ICLR 2026 Papers — Page 52
International Conference on Learning Representations · 5356 papers
Variance-Dependent Regret Lower Bounds for Contextual Bandits
Jiafan He (University of California, Los Angeles), Quanquan Gu (University of California, Los Angeles)
Optimization
🎯 What it does: This paper constructs hard instances for linear contextual bandits under arbitrary (pre-determined or adaptive) noise variance sequences, establishing variance-related lower bounds that match known upper bounds, and proving that strong adversaries cannot guarantee lower bounds;
Variation in Verification: Understanding Verification Dynamics in Large Language Models
Yefan Zhou (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Investigates the verifiability of large language models in environments without reference answers, systematically explores the impact of problem difficulty, generator and verifier capabilities on verification effectiveness, and evaluates their application in test-time scaling (TTS)
Variation-aware Flexible 3D Gaussian Editing
Hao Qin (Zhejiang University), Qiang Zhu (Zhejiang University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelGaussian SplattingTextPoint CloudMesh
🎯 What it does: Propose a variational prediction-based 3D Gaussian splatting (3DGS) editing framework, VF-Editor, which enables spatial and appearance modifications of 3D Gaussian voxels in a single forward pass.
Variational Autoencoding Discrete Diffusion with Enhanced Dimensional Correlations Modeling
Tianyu Xie (Peking University), Cheng Zhang (Peking University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: Propose Variational Autoencoding Discrete Diffusion (VADD), integrating latent variable structures into the reverse process of Masked Diffusion Models (MDMs) to enable the model to implicitly capture correlations between dimensions, thereby improving the quality of samples generated with fewer steps.
Variational Deep Learning via Implicit Regularization
Jonathan Wenger (Columbia University), John Patrick Cunningham
ClassificationExplainability and InterpretabilityComputational EfficiencyImage
🎯 What it does: This paper proposes Implicit Bayesian Variational Inference (IBVI), which achieves variational deep learning by leveraging the implicit regularization of gradient descent, without explicitly using the KL regularization term.
Variational Inference for Cyclic Learning
Zhuojun Zou (Chinese Academy of Sciences), Jie Hao (Guangdong Institute of Artificial Intelligence and Advanced Computing)
Image TranslationObject TrackingTransformerAuto EncoderGenerative Adversarial NetworkImageVideo
🎯 What it does: Proposes a unified variational inference framework that addresses probabilistic modeling and training for symmetric and self-cyclic tasks;
Variational Reasoning for Language Models
Xiangxin Zhou (Sea AI Lab), Tianyu Pang
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposed a variational inference framework that treats the thinking trajectory of language models as latent variables, optimizing reasoning capabilities through multi-trajectory ELBO and forward KL objectives;
VaseVQA-3D: Benchmarking 3D VLMs on Ancient Greek Pottery
Nonghai Zhang (Peking University), Hao Tang (Peking University)
Supervised Fine-TuningReinforcement LearningVision Language ModelImageTextMeshBenchmark
🎯 What it does: Constructed the first 3D Visual Question Answering dataset for ancient Greek pottery, VaseVQA-3D, and trained a specialized Vision-Language Model (VLM) called VaseVLM based on this dataset.
vAttention: Verified Sparse Attention via Sampling
Aditya Desai (University of California Berkeley), Ion Stoica (University of California Berkeley)
Computational EfficiencyTransformerText
🎯 What it does: Propose vAttention, which provides verifiable sparse attention during the inference phase and allows users to set error thresholds via (ε,δ);
vCache: Verified Semantic Prompt Caching
Luis Gaspar Schroeder, Joseph E. Gonzalez
RetrievalLarge Language ModelText
🎯 What it does: Designed and implemented an online learning semantic cache system, vCache, which learns threshold values for each cache vector and guarantees error rates as specified by users, addressing the insufficient reliability and hit rate issues caused by traditional fixed thresholds.
VCWorld: A Biological World Model for Virtual Cell Simulation
Zhijian Wei (Shanghai Jiao Tong University), Shuangjia Zheng (Shanghai Jiao Tong University)
Drug DiscoveryGraph Neural NetworkLarge Language ModelWorld ModelBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose VCWorld, a cell-level white-box simulator based on a biological knowledge graph and LLM reasoning, for predicting drug-induced gene expression changes.
VEAttack: Downstream-agnostic Vision Encoder Attack against Large Vision Language Models
Hefei Mei (City University of Hong Kong), Chang Xu (University of Sydney)
Adversarial AttackTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a gray-box attack method called VEAttack that targets only the visual encoder, achieving non-targeted, low-perturbation attacks on large vision-language models by minimizing the cosine similarity of visual features.
VenusX: Unlocking Fine-Grained Functional Understanding of Proteins
Yang Tan (Shanghai Jiao Tong University), Bingxin Zhou (Shanghai Jiao Tong University)
Representation LearningGraph Neural NetworkTransformerLarge Language ModelBiomedical DataBenchmark
🎯 What it does: Proposed and released the VENUSX benchmark to evaluate the fine-grained functional understanding capabilities of protein representation learning at the residue and fragment levels.
