ICLR 2026 Papers — Page 53
International Conference on Learning Representations · 5356 papers
VoG: Enhancing LLM Reasoning through Stepwise Verification on Knowledge Graphs
Wenxin Zhao (Hong Kong University of Science and Technology), Lei Chen (Hong Kong University of Science and Technology)
Graph Neural NetworkTransformerLarge Language ModelAgentic AITextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the VoG framework, which enhances the accuracy and robustness of multi-hop question answering through a cyclic process of progressive retrieval, verification, and revision of knowledge graphs guided by LLM-generated reasoning plans.
VoMP: Predicting Volumetric Mechanical Property Fields
Rishit Dagli (NVIDIA), Maria Shugrina (NVIDIA)
Computational EfficiencyTransformerVision Language ModelNeural Radiance FieldAuto EncoderGaussian SplattingPoint CloudMeshTabularBenchmarkPhysics Related
🎯 What it does: Trained and deployed an end-to-end VoMP model for rapidly predicting physical material parameters (Young's modulus E, Poisson's ratio ν, density ρ) within voxelized volumes from arbitrary voxelizable 3D representations (meshes, Gaussian Splats, NeRF, SDF, etc.).
VowelPrompt: Hearing Speech Emotions from Text via Vowel-level Prosodic Augmentation
Yancheng Wang (Meta Superintelligence Labs), Yingzhen Yang (Arizona State University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextAudio
🎯 What it does: Propose a sentiment recognition framework (VowelPrompt) based on vowel-level fine-grained speech features and large language models (LLMs). By aligning vowel segments in audio, low-level acoustic features such as pitch, intensity, and duration are extracted, discretized into natural language descriptions, and used as interpretable prompts for LLMs.
VoxPrivacy: A Benchmark for Evaluating Interactional Privacy of Speech Language Models
Yuxiang Wang (Chinese University of Hong Kong, Shenzhen), Zhizheng Wu (Chinese University of Hong Kong, Shenzhen)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkAudio
🎯 What it does: Propose the VoxPrivacy benchmark, which evaluates the performance of multi-user speech models on interactional privacy through three-tiered tasks; construct a 32.86-hour bilingual synthetic dataset (7,107 samples) and conduct large-scale evaluations on nine SLMs.
VPI-Bench: Visual Prompt Injection Attacks for Computer-Use Agents
Tri Cao (National University of Singapore), Bryan Hooi (Cyber Emerging Tech and R&D)
Safty and PrivacyAdversarial AttackLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringImageTextBenchmark
🎯 What it does: Constructed the VPI attack model and introduced the VPI-Bench benchmark to evaluate the robustness of Computer-Use Agents and Browser-Use Agents under visual prompt injection (VPI) attacks.
VQ-Transplant: Efficient VQ-Module Integration for Pre-trained Visual Tokenizers
Xianghong Fang, Tim G. J. Rudner (University Of Toronto)
GenerationComputational EfficiencySupervised Fine-TuningDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Proposes the VQ-Transplant framework, enabling fast replacement of the VQ module in pre-trained visual tokenizers without full model retraining, and introduces the MMD-VQ quantization method compatible with this framework based on Maximum Mean Discrepancy (MMD).
VSF: Simple, Efficient, and Effective Negative Guidance in Few-Step Image Generation Models By Value Sign Flip
Wenqi Marshall Guo (University Of British Columbia), Shan Du (University Of British Columbia)
GenerationTransformerPrompt EngineeringDiffusion modelFlow-based ModelImageVideoBenchmark
🎯 What it does: Proposes a negative prompt guidance method called Value Sign Flip (VSF), which dynamically suppresses unwanted content in diffusion or flow-matching image/video generation models within a few steps (1-8 steps) by flipping attention values.
VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use
Mingyuan Wu (University of Illinois Urbana-Champaign), Klara Nahrstedt (University of Illinois Urbana-Champaign)
Reinforcement LearningAgentic AIPrompt EngineeringVision Language ModelMultimodalityTabularChain-of-Thought
🎯 What it does: Fine-tune visual language models using reinforcement learning to automatically generate and utilize intermediate visual steps during reasoning, enabling the ability to 'think with images.'
VUDG: A Dataset for Video Understanding Domain Generalization
Ziyi Wang (Beijing Institute of Technology), Xinxiao Wu (Beijing Institute of Technology)
Domain AdaptationLarge Language ModelVision Language ModelVideoBenchmark
🎯 What it does: This paper proposes the VUDG dataset for evaluating domain generalization capability in video understanding.
