CVPR 2026 Papers — Page 40

VIRO: Robust and Efficient Neuro-Symbolic Reasoning with Verification for Referring Expression Comprehension

Hyejin Park (Pohang University of Science and Technology), Jungseul Ok (Pohang University of Science and Technology)

🎯 What it does: Propose the VIRO (Verification-Integrated Reasoning Operators) framework, which embeds a lightweight verifier in each step of neuro-symbolic reasoning to explicitly handle no-target scenarios.

VIRST: Video-Instructed Reasoning Assistant for SpatioTemporal Segmentation

Jihwan Hong (Seoul National University), Jaeyoung Do (Seoul National University)

🎯 What it does: proposes the VIRST framework, which unifies global video reasoning and pixel-level segmentation within a single vision-language model;

Virtual Full-stack Scanning of Brain MRI via Imputing Any Quantised Code

Yicheng Wu (Imperial College London), Jianfei Cai (Monash University)

🎯 What it does: Propose the CodeBrain framework, which performs a unified 'virtual full-stack scanning' to complete missing modalities in brain MRI;

Virtual Immunohistochemistry Staining with Dual-Aligned Multi-Task Feature Guidance

Shigeng Xie (Dalian University Of Technology), Fengyu Cong (Dalian University Of Technology)

🎯 What it does: This paper proposes a virtual immunohistochemistry (VIS) model based on bidirectional alignment multi-task features, which significantly improves the quality of virtual staining by utilizing auxiliary task features to guide the generator network at the feature level.

Virtual Nodes Guided Dynamic Graph Neural Network for Brain Tumor Segmentation with Missing Modalities

Sha Tao (University of Science and Technology Beijing), Chao Yao (University of Science and Technology Beijing)

🎯 What it does: Propose a single-stage framework based on graph neural networks, utilizing virtual nodes and dynamic edge connections to achieve brain tumor segmentation under missing modalities.

VirtueBench: Evaluating Trustworthiness under Uncertainty in Long Video Understanding

Xueqing Yu (Peking University), Zhenheng Yang (ByteDance)

🎯 What it does: Propose the VirtueBench benchmark to evaluate the honest rejection behavior of long video understanding models when critical information is missing.

VisiLock: Authorizing Instruction-based Image editing with Dual Score Distillation

Van Thanh Le (Northeastern University), Yun Fu (Northeastern University)

🎯 What it does: Propose Visilock, enabling instruction-guided image editing models to access permissions through visual key locking.

Vision Foundation Models Can Be Good Tokenizers for Latent Diffusion Models

Tianci Bi (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)

🎯 What it does: Propose VFM-VAE that directly uses a frozen visual foundation model (VFM) as the visual tokenizer of LDM, and designs a multi-scale fusion and advanced resolution reconstruction decoder to achieve high-quality reconstruction.

VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions

Adrian Bulat (Samsung AI Cambridge), Georgios Tzimiropoulos (Samsung AI Cambridge)

🎯 What it does: VISOR significantly reduces inference cost by sparsifying the interaction layers between images and text in large-scale vision-language models without losing visual information.

Vision Transformers Need More Than Registers

Cheng Shi (University of Hong Kong), Sibei Yang (Sun Yat-sen University)

🎯 What it does: Proposes LaSt-ViT, a frequency-aware selective aggregation mechanism that utilizes CLS tokens to aggregate stable foreground patches, thereby eliminating the 'lazy aggregation' and sparse feature anomalies that appear in ViT under different supervision modes;

Vision-Language Attribute Disentanglement and Reinforcement for Lifelong Person Re-Identification

Kunlun Xu (Peking University), Jiahuan Zhou (Peking University)

🎯 What it does: Achieve lifetime person re-identification (LReID) using pre-trained vision-language models (VLM), enhancing cross-domain knowledge transfer and memory retention by explicitly separating global and local attributes.

Vision-Language Model Guided Source-Free Domain Adaptation via Optimal Transport

Shuo Han (Xidian University), Xiangrong Zhang (Xidian University)

🎯 What it does: Leverage pre-trained vision-language models (VLM) as external semantic priors, combining optimal transport (OT) alignment with bidirectional distillation to achieve source-agnostic domain adaptation.

Vision-Oriented Lightweight Neural Architecture Search with Budget-Adaptive Evaluation

Yi Fan (Nanjing University), Yu-Bin Yang (Nanjing University)

🎯 What it does: Proposed a vision-oriented lightweight NAS framework that utilizes six low-cost visual probe tasks and a budget-adaptive evaluator to rapidly assess network performance with extremely low training costs;

Vision-Speech Models: Teaching Speech Models to Converse about Images

Amélie Royer (Kyutai), Patrick Pérez (Kyutai)

🎯 What it does: Developed a multi-modal speech-visual dialogue model called MoshiVis, capable of naturally discussing image content in real-time bidirectional speech dialogues. The model achieves image information injection by adding a lightweight visual adaptation module to the pre-trained speech LLM Moshi, and supports seamless switching between image-related and general topics during conversations.

VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image Synthesis

Meng Chu, Jiaya Jia

🎯 What it does: This paper proposes the VisionDirector framework and the LongGoalBench benchmark, which use a vision-language model (VLM) planner to perform closed-loop control of image generation and editing under long-textual goal prompts.

VisionLeaf: Entropy-Guided Leaf-First Reasoning for Efficient and Accurate Think-with-Image

Haokun Gui (Hong Kong University Of Science And Technology), Jiaya Jia (Hong Kong University Of Science And Technology)

🎯 What it does: Propose VisionLeaf, an entropy-guided leaf-priority tree reasoning framework, to improve the inference efficiency and accuracy of vision-language models in multi-step tool calling (think-with-image) tasks.

VisMem: Latent Vision Memory Unlocks Potential of Vision-Language Models

Xinlei Yu (Zhejiang University), Shuicheng Yan (National University of Singapore)

🎯 What it does: Proposed a visual memory framework called VisMem based on cognitive theory, which injects short-term perceptual memory and long-term semantic memory into visual language models, and dynamically inserts latent visual memories into the autoregressive inference process through special memory invocation tokens.

VisPlay: Self-Evolving Vision-Language Models

Yicheng He (University of Illinois Urbana Champaign), Yonghui Yang (National University of Singapore)

🎯 What it does: Proposes VisPlay, a self-evolving RL framework that enables vision-language models to autonomously generate questions and answers on unlabeled images, continuously enhancing reasoning capabilities.

VisRef: Visual Refocusing while Thinking Improves Test-Time Scaling in Multi-Modal Large Reasoning Models

Soumya Suvra Ghosal (University of Maryland), Qin Zhang (Physion Labs)

🎯 What it does: Maintain visual grounding in multimodal models through dynamic re-injection of visual tokens during the reasoning phase;

VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs

Brigitta Malagurski Törtei (Technology Innovation Institute), Sanath Narayan (Technology Innovation Institute)

🎯 What it does: Propose VisRes Bench, a three-layer visual reasoning benchmark consisting of 19,000 natural images, designed to evaluate the perception and reasoning capabilities of vision-language models.

VISTA: A Test-Time Self-Improving Video Generation Agent

Do Xuan Long (Google), Sercan Ö. Arik (Google)

🎯 What it does: Proposed an adaptive multi-agent framework called VISTA, which achieves self-improvement in text-to-video generation by decomposing user prompts, generating candidate videos, performing multi-dimensional critique on the best video, and automatically rewriting prompts during inference.

Vista4D: Video Reshooting with 4D Point Clouds

Kuan Heng Lin, Ning Yu

🎯 What it does: Developed a 4D point cloud-based view re-synthesis framework called Vista4D, which can re-synthesize input videos with different camera trajectories and perspectives while maintaining dynamic consistency.

ViStoryBench: Comprehensive Benchmark Suite for Story Visualization

Cailin Zhuang (ShanghaiTech University), Chi Zhang (Westlake University)

🎯 What it does: Propose the ViStoryBench benchmark, constructing 80 multi-shot story sets with role reference images and scripts, defining 12-dimensional automatic evaluation metrics, and conducting systematic evaluations of 30+ visual narrative models.

Visual Diffusion Models are Geometric Solvers

Nir Goren (Tel Aviv University), Daniel Cohen-Or (Tel Aviv University)

🎯 What it does: This paper applies the Visual Diffusion Model as a solver for geometric problems, directly generating approximate solutions in pixel space that satisfy geometric constraints.

Visual Document Understanding and Reasoning: A Multi-Agent Collaboration Framework with Agent-Wise Adaptive Test-Time Scaling

Xinlei Yu (National University Of Singapore), Xiaobin Hu (National University Of Singapore)

🎯 What it does: Propose the MACT framework, decomposing visual document understanding and reasoning into four specialized agents (planning, execution, judgment, and answering), and designing adaptive reasoning time scaling and hybrid reward modeling at the agent level.

Visual Grounding for Object Questions

Martin Nicolas Everaert (EPFL), Vidya Narayanan (Amazon Inc)

🎯 What it does: Propose a new visual localization task: Visual Grounding for Object Questions (VGOQ), which involves locating visual evidence or contextual regions that help answer open-ended object-related questions given an image and the question; simultaneously, two automatically generated synthetic datasets (VizWiz‑VGOQ and ABO‑VGOQ) are constructed, and a lightweight CLIP‑Seg‑based model is trained to address this task.

