EMNLP 2025 Papers — Page 18
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Unpacking Let Alone: Human-Scale Models Generalize to a Rare Construction in Form but not Meaning
Wesley Scivetti (Georgetown University), Nathan Schneider (Georgetown University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Systematically evaluate the grammatical forms and semantic meanings of the rare construction LET-ALONE in English using a human-scale Transformer language model (OPT), constructing a templated minimal contrast benchmark and measuring the model's syntactic sensitivity with the SLOR metric.
Unraveling Interwoven Roles of Large Language Models in Authorship Privacy: Obfuscation, Mimicking, and Verification
Tuc Nguyen (Indiana University), Thai Le (Indiana University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a unified framework to systematically investigate three tasks related to author identity privacy in large language models: author identity obfuscation (AO), author identity simulation (AM), and author identity verification (AV), and quantify their mutual influences in single-step and iterative processes.
Unstructured Evidence Attribution for Long Context Query Focused Summarization
Dustin Wright (University of Copenhagen), David Jurgens (University of Michigan)
GenerationData SynthesisSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: In the task of long-text query-oriented summarization, this work first explores the use of arbitrary-length (unstructured) evidence fragments for citation and proposes a generative dataset called SUnsET to train LLMs to replicate and cite these evidence fragments.
Unsupervised Concept Vector Extraction for Bias Control in LLMs
Hannah Cyberey (University of Virginia), David Evans (University of Virginia)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningLarge Language ModelText
🎯 What it does: Investigated an unsupervised concept vector extraction method to activate steering of gender (and race) concepts within LLMs, reducing bias through projection interventions during inference.
Unsupervised Hallucination Detection by Inspecting Reasoning Processes
Ponhvoan Srey (Nanyang Technological University), Anh Tuan Luu (Nanyang Technological University)
Anomaly DetectionRepresentation LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed the IRIS framework, which utilizes the internal states of LLMs for unsupervised hallucination detection.
Unsupervised Word-level Quality Estimation for Machine Translation Through the Lens of Annotators (Dis)agreement
Gabriele Sarti (University of Groningen), Arianna Bisazza (University of Groningen)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: This paper analyzes the internal representations of machine translation models and uncertainty quantification methods, investigates and evaluates a series of unsupervised word-level quality estimation (WQE) metrics, and compares them with supervised methods and multi-annotator human annotations.
Unveiling Internal Reasoning Modes in LLMs: A Deep Dive into Latent Reasoning vs. Factual Shortcuts with Attribute Rate Ratio
Yiran Yang (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigated two modes in multi-hop questions: latent reasoning and fact shortcuts, constructed a specialized dataset, and proposed the ARR metric to distinguish them.
Unveiling the Response of Large Vision-Language Models to Visually Absent Tokens
Sohee Kim (KAIST AI), Eunho Yang (KAIST AI)
Anomaly DetectionExplainability and InterpretabilityTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: The study found that large vision-language models (LVLMs) incorrectly infer the presence of objects in images when the input text contains words lacking visual support. It proposes constructing a visual absence detection module using visual absence-aware neurons (VA neurons) within the model's internal FFN, thereby correcting model outputs in binary classification QA tasks and open-ended generation tasks.
User Feedback in Human-LLM Dialogues: A Lens to Understand Users But Noisy as a Learning Signal
Yuhan Liu (New York University), Eunsol Choi (New York University)
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Studied implicit user feedback in human-computer dialogue, proposed an efficient detection method, and explored utilizing feedback content to improve large model training
Utility-Focused LLM Annotation for Retrieval and Retrieval-Augmented Generation
Hengran Zhang (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences)
RetrievalTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper uses large language models (LLMs) to automatically label the practicality of documents in a retrieval corpus, trains a retriever with these labels, and evaluates its performance in retrieval and retrieval-augmented generation (RAG) tasks.
V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models
Qidong Wang (Tongji University), Ming Jiang (University of Wisconsin-Madison)
Explainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Propose the V-SEAM framework, combining visual semantic editing and attention modulation to perform causal interpretability analysis on vision-language models and enhance VQA performance
V-VAE: A Variational Auto Encoding Framework Towards Fine-Grained Control over Human-Like Chat
Qi Lin (University Of Electronic Science And Technology Of China), Bin Dai (Qingyao Intelligence)
GenerationLarge Language ModelSupervised Fine-TuningAuto EncoderTextBenchmark
🎯 What it does: Propose the V-VAE framework, leveraging variational autoencoders to achieve fine-grained, interpretable control over human-like chatting, and construct a high-quality HumanChatData dataset and HumanChatBench evaluation benchmark.
Value Profiles for Encoding Human Variation
Taylor Sorensen (University of Washington), Verena Rieser (Google DeepMind)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: The study constructs natural language value profiles and combines them with encoder-decoder autoencoders from prompt-based large language models to capture and explain differences among individual reviewers; it also proposes clustering methods based on value profiles and information-theoretic methods to assess available information.
