EMNLP 2025 Papers with AI Summaries
Conference on Empirical Methods in Natural Language Processing · 1809 papers
→ EMNLP 2025 papers with code (593)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 1809 EMNLP 2025 papers by keyword, author or institution —
free trial on arXivSub.
‘Rich Dad, Poor Lad’: How do Large Language Models Contextualize Socioeconomic Factors in College Admission ?
Huy Nghiem (University of Maryland), Hal Daumé III (University of Maryland)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextTabularChain-of-Thought
🎯 What it does: This paper constructs 30,000 synthetic applicant profiles based on real-world correlations from Common App, conducting a large-scale audit of university admissions using four open-source LLMs under two decision modes (rapid decision-making and chain-of-thought).
“Feels Feminine to Me”: Understanding Perceived Gendered Style through Human Annotations
Hongyu Chen (University of Stuttgart), Agnieszka Falenska (University of Stuttgart)
Explainability and InterpretabilityData-Centric LearningText
🎯 What it does: Constructed and annotated a new corpus, collecting 510 texts with 5,100 human five-point rating scores (ranging from very feminine to very masculine) for text gendered style, and analyzed the relationship between human cognition and text features, as well as annotators' sociodemographic characteristics.
“I’ve Decided to Leak”: Probing Internals Behind Prompt Leakage Intents
Jianshuo Dong (Tsinghua University), Han Qiu (Tsinghua University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Analyze whether large language models internally pre-encode prompt leakage intentions using probe technology and predict leakage risks before generation.
“Mm, Wat?” Detecting Other-initiated Repair Requests in Dialogue
Anh Ha Ngo (INRIA Paris), Chloé Clavel (CNRS)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelMultimodality
🎯 What it does: Propose a multimodal model to automatically detect the initiation segments of other-initiated repair (OIR) in Dutch task-oriented dialogues, enhancing detection performance by integrating manually designed linguistic and acoustic features with pre-trained embeddings.
(Almost) Free Modality Stitching of Foundation Models
Jaisidh Singh (University of Tübingen), Antonio Orvieto (ELLIS Institute Tübingen)
Hyperparameter SearchTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposed Hypernetwork Model Alignment (HYMA), a method that utilizes hypernetworks to generate and train all unimodal pair connectors in one go, thereby achieving efficient selection and concatenation of multimodal foundation models.
\mathrm{Wojood^{Relations}}: Arabic Relation Extraction Corpus and Modeling
Alaa Aljabari (Birzeit University), Mustafa Jarrar (Hamad Bin Khalifa University)
ClassificationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the largest Arabic relation extraction corpus, Wojood Relations, and proposed two new relation extraction methods: NLI-RE and GPT-Joint.
\textit{Do It Yourself (DIY)}: Modifying Images for Poems in a Zero-Shot Setting Using Weighted Prompt Manipulation
Sofia Jamil (Indian Institute of Technology Patna), Joseph K J (Adobe Research)
GenerationLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: Propose a weighted prompt manipulation technique for zero-shot poetry image generation, which dynamically adjusts the weights of prompts to control the attention distribution of diffusion models, thereby enhancing the semantic consistency and aesthetic quality of the visual output.
\texttt{Droid}: A Resource Suite for AI-Generated Code Detection
Daniil Orel, Preslav Nakov (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationData SynthesisAnomaly DetectionGraph Neural NetworkTransformerContrastive LearningTextBenchmark
🎯 What it does: Constructed a million-scale code dataset DroidCollection covering 43 generative models, 7 languages, and 3 programming domains, and trained two Encoder-Only detectors DroidDetect based on this dataset;
3DS: Medical Domain Adaptation of LLMs via Decomposed Difficulty-based Data Selection
Hongxin Ding (Peking University), Yasha Wang (Peking University)
Domain AdaptationTransformerSupervised Fine-TuningPrompt EngineeringBiomedical Data
🎯 What it does: Propose a two-stage model-centric data selection framework called 3DS for medical domain adaptation fine-tuning of LLMs.
3MDBench: Medical Multimodal Multi-agent Dialogue Benchmark
Ivan Sviridov (Sber AI Lab), Andrey Savchenko (Sber AI Lab)
Convolutional Neural NetworkLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Proposed 3MDBench—a multimodal, multi-agent telemedicine dialogue benchmark for evaluating the performance of large-scale vision-language models (LVLM) in diagnosis and communication.
3R: Enhancing Sentence Representation Learning via Redundant Representation Reduction
Longxuan Ma (kunming university of science and technology), Zhengtao Yu (kunming university of science and technology)
Representation LearningTransformerContrastive LearningText
🎯 What it does: To address redundant information in sentence representation learning, the 3R method is proposed: dynamically reducing redundancy in sentence vectors through constructing redundancy sentences based on high-frequency words, batch-level dimensional redundancy identification, and dimension-level redundancy reduction regularization.
A Case Against Implicit Standards: Homophone Normalization in Machine Translation for Languages that use the Ge’ez Script.
