🎯 What it does: Proposed a public pediatric inner ear CT dataset (CIED) and designed a pluggable Domain Knowledge Guided Tuning (DKGT) framework to enhance structural anomaly detection, post-operative hearing prediction, and anatomical segmentation of inner ear CT in data-scarce scenarios.
🎯 What it does: To deploy large video generation models on mobile devices, we propose Turbo-VAED, a lightweight VAE decoder that reduces parameter redundancy using 3D depthwise separable convolutions, achieves mobile-friendly upsampling through decoupled 3D pixel shuffling, and rapidly transfers the teacher model via decoder-only distillation.
🎯 What it does: Propose a time-weighted contrastive reward learning framework (TW-CRL) that utilizes successful and failed demonstrations to learn dense reward functions from sparse rewards in periodic tasks.
TweezeEdit: Consistent and Efficient Image Editing with Path Regularization
Jianda Mao (Hong Kong University of Science and Technology), Kani Chen (Hong Kong University of Science and Technology)
CodeGenerationDiffusion modelImage
🎯 What it does: Proposes TweezeEdit, a parameter-free and inversion-free image editing framework based on consistency models, which can achieve target prompt alignment while preserving the semantics of the source image;
Two Constraint Compilation Methods for Lifted Planning
Periklis Mantenoglou (rebro University), Pedro Zuidberg Dos Martires (rebro University)
CodeComputational EfficiencyBenchmark
🎯 What it does: Propose two lifted constraint compilation methods, LiftedTCORE and LCC, to handle qualitative state trajectory constraints in PDDL, avoiding global variable substitution across the problem domain.
🎯 What it does: Proposed the U2B framework to eliminate bias caused by graph scale imbalance in graph classification tasks, employing a two-stage distillation-refinement approach with node-level and graph-level transformers.
🎯 What it does: This work proposes the UCPO framework, which leverages preference optimization to directly embed constraints into neural combinatorial optimizers, achieving complex constraint solving without altering the network architecture.
UDA: Unsupervised Debiasing Alignment for Pair-wise LLM-as-a-Judge
Yang Zhang (ZhipuAI), Jie Tang (Westlake University)
CodeTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper proposes an unsupervised debiasing alignment (UDA) framework that dynamically corrects the Elo scores of LLM-as-a-judge to reduce evaluation discrepancies caused by self-preference.
UI-R1: Enhancing Efficient Action Prediction of GUI Agents by Reinforcement Learning
Zhengxi Lu (Zhejiang University), Hongsheng Li (Zhejiang University)
CodeComputational EfficiencyLarge Language ModelReinforcement LearningVision-Language-Action ModelMultimodality
🎯 What it does: Propose the UI-R1 framework, leveraging rule-based reinforcement learning to enhance the reasoning and execution capabilities of multimodal large language models in GUI action prediction tasks, and design a two-phase efficient inference scheme called UI-R1-E.
🎯 What it does: Propose UMNet, a hyperspectral image fusion network based on a dual U-Net backbone, spatial-spectral recursive fusion unit (SRFU), and non-negative matrix factorization memory interaction unit (SMIU), achieving high spatial and spectral fidelity through two-stage uncertainty-guided loss.
Unbiased Rectification for Sequential Recommender Systems Under Fake Orders
Qiyu Qin (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
CodeRecommendation SystemLarge Language ModelContrastive LearningSequential
🎯 What it does: This paper proposes the DITaR framework for unbiased correction of fake orders embedded in real user sequences in sequential recommendation systems.
Uncertainty-Based Methods for Automated Process Reward Data Construction and Output Aggregation in Mathematical Reasoning
Jiuzhou Han (Monash University), Ehsan Shareghi (Monash University)
CodeData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose an uncertainty-driven process reward model (PRM) data construction and annotation framework, and design two uncertainty-aware aggregation methods (Hybrid Majority Reward Vote and Weighted Reward-Frequency Vote) that integrate Majority Vote with PRM for mathematical reasoning.
Uncertainty-Guided View-Strength-Aware Feature Utilization for Multi-View Classification
Li Lv (Shanxi University), Xinyan Liang (Taiyuan University of Science and Technology)
CodeClassification
🎯 What it does: Designed and implemented the UVF framework, which guides view-intensity-aware feature utilization and fusion through uncertainty estimation, achieving multi-view classification.