Variance Sensitivity Induces Attention Entropy Collapse and Instability in Transformers
Jonghyun Hong (Hanyang University), Sungyoon Lee (Hanyang University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper investigates the attention entropy collapse problem caused by the high variance sensitivity of softmax in Transformers, and proves that entropy-stable attention methods (such as ReLU kernel attention and QK-LayerNorm) can prevent entropy collapse, thereby improving training stability.
VC4VG: Optimizing Video Captions for Text-to-Video Generation
Yang Du (Renmin University of China), Qin Jin (Renmin University of China)
GenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Propose the VC4VG framework, which decomposes video subtitles into five dimensions (subject attributes, environment, actions, camera parameters, style) and optimizes them to enhance the training effectiveness of text-to-video models.
VCSearch: Bridging the Gap Between Well-Defined and Ill-Defined Problems in Mathematical Reasoning
Shi-Yu Tian (Nanjing University), Yu-Feng Li (Nanjing University)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the VCSEARCH framework, which utilizes formal language and variable-constraint dynamic search technology to detect missing or contradictory conditions in mathematical problems without requiring additional training; constructed the PMC benchmark containing over 5,000 ill-defined math problems;
VEHME: A Vision-Language Model For Evaluating Handwritten Mathematics Expressions
Thu Phuong Nguyen (Ulsan National Institute of Science and Technology), Taehwan Kim (Ulsan National Institute of Science and Technology)
RecognitionData SynthesisLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageText
🎯 What it does: Proposed an end-to-end vision-language model called VEHME for automatically assessing the correctness of handwritten mathematical solutions and locating errors.
VELA: An LLM-Hybrid-as-a-Judge Approach for Evaluating Long Image Captions
Kazuki Matsuda (Keio University), Komei Sugiura (Keio University)
TransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose an automatic evaluation metric called VELA specifically for assessing long image descriptions generated by multimodal large language models, which evaluates from three aspects: descriptiveness, relevance, and fluency.
VerIF: Verification Engineering for Reinforcement Learning in Instruction Following
Hao Peng (Tsinghua University), Juanzi Li (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose VERIF, a hybrid approach combining rule-based code verification with large inference models (e.g., QwQ-32B) verification, applied to instruction-following reinforcement learning;
VeriFact: Enhancing Long-Form Factuality Evaluation with Refined Fact Extraction and Reference Facts
Xin Liu (University of Michigan), Lu Wang (University of Michigan)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the VERIFACT framework to improve fact extraction and verification in long texts, and construct the FACTRBENCH benchmark to simultaneously evaluate precision and recall.
VeriLocc: End-to-End Cross-Architecture Register Allocation via LLM
Lesheng Jin (UC San Diego), Jingbo Shang (UC San Diego)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: Propose VERILOCC, an end-to-end framework that treats register allocation as a sequence-to-sequence translation task, combining LLM, static analysis, and SMT verification to achieve cross-GPU architecture register allocation.
VERITAS: Leveraging Vision Priors and Expert Fusion to Improve Multimodal Data
Tingqiao Xu (Shanghai Jiao Tong University), Jiayu Chen (Fudan University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: This study proposes the VERITAS pipeline, systematically integrating visual priors, three large multimodal model evaluations and statistical fusion, as well as a lightweight GRPO evaluation model, to enhance the quality of supervised fine-tuning data while reducing hallucinations and factual errors;
VersaTune: An Efficient Data Composition Framework for Training Multi-Capability LLMs
Keer Lu (Peking University), Wentao Zhang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
🎯 What it does: Propose a data combination framework called VersaTune during the supervised fine-tuning phase to enhance the overall performance of large language models in multi-domain tasks.
Viability of Machine Translation for Healthcare in Low-Resourced Languages
Hellina Hailu Nigatu (UC Berkeley), Monojit Choudhury (MBZUAI)
TransformerLarge Language ModelPrompt EngineeringBiomedical DataElectronic Health Records
🎯 What it does: Explored the feasibility of general machine translation tools in medical scenarios for low-resource languages (Amharic, Tigrinya), constructed an evaluation dataset, established an error classification system, assessed translation errors, and investigated the effectiveness of preprocessing interventions (source sentence rewriting, translation mediation through Arabic).
ViClaim: A Multilingual Multilabel Dataset for Automatic Claim Detection in Videos
Patrick Giedemann (Zurich University of Applied Sciences), Mark Cieliebak (Zurich University of Applied Sciences)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the ViClaim dataset, utilizing multilingual short video transcripts for sentence-level multi-label claim detection;
Video Compression Commander: Plug-and-Play Inference Acceleration for Video Large Language Models
Xuyang Liu (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
CompressionTransformerVision Language ModelVideo
🎯 What it does: Proposes VidCom 2, a pluggable video token compression framework based on frame uniqueness adaptivity, for accelerating VideoLLM inference.
Video-RTS: Rethinking Reinforcement Learning and Test-Time Scaling for Efficient and Enhanced Video Reasoning
Ziyang Wang (UNC Chapel Hill), Mohit Bansal (UNC Chapel Hill)
Computational EfficiencyLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: For long video reasoning tasks, a fully reinforcement learning (GRPO)-based training scheme is proposed, combined with a sparse-to-dense testing scaling (Sparse‑to‑Dense Video TTS) strategy to adaptively allocate inference computation.
