🎯 What it does: This paper proposes a method called HIERAMP, which utilizes a visual autoregressive model (VAR) to inject learnable class tokens at multiple scales and enhances attention at each scale, thereby guiding the model to focus on semantically important regions during the dataset distillation process and generating more discriminative synthetic samples.
Hierarchical Action Learning for Weakly-Supervised Action Segmentation
Junxian Huang (Guangdong University of Technology), Shenghua Gao (University of Hong Kong)
CodeSegmentationTransformerVideoBenchmark
🎯 What it does: Proposed the Hierarchical Action Learning (HAL) model, which models weakly supervised action segmentation using hierarchical causal generative processes and pyramid Transformers, addressing issues of over-segmentation and boundary noise in traditional methods.
Hierarchical Attacks for Multi-Modal Multi-Agent Reasoning
Hao Zhou (JD.com), Ai Han (JD.com)
CodeAdversarial AttackLarge Language ModelAgentic AIMultimodality
🎯 What it does: This paper proposes a hierarchical attack framework called HAM³ to systematically evaluate the vulnerability of multi-modal multi-agent systems in three layers: perception, communication, and reasoning, and experiments are conducted on the GQA visual question answering task.
Hierarchical Concept Embedding & Pursuit for Interpretable Image Classification
Nghia Nguyen (University of Pennsylvania), René Vidal (University of Pennsylvania)
CodeClassificationExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes the Hierarchical Concept Embedding & Pursuit (HCEP) framework, which constructs concept embeddings in the latent space of vision-language models by leveraging semantic hierarchical structures, and recovers the conceptual paths contained in images through hierarchical sparse coding (Hierarchical Beam OMP), thereby achieving interpretable image classification.
HiF-VLA: Hindsight, Insight and Foresight through Motion Representation for Vision-Language-Action Models
Minghui Lin, Donglin Wang (Westlake University)
CodeRobotic IntelligenceTransformerLarge Language ModelMixture of ExpertsVision-Language-Action ModelOptical FlowVideoText
🎯 What it does: Propose HiF-VLA, a vision-language-action model that achieves bidirectional temporal reasoning through motion vectors, significantly improving the continuity and consistency of long-horizon manipulation tasks.
HiFICL: High-Fidelity In-Context Learning for Multimodal Tasks
Xiaoyu Li (University of Electronic Science and Technology of China), Zihan Xiong (University of Electronic Science and Technology of China)
CodeMeta LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose HiFICL, achieving In-Context Learning (ICL) mechanism for high-fidelity multimodal models through learnable virtual key-value pairs and low-rank decomposition, and present it as a parameter-efficient fine-tuning method.
🎯 What it does: Proposed an event camera-based turbulence elimination method called EHETM, which utilizes high temporal resolution events to extract fine-grained motion information within a short time, achieving high-quality image recovery with low frame rates.
🎯 What it does: Investigated the detection of critical configurations (i.e., scenarios causing ambiguity or instability in fundamental matrix estimation) using eight point correspondences across two views, and proposed a novel discriminative method based on homotopy projective transformation.
HOPS: Hierarchical Open-vocabulary Part Segmentation with Attention-Aware Filtering and Affinity-Guided Enhancement
Xinlong Li (Tianjin University), Wei Feng (Tianjin University)
CodeSegmentationTransformerVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Proposes HOPS, a hierarchical open-vocabulary part segmentation framework, combining the Attention-Aware Filtering Module (AFM) and Affinity-Guided Enhancement Module (AEM) to address object over-segmentation and part under-segmentation issues.
HulluEdit: Single-Pass Evidence-Consistent Subspace Editing for Mitigating Hallucinations in Large Vision-Language Models
Yangguang Lin (Beijing University of Posts and Telecommunications), Jitao Sang (Beijing University of Posts and Telecommunications)
CodeExplainability and InterpretabilityComputational EfficiencyVision Language ModelMultimodality
🎯 What it does: Propose a single-channel, reference-free orthogonal subspace editing method called HulluEdit to address the problem of object hallucination in large vision-language models.
