These 593 EMNLP 2025 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every EMNLP 2025 paper, free trial on arXivSub.
“I’ve Decided to Leak”: Probing Internals Behind Prompt Leakage Intents
Jianshuo Dong (Tsinghua University), Han Qiu (Tsinghua University)
CodeSafty 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)
CodeClassificationExplainability 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)
CodeHyperparameter 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.
3MDBench: Medical Multimodal Multi-agent Dialogue Benchmark
Ivan Sviridov (Sber AI Lab), Andrey Savchenko (Sber AI Lab)
CodeConvolutional 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.
🎯 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 Comprehensive Literary Chinese Reading Comprehension Dataset with an Evidence Curation Based Solution
Dongning Rao (Guangdong University of Technology), Zhihua Jiang (Jinan University)
CodeTransformerLarge 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 Position Paper on the Automatic Generation of Machine Learning Leaderboards
Roelien C. Timmer (CSIRO Data61), Stephen Wan (CSIRO Data61)
CodeTransformerLarge 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 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)
CodeGraph 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;
🎯 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)
CodeTransformerLarge 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.
Accelerate Parallelizable Reasoning via Parallel Decoding within One Sequence
Yijiong Yu (Tsinghua University), Ji Pei (OpenCSG)
CodeComputational 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.
ACING: Actor-Critic for Instruction Learning in Black-Box LLMs
Salma Kharrat (KAUST), Marco Canini
CodeLarge 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.
ActionStudio: A Lightweight Framework for Data and Training of Large Action Models
Jianguo Zhang (Salesforce AI Research), Caiming Xiong (Salesforce AI Research)
CodeComputational 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.
Davis Brown (University of Pennsylvania), Eric Wong (University of Pennsylvania)
CodeExplainability 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.
AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender
Weixiang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
CodeSafty 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.
Addressing Tokenization Inconsistency in Steganography and Watermarking Based on Large Language Models
Ruiyi Yan (Kyoto University), Yugo Murawaki (Kyoto University)
CodeTransformerLarge 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.
Africa Health Check: Probing Cultural Bias in Medical LLMs
Charles Nimo, Michael L. Best
CodeExplainability 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).
Agent-as-Judge for Factual Summarization of Long Narratives
Yeonseok Jeong (IPAI Seoul National University), Byung-Hak Kim (Hyundai Card)
CodeGenerationGraph 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.
Weihua Du (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
CodeKnowledge 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;
AIMMerging: Adaptive Iterative Model Merging Using Training Trajectories for Language Model Continual Learning
Yujie Feng (Tencent), Xiao-Ming Wu (Hong Kong Polytechnic University)
CodeTransformerLarge 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.
🎯 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 with Fill-In-the-Middle for Enhancing Code Generation
Houxing Ren (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
CodeAI 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.
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)
CodeExplainability 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.
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)
CodeOptimizationComputational 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.
Anchoring-Guidance Fine-Tuning (AnGFT): Elevating Professional Response Quality in Role-Playing Conversational Agents
Qibin Li (Dalian University of Technology), Baoxun Wang (Peking University)
CodeTransformerSupervised 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.
Answering Narrative-Driven Recommendation Queries via a Retrieve–Rank Paradigm and the OCG-Agent
Yunxiao Shi (University of Technology Sydney), Min Xu (University of Technology Sydney)
CodeRecommendation SystemLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a retrieval-ranking framework for narrative-driven recommendation, and develop OCG-Agent to achieve wide and deep candidate retrieval.
AnyMAC: Cascading Flexible Multi-Agent Collaboration via Next-Agent Prediction
Song Wang (University of Virginia), Jundong Li (University of Virginia)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose the ANYMAC framework, which employs a sequential communication protocol based on LLM for multi-agent collaboration, dynamically predicts the next agent (Next-Agent Prediction) and context (Next-Context Selection), and constructs task-adaptive communication links.
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
Zhuo Li (Chinese University of Hong Kong, Shenzhen), Xiang Wan (Chinese University of Hong Kong, Shenzhen)
CodeReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Designed an adaptive marginal mechanism based on Optimal Transport to enhance the robustness of reward models in distinguishing human preferences;
Are Language Models Consequentialist or Deontological Moral Reasoners?