Video2Roleplay: A Multimodal Dataset and Framework for Video-Guided Role-playing Agents
Xueqiao Zhang (Zhejiang University), Yawei Luo (Zhejiang University)
TransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose dynamic role configuration and introduce the video modality into role-playing agents, constructing a large-scale multi-modal dataset and designing a corresponding framework
VideoEraser: Concept Erasure in Text-to-Video Diffusion Models
Naen Xu (Zhejiang University), Shouling Ji (Zhejiang University)
GenerationPrompt EngineeringDiffusion modelImageVideoTextMultimodality
🎯 What it does: Propose VideoEraser, a training-free, plug-and-play concept elimination framework that can suppress the generation of undesirable or copyright-infringing content in text-to-video diffusion models;
VideoPASTA: 7K Preference Pairs That Matter for Video-LLM Alignment
Yogesh Kulkarni (Arizona State University), Pooyan Fazli (Arizona State University)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelVideoText
🎯 What it does: Propose the VideoPASTA framework, achieving alignment optimization for Video-LLMs by constructing 7,000 high-quality aligned and adversarial preference pairs.
ViDoRAG: Visual Document Retrieval-Augmented Generation via Dynamic Iterative Reasoning Agents
Qiuchen Wang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
GenerationRetrievalTransformerAgentic AIVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose ViDoRAG—a multi-agent retrieval-augmented generation framework for visually rich documents, and construct a new ViDoSeek dataset
ViLBench: A Suite for Vision-Language Process Reward Modeling
Haoqin Tu (University Of California Santa Cruz), Cihang Xie (University Of California Santa Cruz)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper conducts benchmark evaluations of existing visual-language large models (VLLM) using reward models (ORM and PRM), proposes the VILBENCH benchmark specifically requiring fine-grained process rewards, constructs the ViLReward-73K dataset with 73.6K step-by-step reward samples, and trains a 3B-parameter ViLPRM, demonstrating its advantages on VILBENCH and other reward benchmarks.
ViPE: Visual Perception in Parameter Space for Efficient Video-Language Understanding
Shichen Lu (Beihang University), Jing Liu (Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Propose ViPE, a video-to-parameter alignment paradigm that directly injects video information into the weights of large language models (LLMs), eliminating the need for a large number of visual tokens.
VISaGE: Understanding Visual Generics and Exceptions
Stella Frank (University of Copenhagen), Emily Allaway (University of Edinburgh)
Anomaly DetectionExplainability and InterpretabilityLarge Language ModelVision Language ModelImage
🎯 What it does: This paper constructs the VISaGE dataset and uses it to evaluate priority bias in visual language models when handling general attributes and exception instances;
VisBias: Measuring Explicit and Implicit Social Biases in Vision Language Models
Jen-tse Huang (University of Southern California), Jieyu Zhao (University of Southern California)
Explainability and InterpretabilityVision Language ModelImageTextBenchmark
🎯 What it does: Developed the VISBIAS benchmark for simultaneously evaluating explicit and implicit social biases in vision-language models;
VisCRA: A Visual Chain Reasoning Attack for Jailbreaking Multimodal Large Language Models
Bingrui Sima (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Developed a jailbreak method called VisCRA that leverages visual chain-of-thought reasoning, combining attention-guided masking with two-stage reasoning induction to attack multimodal large language models.
VisEscape: A Benchmark for Evaluating Exploration-driven Decision-making in Virtual Escape Rooms
Seungwon Lim (Yonsei University), Youngjae Yu (Seoul National University)
Large Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the VisEscape benchmark, which includes 20 virtual escape rooms, to evaluate the decision-making capabilities of multi-modal AI in scenarios requiring self-guided exploration, information collection, and iterative reasoning, and design memory management and reasoning modules to enhance model performance.
VisFinEval: A Scenario-Driven Chinese Multimodal Benchmark for Holistic Financial Understanding
Zhaowei Liu (Shanghai University of Finance and Economics), Liwen Zhang (Shanghai University of Finance and Economics)
Large Language ModelVision Language ModelImageTextMultimodalityBenchmarkFinance Related
🎯 What it does: Constructed and released VisFinEval, a Chinese financial multimodal evaluation benchmark, covering 15,848 QA pairs, involving 8 types of financial images, hierarchically categorized by pre, mid, and post office scenarios, systematically assessing the overall financial understanding and reasoning capabilities of multimodal large language models (MLLM).
Vision-and-Language Navigation with Analogical Textual Descriptions in LLMs
Yue Zhang (Michigan State University), Parisa Kordjamshidi (Michigan State University)
Autonomous DrivingTransformerLarge Language ModelPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmark
🎯 What it does: Introduce an analogy reasoning module in vision-and-language navigation (VLN) driven by large language models, enhancing contextual understanding for navigation decisions by comparing multi-perspective images to generate distinctive scene and spatial descriptions.
