EMNLP 2025 Papers — Page 11
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Minimal, Local, and Robust: Embedding-Only Edits for Implicit Bias in T2I Models
Feng He (University of Sheffield), Zhixue Zhao (University of Sheffield)
GenerationSupervised Fine-TuningVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes a method called Embedding-only Editing (EMBEDIT) that modifies only the word embeddings to correct implicit assumptions and biases in text-to-image (T2I) models, without affecting other knowledge in the model.
Mining the Past with Dual Criteria: Integrating Three types of Historical Information for Context-aware Event Forecasting
Rong Ma (Xinjiang Technical Institute of Physics & Chemistry), Xinyue Wang (Hohai University)
Graph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningGraphSequentialRetrieval-Augmented Generation
🎯 What it does: Proposes the ITHI method, systematically integrating sequential event information, periodic repetitive event information, and query-related historical information, and filters out high-quality relevant historical events through dual-standard constraints (event semantics + fact structure self-supervised filtering);
MIO: A Foundation Model on Multimodal Tokens
Zekun Moore Wang (Beihang University), Wenhao Huang (01.AI)
TransformerLarge Language ModelSupervised Fine-TuningImageVideoTextMultimodalityAudio
🎯 What it does: Developed and trained MIO, a fully autoregressive foundational model supporting four modalities (text, image, speech, and video), capable of performing multimodal interaction and interleaved sequence generation.
MIRROR: Multimodal Cognitive Reframing Therapy for Rolling with Resistance
Subin Kim (KT Corporation), Gary Lee (POSTECH)
Data SynthesisTransformerVision Language ModelMultimodality
🎯 What it does: Studied a cognitive restructuring therapy that integrates visual and textual modalities, utilizing the synthetic multimodal dataset MIRROR to train vision-language models for identifying and addressing client resistance.
Mitigating Biases in Language Models via Bias Unlearning
Dianqing Liu (University of Science and Technology of China), Zhendong Mao (State Key Laboratory of Communication Content Cognition People's Daily Online)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the BiasUnlearn framework, which employs a dual-path unlearning mechanism (forgetting stereotypes + retaining anti-stereotypes) to debias large language models; meanwhile, adversarial forgetting sets and dynamic dataset exchange are used to prevent bias polarization reversal.
Mitigating Catastrophic Forgetting in Large Language Models with Forgetting-aware Pruning
Wei Huang (Ant Group), Yinggui Wang (Ant Group)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a forget-aware pruning metric (FAPM) that alleviates catastrophic forgetting in LLM fine-tuning by simultaneously considering magnitude and relative change ratio on task vectors, without modifying the training process or model architecture.
Mitigating Hallucinations in Large Vision-Language Models via Entity-Centric Multimodal Preference Optimization
Jiulong Wu, Min Zhang (Soochow University)
OptimizationReinforcement Learning from Human FeedbackVision Language ModelDiffusion modelMultimodality
🎯 What it does: Proposes an entity-centric multimodal preference optimization (EMPO) method to reduce hallucinations in large vision-language models.
Mitigating Hallucinations in LM-Based TTS Models via Distribution Alignment Using GFlowNets
Chenlin Liu (Harbin Institute Of Technology), Jiqing Han (Harbin Institute Of Technology)
GenerationTransformerSupervised Fine-TuningReinforcement LearningFlow-based ModelAudio
🎯 What it does: Proposes the GOAT framework, which utilizes GFlowNets to post-train LM-based TTS models, reducing hallucinations in generated text.
Mitigating Hallucinations in Vision-Language Models through Image-Guided Head Suppression
Sreetama Sarkar (University of Southern California), Souvik Kundu (Intel Labs)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelImageMultimodality
🎯 What it does: Propose a dynamic attention head suppression (SPIN) during the inference phase of vision-language models to reduce hallucination generation.
Mitigating the Privacy Issues in Retrieval-Augmented Generation (RAG) via Pure Synthetic Data
Shenglai Zeng (Michigan State University), Jiliang Tang (University of Arizona)
Data SynthesisRetrievalSafty and PrivacyLarge Language ModelAgentic AITextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose a two-stage synthetic data generation framework called SAGE, which uses synthetic data instead of real retrieval data to reduce privacy leakage risks in RAG systems.
Mixing Inference-time Experts for Enhancing LLM Reasoning
Soumya Sanyal (University of Southern California), Xiang Ren (University of Southern California)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextChain-of-Thought
🎯 What it does: Proposes the MIXIE framework, which dynamically mixes multiple expert models during inference to enhance LLM reasoning quality, particularly the coherence and consistency of chain-of-thought reasoning.
MixLoRA-DSI: Dynamically Expandable Mixture-of-LoRA Experts for Rehearsal-Free Generative Retrieval over Dynamic Corpora
Tuan-Luc Huynh (Monash University), Thanh-Toan Do (Monash University)
RetrievalTransformerMixture of ExpertsText
🎯 What it does: Propose the MixLoRA-DSI framework, achieving a non-replayable, dynamically scalable generative retrieval index that appends only necessary LoRA experts when new documents are added.