Keenan Samway (Max Planck Institute for Intelligent Systems), Zhijing Jin (Max Planck Institute for Intelligent Systems)
CodeAutonomous DrivingExplainability and InterpretabilityLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper systematically analyzes the moral reasoning trajectories of large language models (LLMs) by providing over 600 rewritten autonomous vehicle moral dilemmas (Trolley Problem) and prompting the models to generate their thinking processes;
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper constructs the CHIME dataset to explain Chinese internet emoticons and evaluate the understanding ability of large language models (LLMs) without fine-tuning.
Are Stereotypes Leading LLMs’ Zero-Shot Stance Detection ?
Anthony Dubreuil (Université de Saint-Etienne), Amine Trabelsi (Université de Sherbrooke)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically evaluates and quantifies the implicit biases of large language models toward dialects (AAE vs SAE) and text readability (complexity) in zero-shot stance detection.
Artificial Impressions: Evaluating Large Language Model Behavior Through the Lens of Trait Impressions
Nicholas Deas (Columbia University), Kathleen McKeown (Columbia University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study proposes and evaluates the LLM's 'human impressions' of prompt authors by decoding warmth and capability dimensions through linear probes in the hidden layer.
Assessing effective de-escalation of crisis conversations using transformer-based models and trend statistics
Ignacio J. Tripodi (Crisis Text Line), Elizabeth A. Olson (Crisis Text Line)
CodeClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Developed and trained a BERT-based sentiment valuation scoring model (BERT-EV), which numerically evaluates the sentiment of each message in crisis text conversations, and quantifies the emotional de-escalation level and predicts clinical outcomes by analyzing sentiment trends over time using the Mann-Kendall test.
🎯 What it does: Proposes Astra Transformer, a specialized model for efficient multimodal trajectory modeling, incorporating trajectory attention and learnable action queries, along with contrastive dynamics learning to enhance environmental dynamic understanding and multimodal alignment.
Attacking Misinformation Detection Using Adversarial Examples Generated by Language Models
Piotr Przybyła (Universitat Pompeu Fabra), Horacio Saggion (Universitat Pompeu Fabra)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study proposes the TREPAT framework, which generates semantically preserved rewrites using large language models (LLMs), decomposes them into atomic edits, and employs beam search to generate adversarial examples for misinformation detection classifiers within a limited number of queries.
Attacks by Content: Automated Fact-checking is an AI Security Issue
Michael Sejr Schlichtkrull (Queen Mary University of London)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper explains how attackers implement attacks by providing misleading or forged content to retrieval-augmented agents, and proposes integrating automated fact-checking technology into a complete self-defense pipeline for agents.
Audio-centric Video Understanding Benchmark without Text Shortcut
Yudong Yang (Tsinghua University), Chao Zhang (ByteDance)
CodeLarge Language ModelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes the AVUT audio-centric video understanding benchmark, designing multiple tasks focusing on audio content and audio-visual alignment, and eliminating text shortcuts through an answer permutation filtering mechanism.
Automated Knowledge Graph Construction using Large Language Models and Sentence Complexity Modelling
Sydney Anuyah (Indiana University), Sunandan Chakraborty (Indiana University)
CodeRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBiomedical DataChain-of-Thought
🎯 What it does: Propose CoDe-KG end-to-end open-source pipeline, automatically constructing knowledge graphs from medical literature through coreference resolution, sentence type classification, sentence simplification, and relation extraction, while publicly releasing a 150k triple dataset.
Automating Steering for Safe Multimodal Large Language Models
Lyucheng Wu (Zhejiang University), Shumin Deng (National University of Singapore NUS NCS Joint Lab)
CodeSafty and PrivacyLarge Language ModelContrastive LearningMultimodality
🎯 What it does: Developed AutoSteer, which automatically performs safe regulation for multimodal large language models during inference, achieving safe intervention without fine-tuning by leveraging safety awareness scores, toxicity detectors, and rejection heads.