Vision-Free Retrieval: Rethinking Multimodal Search with Textual Scene Descriptions
Ioanna Ntinou (Queen Mary University of London), Georgios Tzimiropoulos (Queen Mary University of London)
RetrievalLarge Language ModelVision Language ModelContrastive LearningImageText
🎯 What it does: Propose a vision-agnostic single-encoder retrieval framework called LexiCLIP, which converts images into structured text descriptions and performs retrieval entirely within the text domain.
VisiPruner: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal LLMs
Yingqi Fan, Xiaoyu Shen (Ningbo Key Laboratory Of Spatial Intelligence And Digital Derivative)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Proposed a training-free visual token pruning framework called VisiPruner, achieving significant computational efficiency improvements through systematic analysis of hierarchical interactions in multi-modal large language models (MLLM).
VistaWise: Building Cost-Effective Agent with Cross-Modal Knowledge Graph for Minecraft
Honghao Fu (Hong Kong University of Science and Technology (Guangzhou)), Hao Wang (Hong Kong University of Science and Technology (Guangzhou))
Object DetectionConvolutional Neural NetworkLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelVideoTextMultimodalityGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a cost-effective agent framework named VistaWise in Minecraft, leveraging cross-modal knowledge graphs, lightweight object detection, and desktop-level mouse/keyboard skill libraries to accomplish complex tasks using only hundreds of frames of data.
Visual Contextual Attack: Jailbreaking MLLMs with Image-Driven Context Injection
Miao Ziqi (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Adversarial AttackLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes a vision-context-based multimodal large language model (MLLM) reverse attack framework called VisCo Attack, which bypasses security barriers through image-driven context injection.
Visual-Aware Speech Recognition for Noisy Scenarios
Balaji Darur (IIIT Hyderabad), Karan Singla (Whissle Inc)
RecognitionTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: Propose a scalable dataset generation pipeline and a vision-aware speaker diarization model that leverages the correlation between video background and noise to improve transcription quality in noisy environments.
VisualWebInstruct: Scaling up Multimodal Instruction Data through Web Search
Yiming Jia (University of Toronto), Wenhu Chen (University of Waterloo)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Constructed a large-scale multimodal reasoning dataset called VisualWebInstruct based on Google image search, and refined it through two-stage filtering and LLM-assisted refinement to obtain approximately 900,000 question-answer pairs.
VLA-Mark: A cross modal watermark for large vision-language alignment models
Shuliang Liu (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
Safty and PrivacyTransformerVision Language ModelImageTextMultimodality
🎯 What it does: For vision-language alignment large models (VLAMMs), a cross-modal watermarking scheme named VLA-Mark is proposed, which can embed detectable watermarks without compromising visual-textual consistency.
VLASCD: A Visual Language Action Model for Simultaneous Chatting and Decision Making
Zuojin Tang (Zhejiang University), Bin Liu (E-surfing Digital Life Technology Co., Ltd.)
Autonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelVideoText
🎯 What it does: Propose a multi-input multi-output visual-language-action model called MIMO-VLA, capable of simultaneously performing natural language dialogue and real-time driving decision-making.
VLP: Vision-Language Preference Learning for Embodied Manipulation
Runze Liu, Xiu Li (Tsinghua University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Studied a learning framework called VLP based on visual-language preferences, using video and language instructions to generate preference labels instead of expensive human feedback for training reinforcement learning (RL) policies in embedded manipulation tasks.
VocalNet: Speech LLMs with Multi-Token Prediction for Faster and High-Quality Generation
Yuhao Wang (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
GenerationTransformerLarge Language ModelAudio
🎯 What it does: Proposed VocalNet, a speech large language model that integrates a two-stage training framework with multi-token prediction (MTP);
Voice of a Continent: Mapping Africa’s Speech Technology Frontier
AbdelRahim A. Elmadany (University of British Columbia), Muhammad Abdul-Mageed (University of British Columbia)
RecognitionGenerationTransformerSupervised Fine-TuningBenchmarkAudio
🎯 What it does: Built the SimbaBench benchmark, aggregating and unifying 8,605+ hours of speech data across 61 African languages, followed by fine-tuning the Simba series of ASR, TTS, and SLID models on this benchmark, and systematically evaluating their performance.
VoiceBBQ: Investigating Effect of Content and Acoustics in Social Bias of Spoken Language Model
Junhyuk Choi (Chung-Ang University), Bugeun Kim (Chung-Ang University)
TransformerLarge Language ModelTextBenchmarkAudio
🎯 What it does: This study proposes the VoiceBBQ dataset to evaluate social bias in speech language models from both content and acoustic perspectives;
VoiceCraft-X: Unifying Multilingual, Voice-Cloning Speech Synthesis and Speech Editing
Zhisheng Zheng (University of Texas at Austin), David Harwath (University of Texas at Austin)
GenerationData SynthesisTransformerLarge Language ModelTextAudio
🎯 What it does: Propose VoiceCraft-X, a unified autoregressive neural speech encoder language model capable of achieving zero-shot voice cloning TTS and voice editing across 11 languages;
VRoPE: Rotary Position Embedding for Video Large Language Models
Zikang Liu (Institute of Automation Chinese Academy of Sciences), Jing Liu (Institute of Automation Chinese Academy of Sciences)
Representation LearningTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: Proposed a Video Rotation Position Encoding (VRoPE) for video large language models to better capture the spatiotemporal structure of videos and avoid position bias.
