ACL 2025 Papers — Page 17
Annual Meeting of the Association for Computational Linguistics · 1699 papers
UniLR: Unleashing the Power of LLMs on Multiple Legal Tasks with a Unified Legal Retriever
Ang Li, Kun Kuang (Zhejiang University)
Graph Neural NetworkTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose UniLR, a unified legal retriever that provides retrieval support for LLMs across various legal tasks.
Unintended Harms of Value-Aligned LLMs: Psychological and Empirical Insights
Sooyung Choi (Sungkyunkwan University), JinYeong Bak (Microsoft Research Asia)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This study conducts a systematic safety evaluation of large language models (LLMs) aligned with the Schwartz value system, exploring the association between different value dimensions and specific safety risks (such as hate speech, sexual content, political agitation, etc.), and proposes a simple strategy to reduce harmful behaviors by suppressing risk-related values in prompts.
Unique Hard Attention: A Tale of Two Sides
Selim Jerad (ETH Zurich), Ryan Cotterell (ETH Zurich)
Transformer
🎯 What it does: Studied the differences in expressive power between left-priority and right-priority resolution in the unique hard attention (UHA) Transformer under finite precision;
UniRAG: Unified Query Understanding Method for Retrieval Augmented Generation
Rui Li (University of Science and Technology of China), Yu Su (Hefei Comprehensive National Science Center)
GenerationRetrievalRepresentation LearningTransformerContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose UniRAG, a unified query understanding framework that integrates query expansion and encoding into a single decoder LLM, achieving end-to-end query enhancement and representation;
Unleashing LLM Reasoning Capability via Scalable Question Synthesis from Scratch
Yuyang Ding (Soochow University), Min Zhang (Tencent)
Data SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposed and implemented ScaleQuest, a million-scale math reasoning Q&A dataset synthesized from scratch based on a lightweight 7B model, and used this dataset for instruction tuning
Unlocking General Long Chain-of-Thought Reasoning Capabilities of Large Language Models via Representation Engineering
Xinyu Tang (Renmin University of China), Zhiqiang Zhang (Ant Group)
Representation LearningTransformerContrastive LearningTextChain-of-Thought
🎯 What it does: By engineering the internal representations of LLMs, we propose a no-training method called GLoRE, which activates the model's general long-chain reasoning (long CoT) capability.
Unmasking Style Sensitivity: A Causal Analysis of Bias Evaluation Instability in Large Language Models
Jiaxu Zhao (Carnegie Mellon University), Mykola Pechenizkiy (Carnegie Mellon University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Analyze the stability of bias evaluation results in large models across different text styles, and investigate the impact of style transformation on bias scores;
Unraveling LoRA Interference: Orthogonal Subspaces for Robust Model Merging
Haobo Zhang (University of Michigan), Jiayu Zhou (University of Michigan)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a data-driven constraint on the LoRA subspace before fine-tuning (OSRM) to reduce output interference between different tasks, thereby improving the merging effectiveness of multi-task models.
Unraveling the Mechanics of Learning-Based Demonstration Selection for In-Context Learning
Hui Liu (City University of Hong Kong), Haoliang Li (City University of Hong Kong)
ClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper conducts a mechanism analysis of learning-based demonstration selection methods and proposes two low-cost demonstration selection approaches.
Unravelling the Logic: Investigating the Generalisation of Transformers in Numerical Satisfiability Problems
Tharindu Madusanka (University of Manchester), Riza Batista-Navarro (University of Manchester)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigated the generalization capability of Transformer models on numerical satisfiability problems, systematically evaluating their adaptability to variations in vocabulary, numerical values, scale, and noise.
UnSeenTimeQA: Time-Sensitive Question-Answering Beyond LLMs’ Memorization
Md Nayem Uddin (Arizona State University), Chitta Baral (Arizona State University)
TransformerPrompt EngineeringTime SeriesSequentialBenchmarkChain-of-Thought
🎯 What it does: Introduces a data-free time-sensitive question-answering benchmark called UnSeenTimeQA, which evaluates the temporal reasoning capabilities of large language models using synthetic logistics events.
Unsolvable Problem Detection: Robust Understanding Evaluation for Large Multimodal Models
Atsuyuki Miyai (University of Tokyo), Kiyoharu Aizawa (University of Tokyo)
Explainability and InterpretabilitySupervised Fine-TuningPrompt EngineeringMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the Unsolvable Problem Detection (UPD) task, constructing the MM-UPD Bench to evaluate the refusal capability of multi-modal large models in multiple-choice QA scenarios involving missing answers, invalid answer sets, and image-text mismatches
Untie the Knots: An Efficient Data Augmentation Strategy for Long-Context Pre-Training in Language Models
Junfeng Tian (East China Normal University), Debing Zhang (East China Normal University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Proposed and implemented a data augmentation strategy called Untie the Knots (UtK) to efficiently enhance the long-text context capability of large language models without altering the original data mixture.