Visual Personalization Turing Test

Rameen Abdal (Snap Research), Kuan-Chieh Jackson Wang (Snap Research)

🎯 What it does: Propose the Visual Personalization Turing Test (VPTT) framework, which includes a benchmark with 10,000 synthetic personas, a retrieval-augmented generation model called VPRAG, and a visualizable and interpretable VPTT score to evaluate whether generated content is indistinguishable from the creative style of a specific user.

Visual Prototype Conditioned Focal Region Generation for UAV-Based Object Detection

Wenhao Li (Beihang University), Jiaxin Chen (Beihang University)

🎯 What it does: Propose the UAVGen framework, achieving diffusion-based UAV object detection data augmentation through a vision prototype conditional diffusion model and a focused region enhancement data pipeline

Visual-Aware CoT: Achieving High-Fidelity Visual Consistency in Unified Models

Zixuan Ye (Hong Kong University of Science and Technology), Wenhan Luo (Hong Kong University of Science and Technology)

🎯 What it does: Proposes the Visual-Aware CoT (VACoT) framework, incorporating adaptive visual planning and iterative visual correction within a unified model to achieve visual consistency in multi-reference image generation.

Visual-RRT: Finding Paths toward Visual-Goals via Differentiable Rendering

Sebin Lee (KAIST), Sung-Eui Yoon (KAIST)

🎯 What it does: Designed a visual target motion planning method called vRRT, combining differentiable rendering gradients and RRT exploration to enable robot path planning based solely on target images.

VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer

Yanning Hou (National University of Defense Technology), Ke Xu (Anhui University)

🎯 What it does: Proposed a zero-shot anomaly detection framework called VisualAD that uses only visual Transformers. It freezes the ViT and inserts learnable abnormal and normal global tokens into its input sequence. These tokens interact across multiple layers of features through spatially aware cross-attention (SCA) and self-alignment function (SAF), ultimately generating pixel-level anomaly maps and image-level anomaly scores.

VisualOverload: Probing Visual Understanding of VLMs in Really Dense Scenes

Paul Gavrikov (Independent Researcher), Hilde Kuehne (Tübingen AI Center)

🎯 What it does: Proposes VisualOverload, a new visual question answering benchmark focused on fine-grained visual understanding in high-resolution, dense scenes.

ViT$^3$: Unlocking Test-Time Training in Vision

Dongchen Han (Tsinghua University), Gao Huang (Alibaba Group)

🎯 What it does: This paper systematically experiments on Test-Time Training (TTT) design in the visual field, proposes six practical experiences, and builds a linear complexity Vision Test-Time Training (ViT³) model based on these experiences, which can achieve efficient and competitive performance in various visual tasks;

VITAL: Vision-Encoder-centered Pre-training for LMMs in Visual Quality Assessment

Ziheng Jia (Shanghai Jiaotong University), Xiongkuo Min (Shanghai Jiaotong University)

🎯 What it does: Constructed a large-scale machine-annotated dataset containing 458 million visual-language pairs, and developed the VITAL series of multimodal large models centered on visual encoders for visual quality assessment.

ViterbiPlanNet: Injecting Procedural Knowledge via Differentiable Viterbi for Planning in Instructional Videos

Luigi Seminara (University of Catania), Antonino Furnari (University of Catania)

🎯 What it does: For procedural planning tasks in instructional videos, this paper proposes an end-to-end ViterbiPlanNet framework that integrates program knowledge graphs (PKG) into the decoding process via a differentiable Viterbi layer (DVL), enabling joint learning of emission probability prediction from visual inputs and structured planning.

ViTPrompt: Training-Free Prompt Refinement with Visual Tokens for Open-Vocabulary Detection

Yitong Qin (University of Electronic Science and Technology of China), Heng Tao Shen (University of Electronic Science and Technology of China)

🎯 What it does: Propose ViTPrompt, a training-free, two-phase inference open-domain object detection adaptation framework that leverages RoI visual tokens from initial high-confidence detections to enhance text prompts, jointly improving localization and classification.

VIVA: VLM-Guided Instruction-Based Video Editing with Reward Optimization

Xiaoyan Cong (Brown University), Chongyang Ma (ByteDance)

🎯 What it does: Propose VIVA, an instruction-driven video editing framework guided by Vision-Language Models (VLM) and combined with reward optimization.