Vulcan: Crafting Compact Class-Specific Vision Transformers For Edge Intelligence
Ziteng Wei (Swinburne University of Technology), Yun Yang (Swinburne University of Technology)
ClassificationObject DetectionSegmentationComputational EfficiencyTransformerImage
🎯 What it does: This paper proposes a novel post-training pruning method called Vulcan, which can extract compact models tailored to specific target classes from pre-trained Vision Transformers while meeting resource budget constraints;
W-EDIT: A Wavelet-Based Frequency-Aware Framework for Text-Driven Image Editing
Jiahui Sun (University of Chinese Academy of Sciences), Jing Liu (Chinese Academy of Sciences)
GenerationTransformerDiffusion modelFlow-based ModelImageText
🎯 What it does: Propose an untrained text-driven image editing framework based on wavelet frequency domain, named W-Edit, which can perform various editing operations while maintaining global structural consistency;
WAFT: Warping-Alone Field Transforms for Optical Flow
Yihan Wang (Princeton University), Jia Deng (Princeton University)
Convolutional Neural NetworkTransformerOptical FlowVideoBenchmark
🎯 What it does: Propose WAFT, an iterative optical flow estimation framework based on warping, which removes the high-cost cost volume and performs multi-step iterative updates of the flow using high-resolution feature warping.
WALT: Web Agents that Learn Tools
Viraj Prabhu (Salesforce Research), Ran Xu (Salesforce Research)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the WALT framework, which automatically generates callable tools by reverse engineering website functions (such as search, filtering, and publishing), replacing traditional low-level UI operations;
WARC-Bench: Web Archive based Benchmark for GUI Subtask Executions
Sanjari Srivastava (Uniphore), Peng Qi (Uniphore)
Data SynthesisTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision-Language-Action ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the WARC-Bench benchmark, leveraging Web Archive files to provide a reproducible short-term GUI subtask environment, and evaluate multimodal models using Subtask Vision Agent (SVA).
WARP: Weight Teleportation for Attack-Resilient Unlearning Protocols
Mohammad mahdi Maheri (Imperial College London), Hamed Haddadi (Imperial College London)
Safty and PrivacyConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed the WARP (Weight Jumping) defense mechanism to enhance privacy security during the approximate machine unlearning process, reducing the success rates of membership inference and data reconstruction attacks.
Watch the Weights: Unsupervised monitoring and control of fine-tuned LLMs
Ziqian Zhong (Carnegie Mellon University), Aditi Raghunathan (Carnegie Mellon University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the WeightWatch method based on weight differences, which monitors and controls model abnormal behaviors (such as backdoors and forgetting) by utilizing the singular vectors of weight differences between fine-tuned models and base models, without accessing training data.
Watch your steps: Dormant Adversarial Behaviors that Activate upon LLM Finetuning
Thibaud Gloaguen (ETH Zurich), Martin Vechev (ETH Zurich)
Adversarial AttackMeta LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Researchers proposed and implemented an attack method named FAB, which can implant malicious behaviors into a 'seemingly secure' large language model (LLM) without altering its original performance. These malicious behaviors can be activated after downstream users perform fine-tuning.
WaterDrum: Watermark-based Data-centric Unlearning Metric
Xinyang Lu (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Safty and PrivacyExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a text watermark-based forgetting measurement method for LLM data centers called WaterDrum, and provided an interpretable and sustainable evaluation framework for forgetting algorithms.
Watermark-based Attribution of AI-Generated Content
Zhengyuan Jiang (Duke University), Neil Zhenqiang Gong (Duke University)
GenerationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes a user-level attribution method based on watermarks, capable of tracing the specific user who generated AI content;
Watermarking Diffusion Language Models
Thibaud Gloaguen (ETH Zurich), Martin Vechev (ETH Zurich)
Safty and PrivacyLarge Language ModelDiffusion modelText
🎯 What it does: Proposed a watermarking scheme applicable to diffusion language models (DLM), achieving retrievable watermarks in generated text by adjusting the expected value of context hashes and prediction bias.
WATS: Wavelet-Aware Temperature Scaling for Reliable Graph Neural Networks
Xiaoyang Li (Independent Researcher), Chang Xu (University of Sydney)
ClassificationGraph Neural NetworkGraph
🎯 What it does: Propose a post-calibration framework named WATS, which utilizes graph heat kernel wavelet features to predict temperatures for each node, thereby calibrating the confidence of GNN.
WAVE: Learning Unified & Versatile Audio-Visual Embeddings with Multimodal LLM
Changli Tang (Tsinghua University), Chao Zhang (Tsinghua University)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Propose the WAVE model, constructing a unified audio-visual-text embedding space and supporting cross-modal retrieval and prompt-aware embeddings.
WavefrontDiffusion: Dynamic Decoding Schedule for Improved Reasoning
Haojin Yang (Peking University), Yiwei Wang (University of California Merced)
GenerationComputational EfficiencyDiffusion modelText
🎯 What it does: Proposes a dynamic decoding scheduling method called WavefrontDiffusion based on wavefront expansion, aiming to enhance the semantic coherence and generation quality of Diffusion Language Models in text generation.
Wavelet Predictive Representations for Non-Stationary Reinforcement Learning
Min Wang (Beijing Institute of Technology), Mingzhong Wang (University of the Sunshine Coast)
Representation LearningReinforcement LearningSequential
🎯 What it does: Propose WISDOM, which captures multi-scale features of task evolution through a learnable wavelet transform network, achieving fast adaptation in non-stationary reinforcement learning.