Mixture of Languages: Improved Multilingual Encoders Through Language Grouping
João Maria Janeiro (FAIR at Meta), Loic Barrault
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a novel multilingual pre-trained encoder named Mixture of Languages (MoL), which groups languages by similarity and designs specialized sparse activation layers for each group, maintaining the same number of parameters during inference as traditional dense models;
Mixture of Length and Pruning Experts for Knowledge Graphs Reasoning
Enjun Du (Hong Kong University of Science and Technology (Guangzhou)), Yongqi Zhang (Hong Kong University of Science and Technology (Guangzhou))
Computational EfficiencyRepresentation LearningGraph Neural NetworkMixture of ExpertsGraphBenchmark
🎯 What it does: Proposes the MoKGR framework, achieving path length adaptation and path pruning personalization in knowledge graph reasoning through a hybrid expert mechanism.
Mixture of Weight-shared Heterogeneous Group Attention Experts for Dynamic Token-wise KV Optimization
Guanghui Song (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Xitong Gao (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Design a mixSGA, a weight-sharing based multi-expert group attention mechanism in autoregressive language models, dynamically allocating different KV cache sizes for each token to significantly reduce computational and memory overhead while maintaining full context.
Mixture-of-Clustered-Experts: Advancing Expert Specialization and Generalization in Instruction Tuning
Sugyeong Eo, Heuiseok Lim (Korea University)
AI Code AssistantTransformerSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Propose a two-stage sparse expert network called Mixture-of-Clustered-Experts (MoCE), which first assigns expert groups based on sequence embedding clustering and then performs token-level routing within groups for instruction tuning.
ML-Promise: A Multilingual Dataset for Corporate Promise Verification
Yohei Seki (University of Tsukuba), Chung-Chi Chen (National Taipei University)
ClassificationRetrievalTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Proposed the 'Promise Verification' task and constructed the first multilingual (Chinese, English, French, Japanese, Korean) ESG report dataset ML-Promise, covering three industries and three companies with a total of 3,010 instances; each instance was annotated with four categories (Promise Identification, Actionable Evidence, Clarity of the Promise-Evidence Pair, Timing for Verification).
MLWQ: Efficient Small Language Model Deployment via Multi-Level Weight Quantization
Chun Hu (Wuhan University), Qingan Li (Wuhan University)
Computational EfficiencyTransformerText
🎯 What it does: Propose a multi-layer weight quantization (MLWQ) method for efficiently deploying small language models on resource-constrained devices.
MMAG: Multimodal Learning for Mucus Anomaly Grading in Nasal Endoscopy via Semantic Attribute Prompting
Xinpan Yuan (Hunan University Of Technology), Xu Zhang (Hunan University Of Technology)
ClassificationAnomaly DetectionTransformerPrompt EngineeringVision Language ModelImageBiomedical Data
🎯 What it does: This paper proposes a multi-modal learning-based framework for grading nasal endoscopy mucus abnormalities, named MMAG, which first locates mucus features through attribute prompts and then maps them to severity levels for grading;
MMAPG: A Training-Free Framework for Multimodal Multi-hop Question Answering via Adaptive Planning Graphs
Yiheng Hu (University of New South Wales), Liming Zhu (CSIRO Data61)
Graph Neural NetworkLarge Language ModelVision Language ModelMultimodalityTabularRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose MMAPG—a training-agnostic, adaptive planning graph-based framework for multimodal multi-hop question answering.
MMDocIR: Benchmarking Multimodal Retrieval for Long Documents
Kuicai Dong (Huawei Technologies), Yong Liu (Huawei Technologies)
RetrievalTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the MMDOCIR benchmark, covering dual-task retrieval (page-level and layout-level) and providing high-quality evaluation and training sets.
MMLU-ProX: A Multilingual Benchmark for Advanced Large Language Model Evaluation
Weihao Xuan (University of Tokyo), Irene Li (University of Tokyo)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a multilingual reasoning benchmark MMLU-ProX covering 29 languages with 11,800 questions;
MobiZO: Enabling Efficient LLM Fine-Tuning at the Edge via Inference Engines
Lei Gao (University of Southern California), Murali Annavaram (University of Southern California)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose MobiZO, a zeroth-order optimization-based on-device LLM fine-tuning framework, which utilizes the MP-LoRA module to achieve multi-perturbation parallelism and combines outer-loop and inner-loop parallelization, ultimately realizing efficient, low-memory on-device fine-tuning.
ModalPrompt: Towards Efficient Multimodal Continual Instruction Tuning with Dual-Modality Guided Prompt
Fanhu Zeng (Chinese Academy Of Sciences), Cheng-Lin Liu (Chinese Academy Of Sciences)
Computational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose the ModalPrompt framework, achieving multi-modal continuous instruction tuning through prompt learning guided by dual-modal (visual + text) features, significantly reducing forgetting and controlling inference complexity.
Model Consistency as a Cheap yet Predictive Proxy for LLM Elo Scores
Ashwin Ramaswamy (Independent), Ermal Rrapaj (Lawrence Berkeley National Laboratory)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose a low-cost model intelligence estimation metric called Consistency Score by measuring the consistency (variance) of LLMs in contrast tasks.