HumanBA: Human-Aware Bundle Adjustment via Global Human-Camera Decoupling
Fengyuan Yang (National University of Singapore), Angela Yao (National University of Singapore)
CodePose EstimationOptimizationSimultaneous Localization and MappingVideoMesh
🎯 What it does: This paper proposes HumanBA, a method that decouples dynamic human motion into camera motion and human body motion, and incorporates pseudo-static human joints as constraints into Bundle Adjustment;
CodeData SynthesisTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose the HUMANVBENCH video benchmark, which includes 16 fine-grained human-centric tasks, and develop an automated annotation and multiple-choice question generation pipeline.
Hybrid Token Compression for Vision-Language Models
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
CodeCompressionRepresentation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose HTC-VLM, a hybrid token compression framework that can compress visual images into a single token while preserving semantic and detail information.
🎯 What it does: Designed and implemented a hypergraph-based point cloud completion framework, Hyper-PCN, which achieves complete and detailed 3D reconstruction by modeling high-order associations in incomplete point clouds.
Ziheng Chen (University of Trento), Nicu Sebe (University of Trento)
CodeClassificationImageGraphBiomedical Data
🎯 What it does: Proposed hyperbolic space polynomial logistic regression (BMLR) and fully connected layer (BFC) based on the Busemann function, which can directly operate on the Poincaré ball and Lorentz hyperbolic model, avoiding the parameter redundancy and low batch efficiency of traditional hyperplane MLR.
Hypergraph-State Collaborative Reasoning for Multi-Object Tracking
Zikai Song (Huazhong University of Science and Technology), Xinchao Wang (La Trobe Uniersity)
CodeObject TrackingGraph Neural NetworkVideo
🎯 What it does: Propose a collaborative motion estimation framework based on hypergraphs and state space models (HyperSSM) for motion prediction and association in multi-target tracking.
HyperST: Hierarchical Hyperbolic Learning for Spatial Transcriptomics Prediction
Chen Zhang (National Institute for Data Science in Health and Medicine, Xiamen University), Rongshan Yu (National Institute for Data Science in Health and Medicine, Xiamen University)
CodeRepresentation LearningSupervised Fine-TuningContrastive LearningImageBiomedical Data
🎯 What it does: Propose the HyperST framework, which utilizes multi-level image and gene feature extraction to align in a hyperbolic space for predicting spatial transcriptomics data.
I'm a Map! Interpretable Motion-Attentive Maps: Spatio-Temporally Localizing Concepts in Video Diffusion Transformers
Youngjun Jun (Yonsei University), Seong Jae Hwang (Yonsei University)
CodeSegmentationExplainability and InterpretabilityTransformerDiffusion modelVideo
🎯 What it does: For video diffusion Transformers, a training-free and gradient-free motion attention mapping (IMAP) is proposed, which can perform spatiotemporal localization of any motion concept in videos.
IAG: Input-aware Backdoor Attack on VLM-based Visual Grounding
Junxian Li (Shanghai Jiao Tong University), Di Zhang (Fudan University)
CodeAdversarial AttackConvolutional Neural NetworkVision Language ModelMultimodality
🎯 What it does: This paper proposes a text-guided input-aware reverse gated attack (IAG), which can dynamically generate stealthy triggers in visual-linguistic models (VLM) for visual localization tasks, forcing the model to ignore user queries and forcibly localize any specified target.
🎯 What it does: Proposed a generalizable 3D Gaussian scattering model called IDESplat, which achieves multi-level fusion of depth probability through iterative warp and depth probability boosting unit (DPBU), thereby accurately predicting Gaussian means;
IDperturb: Enhancing Variation in Synthetic Face Generation via Angular Perturbations
Fadi Boutros, Naser Damer (Fraunhofer Institute for Computer Graphics Research IGD)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposed the IDPERTURB method, which enhances intra-class diversity of synthetic faces by performing angle-restricted spherical cap sampling within the identity embedding space, generating more generalizable training data.