Unveil: Unified Visual-Textual Integration and Distillation for Multi-modal Document Retrieval
Hao Sun (Peking University), Yan Zhang (Peking University)
RetrievalKnowledge DistillationRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a unified visual-text embedding framework called Unveil, which can integrate OCR text and visual information during retrieval, and achieve efficient retrieval without OCR by transferring semantic understanding from the teacher (vision + text) model to the student model (vision only) through knowledge distillation.
Unveiling Attractor Cycles in Large Language Models: A Dynamical Systems View of Successive Paraphrasing
Zhilin Wang (Shanghai AI Laboratory), Yue Zhang (Westlake University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study treats multi-round rewriting as a discrete dynamical system, discovering that large language models converge to a 2-period limit cycle during continuous rewriting, thereby restricting text diversity.
Unveiling Cultural Blind Spots: Analyzing the Limitations of mLLMs in Procedural Text Comprehension
Amir Hossein Yari (Sharif University of Technology), Fajri Koto (Mohamed bin Zayed University of Artificial Intelligence)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Proposed and constructed the multilingual, multicultural program text understanding benchmark CAPTex, systematically evaluating the program reasoning capabilities of multilingual large language models (mLLMs) in cultural contexts.
Unveiling Environmental Impacts of Large Language Model Serving: A Functional Unit View
Yanran Wu (Purdue University), Yi Ding (Purdue University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the FUEL framework, which standardizes the evaluation of carbon emissions in LLM services using functional units (FU), and validate its feasibility through three case studies on model size, quantization, and hardware.
Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders
Boyi Deng (University of Science and Technology of China), Fuli Feng (University of Science and Technology of China)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Use sparse autoencoders (SAE) to analyze activations in large language models (LLMs), propose a monolingualness metric to identify language-specific features, conduct directional ablation experiments to demonstrate the critical role of these features in monolingual performance, and leverage these features to improve steering vectors for language control.
Unveiling Privacy Risks in LLM Agent Memory
Bo Wang (Michigan State University), Pengfei He (Michigan State University)
Safty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBiomedical DataElectronic Health Records
🎯 What it does: Propose a black-box memory extraction attack method called MEXTRA targeting the memory module of large language model (LLM) agents, achieving leakage of private interaction records through carefully designed attack prompts and an automated prompt generator.
Unveiling the Potential of BERT-family: A New Recipe for Building Scalable, General and Competitive Large Language Models
Yisheng Xiao (Soochow University), Min Zhang (Soochow University)
TransformerLarge Language ModelText
🎯 What it does: Proposed BiGLM—a bidirectional universal language model based on the BERT family—and designed new pre-training tasks and multiple enhancement strategies, demonstrating that encoder-only models can achieve comparable performance to autoregressive LLMs.
Unveiling the Power of Source: Source-based Minimum Bayes Risk Decoding for Neural Machine Translation
Boxuan Lyu (Institute of Science Tokyo), Manabu Okumura (Nara Institute of Science and Technology)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a source-sentence based Minimum Bayes Risk (sMBR) decoding method that uses multiple approximated sources (generated through synonym rewriting or back-translation) as support hypotheses, and employs reference-free quality estimation (QE) metrics to evaluate the quality of candidate translations, thereby improving neural machine translation (NMT) decoding.
UORA: Uniform Orthogonal Reinitialization Adaptation in Parameter Efficient Fine-Tuning of Large Models
Xueyan Zhang (University of Waterloo), Yining Wang (University of Toronto)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningImageText
🎯 What it does: Proposes UORA, a parameter-efficient fine-tuning method that uses interpolation reinitialization on frozen projection matrices.
Upcycling Instruction Tuning from Dense to Mixture-of-Experts via Parameter Merging
Tingfeng Hui, Sen Su (Beijing University Of Posts And Telecommunications)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposes Upcycling Instruction Tuning (UpIT), which efficiently and flexibly upgrades dense pre-trained models into Mixture-of-Experts (MoE) instruction models by leveraging intermediate checkpoints generated during dense model training as specialized experts, and introducing genetic algorithm parameter merging and router pre-optimization.
UrbanVideo-Bench: Benchmarking Vision-Language Models on Embodied Intelligence with Video Data in Urban Spaces
Baining Zhao (Tsinghua University), Yong Li (Tsinghua University)
Vision Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This study constructs the UrbanVideo-Bench benchmark, specifically designed to evaluate embodied cognitive capabilities in first-person videos within urban aerial environments.
