ACL 2025 Papers with AI Summaries
Annual Meeting of the Association for Computational Linguistics · 1699 papers
→ ACL 2025 papers with code (518)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 1699 ACL 2025 papers by keyword, author or institution —
free trial on arXivSub.
“Give Me BF16 or Give Me Death”? Accuracy-Performance Trade-Offs in LLM Quantization
Eldar Kurtic (Red Hat AI), Dan Alistarh (Red Hat AI)
Computational EfficiencyTransformerTextBenchmark
🎯 What it does: Investigated the accuracy, performance, and cost trade-offs of quantization formats such as FP8, INT8, and INT4 on the Llama-3.1 large model.
“What do you call a dog that is incontrovertibly true? Dogma”: Testing LLM Generalization through Humor
Alessio Cocchieri (University of Bologna), Gianluca Moro (University of Bologna)
GenerationData SynthesisTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Created and evaluated PHUNNY, a structured pun-based QA benchmark based on single-step affix rewriting, examining the generalization ability of large language models in humor reasoning and generation.
“Yes, My LoRD.” Guiding Language Model Extraction with Locality Reinforced Distillation
Zi Liang (Hong Kong Polytechnic University), Haibo Hu (Hong Kong Polytechnic University)
Knowledge DistillationAdversarial AttackTransformerReinforcement LearningContrastive LearningText
🎯 What it does: This paper proposes LoRD, a model extraction attack method for large-scale language models, which efficiently steals the target model through local reinforcement distillation.
(RSA)²: A Rhetorical-Strategy-Aware Rational Speech Act Framework for Figurative Language Understanding
Cesare Spinoso-Di Piano (McGill University), Jackie CK Cheung
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a new Rhetorical-Strategy-Aware RSA (RSA 2) framework that can explain non-literal language such as metaphors and irony without explicitly modeling speakers' motivations;
\mathcal{A}^3: Automatic Alignment Framework for Attributed Text Generation
Yue Wang (Soochow University), Min Zhang (Ant Group)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the A³ automatic alignment framework for generating high-quality citation-aware QA pairs without manual annotation, supporting two training stages (SFT and PO) to improve LLM citation recall and precision.
\phi-Decoding: Adaptive Foresight Sampling for Balanced Inference-Time Exploration and Exploitation
Fangzhi Xu (Xi'an Jiaotong University), Zhiyong Wu (Shanghai AI Lab)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a self-adaptive reasoning decoding strategy φ-Decoding, which does not rely on external reward models and is based on lookahead sampling and clustering, to dynamically balance exploration and exploitation during reasoning, thereby improving the reasoning accuracy of LLMs.
\textit{L-CiteEval}: A Suite for Evaluating Fidelity of Long-context Models
Zecheng Tang (Soochow University), Min Zhang (Chinese University of Hong Kong)
Large Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the L-CiteEval evaluation suite, which assesses long-text models (LCM) in terms of generation quality and context fidelity (fidelity), covering 11 tasks across 5 categories, and supporting context lengths from 8K to 48K.
500xCompressor: Generalized Prompt Compression for Large Language Models
Zongqian Li (University of Cambridge), Nigel Collier (University of Cambridge)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: Propose 500xCompressor, a soft prompting compression method that can compress long texts into a single special token while retaining most of the original functionality.
A Drop-In Solution for On-the-Fly Adaptation of Speculative Decoding in Large Language Models
Jiesong Liu (North Carolina State University), Xipeng Shen (North Carolina State University)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: The study proposes a speculative decoding method that does not require offline benchmarks or training, enabling real-time adjustment during inference to enhance LLM inference speed.
A Dual-Mind Framework for Strategic and Expressive Negotiation Agent
Yutong Liu (Jilin University), Hao Xu (Jilin University)
TransformerLarge Language ModelText
🎯 What it does: Proposes the Dual-Mind Negotiation Agent (DMNA) framework, combining intuitive fast experience response with deliberate multi-dimensional reflection to enhance negotiation agents' strategy planning and expression quality.
A Dual-Perspective NLG Meta-Evaluation Framework with Automatic Benchmark and Better Interpretability
Xinyu Hu (Wangxuan Institute of Computer Technology, Peking University), Xiaojun Wan (Wangxuan Institute of Computer Technology, Peking University)
GenerationExplainability and InterpretabilityLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Proposed a dual-perspective NLG meta-evaluation framework (global perspective and local perspective), and implemented an automatic benchmark construction method without requiring human annotation;
A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning
Zhiyu Zhang (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
GenerationTransformerPrompt EngineeringDiffusion modelGraph
🎯 What it does: Proposed a deep generative adaptive replay continuous learning framework named DGAR to alleviate catastrophic forgetting in temporal knowledge graph reasoning.
A Little Human Data Goes A Long Way
Dhananjay Ashok (Information Sciences Institute, University of Southern California), Jonathan May (Information Sciences Institute, University of Southern California)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Evaluate the impact on model performance in fact verification and evidence-based question answering tasks when replacing a large proportion of synthetic data with a small amount of human-annotated data (125–200 examples), and quantify the cost-performance trade-off.