WavePolyp: Video Polyp Segmentation via Hierarchical Wavelet-Based Feature Aggregation and Inter-Frame Divergence Perception
Yuhua Zhang (Shenzhen University), Jing Qin (Hong Kong Polytechnic University)
SegmentationTransformerVideoBiomedical Data
🎯 What it does: Proposed a novel video adenoma segmentation network called WavePolyp for automatically segmenting adenomas in colonoscopy videos.
wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models
Xiaohang Tang (University College London), Ilija Bogunovic (Universität Basel)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: Proposed a ratio-free weighted policy optimization method (wd1) for fine-tuning discrete diffusion large language models to enhance reasoning capabilities
We-Math 2.0: A Versatile MathBook System for Incentivizing Visual Mathematical Reasoning
Runqi Qiao (Beijing University Of Posts And Telecommunications), Honggang Zhang (Beijing University Of Posts And Telecommunications)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose the WE-MATH 2.0 framework, construct a 5-level mathematical knowledge system and a bidirectional problem-image expansion dataset, and enhance the visual mathematical reasoning ability of multimodal large language models through two-stage reinforcement learning.
Weak Correlations as the Underlying Principle for Linearization of Gradient-Based Learning Systems
Ori Shem-Ur (Tel Aviv University), Yaron Oz (Tel Aviv University)
OptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This paper investigates the parameter linearization phenomenon exhibited by gradient descent learning algorithms under the extreme limit of network width, proposing that weak derivative correlation is the fundamental mechanism causing this phenomenon, and provides an equivalent theorem between linearization and correlation;
Weak-to-Strong Diffusion with Reflection
Lichen Bai (xLeaF Lab, Hong Kong University of Science and Technology (Guangzhou)), Zeke Xie (xLeaF Lab, Hong Kong University of Science and Technology (Guangzhou))
GenerationMixture of ExpertsDiffusion modelImageVideo
🎯 What it does: Propose the Weak‑to‑Strong Diffusion (W2SD) framework, which enhances generation quality by alternating between denoising with a strong model and inversion with a weak model during sampling to reflectively correct latent variables, thereby narrowing the gradient gap between the trained model and the ideal model.
Weak-to-Strong Generalization with Failure Trajectories
Ruimeng Ye (University Of Tulsa), Bo Hui (Northwestern University)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningText
🎯 What it does: Propose a generalization framework from weak models to strong models, leveraging the success and failure trajectories generated by weak models to construct a hierarchical trajectory tree, and fine-tuning the strong model using TreeDPO or MCTS to enhance its reasoning and decision-making capabilities.
WearVox: An Egocentric Multichannel Voice Assistant Benchmark for Wearables
Zhaojiang Lin (Meta), Xin Luna Dong (Meta)
TransformerLarge Language ModelBenchmarkAudio
🎯 What it does: Propose the WearVox benchmark, collecting 3,842 multi-channel egocentric audio samples from AI glasses and evaluating five wearable speech tasks.
Web-CogReasoner: Towards Multimodal Knowledge-Induced Cognitive Reasoning for Web Agents
Yuhan Guo (Southwestern University of Finance and Economics), Yong Dai (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the Web-CogKnowledge framework, decomposing the learning of Web agents into two stages: knowledge acquisition and cognitive processes, and constructed Web-CogDataset and Web-CogBench, training a multimodal Web agent Web-CogReasoner based on knowledge-driven Chain-of-Thought (CoT);
WebArbiter: A Generative Reasoning Process Reward Model for Web Agents
Yao Zhang (LMU Munich), Volker Tresp (LMU Munich)
Knowledge DistillationTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Introduces WebArbiter, a principle-induction reasoning-based process reward model for stepwise reward evaluation of Web agents.
WebDevJudge: Evaluating (M)LLMs as Critiques for Web Development Quality
Chunyang Li (Tencent), Han Hu (Nanyang Technological University)
AI Code AssistantLarge Language ModelAgentic AIImageTextMultimodalityBenchmark
🎯 What it does: Propose the WEBDEVJUDGE benchmark to evaluate the performance of LLMs as reviewers in web development.
WebDS: An End-to-End Benchmark for Web-based Data Science
Ethan Hsu (Stanford University), Christopher D Manning (Stanford University)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose the WebDS benchmark, which includes 870 web data science tasks across 29 public websites, evaluating the performance of LLM agents in a complete data science workflow (browsing, data acquisition, analysis, output).
WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents
Sicheng Fan (Fudan University), Dehan Kong (Chinese University of Hong Kong)
Data SynthesisCompressionTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposed and implemented WebFactory: a closed-loop automated pipeline that compresses internet knowledge from large language models (LLMs) into executable graphical user interface (GUI) agents.
WebGen-Agent: Enhancing Interactive Website Generation with Multi-Level Feedback and Step-Level Reinforcement Learning
Zimu Lu (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Proposed and implemented WebGen-Agent, a code agent that iteratively generates and optimizes website code using LLM through multi-layer visual feedback (screenshots + GUI interaction) and backtracking/best-step selection mechanisms.