IEBGL:An Interpretability-Enhanced Brain Graph Learning Framework with LLM-Instructed Topology and Literature-Augmented Semantics
Yihang Duan (Nanjing Forestry University), Li Zhang (Nanjing Forestry University)
CodeClassificationExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraphBiomedical DataMagnetic Resonance ImagingRetrieval-Augmented Generation
🎯 What it does: Propose an interpretable enhanced brain mapping learning framework, IEBGL, which improves the topology and node features of rs-fMRI brain networks for brain disease diagnosis through Topology Reconstruction via LLM Instructions (LITR) and Semantic Aggregation of Medical Literature (LASA).
IF-Bench: Benchmarking and Enhancing MLLMs for Infrared Images with Generative Visual Prompting
Tao Zhang (MAIS, Institute of Automation), Chunhong Pan (MAIS, Institute of Automation)
CodeLarge Language ModelPrompt EngineeringMixture of ExpertsVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This study first constructs the first high-quality infrared image multimodal understanding benchmark, IF-Bench, containing 499 infrared images and 680 visual question answering (VQA) pairs; subsequently, it systematically evaluates the infrared image understanding capabilities of over 40 open-source and closed-source multimodal large language models (MLLMs); finally, it proposes a training-agnostic generated visual prompting (GenViP) method, which converts infrared images into semantically aligned RGB images using image editing models and inputs them together with the original images, significantly enhancing the performance of various MLLMs on infrared tasks.
IF-Prune: Information-Flow Guided Token Pruning for Efficient Vision-Language Models
Guohao Sun (Snap Inc), Jian Wang (Snap Inc)
CodeComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Proposed an information-flow-based posterior-guided visual token pruning framework called IF-Prune, which estimates the importance of each visual token in a single forward pass using a lightweight information bottleneck fine-tuned small VLM, and applies the pruning results to large VLMs.
Illustrator's Depth: Monocular Layer Index Prediction for Image Decomposition
Nissim Maruani (Inria), Wang Yifan (Adobe Research)
CodeSegmentationDepth EstimationTransformerImage
🎯 What it does: This paper proposes the concept of 'Illustrator's Depth,' utilizing a neural network to predict the layer index for each pixel in a single image, enabling image layering and editing;
🎯 What it does: Propose the Diffusion Preview framework and design ConsistencySolver as a trainable high-order ODE solver to generate high-quality previews with low-step sampling, enabling an interactive image generation workflow.
Imagine Before Concentration: Diffusion-Guided Registers Enhance Partially Relevant Video Retrieval
Jun Li, Bin Chen (Harbin Institute Of Technology)
CodeRetrievalVision Language ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Propose the DreamPRVR framework, which first generates global registration through a text-supervised diffusion model, and then refines video representations using registration-enhanced Gaussian Attention to address partial relevance video retrieval problems.
Imbalanced View Contribution Evaluation and Refinement for Deep Incomplete Multi-View Clustering
Taichun Zhou (National University of Defense Technology), Kunlun He (Chinese PLA General Hospital)
CodeOptimizationExplainability and InterpretabilityRepresentation LearningData-Centric LearningAuto EncoderMultimodality
🎯 What it does: This paper proposes an Imbalanced View Contribution Evaluation and Refinement (ICER) framework specifically designed to address the issue of contribution imbalance caused by missing views in incomplete multi-view clustering.
Improving Calibration in Test-Time Prompt Tuning for Vision-Language Models via Data-Free Flatness-Aware Prompt Pretraining
Hyeonseo Jang (Yonsei University), Kibok Lee (Yonsei University)
CodeOptimizationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose an unsupervised pre-training framework (FPP) that initializes prompts to flatter loss landscape regions before test-time prompt tuning (TPT), aiming to improve the calibration performance of vision-language models.