Hellina Hailu Nigatu (UC Berkeley), Seid Muhie Yimam (University of Hamburg)
Data-Centric LearningTransformerText
🎯 What it does: Investigated the impact of normalizing stem homonyms in languages using the Ge'ez script (such as Amharic, Tigrinya, and Ge'ez) on machine translation models, particularly during different stages including training, zero-shot, cross-lingual transfer, and post-inference normalization.
A Causal Lens for Evaluating Faithfulness Metrics
Kerem Zaman (UNC Chapel Hill), Shashank Srivastava (UNC Chapel Hill)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the CAUSAL DIAGNOSTICITY framework, which generates trustworthy/untrustworthy explanation pairs via knowledge editing to evaluate metrics for the credibility of natural language explanations.
A Comprehensive Framework to Operationalize Social Stereotypes for Responsible AI Evaluations
Aida Mostafazadeh Davani (Google Research), Vinodkumar Prabhakaran (Google Research)
TextGraph
🎯 What it does: Proposed a unified framework for systematically operationalizing social stereotypes to enable responsible AI detection and mitigation in generative AI assessments.
A Comprehensive Literary Chinese Reading Comprehension Dataset with an Evidence Curation Based Solution
Dongning Rao (Guangdong University of Technology), Zhihua Jiang (Jinan University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the largest-scale and most comprehensive ancient Chinese reading comprehension dataset CRISIS, and proposed the VIRTUAL solution based on evidence mining, option shuffling, and AMR sentence segmentation.
A Computational Simulation of Language Production in First Language Acquisition
Yuan Gao (University of Cambridge), Weiwei Sun (University of Cambridge)
GenerationRepresentation LearningTextGraph
🎯 What it does: This paper constructs a Synchronous Hyperedge Replacement Grammar (SHRG) from semantic graphs to a syntax-semantic interface to simulate children's language production, using unsupervised algorithms such as EM to induce statistical syntax from meaning graphs.
A Culturally-diverse Multilingual Multimodal Video Benchmark & Model
Bhuiyan Sanjid Shafique (Mohamed bin Zayed University of Artificial Intelligence), Fahad Shahbaz Khan (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Proposes a multilingual multicultural video LMM benchmark ViMUL-Bench and the corresponding multilingual video LMM model ViMUL, aiming to evaluate and enhance video understanding models' performance across 14 languages and 15 cultural/general domains.
A Fully Probabilistic Perspective on Large Language Model Unlearning: Evaluation and Optimization
Anda Cheng (Ant Group), Yinggui Wang (Ant Group)
OptimizationSafty and PrivacyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a fully probabilistic evaluation framework (FPE) and contrastive embedding loss (CEL) to more strictly evaluate and enhance the zero-shot effectiveness of LLMs.
A Generative Pre-Trained Language Model for Channel Prediction in Wireless Communications Systems
Bo Lin (Tsinghua University), Feifei Gao (Tsinghua University)
TransformerLarge Language ModelTabularTime SeriesBenchmarkPhysics Related
🎯 What it does: This paper proposes viewing wireless channels as words in a language, constructing a time-series channel sequence as a 'channel sentence,' and designing the CP-GPT model based on the Transformer decoder. It utilizes two specialized pre-training tasks (NCP and MCR) to learn the temporal evolution and feature reconstruction of channels, followed by verifying its channel prediction capability in scenarios such as few-shot, time slots, number of antennas, cross-frequency, and cross-antenna on multi-dimensional benchmarks.
A Good Plan is Hard to Find: Aligning Models with Preferences is Misaligned with What Helps Users
Nishant Balepur (University of Maryland), Jordan Lee Boyd-Graber (University of Maryland)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIText
🎯 What it does: Investigated the effectiveness of step-by-step instructions generated by large language models (LLMs) in helping users solve multi-step question-answering tasks, and evaluated whether common alignment signals such as user preferences, reward models, and agent simulations can accurately predict actual helpfulness.
A Graph-Theoretical Framework for Analyzing the Behavior of Causal Language Models
Rashin Rahnamoun (Shahid Beheshti University), Mehrnoush Shamsfard (Shahid Beheshti University)
Explainability and InterpretabilityRepresentation LearningTransformerTextBenchmark
🎯 What it does: Propose a framework based on graph theory, constructing a word transition sampling graph from the output of causal language models to perform macro-level structural analysis of the model's generation behavior.
A Head to Predict and a Head to Question: Pre-trained Uncertainty Quantification Heads for Hallucination Detection in LLM Outputs
Artem Shelmanov (MBZUAI), Timothy Baldwin (MBZUAI)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and pre-trained an auxiliary module (UQ Head) for claim-level hallucination detection in text generated by large language models, without requiring retraining or modification of the original model;
A Knowledge-driven Adaptive Collaboration of LLMs for Enhancing Medical Decision-making
Xiao Wu (Mohamed bin Zayed University of Artificial Intelligence), Yutong Xie (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelAgentic AIMixture of ExpertsTextBiomedical DataChain-of-Thought
🎯 What it does: Proposes the KAMAC framework, leveraging knowledge-driven dynamic multi-agent collaboration to enhance medical decision-making quality.
