EMNLP 2024 Papers — Page 8
Conference on Empirical Methods in Natural Language Processing · 1268 papers
M3D: MultiModal MultiDocument Fine-Grained Inconsistency Detection
Chia-Wei Tang (Virginia Tech), Chris Thomas (Virginia Tech)
Anomaly DetectionGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes a multi-modal, multi-document fine-grained inconsistency detection method that can infer the truth value for each clause or entity/relationship in a claim and provide corresponding evidence.
M3Hop-CoT: Misogynous Meme Identification with Multimodal Multi-hop Chain-of-Thought
Gitanjali Kumari (Indian Institute of Technology Patna), Asif Ekbal (Indian Institute of Technology Jodhpur)
ClassificationTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodalityChain-of-Thought
🎯 What it does: Construct a framework (M3Hop-CoT) that parses and identifies female hate memes from three dimensions—emotion, intent, and context—by combining the multi-hop chain-of-thought reasoning of large language models with the CLIP vision-language model and scene graphs.
MAgIC: Investigation of Large Language Model Powered Multi-Agent in Cognition, Adaptability, Rationality and Collaboration
Lin Xu (National University of Singapore), Jiashi Feng (National University of Singapore)
TransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented a competition-based multi-agent evaluation framework called MAgIC to quantify seven capabilities of large language models (LLMs) in multi-agent environments, including judgment, reasoning, deception, self-awareness, cooperation, coordination, and rationality.
MAIR: A Massive Benchmark for Evaluating Instructed Retrieval
Weiwei Sun (Carnegie Mellon University), Zhaochun Ren (Leiden University)
RetrievalTextBenchmark
🎯 What it does: This paper proposes MAIR, a large-scale retrieval benchmark covering 126 different IR tasks and containing 805 manually annotated retrieval instructions, aiming to evaluate the generalization and instruction-following capabilities of instruction-tuned retrieval models.
Major Entity Identification: A Generalizable Alternative to Coreference Resolution
Kawshik Manikantan (IIIT Hyderabad), Vineet Gandhi (IIIT Hyderabad)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes a new reference task called Main Entity Identification (MEI), which replaces traditional coreference resolution tasks by predefining the entities of focus.
Make Some Noise: Unlocking Language Model Parallel Inference Capability through Noisy Training
Yixuan Wang (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Introduce noise input during the supervised fine-tuning phase to train a large language model (LLM) capable of parallel reasoning, and propose a tree-structured retrieval-enhanced Jacobi decoding strategy
Making Large Language Models Better Reasoners with Orchestrated Streaming Experiences
Xiangyang Liu (Fudan University), Xipeng Qiu (Fudan University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the RoSE framework, which utilizes a streaming experience pool to continuously collect and organize answered questions and reasoning paths of LLMs, achieving automatic, self-improving reasoning enhancement.
MaPPER: Multimodal Prior-guided Parameter Efficient Tuning for Referring Expression Comprehension
Ting Liu (National University of Defense Technology), Quanjun Yin (Hefei University of Technology)
RecognitionComputational EfficiencyTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a multi-modal prior-guided parameter-efficient fine-tuning framework called MaPPER for referring expression understanding tasks.
MAR: Matching-Augmented Reasoning for Enhancing Visual-based Entity Question Answering
Zhengxuan Zhang (Hong Kong University of Science and Technology), Nan Tang (Hong Kong University of Science and Technology)
RetrievalGraph Neural NetworkPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Matching-Augmented Reasoning (MAR) method, using a matching graph to perform fine-grained retrieval and reasoning on Visual Entity Question Answering (VEQA), addressing the challenges of MLLMs in identifying individual entities.
MARE: Multi-Aspect Rationale Extractor on Unsupervised Rationale Extraction
Han Jiang (Central South University), Jianxin Wang (Central South University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a multi-aspect unsupervised rationale extraction model called MARE, which can simultaneously predict and explain multiple aspects of text within a single model.
MASIVE: Open-Ended Affective State Identification in English and Spanish
Nicholas Deas (Columbia University), Kathleen McKeown (Columbia University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the Affective State Identification (ASI) task, which utilizes generative models to predict emotional states present in text;
MatchTime: Towards Automatic Soccer Game Commentary Generation
Jiayuan Rao (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)
GenerationTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper constructs an automatic football match commentary system by first performing time alignment on existing data and then training a generative model using new data.
Mathador-LM: A Dynamic Benchmark for Mathematical Reasoning on Large Language Models
Eldar Kurtic (ISTA & Neural Magic, Inc.), Dan Alistarh (ISTA & Neural Magic, Inc.)
GenerationData SynthesisTransformerPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the dynamically generated Mathador-LM benchmark to evaluate large language models' mathematical reasoning, rule parsing, planning, and problem-solving capabilities;
Matryoshka-Adaptor: Unsupervised and Supervised Tuning for Smaller Embedding Dimensions
Jinsung Yoon (Google Cloud AI), Tomas Pfister (Google Cloud AI)
Computational EfficiencyRepresentation LearningLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: Proposes a novel tuning framework called Matryoshka-Adaptor for customizing embeddings of large language models (LLMs), aiming to significantly reduce the embedding dimensionality while maintaining performance, thereby enhancing computational efficiency and cost-effectiveness.
mDPO: Conditional Preference Optimization for Multimodal Large Language Models
Fei Wang (University of Southern California), Muhao Chen (University of California, Davis)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose MDPO, improving the direct preference optimization (DPO) method for multimodal LLMs to better utilize visual information.
