ACL 2025 Papers — Page 10
Annual Meeting of the Association for Computational Linguistics · 1699 papers
Less, but Better: Efficient Multilingual Expansion for LLMs via Layer-wise Mixture-of-Experts
Xue Zhang (Beijing Jiaotong University), Jie Zhou (Tencent Inc)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: By analyzing the language similarity of hidden states across different layers, we propose LayerMoE, which allocates varying numbers of MoE experts per layer and adds classifiers before high-similarity layers, achieving efficient expansion for new languages while maintaining the performance of the original language;
LETS-C: Leveraging Text Embedding for Time Series Classification
Rachneet Kaur (J.P. Morgan), Manuela Veloso (J.P. Morgan)
ClassificationRepresentation LearningConvolutional Neural NetworkLarge Language ModelTextTime SeriesBenchmark
🎯 What it does: Convert time series into vectors using pre-trained text embedding models, then perform temporal classification using a lightweight CNN-MLP classification head.
Leveraging Dual Process Theory in Language Agent Framework for Real-time Simultaneous Human-AI Collaboration
Shao Zhang (Shanghai Jiao Tong University), Ying Wen (Shanghai Jiao Tong University)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: Propose a language agent framework named DPT-Agent that integrates dual process theory to achieve real-time parallel human-machine collaboration.
Leveraging Human Production-Interpretation Asymmetries to Test LLM Cognitive Plausibility
Suet-Ying Lam (UMass Amherst), Rob Voigt (Northwestern University)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringText
🎯 What it does: Explores the production-explanation asymmetry in LLMs during sentence generation and explanation, conducting experiments on pronoun generation and explanation deviations using implicit causal verbs
Leveraging In-Context Learning for Political Bias Testing of LLMs
Patrick Haller (University of Zurich), Lena Ann Jäger (University of Zurich)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Introduce the Questionnaire Modeling task, using human questionnaire answers as context to evaluate political bias in LLMs.
Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs
Ananth Muppidi (IIIT Hyderabad), Sambaran Bandyopadhyay (Adobe Research)
ClassificationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposed an input-adaptive soft prompt framework called ID-SPAM, which generates soft prompts using self-attention and inserts them into a single layer, achieving parameter-efficient fine-tuning and zero-shot task/domain transfer.
LexCLiPR: Cross-Lingual Paragraph Retrieval from Legal Judgments
Rohit Upadhya (Technical University of Munich), Santosh T.y.s.s
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed the LexCLiPR cross-lingual legal case paragraph retrieval dataset, and evaluated multiple retrieval models under zero-shot and fine-tuned settings.
LexGen: Domain-aware Multilingual Lexicon Generation
Ayush Maheshwari (NVIDIA), Ganesh Ramakrishnan (Indian Institute of Technology Bombay)
GenerationData SynthesisTransformerLarge Language ModelTextBenchmark
🎯 What it does: Generate dictionary entries for multiple domains and six Indian languages to address the scarcity of specialized vocabulary.
Lexical Diversity-aware Relevance Assessment for Retrieval-Augmented Generation
Zhange Zhang (Beihang University), Xianglong Liu (Beihang University)
GenerationRetrievalTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: The paper proposes a vocabulary diversity-aware retrieval-augmented generation framework called DRAG, which enhances the factual accuracy of large language models (LLMs) by leveraging fine-grained query splitting, relevance assessment, and risk-guided sparse correction.
Lexical Recall or Logical Reasoning: Probing the Limits of Reasoning Abilities in Large Language Models
Henrike Beyer (University of Dundee), Chris Reed (University of Dundee)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed and evaluated a logic grid puzzle benchmark named Mystery-Zebra, containing 4,290 puzzles, designed to test LLMs' reasoning capabilities under lexical interference and formal difficulty levels.
LexKeyPlan: Planning with Keyphrases and Retrieval Augmentation for Legal Text Generation: A Case Study on European Court of Human Rights Cases
Santosh T.y.s.s, Elvin Quero Hernandez (Technical University of Munich)
GenerationRetrievalTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose a three-stage legal text generation framework called LexKeyPlan, which first generates legal concept key phrases as a prospective content plan, then uses this plan to retrieve external documents, and finally generates legal reasoning text.
LexTempus: Enhancing Temporal Generalizability of Legal Language Models Through Dynamic Mixture of Experts
Santosh T.y.s.s, Tuan-Quang Vuong (Technical University of Munich)
Domain AdaptationTransformerMixture of ExpertsText
🎯 What it does: Propose the LexTempus dynamic expert mixture model for adaptive learning with temporal generalization in legal language models.
Library-Like Behavior In Language Models is Enhanced by Self-Referencing Causal Cycles
Munachiso S Nwadike, Kentaro Inui (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Proposed and verified the self-referential causal loop (RECALL) mechanism, utilizing cycle tokens to enable large language models (LLMs) to transcend sequential limitations and overcome the 'reversal curse'.