Uncovering and Mitigating Destructive Multi-Embedding Attacks in Deepfake Proactive Forensics
Lixin Jia (Xinjiang University), Gaobo Yang (Hefei University of Technology)
CodeAnomaly DetectionSafty and PrivacyAdversarial AttackImage
🎯 What it does: This paper first reveals the vulnerability of multi-embedding attacks (MEA) in active deepfake forensics and proposes an adversarial interference simulation (AIS) training paradigm to enhance watermark robustness.
Uncovering and Mitigating Transient Blindness in Multimodal Model Editing
XiaoQi Han, Jeff Z. Pan (Shanxi University)
CodeExplainability and InterpretabilityAdversarial AttackData-Centric LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes a dynamic evaluation framework called De-VQA for multi-modal model editing (MMED), utilizing three locality dimensions (random image, no image, and consistent image, totaling seven categories of data) to systematically detect and quantify the model's dependency on visual information after editing, identifying and mitigating the 'Transient Blindness' phenomenon.
Uncovering Hidden Degeneration: A Physics-Guided Bidirectional Inference Framework for Industrial Time Series Prediction
Xingwang Li (Southwest Jiaotong University), Qiang Duan (Pennsylvania State University)
CodeAnomaly DetectionTransformerTime SeriesPhysics Related
🎯 What it does: Propose a physics-guided bidirectional reasoning framework that combines continuous damage mechanics simulation with maximum entropy inference to predict hidden degradation processes in industrial time series.
Understanding and Enhancing Differentiable Architecture Search from Information Bottleneck Perspective
Haidong Kang, Bo Yi (Northeastern University)
CodeNeural Architecture SearchImageBenchmark
🎯 What it does: Proposed a Batch Entropy-decay Regularization (BER) method based on the information bottleneck theory to help differential architecture search (DAS) avoid performance collapse.
Understanding the Impact of Proportionality in Approval-Based Multiwinner Elections
Niclas Boehmer (Hasso Plattner Institute, University of Potsdam), Jannik Peters (National University of Singapore)
CodeTabular
🎯 What it does: This paper studies the practical impact of the proportionality axiom in approval-based multi-choice committee elections, combining algorithmic analysis with experimental evaluation to assess its constraints on selectable committees and candidate importance.
🎯 What it does: Propose a two-stage unsupervised cross-view geolocation framework named UniABG, first eliminating perspective differences through perspective-aware adversarial bridging, then improving cross-view correspondences via heterogeneous graph filtering correction.
🎯 What it does: To address the problem of poor 3D instance segmentation caused by inconsistent 2D instance segmentation labels across multiple views, this paper proposes a unified framework called UniC-Lift. It introduces learnable vector embeddings for each Gaussian atom in 3D Gaussian Splatting (3DGS), achieves consistent segmentation through alignment, triplet loss, and spatial smoothing regularization, and finally obtains 3D instance labels directly via a simple 'Embedding-to-Label' decoding process.
🎯 What it does: Proposed a Unified Interactive Consistent Learning (UICL) framework to address object detection under single-source domain generalization.
🎯 What it does: Proposed a unified minimax optimization framework called UPO for jointly estimating and correcting propensity scores in recommendation systems to eliminate bias caused by MNAR.
Unified Mixture-of-Experts Framework for Joint Cardiac and Vascular Ultrasound Analysis and Report Generation
Bin Pu (Hunan University), Kenli Li (Hunan University)
CodeGenerationExplainability and InterpretabilityLarge Language ModelMixture of ExpertsImageTextMultimodalityBiomedical DataUltrasound
🎯 What it does: This paper proposes the ECV framework, achieving joint parameter measurement and report generation for cardiac and vascular ultrasound images.
🎯 What it does: Propose a zero-shot inference lifelong person re-identification method named Unified Representation Causal Prompt Distillation (URCPD), combining feature decoupled style transfer and causal prompt distillation, which enhances cross-domain generalization while significantly alleviating catastrophic forgetting.
🎯 What it does: Propose a unified multi-view subspace clustering method called UVELRS, which first extracts consistent cross-view representations from self-representation matrices of each view, and then constructs a tensor and jointly optimizes it via tensor total variation Schatten p norm.
🎯 What it does: Proposes UniHR, a unified hierarchical knowledge graph representation learning framework that can convert three complex facts—hyper-relational, temporal, and nested—into triplet forms. It learns semantic and structural information of facts and entities through internal and external message passing, ultimately completing link prediction for multi-type knowledge graphs.
UniMM-V2X: MoE-Enhanced Multi-Level Fusion for End-to-End Cooperative Autonomous Driving
Ziyi Song (Tsinghua University), Zhisheng Niu (Tsinghua University)
CodeAutonomous DrivingTransformerMixture of ExpertsMultimodality
🎯 What it does: Propose the UniMM-V2X framework to achieve end-to-end autonomous driving with multi-agent collaboration at perception and prediction levels.