MEANT: Multimodal Encoder for Antecedent Information
Benjamin Irving, Annika Marie Schoene (Northeastern University)
ClassificationTransformerVision Language ModelImageTextMultimodalityTime SeriesFinance Related
🎯 What it does: Proposed the MEANT model—a multimodal encoder integrating language, visual, and temporal dimensions for predicting momentum signals in stock markets, and released a new dataset called TempStock.
Measuring Psychological Depth in Language Models
Fabrice Y Harel-Canada (University of California Los Angeles), Nanyun Peng
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed and validated the Psychological Depth Scale (PDS) to assess the psychological depth of large language models (LLMs) and human-authored short stories in terms of emotional resonance, immersion, emotional arousal, authenticity, and narrative complexity.
MedAdapter: Efficient Test-Time Adaptation of Large Language Models Towards Medical Reasoning
Wenqi Shi (Georgia Tech), May Dongmei Wang
Domain AdaptationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataChain-of-Thought
🎯 What it does: Propose MedAdapter, a lightweight test-time adapter that leverages BERT-sized models to rank candidate answers generated by LLMs, achieving efficient adaptation in medical reasoning tasks;
MedCoT: Medical Chain of Thought via Hierarchical Expert
Jiaxiang Liu (Zhejiang University), Zuozhu Liu (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsVision Language ModelMultimodalityBiomedical DataChain-of-Thought
🎯 What it does: Propose MedCoT, a hierarchical expert verification-based medical vision-and-language question-answering model;
Media Attitude Detection via Framing Analysis with Events and their Relations
Jin Zhao (Brandeis University), Nianwen Xue (Brandeis University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By extracting events, event coreference, and causal relationships between events, three framework device encodings are constructed, which are used as inputs for news media attitude detection.
Medical Adaptation of Large Language and Vision-Language Models: Are We Making Progress?
Daniel P Jeong, Michael Oberst (Johns Hopkins University)
Domain AdaptationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBiomedical Data
🎯 What it does: Compared the performance of medical LLMs and VLMs obtained through domain-adapted pretraining (DAPT) with their original base models on zero/one-shot medical question answering (QA) tasks.
MediTOD: An English Dialogue Dataset for Medical History Taking with Comprehensive Annotations
Vishal Vivek Saley, Mausam .
Data-Centric LearningTransformerLarge Language ModelTextBiomedical DataBenchmark
🎯 What it does: Developed an English doctor-patient dialogue dataset called MediTOD, and constructed a Comprehensive Medical Attribute Schema (CMAS) using a questionnaire-based annotation method, achieving precise association between symptoms and attributes;
MedReadMe: A Systematic Study for Fine-grained Sentence Readability in Medical Domain
Chao Jiang (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataBenchmark
🎯 What it does: Propose the MEDREADME dataset, containing 4,520 medical sentences with readability scores and fine-grained complex span annotations, and conduct in-depth analysis through the 'Google-Easy'/'Google-Hard' classification of medical terminology; perform systematic experiments on medical text readability, investigating the impact of vocabulary, syntax, and terminology on difficulty; improve traditional readability formulas by introducing term counting to enhance correlation with human evaluation; train and evaluate complex span identification models.
MemeCLIP: Leveraging CLIP Representations for Multimodal Meme Classification
Siddhant Bikram Shah (Northeastern University), Haohan Wang (University of Illinois Urbana-Champaign)
ClassificationTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes the MemeCLIP framework, which achieves multi-faceted (hate, target, stance, humor) meme classification using a frozen CLIP model and lightweight modules.
Memorize Step by Step: Efficient Long-Context Prefilling with Incremental Memory and Decremental Chunk
Zhiyuan Zeng (Fudan University), Xipeng Qiu (Fudan University)
CompressionComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Studied the efficiency issues of large language models during the long-text pre-filling phase, proposing two dynamic strategies, Incremental Memory and Decremental Chunk, to improve KV cache compression, enhance inference speed, and reduce GPU memory usage.
Memory-Efficient Fine-Tuning of Transformers via Token Selection
Antoine Simoulin (Meta AI), Grey Yang (Meta AI)
Computational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Proposed the TOKEN-TUNE method, which significantly reduces the activation memory storage requirements during fine-tuning of Transformer models by computing gradients for only a portion of input tokens during backpropagation.
Mentor-KD: Making Small Language Models Better Multi-step Reasoners
Hojae Lee (Korea University), SangKeun Lee (Korea University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Designed and verified the Mentor-KD framework, which combines medium-scale task-specific mentor models with large LLM teachers' chain-of-thought (CoT) results to generate additional multi-step reasoning samples and soft labels, further enhancing the performance of small language models on complex reasoning tasks.