Model Unlearning via Sparse Autoencoder Subspace Guided Projections
Xu Wang (University of Hong Kong), Difan Zou (University of Hong Kong)
Safty and PrivacyRepresentation LearningAuto EncoderText
🎯 What it does: Propose a subspace projection forgetting framework named SSPU based on SAE features for interpretable and robust knowledge deletion in the parameter space.
Model-based Large Language Model Customization as Service
Zhaomin Wu (National University Of Singapore), Qiang Yang (National University Of Singapore)
Data SynthesisSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelTabular
🎯 What it does: Propose a framework called Llamdex, allowing clients to upload pre-trained domain expert models instead of sensitive data, thereby enabling customized services for large language models (LLMs).
Model-Based Ranking of Source Languages for Zero-Shot Cross-Lingual Transfer
Abteen Ebrahimi (University of Colorado Boulder), Katharina von der Wense (University of Colorado Boulder)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes NN-RANK, an algorithm that ranks source languages by leveraging hidden representations from multilingual models for zero-shot cross-lingual transfer;
ModelCitizens: Representing Community Voices in Online Safety
Ashima Suvarna (University of California Los Angeles), Saadia Gabriel (University of California Los Angeles)
ClassificationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Constructed the MODELCITIZENS dataset, collecting 6,800 social media posts annotated for toxicity by members and non-members of the target groups, while enhancing with dialogue contexts generated by LLMs;
Modeling Bottom-up Information Quality during Language Processing
Cui Ding (University of Zürich), Ethan Wilcox (Georgetown University)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelTextMultimodality
🎯 What it does: Measuring the mutual information between visual input and word identity using information-theoretic methods, verifying its impact on reading time, and testing through experiments obscuring the upper/lower halves of words combined with bilingual human reading data in Chinese and English.
ModRWKV: Transformer Multimodality in Linear Time
Jiale Kang (Yuanshi Inc), Zhouran Ji (Yuanshi Inc)
Computational EfficiencyRecurrent Neural NetworkLarge Language ModelVision Language ModelImageTextMultimodalityTime SeriesAudio
🎯 What it does: Construct a linear RNN multimodal framework ModRWKV based on RWKV7, supporting pluggable modal encoders for cross-modal learning.
Molecular String Representation Preferences in Pretrained LLMs: A Comparative Study in Zero- & Few-Shot Molecular Property Prediction
George Arthur Baker (University of Colorado Boulder), Katharina von der Wense (University of Colorado Boulder)
Representation LearningDrug DiscoveryTransformerLarge Language ModelBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Evaluate the zero-shot and few-shot reasoning performance of four mainstream large language models in molecular property prediction tasks using five molecular string representations: SMILES, DeepSMILES, SELFIES, InChI, and IUPAC names.
MolErr2Fix: Benchmarking LLM Trustworthiness in Chemistry via Modular Error Detection, Localization, Explanation, and Correction
Yuyang Wu (Carnegie Mellon University), Olexandr Isayev (Carnegie Mellon University)
Explainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelTextGraphBiomedical DataBenchmark
🎯 What it does: Propose the MOLERR2FIX benchmark to evaluate the ability of large language models in detecting, locating, explaining, and correcting errors in chemical descriptions.
MoLoRAG: Bootstrapping Document Understanding via Multi-modal Logic-aware Retrieval
Xixi Wu (Chinese University of Hong Kong), Hong Cheng (Chinese University of Hong Kong)
RetrievalGraph Neural NetworkVision Language ModelMultimodalityGraphRetrieval-Augmented Generation
🎯 What it does: Propose MoLoRAG, a logic-aware multimodal multi-page document retrieval framework, which achieves efficient document page retrieval using graph traversal methods and feeds the retrieval results into an LVLM for question answering.
MoMoE: Mixture of Moderation Experts Framework for AI-Assisted Online Governance
Agam Goyal (University of Illinois Urbana-Champaign), Eshwar Chandrasekharan (University of Illinois Urbana-Champaign)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposes MoMoE—a cross-community, modular hybrid expert framework for AI-assisted online community content governance;
Mondrian: A Framework for Logical Abstract (Re)Structuring
Elizabeth Grace Orwig (Yonsei University), Yo-Sub Han (Yonsei University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Mondrian framework, which reorganizes academic abstracts according to the ABT (And-But-Therefore) structure, and introduces the EB-DTW metric to evaluate structural conformity;
MoR: Better Handling Diverse Queries with a Mixture of Sparse, Dense, and Human Retrievers
Jushaan Singh Kalra (Carnegie Mellon University), Tongshuang Wu (Carnegie Mellon University)
RetrievalMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the Mixture of Retrievers (MoR) framework, which dynamically assigns and fuses results from multiple sparse/dense retrievers or even human retrievers for each query in retrieval-augmented generation (RAG) tasks.