Improving Diffusion Generalization with Weak-to-Strong Segmented Guidance
Liangyu Yuan (Westlake University), Chi Zhang (Westlake University)
CodeGenerationDiffusion modelImage
🎯 What it does: The paper proposes a Segmented Guidance (SGG) mechanism that combines conditional dependent and conditional independent guidance to enhance the generalization performance of diffusion models and migrate the weak-to-strong guidance principle into the training phase;
🎯 What it does: This paper proposes a sparse autoencoder based on Transformer (Sparsemax SAE), which utilizes the sparsemax attention mechanism to automatically determine the required number of concepts in each feature vector during the autoencoding process, thereby achieving higher quality concept learning and reconstruction.
Improving Vision-language Models with Perception-centric Process Reward Models
Yingqian Min (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeRetrievalReinforcement Learning from Human FeedbackReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose Perceval, a perception-centric process reward model, to achieve word-by-word error localization and correction in the reasoning chain of visual language models;
🎯 What it does: Propose the IM-S3 framework, which utilizes diffusion inversion matching for fine-tuning and achieves untrained class discrimination through subgroup center similarity during the sampling stage, thereby enhancing the distribution coverage and discriminability of diffusion models in dataset distillation.
🎯 What it does: This paper proposes Pixio, a self-supervised visual pre-training method based on Masked AutoEncoder, which utilizes the self-supervised screening strategy MetaCLIP-S on 2B web-crawled images. It further improves the decoder depth, occlusion block size, and the number of class tokens in MAE to learn richer spatial structural representations.
InfiniBench: Infinite Benchmarking for Visual Spatial Reasoning with Customizable Scene Complexity
Haoming Wang (University of Pittsburgh), Wei Gao (University of Pittsburgh)
CodeGenerationData SynthesisOptimizationLarge Language ModelAgentic AIVision Language ModelVideoMultimodalityMeshBenchmark
🎯 What it does: By integrating LLM agents, clustering-based layout optimization, and task-aware camera trajectory generation methods, InfiniBench can automatically generate infinitely customizable, physically feasible, and highly complex 3D scenes, rendering them into multi-frame videos suitable for vision-language models (VLMs).
🎯 What it does: This paper proposes InfiniDepth, which converts monocular depth estimation into depth prediction in a continuous space using neural implicit fields, thereby achieving depth estimation at arbitrary resolutions with fine-grained details.
🎯 What it does: Proposed a multi-modal interactive learning framework DMIL based on information theory, which can adaptively decompose and reinforce redundant, unique, and synergistic information;
Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings
Yunxiang Peng (University of Delaware), Xi Peng (University of Delaware)
CodeClassificationExplainability and InterpretabilityTransformerImage
🎯 What it does: This paper proposes to evaluate model generalization ability by leveraging the internal circuit structure of Vision Transformers, and introduces two novel metrics: DDB for model pre-selection and CSS for post-deployment performance monitoring.
🎯 What it does: Proposed the INSIGHT Bench, focusing on the 'on-site guidance normalization' task involving text/symbol guidance directly embedded on objects for robots, and constructed the corresponding simulation environment and data generation framework.
🎯 What it does: Proposed a complete three-module pipeline OA-VAT to address instance-level distinction and occlusion recovery in visual active tracking.
Instruction-Guided Lesion Segmentation for Chest X-rays with Automatically Generated Large-Scale Dataset
Geon Choi (KAIST), Edward Choi (KAIST)
CodeSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBiomedical Data
🎯 What it does: Constructed the MIMIC-ILS large-scale instruction-oriented chest X-ray lesion segmentation dataset, and trained the ROSALIA model on it to achieve precise segmentation and text generation based on natural language instructions.
InternVideo-Next: Towards World-Understanding Video Models
Chenting Wang (Shanghai Jiao Tong University), Limin Wang (Shanghai Jiao Tong University)
CodeRecognitionRetrievalVision Language ModelDiffusion modelContrastive LearningWorld ModelVideoMultimodalityBenchmark
🎯 What it does: Propose InternVideo-Next, a two-stage self-supervised video pre-training framework that combines pixel reconstruction with latent prediction, leveraging a semantic-guided diffusion decoder and frozen teacher's latent prediction to learn world knowledge.