User-side Model Consistency Monitoring for Open Source Large Language Models Inference Services
Qijun Miao (Tsinghua University), Zhixuan Fang (Tsinghua University)
Anomaly DetectionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a user-oriented LLM consistency monitoring paradigm and designs a token confidence metric based on logits difference, enabling detection of whether the server-side model has been downgraded through a single forward pass;
Using Information Theory to Characterize Prosodic Typology: The Case of Tone, Pitch-Accent and Stress-Accent
Ethan Wilcox, Tamar I Regev
Explainability and InterpretabilityRepresentation LearningTransformerTextAudio
🎯 What it does: This paper uses an information theory framework to calculate the mutual information between pitch contours and text across ten languages (including tonal, pitch-accented, and stress languages), exploring the discriminative power of pitch for word meaning differentiation;
Using Shapley interactions to understand how models use structure
Divyansh Singhvi (Independent), Naomi Saphra (Harvard University)
Explainability and InterpretabilityTransformerLarge Language ModelTextMultimodalityAudio
🎯 What it does: This paper studies the nonlinear interactions between internal features of language models and speech models by calculating the Shapley Taylor interaction index, revealing how they encode syntax, metaphor, and phonetic interactions.
Using Subtext to Enhance Generative IDRR
Zhipang Wang (Soochow University), Guodong Zhou (Soochow University)
RecognitionKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Enhance the generative implicit discourse relation recognition (IDRR) model using subtext.
UTBoost: Rigorous Evaluation of Coding Agents on SWE-Bench
Boxi Yu (Chinese University of Hong Kong Shenzhen), Daniel Kang (University of Illinois Urbana Champaign)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Developed the UTBoost framework to automatically expand SWE-Bench test cases and improve the parser through intramorphic testing.
V-Oracle: Making Progressive Reasoning in Deciphering Oracle Bones for You and Me
Runqi Qiao (Beijing University of Posts and Telecommunications), Honggang Zhang (WeChat Vision, Tencent Inc.)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Convert the oracle interpretation (oracle script) task into a Visual Question Answering (VQA) paradigm, proposing a progressive framework named V-Oracle with a step-by-step reasoning chain and multi-dimensional data augmentation. The Oracle-Bench evaluation benchmark was constructed based on expert annotations, containing 2,834 samples, 13 subdomains, 6 domains, and 3 glyph principles.
Value Portrait: Assessing Language Models’ Values through Psychometrically and Ecologically Valid Items
Jongwook Han (Seoul National University), Yohan Jo (Seoul National University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the Value Portrait benchmark, constructing 520 question-answer pairs by evaluating human similarity to Schwartz's basic values and Big Five traits, to assess the value orientations of 44 LLMs.
Value Residual Learning
Zhanchao Zhou (Zhejiang University), Zhenzhong Lan (Westlake University)
Computational EfficiencyRepresentation LearningTransformerText
🎯 What it does: This paper proposes ResFormer by adding a residual connection to the first-layer value vector before attention to enhance information propagation in deep networks; it also introduces SVFormer which shares value vectors to significantly reduce KV cache.
Value-Spectrum: Quantifying Preferences of Vision-Language Models via Value Decomposition in Social Media Contexts
Jingxuan Li (University of California, Los Angeles), Ying Nian Wu (University of California, Los Angeles)
Recommendation SystemExplainability and InterpretabilityPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the Value-Spectrum benchmark, utilizing VLM agents to automatically capture short video screenshots from TikTok, YouTube Shorts, and Instagram Reels, and implementing retrieval based on Schwartz value dimensions through a CLIP vector database; subsequently, evaluated the inherent preferences and plasticity of VLMs in value orientations using designed questionnaires and role-play strategies.
Velocitune: A Velocity-based Dynamic Domain Reweighting Method for Continual Pre-training
Zheheng Luo (University of Manchester), Peng Cheng (University of Manchester)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose the Velocitune framework, which dynamically adjusts domain sampling ratios during continuous pre-training based on the learning speeds of each domain to balance multi-domain learning progress;
Veracity Bias and Beyond: Uncovering LLMs’ Hidden Beliefs in Problem-Solving Reasoning
Yue Zhou (University of Illinois Chicago), Barbara Di Eugenio (University of Illinois Chicago)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper evaluates whether large language models exhibit 'veracity bias'—associating the correctness of solutions with race or gender identity—in mathematical, programming, common-sense reasoning, and writing tasks through two types of experiments (Attribution Bias and Evaluation Bias). It also discovers that models automatically use stereotypical colors in code generation experiments. The experiments rely solely on prompts and annotations without employing role-playing or social contexts.
VF-Eval: Evaluating Multimodal LLMs for Generating Feedback on AIGC Videos
Tingyu Song (University of Chinese Academy of Sciences), Yilun Zhao (Yale University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes the VF-EVAL benchmark to evaluate the feedback generation and reasoning capabilities of multimodal large language models on AI-generated videos.
VideoVista-CulturalLingo: 360° Horizons-Bridging Cultures, Languages, and Domains in Video Comprehension
Xinyu Chen (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
RecognitionObject DetectionTransformerLarge Language ModelVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Propose the VideoVista-CulturalLingo video evaluation benchmark, covering multilingual, multicultural, and multidomain videos, and design four task categories: event, cultural, object, and scientific; employ an automated annotation framework combining LLMs and visual tools, supplemented by human review, to generate 3,134 QA pairs.
