document relevance learning to generate high-quality DocID, thereby enhancing the performance of generative retrieval (GR).
Multi-Modality Expansion and Retention for LLMs through Parameter Merging and Decoupling
Junlin Li (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityAudio
🎯 What it does: Propose an untrained multi-modal extension and preservation method called MMER, which utilizes parameter merging and decoupling mechanisms to integrate multi-modal encoder parameters with language model parameters, expanding the multi-modal capabilities of LLMs while maintaining original performance without increasing inference parameters, thereby alleviating catastrophic forgetting.
Multi-perspective Alignment for Increasing Naturalness in Neural Machine Translation
Huiyuan Lai (University of Groningen), Antonio Toral (Universitat d'Alacant)
GenerationTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a multi-perspective alignment framework that enhances the fluency of neural machine translation (NMT) while preserving content through reinforcement learning, verifying the effectiveness of the method in the English-to-Dutch literary translation task.
Multi-task Adversarial Attacks against Black-box Model with Few-shot Queries
Wenqiang Wang (Sun Yat-sen University), Xiaochun Cao (Sun Yat-sen University)
ClassificationAdversarial AttackTransformerText
🎯 What it does: Proposed a few-query attack method CEMA for black-box multi-task text models, achieving cross-task attacks through deep label substitution models and ensemble attacks.
MultiAgentBench : Evaluating the Collaboration and Competition of LLM agents
Kunlun Zhu (University of Illinois Urbana Champaign), Jiaxuan You (University of Illinois Urbana Champaign)
TransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Proposed the MultiAgentBench benchmark and the MARBLE framework to evaluate the performance of large language models in multi-agent collaborative and competitive scenarios
Multilingual Arbitration: Optimizing Data Pools to Accelerate Multilingual Progress
Ayomide Odumakinde (Cohere), Sara Hooker (Cohere Labs)
Computational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a multilingual arbitration method that generates better training data by dynamically routing samples among multiple teacher models.
Multilingual Encoder Knows more than You Realize: Shared Weights Pretraining for Extremely Low-Resource Languages
Zeli Su (Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE), Yushuang Dong (Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE)
GenerationTransformerAuto EncoderText
🎯 What it does: For extremely low-resource languages, a shared-weight encoder-decoder framework named XLM-SWCM is constructed to efficiently generate text.
Multilingual Gloss-free Sign Language Translation: Towards Building a Sign Language Foundation Model
Sihan Tan (Institute of Science Tokyo), Kazuhiro Nakadai (Institute of Science Tokyo)
RecognitionImage TranslationTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Propose a multilingual gloss-free sign language translation model, Sign2(LID+Text), which simultaneously predicts word-level sign language identification (LIDtok) and sign-to-text CTC alignment, achieving one-to-one, one-to-many, and many-to-many multilingual sign language translation.
Multilingual Text-to-Image Generation Magnifies Gender Stereotypes
Felix Friedrich (TU Darmstadt), Alexander Fraser (Munich Center for Machine Learning)
GenerationVision Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: Constructed a multilingual text-to-image generation gender bias evaluation benchmark MAGBIG, used it to assess gender bias in five multilingual T2I models, and tested the mitigation effect of neutral prompts.
Multimodal Coreference Resolution for Chinese Social Media Dialogues: Dataset and Benchmark Approach
Xingyu Li (Soochow University), Guohong Fu (Soochow University)
Object DetectionObject TrackingRetrievalConvolutional Neural NetworkTransformerSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Created the TikTalkCoref Chinese multimodal coreference dataset and proposed a baseline method and experimental framework.
Multimodal Pragmatic Jailbreak on Text-to-image Models
Tong Liu (LMU Munich), Jindong Gu (University of Oxford)
Adversarial AttackLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: Propose and evaluate a new multimodal practical jailbreak that tricks text-to-image models into generating images containing visual text, leading to unsafe content when combined with text.
Multimodal Transformers are Hierarchical Modal-wise Heterogeneous Graphs
Yijie Jin (Shanghai University), Cangzhi Zheng (Shanghai University)
RecognitionRepresentation LearningGraph Neural NetworkTransformerMultimodality
🎯 What it does: This paper proposes a new multi-modal Transformer framework called GsiT, which treats multi-modal fusion as a hierarchical modal heterogeneous graph (HMHG), achieving parameter compression through shared weights;
Multiple LLM Agents Debate for Equitable Cultural Alignment
Dayeon Ki (University of Maryland), Marine Carpuat (University of Maryland)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose a multi-LLM debate framework where two large language models engage in debates within cultural contexts to collaboratively reach a final decision.