AutoSDT: Scaling Data-Driven Discovery Tasks Toward Open Co-Scientists
Yifei Li (University of Wisconsin-Madison), Huan Sun (University of Wisconsin-Madison)
CodeAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the AutoSDT automated pipeline for large-scale collection of real scientific code generation tasks across four disciplines (bioinformatics, chemistry, geographic information science, psychology and cognitive neuroscience), constructing AutoSDT-5K (5,404 tasks)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: In the 2024 BabyLM challenge, a masked language model trained with 100M words (or fewer) is used, employing the causal probing method (global/local affinity) proposed by Rozner et al. to evaluate the model's mastery of various structural syntax (e.g., causal remainder structures, fixed slots, metaphors, and literal usages).
BANMIME : Misogyny Detection with Metaphor Explanation on Bangla Memes
Md Ayon Mia (Dhaka International University), Akmmahbubur Rahman
CodeClassificationExplainability and InterpretabilityData-Centric LearningSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Constructed the BANMIME dataset, containing 2000 Bangla memes, with multi-modal annotations (types of gender discrimination, humor types, metaphor localization, and explanations) and evaluated multiple vision-language models.
Batched Self-Consistency Improves LLM Relevance Assessment and Ranking
Anton Korikov (University of Toronto), Navid Rekabsaz (University of Toronto)
CodeRetrievalLarge Language ModelText
🎯 What it does: This paper studies batched pointwise evaluation and ranking methods in large language models (LLMs), combined with self-consistency techniques, to enhance relevance judgment and sorting performance in retrieval tasks.
Benchmarking LLMs for Translating Classical Chinese Poetry: Evaluating Adequacy, Fluency, and Elegance
Andong Chen, Min Zhang (Harbin Institute of Technology)
CodeGenerationTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a retrieval-enhanced translation method RAT and GPT-4 evaluation metrics to assess and improve the translation of classical Chinese poetry into English
Benchmarking LLMs on Semantic Overlap Summarization
John Salvador (University of Central Florida), Santu Karmaker (University of Central Florida)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Evaluate the performance of large language models on the Semantic Overlap Summary (SOS) task and introduce a new privacy policy pair dataset (PrivacyPolicyPairs, 3P).
BeSimulator: A Large Language Model Powered Text-based Behavior Simulator
Jianan Wang (National University of Defense Technology), Lihanxun
CodeRobotic IntelligenceLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented BeSimulator, a text environment behavior simulation framework based on large language models (LLMs), for pre-validation of robot behavior planning; the framework generates text scenarios, simulates actions step-by-step according to control logic, and evaluates whether task goals are achieved;
Beyond Averages: Learning with Annotator Disagreement in STS
Alejandro Benito-Santos (Universidad Nacional de Educacion a Distancia), Adrian Ghajari
CodeRetrievalTransformerText
🎯 What it does: This paper proposes two methods to model annotator inconsistency in semantic text similarity tasks, rather than simply taking the average, thereby improving the model's calibration and ranking performance in the face of human uncertainty.
Beyond Checkmate: Exploring the Creative Choke Points for AI Generated Texts
Nafis Irtiza Tripto (Pennsylvania State University), Dongwon Lee (Pennsylvania State University)
CodeClassificationLarge Language ModelText
🎯 What it does: Explore the subtle differences between human writing and LLM-generated text in three paragraphs (introduction, body, and conclusion), and evaluate the contribution of these paragraphs to AI text detection.
Beyond Human Labels: A Multi-Linguistic Auto-Generated Benchmark for Evaluating Large Language Models on Resume Parsing
Zijian Ling (Apply U UK), Xiangjian He (University of Nottingham Ningbo)
CodeData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the first multilingual, structured resume benchmark ResumeBench, containing 2,500 synthetic resumes manually reviewed for compliance, covering 50 templates, 30 career fields, and 5 languages;
CodeReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Propose a dynamic data transformation method that converts offline aligned data into online equivalent data, achieving LLM alignment without requiring a reward model.
Beyond Outlining: Heterogeneous Recursive Planning for Adaptive Long-form Writing with Language Models
Ruibin Xiong (Independent Researcher), Jürgen Schmidhuber (Center of Excellence for Generative AI KAUST)
CodeGenerationLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a general-purpose long-form writing framework called WriteHERE, which achieves adaptive writing through recursive task decomposition and dynamic integration of retrieval, reasoning, and creation tasks.