VKG-QA: Visual Knowledge Graph-based Question Answer for Large Multimodal Models

Yuntao Du (Shandong University), Lizhen Cui (Shandong University)

🎯 What it does: Proposes a visual knowledge graph-based question answering benchmark, VKG-QA, systematically evaluating the ability of large-scale multimodal models in visual structured knowledge reasoning.

VL-Eraser: Vacuum Distillation for Machine Unlearning in Vision-Language Models

Yili Wang (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)

🎯 What it does: Propose a two-stage VLM machine unlearning method called VL-Eraser: first, unwanted knowledge is migrated to a low-rank adapter (LoRA) through vacuum distillation, followed by complete elimination of knowledge via parameter arithmetic deletion.

VL-RouterBench: A Benchmark for Vision-Language Model Routing

Zhehao Huang (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)

🎯 What it does: This paper proposes VL-RouterBench, a unified benchmark for evaluating vision-language model (VLM) routing systems, along with a complete data construction and evaluation toolchain.

VLA Models Are More Generalizable Than You Think: Revisiting Physical and Spatial Modeling

Weiqi Li (Sun Yat-sen University), Guangrun Wang (Sun Yat-sen University)

🎯 What it does: To address the degradation of vision-language-action (VLA) model robustness caused by perspective changes and visual disturbances, this paper proposes a "one-shot adaptation" framework: Feature Token Modulation (FTM) and Feature Linear Adaptation (FLA), which significantly restores and enhances model performance under new perspectives and various visual disturbances with only a minimal number of learnable parameters.

VLIC: Vision-Language Models As Perceptual Judges for Human-Aligned Image Compression

Kyle Sargent (Stanford University), Jason Y. Zhang (Google Research)

🎯 What it does: Propose an image compression method called VLIC based on diffusion autoencoders, followed by zero-shot human perception judgment using Vision-Language Models (VLM), and post-training the model with Diffusion DPO to enhance the subjective quality of compressed images.

VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction

Zhiwen Fan (University Of Texas At Austin), Rakesh Ranjan (Meta)

🎯 What it does: This paper proposes the VLM-3R framework, integrating 3D reconstruction instruction fine-tuning into a monocular video vision-language model, enabling it to directly recover implicit 3D structure from videos and perform spatial and spatiotemporal reasoning.

VLM-Guided Group Preference Alignment for Diffusion-based Human Mesh Recovery

Wenhao Shen (Nanyang Technological University), Guosheng Lin (SenseTime Research)

🎯 What it does: Construct a critic agent based on a Visual Language Model (VLM) that scores multiple human mesh predictions generated from a single RGB image, and use these scores to build an unlabeled group preference dataset. Subsequently, fine-tune a diffusion-based human mesh recovery (HMR) model through a group preference alignment framework (improved GRPO), thereby enhancing the physical feasibility and consistency with the image of predictions.

VLM-Loc: Localization in Point Cloud Maps via Vision-Language Models

Shuhao Kang (Nankai University), Yun Liu (Nankai University)

🎯 What it does: Proposed the VLM-Loc framework, which leverages visual language models (VLM) to map natural language descriptions to precise locations in point cloud maps, achieving text-to-point cloud (T2P) localization.

VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm

Zhenkai Wu (Huawei Technologies), Xinghao Chen (Huawei Technologies)

🎯 What it does: Propose a training-free visual-language model (VLM) token pruning method called VLM-Pruner, which significantly reduces the number of visual tokens while maintaining model performance.

VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models

Juncan Deng (Zhejiang University), Kejie Huang (Zhejiang University)

🎯 What it does: This paper proposes a post-training quantization method for large-scale vision-language models, VLM-PTQ, which can efficiently compress the model at low bit rates.

VLM4RSDet: Collaborative Optimization with Vision-Language Model for Enhancing Remote Sensing Object Detection

Shuohao Shi (National University of Defense Technology), Xin Xu (National University of Defense Technology)

🎯 What it does: Propose a collaborative optimization framework called VLM4RSDet, which jointly trains a traditional closed-set remote sensing object detector with a vision-language model. During inference, only the detector is retained to avoid additional computational costs.

VMD-FACT: A New Video Dataset and MLLM-based method for Detecting Realistic AI-Generated Video Misinformation

Yongkang Zhang (Beihang University), Yan Wang (Zhongguancun Laboratory)

🎯 What it does: Proposed the RAVM dataset and the IEEG method for detecting AI-generated video rumors in real-world scenarios;

VMonarch: Efficient Video Diffusion Transformers with Structured Attention

Cheng Liang (Nanjing University), Limin Wang (Nanjing University)

🎯 What it does: The paper proposes a new efficient attention mechanism called VMonarch in video diffusion Transformers, using a spatiotemporal structured Monarch matrix to approximate sparse attention.