Message Passing on Semantic-Anchor-Graphs for Fine-grained Emotion Representation Learning and Classification
Pinyi Zhang (East China Normal University), Kai Zhang (Fudan University)
ClassificationRepresentation LearningGraph Neural NetworkLarge Language ModelTextGraph
🎯 What it does: Propose SEAN-GNN, which constructs a fixed-size graph using semantic anchors to capture semantic distribution in sentences and temporal relationships between words, and achieves sentiment feature fusion through message passing;
MetaGPT: Merging Large Language Models Using Model Exclusive Task Arithmetic
Yuyan Zhou (Baichuan Inc.), Weipeng Chen (Baichuan Inc.)
Meta LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a MetaGPT method that achieves multi-task learning by fusing LLMs through model-specific task arithmetic without using any data.
MetaReflection: Learning Instructions for Language Agents using Past Reflections
Priyanshu Gupta (Microsoft), Sherry Shi (Microsoft)
Meta LearningLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose an offline reinforcement learning method called METAREFLECTION, which enhances the performance of language agents by converting past self-reflection into transferable meta-reflection rules and constructing semantic memory.
Methods of Automatic Matrix Language Determination for Code-Switched Speech
Olga Iakovenko (University of Sheffield), Thomas Hain (University of Sheffield)
ClassificationConvolutional Neural NetworkTransformerTextAudio
🎯 What it does: This study constructs and evaluates three matrix language identity (MLID) discrimination systems based on text and audio using the matrix language framework (MLF) theory, and compares them with traditional language identification (LID) methods;
Metrics for What, Metrics for Whom: Assessing Actionability of Bias Evaluation Metrics in NLP
Pieter Delobelle (KU Leuven), Zeerak Talat (Mohamed bin Zayed University of Artificial Intelligence)
TextReview/Survey Paper
🎯 What it does: This paper introduces the concept of 'actionability,' which measures whether bias metric results can support practical decision-making and interventions, and outlines six key requirements for actionability. Subsequently, a systematic review of 146 NLP bias measurement papers was conducted to examine whether they meet these requirements, revealing widespread issues such as missing information on motivation, construct, and reliability in current research. Based on this, the paper provides improvement recommendations.
MIBench: Evaluating Multimodal Large Language Models over Multiple Images
Haowei Liu (Institute of Automation, Chinese Academy of Sciences), Weiming Hu (Institute of Automation, Chinese Academy of Sciences)
TransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a large-scale multi-image benchmark, MIBench, to evaluate the capabilities of multimodal large language models (MLLMs) in multi-image scenarios, covering three major scenarios (multi-image instruction, cross-modal knowledge search, and multimodal context learning) and 13 tasks, totaling approximately 13K high-quality samples.
Middleware for LLMs: Tools Are Instrumental for Language Agents in Complex Environments
Yu Gu (Ohio State University), Yu Su (Cisco Research)
TransformerLarge Language ModelAgentic AIGraphTabularChain-of-Thought
🎯 What it does: Proposed a 'Middleware' tool framework that utilizes specially designed tools to help large language models (LLMs) proactively explore and acquire information in complex environments such as databases and knowledge graphs, thereby better completing tasks.
MIND: Multimodal Shopping Intention Distillation from Large Vision-language Models for E-commerce Purchase Understanding
Baixuan Xu (Hong Kong University of Science and Technology), Yangqiu Song (Amazon.com Inc)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Propose the MIND framework, leveraging large vision-language models to generate multimodal (text-image) shopping intents based on product co-purchase records, and automatically filter them using a human-like role-aware filter, constructing a 1.26M intention knowledge base.
MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents
Liyan Tang (University of Texas at Austin), Greg Durrett (University of Texas at Austin)
Data SynthesisComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The study proposes a lightweight fact-checking model called MiniCheck based on synthetic training data, which can verify facts in LLM-generated content against supporting documents without relying on sentence splitting or context resolution.
MiniConGTS: A Near Ultimate Minimalist Contrastive Grid Tagging Scheme for Aspect Sentiment Triplet Extraction
Qiao Sun (Shanghai Artificial Intelligence Laboratory), Qinying Gu (Shanghai Artificial Intelligence Laboratory)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Propose the MiniConGTS method, combining a minimalist grid tagging scheme and token-level contrastive learning to achieve end-to-end efficient extraction in the Aspect Sentiment Triplet Extraction (ASTE) task.
MIPD: Exploring Manipulation and Intention In a Novel Corpus of Polish Disinformation
Arkadiusz Modzelewski (Polish-Japanese Academy of Information Technology), Adam Wierzbicki (Polish-Japanese Academy of Information Technology)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper constructs the Polish MIPD dataset, collecting and annotating 15,356 web articles with annotations on credibility, manipulation techniques, intent types, and topics, and trains various baseline models on this dataset.
MirrorStories: Reflecting Diversity through Personalized Narrative Generation with Large Language Models
Sarfaroz Yunusov (Brock University), Ali Emami (Brock University)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: The study utilizes large language models to generate personalized 'mirror stories' and constructs a corpus of 1500 short stories, personalized based on name, gender, age, ethnicity, interests, and story moral;
MisinfoEval: Generative AI in the Era of “Alternative Facts”
Saadia Gabriel (University of California, Los Angeles), Asuman E. Ozdaglar
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextTabular
🎯 What it does: Design and evaluate a misinformation intervention method (MisinfoEval) generated by large language models (LLMs) in a simulated social media environment, verifying its effectiveness through two-phase experiments.