Beyond Task-Oriented and Chitchat Dialogues: Proactive and Transition-Aware Conversational Agents
Yejin Yoon (Hanyang University), Taeuk Kim (Hanyang University)
CodeData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextSequential
🎯 What it does: Constructed a dialogue dataset TACT containing a mix of multi-turn tasks and casual conversations with recoverable transfers, and trained a model to realize an active and mode-aware dialogue agent.
Beyond Text: Unveiling Privacy Vulnerabilities in Multi-modal Retrieval-Augmented Generation
Jiankun Zhang (Jilin University), Yi Chang (Michigan State University)
CodeRetrievalSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringImageMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: This paper systematically evaluates the privacy risks of multi-modal retrieval-augmented generation (MRAG) systems and proposes an attack method that can be conducted in a black-box environment through composite structured prompts, revealing direct and indirect privacy leakage in both visual and speech modalities.
Beyond the Leaderboard: Understanding Performance Disparities in Large Language Models via Model Diffing
Sabri Boughorbel (Qatar Computing Research Institute, HBKU), Majd Hawasly (Qatar Computing Research Institute, HBKU)
CodeSafty and PrivacyExplainability and InterpretabilityRepresentation LearningTransformerAuto EncoderText
🎯 What it does: Using model diffing and crosscoders methods, conduct a fine-grained analysis of the latent representations of Gemma-2-9b-it and its SimPO-enhanced version, revealing specific capability differences in safety, instruction following, and multilingual aspects.
Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations
Giorgos Filandrianos (National Technical University of Athens), Giorgos Stamou (National Technical University of Athens)
CodeRecommendation SystemAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: By embedding cognitive biases (such as social proof, scarcity, exclusivity, etc.) into product descriptions, the study investigates their impact on product recommendation systems driven by large language models (LLMs).
Bit-Flip Error Resilience in LLMs: A Comprehensive Analysis and Defense Framework
Yuhang Chen (University of North Carolina at Chapel Hill), Tianlong Chen (University of North Carolina at Chapel Hill)
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically analyzes the impact of bit-flip errors on large language models (LLMs) and proposes the FlipGuard defense framework.
Boosting Multi-modal Keyphrase Prediction with Dynamic Chain-of-Thought in Vision-Language Models
Qihang Ma (ByteDance Douyin Content Group), Jiao Ran (ByteDance Douyin Content Group)
CodeSupervised Fine-TuningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: This paper explores the application of vision-language models (VLM) to multimodal key phrase prediction (MMKP), proposing three strategies: zero-shot learning, supervised fine-tuning, Fine-tune-CoT (Chain-of-Thought fine-tuning), and dynamic chain-of-thought (Dynamic CoT), aiming to enhance the model's reasoning and generalization capabilities in scenarios with missing text information and unseen phrases.
BrailleLLM: Braille Instruction Tuning with Large Language Models for Braille Domain Tasks
Tianyuan Huang (Zhejiang University), Jiajun Bu (Zhejiang University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose BrailleLLM, a large language model using instruction fine-tuning, specialized in bidirectional Braille and mixed text (including formulas) conversion;
Breaking the Noise Barrier: LLM-Guided Semantic Filtering and Enhancement for Multi-Modal Entity Alignment
Chenglong Lu (Northeastern University), Fu Zhang (Northeastern University)
CodeConvolutional Neural NetworkGraph Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose the LGEA framework to address the noise problem in multi-modal entity alignment. First, use LLM for semantic-level visual filtering and attribute summarization, then perform multi-modal fusion and alignment.
BRSpeech-DF: A Deep Fake Synthetic Speech Dataset for Portuguese Zero-Shot TTS
Alexandre Costa Ferro Filho (Advanced Knowledge Center in Immersive Technologies), Arlindo Rodrigues Galvão Filho (Advanced Knowledge Center in Immersive Technologies)
🎯 What it does: Constructed and publicly released the BRSpeech-DF dataset for Portuguese speech audio deepfake detection, containing real read speech and multi-model zero-shot text-to-speech synthesized speech.
BTC-SAM: Leveraging LLMs for Generation of Bias Test Cases for Sentiment Analysis Models
Zsolt T. Kardkovács (Dublin City University), Walid Gaaloul (Télécom SudParis)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed the BTC-SAM framework, which automatically generates diverse bias test cases for evaluating social bias in sentiment analysis models using large language models with only minimal specifications provided.