Vocabulary Scaling Law: Tuning Open-vocabulary Predictors for Their Openness

Ziliang Chen (Peng Cheng Laboratory), Xipeng Chen (Peng Cheng Laboratory)

🎯 What it does: This paper proposes a framework called SVFT, based on submodular optimization, for selecting a subset of vocabulary and fine-tuning class names to enhance CLIP's stability and scalability under open-vocabulary settings.

VoDaSuRe: A Large-Scale Dataset Revealing Domain Shift in Volumetric Super-Resolution

August Leander Høeg (Technical University of Denmark), Anders Bjorholm Dahl (Technical University of Denmark)

🎯 What it does: This paper proposes a large voxel super-resolution dataset called VoDaSuRe and evaluates various existing SR models on this dataset, demonstrating significant domain differences between models trained on simulated downsampled data and real low-resolution scans.

VOLD: Reasoning Transfer from LLMs to Vision-Language Models via On-Policy Distillation

Walid Bousselham (University of Tuebingen), Cordelia Schmid (Inria)

🎯 What it does: Propose the VOLD framework, which transfers the reasoning capabilities of a text LLM teacher to vision-language models through a two-stage training process (SFT alignment + GRPO + on-policy distillation).

Volumetric Functional Maps

Filippo Maggioli (Pegaso University), Marco Livesu (CNR IMATI)

🎯 What it does: Achieved the first functional mapping framework for volumetric grids, computing high-quality correspondences between 3D volumes and validated in multiple applications including segmentation transfer, mesh connectivity migration, and texture synthesis.

VOSR: A Vision-Only Generative Model for Image Super-Resolution

Rongyuan Wu (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)

🎯 What it does: Proposed VOSR, a fully vision-based generative super-resolution model.

Voxify3D: Pixel Art Meets Volumetric Rendering

Yi-Chuan Huang (National Yang Ming Chiao Tung University), Yu-Lun Liu (National Yang Ming Chiao Tung University)

🎯 What it does: Propose a two-stage differentiable framework that converts 3D meshes into voxel art with a pixel art style, integrating orthogonal pixel art supervision, CLIP semantic alignment, and palette-based Gumbel-Softmax discrete quantization.

VoxTell: Free-Text Promptable Universal 3D Medical Image Segmentation

Maximilian Rokuss (German Cancer Research Center), Klaus Maier-Hein (German Cancer Research Center)

🎯 What it does: Propose VoxTell, a vision-language model capable of directly generating 3D medical image segmentation masks from arbitrary natural language prompts.

VQ-VA World: Towards High-Quality Visual Question-Visual Answering

Chenhui Gou (Monash University), Hamid Rezatofighi (Monash University)

🎯 What it does: Proposed the Visual Question-Visual Answer (VQ-VA) task and constructed a large-scale dataset VQ-VA World along with a human-curated evaluation benchmark IntelligentBench.

VQRAE: Representation Quantization Autoencoders for Multimodal Understanding, Generation and Reconstruction

Sinan Du (Tsinghua University), Chun Yuan (Tsinghua University)

🎯 What it does: Propose VQRAE, a vector-quantized representation autoencoder that achieves unified visual representation generation, understanding, and reconstruction;

VRCLIP: Multimodal Canonical Correlation Alignment for CLIP-Driven Vision-Radio Person Re-Identification

Rui Zhang (University of Science and Technology of China), Yan Chen (University of Science and Technology of China)

🎯 What it does: Proposed the VRCLIP framework, which jointly uses RGB images and low-frequency RF signals for cross-modal person re-identification;

VRR-QA: Visual Relational Reasoning in Videos Beyond Explicit Cues

Sirnam Swetha (University of Central Florida), Mubarak Shah (University of Central Florida)

🎯 What it does: Constructed VRR-QA, a new benchmark focusing on implicit visual relationship reasoning, manually selecting 1k question-answer pairs from 1000 movies and classifying them into 9 categories of reasoning types.

VS-Bench: Evaluating VLMs for Strategic Abilities in Multi-Agent Environments

Zelai Xu, Yu Wang

🎯 What it does: Proposes VS-BENCH—a benchmark comprising ten visually driven multi-agent games to evaluate vision-language models (VLMs) in cooperative, competitive, and mixed-motivation environments, assessing their perception, strategic reasoning, and decision-making capabilities.

VSRELL: A Simple Baseline for Video Super-Resolution and Enhancement in Low-Light Environment

Yanming Hui (Tianjin University), Bingqin Lv (Tianjin University)

🎯 What it does: Propose an end-to-end joint video super-resolution and low-light enhancement framework called VSRELL, which can directly restore low-light, low-resolution videos into high-resolution normally illuminated sequences.