MultiSocial: Multilingual Benchmark of Machine-Generated Text Detection of Social-Media Texts
Dominik Macko (Kempelen Institute of Intelligent Technologies), Ivan Srba (Kempelen Institute of Intelligent Technologies)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Created and released a large-scale benchmark dataset for detecting machine-generated text on social media, MultiSocial, which is multi-platform, multi-language (22 languages), and multi-generator (7 LLMs). This dataset was used to evaluate the performance of 17 state-of-the-art detection methods (statistical zero-shot, pre-trained, and fine-tuned categories) across languages, platforms, and generators.
MuSC: Improving Complex Instruction Following with Multi-granularity Self-Contrastive Training
Hui Huang (Harbin Institute of Technology), Tiejun Zhao (Harbin Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: Proposed a multi-grained self-contrast training framework named MuSC to enhance large language models' ability to follow complex instructions;
MUSTS: MUltilingual Semantic Textual Similarity Benchmark
Tharindu Ranasinghe (Lancaster University), Ruslan Mitkov (Lancaster University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a multilingual semantic text similarity benchmark called MUSTS, covering 13 languages (including low-resource languages), and evaluates over 25 unsupervised and supervised methods on this benchmark.
Mutual-Taught for Co-adapting Policy and Reward Models
Tianyuan Shi (Sun Yat-sen University), Ming Yan (Alibaba Group)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: Proposes the Mutual-Taught method, which employs a self-training Expectation-Maximization (EM) approach to enable mutual improvement between the policy model and the reward model.
My Words Imply Your Opinion: Reader Agent-Based Propagation Enhancement for Personalized Implicit Emotion Analysis
Jian Liao (Shanxi University), JianXing Zheng
RecognitionRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelAgentic AITextGraphTabular
🎯 What it does: Propose the RAPPIE model to address the personalized implicit emotion analysis (PIEA) task, generating reader agents via LLM to simulate reader feedback and enhancing emotion recognition by combining role-aware multi-view graph learning.
Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
Jingyang Yuan (Peking University), Wangding Zeng (DeepSeek-AI)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes NSA (Natively Trainable Sparse Attention), a sparse mechanism that achieves efficient attention for long contexts during both training and inference phases.
Navigating Rifts in Human-LLM Grounding: Study and Benchmark
Omar Shaikh (Stanford University), Eric Horvitz (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper systematically studies grounding behavior in human-large language model (LLM) dialogues, constructs a classification of grounding behavior, uses GPT for automatic annotation and trains a prediction model, and finally creates the RIFTS benchmark based on prediction results to evaluate LLM grounding capabilities, while proposing a simple intervention method based on the predictor.
Negative Matters: Multi-Granularity Hard-Negative Synthesis and Anchor-Token-Aware Pooling for Enhanced Text Embeddings
Tengyu Pan (Tsinghua University), Jianyong Wang (Tsinghua University)
RetrievalRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Proposes a multi-grained hard negative sample synthesis framework and Anchor Token Aware Pooling (ATA) method to enhance the representation quality of text embedding models.
Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset
Dan Su (NVIDIA), Bryan Catanzaro (NVIDIA)
Data SynthesisData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By improving the Common Crawl data extraction and filtering process, integrating model-based benchmark quality classifier ensembling, generating synthetic data through rephrasing of low-quality text, and reducing traditional heuristic filtering, we constructed a long-term pre-training dataset Nemotron-CC (containing 4.4T deduplicated tokens and 1.9T synthetic tokens) with a total size of 6.3T, and provided a high-quality subset Nemotron-CC-HQ with 1.1T.
Neural Incompatibility: The Unbridgeable Gap of Cross-Scale Parametric Knowledge Transfer in Large Language Models
Yuqiao Tan (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Jun Zhao (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Investigate parameter knowledge transfer (PKT) methods for cross-scale large language models (LLMs), proposing Pre-Align PKT and Locate-Then-Align (LaTen) techniques to achieve parameter space alignment and knowledge injection