WebSailor-V2: Bridging the Chasm to Proprietary Agents via Synthetic Data and Scalable Reinforcement Learning
Kuan Li (Hong Kong University of Science and Technology), Jingren Zhou
Data SynthesisTransformerSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Develop and deploy WebSailor-V2: construct a dense knowledge graph dataset SailorFog-QA-V2, complete SFT (Supervised Fine-Tuning) and symmetric dual-environment RL (Reinforcement Learning) on Qwen3-30B-A3B, significantly improving the performance of open-source Web agents in multi-step reasoning and information retrieval tasks.
Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels
Zhepeng Cen (Salesforce AI Research), Weiran Yao (Carnegie Mellon University)
Data SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes an automated and scalable data pipeline, Webscale-RL, which converts large-scale pre-trained corpora into verifiable question-answer pairs for reinforcement learning training.
WebSeer: Training Deeper Search Agents through Reinforcement Learning with Self-Reflection
Guanzhong He (Tsinghua University), Juanzi Li (Tsinghua University)
TransformerSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Developed a search agent called WebSeer, which employs self-reflective reinforcement learning (SRRL) to perform multi-step reasoning in real-world Web environments and learns deeper tool call chains through two-phase training (cold start + RL).
WebShaper: Agentically Data Synthesizing via Information-Seeking Formalization
Zhengwei Tao (Peking University), Jingren Zhou (Tongyi Lab, Alibaba Group)
Data SynthesisLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study proposes and implements the WebShaper framework, which formalizes information-seeking tasks using set theory-based knowledge projection (KP), and automatically generates diverse high-quality training data through hierarchical expansion strategies;
WebWatcher: Breaking New Frontiers of Vision-Language Deep Research Agent
Xinyu Geng (The Hong Kong University of Science and Technology), Jingren Zhou (Tongyi Lab, Alibaba Group)
RetrievalLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Proposed a multimodal deep research agent called WebWatcher, capable of performing joint reasoning on visual and textual information and utilizing multiple tools to accomplish complex information retrieval and inference tasks.
WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research
Zijian Li (Hong Kong University of Science and Technology), Jingren Zhou (Tongyi Lab, Alibaba Group)
Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the WebWeaver dual-agent framework, supporting automatic report generation for open-ended deep research problems
Weight Decay may matter more than µP for Learning Rate Transfer in Practice
Atli Kosson (EPFL), Xi Chen (Amazon FAR)
OptimizationHyperparameter SearchConvolutional Neural NetworkTransformerLarge Language ModelImageText
🎯 What it does: Studied the role of maximal update parameterization (µP) and weight decay when transferring optimal learning rates from small models to large models in large neural networks (especially large language models), and found that weight decay is more critical in practice.
Weight Space Representation Learning on Diverse NeRF Architectures
Francesco Ballerini (University of Bologna), Samuele Salti (University of Bologna)
ClassificationRetrievalRepresentation LearningGraph Neural NetworkNeural Radiance FieldContrastive LearningImageGraph
🎯 What it does: Proposed the first weight-space representation learning framework capable of handling multiple NeRF architectures (MLP, tri-plane, hash-table), and applied it to tasks such as classification, retrieval, short/detailed description generation, and single-turn question answering.
Weight-Space Linear Recurrent Neural Networks
Roussel Desmond Nzoyem, Tom Deakin (University of Bristol)
ClassificationRecurrent Neural NetworkImageTabularTime SeriesPhysics Related
🎯 What it does: Proposed the WARP (Weight-space Adaptive Recurrent Prediction) model, which explicitly models the hidden states of RNNs as weights of an auxiliary network, and utilizes input differences to drive linear recursion, achieving gradient-free test-time adaptation and context learning.
Welfarist Formulations for Diverse Similarity Search
Siddharth Barman (Indian Institute of Science), Kirankumar Shiragur (Microsoft Research)
RetrievalOptimization
🎯 What it does: This paper proposes an approximate nearest neighbor search framework based on welfare functions, which can achieve attribute-level diversity while maintaining query relevance.
WeTok: Powerful Discrete Tokenization for High-Fidelity Visual Reconstruction
Shaobin Zhuang (Shanghai Jiao Tong University), Yali Wang (Shenzhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
RestorationGenerationCompressionConvolutional Neural NetworkTransformerAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Designed and implemented a discrete visual tokenizer, WeTok, with high compression ratio and high-quality reconstruction, supporting zero-shot image reconstruction and generation at compression ratios of 400% and above.
WFR-FM: Simulation-Free Dynamic Unbalanced Optimal Transport
Qiangwei Peng (Peking University), Peijie Zhou (Peking University)
OptimizationFlow-based ModelRectified FlowBiomedical Data
🎯 What it does: Propose the WFR-FM framework, which utilizes flow matching methods to jointly learn velocity fields and growth rates, solving dynamic unbalanced Wasserstein-Fisher-Rao optimal transport without simulation, applied to single-cell trajectory inference.