A Middle Path for On-Premises LLM Deployment: Preserving Privacy Without Sacrificing Model Confidentiality
Hanbo Huang (Shanghai Jiao Tong University), Shiyu Liang (Shanghai Jiao Tong University)
Safty and PrivacyKnowledge DistillationAdversarial AttackLarge Language ModelText
🎯 What it does: This paper proposes a semi-open framework called SOLID for locally deploying large language models (LLMs), which can provide high customizability while ensuring model confidentiality.
A Multi-Agent Framework with Automated Decision Rule Optimization for Cross-Domain Misinformation Detection
Hui Li (Xiamen University), Jinsong Su (Xiamen University)
Domain AdaptationAnomaly DetectionLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a multi-agent framework MARO that uses LLM agents to perform multi-dimensional analysis and automatically optimize decision rules for cross-domain misinformation detection;
A Multi-Level Benchmark for Causal Language Understanding in Social Media Discourse
Xiaohan Ding (Virginia Tech), Eugenia Rho (Virginia Tech)
ClassificationRecognitionGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the CausalTalk multi-level benchmark, which includes four annotation tasks: binary classification of Reddit COVID-19 discussion posts, explicit/implicit causal identification, causal span extraction, and causal gist generation, covering both gold and silver standard annotations.
A Multilingual, Culture-First Approach to Addressing Misgendering in LLM Applications
Sunayana Sitaram (Microsoft Research India), Si-Qing Chen
Safty and PrivacyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper develops language-specific misgendering guardrails for 42 languages and dialects based on participatory design, validates their effectiveness in meeting transcription summarization tasks, and constructs and publicly releases a synthetic dataset with human validation and evaluation data.
A Necessary Step toward Faithfulness: Measuring and Improving Consistency in Free-Text Explanations
Lingjun Zhao (University of Maryland), Hal Daumé III (University of Maryland)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a metric (PEX) to measure the consistency between free-text explanations and predictions, and utilizes Direct Preference Optimization (DPO) to train large language models, aiming to improve the consistency and credibility of generated explanations.
A Position Paper on the Automatic Generation of Machine Learning Leaderboards
Roelien C. Timmer (CSIRO Data61), Stephen Wan (CSIRO Data61)
TransformerLarge Language ModelPrompt EngineeringTabularReview/Survey PaperBenchmark
🎯 What it does: This paper provides a systematic review of research on automatically generating machine learning leaderboards (ALG), proposes a unified conceptual framework and evaluation criteria, and highlights existing challenges and future directions.
A Probabilistic Inference Scaling Theory for LLM Self-Correction
Zhe Yang (Peking University), Zhifang Sui (Peking University)
TransformerLarge Language ModelText
🎯 What it does: Propose a probabilistic inference scale theory to describe and predict the evolution of accuracy of large language models (LLMs) during multi-round self-correction, and verify the alignment between theoretical curves and empirical curves through experiments.
A Rigorous Evaluation of LLM Data Generation Strategies for Low-Resource Languages
Tatiana Anikina (German Research Institute for Artificial Intelligence), Simon Ostermann (German Research Institute for Artificial Intelligence)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This study systematically evaluates various strategies for text generation using large language models (LLMs) in low-resource language environments and compares their impact on downstream task performance.
A Sequential Multi-Stage Approach for Code Vulnerability Detection via Confidence- and Collaboration-based Decision Making
Chung-Nan Tsai (Lam Research Japan GK), Ching-Sheng Lin (Tunghai University)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose a sequential multi-stage code vulnerability detection framework called ConColl based on large language models, which dynamically switches between three reasoning modes: single agent, retrieval augmented generation (RAG), and multi-agent collaboration based on confidence.
A Simple Yet Effective Method for Non-Refusing Context Relevant Fine-grained Safety Steering in LLMs
Shaona Ghosh (NVIDIA), Christopher Parisien (NVIDIA)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed a reasoning-time safety guidance method called SAFESTEER based on activation vectors, which injects offsets into intermediate layers using category-level safety vectors to achieve fine-grained safety regulation of LLM outputs.
A Survey of Link Prediction in N-ary Knowledge Graphs
Jiyao Wei (University of Chinese Academy of Sciences), Xueqi Cheng (University of Chinese Academy of Sciences)
Graph Neural NetworkGraphReview/Survey Paper
🎯 What it does: Reviews and systematically evaluates the current state of research on link prediction in N-ary knowledge graphs (NKG), providing method classification, performance comparison, and application scenarios;
A Symbolic Adversarial Learning Framework for Evolving Fake News Generation and Detection
Chong Tian (MBZUAI), Xiuying Chen (MBZUAI)
GenerationData SynthesisAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringGenerative Adversarial NetworkText
🎯 What it does: Propose the Symbolic Adversarial Learning Framework (SALF), which co-evolves a generator and detector through adversarial symbolic learning to generate and detect more complex fake news.
A Systematic Analysis of Base Model Choice for Reward Modeling
Kian Ahrabian (University of Southern California), Jay Pujara (University of Southern California)
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Systematically analyzed the impact of base model selection, training stages, and pre-training data distribution on reward model performance, verifying that base model choices can improve reward model performance by up to 14% across 40 models of varying scales.