ViGiL3D: A Linguistically Diverse Dataset for 3D Visual Grounding
Austin Wang (Simon Fraser University), Angel X Chang (Simon Fraser University)
Data SynthesisLarge Language ModelPrompt EngineeringVision Language ModelTextPoint CloudMeshBenchmark
🎯 What it does: Construct and release the ViGiL3D diagnostic 3D visual localization dataset to evaluate models' localization capabilities under diverse language prompts.
VISA: Retrieval Augmented Generation with Visual Source Attribution
Xueguang Ma (University Of Waterloo), Jimmy Lin (University Of Waterloo)
GenerationRetrievalSupervised Fine-TuningVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-augmented generation (RAG) system called VISA for visual source attribution, which simultaneously generates answers on retrieved document screenshots and provides bounding boxes highlighting the evidence supporting the answers, enabling visual evidence localization.
Visual Evidence Prompting Mitigates Hallucinations in Large Vision-Language Models
Wei Li, Jieping Ye (Independent Researcher)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a training-free, model-agnostic visual evidence prompting (VEP) method, which uses object detection and scene graph information generated by small visual models as prompts to help large vision-language models (LVLMs) better understand images, thereby reducing hallucinatory outputs.
VisuoThink: Empowering LVLM Reasoning with Multimodal Tree Search
Yikun Wang (Fudan University), Xipeng Qiu (Fudan University)
Vision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposes VisuoThink, a vision-and-language integrated tree search framework designed to enhance the performance of large-scale vision-language models on complex reasoning tasks.
VITAL: A New Dataset for Benchmarking Pluralistic Alignment in Healthcare
Anudeex Shetty (University of Melbourne), Usman Naseem (Macquarie University)
Prompt EngineeringMixture of ExpertsTextBiomedical DataBenchmark
🎯 What it does: Constructed the VITAL diverse value alignment evaluation dataset in the healthcare domain, and benchmarked eight LLMs under three alignment modes
VLM2-Bench: A Closer Look at How Well VLMs Implicitly Link Explicit Matching Visual Cues
Jianshu Zhang (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
Object TrackingRetrievalLarge Language ModelPrompt EngineeringVision Language ModelImageVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the VLM2-Bench benchmark to evaluate the ability of vision-language models (VLMs) to match and track visual cues in multi-image/video scenarios.
VLMInferSlow: Evaluating the Efficiency Robustness of Large Vision-Language Models as a Service
Xiasi Wang (Hong Kong University of Science and Technology), Yuan Yao (Hong Kong University of Science and Technology)
Computational EfficiencyAdversarial AttackVision Language ModelImage
🎯 What it does: Propose VLMInferSlow, a black-box efficiency attack method that significantly increases VLM inference time by generating imperceptible perturbations through gradient-free optimization.
VLSBench: Unveiling Visual Leakage in Multimodal Safety
Xuhao Hu (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Constructed a multi-modal safety benchmark named VLSBench without visual security information leakage, and revealed the existing VSIL issues in current benchmarks through experiments.
VMLU Benchmarks: A comprehensive benchmark toolkit for Vietnamese LLMs
Cuc Thi Bui (Zalo AI), Le-Minh Nguyen (Japan Advanced Institute of Science and Technology)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the VLMU Benchmarks, which include four tasks specifically designed for Vietnamese (Vi-MQA, Vi-SQuAD, Vi-DROP, Vi-Dialog), to comprehensively evaluate the knowledge, reading, reasoning, and dialogue capabilities of Vietnamese LLMs.
VoxEval: Benchmarking the Knowledge Understanding Capabilities of End-to-End Spoken Language Models
Wenqian Cui (Chinese University of Hong Kong), Irwin King (LIGHTSPEED STUDIOS)
Large Language ModelTextBenchmarkChain-of-ThoughtAudio
🎯 What it does: Designed and released the VoxEval benchmark to evaluate the knowledge understanding capabilities of end-to-end speech large models (SLMs), and systematically tested the performance of multiple existing SLMs.
VQAGuider: Guiding Multimodal Large Language Models to Answer Complex Video Questions
Yuyan Chen (Fudan University), Qingpei Guo (Ant Group)
Object DetectionObject TrackingTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Proposes the definition of the complex video question answering (Complex VQA) task, constructs a dedicated CVQA dataset, and designs the VQAGuider framework to guide multi-modal large language models (MLLMs) to decompose, match VideoAPI, and plan paths by breaking down complex video questions into atomic visual tasks (e.g., video object detection, tracking, action recognition, etc.).
VReST: Enhancing Reasoning in Large Vision-Language Models through Tree Search and Self-Reward Mechanism
Congzhi Zhang, Weijiang Yu (Sun Yat Sen University)
Reinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a training-free method called VReST, which enhances the chain-of-thought (CoT) performance of large vision-language models (LVLM) in complex visual reasoning tasks through Monte Carlo Tree Search (MCTS) and a multi-modal self-reward mechanism.
Vulnerability of LLMs to Vertically Aligned Text Manipulations
Zhecheng Li (University of California San Diego), Kai-Wei Chang (University of California Los Angeles)
ClassificationTransformerTextChain-of-Thought
🎯 What it does: Evaluate the robustness of large language models (LLMs) under vertically aligned text inputs, and explore the resulting performance degradation and potential risks.