Walk and Read Less: Improving the Efficiency of Vision-and-Language Navigation via Tuning-Free Multimodal Token Pruning
Wenda Qin (Boston University), Margrit Betke (Boston University)
Computational EfficiencyLarge Language ModelVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: This paper proposes a multi-modal token pruning framework called Navigation-Aware Pruning (NAP), which integrates Background View Pruning (BGP), Backtracking Node Pruning (BTP), and Vocabulary-based Instruction Pruning (VPP). It significantly reduces the number of visual, textual, and historical node tokens in Vision-and-Language Navigation (VLN) tasks while maintaining or even improving navigation success rates.
WangchanThaiInstruct: An instruction-following Dataset for Culture-Aware, Multitask, and Multi-domain Evaluation in Thai
Peerat Limkonchotiwat (AI Singapore), Sarana Nutanong (VISTEC)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed a dataset of 35,014 Thai instruction-response pairs covering four professional domains and seven task categories, and conducted zero-shot evaluation and instruction fine-tuning experiments.
Warm Up Before You Train: Unlocking General Reasoning in Resource-Constrained Settings
Safal Shrestha (New York University Abu Dhabi), Keith W. Ross (New York University Abu Dhabi)
Computational EfficiencyKnowledge DistillationSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose a two-stage training framework: first, warm up the model using the long-chain reasoning of Knights & Knaves logic puzzles through distillation, then fine-tune it with a small number of target domain samples via RLVR, enhancing reasoning performance in resource-constrained scenarios.
Waste-Bench: A Comprehensive Benchmark for Evaluating VLLMs in Cluttered Environments
Muhammad Ali (Mohamed Bin Zayed University of Artificial Intelligence), Salman Khan (Mohamed Bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Constructed a benchmark named Waste-Bench to evaluate the robustness and reasoning capabilities of vision-language large models (VLLM) in real-world cluttered garbage classification environments.
We Need to Measure Data Diversity in NLP — Better and Broader
Dong Nguyen (Utrecht University), Esther Ploeger (Aalborg University)
TextReview/Survey Paper
🎯 What it does: This paper reviews and systematizes the concepts, methods, and challenges of measuring data diversity in NLP, proposes a measurement framework from a multi-dimensional, interdisciplinary perspective, and provides six actionable recommendations.
We Politely Insist: Your LLM Must Learn the Persian Art of Taarof
Nikta Gohari Sadr (Brock University), Ali Emami
Large Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Introduce the TAAROFBENCH benchmark to evaluate the understanding and generation capabilities of LLMs regarding the Persian cultural politeness form Taarof, and assess and adapt multiple LLMs.
Weaver: Interweaving SQL and LLM for Table Reasoning
Rohit Khoja (Arizona State University), Vivek Gupta (Arizona State University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMultimodalityTabularFinance Related
🎯 What it does: Proposed Weaver, a modular and interpretable framework that dynamically switches between SQL and LLM to address table QA tasks involving mixed structured and unstructured data.
Web Intellectual Property at Risk: Preventing Unauthorized Real-Time Retrieval by Large Language Models
Yisheng Zhong (George Mason University), Zhuangdi Zhu (George Mason University)
Safty and PrivacyLarge Language ModelText
🎯 What it does: Proposes a defense framework that leverages LLM's semantic understanding capabilities by embedding protective strategies into webpage HTML to prevent LLMs from extracting and reconstructing webpage IP addresses during real-time retrieval.
WebAgent-R1: Training Web Agents via End-to-End Multi-Turn Reinforcement Learning
Zhepei Wei (University of Virginia), Lihong Li (Amazon)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextSequentialBenchmarkChain-of-Thought
🎯 What it does: Proposed an end-to-end multi-round reinforcement learning framework called WEBAGENT-R1 for training web agents, enabling large language models to learn task completion through interaction with real web environments.
WebEvolver: Enhancing Web Agent Self-Improvement with Co-evolving World Model
Tianqing Fang (Tencent AI Lab), Dong Yu (Tencent AI Lab)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIWorld ModelTextBenchmark
🎯 What it does: By constructing a self-improving framework called WebEvolver, which co-trains an agent model with a co-evolving world model, enabling Web agents to continuously enhance their performance with the help of self-generated trajectories and multi-step simulation reasoning.
WebInject: Prompt Injection Attack to Web Agents
Xilong Wang (Duke University), Neil Zhenqiang Gong (Duke University)
Adversarial AttackConvolutional Neural NetworkLarge Language ModelAgentic AIPrompt EngineeringImageMultimodality
🎯 What it does: Proposes a web prompt injection attack called WebInject, which uses minimal perturbations to the original pixels of a webpage to induce a web agent based on multi-modal large language models to execute operations specified by the attacker.