Mitigate Extrinsic Social Bias in Pre-trained Language Models via Continuous Prompts Adjustment
Yiwei Dai (Jilin University), Xin Wang (Jilin University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a continuous prompt adjustment method called CPAD, which performs external debiasing in downstream NLU tasks by generating a continuous token list from the full vocabulary, avoiding the limitations of manually crafted vocabularies.
Mitigating Frequency Bias and Anisotropy in Language Model Pre-Training with Syntactic Smoothing
Richard Diehl Martinez (University of Cambridge), Lisa Beinborn (University of Cambridge)
Representation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a 'Syntactic Smoothing' method that smooths backpropagation signals for semantically similar words during the pre-training phase to reduce language models' bias toward word frequency and alleviate anisotropy in the representation space.
Mitigating Language Bias of LMMs in Social Intelligence Understanding with Virtual Counterfactual Calibration
Peng Chen (Tianjin University), Zhiyong Feng (Tianjin University)
TransformerLarge Language ModelMultimodality
🎯 What it does: This paper proposes an output distribution calibration network and a virtual counterfactual augmentation framework to reduce language bias in social intelligent question answering for multimodal models.
Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation
Yongsen Zheng (Nanyang Technological University), Kwok-Yan Lam (Nanyang Technological University)
Recommendation SystemGraph Neural NetworkTransformerContrastive LearningTextGraph
🎯 What it does: Propose the HiCore framework, which leverages multi-channel hypergraphs to learn multi-layer user interests and enhances both recommendation and dialogue tasks in dialogue recommendation systems through a self-supervised approach, aiming to alleviate the worsening Matthew effect over time.
Mitigating Open-Vocabulary Caption Hallucinations
Assaf Ben-Kish (Tel Aviv University), Hadar Averbuch-Elor (Tel Aviv University)
GenerationLarge Language ModelReinforcement LearningDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the OpenCHAIR benchmark and the MOCHa method for quantifying and mitigating hallucinations in open-vocabulary image captioning.
Mitigating the Alignment Tax of RLHF
Yong Lin (Princeton University), Tong Zhang (University Of Illinois Urbana Champaign)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper systematically studies the 'alignment tax' (forgetting of pre-training capabilities) that occurs during the RLHF process, compares various methods to mitigate forgetting, and finds that simple model averaging (MA) achieves the optimal Pareto frontier between alignment rewards and forgetting. Based on this finding, the authors propose heterogeneous model averaging (HMA), which further improves alignment rewards and reduces forgetting by assigning different averaging proportions to different Transformer layers. Experiments are conducted on OpenLLaMA-3B and Mistral-7B, covering multiple RLHF algorithms (RSF, PPO, DPO).
Mitigating the Impact of Reference Quality on Evaluation of Summarization Systems with Reference-Free Metrics
Théo Gigant (Université Paris-Saclay, CNRS, CentraleSupélec, Laboratoire des signaux et systèmes), Frederic Dufaux (Université Paris-Saclay, CNRS, CentraleSupélec, Laboratoire des signaux et systèmes)
GenerationRetrievalLarge Language ModelTextBenchmark
🎯 What it does: Propose a no-reference text summarization quality evaluation metric by measuring the relevance of the summary through tf-idf weighted important n-grams from the source document and overlapping vocabulary with the summary.
Mitigating the Language Mismatch and Repetition Issues in LLM-based Machine Translation via Model Editing
Weichuan Wang (City University of Hong Kong), Ying Wei (Zhejiang University)
GenerationTransformerLarge Language ModelText
🎯 What it does: When using large language models (LLM) for machine translation, language mismatch and repetition errors frequently occur. The authors attempt to mitigate these errors through model editing techniques (Function Vectors and Knowledge Neurons). However, direct model editing has limited effectiveness or may negatively impact overall translation quality. To address this, the study proposes identifying modules across different language settings, taking their intersection to screen attention heads and FFN neurons genuinely associated with errors, and then performing targeted edits. This approach significantly reduces error rates while maintaining or improving overall translation quality.
Mitigating Training Imbalance in LLM Fine-Tuning via Selective Parameter Merging
Yiming Ju (Beijing Academy of Artificial Intelligence), Zheng Zhang (Beijing Academy of Artificial Intelligence)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study finds that the order of training samples during LLM fine-tuning leads to significant training imbalance, and mitigates this issue by fusing fine-tuned models trained under different data orders.
MiTTenS: A Dataset for Evaluating Gender Mistranslation
Kevin Robinson (Google DeepMind), Jasmijn Bastings (Google DeepMind)
Data SynthesisLarge Language ModelTextBenchmark
🎯 What it does: Proposed the MiTTenS dataset to assess translation systems' errors in gender translation, covering 26 languages and including 13 evaluation sets specifically designed for gender mis-translation.
MixGR: Enhancing Retriever Generalization for Scientific Domain through Complementary Granularity
Fengyu Cai (Technical University of Darmstadt), Heinz Koeppl (Technical University of Darmstadt)
RetrievalTransformerContrastive LearningTextBenchmark
🎯 What it does: By splitting queries into sub-queries, documents into propositions, and computing similarity at multiple granularities followed by RRF fusion, the generalization performance of document retrieval in scientific domains is improved.