A Systematic Survey of Automatic Prompt Optimization Techniques
Kiran Ramnath (Amazon Web Services), Lin Lee Cheong (Amazon Web Services)
OptimizationLarge Language ModelReinforcement LearningPrompt EngineeringGenerative Adversarial NetworkImageTextMultimodalityReview/Survey PaperBenchmark
🎯 What it does: Conduct a systematic review of the Automatic Prompt Optimization (APO) field, proposing a five-dimensional unified framework, meticulously classifying and organizing existing methods, summarizing key technologies and evaluation metrics, and identifying future research directions.
A Text-Based Recommender System that Leverages Explicit Affective State Preferences
Tonmoy Hasan (University of North Carolina at Charlotte), Razvan Bunescu (University of North Carolina at Charlotte)
Recommendation SystemTransformerTextMultimodality
🎯 What it does: Built a text-based recommendation system ACRec that recommends books based on users' explicitly expressed fine-grained affective-cognitive (AC) preferences.
A Training-Free Length Extrapolation Approach for LLMs: Greedy Attention Logit Interpolation
Yan Li (University of Sydney), Caren Han (University of Melbourne)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a training-agnostic length extrapolation method called Greedy Attention Logit Interpolation (GALI), which mitigates the position out-of-distribution (O.O.D.) problem by greedily reallocating local position IDs and performing interpolation at the attention logit level when input length exceeds the training window.
AbsVis – Benchmarking How Humans and Vision-Language Models “See” Abstract Concepts in Images
Tarun Tater (University of Stuttgart), Sabine Schulte im Walde (University of Stuttgart)
Explainability and InterpretabilitySupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: This paper constructs the AbsVis dataset, containing 675 images and 225 corresponding abstract/middle/concrete nouns, and collects human and two visual language models (Qwen, Llava) concept-explanation pairs for each image. Further, 2,680 human preference judgments are obtained via Best-Worst Scaling (BWS). Subsequently, preference information is used to fine-tune models through Direct Preference Optimization (DPO) to improve alignment with human-preferred concept explanations.
Accelerate Parallelizable Reasoning via Parallel Decoding within One Sequence
Yijiong Yu (Tsinghua University), Ji Pei (OpenCSG)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a method that achieves parallel decoding using a tree-shaped attention mask within a single sequence, automatically identifying parallelizable inference steps. It generates only the titles first and then completes the subsequent content in parallel, significantly accelerating parallelizable tasks.
Accelerated Test-Time Scaling with Model-Free Speculative Sampling
Woomin Song (KAIST), Sravan Babu Bodapati (Amazon AGI)
Computational EfficiencyText
🎯 What it does: Propose STAND, a model-free acceleration method during inference.
AccessEval: Benchmarking Disability Bias in Large Language Models
Srikant Panda (Oracle AI), Hitesh Laxmichand Patel (Oracle AI)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper constructs the AccessEval benchmark, collecting neutral queries and disability-aware queries from 6 real-world scenarios and 9 disability types, and uses them to evaluate disability bias in 21 large language models.
ACING: Actor-Critic for Instruction Learning in Black-Box LLMs
Salma Kharrat (KAUST), Marco Canini
Large Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose the ACING Actor-Critic reinforcement learning framework for automating prompt optimization in black-box large language models, capable of searching infinite instruction spaces without gradient information and under limited query budgets.
AcT2I: Evaluating and Improving Action Depiction in Text-to-Image Models
Vatsal Malaviya (Arizona State University), Chitta Baral (Arizona State University)
GenerationData SynthesisKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageTextBenchmark
🎯 What it does: Developed the AcT2I benchmark to evaluate the capability of text-to-image models in action depiction, and enhanced prompts in three dimensions (spatial, emotional, and temporal) using LLM (GPT-4) to improve the accuracy of image generation.
ActionStudio: A Lightweight Framework for Data and Training of Large Action Models
Jianguo Zhang (Salesforce AI Research), Caiming Xiong (Salesforce AI Research)
Computational EfficiencyData-Centric LearningLarge Language ModelVision-Language-Action ModelTextSequentialBenchmark
🎯 What it does: Proposed and implemented ActionStudio, a lightweight and scalable data and training framework for constructing and training large-scale action models.
Active Layer-Contrastive Decoding Reduces Hallucination in Large Language Model Generation
Hongxiang Zhang (Purdue University), Tianyi Zhang (Purdue University)
GenerationTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Propose an Active Hierarchical Contrastive Decoding (ActLCD) based on reinforcement learning, dynamically deciding whether to activate hierarchical contrast during LLM generation to enhance factual accuracy.
AdamS: Momentum Itself Can Be A Normalizer for LLM Pretraining and Post-training
Huishuai Zhang (Wangxuan Institute of Computer Technology, Peking University), Luoxin Chen (Wangxuan Institute of Computer Technology, Peking University)
OptimizationComputational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed an optimizer called AdamS, which can replace AdamW for pretraining and post-training of large language models, eliminating the second moment estimation, significantly reducing memory usage and communication overhead.
Adapting Bias Evaluation to Domain Contexts using Generative Models
Tamara Quiroga (University of Chile), Valentin Barriere (National Center for Artificial Intelligence)
Data SynthesisDomain AdaptationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a framework that utilizes large language model (LLM) prompts to automatically rewrite traditional template-based bias assessment datasets into target domains (e.g., Twitter, Wikipedia Talk).