Morables: A Benchmark for Assessing Abstract Moral Reasoning in LLMs with Fables
Matteo Marcuzzo (Ca Foscari University of Venice), Mohammad Taher Pilehvar (Cardiff University)
Adversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the MORABLES benchmark by leveraging historical literary fables and their corresponding moral sentences, designing multiple-choice reasoning tasks with adversarial versions;
Moral Framing in Politics (MFiP): A new resource and models for moral framing
Ines Rehbein (University of Mannheim), Simone Paolo Ponzetto (University of Mannheim)
ClassificationRecognitionTransformerContrastive LearningText
🎯 What it does: Constructed and annotated a new 'Moral Framework' corpus (MFiP) from German parliamentary debate texts, with fine-grained labels for moral framework types, narrative roles, and moral foundations for each text segment; trained and evaluated models for framework identification, type classification, and moral foundation classification based on this corpus, further exploring the effectiveness of data augmentation and contrastive learning in this task.
Morpheme Induction for Emergent Language
Brendon Boldt (Carnegie Mellon University), David R. Mortensen (Carnegie Mellon University)
Representation LearningText
🎯 What it does: This paper proposes a greedy algorithm named CSAR (Count-Select-Ablate-Repeat) for inducing minimal form-meaning pairs (morphemes) from parallel corpora (sentences and corresponding meanings).
MOSAIC: Modeling Social AI for Content Dissemination and Regulation in Multi-Agent Simulations
Genglin Liu (University of California, Los Angeles), Saadia Gabriel (University of California, Los Angeles)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AITextTabular
🎯 What it does: Proposed and implemented MOSAIC—a multi-agent social network simulation framework based on large language models (LLMs)—to study content propagation, user engagement, and misinformation diffusion;
MoSEs: Uncertainty-Aware AI-Generated Text Detection via Mixture of Stylistics Experts with Conditional Thresholds
Junxi Wu (Nankai University), Shu-Tao Xia (Tsinghua University)
ClassificationAnomaly DetectionMixture of ExpertsText
🎯 What it does: Propose the MoSEs framework, leveraging a multi-style reference library, style router, and conditional threshold estimation to achieve style-aware uncertainty detection for AI-generated text.
MoVa: Towards Generalizable Classification of Human Morals and Values
Ziyu Chen (Australian National University), Lexing Xie (Australian National University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This study constructs the MoVa resource, aiming to perform generalizable classification of human morality and values in text, and provides a multi-framework, multi-dataset evaluation benchmark.
MovieCORE: COgnitive REasoning in Movies
Gueter Josmy Faure (National Taiwan University), Winston H. Hsu (National Taiwan University)
Data-Centric LearningTransformerLarge Language ModelAgentic AIVision Language ModelVideoTextBenchmark
🎯 What it does: This paper proposes the MovieCORE dataset, aiming to conduct System-2 level question-answering evaluation through deep narrative and emotional reasoning in movie videos, and develops a collaborative annotation process based on multiple large language models and an ACE post-processing module.
MPCG: Multi-Round Persona-Conditioned Generation for Modeling the Evolution of Misinformation with LLMs
Chong Jun Rong Brian (National University of Singapore), Anthony Kum Hoe Tung (National University of Singapore)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed and implemented a multi-round character-conditioned generation framework, MPCG, using large language models (LLMs) to simulate the gradual evolution and reconstruction of misinformation across different political stances.
MPRF: Interpretable Stance Detection through Multi-Path Reasoning Framework
ZhaoDan Zhang, Xueqi Cheng (Chinese Academy of Sciences)
ClassificationExplainability and InterpretabilityLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed a multi-path reasoning framework (MPRF) that generates, evaluates, optimizes, and fuses multiple reasoning paths to enhance the accuracy, robustness, and interpretability of stance detection.
MR. Judge: Multimodal Reasoner as a Judge
Renjie Pi (HKUST), Meng Cao (Apple)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the MR.Judge framework, which uses multimodal large language models (MLLM) to evaluate and explain the strengths and weaknesses of different candidate answers through multi-choice reasoning;
MrGuard: A Multilingual Reasoning Guardrail for Universal LLM Safety
Yahan Yang (University of Pennsylvania), Insup Lee (University of Pennsylvania)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose a multilingual reasoning safety guard (MrGuard) to detect and filter unsafe prompts across languages.
MS-RAG: Simple and Effective Multi-Semantic Retrieval-Augmented Generation
Xiaozhou You (Ant Group), Lihong Gu (Ant Group)
RetrievalComputational EfficiencyLarge Language ModelTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a multi-semantic retrieval-augmented generation (MS-RAG) system, utilizing multi-layer indexing combining knowledge graphs and dense vectors, along with hybrid recall and multi-semantic re-ranking, significantly enhancing retrieval and question-answering performance.
MuCAL: Contrastive Alignment for Preference-Driven KG-to-Text Generation
Yifei Song (CNRS/LORIA), Claire Gardent (CNRS/LORIA)
GenerationTransformerLarge Language ModelContrastive LearningTextGraph
🎯 What it does: A multilingual alignment model named MuCAL was constructed to map knowledge graphs (KG) and text into a shared semantic space, and this model is used to automatically generate preference data for training KG-to-Text generators;
MUCAR: Benchmarking Multilingual Cross-Modal Ambiguity Resolution for Multimodal Large Language Models
Xiaolong Wang (Tsinghua University), Yang Liu (Tsinghua University)
Agentic AIPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper constructs MUCAR, a multilingual, cross-modal ambiguity resolution benchmark, evaluates and compares the ambiguity resolution performance of 19 large language models in image-text context, and proposes an explicit reasoning framework based on agents.