Mixture-of-Modules: Reinventing Transformers as Dynamic Assemblies of Modules
Zhuocheng Gong (Peking University), Rui Yan (Renmin University of China)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: Redesign the Transformer by decomposing it into dynamically selectable modules, including multi-head attention, feed-forward networks, and skip modules, which are assembled on-demand during the forward pass, breaking away from the traditional depth-sequential structure.
Mixture-of-Skills: Learning to Optimize Data Usage for Fine-Tuning Large Language Models
Minghao Wu (Monash University), Reza Haf (Monash University)
OptimizationData-Centric LearningSupervised Fine-TuningReinforcement LearningText
🎯 What it does: When fine-tuning large language models across multiple datasets, reinforcement learning is utilized to automatically learn and dynamically adjust the sampling probabilities of different datasets, achieving an optimal data usage strategy;
Mixture-of-Subspaces in Low-Rank Adaptation
Taiqiang Wu (University of Hong Kong), Ngai Wong (University of Hong Kong)
Computational EfficiencyRepresentation LearningMixture of ExpertsImageTextMultimodalityBenchmark
🎯 What it does: Propose the MoSLoRA method, which splits the low-rank decomposition of LoRA into subspaces and introduces a learnable mixer to more flexibly integrate subspaces, achieving more efficient parameterized fine-tuning.
MLLM-Protector: Ensuring MLLM’s Safety without Hurting Performance
Renjie Pi (Hong Kong University of Science and Technology), Tong Zhang (University of Illinois at Urbana-Champaign)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBenchmark
🎯 What it does: Proposed a pluggable defense framework called MLLM-Protector to defend against malicious visual inputs that cause multimodal large language models (MLLMs) to generate harmful content
MMNeuron: Discovering Neuron-Level Domain-Specific Interpretation in Multimodal Large Language Model
Jiahao Huo (Hong Kong University of Science and Technology Guangzhou), Xuming Hu (Hong Kong University of Science and Technology Guangzhou)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper calculates the domain activation probability entropy (DAPE) of neurons in multimodal large language models, identifies and marks a small number (<1%) of domain-specific neurons, and evaluates their impact on visual question answering (VQA) tasks through deactivation experiments and logit lens analysis, revealing that existing models underutilize domain-specific information.
MMoE: Enhancing Multimodal Models with Mixtures of Multimodal Interaction Experts
Haofei Yu (Carnegie Mellon University), Paul Pu Liang (Massachusetts Institute of Technology)
Representation LearningTransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Propose the MMOE (Multimodal Mixtures of Experts) framework, which trains specialized expert models for different types of multimodal interactions (redundancy, uniqueness, collaboration) and dynamically fuses experts based on input during inference to enhance multimodal model performance.
MMTE: Corpus and Metrics for Evaluating Machine Translation Quality of Metaphorical Language
Shun Wang (University of Sheffield), Chenghua Lin (University of Sheffield)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This study constructs a multilingual metaphor translation evaluation corpus (MMTE) and proposes a systematic human evaluation framework for fine-grained assessment of metaphor translation quality between English-Chinese and English-Italian;
MoCoKGC: Momentum Contrast Entity Encoding for Knowledge Graph Completion
Qingyang Li (Guilin University of Electronic Technology), Yuchu Qin (Guilin University of Electronic Technology)
Representation LearningTransformerPrompt EngineeringContrastive LearningGraph
🎯 What it does: Propose the MoCoKGC model, which utilizes momentum contrastive learning combined with entity-relation encoding, entity encoding, and momentum entity encoding to achieve knowledge graph completion.
Model Balancing Helps Low-data Training and Fine-tuning
Zihang Liu (Dartmouth College), Yaoqing Yang (Dartmouth College)
OptimizationTransformerSupervised Fine-TuningTextBiomedical Data
🎯 What it does: Under low-data training and fine-tuning scenarios, this paper leverages Heavy-Tail Self-Regularization (HT-SR) theory to detect imbalanced training quality across model layers, and proposes the TempBalance method based on hierarchical learning rate scheduling to balance the training progress of each layer, thereby enhancing model performance.
Model Editing Harms General Abilities of Large Language Models: Regularization to the Rescue
Jia-Chen Gu (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The paper studies the potential side effects of large language models (LLMs) on their general capabilities (reasoning, question answering, sentiment analysis, etc.) during knowledge editing, and proposes a regularization method called RECT to mitigate overfitting caused by editing.
Model Internals-based Answer Attribution for Trustworthy Retrieval-Augmented Generation
Jirui Qi (University of Groningen), Arianna Bisazza (University of Amsterdam)
GenerationExplainability and InterpretabilityTextRetrieval-Augmented Generation
🎯 What it does: In retrieval-augmented generation (RAG), a framework for answer attribution based on internal model information called MIRAGE is proposed, which can perform credible and interpretable attribution of generated answers based on context.
Model-based Preference Optimization in Abstractive Summarization without Human Feedback
Jaepill Choi (Seoul National University), Taesup Kim (Seoul National University)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Construct an automated preference optimization process by using summaries generated by the model under different decoding strategies (deterministic beam search and stochastic temperature sampling) as preference pairs, enhancing the accuracy and reliability of abstractive summarization.