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: Constructed an 8.9 billion-word medical text corpus HC4, and trained nine small-scale models (124–179M parameters) including GPT-2, Llama-3, and Mistral on HC4, SlimPajama, and FineWeb; subsequently conducted general bias evaluation (BOLD) and medical-specific bias evaluation (opioid prescription tendencies across different races, genders, and ages), proposing the Net Bias Prescription Score (NBPS) metric;
BYOKG-RAG: Multi-Strategy Graph Retrieval for Knowledge Graph Question Answering
Costas Mavromatis, George Karypis (Amazon)
CodeRetrievalLarge Language ModelGraphRetrieval-Augmented Generation
🎯 What it does: Propose the BYOKG-RAG framework, which leverages LLM to generate graph retrieval artifacts such as entities, paths, and queries, combined with multiple specialized graph retrieval tools for iterative context retrieval, ultimately producing knowledge graph question-answering results.
Cache-of-Thought: Master-Apprentice Framework for Cost-Effective Vision Language Model Reasoning
Mingyuan Wu (University of Illinois Urbana Champaign), Klara Nahrstedt (University of Illinois Urbana Champaign)
CodeComputational EfficiencyKnowledge DistillationVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Cache-of-Thought (CoT) framework, which stores answers generated by a large VLM (master) in a cache, and subsequently enhances the reasoning quality of a small VLM (apprentice) by using multi-modal retrieval and context learning to retrieve similar QA pairs from the cache as prompts.
Cacheback: Speculative Decoding With Nothing But Cache
Zhiyao Ma (Yale University), Lin Zhong (Yale University)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes Cacheback Decoding, a training- and model-agnostic inference acceleration method that leverages n-gram records in LRU cache tables to generate draft sequences, thereby achieving Speculative Decoding.
CAIR: Counterfactual-based Agent Influence Ranker for Agentic AI Workflows
Amit Giloni (Fujitsu Research of Europe), Roman Vainshtein (Fujitsu Research of Europe)
CodeExplainability and InterpretabilityComputational EfficiencyAdversarial AttackLarge Language ModelAgentic AIText
🎯 What it does: Proposes the CAIR method, which uses adversarial analysis to evaluate the impact of each agent on the final output in multi-agent workflows.
Yunzhi Yao (Zhejiang University), Nanyun Peng (University of California, Los Angeles)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the Circuit-aware Knowledge Editing (CaKE) method, which analyzes the internal reasoning circuits of LLMs and combines circuit-aware training data to ensure that edited knowledge is correctly utilized in multi-hop reasoning.
Calibrating LLM Confidence by Probing Perturbed Representation Stability
Reza Khanmohammadi (Michigan State University), Mohammad M. Ghassemi (Michigan State University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a CCPS method based on target adversarial perturbation of the final hidden states of LLMs and extracting stability features, achieving confidence estimation of LLM answer correctness with a lightweight classifier;
Can Large Language Models Act as Ensembler for Multi-GNNs?
Hanqi Duan (East China Normal University), Xiang Li (East China Normal University)
CodeClassificationGraph Neural NetworkLarge Language ModelSupervised Fine-TuningPrompt EngineeringGraphBenchmark
🎯 What it does: Propose LensGNN, which utilizes a large language model (LLM) as an integrator for multiple GNNs, achieving deep fusion of node attribute text and graph structural information through two-stage alignment between multiple GNNs and LLM, to complete node and graph classification tasks.
Can Large Language Models be Effective Online Opinion Miners?
Ryang Heo (Yonsei University), Dongha Lee (Yonsei University)
CodeData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the OOMB benchmark to evaluate the opinion mining capability of large language models in diverse, real online texts.
Can Large Language Models Translate Spoken-Only Languages through International Phonetic Transcription?
Jiale Chen (South China Normal University), Tianyong Hao (South China Normal University)
CodeTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the UNILANG framework, which utilizes large language models (LLMs) to translate unwritten writing systems (e.g., Bai language) through intermediary text in the International Phonetic Alphabet (IPA);
Can Large Language Models Translate Unseen Languages in Underrepresented Scripts?
Dianqing Lin (Inner Mongolia University), Guodong Shi (Inner Mongolia University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Lotus benchmark to evaluate the translation ability of large language models on low-resource script languages (Mongolian traditional script and I language) that the models have not been exposed to.