VER: Vision Expert Transformer for Robot Learning via Foundation Distillation and Dynamic Routing
Yixiao Wang (University Of California Berkeley), Masayoshi Tomizuka
Knowledge DistillationRobotic IntelligenceTransformerReinforcement LearningMixture of ExpertsVision-Language-Action ModelDiffusion modelFlow-based ModelImage
🎯 What it does: Developed a Vision Expert Transformer (VER) by distilling multiple visual foundation models into an expert library, and using a lightweight router to dynamically select the most relevant experts in downstream robotic tasks;
VeriCoT: Neuro-symbolic Chain-of-Thought Validation via Logical Consistency Checks
Yu Feng (University of Pennsylvania), Huzefa Rangwala (Amazon Web Services)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Proposes VERICOT, a neuro-symbolic framework that automatically formalizes LLM Chain-of-Thought (CoT) steps into first-order logic and verifies their logical consistency through SMT solvers.
VeriEquivBench: An Equivalence Score for Ground-Truth-Free Evaluation of Formally Verifiable Code
Lingfei Zeng (Huazhong University of Science and Technology), Jie Fu (Shanghai Artificial Intelligence Laboratory)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Constructed the VeriEquivBench benchmark, containing 2,389 complex algorithm problems along with their natural language descriptions, Python and Dafny implementations, unit tests, and formal specifications, and proposed an equivalence score evaluation method without relying on existing benchmarks.
Verification and Co-Alignment via Heterogeneous Consistency for Preference-Aligned LLM Annotations
Cheng Chen (Agency for Science, Technology and Research), Ivor Tsang (Agency for Science, Technology and Research)
ClassificationRepresentation LearningData-Centric LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed Heterogeneous Consistency Co-Alignment (HCC), a training-agnostic framework that verifies and co-aligns unlabeled text through mutual cross-validation between knowledge-rich LLMs and task-specific lightweight models.
Verification of the Implicit World Model in a Generative Model via Adversarial Sequences
András Balogh (University of Szeged), Márk Jelasity (University of Szeged)
GenerationAdversarial AttackTransformerLarge Language ModelWorld ModelSequential
🎯 What it does: Investigated the reliability of implicit world models in generative sequence models, proposing adversarial generation of legal game sequences to uncover model errors;
Verifier-Constrained Flow Expansion for Discovery Beyond the Data
Riccardo De Santi (ETH Zurich), Andreas Krause (ETH Zurich)
GenerationDrug DiscoveryFlow-based ModelBiomedical DataOrdinary Differential Equation
🎯 What it does: Propose a method that utilizes a validator to perform global or local entropy maximization expansion on pre-trained flow models, enabling the generated samples to cover a broader effective design space while maintaining validity.
Verifier-free Test-Time Sampling for Vision-Language-Action Models
Suhyeok Jang (KAIST), Jinwoo Shin (KAIST)
Robotic IntelligenceVision-Language-Action ModelMultimodality
🎯 What it does: Proposed MG-Select, a framework for in-model distribution-based validation-free scaling during testing, to enhance the performance of vision-language-action models in high-precision robotic control tasks.
VERIFY: A Novel Multi-Domain Dataset Grounding LTL in Contextual Natural Language via Provable Intermediate Logic
Paapa Kwesi Quansah (Baylor University), Ernest Bonnah (Baylor University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper introduces the VERIFY dataset, containing over 200,000 triplets of linear temporal logic (LTL)–intermediate temporal logic (ITL)–natural language (NL), and constructs a complete pipeline for generating ITL from LTL and subsequently generating context-aware NL. The dataset spans 13 application domains, employing formal verification, provably correct ITL syntax, and a multi-stage quality assurance process combining large language model (LLM) generation and LLM evaluation.
VerifyBench: Benchmarking Reference-based Reward Systems for Large Language Models
Yuchen Yan (Zhejiang University), Yueting Zhuang (Zhejiang University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Constructed and evaluated a reward system benchmark called VerifyBench and its more challenging variant VerifyBench-Hard, specifically designed for reference answer verification, to measure the accuracy of large language models in reasoning tasks for answer validation.
Verifying Chain-of-Thought Reasoning via Its Computational Graph
Zheng Zhao (FAIR at Meta), Nicola Cancedda (FAIR at Meta)
Explainability and InterpretabilityAuto EncoderTextChain-of-Thought
🎯 What it does: Propose a white-box method called Circuit-based Reasoning Verification (CRV), which replaces the MLP of LLM with a sparse interpretable Transcoder, constructs attribution graphs for reasoning steps, extracts their structural features, and uses them to determine the correctness of reasoning steps.
VERINA: Benchmarking Verifiable Code Generation
Zhe Ye (University of California, Berkeley), Dawn Song (University of California, Berkeley)
GenerationTextBenchmark
🎯 What it does: Proposed VERINA, a high-quality Lean benchmark covering code, specifications, and proof generation for evaluating verifiable code generation;
VeriRole: Verifiable Role-Awareness through Hint-Guided Reinforcement Learning
Zongsheng Wang (Renmin University of China), Baoxun Wang (Tencent)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the VeriRole framework, which extracts verifiable facts using a 'hint' mechanism and enhances the role awareness and consistency of role-play dialogue models through Verifiable Role Awareness Reward (VRAR) combined with reinforcement learning.
Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning
Hao Tan (University of Chinese Academy of Sciences), Zhen Lei (University of Chinese Academy of Sciences)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Constructed the HydraFake dataset containing diverse real and forged images, and developed the VERITAS deepfake detector based on a multi-modal large language model (MLLM).
VeriTrail: Closed-Domain Hallucination Detection with Traceability
Dasha Metropolitansky (Microsoft Research), Jonathan Larson (Microsoft Research)
Anomaly DetectionLarge Language ModelTextBenchmark
🎯 What it does: This study addresses the problem of closed-domain hallucinations in multi-step (MGS) language model generation, proposing the VeriTrail method to achieve hallucination detection and traceability.
VFScale: Intrinsic Reasoning through Verifier-Free Test-time Scalable Diffusion Model
Tao Zhang (Zhejiang University), Tailin Wu (Westlake University)
GenerationReinforcement LearningDiffusion modelContrastive Learning
🎯 What it does: This paper proposes the Verifier-free Test-time Scalable Diffusion Model (VFScale), which achieves scalable inference capabilities during testing by increasing the number of samples, thereby enabling the completion of more complex reasoning tasks.
VGR: Visual Grounded Reasoning
Jiacong Wang (School of Artificial Intelligence, University of Chinese Academy of Sciences), Jun Xiao (School of Artificial Intelligence, University of Chinese Academy of Sciences)
RetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented the VGR (Visual Grounded Reasoning) model, enabling multimodal large language models to actively retrieve and replay visual memories during reasoning, thereby significantly enhancing visual reasoning performance.
VibeVoice: Expressive Podcast Generation with Next-Token Diffusion
Zhiliang Peng (Microsoft Research), Furu Wei (Microsoft Research)
GenerationTransformerLarge Language ModelDiffusion modelAuto EncoderTextBenchmarkAudio
🎯 What it does: Propose the VIBEVOICE model to achieve zero-shot, expressive, long-form, multi-speaker podcast speech synthesis
Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction–Reasoning Synergy
Haijier Chen (Wuhan University), Jingrong Wang (Shenzhen University)
Depth EstimationKnowledge DistillationTransformerLarge Language ModelVision Language ModelVideoTextPoint CloudBenchmark
🎯 What it does: Reconstruct 3D scene geometry directly from monocular videos and fuse geometric features with semantic features through Cross-Task Adapter to build a multimodal large language model capable of 3D vision-language reasoning;
Vid2World: Crafting Video Diffusion Models to Interactive World Models
Siqiao Huang (Tsinghua University), Mingsheng Long (Tsinghua University)
GenerationDiffusion modelWorld ModelVideo
🎯 What it does: Convert large-scale pre-trained video diffusion models into interactive world models to achieve autoregressive, action-conditioned generation.
VidBridge-R1: Bridging QA and Captioning for RL-based Video Understanding Models with Intermediate Proxy Tasks
Xinlong Chen (Chinese Academy of Sciences), Liang Wang (Chinese Academy of Sciences)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: Propose the VidBridge-R1 framework, which addresses the conflict between video question answering and caption generation during reinforcement learning training by introducing two intermediate proxy tasks, DarkEventInfer and MixVidQA, to build a unified video understanding model.
Video Scene Segmentation with Genre and Duration Signals
Jungu Cho (CJ Corporation), Hae-Gon Jeon (CJ Corporation)
SegmentationTransformerContrastive LearningVideoTextBenchmark
🎯 What it does: Propose to use movie genres and shot duration signals for video scene segmentation, and design corresponding training and inference strategies.
Video Unlearning via Low-Rank Refusal Vector
Simone Facchiano (Sapienza University Of Rome), Fabio Galasso (Sapienza University Of Rome)
Safty and PrivacyDiffusion modelContrastive LearningVideoMultimodalityBenchmark
🎯 What it does: Proposed a training-agnostic closed-form weight update framework to permanently eliminate specified unsafe concepts in video diffusion models.
Video-As-Prompt: Unified Semantic Control for Video Generation
Yuxuan Bian (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
GenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderVideoMultimodality
🎯 What it does: Propose Video-As-Prompt (VAP), treating reference videos as prompts, achieving unified semantic video generation through pluggable context control using Mixture-of-Transformers.
Video-GPT via Next Clip Diffusion
Shaobin Zhuang (Shanghai Jiao Tong University), Yali Wang (Chinese Academy Of Sciences)
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderVideo
🎯 What it does: Propose Video-GPT, a video generation and prediction framework that treats video clips as visual words and performs self-supervised pre-training through next-clip diffusion.
Video-KTR: Reinforcing Video Reasoning via Key Token Attribution
Ziyue Wang (ByteDance), Xudong Jiang (Nanyang Technological University)
Explainability and InterpretabilityReinforcement LearningVision Language ModelVideo
🎯 What it does: Propose the Video-KTR framework, which identifies critical tokens in video reasoning by conducting comparative analysis of three factors: visual, temporal, and uncertainty, and updates only these tokens in reinforcement learning;
Video-LevelGauge: Investigating Contextual Positional Bias in Video Language Models.