WebMMU: A Benchmark for Multimodal Multilingual Website Understanding and Code Generation
Rabiul Awal (ServiceNow), Sai Rajeswar (ServiceNow)
AI Code AssistantTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose WebMMU, a multimodal multilingual benchmark for website understanding and code generation, encompassing three tasks: website VQA, mockup-to-code, and code editing.
Weight-Aware Activation Sparsity with Constrained Bayesian Optimization Scheduling for Large Language Models
Ming Wang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose an untrained activation sparsity framework WAS, which uses weight information and activation values together to determine thresholds, and employs constrained Bayesian optimization to assign different sparsity rates to each Transformer block, while achieving a GPU kernel that supports non-uniform sparsity, significantly accelerating LLM inference.
Weights-Rotated Preference Optimization for Large Language Models
Chenxu Yang (Chinese Academy of Sciences), Weiping Wang (Baidu Inc.)
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the RoPO algorithm, which explicitly constrains intermediate hidden layer weights using multi-granularity orthogonal matrix rotation, while maintaining DPO's KL divergence constraint in the output layer, mitigating DPO's reward hijacking problem and enhancing expressiveness.
What are Foundation Models Cooking in the Post-Soviet World?
Anton Lavrouk (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Construct the BORSCH dataset to evaluate the understanding of post-Soviet food culture knowledge by Russian and Ukrainian foundational models.
What Do Indonesians Really Need from Language Technology? A Nationwide Survey
Muhammad Dehan Al Kautsar (Mohamed bin Zayed University of Artificial Intelligence), Fajri Koto (Mohamed bin Zayed University of Artificial Intelligence)
RetrievalSafty and PrivacyTextReview/Survey Paper
🎯 What it does: A nationwide questionnaire was conducted, collecting 861 responses from users of over 700 languages in Indonesia regarding their needs and concerns about language technologies.
What Makes a Good Reasoning Chain? Uncovering Structural Patterns in Long Chain-of-Thought Reasoning
Gangwei Jiang (University of Science and Technology of China), Defu Lian (University of Science and Technology of China)
Explainability and InterpretabilityGraph Neural NetworkLarge Language ModelTextGraphBenchmarkChain-of-Thought
🎯 What it does: Built a framework called LCoT2Tree that automatically converts long-chain reasoning (LCoT) into a hierarchical tree structure, using this tree structure to perform structured analysis of the LLM reasoning process.
What You Read Isn’t What You Hear: Linguistic Sensitivity in Deepfake Speech Detection
Binh Nguyen (Independent Researcher), Thai Le (Indiana University)
Adversarial AttackTextAudio
🎯 What it does: Propose and study the TAPAS framework, which generates adversarial samples by performing slight semantic-preserving word replacement on text transcriptions to attack voice anti-fraud detection systems.
What You See is What You Ask: Evaluating Audio Descriptions
Divy Kala (IIIT Hyderabad), Makarand Tapaswi (IIIT Hyderabad)
Large Language ModelMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes the ADQA (Audio Description Question Answering) benchmark, which evaluates audio description (AD) generation methods through multiple-choice QA tasks. It quantifies the subjectivity of AD by aligning and analyzing two independent AD tracks from the same movie.
What’s in a prompt? Language models encode literary style in prompt embeddings
Raphaël Sarfati (Cornell University), Christopher Earls (Cornell University)
ClassificationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Conduct a geometric analysis of deep embeddings from short literary passages in large language models, demonstrating that these embeddings can capture and distinguish authors' writing styles.
When Annotators Disagree, Topology Explains: Mapper, a Topological Tool for Exploring Text Embedding Geometry and Ambiguity
Nisrine Rair (Université de Reims Champagne-Ardenne), Emmanuel Chochoy (Chochoy Conseil)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Visualize and quantitatively analyze the embedding space of RoBERTa-Large on the MD-Offense dataset using the topological data analysis tool Mapper, investigating how the model constructs ambiguous representations in the presence of annotator disagreements.
When Audio and Text Disagree: Revealing Text Bias in Large Audio-Language Models
Cheng Wang (National University Of Singapore), Tianwei Zhang (Nanyang Technological University)
ClassificationRecognitionTransformerSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkAudio
🎯 What it does: Constructed the MCR-BENCH benchmark to evaluate the performance of large audio-language models (LALMs) when audio and text information conflict, revealing their strong text bias.
When Big Models Train Small Ones: Label-Free Model Parity Alignment for Efficient Visual Question Answering using Small VLMs
Abhirama Subramanyam Penamakuri (Indian Institute of Technology Jodhpur), Anand Mishra (Indian Institute of Technology Jodhpur)
Knowledge DistillationTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Developed a Model Parity Aligner (MPA) framework that leverages unlabeled images and high-quality pseudo-labels generated by a large-scale visual language model (L-VLM) to specifically enhance the visual question answering (VQA) performance of small-scale visual language models (S-VLM).