Adaptively profiling models with task elicitation
Davis Brown (University of Pennsylvania), Eric Wong (University of Pennsylvania)
Explainability and InterpretabilityTextChain-of-Thought
🎯 What it does: Proposes the Task Elicitation framework, which automatically generates and clusters natural language tasks using the target model's chain-of-thought reasoning (CoT), thereby describing the model's failure modes.
AdaptThink: Reasoning Models Can Learn When to Think
Jiajie Zhang (Tsinghua University), Juanzi Li (Tsinghua University)
Computational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: Studied dynamically selecting 'thinking' and 'no-thinking' modes in large-scale reasoning models to improve efficiency and performance.
AdaRewriter: Unleashing the Power of Prompting-based Conversational Query Reformulation via Test-Time Adaptation
Yilong Lai (Southeast University), Deyu Zhou (Southeast University)
RetrievalDomain AdaptationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes AdaRewriter, a framework that rewrites conversational queries using a result-supervised reward model in the best-N context and adapts during testing.
AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender
Weixiang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed an adaptive activation guidance method called AdaSteer, aimed at enhancing large language models (LLMs) defense against jailbreak attacks while maintaining their ability to process benign inputs.
Add-One-In: Incremental Sample Selection for Large Language Models via a Choice-Based Greedy Paradigm
Zhuo Li (Shenzhen International Center for Industrial and Applied Mathematics), Jinpeng Hu (Hefei University of Technology)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmark
🎯 What it does: This paper proposes a 'choice-based' greedy framework that directly evaluates the contribution of candidate samples to the selected subset during incremental iterations using large language models (LLMs), thereby selecting high-quality and diverse training samples, significantly reducing the need for full data traversal.
Addressing Tokenization Inconsistency in Steganography and Watermarking Based on Large Language Models
Ruiyi Yan (Kyoto University), Yugo Murawaki (Kyoto University)
TransformerLarge Language ModelText
🎯 What it does: This paper addresses the tokenization inconsistency (TI) problem that occurs during the transmission of steganographic texts and watermarked texts generated by large language models (LLMs), proposing a progressive verification method for steganography and a backward rollback method for watermarks.
Advancing Arabic Diacritization: Improved Datasets, Benchmarking, and State-of-the-Art Models
Abubakr Mohamed (Qatar Computing Research Institute, Hamad Bin Khalifa University), Hamdy Mubarak (Qatar Computing Research Institute, Hamad Bin Khalifa University)
RestorationRecurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: This paper proposes an improved method for Arabic diacritization, including quality analysis and cleaning of large-scale diacritized corpora, constructing a multi-reference evaluation system WikiNews-2024, and developing a BiLSTM model that preserves user-provided diacritics.
Advancing Fine-Grained Visual Understanding with Multi-Scale Alignment in Multi-Modal Models
Wei Wang (University of Science and Technology of China), Li Xiao (University of Science and Technology of China)
Data SynthesisRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a multi-scale fine-grained visual knowledge alignment method, systematically aligning the text descriptions, coordinates, and image information of objects, and gradually enhancing the model's localization and global understanding capabilities through a three-stage training process.
Advancing Oversight Reasoning across Languages for Audit Sycophantic Behaviour via X-Agent
Giulia Pucci (University of Aberdeen), Leonardo Ranaldi (University of Edinburgh)
Anomaly DetectionExplainability and InterpretabilityLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Propose the X-Agent framework, which monitors and corrects LLM's flattery behavior through a supervised reasoning layer.
Adversarial Attacks Against Automated Fact-Checking: A Survey
Fanzhen Liu (Macquarie University), Quan Z. Sheng (Macquarie University)
Adversarial AttackLarge Language ModelTextReview/Survey Paper
🎯 What it does: Provide a comprehensive review of adversarial attacks on automated fact-checking (AFC) systems, propose a unified attack classification framework, and summarize existing attack and defense techniques.
AesBiasBench: Evaluating Bias and Alignment in Multimodal Language Models for Personalized Image Aesthetic Assessment
Kun Li (City University of Hong Kong), Yuzhi Zhao (Hunan University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringImageMultimodalityBenchmark
🎯 What it does: Proposes the AesBiasBench benchmark to evaluate stereotypical biases and alignment of multi-modal large language models (MLLM) in personalized image aesthetic assessment (PIAA) tasks.
Africa Health Check: Probing Cultural Bias in Medical LLMs
Charles Nimo, Michael L. Best
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmark
🎯 What it does: Evaluated cultural bias in multiple medical large language models (LLMs) within the context of African traditional herbal medicine, constructed a question-answering dataset based on PubMed with over 130 country-herb pairs, and proposed a dual evaluation framework with a black-box Cultural Bias Score (CBS) and a white-box Cultural Bias Attribution (CBA).
AFRIDOC-MT: Document-level MT Corpus for African Languages
Jesujoba Oluwadara Alabi, Dietrich Klakow (Saarland University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the AFRIDOC-MT document-level multi-parallel translation corpus, covering English and five African languages (Amharic, Hausa, Swahili, Yorùbá, Zulu), spanning two domains (health and information technology), followed by benchmark experiments on multiple NMT and LLM models.