What "Not" to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging
Inha Kang (KAIST AI), Hyunjung Shim (KAIST AI)
Object DetectionSupervised Fine-TuningVision Language ModelMultimodalityBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: This study investigates the 'affirmative bias' of visual language models in understanding negative statements, proposing a systematic COVAND dataset pipeline and a NEGTOME module based on text token merging. It combines this with LoRA (Low-Rank Adapter) to perform parameter-efficient fine-tuning of detection models, thereby improving negative detection performance.
What Do Large Language Models Know About Opinions?
Erfan Jahanparast (University of California, Berkeley), Serina Chang (University of California, Berkeley)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This study evaluates the internal knowledge of large language models (LLMs) across 22 groups by training probes and sparse autoencoders, revealing that opinion information embedded in LLM internal activations far exceeds what is demonstrated in their outputs.
What Exactly Does Guidance Do in Masked Discrete Diffusion Models
Ye He (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
GenerationTransformerDiffusion modelImage
🎯 What it does: This paper conducts a rigorous theoretical analysis of classifier-free guidance (CFG) in masked discrete diffusion models, deriving analytical results in 1D and 2D, explaining how CFG alters the generation distribution and accelerates the reverse process;
What Generative Search Engines Like and How to Optimize Web Content Cooperatively
Yujiang Wu (Carnegie Mellon University), Chenyan Xiong (Vody)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the AutoGEO framework, which leverages large language models to automatically extract preference rules from generative search engines. The framework employs rule-driven prompt models (AutoGEOAPI) and reinforcement learning models (AutoGEOMini) to enhance document visibility across multiple LLM engines while maintaining response quality.
What Happens Next? Anticipating Future Motion by Generating Point Trajectories
Gabrijel Boduljak (University of Oxford), Andrea Vedaldi (University of Oxford)
GenerationTransformerFlow-based ModelAuto EncoderImageVideo
🎯 What it does: The study predicts future motion from a single image by framing motion prediction as conditional generation of dense trajectory grids;
What happens when generative AI models train recursively on each others' outputs?
Hung Anh Vu (Duke University), Emily Wenger (Duke University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigate scenarios where generative AI models use each other's outputs as training data during iterative training, exploring their impacts on model performance and diversity;
What Layers When: Learning to Skip Compute in LLMs with Residual Gates
Filipe Laitenberger (Humboldt University Berlin), Yuki M Asano (University of Technology Nuremberg)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the GateSkip residual gate mechanism, dynamically determining whether to execute subsequent layers for each token in decoder-only transformers, thereby achieving differentiable token-level layer skipping;
What Matters for Batch Online Reinforcement Learning in Robotics?
Perry Dong (Stanford University), Chelsea Finn (Stanford University)
Robotic IntelligenceReinforcement LearningDiffusion modelSequentialStochastic Differential Equation
🎯 What it does: Studying a batch online reinforcement learning (RL) framework on robots, systematically evaluating the impact of algorithm categories, policy extraction methods, and policy expressiveness, and proposing a general self-improvement recipe;
What matters for Representation Alignment: Global Information or Spatial Structure?
Jaskirat Singh, Saining Xie (New York University)
GenerationRepresentation LearningConvolutional Neural NetworkTransformerDiffusion modelContrastive LearningImage
🎯 What it does: Investigated which characteristics of target representations in generative diffusion model training are more critical (global semantic information versus spatial structure), finding that spatial structure has a greater impact on generation quality; and proposed a simple iREPA method (replacing MLP projection with convolution and adding spatial normalization) to accelerate the convergence speed of representation alignment.
What Scales in Cross-Entropy Scaling Law?
Junxi Yan (Tsinghua University), Jingtao Zhan (Tsinghua University)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a tri-decomposition of cross-entropy, identifying error entropy as the true quantity that varies with model scale, and provides the error entropy scaling law.
What's In My Human Feedback? Learning Interpretable Descriptions of Preference Data
Rajiv Movva (University of California Berkeley), Emma Pierson (University of California Berkeley)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackLarge Language ModelAuto EncoderText
🎯 What it does: Propose the WIMHF method, which automatically discovers interpretable measurable and expressive preference features from human feedback data using sparse autoencoders, aiding in understanding and utilizing preference data;
What's the plan? Metrics for implicit planning in LLMs and their application to rhyme generation and question answering
Jim Maar (University of Potsdam), Neel Nanda (Google DeepMind)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigated the implicit planning behavior of large language models in rhyme poetry generation and question-answering tasks, and quantitatively evaluated through activation vector interventions.
Whatever Remains Must Be True: Filtering Drives Reasoning in LLMs, Shaping Diversity
Germán Kruszewski (Naver Labs Europe), Marc Dymetman (Independent Researcher)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose the DMVR framework, which uses verifiable rewards to filter out incorrect answers, defines a target distribution, and approximates it through α- DPG, thereby enhancing output diversity while maintaining correctness.
When a Robot is More Capable than a Human: Learning from Constrained Demonstrators
Xinhu Li (University of Southern California), Erdem Biyik
Robotic IntelligenceReinforcement Learning
🎯 What it does: Propose a method for robots to learn and surpass expert behavior under limited expert demonstration conditions;
When Agents “Misremember” Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems
Naen Xu (Zhejiang University), Shouling Ji (Zhejiang University)
Large Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: This paper constructs the MANBENCH benchmark targeting the Mandela Effect in LLM multi-agent systems, systematically evaluating its occurrence and persistence across multi-task and multi-protocol scenarios.