When Life Gives You Samples: The Benefits of Scaling up Inference Compute for Multilingual LLMs
Ammar Khairi (Cohere Labs), Sara Hooker (Cohere)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Studies computational amplification strategies for multilingual, multitask large language models, proposes sampling and selection methods tailored for multilingual environments, and verifies their effectiveness across different tasks (open-ended generation, mathematical reasoning, machine translation)
When Long Helps Short: How Context Length in Supervised Fine-tuning Affects Behavior of Large Language Models
Yingming Zheng (Shanghai Jiao Tong University), Lu Chen (Shanghai Jiao Tong University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigate the impact of long-text supervised fine-tuning (SFT) on the performance of large language models in short-text tasks, and systematically analyze the effects of long and short contexts on the internal behaviors of MHA and FFN modules.
When Truthful Representations Flip Under Deceptive Instructions?
Xianxuan Long (Case Western Reserve University), Pan Li (Case Western Reserve University)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: Studied changes in the internal representations of LLMs under deceptive instructions, and analyzed differences between honest and deceptive patterns through linear probes and sparse autoencoders.
When Words Smile: Generating Diverse Emotional Facial Expressions from Text
Haidong Xu (Harbin Institute of Technology), Hao Fei (National University of Singapore)
GenerationData SynthesisTransformerAuto EncoderTextMultimodalityMesh
🎯 What it does: Proposed an end-to-end text-to-expression generation model called CTEG, achieving diverse, smooth, and emotionally consistent 3D expression generation.
Where Confabulation Lives: Latent Feature Discovery in LLMs
Thibaud Ardoin (Freie Universität Berlin), Gerhard Wunder (Freie Universität Berlin)
Explainability and InterpretabilityTransformerPrompt EngineeringText
🎯 What it does: This paper extracts and re-projects a sparse, interpretable 'confabulation' direction from the LLM activation space by comparing dialog prompts of known and unknown entities, using Sparse Principal Component Analysis (SPCA), and injects this vector into intermediate layers to achieve controllable regulation of the model's confabulation behavior.
Where to show Demos in Your Prompt: A Positional Bias of In-Context Learning
Kwesi Adu Cobbina (University of Maryland), Tianyi Zhou (University of Maryland)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigate the impact of demo position on model performance in the context learning of large language models; propose and systematically evaluate DPP (demo position) bias; define two task-agnostic metrics, ACC/CHANGE and PRED/CHANGE; conduct large-scale experiments on eight categories of tasks across ten open-source LLMs; analyze the relationship between model scale, task type, and optimal position; provide practical recommendations and mitigation directions.
Which Word Orders Facilitate Length Generalization in LMs? An Investigation with GCG-Based Artificial Languages
Nadine El-Naggar (Mohamed bin Zayed University of Artificial Intelligence), Ted Briscoe (Mohamed bin Zayed University of Artificial Intelligence)
Data SynthesisRecurrent Neural NetworkTransformerLarge Language ModelTextSequential
🎯 What it does: Investigate the generalization ability of language models under different word orders using artificial language generation experiments based on Generalized Categorial Grammar (GCG);
Whisper-UT: A Unified Translation Framework for Speech and Text
Cihan Xiao (Johns Hopkins University), Sanjeev Khudanpur (Johns Hopkins University)
TransformerSupervised Fine-TuningPrompt EngineeringTextMultimodalityAudio
🎯 What it does: Achieved a unified framework for speech recognition, speech translation, text translation, and multimodal translation within a single model, named Whisper-UT.
Who Holds the Pen? Caricature and Perspective in LLM Retellings of History
Lubna Zahan Lamia (University of Dhaka), Md Mosaddek Khan (University of Dhaka)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Study the simulation of character perspectives in historical narratives by LLMs, with a systematic evaluation of identity distortion (exaggeration, conflict, passive voice, toxicity)
Who is in the Spotlight: The Hidden Bias Undermining Multimodal Retrieval-Augmented Generation
Jiayu Yao (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
RetrievalMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper systematically studies the bias caused by the order of retrieved evidence in multimodal retrieval-augmented generation (RAG) systems and quantifies this bias through experimental and visualization analysis.
Why and How LLMs Benefit from Knowledge Introspection in Commonsense Reasoning
Chengfeng Zhao (Chinese Academy of Sciences), Kang Liu (Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper investigates the effectiveness of large language models in common-sense reasoning tasks and proposes an untrained adaptive introspection strategy.
Why Do Some Inputs Break Low-Bit LLM Quantization?
Ting-Yun Chang (University of Southern California), Robin Jia (University of Southern California)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Investigate why 3-4 bit weight quantization in large language models leads to significant errors on certain inputs and systematically analyze the sources and mechanisms of these errors.
Why Stop at One Error? Benchmarking LLMs as Data Science Code Debuggers for Multi-Hop and Multi-Bug Errors
Zhiyu Yang (Singapore Management University), Yang Deng (Singapore Management University)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Propose DSDBench—a debugging benchmark for multi-hop and multi-defect runtime errors in data science code;
WildDoc: How Far Are We from Achieving Comprehensive and Robust Document Understanding in the Wild?