Intervention-Aware Multiscale Representation Learning from Imaging Phenomics and Perturbation Transcriptomics
Jiayuan Chen, Ping Zhang (The Ohio State University)
CodeKnowledge DistillationRepresentation LearningDrug DiscoveryTransformerAuto EncoderContrastive LearningImageBiomedical Data
🎯 What it does: Propose an intervention-aware distillation framework (TIDE) that leverages transcriptomic information to guide microscopic image feature learning, achieving mechanistic representation of drug interventions.
CodeAnomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper systematically evaluates the effectiveness and interpretability of various self-supervised features in audio-visual deepfake detection, constructing a multi-dimensional evaluation framework including linear probing, anomaly detection, spatiotemporal explanation, and complementary analysis.
CodeTransformerLarge Language ModelReinforcement LearningVision Language ModelAuto EncoderWorld ModelOptical FlowImageVideoTextMultimodalityBenchmarkPhysics Related
🎯 What it does: Explored a general method for acquiring physical and causal reasoning through interactive experience, and proposed the IPR framework.
🎯 What it does: Propose Iris, a two-stage Priors-to-Geometry structure based on deterministic diffusion models, which first uses real images to guide low-frequency layout and then refines high-frequency geometry with synthetic data, achieving unsupervised monocular depth estimation.
🎯 What it does: Propose a bin-free dataset quantization framework called BGFDQ, which utilizes KNN neighbor identification and neighbor-aware core subset selection, combined with adaptive semantic transfer patch dropping, to achieve efficient and scalable image dataset compression.
🎯 What it does: The paper analyzes the intra-modal alignment insufficiency caused by the CLIP projector, proposing IsoCLIP which enhances intra-modal similarity through spectral decomposition and projection into an isotropic subspace in a training-free manner.
🎯 What it does: Propose iSplat, an iterative feed-forward 3D Gaussian splatting framework based on GRU loops, which simultaneously optimizes geometry and appearance through multi-step self-correction.
It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models
Anne Harrington (UC Berkeley), Alexei A. Efros (UC Berkeley)
CodeGenerationDiffusion modelImage
🎯 What it does: This paper proposes a method that directly optimizes the initial noise via gradient optimization on a pre-trained diffusion model to eliminate mode collapse and enhance the diversity of generated images.
Jailbreaking Vision-Language Models via Dissonance-Guided Suffix Optimization and Image-Phrase Injection
Jiacheng Pi (University of Science and Technology of China), Wenjie Ruan (University of Science and Technology of China)
CodeAdversarial AttackPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Propose an attack framework that simultaneously targets text suffixes and image phrases to achieve efficient jailbreak in visual language models.
Yifan Yang (OPPO AI Center), Dan Meng (OPPO AI Center)
CodeTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImage
🎯 What it does: Proposed a multi-attribute comprehensive aesthetic evaluation method called JoPPO based on vision-language models. It first injects compositional priors through SFT, then achieves consistent optimization of attributes and overall scores using joint conditional probability reinforcement learning, capable of outputting attribute scores, overall scores, and natural language explanations.
JUMP-Hand: Learning Joint-wise Uncertainty to Gate Mixture of View Experts for Multi-View 3D Hand Reconstruction
Haohong Kuang (Huazhong University of Science and Technology), Joey Tianyi Zhou (Huazhong University of Science and Technology)
CodePose EstimationConvolutional Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: Proposes a multi-view 3D hand reconstruction method called JUMP-Hand, which utilizes a joint uncertainty gating mechanism to integrate expert information from different views.
KaLOS finds Consensus: A Meta-Algorithm for Evaluating Inter-Annotator Agreement in Complex Vision Tasks
David Tschirschwitz (Bauhaus-Universität Weimar), Volker Rodehorst (Bauhaus-Universität Weimar)
CodeExplainability and InterpretabilityData-Centric LearningImageBiomedical Data
🎯 What it does: This paper proposes the K LOS Meta-Algorithm for unified, interpretable evaluation of annotation consistency across multiple annotators in complex visual tasks (e.g., object detection, instance segmentation).