LIFBench: Evaluating the Instruction Following Performance and Stability of Large Language Models in Long-Context Scenarios
Xiaodong Wu (East China Normal University), Wei Zhang (East China Normal University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes two evaluation frameworks, LIFBENCH and LIFEVAL, for systematically assessing the instruction-following ability and stability of large language models (LLMs) in long-text scenarios, with extensive experiments conducted on 20 mainstream LLMs.
Limited Generalizability in Argument Mining: State-Of-The-Art Models Learn Datasets, Not Arguments
Marc Feger (Heinrich-Heine-University Düsseldorf), Stefan Dietze (GESIS - Leibniz Institute for the Social Sciences)
ClassificationAdversarial AttackTransformerContrastive LearningTextBenchmark
🎯 What it does: This paper conducts a large-scale re-evaluation of 17 publicly available sentence-level datasets in the argument mining field, testing the generalization capability of current models such as BERT across different datasets.
Limited-Resource Adapters Are Regularizers, Not Linguists
Marcell Fekete (Aalborg University), Heather Lent (Aalborg University)
Knowledge DistillationTransformerSupervised Fine-TuningText
🎯 What it does: This paper combines adapter 'souping' (weight averaging) with cross-attention fine-tuning (CA-FT) in low-resource Creole translation tasks, and demonstrates that adapters primarily serve a regularization role rather than conveying linguistic information in this context.
Linguistic Generalizability of Test-Time Scaling in Mathematical Reasoning
Guijin Son (Yonsei University), James Thorne (KAIST AI)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper investigates the language generalization of three test-time scaling methods (Outcome Reward Modeling, Process Reward Modeling, Budget Forcing) on competition-level math reasoning tasks across 55 languages, and constructs and releases the multilingual competition-level math benchmark MCLM as well as the multilingual reasoning LLM MR1-1.5B.
Literary Evidence Retrieval via Long-Context Language Models
Katherine Thai (UMass Amherst), Mohit Iyyer (University of Maryland)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Explored how to leverage long-context large language models (LLMs) to complete literary evidence retrieval tasks, i.e., automatically generating missing citations given complete novel texts and missing literary criticism fragments.
Literature Meets Data: A Synergistic Approach to Hypothesis Generation
Haokun Liu (University of Chicago), Chenhao Tan (University of Chicago)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposes an LLM-driven hypothesis generation framework (HYPOREFINE) that combines literature knowledge with data-driven methods, achieving more generalizable and interpretable hypothesis generation;
Llama See, Llama Do: A Mechanistic Perspective on Contextual Entrainment and Distraction in LLMs
Jingcheng Niu (University of Toronto), Amir H. Abdi (Microsoft)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study identifies and systematizes the 'contextual forcing' phenomenon, where language models assign higher logits/probabilities to any word that appears in the context, regardless of its relevance; and proposes a distinguishable 'forcing head' mechanism, proving that disabling these heads significantly reduces interference.
LLaMA-Omni 2: LLM-based Real-time Spoken Chatbot with Autoregressive Streaming Speech Synthesis
Qingkai Fang (Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences), Yang Feng (Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences)
GenerationData SynthesisTransformerLarge Language ModelFlow-based ModelTextAudio
🎯 What it does: Propose the LLaMA-Omni 2 speech-language model to enable real-time high-quality voice interaction.
LlamaDuo: LLMOps Pipeline for Seamless Migration from Service LLMs to Small-Scale Local LLMs
Chansung Park (Electronics and Telecommunications Research Institute), Jing Tang (Hong Kong University of Science and Technology)
Data SynthesisComputational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed LlamaDuo, an automated LLMOps workflow for transferring knowledge from cloud-based service LLMs to smaller, locally deployable LLMs, ensuring service continuity in offline or restricted environments.
LLaMAs Have Feelings Too: Unveiling Sentiment and Emotion Representations in LLaMA Models Through Probing
Dario Di Palma (Politecnico di Bari), Tommaso Di Noia (Politecnico di Bari)
ClassificationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Probe the hidden layers of Llama 3.2-1B/3B/8B models to study the hierarchical distribution of sentiment (positive/negative polarity) and emotion (happiness, sadness, anger, fear, love, surprise) information, and based on this, propose SENTRILLAMA by selecting the most representative layers and replacing the language modeling head to build a lightweight sentiment classification model.
LLäMmlein: Transparent, Compact and Competitive German-Only Language Models from Scratch
Jan Pfister (Julius-Maximilians-Universität Würzburg), Andreas Hotho (Julius-Maximilians-Universität Würzburg)
GenerationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Train and publicly release two full German decoder language models, LLäMmlein 120M and 1B, from scratch, providing complete training data, code, distributed training details, iterative checkpoints, and system evaluations, aiming to establish a reproducible and transparent baseline for the German NLP research community.