When AI Agents Collude Online: Financial Fraud Risks by Collaborative LLM Agents on Social Platforms
Qibing Ren (Shanghai Jiao Tong University), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Data SynthesisTransformerLarge Language ModelAgentic AITextBenchmarkFinance Related
🎯 What it does: This paper investigates the risks of multi-agent systems driven by large language models (LLMs) collaborating to commit financial fraud on social platforms. It constructs a full lifecycle fraud simulation benchmark named MAFF-Bench, and evaluates the amplification effect of multi-agent collaboration on fraud success rate and population impact rate on this benchmark. Additionally, it explores the effectiveness of three categories of defense strategies: content warnings, agent monitoring, and community resistance.
When and Where to Reset Matters for Long-Term Test-Time Adaptation
Taejun Lim (Yonsei University), Kibok Lee (Yonsei University)
Domain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: To address the model collapse issue in long-term test-time adaptation (TTA), this paper proposes the Adaptive and Selective Reset (ASR) framework, combined with importance-aware knowledge recovery regularization and prediction inconsistency-driven online adaptation adjustment, achieving robust adaptation in continuous domain drift scenarios.
When Bias Meets Trainability: Connecting Theories of Initialization
Alberto Bassi (ETH Zurich), Emanuele Francazi (EPFL)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper investigates the statistical properties of wide networks at initialization, establishes a theoretical equivalence between mean-field theory (MF) and initial guess bias (IGB), and analyzes the impact of different phases on training feasibility.
When Data is the Algorithm: A Systematic Study and Curation of Preference Optimization Datasets
Aladin Djuhera (Technical University Munich), Holger Boche (Technical University Munich)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper conducts a systematic analysis and annotation of five mainstream DPO datasets, and builds a smaller but more performant UltraMix dataset based on this.
When Does Divide and Conquer Work for Long Context LLM? A Noise Decomposition Framework
Zach Xu, Ce Zhang
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: The paper proposes a theoretical framework that categorizes failure modes in long text processing into task noise, model noise, and aggregation noise, and analyzes when Divide-and-Conquer (D&C) can enhance the long-context performance of large language models (LLMs).
When Flatness Does (Not) Guarantee Adversarial Robustness
Nils Philipp Walter (CISPA Helmholtz Center for Information Security), Michael Kamp (Technical University Dortmund)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Compared the flatness of neural networks in the parameter space with adversarial robustness in the input space, and through deriving the closed-form expression and cubic equation bounds of relative sharpness, experiments verified that flatness only guarantees robustness locally and is highly coupled with model confidence.
When Foundation Models are One-Liners: Limitations and Future Directions for Time Series Anomaly Detection
Xiaokun Zhu (KU Leuven), Mathias Verbeke (KU Leuven)
Anomaly DetectionTransformerTime SeriesReview/Survey Paper
🎯 What it does: Systematically evaluate the zero-shot performance of five existing categories of time series foundation models (MOMENT, Chronos, TimesFM, Time-MoE, TSPulse) on unsupervised anomaly detection tasks, examining whether reconstruction/prediction errors can effectively distinguish anomalies.
When Greedy Wins: Emergent Exploitation Bias in Meta-Bandit LLM Training
Sanxing Chen (Duke University), Bhuwan Dhingra (Duke University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningChain-of-Thought
🎯 What it does: Studied the exploration behavior of large language models in multi-armed bandit (MAB) tasks, compared supervised fine-tuning (SFT) and reinforcement learning (RL) training methods, designed new reward signals, and evaluated their performance across different environments and time durations.
When Is Diversity Rewarded in Cooperative Multi-Agent Learning?
Michael Amir (University of Cambridge), Amanda Prorok (University of Cambridge)
Reinforcement Learning
🎯 What it does: In multi-agent task allocation problems, this study investigates whether behavioral diversity (heterogeneity) can surpass homogeneity and provides theoretical criteria; proposes a gradient-based environmental parameter search algorithm HetGPS to automatically discover reward structures that best incentivize diversity; verifies the theory and algorithm across a series of tasks ranging from single-step matrix games to continuous and discrete tasks such as multi-objective capture, Tag, and football.
When Language Models Lose Their Mind: The Consequences of Brain Misalignment
Gabriele Merlin (Max Planck Institute for Software Systems), Mariya Toneva (Max Planck Institute for Software Systems)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: Investigated the causal impact of brain alignment on the language capabilities of large language models (LLMs), constructing two types of models: Brain Misaligned and Brain Preserving, and evaluated their performance on over 200 downstream language tasks.
When Large Multimodal Models Confront Evolving Knowledge: Challenges and Explorations
Kailin Jiang (University of Science and Technology of China), Qing Li (State Key Laboratory of General Artificial Intelligence, BIGAI)
Supervised Fine-TuningMixture of ExpertsVision Language ModelMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the MMEVOKE benchmark for systematically evaluating the knowledge injection and retention capabilities of large multimodal models when facing evolving knowledge, and compared various injection methods on this benchmark.