Modeling Layout Reading Order as Ordering Relations for Visually-rich Document Understanding
Chong Zhang (Fudan University), Tao Gui (Shanghai Key Laboratory of Intelligent Information Processing)
TransformerVision Language ModelTextMultimodalityBenchmark
🎯 What it does: This paper redefines the reading order prediction task for visual-rich documents, modeling reading order as a sorting relationship between layout elements rather than traditional full permutation sequences;
Modeling Nonnative Sentence Processing with L2 Language Models
Tatsuya Aoyama (Georgetown University), Nathan Schneider (Georgetown University)
TransformerLarge Language ModelText
🎯 What it does: Studied pretraining GPT‑2 on multiple first languages (L1), then continuing pretraining on a second language (L2, i.e., English) to build an L2 language model (L2LM) that simulates second language acquisition, and used these models to evaluate their similarity to humans in sentence processing (word-level prediction).
Modeling User Preferences with Automatic Metrics: Creating a High-Quality Preference Dataset for Machine Translation
Sweta Agrawal, Andre Martins
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes and constructs the MT-PREF translation preference dataset, combining the automatic quality evaluation metric XCOMET-XL+XXL with human ratings from professional translators to automatically induce translation preference relationships. The dataset is used to perform preference learning (CPO, DPO, etc.) on the TOWER decoder-only large language model (LLM), thereby improving machine translation quality.
MoDULA: Mixture of Domain-Specific and Universal LoRA for Multi-Task Learning
Yufei Ma (Northwestern Polytechnical University), Libin Yang (Northwestern Polytechnical University)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsTextFinance Related
🎯 What it does: Proposes the MoDULA framework, which employs a three-stage training Mixture of Experts (MoE) parameter-efficient fine-tuning method, combining general experts and domain-specific experts, and provides two variants: MoDULA-Flan and MoDULA-Res.
Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaboration
Shangbin Feng (University of Washington), Yulia Tsvetkov (University of Washington)
Large Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Proposes the MODULAR PLURALISM framework, achieving multi-dimensional diversified alignment by collaborating large LLMs with multiple specialized community small LLMs.
MolTRES: Improving Chemical Language Representation Learning for Molecular Property Prediction
Jun-Hyung Park (Hankuk University of Foreign Studies), SangKeun Lee (Korea University)
Representation LearningDrug DiscoveryTransformerGenerative Adversarial NetworkTextBiomedical Data
🎯 What it does: Propose the MolTRES framework, which trains SMILES Transformers using generator-discriminator training and substructure masking pre-training, and integrates scientific literature embeddings to enhance molecular property prediction.
Moral Foundations of Large Language Models
Marwa Abdulhai, Natasha Jaques (University of Cambridge)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: This study evaluates the ethical tendencies of popular large language models (LLMs) using Moral Foundations Theory (MFT), exploring the cultural/political biases, context consistency, manipulability, and impact on downstream tasks (e.g., donation decisions) introduced by their training data.
More DWUGs: Extending and Evaluating Word Usage Graph Datasets in Multiple Languages
Dominik Schlechtweg, Nina Tahmasebi (University of Stuttgart)
Data-Centric LearningGraph Neural NetworkGraphTabularBenchmark
🎯 What it does: Expand and verify the largest Word Usage Graph (WUG) dataset from SemEval-2020 Task 1: Add two additional rounds of annotation (totaling six rounds) on DWUG (English, German, Swedish), further intensifying the graph structure; supplement annotations for DiscoWUG; conduct comparative experiments on DWUG DE Sense; verify reproducibility through resampling and re-annotation.
More Insightful Feedback for Tutoring: Enhancing Generation Mechanisms and Automatic Evaluation
Wencke Liermann (Electronics and Telecommunications Research Institute), Kong Joo Lee (Chungnam National University)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study proposes two training extensions based on generating personalized student feedback: ① utilizing KL regularization to internalize key sentence information, ② adopting direct preference optimization (DPO) based on NLI to reduce entailment between student answers and generated feedback.
More Than Catastrophic Forgetting: Integrating General Capabilities For Domain-Specific LLMs
Chengyuan Liu (Zhejiang University), Fei Wu (Zhejiang University)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The paper proposes the General Capabilities Integration (GCI) framework, constructs three practical tasks and datasets in the legal domain, and introduces the ALoRA adapter, which combines multi-head attention with LoRA to achieve synergy between general capabilities and domain knowledge;
MORPHEUS: Modeling Role from Personalized Dialogue History by Exploring and Utilizing Latent Space
Yihong Tang, Yuexian Hou (China Mobile Communication Group Tianjin Co., Ltd.)
GenerationTransformerAuto EncoderContrastive LearningText
🎯 What it does: A three-stage training framework named MORPHEUS is constructed to learn role information from dialogue history in a latent space, enabling personalized response generation without external role data.