LLaSE-G1: Incentivizing Generalization Capability for LLaMA-based Speech Enhancement
Boyi Kang (Northwestern Polytechnical University), Lei Xie (Northwestern Polytechnical University)
RestorationTransformerLarge Language ModelAudio
🎯 What it does: Built a general-purpose speech enhancement model LLaSE-G1 based on LLaMA, which uniformly handles tasks such as noise suppression, packet loss compensation, target speaker extraction, echo cancellation, and speech separation.
LLaVA Steering: Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steering
Jinhe Bi (Ludwig Maximilian University of Munich), Yunpu Ma (Ludwig Maximilian University of Munich)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes the Modality Linear Representation-Steering (MoReS) method, which rebalances visual and textual modalities during the visual instruction tuning phase by applying linear subspace mapping to visual representations, and builds the LLaVA Steering model along with a visualization evaluation platform based on this.
LLM Agents Making Agent Tools
Georg Wölflein (EKFZ for Digital Health TU Dresden), Jakob Nikolas Kather (EKFZ for Digital Health TU Dresden)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This study proposes the TOOLMAKER framework, which can automatically convert GitHub code repositories corresponding to scientific papers into tools callable by large language models (LLMs), and evaluates its performance through the TM-BENCH benchmark.
LLM as a Broken Telephone: Iterative Generation Distorts Information
Amr Mohamed (Mohamed bin Zayed University of Artificial Intelligence), Guokan Shang (Mohamed bin Zayed University of Artificial Intelligence)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigates the information distortion issue in large language models during iterative generation processes (such as consecutive translation and rewriting), constructing multiple chain experiments to verify distortion mechanisms and mitigation strategies.
LLM as Entity Disambiguator for Biomedical Entity-Linking
Christophe Ye (Georgia Institute of Technology), Cassie S. Mitchell (Georgia Institute of Technology)
RetrievalTransformerLarge Language ModelBiomedical DataRetrieval-Augmented Generation
🎯 What it does: In the biomedical entity linking task, this paper proposes using a large language model (LLM) as the core entity disambiguator, directly embedding it into the subsequent disambiguation steps after candidate generation, thereby improving linking accuracy without any fine-tuning.
LLM Braces: Straightening Out LLM Predictions with Relevant Sub-Updates
Ying Shen (University of Illinois Urbana-Champaign), Lifu Huang (University of California Davis)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the LLMBRACES method, which dynamically adjusts the contribution of sub-updates in the Transformer's Feed-Forward layer by calculating scores related to input or target attributes to enhance prediction accuracy and controllability.
LLM Meets Scene Graph: Can Large Language Models Understand and Generate Scene Graphs? A Benchmark and Empirical Study
Dongil Yang (Yonsei University), Jinyoung Yeo (Yonsei University)
GenerationData SynthesisTransformerLarge Language ModelTextGraphBenchmarkChain-of-Thought
🎯 What it does: Proposes a scene graph benchmark (TSG Bench) tailored for large language models (LLMs) to evaluate their capabilities in two major tasks: scene graph understanding and generation.
LLM-based Rumor Detection via Influence Guided Sample Selection and Game-based Perspective Analysis
Zhiliang Tian (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a rumor detection method based on large language models (LLM), leveraging influence-guided sample selection and multi-perspective collaborative game theory to enhance detection performance.
LLM-Guided Semantic-Aware Clustering for Topic Modeling
Jianghan Liu (Southeast University), Yining Li (Southeast University)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes LiSA, a topic model that combines large language models (LLMs) with clustering methods. It first generates candidate topic words for each document using LLMs, then performs clustering on both documents and candidate topic words. It utilizes LLMs for mapping and validation, and aligns the document-level and topic-level semantic spaces globally through a collaborative enhancement phase, ultimately achieving more accurate topic distributions.
LLM-Powered Test Case Generation for Detecting Bugs in Plausible Programs
Kaibo Liu (Peking University), Gang Huang (Peking University)
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes TrickCatcher, a test case generation method based on large language models, for detecting tricky bugs in trustworthy programs.
LLM\timesMapReduce: Simplified Long-Sequence Processing using Large Language Models
Zihan Zhou (Xiamen University), Maosong Sun (Tsinghua University)
Data SynthesisLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Designed a training-agnostic chunked MapReduce framework, LLM×MapReduce, enabling short-context LLMs to process long texts, and constructed the Pyramid-Align dataset based on this framework;
LLMs + Persona-Plug = Personalized LLMs
Jiongnan Liu (Renmin University of China), Zhicheng Dou (Renmin University of China)
Recommendation SystemComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a lightweight, pluggable personalized LLM method called PPlug, achieving parameter-free personalized generation by encoding user historical behavior into a single user embedding and appending it to the LLM input.