An-Lan Wang (ByteDance), Can Huang (ByteDance)
TransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposed the WildDoc benchmark, collected over 12,000 document images captured in real-world environments, and evaluated the document understanding capabilities of existing multimodal large language models (MLLMs) on this benchmark.
WildScore: Benchmarking MLLMs in-the-Wild Symbolic Music Reasoning
Gagan Mundada (University of California, San Diego), Junda Wu (University of California, San Diego)
Large Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the WildScore benchmark, constructing a multimodal symbolic music reasoning dataset using real sheet music and Reddit community questions, and evaluated the performance of large multimodal language models.
Will It Still Be True Tomorrow? Multilingual Evergreen Question Classification to Improve Trustworthy QA
Sergey Pletenev (Skoltech), Viktor Moskvoretskii (EPFL)
ClassificationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs a multilingual sustainability issue classification dataset called EverGreenQA, trains and evaluates a lightweight multilingual classifier EG-E5, and applies sustainability discrimination to self-knowledge estimation, QA dataset filtering, and GPT-4o retrieval behavior explanation.
WISE: Weak-Supervision-Guided Step-by-Step Explanations for Multimodal LLMs in Image Classification
Yiwen Jiang (Monash University), Zongyuan Ge (Monash University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageChain-of-Thought
🎯 What it does: Proposed a weakly supervised guided progressive explanation method called WISE, which transforms the concept representations of a concept bottleneck model into explainable multi-modal chain-of-thought (MCoT), enhancing the interpretability and accuracy of multimodal large language models in image classification tasks.
Women, Infamous, and Exotic Beings: A Comparative Study of Honorific Usages in Wikipedia and LLMs for Bengali and Hindi
Sourabrata Mukherjee (Mohamed bin Zayed University of Artificial Intelligence), Monojit Choudhury (Mohamed bin Zayed University of Artificial Intelligence)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper investigates the usage patterns of third-person honorifics in Hindi and Bengali Wikipedia articles, and constructs a detection and comparison framework for large language models (LLMs) in terms of honorific usage during generation and translation tasks.
Word Salad Chopper: Reasoning Models Waste A Ton Of Decoding Budget On Useless Repetitions, Self-Knowingly
Wenya Xie (University of Minnesota), Zirui Liu (University of Minnesota)
OptimizationComputational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Proposes a lightweight, plug-and-play WordSaladChopper (WSC) to detect and eliminate meaningless repetitive text ('word salad') generated by large inference models in real-time, thereby compressing inference length and reducing costs.
Words Like Knives: Backstory-Personalized Modeling and Detection of Violent Communication
Jocelyn J Shen (Massachusetts Institute of Technology), Hae Won Park (Massachusetts Institute of Technology)
ClassificationData SynthesisTransformerLarge Language ModelText
🎯 What it does: Built a framework for detecting relationship breakdowns based on nonviolent communication theory, generated and evaluated the PERSONACONFLICTS CORPUS dataset, and investigated the impact of relational context on human and model conflict perception.
X-CoT: Explainable Text-to-Video Retrieval via LLM-based Chain-of-Thought Reasoning
Prasanna Reddy Pulakurthi (Rochester Institute of Technology), Zhiqiang Tao (Rochester Institute of Technology)
RetrievalTransformerLarge Language ModelVideoTextChain-of-Thought
🎯 What it does: Propose X-CoT, an LLM-based chain-of-thought explanatory text-to-video retrieval framework that replaces traditional cosine similarity ranking with reasoning-based explanations.
X-FLoRA: Cross-modal Federated Learning with Modality-expert LoRA for Medical VQA
Min Hyuk Kim (Chonnam National University), Seok Bong Yoo (Chonnam National University)
Image TranslationFederated LearningVision Language ModelGenerative Adversarial NetworkImageTextMultimodalityBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Propose a cross-modal federated learning framework X-FLoRA, leveraging modal expert LoRA and heterogeneous text-driven image translation to enable clients holding only single-modal data to generate images of another modality without sharing raw data, and perform federated fine-tuning of medical visual question answering (VQA) models based on this.
XAutoLM: Efficient Fine-Tuning of Language Models via Meta-Learning and AutoML
Ernesto L. Estevanell-Valladares (University of Alicante), Ruslan Mitkov (University of Lancaster)
OptimizationComputational EfficiencyHyperparameter SearchMeta LearningTransformerSupervised Fine-TuningText
🎯 What it does: Proposes the XAutoLM framework, which achieves efficient and resource-friendly model and hyperparameter search for fine-tuning language models through meta-learning-driven AutoML;
xCoRe: Cross-context Coreference Resolution
Giuliano Martinelli (Sapienza University of Rome), Roberto Navigli (Sapienza University of Rome)
TransformerText
🎯 What it does: The paper proposes a unified cross-context coreference resolution framework named xCoRe, supporting short texts, long texts, and cross-document coreference tasks.