VT-Intrinsic: Physics-Based Decomposition of Reflectance and Shading using a Single Visible-Thermal Image Pair

Zeqing Yuan (Carnegie Mellon University), Srinivasa G. Narasimhan (Carnegie Mellon University)

🎯 What it does: Using a single visible-infrared image pair, this work decomposes albedo and shading under physical constraints.

VULCAN: Tool-Augmented Multi Agents for Iterative 3D Object Arrangement

Zhengfei Kuang (Stanford University), Sanghyun Woo (Google)

🎯 What it does: Proposed a multi-step text-to-3D object arrangement system named VULCAN, leveraging a multi-agent architecture, visual tools, and constraint solvers to achieve a complete closed-loop from user instructions to precise object arrangement.

VVS: Accelerating Speculative Decoding for Visual Autoregressive Generation via Partial Verification Skipping

Haotian Dong (Shenzhen International Graduate School Tsinghua University), Zhi Wang (Shenzhen International Graduate School Tsinghua University)

🎯 What it does: Proposed a new visual autoregressive generation acceleration framework called VVS, which significantly reduces the number of forward passes of the target model by partially skipping verification steps during inference.

W2W: Language-Model-Based Trajectory Prediction with Reinforcement Learning

Zirui Xu (Changzhou University), Shaobo Shen (Changzhou University)

🎯 What it does: Propose Write-to-Walk (W2W), which converts observed trajectories and social interactions into interpretable natural language prompts, and achieves pedestrian trajectory prediction through two-stage training with T5-Small (full-parameter supervised fine-tuning + LoRA+PPO reinforcement learning).

WaDi: Weight Direction-aware Distillation for One-step Image Synthesis

Lei Wang (Nankai University), Jian Yang (Nankai University)

🎯 What it does: Proposed the WaDi framework, which significantly improves single-step text generation speed and quality through weight-direction-aware distillation of first-order diffusion models using the low-rank rotation adapter LoRaD.

WalkGPT: Grounded Vision-Language Conversation with Depth-Aware Segmentation for Pedestrian Navigation

Rafi Ibn Sultan (Wayne State University), Dongxiao Zhu

WAM-Flow: Parallel Coarse-to-Fine Motion Planning via Discrete Flow Matching for Autonomous Driving

Yifang Xu (Fudan University), Siyu Zhu (Fudan University)

🎯 What it does: WAM-Flow treats vehicle trajectory planning as a discrete flow matching problem, achieving a tunable coarse-to-fine reasoning process through parallel denoising.

Wan-Weaver: Interleaved Multi-modal Generation via Decoupled Training

Jinbo Xing (Tongyi Lab), Yujiu Yang (Tsinghua University)

🎯 What it does: Propose a unified multimodal model named Wan-Weaver, capable of generating long text and image-interleaved multimodal content without requiring real interleaved data.

Wanderland: Geometrically Grounded Simulation for Open-World Embodied AI

Xinhao Liu (New York University), Chen Feng (New York University)

🎯 What it does: Designed and implemented WANDERLAND, a multi-sensor handheld device-based real-to-sim conversion framework for constructing geometrically faithful and photorealistic open-world simulation environments, and released the corresponding dataset.

Watch and Learn: Learning to Use Computers from Online Videos

Chan Hee Song (Ohio State University), Tomas Pfister (Google Cloud AI Research)

🎯 What it does: Built the Watch & Learn framework, which automatically generates 53K executable UI trajectories using large-scale internet tutorial videos and inverse dynamics models, and employs them to enhance the performance of CUAs in multi-task computer interaction tasks.

WaTeRFlow: Watermark Temporal Robustness via Flow Consistency

Utae Jeong (Korea University), Sangpil Kim (Korea University)

🎯 What it does: Improving the robustness of watermarks during image-to-video (I2V) generation, the WaTeRFlow framework is proposed to achieve reliable recovery of watermarks in video frames.

Wave-Former: Through-Occlusion 3D Reconstruction via Wireless Shape Completion

Laura Dodds (Massachusetts Institute of Technology), Fadel Adib (Massachusetts Institute of Technology)

🎯 What it does: Propose Wave-Former, a method that uses mmWave signals to reconstruct the 3D shape of completely occluded objects, converting raw wireless signals into complete 3D point clouds.