Hou Xia (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
Large Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes and implements the Video-LevelGauge benchmark for systematically evaluating biases of large video-language models (LVLMs) at different context positions;
Video-STAR: Reinforcing Open-Vocabulary Action Recognition with Tools
Zhenlong Yuan (Alibaba Group), Shuo Li (Case Western Reserve University)
RecognitionConvolutional Neural NetworkTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the Video-STAR framework, achieving open-vocabulary action recognition through contextual sub-action decomposition and tool-enhanced reinforcement learning.
VideoAgentTrek: Computer-Use Pretraining from Unlabeled Videos
Dunjie Lu (The University of Hong Kong), Tao Yu (The University of Hong Kong)
Representation LearningData-Centric LearningRobotic IntelligenceConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: Construct a scalable pre-training pipeline for computer usage agents by converting publicly available screen recording videos into structured training data through an automated inverse dynamics module.
VideoAnchor: Reinforcing Subspace-Structured Visual Cues for Coherent Visual-Spatial Reasoning
Zhaozhi Wang (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
Representation LearningTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper investigates the shortcomings of multimodal large language models (MLLMs) in visual-spatial reasoning, proposing a no-training, plug-and-play VideoAnchor module. It leverages the self-expressive properties of sparse subspace clustering (SSC) to enhance attention during inference, thereby improving the consistency of visual cues and the accuracy of spatial reasoning.
VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling
Xinhao Li (Nanjing University), Limin Wang (Nanjing University)
CompressionTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: Proposed a hierarchical video compression method called HiCo, combined with multi-stage short-long learning, the LongVid dataset, and the multi-hop pinhole video sandpile evaluation, to build the VideoChat-Flash video multimodal large language model, which can efficiently process long videos and achieve state-of-the-art performance on multiple benchmarks.
VideoJudge: Bootstrapping Enables Scalable Supervision of MLLM-as-a-Judge for Video Understanding
Abdul Waheed (Carnegie Mellon University), Bhiksha Raj (Carnegie Mellon University)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Propose VideoJudge — a self-supervised framework based on a generator-evaluator loop, which trains multimodal large language models (MLLMs) as judges for video understanding tasks, generating point scores and pairwise evaluations, and automatically generating instance-specific evaluation rubrics during testing.
VideoMathQA: Benchmarking Mathematical Reasoning via Multimodal Understanding in Video
Hanoona Abdul Rasheed, Fahad Shahbaz Khan (MBZUAI)
TransformerVision Language ModelVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: Introduce the VideoMathQA benchmark to evaluate models' ability to perform cross-temporal mathematical reasoning in multimodal educational videos.
VideoMind: A Chain-of-LoRA Agent for Temporal-Grounded Video Reasoning
Ye Liu (Hong Kong Polytechnic University), Mike Zheng Shou (National University of Singapore)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAgentic AIVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose VideoMind, a video-language agent capable of achieving temporally interpretable video reasoning through four roles: planning, localization, verification, and answering.
VideoNSA: Native Sparse Attention Scales Video Understanding
Enxin Song (University of California, San Diego), Zhuowen Tu (University of California, San Diego)
RecognitionComputational EfficiencyTransformerVision Language ModelVideo
🎯 What it does: This work proposes VideoNSA, a framework that migrates Native Sparse Attention (NSA) to video-language models, significantly reducing attention computation while preserving video frame information;
VideoPhy-2: A Challenging Action-Centric Physical Commonsense Evaluation in Video Generation
Hritik Bansal (University of California Los Angeles), Kai-Wei Chang (University of California Los Angeles)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextBenchmarkPhysics Related
🎯 What it does: Proposed the VIDEOPHY-2 dataset and an automatic evaluator to assess the physical common sense compliance of text-to-video generation models in multi-event action scenarios
VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning?
Yuanxin Liu (Peking University), Xu Sun (Peking University)
Vision Language ModelVideoTextBenchmarkChain-of-Thought
🎯 What it does: Proposed VIDEOREASONBENCH, a benchmark for evaluating the complex video reasoning capabilities of multimodal large language models (MLLMs) in visually centered tasks.
VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning
Yang Ding (Tsinghua University), Yujiu Yang (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVideoMultimodality
🎯 What it does: Propose the VideoZoomer framework, enabling multimodal large language models to dynamically invoke time-domain scaling tools during long video reasoning processes, obtaining critical details through multi-round interactive communication.