LLMs can be easily Confused by Instructional Distractions
Yerin Hwang (Seoul National University), Kyomin Jung (Seoul National University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes a new benchmark called DIM-Bench to evaluate the instruction-following ability of large language models in 'instructional distraction' scenarios;
LLMs can Perform Multi-Dimensional Analytic Writing Assessments: A Case Study of L2 Graduate-Level Academic English Writing
Zhengxiang Wang (Stony Brook University), Owen Rambow (Stony Brook University)
ClassificationGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigated the performance of large language models in multidimensional analytical writing assessment, using academic review texts written by L2 graduate students for scoring and comment generation;
LLMs Can Simulate Standardized Patients via Agent Coevolution
Zhuoyun Du (Zhejiang University), Haochao Ying (Sun Yat-sen University)
TransformerLarge Language ModelAgentic AITextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: This paper proposes EvoPatient, a multi-agent co-evolution framework that automatically simulates and records dialogues between patient and doctor agents generated by LLMs during simulated medical consultations. Through self-supervised mechanisms, it continuously enhances the standardization of patient presentations and the professionalism of doctors' questioning, achieving standardized patient (SP) simulation and physician training without human supervision.
LLMs Caught in the Crossfire: Malware Requests and Jailbreak Challenges
Haoyang Li (Institute of Artificial Intelligence (TeleAI) China Telecom), Xuelong Li (Institute of Artificial Intelligence (TeleAI) China Telecom)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This study constructs the MalwareBench benchmark dataset and evaluates the security performance of 29 mainstream large language models (LLMs) under malicious code generation and black-box jailbreak attacks.
LLMs instead of Human Judges? A Large Scale Empirical Study across 20 NLP Evaluation Tasks
Anna Bavaresco (University of Amsterdam), Alberto Testoni (Amsterdam UMC)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study conducted a large-scale empirical analysis of the judgment results of 11 large language models across 20 public NLP evaluation datasets, and released the scalable JUDGE-BENCH benchmark.
LLMs know their vulnerabilities: Uncover Safety Gaps through Natural Distribution Shifts
Qibing Ren (Shanghai Jiao Tong University), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the ActorBreaker method, which constructs multi-layer actor networks based on actor network theory, mines actors related to harmful prompts from pre-trained knowledge, and detects safety vulnerabilities in aligned LLMs through self-generated multi-round natural distribution shift attacks.
LLMs syntactically adapt their language use to their conversational partner
Florian Kandra (Saarland University), Alexander Koller (Saarland University)
GenerationTransformerLarge Language ModelText
🎯 What it does: Investigated whether large language models (LLMs) adapt syntactically in dialogues as humans do, constructing a custom LLM dialogue corpus and measuring repetition and consistency of syntactic structures.
LLMs Trust Humans More, That’s a Problem! Unveiling and Mitigating the Authority Bias in Retrieval-Augmented Generation
Yuxuan Li (Southern University of Science and Technology), Xuetao Wei (Southern University of Science and Technology)
Explainability and InterpretabilityTransformerTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper investigates the Authority Bias in large language models within Retrieval-Augmented Generation (RAG) systems, where models tend to trust user-provided information over knowledge in retrieval databases even when the user input contradicts factual information.
LocAgent: Graph-Guided LLM Agents for Code Localization
Zhaoling Chen (Yale University), Xingyao Wang (All Hands AI)
AI Code AssistantGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: An efficient code localization framework called LOCAGENT is proposed by constructing a heterogeneous directed graph of the codebase and enabling LLM agents to perform multi-hop reasoning.
Locate-and-Focus: Enhancing Terminology Translation in Speech Language Models
Suhang Wu (Xiamen University), Jinsong Su (Xiamen University)
TransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Propose a Locate-and-Focus method that first locates the audio segments corresponding to terms in speech, then improves the term translation accuracy in speech translation by making the model focus on external translation knowledge through audio replacement and label prompts.
Logic-Regularized Verifier Elicits Reasoning from LLMs
Xinyu Wang (East China Normal University), Xuelong Li (China Telecom)
Explainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextChain-of-Thought
🎯 What it does: Propose an unsupervised validator called LOVER that leverages internal activations of LLMs and evaluates the correctness of reasoning paths through logical constraints.
Logical forms complement probability in understanding language model (and human) performance
Yixuan Wang (University of Chicago), Freda Shi (University of Waterloo)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Systematically evaluate the reasoning capabilities of large language models (LLMs) in propositional and modal logic, focusing on hypothetical syllogisms and disjunctive syllogisms (e.g., modus ponens, modus tollens), and compare performance differences between LLMs and humans across various logical forms and modalities using human experimental data.