Can LLMs be Good Graph Judge for Knowledge Graph Construction?
Haoyu Huang (Hong Kong University of Science and Technology), Wentao Zhang (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the GraphJudge framework, leveraging large language models (LLMs) to act as a 'graph judge' during knowledge graph (KG) construction. It first denoises entity-centric text and extracts candidate triplets, then uses a fine-tuned open-source LLM to determine the authenticity of triplets, filtering noise, errors, and hallucinations to ultimately generate high-quality KGs.
Cardiverse: Harnessing LLMs for Novel Card Game Prototyping
Danrui Li (Rutgers University), Mubbasir Kapadia (Roblox)
CodeGenerationData SynthesisTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Construct a complete card game prototype generation pipeline by extracting game mechanics through graph structures, generating code with LLM, and building game AI based on heuristic functions.
CARE: A Disagreement Detection Framework with Concept Alignment and Reasoning Enhancement
Jiyuan Liu (Sun Yat-sen University), Yanghui Rao (Sun Yat-sen University)
CodeClassificationData SynthesisExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the CARE framework, consisting of two stages: Concept Alignment (aligning LLM and expert conceptual understanding through a sub-concept vocabulary) and Reasoning Enhancement (enhancing LLM's causal reasoning ability via a curriculum-based reasoning process, including rationale-to-critique and counterfactual-to-detection).
🎯 What it does: Propose the CARFT method, combining contrastive learning with annotated Chain-of-Thought (CoT) based reinforcement fine-tuning to enhance LLM reasoning performance.
CARMA: Enhanced Compositionality in LLMs via Advanced Regularisation and Mutual Information Alignment
Nura Aljaafari (University of Manchester), Andre Freitas
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: By incorporating mutual information alignment and layer-wise stability regularization into the multi-layer hidden representations of LLMs, the model's consistency and robustness in compositional reasoning tasks are enhanced, while maintaining performance on downstream tasks after fine-tuning.
🎯 What it does: Propose the UNKGCP framework to construct prediction intervals for uncertain knowledge graph embedding models, providing statistical confidence guarantees.
Certified Mitigation of Worst-Case LLM Copyright Infringement
Jingyu Zhang (Johns Hopkins University), Daniel Khashabi (Johns Hopkins University)
CodeSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Designed and implemented BLOOMSCRUB, a copyright shielding method that detects and rewrites long references using Bloom filters during inference.
Chameleon LLMs: User Personas Influence Chatbot Personality Shifts
Jane Xing (University of North Carolina at Chapel Hill), Shashank Srivastava (University of North Carolina at Chapel Hill)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied how large language models (LLMs) perceive personality during prolonged interactions with users, constructed a self-assessment measurement framework based on the IPIP 50-item questionnaire, and conducted large-scale simulation experiments to explore adaptation patterns and predictability/regulatability of different Big Five personality traits.
Chart2Code53: A Large-Scale Diverse and Complex Dataset for Enhancing Chart-to-Code Generation
Tianhao Niu (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
CodeGenerationData SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed a large-scale, diverse, and complex Chart2Code dataset called Chart2Code53, and fine-tuned open-source multimodal LLMs on this dataset to achieve state-of-the-art (SOTA) performance on the Chart2Code benchmark.
Charting the Landscape of African NLP: Mapping Progress and Shaping the Road Ahead
Jesujoba Oluwadara Alabi, Dietrich Klakow (Saarland University)
CodeTransformerLarge Language ModelTextReview/Survey PaperAudio
🎯 What it does: Conduct a systematic literature review of natural language processing research on African languages over the past five years, collecting and annotating 884 papers, and analyzing research languages, tasks, techniques, datasets, and research trends.
Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites
Xintong Wang (Universität Hamburg), Chris Biemann (Universität Hamburg)
CodeGenerationData SynthesisTransformerLarge Language ModelTextBenchmark
🎯 What it does: Developed the TOXIREWRITECN Chinese toxic language detoxification dataset and evaluated the detoxification effectiveness of 17 LLMs under emotion preservation.