A Measure of the System Dependence of Automated Metrics
Pius Von Däniken, Mark Cieliebak (Zurich University Of Applied Sciences)
Text
🎯 What it does: Proposed and evaluated a new metric method (SysDep) to quantify the degree of bias of automatic machine translation evaluation metrics across different systems.
A Modular Approach for Clinical SLMs Driven by Synthetic Data with Pre-Instruction Tuning, Model Merging, and Clinical-Tasks Alignment
Jean-Philippe Corbeil (Microsoft Healthcare & Life Sciences), Paul Vozila (Microsoft Healthcare & Life Sciences)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningBiomedical DataBenchmark
🎯 What it does: Proposed the MediPhi multi-module framework, which elevates small language models to clinical levels through pre-instruction fine-tuning, model merging, and synthetic instruction data alignment.
A Multi-Agent Framework for Mitigating Dialect Biases in Privacy Policy Question-Answering Systems
Đorđe Klisura (University of Texas at San Antonio), Anthony Rios (University of Texas at San Antonio)
Data-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This paper proposes a multi-agent framework that improves the performance of privacy policy question-answering systems across different English dialects by translating user non-standard English queries into standard American English (SAE) and integrating privacy policy expertise;
A Multi-persona Framework for Argument Quality Assessment
Bojun Jin (Harbin Institute of Technology), Ruifeng Xu (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a multi-perspective framework called MPAQ that utilizes large language models to simulate different assessor personalities, generating argument quality scores and reasons from multiple viewpoints.
A Mutual Information Perspective on Knowledge Graph Embedding
Jiang Li (Inner Mongolia University), Guanglai Gao (Inner Mongolia University)
Representation LearningContrastive LearningGraph
🎯 What it does: Propose a knowledge graph embedding (KGE) framework based on mutual information maximization to enhance semantic representations of entities and relations, while reducing intra-group similarity in 1-N and N-1 relations.
A New Formulation of Zipf’s Meaning-Frequency Law through Contextual Diversity
Ryo Nagata (Konan University), Kumiko Tanaka-Ishii (Waseda University)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: This paper reinterprets Zipf's Law of Meaning Frequency by measuring the variability in the direction of context vectors, avoiding the use of a dictionary.
A Parameter-Efficient and Fine-Grained Prompt Learning for Vision-Language Models
Yongbin Guo (South China University of Technology), C.L.Philip Chen
RetrievalTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposed the fine-grained prompt learning framework DoPL, achieving visual-text fine-grained semantic alignment with only 0.12% tunable parameters through non-parametric fine-grained prompt generation (DPG) and low-entropy information concentration theory, improving cross-modal understanding performance.
A Reality Check on Context Utilisation for Retrieval-Augmented Generation
Lovisa Hagström (Chalmers University of Technology), Isabelle Augenstein (University of Copenhagen)
GenerationData SynthesisTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a real-world retrieval-augmented generation (RAG) evaluation framework and constructs a new dataset called DRUID to study the actual usage of retrieval contexts in RAG.
A Self-Denoising Model for Robust Few-Shot Relation Extraction
Liang Zhang (Xiamen University), Jinsong Su (Xiamen University)
ClassificationMeta LearningGraph Neural NetworkTransformerText
🎯 What it does: This paper studies the problem of label noise in the support set of few-shot relation extraction, proposing a self-de-noising model that can automatically correct erroneous labels of support instances and enhance model robustness.
A Silver Bullet or a Compromise for Full Attention? A Comprehensive Study of Gist Token-based Context Compression
Chenlong Deng (Renmin University of China), Zhicheng Dou (Tencent)
CompressionComputational EfficiencyTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This paper conducts a systematic empirical evaluation of context compression methods based on gist tokens, exploring whether they can replace full attention models and identifying the main failure modes caused by compression, and proposes two improvement strategies based on this.
A Spatio-Temporal Point Process for Fine-Grained Modeling of Reading Behavior
Francesco Ignazio Re (ETH Zürich), Ryan Cotterell (ETH Zürich)
Time Series
🎯 What it does: Proposes a marked point process model based on the spatio-temporal Hawkes process to simultaneously predict fixation start time, spatial location, and duration during reading.
A Statistical and Multi-Perspective Revisiting of the Membership Inference Attack in Large Language Models
Bowen Chen (University of Tokyo), Yusuke Miyao (University of Tokyo)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper reassesses membership inference attacks (MIA) in large language models through large-scale statistical experiments, revealing that their performance varies with model scale, domain, text length, and threshold selection.
A Strategic Coordination Framework of Small LMs Matches Large LMs in Data Synthesis
Xin Gao (Shanghai Artificial Intelligence Laboratory), Conghui He (Shanghai Artificial Intelligence Laboratory)
Data SynthesisTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a multi-small-model collaborative framework GRA, which replaces a single large model by dividing tasks into three stages: generator, reviewer, and adjudicator, to accomplish data synthesis and quality control;
A Survey of Post-Training Scaling in Large Language Models
Hanyu Lai (Tsinghua University), Jie Tang (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningTextReview/Survey Paper
🎯 What it does: Reviews the motivations, classifications, scalable implementation methods, and the latest progress in application scenarios such as mathematics, code generation, and agent execution for post-training expansion techniques (SFT, RLxF, TTC).