LogicPro: Improving Complex Logical Reasoning via Program-Guided Learning
Jin Jiang (Peking University), Liangcai Gao (Peking University)
Data SynthesisExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose the LogicPro method, which generates high-quality, scalable complex logical reasoning data by leveraging LeetCode algorithm problems and their program solutions, ultimately synthesizing 540K text reasoning samples;
LoGU: Long-form Generation with Uncertainty Expressions
Ruihan Yang (Fudan University), Deqing Yang (Fudan University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Investigated how large language models express uncertainty in long-text generation and proposed the LoGU (Long-form Generation with Uncertainty) task;
LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks
Yushi Bai (Tsinghua University), Juanzi Li (Tsinghua University)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose LongBench v2, a multi-task benchmark for long-text understanding and reasoning, containing 503 questions with lengths ranging from 8k to 2M words, covering six major task categories.
LongDocURL: a Comprehensive Multimodal Long Document Benchmark Integrating Understanding, Reasoning, and Locating
Chao Deng (MAIS, Institute of Automation, Chinese Academy of Sciences), Cheng-Lin Liu (Taobao & Tmall Group of Alibaba)
TransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the LongDocURL long-document multimodal question-answering benchmark, covering three tasks (understanding, reasoning, localization) with 20 subtasks, collecting 2,325 QA pairs across 33,000 pages of documents.
LongRecipe: Recipe for Efficient Long Context Generalization in Large Language Models
Zhiyuan Hu (National University Of Singapore), Bryan Hooi (Nanjing University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes the LongRecipe method, which utilizes influence word analysis, position index transformation, and training optimization strategies to efficiently extend the long context window of large language models while maintaining their original general capabilities.
LongReD: Mitigating Short-Text Degradation of Long-Context Large Language Models via Restoration Distillation
Zican Dong (Renmin University of China), Weipeng Chen (Baichuan Inc)
Knowledge DistillationTransformerLarge Language ModelText
🎯 What it does: To address the issue of performance degradation on short text tasks caused by expanding the context window in large language models, this paper proposes a new training framework called LongReD, aiming to simultaneously enhance long text processing capabilities while maintaining short text performance.
LongReward: Improving Long-context Large Language Models with AI Feedback
Jiajie Zhang (Tsinghua University), Juanzi Li (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose the LongReward method, which uses a pre-trained LLM as a judge to score four dimensions of the long context model's answers, and employs DPO for reinforcement learning to enhance long-text capabilities.
LongSafety: Evaluating Long-Context Safety of Large Language Models
Yida Lu (Tsinghua University), Minlie Huang (Tsinghua University)
Safty and PrivacyTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes the LONGSAFETY benchmark to evaluate the safety of large language models in open-ended long-context tasks.
Look Both Ways and No Sink: Converting LLMs into Text Encoders without Training
Ziyong Lin (State Key Laboratory of General Artificial Intelligence, BIGAI), Zixia Jia (State Key Laboratory of General Artificial Intelligence, BIGAI)
Representation LearningTransformerLarge Language ModelTextFinance Related
🎯 What it does: Without additional training, pre-trained large language models are converted into high-performance text encoders by modifying the attention mask and masking the first token.
Lost in Literalism: How Supervised Training Shapes Translationese in LLMs
Yafu Li (Shanghai AI Laboratory), Yue Zhang (Westlake University)
GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Systematically evaluate over-literal translation (translationese) in LLM translations and investigate its root causes from the supervised fine-tuning (SFT) stage.
Lost in Multilinguality: Dissecting Cross-lingual Factual Inconsistency in Transformer Language Models
Mingyang Wang (Bosch Center for Artificial Intelligence), Hinrich Schuetze
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Investigate the phenomenon of factual inconsistencies in multilingual language models when answering semantic equality prompts, and analyze their internal causes using mechanism interpretability methods; propose a linear shortcut to skip the final layer's language conversion errors to improve cross-lingual consistency and accuracy.
Lost in the Context: Insufficient and Distracted Attention to Contexts in Preference Modeling
Shihan Dou (Fudan University), Xuanjing Huang (Fudan University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper investigates the attention issue in RLHF where reward models ignore context, proposing the AttnRM framework to improve reward model performance by strengthening attention on key context fragments.
Low-Bit Quantization Favors Undertrained LLMs
Xu Ouyang (University of Virginia), Dong Yu (Tencent AI Lab)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies the impact of low-bit quantization on large language models, finding that low-bit quantization is more suitable for under-trained models, and derives scaling laws of quantization loss by analyzing over 1500 quantization checkpoints.
LPOI: Listwise Preference Optimization for Vision Language Models
Fatemeh Pesaran Zadeh (Seoul National University), Gunhee Kim (Seoul National University)
Object DetectionOptimizationSupervised Fine-TuningReinforcement LearningPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Propose a list-based ranking preference optimization method (LPOI), which automatically generates obstructed samples of key objects and interpolates to construct sorted lists, training vision-language models to reduce hallucinations and improve alignment with visual information.