Multi-Document Event Extraction Using Large and Small Language Models
Qingkai Min (Zhejiang University), Yue Zhang (Tsinghua University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a collaborative framework that first uses a small language model (SLM) to complete local subtasks such as event triggering, coreference, and argument extraction, and then employs a large language model (LLM) for multi-step reasoning, ultimately achieving cross-document event aggregation and normalization.
Multi-Domain Explainability of Preferences
Nitay Calderon (Technion), Roi Reichart (Technion)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a fully automated multi-domain preference explanation framework: first, using LLM to discover concepts and generate concept definitions, then representing each query-reply triplet as a concept vector, followed by learning shared and domain-specific concept weights through a hierarchical multi-domain regression (HMDR) model, thereby providing local and global explanations for 12 preference mechanisms such as human preferences, LLM judgments, and reward models.
Multi-Frequency Contrastive Decoding: Alleviating Hallucinations for Large Vision-Language Models
Bingqian Liu (Northeastern University), Jingwei Cheng (Northeastern University)
Explainability and InterpretabilityVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the Multi-Frequency Contrastive Decoding (MFCD) method, leveraging the differences between high-frequency and low-frequency image features to suppress object hallucinations in large vision-language models.
Multi-LMentry: Can Multilingual LLMs Solve Elementary Tasks Across Languages?
Luca Moroni (Sapienza University of Rome), Marta Villegas (Barcelona Supercomputing Center)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper created Multi-LMentry, expanding the LMentry benchmark for the first time to nine languages (including low-resource languages) and redesigning and generating 25 basic tasks for each language, totaling nearly 1 million samples.
Multi-Modal Framing Analysis of News
Arnav Arora (University of Copenhagen), Isabelle Augenstein (University of Copenhagen)
ClassificationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Leverages large language models and vision-language models to conduct multimodal, multilabel framework analysis on American news articles (including text and images), constructing a 500k news dataset providing textual, visual, and thematic framework labels, supporting in-depth research on news bias.
Multi-perspective Analysis of Large Language Model Domain Specialization: An Experiment in Accounting Audit Procedures Generation
Yusuke Noro
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextFinance RelatedRetrieval-Augmented Generation
🎯 What it does: This paper conducts experiments on the accounting audit procedure generation task (AAPG), comparing three domain-specialization methods: supervised fine-tuning (SFT-IT, SFT-CV), context learning (ICL), and combinations of both, to explore their differences in text generation characteristics.
Multi-view-guided Passage Reranking with Large Language Models
Jeongwoo Na (Sungkyunkwan University), Jongwuk Lee (Sungkyunkwan University)
RetrievalComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a non-generative large language model re-ranker called MVP based on multi-perspective soft prompts and anchor decoding, to efficiently evaluate all candidate paragraphs in one go.
MultiAgentESC: A LLM-based Multi-Agent Collaboration Framework for Emotional Support Conversation
Yangyang Xu (University of Science and Technology of China), Xun Yang (University of Science and Technology of China)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Propose a training-agnostic multi-agent collaboration framework named MultiAgentESC, which simulates the three-stage process of emotional support dialogues (dialogue analysis, strategy reasoning, and response generation)
MultiDocFusion : Hierarchical and Multimodal Chunking Pipeline for Enhanced RAG on Long Industrial Documents
Joongmin Shin (Korea University), Heuiseok Lim (Korea University)
RetrievalTransformerLarge Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Built a multi-modal document chunking pipeline called MultiDocFusion, specifically designed for long industrial documents, performing visual layout parsing, OCR text extraction, LLM-based chapter-level parsing (DSHP-LLM), and depth-first search (DFS) chunking, ultimately enhancing retrieval and QA performance.
MULTIGUARD: An Efficient Approach for AI Safety Moderation Across Languages and Modalities
Sahil Verma (University of Washington), Chandan Singh (Microsoft)
ClassificationSafty and PrivacyRepresentation LearningTransformerLarge Language ModelImageTextMultimodalityAudio
🎯 What it does: This paper proposes OMNIGUARD, a method that utilizes the internal representations of large language models/multimodal models for detecting harmful prompts;
Multilingual Dialogue Generation and Localization with Dialogue Act Scripting
Justin Vasselli (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
GenerationTransformerLarge Language ModelText
🎯 What it does: Proposed Dialogue Act Script (DAS), a framework for encoding, localizing, and generating multilingual dialogues by abstracting dialogue intentions;
Multilingual Federated Low-Rank Adaptation for Collaborative Content Anomaly Detection across Multilingual Social Media Participants
Jiaxin Li (Huaihua University), Xiaoci Zhang (Heidelberg University)
Anomaly DetectionFederated LearningSupervised Fine-TuningText
🎯 What it does: Propose the MuLA-F framework to achieve LoRA weight separation and orthogonal optimization in multilingual federated learning, aiming to enhance anomaly detection of multilingual social media content.