A Survey on Efficient Large Language Model Training: From Data-centric Perspectives
Junyu Luo (Peking University), Ming Zhang (Peking University)
Knowledge DistillationData-Centric LearningTransformerLarge Language ModelReview/Survey Paper
🎯 What it does: This paper provides a systematic review of data-efficient methods in the post-training of large language models, proposing a 'data flywheel' framework based on data centers and a five-category taxonomy.
A Survey on Foundation Language Models for Single-cell Biology
Fan Zhang (Tencent Jarvis Lab), Xian Wu (Tencent Jarvis Lab)
TransformerLarge Language ModelBiomedical DataReview/Survey Paper
🎯 What it does: Systematically reviews foundational language models in the field of single-cell biology (including single-cell pre-trained models PLMs and large language models LLMs), providing a detailed overview of data labeling, pre-training paradigms, single-cell to text conversion, and optimization methods, and summarizes the applications of these models in various analysis tasks at the cell-level, gene-level, drug, spatial, and other levels.
A Survey on Patent Analysis: From NLP to Multimodal AI
Homaira Huda Shomee (University of Illinois Chicago), Sourav Medya (University of Illinois Chicago)
ClassificationGenerationRetrievalTransformerLarge Language ModelTextMultimodalityReview/Survey Paper
🎯 What it does: Reviews the main tasks of patent analysis (classification, retrieval, quality assessment, and generation), and systematically summarizes the applications and methods of NLP, Multimodal AI, and LLM in these tasks in recent years;
A Systematic Study of Compositional Syntactic Transformer Language Models
Yida Zhao (ShanghaiTech University), Kewei Tu (Ant Group)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed and unified the Compositional Syntax Language Model (SLM) framework based on constituent syntax, and explored and experimented with over ten model variants under this framework
A Text is Worth Several Tokens: Text Embedding from LLMs Secretly Aligns Well with The Key Tokens
Zhijie Nie (Beihang University), Zhanyu Wu (Beihang University)
RetrievalExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper investigates and proves that vectors generated by large language models (LLMs) during the text embedding phase align with key tokens in the input text after passing through the decoding layer, and explains this phenomenon through spectral analysis, further leveraging this discovery to achieve sparse retrieval and model interpretability.
A Theory of Response Sampling in LLMs: Part Descriptive and Part Prescriptive
Sarath Sivaprasad (CISPA Helmholtz Center for Information Security), Mario Fritz (TCS Research)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigate the heuristic mechanisms in the sampling decision process of large language models (LLMs), proposing two component models: descriptive and prescriptive.
A Training-free LLM-based Approach to General Chinese Character Error Correction
Houquan Zhou (Soochow University), Min Zhang (Soochow University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the general Chinese character error correction (C2EC) task and construct a high-quality real error dataset, while extending an unsupervised prompt-free training framework to handle insertion, deletion, and substitution errors.
A Triple-View Framework for Fine-Grained Emotion Classification with Clustering-Guided Contrastive Learning
Junqing Gong (Nankai University), Wei Shen (Nankai University)
ClassificationTransformerContrastive LearningText
🎯 What it does: Proposed a three-view framework called TACO, which jointly learns instance-label, instance-instance, and label-label embeddings to address the challenges of distinguishing similar sentiments and mitigating long-tail sentiment bias in fine-grained sentiment classification.
A Troublemaker with Contagious Jailbreak Makes Chaos in Honest Towns
Tianyi Men (Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a large-scale, multi-topology, text-based multi-agent attack evaluation framework (TMCHT) and introduces a propagative adversarial attack method, ARCJ, targeting independent memory systems to address the 'toxicity fading' problem of memory attacks in non-complete graph structures and large-scale systems.
A Unified Agentic Framework for Evaluating Conditional Image Generation
Jifang Wang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
GenerationLarge Language ModelAgentic AIImageTextMultimodalityBenchmark
🎯 What it does: Built a unified agentic framework CIGEVAL for evaluating seven conditional image generation tasks;
A Variational Approach for Mitigating Entity Bias in Relation Extraction
Samuel Mensah (JP Morgan), Charese Smiley (JP Morgan)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextBiomedical DataFinance Related
🎯 What it does: The study proposes an entity debiasing method based on the variational information bottleneck to reduce over-reliance on entities in relation extraction models.
A-TASC: Asian TED-Based Automatic Subtitling Corpus
Yuhan Zhou (University of Tokyo), Naoki Yoshinaga (University of Tokyo)
Data-Centric LearningTextBenchmarkAudio
🎯 What it does: Proposed A-TASC, an Asian TED benchmark automatic captioning dataset covering Chinese, Japanese, Korean, and Vietnamese with approximately 800 hours of data; developed SacreSubER evaluation metric to adapt to non-space languages; evaluated performance of end-to-end and pipeline captioning systems.