LSSF: Safety Alignment for Large Language Models through Low-Rank Safety Subspace Fusion
Guanghao Zhou (East China Normal University), Jun Zhou (Ant Group)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a post-hoc safety realignment framework LSSF, which restores compromised safety after fine-tuning through low-rank safety subspace fusion.
M-MAD: Multidimensional Multi-Agent Debate for Advanced Machine Translation Evaluation
Zhaopeng Feng (Zhejiang University), Zuozhu Liu (Zhejiang University)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose the M-MAD framework to realize multi-dimensional multi-agent debate-based LLM-as-a-judge machine translation evaluation.
M-RewardBench: Evaluating Reward Models in Multilingual Settings
Srishti Gureja (Cohere Labs), Marzieh Fadaee (Cohere Labs)
Reinforcement Learning from Human FeedbackReinforcement LearningTextBenchmark
🎯 What it does: Constructed a multilingual reward model evaluation benchmark called M-REWARDBENCH, covering 23 languages and tasks including chat, safety, reasoning, and translation, and conducted systematic evaluations of multiple existing reward models.
M2RC-EVAL: Massively Multilingual Repository-level Code Completion Evaluation
Jiaheng Liu (Nanjing University), Bo Zheng (Alibaba Group)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a repository-level code completion benchmark, MRC-EVAL 2, covering 18 mainstream programming languages, providing two fine-grained annotations (bucket-level and semantic-level), while organizing a multilingual instruction corpus, MRC-INSTRUCT 2, to enhance the completion capabilities of existing code LLMs.
M2S: Multi-turn to Single-turn jailbreak in Red Teaming for LLMs
Junwoo Ha (AIM Intelligence), Suhyun Kim (Kyung Hee University)
Safty and PrivacyComputational EfficiencyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a framework (M2S) that converts multi-turn conversational jailbreaks into single-round structured prompts, reducing human effort and improving attack success rates.
M³GQA: A Multi-Entity Multi-Hop Multi-Setting Graph Question Answering Benchmark
Boci Peng (Peking University), Yan Zhang (Peking University)
TransformerLarge Language ModelGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and constructed M3GQA, a multi-entity, multi-hop, multi-setting graph question answering benchmark containing diverse queries, answers, and semantically correct reasoning paths.
MaCP: Minimal yet Mighty Adaptation via Hierarchical Cosine Projection
Yixian Shen (University of Amsterdam), Anuj Pathania (University of Amsterdam)
Domain AdaptationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a Hierarchical Cosine Projection (MaCP) based on Discrete Cosine Transform (DCT) for parameter-efficient fine-tuning of large language models while maintaining accuracy.
MadaKV: Adaptive Modality-Perception KV Cache Eviction for Efficient Multimodal Long-Context Inference
Kunxi Li (Zhejiang University), Fei Wu (Shanghai Jiao Tong University)
Computational EfficiencyVision Language ModelMultimodalityBenchmark
🎯 What it does: Designed a KV cache compression strategy called MadaKV for multi-modal long-context reasoning, achieving adaptive modality-aware cache eviction.
MAGNET: Augmenting Generative Decoders with Representation Learning and Infilling Capabilities
Savya Khosla (Adobe Research), Jing Shi (Adobe Research)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: In MAGNET, the authors design and implement a method that enables a decoder LLM with causal attention to simultaneously obtain text encoding, missing paragraph filling, and retention generation capabilities through self-supervised learning.
Magnet: Multi-turn Tool-use Data Synthesis and Distillation via Graph Translation
Fan Yin (University of California, Los Angles), Tomas Pfister (Google)
Data SynthesisKnowledge DistillationGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: Built a graph-translation-based multi-round tool usage data synthesis and distillation framework named MAGNET to enhance large language models' multi-round function calling capabilities.
MAIN-RAG: Multi-Agent Filtering Retrieval-Augmented Generation
Chia-Yuan Chang (Texas A&M University), Na Zou (University of Houston)
GenerationRetrievalLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Propose a training-free, multi-agent retrieval-augmented generation framework named MAIN-RAG, which significantly improves the accuracy and reliability of retrieval-augmented generation (RAG) answers by having multiple LLM agents jointly filter, score, and sort retrieved documents.
Make Imagination Clearer! Stable Diffusion-based Visual Imagination for Multimodal Machine Translation
Andong Chen (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Image TranslationGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose the IMAGE framework, which generates images from source sentences using Stable Diffusion and then translates them using a multimodal LLM.
Making FETCH! Happen: Finding Emergent Dog Whistles Through Common Habitats
Kuleen Sasse (Johns Hopkins University), Mark Dredze (Johns Hopkins University)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the FETCH! task, automatically discovering emerging dog whistles in large-scale social media corpora and constructing a benchmark evaluation system.