Wavelet-based Frame Selection by Detecting Semantic Boundary for Long Video Understanding

Wang Chen (Xiamen University), Xiawu Zheng (Xiamen University)

🎯 What it does: For long video understanding, an untrained wavelet transform framework named WFS-SB is proposed, achieving frame selection by detecting semantic boundaries

Wavelet-Driven 3D Anomaly Detection under Pose-Agnostic and Sparse-View

Mingwen Shao (Shenzhen University of Advanced Technology), Ling Jian (China University of Petroleum (East China))

🎯 What it does: This paper proposes Wave-Pose3D, which constructs a complete three-stage framework for the task of pose-agnostic anomaly detection under sparse views.

Weakly Supervised Video Anomaly Detection with Anomaly-Connected Components and Intention Reasoning

Yu Wang (Tongji University), Shengjie Zhao (Tongji University)

🎯 What it does: Propose the LAS-VAD framework for weakly supervised video anomaly detection, integrating anomaly connected components (ACC) with intention awareness (IAM) mechanisms, and combining anomaly attribute information to improve detection accuracy.

WeatherCity: Urban Scene Reconstruction with Controllable Multi-Weather Transformation

Wenhua Wu (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)

🎯 What it does: Propose WeatherCity, an editable high-fidelity 4D urban scene reconstruction and controllable multi-weather transformation framework.

WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation

Wei Chow (National University Of Singapore), Tat-Seng Chua (National University Of Singapore)

🎯 What it does: Proposed WEAVE, the first dataset and evaluation benchmark for multimodal understanding and generation in multi-round interactions

WeaveTime: Streaming from Earlier Frames into Emergent Memory in VideoLLMs

Yulin Zhang (ShanghaiTech University), Sibei Yang (University of Hong Kong)

🎯 What it does: Propose the WeaveTime framework, addressing the issue of temporal irrelevance in streaming Video-LLM to enable online video question answering.

WebChain: A Large-Scale Human-Annotated Dataset of Real-World Web Interaction Traces

Sicheng Fan (Fudan University), Dehan Kong (IMean AI)

🎯 What it does: This study constructs WebChain, the largest-scale, fully human-annotated real-world web interaction trajectory dataset, and proposes the Dual Mid-Training training scheme based on this data, achieving SOTA performance on WebChainBench and multiple public GUI benchmarks.

WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks

Hao Bai (Microsoft), Spencer Whitehead (Microsoft)

🎯 What it does: Built WebGym, an open-source training environment containing nearly 300,000 real web tasks, supporting evaluation, and asynchronous rollout.

WeDetect: Fast Open-Vocabulary Object Detection as Retrieval

Shenghao Fu (Tencent Inc), Wei-Shi Zheng (Tencent Inc)

🎯 What it does: Proposed a novel retrieval-based open-vocabulary object detection framework called WeDetect, and extended it to a unified proposal generator WeDetect-Uni and an LLM-driven referential expression understanding model WeDetect-Ref.

Weight Space Representation Learning via Neural Field Adaptation

Zhuoqian Yang (École Polytechnique Fédérale de Lausanne), Sabine Süsstrunk (École Polytechnique Fédérale de Lausanne)

🎯 What it does: Generate weight space representations applicable to reconstruction, generation, and discrimination tasks by adapting pretrained base neural fields through multiplicative LoRA.

WeMMU: Enhanced Bridging of Vision-Language Models and Diffusion Models via Noisy Query Tokens

Jian Yang (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)

🎯 What it does: Propose a framework named WeMMU that efficiently bridges a frozen Vision-Language Model with a tunable diffusion model for unified multimodal tasks such as text-to-image generation and image editing, utilizing Noisy Query Tokens and a VAE branch.

What Are You Doing? A Closer Look at Controllable Human Video Generation

Emanuele Bugliarello (Google DeepMind), Cordelia Schmid (Google DeepMind)

🎯 What it does: Created and released the controllable human video generation benchmark WYD, providing fine-grained multi-class annotations and corresponding evaluation frameworks

What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models

Yingqi Fan (Eastern Institute of Technology), Xiaoyu Shen (Eastern Institute of Technology)

🎯 What it does: This study constructs the EmbedLens probe framework to perform fine-grained analysis of visual tokens in multi-modal large language models (MLLMs), revealing that visual tokens can be categorized into three types: sink, dead, and alive, and demonstrating that alive tokens carry most of the image semantics. Further quantitative experiments show that visual self-attention and feed-forward networks inside MLLMs contribute little to most tasks, and shallow processing has minimal impact on visual tokens, suggesting that visual inputs should be directly injected into intermediate layers.

What Is It Like to Be a Noise? An Entropy-based Gaussian Noise Regularization for Diffusion Models

Pascal Chang (Eth Zurich), Studios___ Switzerland 0009-0002-4133-4309

🎯 What it does: Proposed an entropy-based Gaussian regularization method to optimize the noise vector during the inference of diffusion models, ensuring the optimized noise maintains a typical Gaussian distribution and preventing artifacts and reward hijacking.