AAD-LLM: Neural Attention-Driven Auditory Scene Understanding
Xilin Jiang (Columbia University), Nima Mesgarani (Columbia University)
RecognitionRecurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringBiomedical DataChain-of-ThoughtAudio
🎯 What it does: Propose a system called AAD-LLM that decodes EEG signals into the listener's attention and integrates it into an auditory large language model (LLM) for listener-intent-driven speech understanding and answering in multi-speaker scenarios.
AbGen: Evaluating Large Language Models in Ablation Study Design and Evaluation for Scientific Research
Yilun Zhao (Yale NLP Lab), Arman Cohan (TCS Research)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the ABGEN benchmark to evaluate the ability of large language models (LLMs) in designing ablation experiments in scientific research;
Accelerating Dense LLMs via L0-regularized Mixture-of-Experts
Zhenyu Zhang (YZW), Meng Chen (Wise AI)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Construct a lightweight Mixture-of-Experts (MoE) model using L0 regularization, accelerating inference of large language models with a small corpus of only 30B tokens while maintaining or improving performance.
Accurate KV Cache Quantization with Outlier Tokens Tracing
Yi Su (Soochow University), Min Zhang (Huawei Cloud)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper addresses the quantization issue of LLM KV cache, finding that a small number of abnormal tokens significantly affect quantization accuracy. It proposes the OTT method to dynamically identify and exclude these outlier tokens, retaining them in full precision to improve the accuracy of 2bit quantization, while reducing memory usage and increasing throughput.
ACECODER: Acing Coder RL via Automated Test-Case Synthesis
Huaye Zeng (University of Waterloo), Wenhu Chen (University of Waterloo)
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Construct the ACECODER framework by leveraging automated large-scale test case synthesis and reward model training to achieve RL fine-tuning of code generation models;
ACORD: An Expert-Annotated Retrieval Dataset for Legal Contract Drafting
Steven H Wang, Roger Wattenhofer (Atticus Project)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed the ACORD expert-annotated contract clause retrieval benchmark dataset, covering 114 queries and 126k clauses, focusing on high-difficulty clauses.
Acoustic Individual Identification of White-Faced Capuchin Monkeys Using Joint Multi-Species Embeddings
Álvaro Vega-Hidalgo (University of Michigan), Rada Mihalcea (University of Michigan)
RecognitionTransformerAudio
🎯 What it does: Using cross-species pre-trained acoustic embeddings (human speech and bird acoustic models) for individual identification of capuchin monkeys (Capuchin) in the Taboga Reserve, Costa Rica, and exploring combination methods of multi-species embeddings.
Acquisition and Application of Novel Knowledge in Large Language Models
Ziyu Shang (Southeast University), Guozheng Li (Southeast University)
Data-Centric LearningTransformerLarge Language ModelTextGraph
🎯 What it does: Propose a knowledge synthesis method based on the theory of biological evolution, construct a new knowledge dataset called NovelHuman, and investigate the intra-sentence position sensitivity of LLMs when acquiring new knowledge. The Permutation-AR (PermAR) framework is introduced to address the sequential limitations of autoregressive models, thereby enhancing knowledge acquisition capabilities.
ACT: Knowledgeable Agents to Design and Perform Complex Tasks
Makoto Nakatsuji (NTT Communication Science Laboratories), Yoshihide Sato (NTT Human Informatics Laboratories)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: Proposes the ACT framework, enabling multiple agents to collaboratively design and execute complex tasks in team meetings through shared structured knowledge, achieving accurate task coordination.
Activating Distributed Visual Region within LLMs for Efficient and Effective Vision-Language Training and Inference
Siyuan Wang (University of Southern California), Zhongyu Wei (Fudan University)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper studies identifying and activating visual regions within large vision-language models (LVLM), proving that updating approximately 25% of sparsely uniformly distributed layers can maintain 99% visual performance while preserving text capabilities.
Activation Steering Decoding: Mitigating Hallucination in Large Vision-Language Models through Bidirectional Hidden State Intervention
Jingran Su (Hong Kong Polytechnic University), Zhaoxiang Zhang (CASIA)
Explainability and InterpretabilityVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose a training-agnostic activation-guided decoding method (ASD) that suppresses hallucination generation in large-scale vision-language models by injecting adversarial direction vectors into intermediate hidden states
ActiView: Evaluating Active Perception Ability for Multimodal Large Language Models
Ziyue Wang (Tsinghua University), Yang Liu (Tsinghua University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes and implements a benchmark named ActiView to evaluate the capabilities of multimodal large language models in active perception (field-of-view switching and zooming). By designing three evaluation pipelines (Zooming, Shifting, Mixed) and constructing a question-answering dataset with multi-perspective, fine-grained prompts, the performance of 30 models is systematically assessed. Results show significant gaps between existing models and humans, especially poor performance in active perception. By comparing full-image input, manual prompts, and model-driven zooming/switching, the necessity and improvement potential of active perception are validated. Meanwhile, the accuracy of view selection and final answer accuracy are deeply analyzed.