Multilingual Language Model Pretraining using Machine-translated Data
Jiayi Wang (University College London), Pontus Stenetorp (University College London)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Translated the high-quality English educational corpus FineWebEdu into 9 languages using NMT, generating a multilingual corpus TransWebEdu with 1.7T words, and pre-trained a 1.3B parameter multilingual model TransWebLLM from scratch based on this; subsequently continued pre-training on a small amount of general web corpus, rewritten multiple-choice question corpus, and code/Q&A data.
Multilingual Pretraining for Pixel Language Models
Ilker Kesen (University of Copenhagen), Desmond Elliott (University of Copenhagen)
ClassificationRecognitionRepresentation LearningTransformerVision Language ModelAuto EncoderImageText
🎯 What it does: This paper proposes and trains a multilingual pixel language model called PIXEL-M4, which is pre-trained using four different scripts (Latin, Devanagari, Chinese characters, and Cyrillic) to enhance cross-lingual transfer capabilities.
Multilingual Prompting for Improving LLM Generation Diversity
Qihan Wang (New York University), Emily Black (New York University)
GenerationTransformerPrompt EngineeringText
🎯 What it does: Propose a multilingual prompting method that significantly enhances the diversity of LLM-generated outputs by repeatedly asking questions under different language and cultural contexts and aggregating the answers.
Multilingual vs Crosslingual Retrieval of Fact-Checked Claims: A Tale of Two Approaches
Alan Ramponi (Fondazione Bruno Kessler), Sara Tonelli (Fondazione Bruno Kessler)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies the multilingual and cross-lingual fact-checking claim retrieval (PFCR) task, compares supervised and unsupervised methods, and explores the effectiveness of negative example sampling and re-ranking strategies in multilingual and cross-lingual scenarios.
Multilinguality Does not Make Sense: Investigating Factors Behind Zero-Shot Cross-Lingual Transfer in Sense-Aware Tasks
Roksana Goworek (Queen Mary University of London), Haim Dubossarsky (Queen Mary University of London)
ClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Investigated the zero-shot cross-lingual transfer effectiveness of multilingual pre-trained models on multilingual perception tasks (word sense disambiguation and semantic change detection), and validated through large-scale experiments across 28 languages that multilingualism is not the key factor enhancing transfer performance.
MultiLogicNMR(er): A Benchmark and Neural-Symbolic Framework for Non-monotonic Reasoning with Multiple Extensions
Yeliang Xiu (Sun Yat-sen University), Yongmei Liu (Sun Yat-sen University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the multi-extension non-monotonic reasoning dataset MultiLogicNMR and its OOD and NL variants, and proposed the neural-symbolic framework MultiLogicNMRer.
MultiMatch: Multihead Consistency Regularization Matching for Semi-Supervised Text Classification
Iustin Sirbu (National University of Science and Technology POLITEHNICA Bucharest), Traian Rebedea (National University of Science and Technology POLITEHNICA Bucharest)
ClassificationTransformerTextMultimodality
🎯 What it does: Propose MultiMatch, a semi-supervised text classification framework combining multi-head co-training, consistency regularization, and pseudo labels, and design a pseudo label weighting module for selection, filtering, and weighting.
MultiMed-ST: Large-scale Many-to-many Multilingual Medical Speech Translation
Khai Le-Duc (University of Toronto), Thanh Nguyen-Tang (New Jersey Institute of Technology)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataAudio
🎯 What it does: The paper systematically constructs a large-scale multilingual medical speech translation dataset called MultiMed-ST and conducts in-depth experiments and analysis in this field.
Multimedia Event Extraction with LLM Knowledge Editing
Jiaao Yu (Beijing University of Posts and Telecommunications), Lanlan Rui (Beijing University of Posts and Telecommunications)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose a multimodal event extraction framework based on LLM knowledge editing, enhancing the understanding of event structure through multi-layer redundant neuron selection and mask editing.
Multimodal Fine-grained Context Interaction Graph Modeling for Conversational Speech Synthesis
Zhenqi Jia, Haizhou Li (Johns Hopkins University)
GenerationGraph Neural NetworkTransformerMultimodality
🎯 What it does: Construct the MFCIG-CSS system, which models word-level semantics and prosody in multi-modal dialogue history using two fine-grained interaction graphs (Semantic Interaction Graph SIG and Prosody Interaction Graph PIG), and injects the encoded interaction features into the speech synthesizer to generate more natural and conversation-like speech.
Multimodal Language Models See Better When They Look Shallower
Haoran Chen (Zhejiang Gongshang University), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative)
RecognitionObject DetectionTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Systematically studied the impact of different hierarchical features in the visual encoder of multimodal large language models (MLLMs) on model performance, and proposed a lightweight fusion method that concatenates shallow, medium, and deep features.
Multimodal Neural Machine Translation: A Survey of the State of the Art
Yi Feng, Vincent Ng (Nanjing University)
Large Language ModelVision Language ModelMultimodalityReview/Survey Paper
🎯 What it does: This paper systematically reviews the current research status of multimodal neural machine translation (MNMT), including task definition, main challenges, datasets, method classification, and evaluation methods.