What Is the Optimal Ranking Score Between Precision and Recall? We Can Always Find It and It Is Rarely F1

Sébastien Piérard (University of Liège), Marc Van Droogenbroeck (University of Liège)

🎯 What it does: This paper studies how to find the optimal ranking metric between precision and recall in binary classification problems. The authors prove that all Fβ scores (i.e., weighted harmonic means) can produce meaningful rankings and construct a set of rankings along a 'shortest path.' Subsequently, they define and compute the optimal β value using Karcher mean and Kendall distance, and provide a closed-form solution.

What Makes Good Synthetic Training Data for Zero-Shot Stereo Matching?

David Yan (Princeton University), Jia Deng (Princeton University)

🎯 What it does: Built a procedural data generator based on Infinigen and Blender, systematically exploring and optimizing various scene, object, material, lighting, and camera baseline parameters to generate the WMGStereo-150k dataset for training stereo matching networks;

What Matters in Practical Learned Image Compression

Kedar Tatwawadi (Apple), Oren Rippel (Apple)

🎯 What it does: Propose a real-time, low-latency image compression stream called PICO that can run on mobile devices, achieving 12MP image encoding in 230 ms and decoding in 150 ms on the iPhone 17 Pro Max.

What Your Features Reveal: Data-Efficient Black-Box Feature Inversion Attack for Split DNNs

Zhihan Ren (Xi'an Jiaotong University), Fan Li (Xi'an Jiaotong University)

🎯 What it does: Studied a black-box feature inversion attack (FIA-Flow), which can high-fidelity recover the original input image from the intermediate features transmitted by Split DNN with only a small number of image-feature pairs.

What's Wrong with Synthetic Data for Scene Text Recognition? A Strong Synthetic Engine with Diverse Simulations and Self-Evolution

Xingsong Ye (Fudan University), Zhineng Chen (Tencent Inc)

🎯 What it does: Propose the UnionST synthesis engine and build three synthetic datasets, UnionST-S, UnionST-P, and UnionST-SP, based on it. Additionally, design a Self-Evolutionary Learning (SEL) framework for semi-supervised training on large-scale unlabeled real data.

When Anonymity Breaks: Identifying Models Behind Text-to-Image Leaderboards

Ali Naseh (University of Massachusetts Amherst), Amir Houmansadr (University of Massachusetts Amherst)

🎯 What it does: This paper investigates the anonymity of text-to-image (T2I) models in voting leaderboards, demonstrating that high-precision de-anonymization is achievable using only image embeddings;

When AVSR Meets Video Conferencing: Dataset, Degradation, and the Hidden Mechanism Behind Performance Collapse

Yihuan Huang (Key Laboratory of Aerospace Information Security and Trusted Computing Ministry of Education), Kai Li (Tsinghua University)

🎯 What it does: Systematically evaluate the performance of AVSR in video conferencing, construct the first multi-modal dataset MLD-VC specifically for video conferencing, and enhance model robustness through fine-tuning on this dataset.

When CLIP Sees More, It Fights Back Harder: Multi-View Guided Adaptive Counterattacks for Test-Time Adversarial Robustness

Sunoh Kim (Dankook University), Daeho Um (University of Seoul)

🎯 What it does: Propose a multi-view method (MAC) for adaptive adversarial attacks on the CLIP model during testing, enhancing adversarial robustness through adversarial recovery guided by multiple views.

When Do Models Actually Decide? Mapping the Layer-Wise Decision Timeline in Pretrained Neural Networks

Minhyeok Lee (Chung-Ang University)

🎯 What it does: Perform layer-wise linear probing on pre-trained ResNet and models such as ViT and ConvNeXt, record the decision layer for each image, and investigate when the model truly makes predictions during forward propagation.

When Lines Meet Textures: Spatial-Frequency Aligned Diffusion Features for Cross-Sparsity Correspondence

Mingrui Zhu (Xidian University), Xinbo Gao (Xidian University)

🎯 What it does: This paper addresses the cross-sparse correspondence problem between sparse lines and texture-rich images, proposing the SFA-DIFT framework that achieves high-precision correspondence by fusing spatial and frequency domain alignment.

When Local Rules Create Global Order: Self-Organized Representation Learning for Latent Diffusion Models

Junrong Lian (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)

🎯 What it does: Proposed a self-organizing representation learning framework called SORL to simultaneously achieve local smoothness and global dispersion of the latent space structure in latent diffusion models.