🎯 What it does: Propose the KLIP metric based on the prior and posterior KL divergence of diffusion models for detecting local distribution shifts in inverse problems.
🎯 What it does: Proposes L3DR, a LiDAR generation framework that combines 2D range view diffusion with a 3D residual regression network to address defects such as depth leakage and wavefronts, enhancing the geometric realism of point clouds.
LacTokGen: Latent Consistency Tokenizer for 1024-pixel Image Generation by 256 Tokens
Qingsong Xie (OPPO AI Center), Haonan Lu (OPPO AI Center)
CodeGenerationTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the Latent Consistency Tokenizer (LacTok) and its autoregressive generation model LacTokGen, which can efficiently reconstruct and generate 1024×1024 pixel images using only 256 discrete tokens;
LangRef3DGS: Natural Language-Guided 3D Referential Segmentation from Partial Observations via 3D Gaussian Splatting
Xulun Ye (Ningbo University), Kun Zhou (Shenzhen University)
CodeSegmentationVision Language ModelContrastive LearningGaussian SplattingPoint Cloud
🎯 What it does: Construct a real-time language-guided 3D segmentation framework based on 3D Gaussian Splatting, which can perform semantic segmentation and discover new categories under partial RGB-D observations.
CodeClassificationObject DetectionSegmentationDomain AdaptationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a cross-modal visual feature enhancement framework that leverages image-level and domain-level language descriptions to guide pre-trained visual models in cross-domain few-shot learning adaptation.
🎯 What it does: Propose a pseudo-supervised framework called LAOF that leverages optical flow constraints to learn robust latent action representations from unlabeled videos.
🎯 What it does: Propose a robust anchor-based ensemble clustering method called RANGE. First, map the base clustering results to a bipartite graph, enhance the reliability of the bipartite graph using high-order fuzzy enhancement, then decompose the similarity matrix in the anchor space to obtain clean clustering structures and residual anomalous structures. Finally, perform k-means on the clean part to achieve consensus clustering; the residual part can be directly used for anomaly detection.
LaSM: Layer-wise Scaling Mechanism for Defending Pop-up Attack on GUI Agents
Zihe Yan (Shanghai Jiao Tong University), Gongshen Liu (Shanghai Jiao Tong University)
CodeAdversarial AttackTransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose a post-training hierarchical scaling mechanism called LaSM to enhance the robustness of GUI agents against pop-up injection attacks.
Yawen Yang (Hefei University of Technology), Meng Wang (Hefei University of Technology)
CodeAnomaly DetectionTransformerImage
🎯 What it does: Proposes a synthetic image detection method based on the consistency differences of mid-level features in visual Transformers, called Layer Transition Discrepancy (LTD).
🎯 What it does: Proposed the LEAD expert and dataset, aligning through visibility, uncertainty, and intention, improving target point conditioning, training TransFuser v6, achieving optimal closed-loop driving performance on CARLA benchmarks.
LEADER: Learning Reliable Local-to-Global Correspondences for LiDAR Relocalization
Jianshi Wu (Fujian Key Laboratory of Urban Intelligent Sensing and Computing), Cheng Wang (Fujian Key Laboratory of Urban Intelligent Sensing and Computing)
CodePose EstimationAutonomous DrivingConvolutional Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposes a LiDAR Absolute Pose Re-localization framework named LEADER, which directly outputs global 6-DoF poses by performing scene coordinate regression on single-frame LiDAR point clouds.
🎯 What it does: Propose an active learning framework LH3D based on learnability for monocular roadside 3D detection, selecting images that are both easy to annotate and informative under a limited annotation budget.
🎯 What it does: This paper proposes a test-time constrained optimization framework that leverages camera pose, intrinsic parameters, and depth priors through loss constraints rather than network inputs, enhancing the performance of multi-view Transformers in 3D reconstruction and camera pose estimation tasks.