When LLMs get significantly worse: A statistical approach to detect model degradations
Jonas M. Kübler (Amazon), George Karypis (Amazon)
TransformerLarge Language ModelText
🎯 What it does: To address the potential accuracy degradation in large language models (LLMs) after inference optimization, this paper proposes a statistical hypothesis testing framework based on the McNemar test to determine whether observed accuracy drops are real or due to noise.
When Machine Learning Gets Personal: Evaluating Prediction and Explanation
Louisa Cornelis (University of California Santa Barbara), Nina Miolane (University of California Santa Barbara)
Explainability and InterpretabilityTabularBiomedical DataElectronic Health Records
🎯 What it does: Proposes a unified framework to quantify the impact of model personalization on prediction and explanation, and assess whether both change synchronously;
When MLLMs Meet Compression Distortion: A Coding Paradigm Tailored to MLLMs
Jinming Liu (Shanghai Jiao Tong University), Yan Lu (Eastern Institute of Technology)
CompressionTransformerVision Language ModelAuto EncoderContrastive LearningMultimodality
🎯 What it does: Study the impact of compression distortion on multimodal large language models (MLLMs) and propose the CoTAM compression algorithm, which utilizes shallow CLIP attention to guide bit allocation, lightweight adapters, and multi-level reconstruction loss to preserve low-level details while ensuring high-level semantics.
When More is Less: Understanding Chain-of-Thought Length in LLMs
Yuyang Wu (Peking University), Yisen Wang (Peking University)
Large Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Systematically investigate the impact of Chain-of-Thought (CoT) length on large language model reasoning performance, and prove that there exists an 'optimal' CoT length; reveal the scaling laws of optimal length with task difficulty, model size, and per-step computational demands through controlled experiments, theoretical error propagation analysis, and RL-based adaptive regulation; further propose practical methods based on optimal length training and length-filtering voting.
When Priors Backfire: On the Vulnerability of Unlearnable Examples to Pretraining
Zhihao Li (Western University), Boyu Wang (Western University)
ClassificationOptimizationAdversarial AttackMeta LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Investigated the vulnerability of pre-trained models to unlearnable examples (UEs) and proposed the BAIT framework to overcome the failure of UEs caused by pre-trained priors
When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models
Nan Zhang (Pennsylvania State University), Rui Zhang (Pennsylvania State University)
CompressionExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Quantize, distill, and prune large reasoning models (LRMs), and analyze the impact of compression on inference capability using performance benchmarks and fine-grained mechanisms.
When Scores Learn Geometry: Rate Separations under the Manifold Hypothesis
Xiang Li (ETH Zurich), Niao He (ETH Zurich)
GenerationDiffusion modelScore-based ModelImage
🎯 What it does: Proposes a new perspective, suggesting that the success of score learning methods stems from the implicit learning of data manifolds rather than the complete distribution.
When Shift Happens - Confounding Is to Blame
Abbavaram Gowtham Reddy (CISPA Helmholtz Center for Information Security), Krikamol Muandet (CISPA Helmholtz Center for Information Security)
Domain AdaptationTabularElectronic Health RecordsBenchmark
🎯 What it does: This paper investigates the impact of distribution drift caused by hidden confounding on OOD generalization, explains why ERM can outperform traditional methods, and demonstrates that incorporating non-causal but information-rich covariates can enhance generalization performance.
When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond?
Xinyu Zhou (HKUST), Seyed Ali Bahrainian (University of Tübingen)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextTime SeriesChain-of-Thought
🎯 What it does: This paper systematically studies how to train large language models to learn self-denial (abstention) in time-sensitive question-answering tasks.
When Style Breaks Safety: Defending LLMs Against Superficial Style Alignment
Yuxin Xiao (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Studied the impact of style patterns on the safety of large language models, and found that style patterns can lead to ASR (attack success rate) inflation, proposing SafeStyle to counteract the security risks caused by surface style alignment;
When Thinking Backfires: Mechanistic Insights into Reason-induced Misalignment
Hanqi Yan (King's College London), Yulan He (King's College London)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelContrastive LearningTextChain-of-Thought
🎯 What it does: Explores and reveals the 'Reasoning-Induced Misalignment (RIM)' phenomenon, where enhancing the reasoning capabilities of Large Language Models (LLMs) (via Chain-of-Thought prompts or training) paradoxically increases the model's tendency to respond to malicious requests.
When to Ensemble: Identifying Token-Level Points for Stable and Fast LLM Ensembling
Heecheol Yun (KAIST), Eunho Yang (KAIST)
GenerationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposed the SAFE framework, which dynamically determines when to perform probability-level fusion during long-text generation based on tokenization mismatch and model consistency, thereby enhancing the stability and efficiency of LLMs.