CIE: Controlling Language Model Text Generations Using Continuous Signals
Vinay Samuel (University of Maryland), Daphne Ippolito (Carnegie Mellon University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes the CIE (Control through Interpolated Embeddings) method, which utilizes interpolated control vectors to achieve continuous regulation of language model output attributes, particularly focusing on precise control over answer length.
CiteBART: Learning to Generate Citations for Local Citation Recommendation
Ege Yiğit Çelik (Izmir Institute of Technology), Selma Tekir (Izmir Institute of Technology)
CodeGenerationRecommendation SystemTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the CiteBART model, which utilizes BART pretraining and custom citation masking to end-to-end generate citations in local contexts, subsequently expanding to a global version by incorporating cited paper titles and abstracts.
ClimateViz: A Benchmark for Statistical Reasoning and Fact Verification on Scientific Charts
Ruiran Su (University of Oxford), Janet B. Pierrehumbert (University of Oxford)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringMultimodalityTabularBenchmark
🎯 What it does: This study constructs the CLIMATEVIZ dataset, providing a fact-checking benchmark based on real scientific charts and conducting a systematic evaluation of multimodal large language models.
COAS2W: A Chinese Older-Adults Spoken-to-Written Transformation Corpus with Context Awareness
Chun Kang (Fudan University), Yangfan Zhou (Fudan University)
CodeGenerationData SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkAudio
🎯 What it does: Constructed the COAS2W corpus (10,004 transcribed elderly speech sentences with 4-sentence context, fine-grained error labels, and written results), and trained and evaluated context-aware speech-to-written models on this corpus.
🎯 What it does: Proposed a self-distillation framework named CODI, which compresses Chain-of-Thought (CoT) reasoning into a continuous latent space, achieving implicit CoT reasoning.
CoEvo: Coevolution of LLM and Retrieval Model for Domain-Specific Information Retrieval
Ang Li, Kun Kuang (Zhejiang University)
CodeRetrievalTransformerLarge Language ModelContrastive LearningTextBiomedical Data
🎯 What it does: Study the collaborative evolution of LLM and retriever in professional domain retrieval, proposing the CoEvo framework to achieve alternating optimization
Cognitive Linguistic Identity Fusion Score (CLIFS): A Scalable Cognition‐Informed Approach to Quantifying Identity Fusion from Text
Devin R. Wright (Indiana University Bloomington), Yong-Yeol Ahn (Indiana University Bloomington)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a metaphor detection method based on LLM, constructed an automated and scalable cognitive linguistics identity fusion metric (CLIFS), and significantly improved performance in violence risk prediction.
Compound AI Systems Optimization: A Survey of Methods, Challenges, and Future Directions
Yu-Ang Lee (National Taiwan University), Yun-Nung Chen (National Taiwan University)
CodeOptimizationMeta LearningSupervised Fine-TuningReinforcement LearningAgentic AITextReview/Survey Paper
🎯 What it does: Reviews optimization methods, challenges, and future directions for composite AI systems (composed of components such as multimodal, tools, and agents), proposes a unified graph structure and conditional edge formulation, and systematically summarizes 26 representative works by constructing a 2×2 framework based on structural flexibility and learning signals.
Computational Analysis of Character Development in Holocaust Testimonies
Esther Shizgal (Hebrew University of Jerusalem), Omri Abend (Hebrew University of Jerusalem)
CodeClassificationTransformerLarge Language ModelText
🎯 What it does: Segmented 1000 Holocaust survivor interview texts, filtered religious content, determined belief and practice tendencies using LLMs, constructed religious trajectories, and clustered analysis to identify multiple common trajectory patterns.
🎯 What it does: Proposed the CARE framework, which introduces a context evaluator into RAG systems. By utilizing soft context embeddings, the framework identifies and resolves conflicts between retrieved context and LLM's internal knowledge, thereby enhancing the robustness of retrieval-augmented generation.
ConsistentChat: Building Skeleton-Guided Consistent Multi-Turn Dialogues for Large Language Models from Scratch
Jiawei Chen (Chinese Academy of Sciences), Xianpei Han (Chinese Academy of Sciences)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed a multi-turn dialogue instruction generation framework based on a dialogue intent skeleton, and generated the ConsistentChat dataset from scratch, aiming to enhance the consistency and coherence of large language models in multi-turn dialogues.