VidGuard-R1: AI-Generated Video Detection and Explanation via Reasoning MLLMs and RL
Kyoungjun Park (University of Texas at Austin), Lili Qiu (University of Texas at Austin)
ClassificationExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a framework named VidGuard-R1 for AI-generated video detection and interpretable reasoning based on a multimodal large language model, achieving video authenticity judgment and explanation through chain-of-thought reasoning and reinforcement learning.
villa-X: Enhancing Latent Action Modeling in Vision-Language-Action Models
Xiaoyu Chen (Tsinghua University), Jiang Bian (Tsinghua University)
Robotic IntelligenceMixture of ExpertsVision-Language-Action ModelDiffusion modelFlow-based ModelVideoMultimodality
🎯 What it does: Proposes Villa-X, a Vision-Language-Latent-Action (ViLLA) framework that integrates physics-aware latent action learning with joint diffusion strategies to achieve cross-modal, cross-robot platform zero-shot pre-training and control.
ViMo: A Generative Visual GUI World Model for App Agents
Dezhao Luo (Queen Mary University of London), Kun Shao (Huawei Noah's Ark Lab)
GenerationTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelWorld ModelImageText
🎯 What it does: Proposed ViMo, a generative visual GUI world model capable of predicting future GUIs of apps in the visual modality, and separated text generation from graphic generation through Symbolic Text Representation (STR), enhancing text readability and visual consistency.
VINCIE: Unlocking In-context Image Editing from Video
Leigang Qu (National University of Singapore), Lu Jiang (ByteDance Seed)
Image HarmonizationGenerationTransformerVision Language ModelDiffusion modelFlow-based ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Developed a multi-round contextual image editing model called VINCIE that can learn solely from raw videos, supporting continuous editing without requiring manually paired editing data.
Vintix II: Decision Pre-Trained Transformer is a Scalable In-Context Reinforcement Learner
Andrei Polubarov (Applied AI Institute), Vladislav Kurenkov (Innopolis University)
Meta LearningTransformerRectified FlowOrdinary Differential Equation
🎯 What it does: This paper extends Decision Pretrained Transformer (DPT) to multi-domain continuous control tasks and achieves scalable context reinforcement learning through flow matching.
ViPER: Empowering the Self-Evolution of Visual Perception Abilities in Vision-Language Models
Juntian Zhang (Renmin University of China), Rui Yan (Wuhan University)
GenerationData SynthesisReinforcement LearningVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: Proposed the ViPER framework, which enables self-evolution of visual perception capabilities in vision-language models (VLM) through self-generated data and two-stage reinforcement learning.
ViPO: Visual Preference Optimization at Scale
Ming Li (University of Central Florida), Chen Chen (ByteDance Seed)
GenerationData SynthesisOptimizationData-Centric LearningVision Language ModelDiffusion modelImageVideo
🎯 What it does: Propose the Poly-DPO method in visual generation and construct the ViPO large-scale preference dataset to address the noise and distribution imbalance issues in existing preference data.
ViPRA: Video Prediction for Robot Actions
Sandeep Routray (Carnegie Mellon University), Deepak Pathak (Carnegie Mellon University)
Robotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelFlow-based ModelAuto EncoderOptical FlowVideoTextMultimodalityBenchmark
🎯 What it does: Propose the ViPRA framework, which learns discrete latent actions centered on motion from unlabeled videos, jointly predicts future visual states and latent actions in a video-language model, and subsequently maps latent actions to continuous high-frequency robot control commands using a flow matching decoder, achieving data-efficient general-purpose robot policies.
Virne: A Comprehensive Benchmark for RL-based Network Resource Allocation in NFV
Tianfu Wang, Hui Xiong
OptimizationGraph Neural NetworkReinforcement LearningGraphTabularBenchmark
🎯 What it does: Proposed and implemented Virne, a unified and scalable NFV-RA benchmark framework, supporting customizable simulations for multiple scenarios, integrating over 30 algorithms (including deep reinforcement learning methods), and providing a complete offline/online evaluation process.
Virtual Community: An Open World for Humans, Robots, and Society
Qinhong Zhou (University of Massachusetts Amherst), Chuang Gan (University of Massachusetts Amherst)
GenerationData SynthesisRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelDiffusion modelWorld ModelImageTextMultimodalityBenchmark
🎯 What it does: Built an open-source Virtual Community platform capable of automatically generating large-scale open-world scenes and multi-agent communities aligned with the real world, achieving multimodal interaction between humans and robots under a unified Genesis physics engine, followed by proposing two multi-agent challenges on community planning and robot collaboration.
VIRTUE: Visual-Interactive Text-Image Universal Embedder
Wei-Yao Wang (Sony Group Corporation), Yuki Mitsufuji (Sony Group Corporation)
SegmentationRetrievalRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the VIRTUE framework, integrating the pre-trained segmentation model SAM2 with a vision-language model, enabling users to encode images at the entity level using visual prompts such as points, boxes, or masks while preserving global context, and constructing a large-scale SCaR retrieval benchmark to evaluate visual interaction capabilities.
VisCoder2: Building Multi-Language Visualization Coding Agents
Yuansheng Ni (University of Waterloo), Wenhu Chen (University of Waterloo)
AI Code AssistantLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: Researchers propose VisCoder2, a multilingual visualization coding agent, and construct a 679K executable visualization code dataset VisCode-Multi-679K and an 888-item multilingual benchmark VisPlotBench, supporting multi-round debugging.