Making LLMs Better Many-to-Many Speech-to-Text Translators with Curriculum Learning
Yexing Du (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: Reframe the speech-to-text translation (S2TT) task as a speech recognition + translation (SRT) task, leveraging the machine translation capability of large language models through a three-stage curriculum learning to achieve multilingual end-to-end S2TT.
Mamba Knockout for Unraveling Factual Information Flow
Nir Endy (Tel Aviv University), Raja Giryes (Tel Aviv University)
Explainability and InterpretabilityTransformerText
🎯 What it does: Migrate the Attention Knockout method from Transformer to Mamba (structured state space model), investigating the flow and localization of factual information in Mamba-1 and Mamba-2.
MAmmoTH-VL: Eliciting Multimodal Reasoning with Instruction Tuning at Scale
Jiawei Guo, Xiang Yue (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed a large-scale visual instruction dataset with 12 million entries containing intermediate reasoning (Chain-of-Thought), and trained the MAmmoTH-VL-8B model based on this dataset.
Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM-Based Multi-Agent System
Haoyang Su (Shanghai Artificial Intelligence Laboratory), Nanqing Dong (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelAgentic AITextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-agent system called VIRSCI based on large language models to simulate collaborative scientific research teams, automatically generating and evaluating research ideas and abstracts.
Map&Make: Schema Guided Text to Table Generation
Naman Ahuja (Arizona State University), Vivek Gupta (Arizona State University)
GenerationLarge Language ModelAgentic AIPrompt EngineeringTextTabularChain-of-Thought
🎯 What it does: Propose the Map&Make framework, which splits text into atomic propositions, infers potential table structures, and generates tables;
MAPLE: Enhancing Review Generation with Multi-Aspect Prompt LEarning in Explainable Recommendation
Ching-Wen Yang (National Cheng Kung University), Hung-Yu Kao (National Tsing Hua University)
Recommendation SystemExplainability and InterpretabilityTransformerPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the MAPLE model, which generates personalized, accurate, and diverse recommendation explanations through multi-aspect prompting learning.
MapNav: A Novel Memory Representation via Annotated Semantic Maps for VLM-based Vision-and-Language Navigation
Lingfeng Zhang (Hong Kong University of Science and Technology), Renjing Xu (Hong Kong University of Science and Technology)
SegmentationTransformerVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Propose an end-to-end visual-language navigation model called MapNav based on Annotated Semantic Map (ASM), directly replacing traditional historical frames with ASM to achieve compression and structuring of memory representation;
MAPoRL: Multi-Agent Post-Co-Training for Collaborative Large Language Models with Reinforcement Learning
Chanwoo Park (MIT), Joo-Kyung Kim (Amazon)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: This paper proposes MAPoRL—a multi-agent post-training framework that significantly improves the final answer accuracy of multi-LLM systems by jointly optimizing the collaborative behavior of LLMs using reinforcement learning in multi-agent dialogue environments.
Mapping 1,000+ Language Models via the Log-Likelihood Vector
Momose Oyama (Kyoto University), Hidetoshi Shimodaira (Kyoto University)
Computational EfficiencyRepresentation LearningLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a method to build model coordinates by calculating the log-likelihood vectors of autoregressive language models on a predefined text set, and uses these coordinates to approximate the KL divergence between models through the Euclidean distance in the log-likelihood space, enabling scalable comparison and visualization of over 1000 language models;
Mapping the Podcast Ecosystem with the Structured Podcast Research Corpus
Benjamin Roger Litterer, Dallas Card (University of Michigan)
TextMultimodalityAudio
🎯 What it does: This paper first constructs and publicly releases the SPORC dataset, which includes 1.1 million podcast transcriptions, covering metadata, audio features, speaker separation, and host/guest role annotations. Based on this dataset, the paper conducts a systematic analysis of podcast content, network structures, and responses to major events.
MAPS: Motivation-Aware Personalized Search via LLM-Driven Consultation Alignment
Weicong Qin (Renmin University of China), Jun Xu (Renmin University of China)
Recommendation SystemTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: By introducing AI consultation texts, extracting users' search intent, and utilizing LLM with Mixture of Attention Experts (MoAE) to map queries and consultations into a unified semantic space, achieving personalized product retrieval and ranking.
Marco-Bench-MIF: On Multilingual Instruction-Following Capability of Large Language
Bo Zeng (Alibaba International Digital Commerce), Kaifu Zhang (Alibaba International Digital Commerce)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed and evaluated a multilingual instruction-following benchmark named Marco-Bench-MIF, covering 30 languages with fine-grained localization tailored to language and cultural nuances;
Marco-o1 v2: Towards Widening The Distillation Bottleneck for Reasoning Models
Huifeng Yin (Alibaba International Digital Commerce), Kaifu Zhang (Alibaba International Digital Commerce)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose generating tree-like Chain-of-Thought (CoT) data from scratch using Monte Carlo Tree Search (MCTS), combined with multi-model collaboration and reflection nodes to create training data closer to human reasoning paths; subsequently, significantly enhance reasoning capabilities on small models through CoT-aware SFT, DPO (including conservative and masking-based variants), and joint loss methods.