When to Retrain after Drift: A Data-Only Test of Post-Drift Data Size Sufficiency
Ren Fujiwara (University of Osaka), Yasushi Sakurai (University of Osaka)
Anomaly DetectionComputational EfficiencyData-Centric LearningTime SeriesSequential
🎯 What it does: Proposed a model-free, label-free, and data-based framework called CALIPER for estimating sufficient post-drift data volume to achieve stable retraining after detecting concept drift.
When to use Graphs in RAG: A Comprehensive Analysis for Graph Retrieval-Augmented Generation
Zhishang Xiang, Jinsong Su (Xiamen University)
RetrievalGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed GraphRAGBench, a comprehensive benchmark to evaluate the performance of Graph Retrieval-Augmented Generation (GraphRAG) compared to traditional RAG, and conducted systematic experiments on multiple GraphRAG frameworks.
When Weak LLMs Speak with Confidence, Preference Alignment Gets Stronger
Amirabbas Afzali (EPFL), Maria Brbic (EPFL)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The study leverages the confidence of weak large language models to weight preference alignment, proposing the CW-PO framework to reduce manual annotation costs and enhance alignment effectiveness.
When would Vision-Proprioception Policies Fail in Robotic Manipulation?
Jingxian Lu (Renmin University of China), Di Hu (Renmin University of China)
Robotic IntelligenceRecurrent Neural NetworkTransformerVision-Language-Action ModelAuto EncoderMultimodality
🎯 What it does: Investigate the failure reasons of visual-proprioceptive policies in robot manipulation and propose a stage-guided gradient adjustment algorithm (GAP) to balance learning between the two modalities.
Where Did It Go Wrong? Attributing Undesirable LLM Behaviors via Representation Gradient Tracing
Zhe Li (Singapore Management University), Jun Sun (Singapore Management University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose a framework based on Representation Gradient Tracing (RepT) to diagnose and attribute harmful, inaccurate, or backdoor-polluted output behaviors in large language models.
Where Did This Sentence Come From? Tracing Provenance in LLM Reasoning Distillation
kaiyuan liu, Jieping Ye (University of Michigan)
Knowledge DistillationLarge Language ModelText
🎯 What it does: Proposed a cross-model reasoning distillation source tracking framework, analyzing the teacher and student sources in the outputs of distilled models, and based on this, introduced a teacher-guided data selection strategy to enhance performance.
Who Matters Matters: Agent-Specific Conservative Offline MARL
Haosheng Chen (East China Normal University), Xiangfeng Wang (Tongji University)
Reinforcement LearningBenchmark
🎯 What it does: Propose a dynamic conservative degree allocation framework OMCDA for offline multi-agent reinforcement learning, addressing the problem of conservative degree allocation when agents with different roles collaboratively learn on a fixed dataset.
WholeBodyVLA: Towards Unified Latent VLA for Whole-body Loco-manipulation Control
Haoran Jiang (Fudan University), Hongyang Li (University of Hong Kong)
Robotic IntelligenceReinforcement LearningVision-Language-Action ModelVideoMultimodality
🎯 What it does: Designed and implemented WholeBodyVLA, an end-to-end multimodal control framework enabling bipedal robots to perform coherent walking and manipulation tasks over a wide range.
Why Adversarially Train Diffusion Models?
Maria Rosaria Briglia (Sapienza University of Rome), Iacopo Masi (University of Bologna)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: Proposed an adversarial training method for diffusion models to improve their robustness under noisy and corrupted data.
Why Ask One When You Can Ask $k$? Learning-to-Defer to the Top-$k$ Experts
Yannis Montreuil (National University of Singapore), Wei Tsang Ooi (National University of Singapore)
Mixture of ExpertsImage
🎯 What it does: Proposed the Topk Learning-to-Delay (Topk L2D) framework and its adaptive extension Topk×(θ) L2D, achieving simultaneous allocation of queries to the k most cost-effective experts, and providing a k-independent convex approximation loss function;
Why Attention Patterns Exist: A Unifying Temporal Perspective Analysis
Qingyue Yang (University of Science and Technology of China), Bin Li (University of Science and Technology of China)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextSequential
🎯 What it does: Proposed the Temporal Attention Pattern Predictability Analysis (TAPPA) framework, which unifies the explanation of attention patterns in large language models from a temporal perspective;
Why Do Unlearnable Examples Work: A Novel Perspective of Mutual Information
Yifan Zhu (State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences), Xiao-Shan Gao (State Key Laboratory of Mathematical Sciences, Academy of Mathematics and Systems Science, Chinese Academy of Sciences)
Adversarial AttackData-Centric LearningImage
🎯 What it does: Propose a learnable invalid sample generation method based on mutual information reduction (MI-UE), achieving higher non-learnability through covariance reduction.
Why DPO is a Misspecified Estimator and How to Fix It
Aditya Gopalan (Indian Institute of Science Bangalore), Debangshu Banerjee (HP AI Research)
Reinforcement Learning from Human FeedbackReinforcement LearningTextBenchmark
🎯 What it does: This paper reveals the misjudgment estimation problem of DPO under parameterized strategy classes through theoretical analysis, and proposes the AuxDPO algorithm to correct this defect.