MULTIVOX: A Benchmark for Evaluating Voice Assistants for Multimodal Interactions
Ramaneswaran Selvakumar (University of Maryland), Dinesh Manocha (University of Maryland)
Vision Language ModelImageVideoMultimodalityBenchmarkAudio
🎯 What it does: Propose the MULTIVOX benchmark, collecting 1000 professionally recorded voice queries with corresponding images/videos to evaluate the context-awareness and answer quality of multimodal language models in integrating speech and vision.
MUSE: MCTS-Driven Red Teaming Framework for Enhanced Multi-Turn Dialogue Safety in Large Language Models
Siyu Yan (East China Normal University), Chenjuan Guo (East China Normal University)
Safty and PrivacyAdversarial AttackReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Proposed the MUSE framework, which includes MUSE-A (based on framework semantics and MCTS) for multi-round jailbreak attacks, and MUSE-D (fine-grained security alignment) for multi-round security defense;
MuseScorer: Idea Originality Scoring At Scale
Ali Sarosh Bangash (University of South Florida), Raiyan Abdul Baten (University of South Florida)
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes MUSESCORER, a fully automated and scalable originality scoring system that leverages LLMs as judges and external retrieval to perform bucketing of creative responses and calculate frequency-based originality scores.
MusKGC: A Flexible Multi-source Knowledge Enhancement Framework for Open-World Knowledge Graph Completion
Xin Song (National University of Defense Technology), Bin Zhou (National University of Defense Technology)
TransformerLarge Language ModelTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-source knowledge-enhanced framework called MusKGC based on large language models to complete knowledge graphs under the open-world assumption;
MuTIS: Enhancing Reasoning Efficiency through Multi Turn Intervention Sampling in Reinforcement Learning
Wenshuo Zhao (Zhejiang University), Linchao Zhu (Zhejiang University)
Computational EfficiencyLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the Multi-Turn Intervention Sampling (MuTIS) framework, which truncates tokens during reasoning based on a token limit and inserts intervention prompts, followed by fine-tuning large language models using reinforcement learning (PPO) to achieve shorter and more accurate reasoning chains.
MUZO: Leveraging Multiple Queries and Momentum for Zeroth-Order Fine-Tuning of Large Language Models
Yuezhang Peng (Shanghai Jiao Tong University), Xie Chen (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the MUZO method, combining multi-query zeroth-order optimization with Adam to achieve low-memory LLM fine-tuning
MythTriage: Scalable Detection of Opioid Use Disorder Myths on a Video-Sharing Platform
Hayoung Jung (Princeton University), Tanu Mitra
ClassificationRecommendation SystemAnomaly DetectionTransformerLarge Language ModelVideoText
🎯 What it does: This paper constructs a large-scale detection and analysis framework for YouTube platform-related misleading myths about opioid use disorder (OUD). It first collaborates with clinical experts to identify 8 key myths and create an expert-annotated golden dataset. Subsequently, it designs the MYTHTRIAGE three-stage annotation pipeline, which uses a lightweight model for preliminary classification and then assigns difficult samples to a high-performance LLM (GPT-4o) for reasoning. Finally, the pipeline is applied to 164K recommended videos to complete approximately 1.3M labels, and quantitatively analyzes the prevalence of myths in search results and recommendation paths.
N-CORE: N-View Consistency Regularization for Disentangled Representation Learning in Nonverbal Vocalizations
Siddhant Bikram Shah, Kristina T. Johnson (Northeastern University)
ClassificationRepresentation LearningTransformerContrastive LearningAudio
🎯 What it does: Proposed a non-verbal speech separation representation learning framework called N-CORE based on multi-view consistency regularization, which can separate emotional and speaker information in non-verbal speech.
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
Kinjal Basu (IBM Research), Pavan Kapanipathi (IBM Research)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose the NESTFUL benchmark to evaluate the performance of large language models on nested API call sequences, covering over 1800 executable multi-level nested calls.
Neural Topic Modeling via Contextual and Graph Information Fusion
Jiyuan Liu (Sun Yat-sen University), Yanghui Rao (Sun Yat-sen University)
Representation LearningAuto EncoderTextGraph
🎯 What it does: This paper proposes a variational autoencoder framework (CGTM) that integrates contextual and graph information for unsupervised topic modeling.
NeuroAda: Activating Each Neuron’s Potential for Parameter-Efficient Fine-Tuning
Zhi Zhang (University of Amsterdam), Ekaterina Shutova (University of Amsterdam)
Computational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes NeuroAda, a parameter-efficient fine-tuning method that introduces a small number of trainable bypass connections for each neuron while freezing the original model weights.
Neuron-Level Differentiation of Memorization and Generalization in Large Language Models
Ko-Wei Huang (National Taiwan University), Shou-De Lin (National Taiwan University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper constructs tasks to distinguish memory and generalization, performing neuron-level behavior separation and intervention on GPT-2 and LoRA fine-tuned LLaMA, proving that memory and generalization correspond to different neuron subsets within the model.
NEXUS: Network Exploration for eXploiting Unsafe Sequences in Multi-Turn LLM Jailbreaks
Javad Rafiei Asl (Old Dominion University), Daniel Takabi (Old Dominion University)
Adversarial AttackLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed a modular framework called NEXUS for constructing, refining, and executing multi-round LLM jailbreak attacks.