MOSEL: 950,000 Hours of Speech Data for Open-Source Speech Foundation Model Training on EU Languages
Marco Gaido (Fondazione Bruno Kessler), Matteo Negri (Fondazione Bruno Kessler)
RecognitionData-Centric LearningTransformerLarge Language ModelAudio
🎯 What it does: Constructed an open-source compliance voice dataset (MOSEL) covering 24 official EU languages with a total of 950k hours, and automatically generated pseudo-labels for 441k hours of unannotated data within it;
MOSEL: Inference Serving Using Dynamic Modality Selection
Bodun Hu (University of Texas at Austin), Aditya Akella (University of Texas at Austin)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Designed and implemented MOSEL, an automated multi-modal inference service system that improves throughput and resource utilization by dynamically selecting input modalities while meeting user-defined accuracy and latency SLOs.
MP2D: An Automated Topic Shift Dialogue Generation Framework Leveraging Knowledge Graphs
Yerin Hwang (Seoul National University), Kyomin Jung (Seoul National University)
GenerationData SynthesisLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed the MP2D framework, which can automatically generate dialog data containing natural topic transitions.
MQuinE: a Cure for “Z-paradox” in Knowledge Graph Embedding
Yang Liu (Baidu Research), Mingming Sun (Beijing Institute of Mathematical Sciences and Applications)
Representation LearningGraph
🎯 What it does: Research and address the Z-paradox problem existing in knowledge graph embedding models, proposing a new model called MQuinE;
MSI-Agent: Incorporating Multi-Scale Insight into Embodied Agents for Superior Planning and Decision-Making
Dayuan Fu (Tsinghua University), Bowen Zhou (Tsinghua University)
Robotic IntelligenceTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes MSI-Agent, an approach that enhances the planning and decision-making capabilities of LLM-based embodied agents by generating, filtering, and utilizing long-term memory through multi-scale insights (general, subtask, environment).
MT-Eval: A Multi-Turn Capabilities Evaluation Benchmark for Large Language Models
Wai-Chung Kwan (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Built and released an evaluation benchmark called MT-Eval for multi-turn dialogue capabilities, covering four types of dialogue patterns (recollection, expansion, refinement, follow-up), and evaluated 10 mainstream LLMs.
MTA4DPR: Multi-Teaching-Assistants Based Iterative Knowledge Distillation for Dense Passage Retrieval
Qixi Lu (Beijing Language and Culture University), Gongbo Tang (Beijing Language and Culture University)
RetrievalKnowledge DistillationTransformerContrastive LearningText
🎯 What it does: Proposed and implemented the MTA4DPR framework, a multi-assistant iterative knowledge distillation method that jointly trains a dense retrieval student model using a teacher model and multiple assistant models.
MTLS: Making Texts into Linguistic Symbols
Wenlong Fei (HeFei University of Technology), Hongbo Li (HeFei University of Technology)
Representation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningImageText
🎯 What it does: By rendering text into glyphs (pixel maps) and constructing symbol embeddings (SSS embedding), then jointly pre-training with pre-trained language models (BERT, RoBERTa), replacing the original token embedding layer to enhance the model's multilingual capabilities.
Multi-Dialect Vietnamese: Task, Dataset, Baseline Models and Challenges
Nguyen Van Dinh (University of Information Technology), Kiet Van Nguyen (University of Information Technology)
ClassificationRecognitionTransformerSupervised Fine-TuningBenchmarkAudio
🎯 What it does: This study proposes the ViMD corpus, covering 63 provincial dialects in Vietnam, and conducts two benchmark experiments on dialect identification and speech recognition based on this dataset.
Multi-expert Prompting Improves Reliability, Safety and Usefulness of Large Language Models
Do Xuan Long (National University of Singapore), Nancy F. Chen (National University of Singapore)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelPrompt EngineeringMixture of ExpertsTextChain-of-Thought
🎯 What it does: Propose the Multi-expert Prompting method, which simulates multiple experts answering input instructions and aggregates these responses through seven sub-tasks within a single round to enhance the reliability, safety, and usefulness of large language models.
Multi-Granularity History and Entity Similarity Learning for Temporal Knowledge Graph Reasoning
Shi Mingcong (Soochow University), Li Qing (NingboTech University)
Representation LearningRecurrent Neural NetworkGraph Neural NetworkGraphTime SeriesBenchmark
🎯 What it does: This paper proposes the MGESL model, which jointly utilizes entity similarity, coarse-grained and fine-grained historical information for temporal knowledge graph reasoning.
Multi-Level Cross-Modal Alignment for Speech Relation Extraction
Liang Zhang (Xiamen University), Jinsong Su (Xiamen University)
TransformerContrastive LearningMultimodalityAudio
🎯 What it does: Proposed a multi-layer cross-modal alignment model (MCAM) for extracting relation triplets from real speech.
Multi-Level Information Retrieval Augmented Generation for Knowledge-based Visual Question Answering
Omar Adjali (Université Paris-Saclay CEA, List), Hervé Le Borgne (Université Paris-Saclay CEA, List)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a multi-layer information retrieval augmented generation framework, MiRAG, for knowledge-based visual question answering tasks.
Multi-LogiEval: Towards Evaluating Multi-Step Logical Reasoning Ability of Large Language Models
Nisarg Patel (Arizona State University), Chitta Baral (Arizona State University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a multi-step logical reasoning evaluation dataset named Multi-LogiEval, covering propositional logic, predicate logic, and non-monotonic logic, containing 33 inference rules and their combinations, with a maximum of 5-step reasoning depth, and evaluated the reasoning capabilities of various large language models using zero-shot chain-of-thought (CoT) assessment.