🎯 What it does: Proposes a method that utilizes adaptive SE(3) B-spline motion bases to explicitly model continuous pose deformations of dynamic Gaussian clouds, achieving high-quality dynamic Gaussian splatting from monocular video;
🎯 What it does: This paper proposes a method to predict richer pre-trained weights (KNOW prediction) by utilizing structured progressive forgetting and its inverse process, and designs a lightweight meta-learning hypernetwork called KNOWN.
CodeClassificationData SynthesisExplainability and InterpretabilityDiffusion modelImageVideo
🎯 What it does: Utilize the automatically obtained provenance information during the synthetic data generation process as auxiliary supervision, employing input gradient guidance (provenance loss) to suppress the model's dependence on non-target regions, thereby enhancing discriminative ability for the target region.
🎯 What it does: Construct a two-stage framework: first, use a VAE to compress sparse tracker trajectories into a 64× temporal compressed dense motion space, then train a conditional flow matching model in this space to achieve long-term motion generation conditioned on text or spatial 'stamps'.
🎯 What it does: This study proposes an adaptive collaborative perception adversarial attack framework called MVIG, which leverages mutual perspective information graphs (MVIG) to capture vulnerabilities leaked by defensive systems, enabling dynamic attacks on collaborative perception systems.
Learning to Focus and Precise Cropping:A Reinforcement Learning Framework with Information Gaps and Grounding Loss for MLLMs
Xuanpu Zhao (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
CodeConvolutional Neural NetworkLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Proposes a two-stage reinforcement learning framework based on information gap and localization loss, enabling multi-modal large language models to actively focus and precisely crop regions of interest in high-resolution visual question answering.
Learning to Reason in 4D: Dynamic Spatial Understanding for Vision Language Models
Shengchao Zhou (University of Hong Kong), Xiaojuan Qi (University of Hong Kong)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoPoint CloudBenchmark
🎯 What it does: Built a dynamic spatial reasoning (DSR) dataset and evaluation framework, and improved vision-language models (VLM) to enhance reasoning capabilities for 4D dynamic scenes.
Learning Transferable Temporal Primitives for Video Reasoning via Synthetic Videos
Songtao Jiang (Zhejiang University), Zuozhu Liu (Zhejiang University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodalityChain-of-Thought
🎯 What it does: Proposes the SynRL framework, which teaches visual language models (VLM) fundamental temporal primitives (direction, speed, state tracking, etc.) using procedurally generated synthetic videos and precise frame-level annotations, and achieves post-training enhancement through chain-of-thought (CoT) and reinforcement learning (GRPO).
🎯 What it does: Proposed a large-scale publicly available surgical video dataset named LEMON (over 4K segments, 938 hours, 85M frames) and trained a new surgical foundation model, LemonFM, based on this dataset.
LF-BVN: Blind-View Network for Self-Supervised Light Field Denoising
Longzhao Guo (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
CodeRestorationDepth EstimationImage
🎯 What it does: This paper proposes a self-supervised light field image denoising framework named LF-BVN (Blind-View Network), which achieves training without clean images by leveraging multi-view consistency in light fields.
🎯 What it does: Design a lightweight dense prediction reader, LiDeRe, on large frozen visual backbones (e.g., DINOv3), leveraging learnable interpolation and content-guided attention to achieve high-resolution, fine-grained pixel-level predictions.
Linear Fundamental Matrix Estimation from 7 or 5 Points
Taci Ata Kucukpinar (University of Missouri), Kannappan Palaniappan (University of Missouri)
CodePose EstimationImage
🎯 What it does: This paper proposes a linear minimal solver for the V-Umlaut configuration, which estimates the fundamental matrix in two-view cameras using five real corresponding points and two virtual midpoints.
Linguistic Priors for Visual Decoupling: Towards Symmetric Vision-Brain Alignment
Dongjun Liu (Hangzhou Dianzi University), Wanzeng Kong (Hangzhou Dianzi University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: Proposes a vision disentanglement method guided by language priors, using text priors to guide the separation of foreground objects and background in visual images, thereby achieving better alignment between visual and brain signals.