AdaDHP: Fine-Grained Fine-Tuning via Dual Hadamard Product and Adaptive Parameter Selection
Han Liu (Dalian University of Technology), Hong Yu (Dalian University of Technology)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose AdaDHP, an efficient fine-tuning method that simultaneously performs row-column Hadamard products on each weight matrix and dynamically selects important parameters.
AdaEdit: Advancing Continuous Knowledge Editing For Large Language Models
Qi Li (Hong Kong University of Science and Technology), Xiaowen Chu (Hong Kong University of Science and Technology)
OptimizationRepresentation LearningTransformerText
🎯 What it does: Propose AdaEdit, a continuous knowledge editing method for large language models, aiming to address the significant performance degradation in knowledge update and retention during continuous editing.
AdamMeme: Adaptively Probe the Reasoning Capacity of Multimodal Large Language Models on Harmfulness
Zixin Chen (Beijing University Of Posts And Telecommunications), Jing Ma (Hong Kong Baptist University)
Safty and PrivacyTransformerLarge Language ModelAgentic AIMultimodalityBenchmark
🎯 What it does: Proposes the AdamMeme framework, which utilizes multi-agent dynamic evaluation to assess the understanding of harmful memes by multi-modal LLMs.
Adapt Once, Thrive with Updates: Transferable Parameter-Efficient Fine-Tuning on Evolving Base Models
Naibin Gu (Chinese Academy of Sciences), Weiping Wang (Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes Trans-PEFT, a technique that enables migration of PEFT modules without requiring retraining after updates to large language model versions.
AdaptAgent: Adapting Multimodal Web Agents with Few-Shot Learning from Human Demonstrations
Gaurav Verma (Georgia Institute of Technology), Manuela Veloso (J.P. Morgan AI Research)
Meta LearningTransformerLarge Language ModelAgentic AIVision Language ModelVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Propose the AdaptAgent framework, enabling multimodal web agents to rapidly adapt to new websites and domains with only 1-2 human demonstrations.
Adaptive and Robust Translation from Natural Language to Multi-model Query Languages
Gengyuan Shi (Tsinghua University), Jiawei Ren (Tsinghua University)
Representation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the text-to-multi-model query language (MMQL) translation task and constructed the first corresponding dataset;
Adaptive Retrieval Without Self-Knowledge? Bringing Uncertainty Back Home
Viktor Moskvoretskii (Skoltech), Alexander Panchenko (Skoltech)
RetrievalComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Evaluated and compared the performance, efficiency, and self-awareness capabilities of 35 adaptive retrieval methods and 27 uncertainty estimation approaches in QA, demonstrating that uncertainty methods significantly reduce retrieval and LLM calls while maintaining accuracy.
Adaptive Tool Use in Large Language Models with Meta-Cognition Trigger
Wenjun Li (Huawei Noah's Ark Lab), Yong Liu (Huawei Noah's Ark Lab)
Computational EfficiencyMeta LearningLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose MeCo as a lightweight metacognitive trigger that determines when an LLM should invoke external tools or retrieval to improve decision-making efficiency.
Addressing Blind Guessing: Calibration of Selection Bias in Multiple-Choice Question Answering by Video Language Models
Olga Loginova (University of Trento), Alexey Kravets (University of Bath)
Explainability and InterpretabilityVision Language ModelVideoTextMultimodality
🎯 What it does: Through systematic experiments on video multiple-choice questions (MCQA), this paper analyzes and quantifies the selection bias of video language models (VLMs) in answer positions, proposing a post-processing calibration method called BOLD (and its weighted version Weighted_BOLD) to reduce bias and improve accuracy.
Advancing Collaborative Debates with Role Differentiation through Multi-Agent Reinforcement Learning
Haoran Li (South China Normal University), Minlie Huang
Large Language ModelReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: Proposed a multi-LLM collaborative reasoning framework with automatic role assignment to enhance performance on mathematical reasoning tasks.
Advancing Sequential Numerical Prediction in Autoregressive Models
Xiang Fei (ByteDance Inc), Can Huang (ByteDance Inc)
RecognitionComputational EfficiencyImageTextTime SeriesSequential
🎯 What it does: Proposed a new sequence-level numerical prediction loss called NTIL, specifically designed for autoregressive models;
Advancing SMoE for Continuous Domain Adaptation of MLLMs: Adaptive Router and Domain-Specific Loss
Liang Zhang (Xiamen University), Jinsong Su (Xiamen University)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsMultimodalityBenchmark
🎯 What it does: Proposes a continuous instruction tuning framework based on Sparse Mixture of Experts (SMoE) for domain incremental learning in multimodal large language models, avoiding catastrophic forgetting.