VisCodex: Unified Multimodal Code Generation via Merging Vision and Coding Models
Lingjie Jiang (Microsoft Research), Furu Wei (Microsoft Research)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose the VisCodex framework, which merges visual models and code models through task vector model merging to build a unified multimodal code generation model.
VisioMath: Benchmarking Figure-based Mathematical Reasoning in LMMs
Can Li (Beijing Normal University), Hua Huang (Beijing Normal University)
Supervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed the VisioMath benchmark, containing 1,800 K-12 math problems, all with high similarity graphical options, testing multi-image reasoning.
Vision Language Models are Biased
An Vo (KAIST), Daeyoung Kim (KAIST)
RecognitionData SynthesisExplainability and InterpretabilityPrompt EngineeringVision Language ModelImageBenchmark
🎯 What it does: Proposed the VLMBias benchmark, systematically evaluating VLMs' visual bias and counting accuracy using natural neutral questions and adversarial counterfactual images across seven domains (animals, logos, flags, chess pieces, chessboards, optical illusions, grid patterns).
Vision-Language-Action Instruction Tuning: From Understanding to Manipulation
Shuai Yang (Zhejiang University), Jiangmiao Pang (Shanghai Artificial Intelligence Laboratory)
Robotic IntelligenceMixture of ExpertsVision Language ModelVision-Language-Action ModelFlow-based ModelMultimodality
🎯 What it does: Propose InstructVLA, a unified vision-language-action model that can achieve precise robot action generation while maintaining the reasoning capabilities of large-scale vision-language models;
Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models
Wenxuan Huang (East China Normal University), Shaohui Lin (East China Normal University)
TransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed a 200K multi-modal chain-of-thought (CoT) dataset without human annotation, and used this dataset to perform cold start initialization for large multi-modal language models (MLLMs); subsequently, adopted reinforcement learning (GRPO) combined with progressive thinking suppression training (PTST), significantly enhancing the model's complex reasoning capabilities through RL training using only 10K multi-modal math samples;
Vision-SR1: Self-Rewarding Vision-Language Model via Reasoning Decomposition and Multi-Reward Policy Optimization
Zongxia Li (Tencent AI Seattle Lab), Dong Yu (University of Maryland)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose Vision-SR1, a self-rewarding reinforcement learning framework that decomposes visual reasoning into visual description and language reasoning, and significantly enhances visual reasoning capabilities by leveraging model self-evaluation to obtain visual rewards.
Vision-Zero: Scalable VLM Self-Evolution via Multi-Agent Self-Play
Qinsi Wang (Duke University), Wentian Zhao (Adobe Inc)
Representation LearningData-Centric LearningReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Propose the Vision-Zero framework, leveraging a multi-agent 'Who's the Spy?' visual self-play game to achieve zero-shot self-evolution for Vision-Language Models (VLM).
VisionLaw: Inferring Interpretable Intrinsic Dynamics from Visual Observations via Bilevel Optimization
Jiajing Lin (Xiamen University), Min Jiang (Xiamen University)
OptimizationExplainability and InterpretabilityLarge Language ModelGaussian SplattingVideoPhysics Related
🎯 What it does: Inferring interpretable essential dynamics directly from multi-view videos through a two-layer optimization framework and LLM-driven evolution;
VisionReasoner: Unified Reasoning-Integrated Visual Perception via Reinforcement Learning
Yuqi Liu (CUHK), Jiaya Jia (SmartMore)
Object DetectionSegmentationReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Propose VisionReasoner, a unified vision-language framework capable of performing multiple visual perception tasks such as detection, segmentation, and counting through a structured reasoning process.
VisionTrim: Unified Vision Token Compression for Training-Free MLLM Acceleration
Hanxun Yu (Zhejiang University), Jianke Zhu (Zhejiang University)
CompressionComputational EfficiencyTransformerVision Language ModelImageVideoBenchmark
🎯 What it does: Propose VisionTrim, a unified training-agnostic visual token compression framework for accelerating inference in multimodal large language models (MLLM);
VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations
Yupeng Xie (Hong Kong University of Science and Technology), Yuyu Luo (Hong Kong University of Science and Technology)
Large Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed the VISJUDGE-BENCH benchmark to evaluate the judgment capabilities of multimodal large language models (MLLM) in visual quality (credibility, expressiveness, aesthetics), and fine-tuned the VISJUDGE model based on this benchmark.
Visual Autoregressive Modeling for Instruction-Guided Image Editing
Qingyang Mao (HiDream.ai Inc.), Tao Mei (University of Science and Technology of China)
Image TranslationGenerationTransformerMultimodality
🎯 What it does: Proposed VAREdit, an instruction-guided image editing method based on a visual autoregressive (VAR) framework, employing a multi-scale 'next scale' prediction strategy.
Visual Compositional Tuning
Xindi Wu (Princeton University), Olga Russakovsky (Princeton University)
Data SynthesisLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed COMPACT, a data generation scheme that enhances the sample complexity of visual instruction tuning by synthesizing multiple atomic visual capabilities, significantly improving data utilization efficiency.