WAFFLE: Fine-tuning Multi-Modal Model for Automated Front-End Development
Shanchao Liang (Purdue University), Lin Tan (Purdue University)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Developed a multimodal large language model fine-tuning strategy called WAFFLE for frontend development, specifically tailored for the task of generating HTML code from UI design.
Wait, that’s not an option: LLMs Robustness with Incorrect Multiple-Choice Options
Gracjan Góral (University of Warsaw), Paweł Budzianowski (University of Warsaw)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: Studying how LLMs demonstrate reflective judgment ability in multiple-choice questions without correct answers
Walk in Others’ Shoes with a Single Glance: Human-Centric Visual Grounding with Top-View Perspective Transformation
Yuqi Bu (Shenzhen Polytechnic University), Qiong Liu (National University of Singapore)
Object DetectionPose EstimationDepth EstimationRetrievalLarge Language ModelVision Language ModelImageTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the Human Viewpoint Visual Localization (HVG) task and the corresponding InterRef dataset, and designs a Top-View Enhanced Perspective Transformation (TEP) method that infers human viewpoints from a single robot perspective, addressing the sensitivity of visual reference frameworks to differences between robot and human viewpoints.
Warmup Generations: A Task-Agnostic Approach for Guiding Sequence-to-Sequence Learning with Unsupervised Initial State Generation
Senyu Li (Mila - Quebec AI Institute), David Ifeoluwa Adelani (Mila - Quebec AI Institute)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose a task-agnostic Warmup Generation framework, where the model first generates unsupervised 'warmup' sequences before generating the final sequence to guide the output.
WarriorCoder: Learning from Expert Battles to Augment Code Large Language Models
Huawen Feng (South China University of Technology), Qi Zhang (Microsoft)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposes the WarriorCoder framework, which generates training data through battles between expert code LLMs, enabling the target model to learn from the winning responses to enhance coding capabilities;
Watching the Watchers: Exposing Gender Disparities in Machine Translation Quality Estimation
Emmanouil Zaranis (Instituto de Telecomunicações), Andre Martins
Prompt EngineeringText
🎯 What it does: Investigated gender bias in quality estimation (QE) metrics and evaluated their impact on the machine translation pipeline
Watermarking Large Language Models: An Unbiased and Low-risk Method
Minjia Mao (University of Delaware), Michael Chau (University of Hong Kong)
GenerationSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposes a novel unbiased watermarking method called Sampling One Then Accepting (STA-1), and extends it to a stronger version named Sampling M Then Accepting (STA-M); the method randomly divides green and red token lists during LLM generation, samples once, and accepts if it falls into the green list or resamples if it falls into the red list, maintaining the original distribution expectation; during the detection phase, it employs a z-test based on green token counting and provides an upper bound for Type-II error using the Gini index; additionally, it provides a low-risk analysis, proving that STA-1 has lower risk in low-entropy scenarios; experiments verify that STA-1 matches or exceeds existing watermarking methods in terms of text quality, detection efficiency, and anti-attack capabilities.
WavRAG: Audio-Integrated Retrieval Augmented Generation for Spoken Dialogue Models
Yifu Chen (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationRetrievalTransformerContrastive LearningTextMultimodalityRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Designed WavRAG, an end-to-end retrieval-augmented generation framework with raw audio input;
We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning?
Runqi Qiao (Beijing University Of Posts And Telecommunications), Honggang Zhang (Beijing University Of Posts And Telecommunications)
Large Language ModelMultimodalityBenchmark
🎯 What it does: Constructed the WE-MATH benchmark, systematically decomposing visual math problems into single-step subproblems and providing fine-grained evaluation;
Weaving Context Across Images: Improving Vision-Language Models through Focus-Centric Visual Chains
Juntian Zhang (Renmin University of China), Rui Yan (Xiaomi Inc.)
Data SynthesisLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes a Focus-Centric Visual Chain to decompose multi-image tasks, enhancing the performance of Vision-Language Models (VLMs) in multi-image reasoning.
WebWalker: Benchmarking LLMs in Web Traversal
Jialong Wu (Southeast University), Fei Huang (Alibaba Group)
RetrievalLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the WebWalkerQA benchmark to evaluate large language models' information retrieval and reasoning capabilities in complex multi-step web browsing, and proposed the WebWalker multi-agent framework to simulate human web navigation;
WET: Overcoming Paraphrasing Vulnerabilities in Embeddings-as-a-Service with Linear Transformation Watermarks
Anudeex Shetty (University of Melbourne), Jey Han Lau (University of Melbourne)
ClassificationRepresentation LearningAdversarial AttackLarge Language ModelText
🎯 What it does: The paper proposes a new plagiarism attack method for Embedding-as-a-Service—using averaged multiple paraphrase embeddings to erase existing watermarks, and designs a watermarking method called WET based on linear transformation that can resist such attacks.