Advancing Zero-shot Text-to-Speech Intelligibility across Diverse Domains via Preference Alignment
Xueyao Zhang (Chinese University of Hong Kong, Shenzhen), Zhizheng Wu (Chinese University of Hong Kong, Shenzhen)
GenerationData SynthesisTransformerReinforcement LearningFlow-based ModelAudio
🎯 What it does: Improving the intelligibility of zero-shot TTS, the Intelligibility Preference Speech Dataset (INTP) was constructed, and post-training was conducted on multiple TTS architectures using Preference Alignment (Direct Preference Optimization DPO and its extensions).
Adversarial Alignment with Anchor Dragging Drift (A^3D^2): Multimodal Domain Adaptation with Partially Shifted Modalities
Jun Sun (Zhejiang Lab), Lingfang Zeng (Zhejiang Lab)
Domain AdaptationRecurrent Neural NetworkTransformerMultimodality
🎯 What it does: Proposed a multi-modal domain adaptation method addressing partial modal invariance (anchor) and partial modal shift (drift), employing a dual alignment strategy to simultaneously achieve drift-drift and anchor-drift matching.
Adversarial Tokenization
Renato Geh, Guy Van Den Broeck
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper investigates the semantic preservation and security evasion phenomena of LLMs when using non-canonical tokenization, proving that attackers can achieve jailbreak, bypass security models, and prompt injection attacks by merely rewriting tokenization (without altering the text). It also proposes a greedy search algorithm called AdvTok to find the required tokenization for attacks and experimentally validates its effectiveness in three attack scenarios.
Adverse Event Extraction from Discharge Summaries: A New Dataset, Annotation Scheme, and Initial Findings
Imane Guellil (University of Edinburgh), Beatrice Alex (University of Edinburgh)
Anomaly DetectionRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health Records
🎯 What it does: This paper constructs a manually annotated adverse event corpus for elderly discharge summaries and provides corresponding annotation guidelines.
AfriMed-QA: A Pan-African, Multi-Specialty, Medical Question-Answering Benchmark Dataset
Charles Nimo (Georgia Institute of Technology), Mercy Nyamewaa Asiedu (SisonkeBiotik)
TextBenchmark
🎯 What it does: Constructed AfriMed-QA, an English medical QA dataset containing 15,275 multiple-choice questions, short-answer questions, and consumer inquiries across 16 African countries and 32 medical specialties, and evaluated 30 LLMs.
AfroCS-xs: Creating a Compact, High-Quality, Human-Validated Code-Switched Dataset for African Languages
Kayode Olaleye (University of Pretoria), Vukosi Marivate (University of Pretoria)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextAgriculture RelatedFinance Related
🎯 What it does: Constructed a manually verified synthetic code-switching dataset AfroCS-xs for four African languages and English, used for fine-tuning LLMs to perform code-switching translation.
AGD: Adversarial Game Defense Against Jailbreak Attacks in Large Language Models
Shilong Pan (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
Safty and PrivacyTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: Propose the AGD (Adversarial Game Defense) defense framework, which dynamically balances the usefulness and safety of large language models against jailbreak attacks by utilizing adversarial correction of abnormal attention weights, variable-sum game regulation of attention head activation, and expert model-guided next token sampling.
Agent-RewardBench: Towards a Unified Benchmark for Reward Modeling across Perception, Planning, and Safety in Real-World Multimodal Agents
Tianyi Men (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Jun Zhao (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences)
TransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Proposed Agent-RewardBench, a benchmark for evaluating the reward modeling capabilities of multimodal large language models (MLLMs) in intelligent agent tasks.
AgentDropout: Dynamic Agent Elimination for Token-Efficient and High-Performance LLM-Based Multi-Agent Collaboration
Zhexuan Wang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Computational EfficiencyGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposes AgentDropout, a dynamic node and edge dropout mechanism to optimize the communication topology of LLM-based multi-agent systems, thereby reducing token consumption and improving task performance.
AgentGym: Evaluating and Training Large Language Model-based Agents across Diverse Environments
Zhiheng Xi (Fudan University), Yu-Gang Jiang (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose the AGENTGYM framework, instruction set AGENTEVAL, trajectory sets AGENTTRAJ/AGENTTRAJ-L, and develop the self-improvement learning method AGENTSTAR for evaluating and training general agents based on LLMs.
Agentic Knowledgeable Self-awareness
Shuofei Qiao (Zhejiang University), Huajun Chen (Zhejiang University)
TransformerSupervised Fine-TuningAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposes the agentic knowledge-based self-awareness (KnowSelf) method, enabling LLM agents to autonomously decide whether to reflect or query knowledge based on contextual situations.
Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning with Agentic Tools
Junde Wu (University of Oxford), Yueming Jin (National University of Singapore)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Propose the Agentic Reasoning framework, integrating external tools (Web-Search, Coding, Mind-Map) into the LLM reasoning process to enhance reasoning and in-depth research capabilities for complex knowledge tasks.