Linking Modality Isolation in Heterogeneous Collaborative Perception
Changxing Liu (Nanjing University of Science and Technology), Siheng Chen (Nanjing University of Science and Technology)
CodeAutonomous DrivingMultimodality
🎯 What it does: Propose the CodeAlign framework to address the modality isolation problem in heterogeneous collaborative perception caused by the lack of co-occurrence data across different modalities, by constructing modality-specific discrete feature spaces through codebooks and achieving cross-modal alignment via feature-code-feature (FCF) translation.
Linking Perception, Confidence and Accuracy in MLLMs
Yuetian Du (Zhejiang University), Qiang Zhu (Zhejiang University)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose Confidence-Driven Reinforcement Learning (CDRL) to calibrate model confidence by incorporating original-noise image pairs and confidence rewards into multi-modal large language models. Based on this confidence, develop the Confidence-Aware Test-Time Scaling (CA-TTS) framework, integrating three modules: Self-Consistency, Self-Reflection, and Self-Check, ultimately significantly enhancing visual reasoning performance.
🎯 What it does: Designed and implemented a lightweight point cloud Transformer called LitePT, integrating convolution and attention modules to provide an efficient and scalable backbone network for point cloud analysis.
LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding
Xuanzhao Dong (Arizona State University), Yalin Wang (Arizona State University)
CodeRecognitionTransformerSupervised Fine-TuningVision Language ModelDiffusion modelBiomedical Data
🎯 What it does: By migrating the large-scale language diffusion model LLaDA to the medical imaging field and combining it with visual instruction tuning, the LLaDA-MedV model was constructed to achieve visual-language understanding in medical imaging.
LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning
Zebin You (Renmin University of China), Chongxuan Li (Renmin University of China)
CodeGenerationLarge Language ModelVision Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Proposed a fully diffusion-based multimodal large language model, LLaDA-V, integrating visual instruction tuning with masked diffusion models.
🎯 What it does: Propose the DiffRender-VLA framework, which utilizes differentiable rendering to convert 3D spatial information into 2D images directly usable by vision-language action (VLA) models, achieving end-to-end gradient flow in robotic manipulation;
🎯 What it does: Propose the LoFA framework, which directly generates complete LoRA weights from user prompts through a two-stage hypernetwork, enabling second-level fast adaptation for visual generation models.
🎯 What it does: Proposes a general backdoor defense method called Logit-Margin Repulsion (LMR), which significantly suppresses the logits of the backdoor class by using selective cross-entropy and logit margin constraints on clean samples, and further eliminates backdoor channels during the pruning stage, ultimately achieving efficient removal of traditional backdoors and conditional backdoors (quantization/trimming-induced).
LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding
Jihao Qiu (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose LongVideo-R1, design an active reasoning agent based on a multimodal large language model (LLM) that can achieve long video question answering under low computational budget.
Love Me, Love My Label: Rethinking the Role of Labels in Prompt Retrieval for Visual In-Context Learning
Tianci Luo (Tsinghua University), Chun Yuan (Tsinghua University)
CodeObject DetectionSegmentationGenerationRetrievalTransformerPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningImage
🎯 What it does: Designed the LaPR framework, incorporating label information into prompt retrieval in Vision-and-Language Context Learning (VICL). A mixture-of-experts network and a query-adaptive router are employed to generate label-aware prompt and query embeddings, improving the quality of prompt selection.
🎯 What it does: Propose a low-rank residual diffusion model (LRDM), which projects residuals into a low-rank subspace during near-field image restoration tasks to achieve more efficient and structured recovery.
LRDUN: A Low-Rank Deep Unfolding Network for Efficient Spectral Compressive Imaging
He Huang (Wuhan University), Wei He (Wuhan University)
CodeRestorationCompressionConvolutional Neural NetworkImagePhysics Related
🎯 What it does: Propose the low-rank deep unfolding network LRDUN to reconstruct compressed hyperspectral images and significantly improve reconstruction quality.
🎯 What it does: Proposes a block reconstruction framework called LS-ViT based on least squares Hessian estimation for low-bit post-training quantization of Vision Transformers;