Visual Jigsaw Post-Training Improves MLLMs
Penghao Wu (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
RecognitionTransformerReinforcement LearningVision Language ModelImageVideoPoint Cloud
🎯 What it does: Proposed a self-supervised pre-training framework called Visual Jigsaw, which enhances the visual perception and understanding capabilities of multi-modal large language models (MLLMs) by enabling them to restore the original order of scrambled image, video, or 3D data fragments.
Visual Multi-Agent System: Mitigating Hallucination Snowballing via Visual Flow
Xinlei Yu (National University of Singapore), Shuicheng YAN
TransformerVision Language ModelImageVideoMultimodalityBenchmark
🎯 What it does: Investigate the 'visual hallucination snowball effect' in multi-agent systems and propose a lightweight, model-agnostic ViF method to alleviate this issue.
Visual Planning: Let's Think Only with Images
Yi Xu (University of Cambridge), Ivan Vulić (University of Cambridge)
TransformerReinforcement LearningImage
🎯 What it does: This paper proposes the 'visual planning' paradigm, which uses large visual models to perform multi-step planning through pure image sequences, and designs a two-stage training framework VPRL based on reinforcement learning;
Visual Prompt-Agnostic Evolution
Junze Wang (University of Science and Technology Beijing), Cong Cong (University of New South Wales)
ClassificationSegmentationTransformerPrompt EngineeringImage
🎯 What it does: Proposes the Prompt-Agnostic Evolution (PAE) framework, treating Visual Prompt Tuning (VPT) as a Koopman-Lyapunov discrete dynamical system, adopting task-aware frequency-domain initialization and shared evolution operators to significantly accelerate convergence and improve accuracy.
Visual Self-Refine: A Pixel-Guided Paradigm for Accurate Chart Parsing
Jinsong Li (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)
RecognitionVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Propose the Visual Self-Refinement (VSR) framework and implement ChartVSR for chart parsing, generating pixel-level localization, visual feedback, iterative correction, then decoding the final structured data, and constructing a challenging ChartP-Bench benchmark;
Visual symbolic mechanisms: Emergent symbol processing in Vision Language Models
Rim Assouel (Mila Quebec AI Institute Universite de Montreal), Taylor Whittington Webb
TransformerVision Language ModelImageMultimodality
🎯 What it does: This paper studies and reveals the implicit symbolic mechanism for achieving visual binding in visual language models, discovering a three-stage architecture driven by spatial indexing (position ID);
VisualPRM400K: An Effective Dataset for Training Multimodal Process Reward Models
Weiyun Wang (Fudan University), Wenhai Wang (Chinese University of Hong Kong)
Data-Centric LearningReinforcement Learning from Human FeedbackReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Construct a multimodal process supervision dataset named VisualPRM400K with 400K samples, and develop the VisualPRM model to evaluate each step of the reasoning process;
VisualPrompter: Semantic-Aware Prompt Optimization with Visual Feedback for Text-to-Image Synthesis
Shiyu Wu (Institute of Automation, Chinese Academy of Sciences), Jing Liu (Institute of Automation, Chinese Academy of Sciences)
GenerationLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Developed a no-training, training-free prompt engineering framework called VisualPrompter, which fine-grained improves the model adaptability and semantic consistency of user prompts through visual feedback self-reflection (SERE) and task-specific optimization (TSPO);
VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models
Weiye Xu (University of Science and Technology of China), Jinguo Zhu (Shanghai Artifcial Intelligence Laboratory)
TransformerLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed the VisuLogic visual reasoning benchmark, containing 1,000 multiple-choice questions verified by humans, covering six reasoning categories: quantification, spatial, localization, attributes, style, and others.
VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning
Hao Yan (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Data SynthesisSupervised Fine-TuningReinforcement LearningImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed the VisuRiddles benchmark and synthesizer, and proposed the Perception-Augmented Visual Reasoner (PAVR) model, aiming to enhance the performance of multimodal large language models on abstract visual reasoning (AVR) tasks.
VITA: Vision-to-Action Flow Matching Policy
Dechen Gao (University of California, Davis), Iman Soltani (University of California, Davis)
Robotic IntelligenceTransformerVision-Language-Action ModelFlow-based ModelAuto EncoderImageMultimodalityOrdinary Differential Equation
🎯 What it does: Proposed a noise-free, vision-condition-free flow-matching visual-to-action strategy called VITA;
VITA: Zero-Shot Value Functions via Test-Time Adaptation of Vision–Language Models
Christos Ziakas (Imperial College London), Alessandra Russo (Imperial College London)
Domain AdaptationRobotic IntelligenceMeta LearningReinforcement LearningVision Language ModelContrastive LearningTime Series
🎯 What it does: Propose the VITA method, which improves the zero-shot value function estimation of VLM through test-time adaptation, addressing the bottlenecks in semantic and temporal reasoning.