Agent-as-Judge for Factual Summarization of Long Narratives
Yeonseok Jeong (IPAI Seoul National University), Byung-Hak Kim (Hyundai Card)
GenerationGraph Neural NetworkTransformerLarge Language ModelAgentic AITextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Agent-as-Judge framework NARRATIVEFACTSCORE based on Character Knowledge Graph (CKG) to assess and improve the factual accuracy of long-form narrative summaries.
Agent-to-Agent Theory of Mind: Testing Interlocutor Awareness among Large Language Models
Younwoo Choi (University of Toronto), Zhijing Jin (MPI)
RecognitionLarge Language ModelText
🎯 What it does: Investigate interlocutor awareness in large language models, revealing their recognition capabilities and behavioral adaptability through systematic evaluation and three case studies.
Agentic-R1: Distilled Dual-Strategy Reasoning
Weihua Du (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
Knowledge DistillationTransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Propose the DualDistill framework, which combines trajectories generated by two complementary teachers (agentic and text-based reasoning) for combined distillation, and trains the student model Agentic‑R1 to dynamically select between using tools or pure text-based reasoning during inference;
AgentPro: Enhancing LLM Agents with Automated Process Supervision
Yuchen Deng, See-Kiong Ng (National University Of Singapore)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes the AgentPro framework, which utilizes Monte Carlo Tree Search (MCTS) to automatically generate step-by-step annotations. During the training process, a Process Reward Model (PRM) evaluates multi-step reasoning from the LLM in a fine-grained manner, and under a rejection sampling mechanism, the best reasoning path is selected to perform reinforcement learning fine-tuning on the LLM.
AI Argues Differently: Distinct Argumentative and Linguistic Patterns of LLMs in Persuasive Contexts
Esra Dönmez (University of Stuttgart), Agnieszka Falenska (University of Stuttgart)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Analyze and compare differences in language features and argument quality between human-generated and LLM-generated persuasive arguments, and construct a lightweight discriminator based on these interpretable features.
AI Chatbots as Professional Service Agents: Developing a Professional Identity
Wenwen Li (Fudan University), Yidong Chai (Hefei University of Technology)
TransformerLarge Language ModelTextBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the LAPI framework, which uses the Health Belief Model for task planning and employs practical entropy optimization to enable LLMs to generate responses consistent with professional identities in medical question-answering.
AI Knows Where You Are: Exposure, Bias, and Inference in Multimodal Geolocation with KoreaGEO
Xiaonan Wang (Yonsei University), Hansaem Kim (Yonsei University)
Safty and PrivacyVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes the KoreaGEO benchmark for evaluating the geolocation accuracy and privacy leakage risks of multimodal vision-language models in Korean street scenes.
AI Sees Your Location—But With A Bias Toward The Wealthy World
Jingyuan Huang (University of Georgia), Jieyu Zhao (University of Southern California)
RecognitionVision Language ModelImageBenchmark
🎯 What it does: This paper systematically evaluates the performance and bias of vision-language models (VLM) in image geolocation by constructing the FAIRLOCATOR benchmark dataset, revealing that models are more accurate in developed and densely populated areas and tend to predict large cities;
AIMMerging: Adaptive Iterative Model Merging Using Training Trajectories for Language Model Continual Learning
Yujie Feng (Tencent), Xiao-Ming Wu (Hong Kong Polytechnic University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose an adaptive iterative model merging framework (AimMerging) based on training trajectories, which dynamically schedules merging timing and frequency through learning and forgetting signals to achieve knowledge retention and transfer in LLM continuous learning.
AIP: Subverting Retrieval-Augmented Generation via Adversarial Instructional Prompt
Saket Sanjeev Chaturvedi (Clemson University), Xiaoyong Yuan (Clemson University)
RetrievalAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the AIP attack, which achieves the induction of retrieval bias toward malicious documents by maliciously rewriting the instructional prompts in RAG systems, without modifying user queries or the model's internal structure.
AIR: Complex Instruction Generation via Automatic Iterative Refinement
Wei Liu (Alibaba Group), Bo Zheng (Alibaba Group)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Proposed the Automatic Iterative Refinement (AIR) framework, which generates initial instructions from documents and iteratively adds constraints through an LLM evaluator to ultimately produce high-quality complex instructions.
Aligning Text/Speech Representations from Multimodal Models with MEG Brain Activity During Listening
Padakanti Srijith (Iiit Hyderabad), Subba Reddy Oota (Microsoft)
Representation LearningTransformerTextMultimodalityMagnetic Resonance ImagingAudio
🎯 What it does: Investigate the alignment between a multimodal (text + speech) pre-trained model and natural story auditory MEG recordings, comparing the brain prediction performance of multimodal and unimodal models at the text and speech embedding levels.
Alignment for Efficient Tool Calling of Large Language Models
Hongshen Xu (X-LANCE Lab, MoE Key Lab Of Artificial Intelligence, AI Institute), Kai Yu (X-LANCE Lab, MoE Key Lab Of Artificial Intelligence, AI Institute)
Computational EfficiencyLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: Propose a multi-objective alignment framework that integrates probabilistic knowledge boundary estimation with dynamic decision-making to enhance the tool calling efficiency of large language models.