What are the Essential Factors in Crafting Effective Long Context Multi-Hop Instruction Datasets? Insights and Best Practices
Zhi Chen (Harbin Institute of Technology), Dahua Lin (Harbin Institute of Technology)
Data SynthesisData-Centric LearningTransformerLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a multi-agent interactive multi-hop generation framework (MIMG) for synthesizing high-quality, long-context, multi-hop instruction datasets;
What Happened in LLMs Layers when Trained for Fast vs. Slow Thinking: A Gradient Perspective
Ming Li (University of Maryland), Tianyi Zhou (University of Maryland)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Investigated the hierarchical gradient behavior of large language models under different training strategies (fast thinking vs slow thinking, correct vs incorrect answers, base models vs instruction-tuned models), and quantified gradient features using spectral metrics such as nuclear norm.
What is Stigma Attributed to? A Theory-Grounded, Expert-Annotated Interview Corpus for Demystifying Mental-Health Stigma
Han Meng (National University of Singapore), Yi-Chieh Lee (National University of Singapore)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Built a mental health stigma interview corpus based on a chatbot, with expert annotations and theory-driven fine-grained labels; simultaneously used this corpus to conduct benchmark experiments on multiple large language models and fine-tuned RoBERTa for stigma detection.
What Is That Talk About? A Video-to-Text Summarization Dataset for Scientific Presentations
Dongqi Liu (Saarland University), Vera Demberg (Max Planck Institute for Informatics)
GenerationTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the VISTA dataset for the text summarization task of scientific conference presentation videos, evaluated multimodal and text models on this dataset, and further introduced a planning-based generation framework to improve summary quality
What Makes a Good Natural Language Prompt?
Do Xuan Long (National University of Singapore), Min-Yen Kan (National University of Singapore)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperChain-of-Thought
🎯 What it does: Conduct a meta-analysis of natural language prompting-related papers from 2022-2025, constructing an evaluation framework with 21 attributes, analyzing the impact of attributes on LLM performance, and verifying the effectiveness of single-attribute improvements in reasoning tasks through experiments.
What Matters in Evaluating Book-Length Stories? A Systematic Study of Long Story Evaluation
Dingyi Yang (Renmin University of China), Qin Jin (Renmin University of China)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes and constructs the first automatic evaluation benchmark for long stories (over 100K tokens), named LongStoryEval, and builds a multi-dimensional evaluation standard based on real reader reviews; on this basis, three long-story evaluation methods (aggregation, incremental update, and summarization) are studied, and an efficient NovelCritique 8B model is developed;
What Really Matters in Many-Shot Attacks? An Empirical Study of Long-Context Vulnerabilities in LLMs
Sangyeop Kim (Coxwave), Kimin Lee (KAIST)
Safty and PrivacyAdversarial AttackTransformerPrompt EngineeringText
🎯 What it does: This paper investigates the security issues of utilizing long contexts for many-shot jailbreaking in large language models (LLMs), systematically evaluates the impact of different context lengths (up to 128K tokens) on attack effectiveness, and analyzes factors such as example density, theme, harm level, and text form on model vulnerability.
What’s the Difference? Supporting Users in Identifying the Effects of Prompt and Model Changes Through Token Patterns
Michael A. Hedderich (LMU Munich), Barbara Plank (LMU Munich)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Propose the Spotlight method, which combines data mining to automatically extract token patterns, helping users identify systematic differences in LLM outputs under prompt or model variations.
When Backdoors Speak: Understanding LLM Backdoor Attacks Through Model-Generated Explanations
Huaizhi Ge (Columbia University), Ruixiang Tang (Rutgers University)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The study utilizes natural language explanations generated by large language models (LLM) to detect and analyze backdoor attacks, exploring differences and internal mechanisms in explanations generated under backdoor triggers.
When GPT Spills the Tea: Comprehensive Assessment of Knowledge File Leakage in GPTs
Xinyue Shen (CISPA Helmholtz Center for Information Security), Yang Zhang (CISPA Helmholtz Center for Information Security)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study systematically evaluates the risk of GPT knowledge file leakage based on the DSPM process, identifying five leakage vectors: metadata, GPT initialization, retrieval, sandbox execution environment, and prompts. Experiments demonstrate that the Code Interpreter can achieve 95.95% original file download success rate;
When Harry Meets Superman: The Role of The Interlocutor in Persona-Based Dialogue Generation
Daniela Occhipinti (Fondazione Bruno Kessler), Malvina Nissim (Fondazione Bruno Kessler)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Investigated how models adapt to both the target speaker and interlocutor's identity information in character-based dialogue generation;
When People are Floods: Analyzing Dehumanizing Metaphors in Immigration Discourse with Large Language Models
Julia Mendelsohn (University of Maryland), Ceren Budak (University of Michigan)
ClassificationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Studied the use of dehumanizing metaphors in immigration discourse and analyzed their relationship with political ideology and tweet interactions (retweets/likes).