Agentic Reward Modeling: Integrating Human Preferences with Verifiable Correctness Signals for Reliable Reward Systems
Hao Peng (Tsinghua University), Juanzi Li (Tsinghua University)
Reinforcement Learning from Human FeedbackLarge Language ModelContrastive LearningTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose 'Agentic Reward Modeling,' which incorporates verifiable correctness signals (factuality and instruction following) into traditional reward models to generate more reliable rewards;
AgentRM: Enhancing Agent Generalization with Reward Modeling
Yu Xia (Tsinghua University), Maosong Sun (Tsinghua University)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Introduce a generalizable reward model AgentRM in LLM-based agents, guiding the policy model to search during testing through the reward model, thereby improving performance on unknown tasks.
Agents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks
Rana Shahroz, Tianlong Chen (University of North Carolina at Chapel Hill)
Adversarial AttackTransformerPrompt EngineeringTextBenchmark
🎯 What it does: The study conducts covert attacks on security mechanisms in multi-agent LLM systems by optimizing prompt propagation paths and employing permutation-invariant escape loss;
AGrail: A Lifelong Agent Guardrail with Effective and Adaptive Safety Detection
Weidi Luo (Ohio State University), Chaowei Xiao (University of Wisconsin-Madison)
Safty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a lifelong learning guardian framework AGrail for adaptive risk detection and safety strategy optimization of LLM agent behaviors; simultaneously constructed the Safe-OS benchmark dataset for operating system agents.
Agri-CM^3: A Chinese Massive Multi-modal, Multi-level Benchmark for Agricultural Understanding and Reasoning
Haotian Wang (Harbin Institute Of Technology), Jingchi Jiang (Harbin Institute Of Technology)
Vision Language ModelImageTextMultimodalityBenchmarkAgriculture RelatedChain-of-Thought
🎯 What it does: Proposed and constructed Agri-CM 3, a multi-modal multi-level benchmark for crop pest and disease management, to evaluate the performance of multi-modal large language models across three levels: perception, hybrid cognitive reasoning, and knowledge application.
AIMSCheck: Leveraging LLMs for AI-Assisted Review of Modern Slavery Statements Across Jurisdictions
Adriana Eufrosina Bora (Mila Quebec AI Institute), Kerrie Mengersen (Queensland University of Technology)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Under the context of the Modern Slavery Act, a compliance checking framework named AIMSCheck was constructed for cross-border corporate disclosure texts, and the AIMS.uk and AIMS.ca datasets for the UK and Canada were released, supporting sentence-level, multi-label classification, token-level explanation, and evidence status tracking;
AIR-Bench: Automated Heterogeneous Information Retrieval Benchmark
Jianlyu Chen (University of Science and Technology of China), Zheng Liu (Beijing Academy of Artificial Intelligence)
RetrievalTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed AIR-BENCH—a fully automated, heterogeneous, and dynamically scalable retrieval evaluation benchmark that leverages large language models to automatically generate multi-task, multi-domain, and multi-lingual test sets;
ALGEN: Few-shot Inversion Attacks on Textual Embeddings via Cross-Model Alignment and Generation
Yiyi Chen (Aalborg University), Johannes Bjerva (Aalborg University)
Representation LearningAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose a cross-model alignment and generation-based few-shot text embedding inversion attack (ALGEN), which can recover the original text with only a minimal number of leaked samples.
Algorithmic Fidelity of Large Language Models in Generating Synthetic German Public Opinions: A Case Study
Bolei Ma (LMU Munich), Matthias Aßenmacher
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Evaluate the algorithmic fidelity of large language models (LLMs) in generating German public opinions (open-ended questions), comparing how different models reproduce subgroup perspectives.
Align-SLM: Textless Spoken Language Models with Reinforcement Learning from AI Feedback
Guan-Ting Lin (Amazon AGI), Ivan Bulyko (Amazon AGI)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAudio
🎯 What it does: Propose the Align-SLM framework, which uses preference optimization (RLAIF+DPO) with AI (LLM) feedback to semantically align speech-language models (SLM) without text, improving performance in speech-to-speech generation tasks.
AlignDistil: Token-Level Language Model Alignment as Adaptive Policy Distillation
Songming Zhang (Beijing Jiaotong University), Jinan Xu (Beijing Jiaotong University)
Knowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Propose AlignDistil, a token-level alignment method based on RLHF and DPO loss, which leverages the teacher distribution to guide the distributed learning of the student LLM;
Aligned but Blind: Alignment Increases Implicit Bias by Reducing Awareness of Race
Lihao Sun (University of Chicago), Xuechunzi Bai (University of Chicago)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerPrompt EngineeringText
🎯 What it does: The paper constructs prompts containing implicit and explicit racial bias to evaluate and compare the differences in explicit and implicit bias between the Llama3 base and aligned models. It then uses activation patches, Self-IE, and other mechanisms to explain internal representations, and designs interventions that inject racial concepts to mitigate implicit bias.