Alignment Quality Index (AQI) : Beyond Refusals: AQI as an Intrinsic Alignment Diagnostic via Latent Geometry, Cluster Divergence, and Layer wise Pooled Representations
Abhilekh Borah (Manipal University Jaipur), Amitava Das (BITS Goa)
Safty and PrivacyRepresentation LearningTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: This paper proposes the Alignment Quality Index (AQI), a safety alignment evaluation metric based on the geometric properties of the model's internal representations, and creates the LITMUS dataset for validation.
Alignment with Fill-In-the-Middle for Enhancing Code Generation
Houxing Ren (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose StructureCoder, which generates fine-grained DPO training samples using Fill-In-The-Middle (FIM), and fully utilizes limited test cases through AST chunking and curriculum training.
Alignment-Augmented Speculative Decoding with Alignment Sampling and Conditional Verification
Jikai Wang (Soochow University), Min Zhang (Soochow University)
GenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposed a training-free alignment-enhanced speculative decoding algorithm called AASD, which includes alignment sampling and conditional verification.
AlignX: Advancing Multilingual Large Language Models with Multilingual Representation Alignment
Mengyu Bu (Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences), Yang Feng (Baidu Inc)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: To address the performance gap of multilingual large language models on non-mainstream languages, we propose the AlignX two-stage representation framework. It first aligns model representations through multilingual semantic alignment at the intermediate layer, then preserves language-specific features by integrating language features at the output layer, followed by multilingual instruction fine-tuning to enhance cross-lingual understanding and generation capabilities.
All for One: LLMs Solve Mental Math at the Last Token With Information Transferred From Other Tokens
Siddarth Mamidanna (University of California Santa Cruz), Yilun Zhou (George Mason University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper conducts fine-grained experiments on the 'mental arithmetic' task of large language models (LLMs), discovering and verifying a minimalist computational subgraph—All-for-One (AF1). This subgraph performs input-specific arithmetic operations only on the last token, while other tokens perform task-generic computations and pass information to the final token in a few layers.
All Roads Lead to Rome: Graph-Based Confidence Estimation for Large Language Model Reasoning
Caiqi Zhang (University of Cambridge), Nigel Collier (University of Cambridge)
Explainability and InterpretabilityLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed a no-training, graph-based confidence estimation method that constructs a directed graph using multiple reasoning chains and quantifies the confidence in the reasoning results of large language models (LLMs) from graph structures (centrality, path convergence, path weights).
ALLabel: Three-stage Active Learning for LLM-based Entity Recognition using Demonstration Retrieval
Zihan Chen (Beihang University), Yue Zhang (Westlake University)
RecognitionRetrievalLarge Language ModelText
🎯 What it does: Proposes the ALLabel framework, which selects the most representative and informative samples through a three-stage active learning process (diversity, similarity, confidence) to build example retrieval corpora for LLMs, thereby achieving efficient domain entity recognition;
AlphaOne: Reasoning Models Thinking Slow and Fast at Test Time
Junyu Zhang (University of Illinois Urbana Champaign), Huan Zhang (University of Illinois Urbana Champaign)
Computational EfficiencyLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposes the ALPHAONE framework, achieving generic regulation of inference progress during the LRM inference phase through α-moment, expected slow thinking scheduling, and post-fast thinking termination.
AMACE: Automatic Multi-Agent Chart Evolution for Iteratively Tailored Chart Generation
Hyuk Namgoong (Chungnam National University), Sangkeun Jung (Chungnam National University)
GenerationLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityTabular
🎯 What it does: This study proposes the AMACE (Automatic Multi-Agent Chart Evolution) framework, which utilizes multi-agent cyclic collaboration to achieve automatic generation and iterative optimization of charts from tables.
Ambiguity Awareness Optimization: Towards Semantic Disambiguation for Direct Preference Optimization
Jian Li (Tencent), Pengfei Xu (Tencent)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Studied the optimization conflicts in Direct Preference Optimization (DPO) caused by semantically similar (ambiguous) content, and proposed Ambiguity Awareness Optimization (AAO), which alleviates the problem by automatically reweighting ambiguous, transitional, and critical tokens.
AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection
Yejin Lee (Yonsei University), Yo-Sub Han (Yonsei University)
ClassificationTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose the AmpleHate method, which identifies explicit targets through pre-trained NER, uses [CLS] to represent implicit targets, calculates attention relationships between targets and context, and directly injects these relationships into sentence representations to achieve implicit hate speech detection.
AMQ: Enabling AutoML for Mixed-precision Weight-Only Quantization of Large Language Models
Sangjun Lee (Pohang University of Science and Technology), Eunhyeok Park (Pohang University of Science and Technology)
OptimizationComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the AMQ framework to achieve automated mixed-precision weight quantization, quickly identifying the optimal bit-width configuration under a given memory budget.