When the LM misunderstood the human chuckled: Analyzing garden path effects in humans and language models
Samuel Joseph Amouyal (Tel Aviv University), Jonathan Berant (Tel Aviv University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextMultimodality
🎯 What it does: Systematically compare the performance of humans and various large language models (LLMs) in understanding specific garden-path sentences (object/subject structure), using the same comprehension tasks (answering comprehension questions, sentence splitting, image generation) to evaluate misunderstanding patterns of both parties;
When to Speak, When to Abstain: Contrastive Decoding with Abstention
Hyuhng Joon Kim (Seoul National University), Taeuk Kim (Hanyang University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposed a training-free contrastive decoding method (Contrastive Decoding with Abstention, CDA), enabling large language models to adaptively generate answers or proactively abstain from answering when lacking relevant parameters or contextual knowledge.
Where Are We? Evaluating LLM Performance on African Languages
Ife Adebara (University of British Columbia), Muhammad Abdul-Mageed (University of British Columbia)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Designed and released the SAHARA benchmark to evaluate the performance of African languages in large language models and conducted an empirical analysis of the relationship between language policy and data availability.
Which Demographics do LLMs Default to During Annotation?
Johannes Schäfer (University of Bamberg), Roman Klinger (University of Bamberg)
Explainability and InterpretabilityTransformerPrompt EngineeringText
🎯 What it does: Investigate the default socio-demographic attributes in LLM text annotations and evaluate their impact on annotation outcomes through socio-demographic prompting and placebo prompting.
Which of These Best Describes Multiple Choice Evaluation with LLMs? A) Forced B) Flawed C) Fixable D) All of the Above
Nishant Balepur (University of Maryland), Jordan Lee Boyd-Graber
Data-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Analyze and criticize the limitations of traditional multiple-choice questions (MCQA) in LLM evaluation, proposing improved generative formats and dataset improvement methods.
WhiSPA: Semantically and Psychologically Aligned Whisper with Self-Supervised Contrastive and Student-Teacher Learning
Rajath Rao (Stony Brook University), H. Andrew Schwartz (Stony Brook University)
Knowledge DistillationRepresentation LearningTransformerContrastive LearningTextMultimodalityAudio
🎯 What it does: Constructed WhiSPA, a model that aligns Whisper's acoustic representations with SBERT text semantics and psychological dimension embeddings through self-supervised contrastive learning, aiming to enable speech models to directly capture deep semantics and psychological information without requiring subsequent text language models.
White Men Lead, Black Women Help? Benchmarking and Mitigating Language Agency Social Biases in LLMs
Yixin Wan (University of California, Los Angeles), Kai-Wei Chang (University of California, Los Angeles)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the Language Agency Bias Evaluation (LABE) benchmark for systematically assessing language agent (agentic/communal) biases related to gender, race, and intersectional identities in large language models' generated text, and trains a high-accuracy language agent classifier using the constructed LAC dataset. Additionally, the paper introduces the Mitigation via Selective Rewrite (MSR) method, which identifies and rewrites communal sentences to enhance agentic properties using the classifier.
Who Writes What: Unveiling the Impact of Author Roles on AI-generated Text Detection
Jiatao Li (Peking University), Xiaojun Wan (Peking University)
ClassificationAnomaly DetectionLarge Language ModelTextBenchmark
🎯 What it does: To address AI-generated text detection, researchers evaluated the bias of detectors across four dimensions: gender of human authors, CEFR level, academic field, and language context, and constructed a new dataset containing 66,794 texts (human writing + 12 types of LLM-generated texts).
Whose Boat Does it Float? Improving Personalization in Preference Tuning via Inferred User Personas
Nishant Balepur (University of Maryland), Jordan Lee Boyd-Graber
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies how to leverage abduction to infer users' underlying personalized features (persona) from preference data, and use these inferred personas to enhance preference tuning of large language models, thereby improving the model's personalized performance.
Why Prompt Design Matters and Works: A Complexity Analysis of Prompt Search Space in LLMs
Xiang Zhang (University of British Columbia), Dujian Ding (University of British Columbia)
Explainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper clarifies the role of prompts in Chain-of-Thought (CoT) reasoning through theoretical analysis and experiments, and proposes a 'Prompt Space and Answer Space' framework to study how to search for optimal prompts in the prompt space to enhance the logical reasoning performance of large language models (LLMs).
Why Safeguarded Ships Run Aground? Aligned Large Language Models’ Safety Mechanisms Tend to Be Anchored in The Template Region
Chak Tou Leong (Hong Kong Polytechnic University), Wenjie Li (Hong Kong Polytechnic University)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerText
🎯 What it does: This paper investigates the over-reliance on template regions during the safety alignment of large language models—referred to as Template-Anchored Safety Alignment (TASA)—and reveals that this issue leads to vulnerabilities during inference.