NILE: Internal Consistency Alignment in Large Language Models
Minda Hu (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the NILE framework, which extracts, revises, and filters the IFT dataset using internal knowledge to enhance the fine-tuning effectiveness of LLMs.
NileChat: Towards Linguistically Diverse and Culturally Aware LLMs for Local Communities
Abdellah El Mekki (University of British Columbia), Muhammad Abdul-Mageed (University of British Columbia)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose a multi-stage framework that generates pre-trained corpora aligned with Egyptian and Moroccan Arabic dialects and their cultural values through machine translation, controlled synthesis generation, and retrieval-augmented data augmentation, and trains a 3B-parameter NileChat LLM based on this data.
NitiBench: Benchmarking LLM Frameworks on Thai Legal Question Answering Capabilities
Pawitsapak Akarajaradwong (VISAI AI), Sarana Nutanong (Vidyasirimedhi Institute of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed NitiBench, a benchmark tailored for Thai legal question answering, comprising manually annotated test sets across two domains (Corporate and Commercial Law and Tax Law), along with proposed metrics for multi-label retrieval and end-to-end (E2E) evaluation.
NL-Debugging: Exploiting Natural Language as an Intermediate Representation for Code Debugging
Weiming Zhang (Shanghai Jiao Tong University), Weinan Zhang (Huawei Noah's Ark Lab)
AI Code AssistantTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposes the NL-DEBUGGING framework, which uses natural language as an intermediate representation for debugging buggy code.
NL2Lean: Translating Natural Language into Lean 4 through Multi-Aspect Reinforcement Learning
Yue Fang (Peking University), Zhi Jin (Peking University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A reinforcement learning-based framework called ReLean is studied to translate natural language into Lean 4 formal statements.
No Need for Explanations: LLMs can implicitly learn from mistakes in-context
Lisa Alazraki (Imperial College London), Max Bartolo (Cohere)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Studied the implicit learning approach in LLMs where only incorrect answers are provided without explanations, and compared it with explicit learning and traditional Chain-of-Thought (CoT)
Noise, Adaptation, and Strategy: Assessing LLM Fidelity in Decision-Making
Yuanjun Feng (University of Lausanne), Yash Raj Shrestha (University of Lausanne)
TransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: Propose a process-oriented evaluation framework that integrates three-stage interventions of 'intrinsicness,' 'instructions,' and 'mimicry' to systematically examine the decision-making behaviors of large language models (LLMs) in second-price auctions and newspaper subscription problems.
Non-Existent Relationship: Fact-Aware Multi-Level Machine-Generated Text Detection
Yang Wu (China Telecom Cloud Computing Research Institute), Jie Wu (China Telecom Cloud Computing Research Institute)
ClassificationGraph Neural NetworkTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a multi-layer feature fusion model (FAML) based on the authenticity of entity relationships for detecting machine-generated text.
NormGenesis: Multicultural Dialogue Generation via Exemplar-Guided Social Norm Modeling and Violation Recovery
Minki Hong (Dongguk University), Jihie Kim (Dongguk University)
GenerationData SynthesisTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the NormGenesis framework for generating and refining multi-turn dialogues that comply with social norms in English, Chinese, and Korean, achieving post-violation repair through the Violation-to-Resolution (V2R) pattern.
NormXLogit: The Head-on-Top Never Lies
Sina Abbasi (Tehran Institute for Advanced Studies, Khatam University), Mohammad Taher Pilehvar (Cardiff University)
ClassificationExplainability and InterpretabilityTransformerText
🎯 What it does: Proposes a generic, gradient-free token importance explanation method called NormXLogit, which combines word embedding norms with model head outputs (LogAt) to evaluate the contribution of each input token to the final prediction.
Not All Parameters Are Created Equal: Smart Isolation Boosts Fine-Tuning Performance
Yao Wang (University of New South Wales), Minlong Peng (Fudan University)
Computational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes the Core Parameter Isolation Fine-tuning (CPI-FT) framework, which first individually fine-tunes and identifies core parameter regions for each task, then clusters tasks based on the overlap of core regions. Subsequently, core parameters are directly merged, and non-core parameters are smoothly fused using SLERP. Finally, lightweight multi-stage fine-tuning is performed on mixed task data, with the core parameters of already learned tasks frozen, addressing the issues of task interference and catastrophic forgetting in multi-task SFT.
Not What the Doctor Ordered: Surveying LLM-based De-identification and Quantifying Clinical Information Loss
Kiana Aghakasiri (University of Alberta), Mohamed Abdalla (University of Alberta)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataElectronic Health RecordsReview/Survey Paper
🎯 What it does: This paper presents a systematic review of research on clinical text de-identification using large language models (LLMs), evaluating multiple de-identification models (e.g., Llama-3.3, ClinicalBERT, Deidentify, Presidio) on two large datasets. It quantifies clinical information retention (CIR) during de-identification using standard classification metrics and a newly proposed CIRE metric, while manually validating and improving existing evaluation methods (e.g., JSC).