VitaBench: Benchmarking LLM Agents with Versatile Interactive Tasks in Real-world Applications
Wei He (Fudan University), Xunliang Cai (Meitaun LongCat Team)
TransformerLarge Language ModelReinforcement LearningAgentic AITextTabularBenchmark
🎯 What it does: Built a multi-tool, multi-task simulation environment called VitaBench, covering three daily life scenarios: food delivery, in-store consumption, and online travel, and generated 400 tasks for it;
ViTSP: A Vision Language Models Guided Framework for Solving Large-Scale Traveling Salesman Problems
Zhuoli Yin (Purdue University), Hua Cai (Purdue University)
OptimizationVision Language ModelImageGraph
🎯 What it does: Proposed the ViTSP framework, which leverages pre-trained vision-language models (VLMs) to identify and partition subproblems in visualized TSP solutions, followed by an offline solver (Concorde) optimizing these subproblems for efficient large-scale TSP solving.
Vivid-VR: Distilling Concepts from Text-to-Video Diffusion Transformer for Photorealistic Video Restoration
Haoran Bai (Alibaba Group), Ying Chen (Alibaba Group)
RestorationGenerationKnowledge DistillationTransformerDiffusion modelAuto EncoderVideoText
🎯 What it does: Vivid-VR is a generative video restoration method based on Diffusion Transformer (DiT), which utilizes ControlNet to perform controllable generation on low-quality videos, enhancing texture realism and temporal consistency.
VL-JEPA: Joint Embedding Predictive Architecture for Vision-language
Delong Chen (Meta FAIR), Pascale Fung (Meta FAIR)
ClassificationRetrievalTransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes VL-JEPA, a non-generative vision-language model based on joint embedding prediction, supporting real-time inference, selective decoding, and handling multiple tasks such as classification, retrieval, and VQA in a single pass.
Vlaser: Vision-Language-Action Model with Synergistic Embodied Reasoning
Ganlin Yang (University of Science and Technology of China), Zhi Hou (Shanghai AI Laboratory)
Robotic IntelligenceTransformerLarge Language ModelMixture of ExpertsVision-Language-Action ModelFlow-based ModelImageTextMultimodality
🎯 What it does: Proposed Vlaser, a foundation model integrating vision-language-action capabilities, combining high-level reasoning with end-to-end robot control;
VLBiMan: Vision-Language Anchored One-Shot Demonstration Enables Generalizable Bimanual Robotic Manipulation
Huayi Zhou (Chinese University of Hong Kong), Kui Jia (Chinese University of Hong Kong)
Robotic IntelligenceVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: Proposes the VLBiMan framework, achieving consistent bimanual manipulation without retraining by leveraging a single human demonstration and a visual-language model;
VLM-Guided Adaptive Negative Prompting for Creative Generation
Shelly Golan, Or Patashnik
GenerationPrompt EngineeringVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: This study proposes a VLM-Guided Adaptive Negative-Prompting method based on vision-language models, which enhances creativity while maintaining class effectiveness by dynamically guiding the generation process away from conventional patterns through real-time querying of the VLM and accumulating negative prompts during the denoising process of diffusion models.
VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning?
Minkyu Kim (Krafton), Dongmin Park (Krafton)
Autonomous DrivingPrompt EngineeringVision Language ModelImageTextBenchmarkChain-of-Thought
🎯 What it does: Designed and released the VLM-SubtleBench benchmark, comprising 13K image pairs with subtle differences, along with question-answering and description tasks, to evaluate visual language models' performance in fine-grained comparative reasoning.
VLM4VLA: Revisiting Vision-Language-Models in Vision-Language-Action Models
Jianke Zhang (Tsinghua University), Jianyu Chen (Tsinghua University)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageText
🎯 What it does: This paper investigates the transfer effects of different vision-language models (VLM) in vision-language-action (VLA) tasks, proposing a lightweight VLM4VLA framework and conducting systematic evaluations on three major simulation benchmarks: Calvin, SimplerEnv, and Libero.
VLMgineer: Vision-Language Models as Robotic Toolsmiths
George Jiayuan Gao (University of Pennsylvania), Dinesh Jayaraman (University of Pennsylvania)
Robotic IntelligencePrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose VLMGINEER, the first complete automated framework for co-designing tools and actions using vision-language models and evolutionary search, evaluated on 12 robotic tool-use tasks (ROBOT_TOOLBENCH).
VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety
Shruti Palaskar (Apple), Joseph Yitan Cheng
Data SynthesisSafty and PrivacyImageTextMultimodalityBenchmark
🎯 What it does: Proposed the VLSU framework, constructed a dataset of 8,187 real image-text pairs tailored for multimodal safety, and conducted systematic evaluation of model performance on safety assessment tasks.
VMDiff: Visual Mixing Diffusion for Limitless Cross-Object Synthesis
Zeren Xiong (Nanjing University of Science and Technology), Jun Li (Nanjing University of Science and Technology)
GenerationData SynthesisVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: Proposes the Visual Mixing Diffusion (VMDiff) framework for fusing two input images into a single coherent innovative object;
VMoBA: Mixture-of-Block Attention for Video Diffusion Models
Jianzong Wu (Peking University), Yunhai Tong (Kling Team, Kuaishou Technology)
GenerationTransformerDiffusion modelVideoBenchmark
🎯 What it does: Propose a sparse attention mechanism for training video diffusion models