MARS: Benchmarking the Metaphysical Reasoning Abilities of Language Models with a Multi-task Evaluation Dataset
Weiqi Wang (HKUST), Yangqiu Song (HKUST)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a three-step discriminative process for metaphysical reasoning to evaluate the reasoning capabilities of language models in the face of changes in situational distributions;
Masking in Multi-hop QA: An Analysis of How Language Models Perform with Context Permutation
Wenyu Huang (School Of Informatics University Of Edinburgh), Jeff Z. Pan (School Of Informatics University Of Edinburgh)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper systematically arranges the order, distance, and completeness of retrieval results in multi-hop question answering tasks to evaluate the performance of different language models (Encoder-Decoder vs Causal Decoder-Only) in contextual reasoning and explores the impact of bidirectional attention on performance.
Masks Can be Learned as an Alternative to Experts
Peiyu Liu, Shuicheng Yan (Skywork AI)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose the Mixture-of-Masks (MoM) method, converting dense LLMs into sparse MoE architecture by learning binary masks to achieve parameter activation sparsification, significantly reducing forward inference cost while maintaining performance.
MasRouter: Learning to Route LLMs for Multi-Agent Systems
Yanwei Yue (Tongji University), Yiyan Qi
OptimizationLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Designed and implemented a multi-agent system routing framework called MasRouter, capable of dynamically configuring collaboration modes, agent roles, and LLM backends for different queries, constructing an efficient multi-agent system.
Math Neurosurgery: Isolating Language Models’ Math Reasoning Abilities Using Only Forward Passes
Bryan R Christ (University of Virginia), Thomas Hartvigsen (University of Virginia)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a method called MathNeuro that identifies math reasoning-related parameters in LLMs using only forward propagation;
MathFusion: Enhancing Mathematical Problem-solving of LLM through Instruction Fusion
Qizhi Pei (Renmin University of China), Rui Yan (Renmin University of China)
Data SynthesisData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: Propose the MathFusion framework, which combines existing mathematical problems to generate new ones through three fusion strategies: sequential, parallel, and conditional, and further fine-tunes LLMs with instruction-based tuning;
MaXIFE: Multilingual and Cross-lingual Instruction Following Evaluation
Yile Liu (Waseda University), Liang Li (OPPO AI Center)
Large Language ModelTextBenchmark
🎯 What it does: Proposed and implemented the MaXIFE benchmark for systematic evaluation of multilingual (23 languages) and cross-lingual instruction-following capabilities;
Maximal Matching Matters: Preventing Representation Collapse for Robust Cross-Modal Retrieval
Hani Alomari (Virginia Tech), Chris Thomas (Virginia Tech)
RetrievalConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a set-based image-text retrieval framework called MaxMatch, which learns multiple embeddings through maximum one-to-one matching to avoid issues of sparse supervision and set collapse.
Maximizing the Effectiveness of Larger BERT Models for Compression
Wen-Shu Fan (Nanjing University), De-Chuan Zhan (Nanjing University)
CompressionKnowledge DistillationTransformerTextBenchmark
🎯 What it does: Propose and implement the MC3KD method, enhancing the compression effectiveness of large BERT models through teacher intermediate layer selection and maximizing the canonical correlation coefficient between student and teacher layers.
MCS-Bench: A Comprehensive Benchmark for Evaluating Multimodal Large Language Models in Chinese Classical Studies
Yang Liu (South China University of Technology), Lianwen Jin (South China University of Technology)
RecognitionTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed and constructed a multimodal evaluation benchmark named MCS-Bench for Chinese classical studies, encompassing seven subfields with 45 fine-grained tasks.
MDCure: A Scalable Pipeline for Multi-Document Instruction-Following
Gabrielle Kaili-May Liu (Yale University), Arman Cohan (Yale University)
GenerationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: MDCure designed a scalable multi-document instruction generation and filtering pipeline, leveraging zero-shot prompts to generate high-quality multi-document instruction pairs, and then fine-tuning LLMs using a multi-objective reward model to enhance their multi-document instruction-following capabilities.
Meaning Beyond Truth Conditions: Evaluating Discourse Level Understanding via Anaphora Accessibility
Xiaomeng Zhu (Yale University), Robert Frank (Yale University)
Data-Centric LearningLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed a discourse-level understanding task based on referential reachability evaluation, and generated a dataset containing approximately 9,816 sentences with structures such as quantifiers, negations, and disjunctive clauses.