Aligning AI Research with the Needs of Clinical Coding Workflows: Eight Recommendations Based on US Data Analysis and Critical Review
Yidong Gan (University of Sydney), Jonathan K. Kummerfeld (University of Sydney)
ClassificationConvolutional Neural NetworkTransformerTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper critically evaluates the commonly used assessment methods in existing automatic encoding research through in-depth analysis of electronic medical record data from the US MIMIC-III and MIMIC-IV datasets. It points out the disconnection between these methods and actual clinical coding needs, and proposes eight specific improvement suggestions based on these findings. Meanwhile, the authors propose several workflow-based novel AI-assisted coding and auditing methods.
Aligning Large Language Models to Follow Instructions and Hallucinate Less via Effective Data Filtering
Shuzheng Si (Tsinghua University), Maosong Sun (Tsinghua University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Developed and validated the NOVA framework, which utilizes offline data filtering techniques to select samples highly consistent with the model's existing knowledge during the LLM instruction fine-tuning phase, thereby reducing hallucinations while maintaining instruction-following capabilities.
Aligning Large Language Models with Implicit Preferences from User-Generated Content
Zhaoxuan Tan (University Of Notre Dame), Meng Jiang (University Of Notre Dame)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the PUGC framework, which leverages the implicit preferences from unlabeled user-generated content (UGC) to generate high-quality preference data for aligning large language models.
Aligning VLM Assistants with Personalized Situated Cognition
Yongqi Li (Wuhan University), Tieyun Qian (Wuhan University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed the PCogAlignBench dataset and proposed the PCogAlign framework, enabling Vision-Language Models (VLMs) to generate contextualized personalized responses based on individual role sets.
AlignMMBench: Evaluating Chinese Multimodal Alignment in Large Vision-Language Models
Yuhang Wu (Tsinghua University), Yuxiao Dong (Zhipu AI)
Large Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextBenchmarkChain-of-Thought
🎯 What it does: Propose AlignMMBench, a Chinese multimodal alignment benchmark and its evaluator CritiqueVLM
All That Glitters is Not Novel: Plagiarism in AI Generated Research
Tarun Gupta (Indian Institute of Science Bengaluru), Danish Pruthi (Indian Institute of Science Bengaluru)
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper systematically evaluates plagiarism in research proposals generated by large language models, finding that approximately 24% to 36% of documents exhibit obvious plagiarism or serious citation misconduct.
Alleviating Distribution Shift in Synthetic Data for Machine Translation Quality Estimation
Xiang Geng (Nanjing University), Shujian Huang (Nanjing University)
Data SynthesisDomain AdaptationTransformerText
🎯 What it does: Utilizing translation references as supervisory signals, combined with constrained beam search to generate diverse synthetic translations, and annotating synthetic data through a dedicated annotator and the Shortest Phrase Coverage Error (SPCE) algorithm to alleviate the distribution shift problem in machine translation quality estimation.
Alleviating Hallucinations from Knowledge Misalignment in Large Language Models via Selective Abstention Learning
Lei Huang (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a novel training objective and inference strategy—SEAL, utilizing a dedicated [REJ] token to achieve selective rejection, thereby mitigating hallucinations in large models caused by knowledge alignment inconsistencies.
AmbiK: Dataset of Ambiguous Tasks in Kitchen Environment
Anastasia Ivanova, Aleksandr Panov
Data-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose and release the AmbiK dataset, which contains 1000 pairs of ambiguous and clear tasks in kitchen environments, and evaluate various LLM-based ambiguous detection methods on this benchmark.
Amplifying Trans and Nonbinary Voices: A Community-Centred Harm Taxonomy for LLMs
Eddie L. Ungless (University of Edinburgh), Remi Denton (Google Research)
ClassificationLarge Language ModelPrompt EngineeringText
🎯 What it does: Collaborated with the transgender and non-binary communities in the United States to build and release a community-centered harm classification system tailored for large language models (Transing Transformers Toolkit, T3).
An Effective Incorporating Heterogeneous Knowledge Curriculum Learning for Sequence Labeling
Xuemei Tang (Hong Kong Polytechnic University), Jinghang Gu (Hong Kong Polytechnic University)
ClassificationKnowledge DistillationGraph Neural NetworkText
🎯 What it does: Propose a two-stage curriculum learning framework to enhance the performance and training speed of sequence labeling models, particularly reducing training burden when fusing heterogeneous knowledge.
An Efficient and Precise Training Data Construction Framework for Process-supervised Reward Model in Mathematical Reasoning
Wei Sun (Institute of Automation, Chinese Academy of Sciences), Jiajun Zhang (Institute of Automation, Chinese Academy of Sciences)
Computational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Built an efficient and accurate process-supervised data construction framework called EpicPRM, and used it to generate the Epic50k dataset with 50k steps for training a Process-supervised Reward Model (PRM) for mathematical reasoning.
An Efficient Task-Oriented Dialogue Policy: Evolutionary Reinforcement Learning Injected by Elite Individuals
Yangyang Zhao (Changsha University of Science and Technology), Shihan Wang (Utrecht University)
Reinforcement LearningText
🎯 What it does: Proposed the EIERL algorithm, combining evolutionary algorithms with deep reinforcement learning, and accelerating task-oriented dialogue policy learning through an adaptive elite individual injection mechanism.