Multi-News+: Cost-efficient Dataset Cleansing via LLM-based Data Annotation
Juhwan Choi (Chung-Ang University), YoungBin Kim (Chung-Ang University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Utilize large language models (LLM) through chain-of-thought and majority voting mechanisms to automatically identify and remove irrelevant or noisy documents from the Multi-News dataset, generating the cleaned MULTI-NEWS+ dataset.
Multi-pass Decoding for Grammatical Error Correction
Xiaoying Wang (Zhengzhou University), Hongfei Xu (Zhengzhou University)
GenerationComputational EfficiencyTransformerText
🎯 What it does: Propose multiple decoding (MPD) combined with a lightweight early stopping mechanism and source information fusion to enhance the performance of grammar error correction (GEC) based on seq2seq models.
Multilingual Topic Classification in X: Dataset and Analysis
Dimosthenis Antypas, Jose Camacho-Collados (Snap Inc)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Introduces the X-Topic dataset, containing 4,000 multi-label tweets in English, Spanish, Japanese, and Greek, for multilingual topic classification research.
Multimodal Clickbait Detection by De-confounding Biases Using Causal Representation Inference
Jianxing Yu (Sun Yat-sen University), Yanghui Rao (Sun Yat-sen University)
ClassificationRepresentation LearningContrastive LearningMultimodality
🎯 What it does: This paper proposes a multimodal clickbait detection framework based on causal representation inference, achieving more robust clickbait detection by separating invariant factors, scenario-specific causal factors, and noise factors.
Multimodal Self-Instruct: Synthetic Abstract Image and Visual Reasoning Instruction Using Language Model
Wenqi Zhang (Zhejiang University), Yueting Zhuang (Zhejiang University)
Data SynthesisLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper combines large language models with code generation to propose a multi-modal self-instruct process, automatically synthesizing abstract images (such as charts, maps, flowcharts, relationship diagrams, dashboards, etc.) and generating corresponding high-quality visual reasoning questions; subsequently, these synthetic data are used to fine-tune LMMs, significantly improving their performance in abstract image understanding and spatial reasoning; furthermore, a benchmark dataset consisting of 11,193 multi-modal instructions is constructed based on this process, systematically evaluating various mainstream LMMs.
Multiple Sources are Better Than One: Incorporating External Knowledge in Low-Resource Glossing
Changbing Yang (University of British Columbia), Miikka Silfverberg (University of British Columbia)
Representation LearningData-Centric LearningRecurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: The study proposes an automated glossing system that integrates sentence-level and word-level translations, external dictionaries, and large language model (LLM) reasoning to address the problem of data scarcity in low-resource languages.
MuMath-Code: Combining Tool-Use Large Language Models with Multi-perspective Data Augmentation for Mathematical Reasoning
Shuo Yin (Ocean University of China), Jinfeng Bai (Ocean University of China)
Data SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Integrate LLMs with tool usage and multi-perspective data augmentation to construct the MuMath-Code dataset, and enhance the mathematical reasoning ability of open-source LLMs through two-stage fine-tuning (first reinforcing Chain-of-Thought (CoT) reasoning, then incorporating code execution).
Muting Whisper: A Universal Acoustic Adversarial Attack on Speech Foundation Models
Vyas Raina (University of Cambridge), Mark Gales (University of Cambridge)
RecognitionAdversarial AttackTransformerAudio
🎯 What it does: Propose a method that prepends a 0.64-second silent adversarial audio segment to any speech, prompting the Whisper ASR model to directly output a special token without recognizing the speech content, thereby achieving a 'silence' attack.
Nash CoT: Multi-Path Inference with Preference Equilibrium
Ziqi Zhang (Westlake University), Donglin Wang (Westlake University)
Large Language ModelTextChain-of-Thought
🎯 What it does: Propose the Nash CoT (Nash Chain of Thought) method, which reduces the number of reasoning paths in multi-path reasoning while maintaining or improving reasoning accuracy by introducing role-playing and a game mechanism between role-playing LLMs and regular LLMs.
Nearest Neighbor Normalization Improves Multimodal Retrieval
Neil Chowdhury (Massachusetts Institute of Technology), Tristan Thrush (Stanford University)
RetrievalContrastive LearningMultimodality
🎯 What it does: Propose a post-training nearest neighbor normalization (NNN) method that does not require additional learning, aimed at improving contrastive learning-based multi-modal retrieval models
Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent
Xiaoyan Yu (Beijing Institute of Technology), Liehuang Zhu (Beijing Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposes the Neeko framework, enabling efficient learning and instant switching for multi-role dialogue agents
Neuron Specialization: Leveraging Intrinsic Task Modularity for Multilingual Machine Translation
Shaomu Tan (University of Amsterdam), Christof Monz (University of Amsterdam)
TransformerTextBenchmark
🎯 What it does: By analyzing the activation patterns of neurons in the feed-forward layer of multilingual translation models, the Neuron Specialization method is proposed to alleviate multilingual interference and enhance knowledge transfer.