WiCkeD: A Simple Method to Make Multiple Choice Benchmarks More Challenging
Ahmed Elhady (HiTZ Center, University of Basque Country), Mikel Artetxe (HiTZ Center, University of Basque Country)
Prompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the WiCkeD method, which generates more challenging benchmarks by randomly replacing one option in multiple-choice questions with 'None of the above,' and evaluates it on 18 open-source LLMs.
WinSpot: GUI Grounding Benchmark with Multimodal Large Language Models
Zheng Hui (Microsoft), Kazuhito Koishida
RecognitionObject DetectionTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Developed the first Windows GUI localization benchmark, WinSpot, using a two-phase automatic annotation and manual verification based on a multi-modal large language model (MLLM), generating over 5,000 coordinate-instruction pairs.
Wizard of Shopping: Target-Oriented E-commerce Dialogue Generation with Decision Tree Branching
Xiangci Li (AWS AI Labs), Shervin Malmasi (Amazon.com Inc)
GenerationRecommendation SystemTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the TRACER method, which leverages LLM and decision tree planning to generate goal-oriented shopping dialogues, and release a large-scale Wizard of Shopping dataset.
Words of Warmth: Trust and Sociability Norms for over 26k English Words
Saif M. Mohammad (National Research Council Canada)
Data-Centric LearningTextTabular
🎯 What it does: Constructed a Warmth (Trust and Sociability) Lexicon covering 26,000+ English words, combined with the Competence dimension to form the complete WCTS Lexicon;
World Modeling Makes a Better Planner: Dual Preference Optimization for Embodied Task Planning
Siyin Wang (Fudan University), Xipeng Qiu (Fudan University)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelWorld ModelMultimodality
🎯 What it does: This paper proposes a Dual Preference Optimization (D²PO) framework by combining large vision-language models with world model learning to improve the accuracy and efficiency of embedded task planning.
Writing Like the Best: Exemplar-Based Expository Text Generation
Yuxiang Liu (University of Illinois at Urbana-Champaign), Kevin Chen-Chuan Chang (University of Illinois at Urbana-Champaign)
GenerationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Designed the Exemplar-Based Expository Text Generation Task, generating new long expository texts using long text examples.
X-TURING: Towards an Enhanced and Efficient Turing Test for Long-Term Dialogue Agents
Weiqi Wu (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
Computational EfficiencyLarge Language ModelText
🎯 What it does: Propose the X-TURING framework, utilizing burst dialogue mode and pseudo-dialogue generation to conduct long-duration Turing tests on large language models (LLMs);
XDAC: XAI-Driven Detection and Attribution of LLM-Generated News Comments in Korean
Wooyoung Go (National Security Research Institute), Yongdae Kim (KAIST)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes the XDAC framework for detecting and attributing LLM-generated comments in Korean news comments.
YESciEval: Robust LLM-as-a-Judge for Scientific Question Answering
Jennifer D’Souza, Quentin Münch (Leibniz Universität Hannover)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposed an open-source evaluation framework named YESciEval, which enables large language models (LLMs) to act as judges in science question-answering (scienceQ&A) tasks, addressing the issues of optimism bias and robustness in LLM evaluation.
Your Model is Overconfident, and Other Lies We Tell Ourselves
Timothee Mickus (University of Helsinki), Raúl Vázquez (ICANS Strasbourg)
Explainability and InterpretabilityTransformerText
🎯 What it does: This paper systematically investigates various metrics for evaluating data difficulty and uncertainty, and explores their relationship with annotator dissensus. By training and evaluating 29 models on two large datasets, ChaosNLI and DynaSent, the authors compare the interrelationships among 11 metrics, including human annotation inconsistency, model self-consistency, training dynamics, and Conformal Prediction (CP); further attempts are made to use soft label training to enhance models' ability to capture human uncertainty.
YuLan-Mini: Pushing the Limits of Open Data-efficient Language Model
Hu Yiwen, Ji-Rong Wen (Renmin University of China)
Data SynthesisComputational EfficiencyData-Centric LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Built and released an open-source foundation model named YuLan-Mini with 2.4B parameters, and provided complete training details
Zero-Shot Text-to-Speech for Vietnamese
Thi Vu (Movian AI), Dat Quoc Nguyen (Movian AI)
GenerationData SynthesisTextAudio
🎯 What it does: Investigated and evaluated three advanced models trained on the PhoAudiobook dataset for zero-shot Vietnamese text-to-speech synthesis.
ZIPA: A family of efficient models for multilingual phone recognition
Jian Zhu (University of British Columbia), David R. Mortensen (Carnegie Mellon University)
RecognitionComputational EfficiencyTransformerAudio
🎯 What it does: This paper proposes the ZIPA model and the IPAPACK++ corpus, aiming to efficiently identify phonemes in multilingual speech.
𝛿-Stance: A Large-Scale Real World Dataset of Stances in Legal Argumentation
Ankita Gupta (University of Massachusetts Amherst), Brendan O’Connor
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper constructs a large-scale legal argument stance dataset /u1D6FF-Stance, containing millions of argument context-case summary-stance triplets, and evaluates various NLP methods based on this task.