Amulet: Putting Complex Multi-Turn Conversations on the Stand with LLM Juries
Sahana Ramnath (University of Southern California), Xiang Ren
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose the AMULET framework, which leverages dialogue acts (DA) and Gricean maxims (MAXIM) to conduct more precise evaluation of the quality and shortcomings of responses in multi-turn human-computer dialogues.
An Empirical Study of LLM Reasoning Ability Under Strict Output Length Constraint
Yi Sun (Tsinghua University), Yunxin Liu (Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Systematically evaluate the performance of 30 large language models on mathematical reasoning tasks under strict output length constraints, comparing direct truncation and early stopping as two length control strategies.
An Empirical Study on Strong-Weak Model Collaboration for Repo-level Code Generation
Shubham Gandhi (Carnegie Mellon University), Carolyn Rose (Carnegie Mellon University)
AI Code AssistantTransformerTextRetrieval-Augmented Generation
🎯 What it does: Studied how to collaborate powerful but expensive language models with cheaper weak models in code generation tasks to reduce overall generation costs without significantly sacrificing performance.
An Interdisciplinary Approach to Human-Centered Machine Translation
Marine Carpuat (University Of Maryland), François Yvon (Sorbonne-Université)
Review/Survey Paper
🎯 What it does: This paper reviews and integrates translation studies with human-computer interaction (HCI) theory, proposing a human-centered machine translation (MT) research framework, and illustrates practical pathways using clinical scenarios as examples.
An Orthogonal High-Rank Adaptation for Large Language Models
Xin Zhang (South China University of Technology), Tong Zhang (South China University of Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose OHoRA, an orthogonal high-order adaptive method based on the information bottleneck theory, which eliminates redundant information in pre-trained weights using QR decomposition and dynamically enhances the rank of the update matrix through Kronecker product, thereby achieving efficient, low-redundancy parameterized fine-tuning.
Analysing Chain of Thought Dynamics: Active Guidance or Unfaithful Post-hoc Rationalisation?
Samuel Lewis-Lim (University of Sheffield), Nikolaos Aletras (University of Sheffield)
Explainability and InterpretabilityKnowledge DistillationLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper dynamically and credibly analyzes the Chain-of-Thought (CoT) in instruction-tuned, reinforcement learning, and distillation-based reasoning models for soft reasoning tasks, exploring the role and reliability of CoT across different models;
Analyzing and Modeling LLM Response Lengths with Extreme Value Theory: Anchoring Effects and Hybrid Distributions
Liuxuan Jiao (University of Cambridge), Yong Li (Tsinghua University)
TransformerLarge Language ModelText
🎯 What it does: Analyze and model the response lengths of large language models (LLMs) under different temperatures and prompting strategies, proposing a GEV-GPD hybrid model based on extreme value theory;
Analyzing the Effects of Supervised Fine-Tuning on Model Knowledge from Token and Parameter Levels
Junjie Ye (Fudan University), Jianping Fan (Lenovo Research)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates the impact of supervised fine-tuning (SFT) on the knowledge of large language models (LLMs), particularly conducting in-depth analysis at both the token and parameter levels in closed-book question answering (CBQA) tasks.
Analyzing Uncertainty of LLM-as-a-Judge: Interval Evaluations with Conformal Prediction
Huanxin Sheng (University of Rochester), Jian Kang (MBZUAI)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Analyze the uncertainty of LLM judges and propose a conformational prediction-based interval evaluation framework
Analyzing values about gendered language reform in LLMs’ revisions
Jules Watson (University of Toronto), Barend Beekhuizen (University of Toronto)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigate how large language models replace gendered role nouns in text revision tasks and the reasons behind their choices, exploring the impact of contextual factors (such as the gender of the person being revised, the explicitness of gender information, and the sentence's gender context) on model decisions.
AnchorAttention: Difference-Aware Sparse Attention with Stripe Granularity
Yu Zhang (Xiamen University), Yiming Zhang (Shanghai Jiao Tong University)
Computational EfficiencyTransformerText
🎯 What it does: Propose AnchorAttention, a difference-aware, stripe-granularity dynamic sparse attention mechanism, to reduce computational costs during the prefill stage of LLMs with long-sequence contexts.
Anchoring-Guidance Fine-Tuning (AnGFT): Elevating Professional Response Quality in Role-Playing Conversational Agents
Qibin Li (Dalian University of Technology), Baoxun Wang (Peking University)
TransformerSupervised Fine-TuningPrompt EngineeringTextFinance Related
🎯 What it does: This paper proposes the Anchoring-Guidance Fine-Tuning (AnGFT) framework, which combines anchoring system prompts with diverse prompts to perform two-stage fine-tuning on role-playing dialogue models, aiming to improve their answer quality in professional domains.
Anecdoctoring: Automated Red-Teaming Across Language and Place
Alejandro Cuevas (Carnegie Mellon University), Madeleine I. G. Daepp (Microsoft Research)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed and implemented an automatic red teaming method called anecdoctoring, which generates adversarial prompts against cross-lingual and cross-regional generative AI by leveraging real-world fact-checking data; the approach clusters to identify major narratives, constructs knowledge graphs, and then generates adversarial prompts using enhanced LLMs; the code is publicly available in the PyRIT framework.