UniMo: Unified Motion Generation and Understanding with Chain of Thought
Guocun Wang (Tsinghua University), Xiaoguang Han (Chinese University of Hong Kong)
CodeGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelAuto EncoderVideoTextMultimodalityChain-of-Thought
🎯 What it does: Construct a unified model named UniMo that integrates text-to-motion (T2M) and motion-to-text (M2T) tasks, leveraging large language models combined with chain-of-thought reasoning and reinforcement learning to achieve complementary improvements in generation and understanding.
UniSketch: A Unified Framework for Parametric Sketch Generation and Constraint Prediction
Jing Lin, Rubin Fan (Wuhan University)
CodeGenerationTransformerImageSequential
🎯 What it does: Proposes a unified framework capable of generating parametric sketches from images, predicting constraints, and performing unconditional sketch synthesis.
🎯 What it does: Propose a general adversarial purification framework UDAP for Stable Diffusion, utilizing DDIM inversion reconstruction error for adversarial noise elimination, and achieving efficient purification through dynamic epoch adjustment.
CodeClassificationAnomaly DetectionBiomedical Data
🎯 What it does: Proposed a multi-view debiasing framework BF-EML based on evidence inference for achieving global seizure detection without retraining.
🎯 What it does: Propose the FoundationSSC framework, achieving high-quality 3D scene completion by dual decoupling at the source and channel layers, leveraging foundation models to provide separated semantic and geometric priors.
Unleashing the Potential of Large Language Models for Text-to-Image Generation Through Autoregressive Representation Alignment
Xing Xie (Chinese Academy of Sciences), Liangqiong Qu (University of Hong Kong)
CodeGenerationTransformerLarge Language ModelVision Language ModelImageBiomedical Data
🎯 What it does: Propose a training framework called ARRA that does not require modifying the architecture of large language models (LLMs), enabling LLMs to generate globally consistent images while maintaining the autoregressive paradigm.
Unleashing the Power of Image-Tabular Self-Supervised Learning via Breaking Cross-Tabular Barriers
Yibing Fu (National University of Singapore), Yueming Jin (National University of Singapore)
CodeConvolutional Neural NetworkTransformerMixture of ExpertsContrastive LearningImageMultimodalityTabularBiomedical DataAlzheimer's Disease
🎯 What it does: Proposed the CITab framework to achieve cross-table image-table self-supervised learning, enhancing the transferability of medical multimodal models.
CodeAutonomous DrivingAdversarial AttackTransformerLarge Language ModelPrompt EngineeringImage
🎯 What it does: Propose the UYE framework to generate physical camouflage patterns that can both mislead deep neural networks and remain stealthy to human perception.
🎯 What it does: This paper investigates the implicit regularization behavior of Sharpness-Aware Minimization (SAM) in multi-core tensor models and proposes an explicit regularization method called Deviation-Aware Scaling (DAS) based on core norm deviation.
🎯 What it does: Developed a general-purpose multi-agent reinforcement learning (MARL) platform called Unreal-MAP based on Unreal Engine, and built an experimental framework that can seamlessly integrate with various algorithm frameworks (e.g., HMAP, PyMARL, HARL). It provides a five-layer modular architecture allowing customizable tasks, maps, teams, entities, and events.
🎯 What it does: Propose an unsupervised motion-compensated implicit neural representation (MoCo-INR) framework for reconstructing high-quality cardiac MRI images from highly undersampled k-t space data.
🎯 What it does: Riner proposed an unsupervised multi-parameter inverse solution method that directly estimates the detector's non-ideal response from raw projection data and reconstructs 3D CBCT images without ring artifacts.
Unveiling the Attribute Misbinding Threat in Identity-Preserving Models
Junming Fu (Sun Yat-sen University), Jianquan Yang (Sun Yat-sen University)
CodeGenerationSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes the Attribute Misbinding Attack, which exploits attention bias in text-to-image diffusion models to cause the model to incorrectly bind sensitive attributes to target identities, thereby generating NSFW images.
Uplift Modeling with Delayed Feedback: Identifiability and Algorithms
Chunyuan Zheng (Peking University), Zhouchen Lin (Peking University)
CodeTabular
🎯 What it does: Studies methods for estimating uplift modeling under delayed feedback scenarios, proposing a joint framework that simultaneously learns latent response times and potential outcomes.
UQ-Bench: A Benchmark for Evaluating Multimodal LLMs on Underwater Image Quality Assessment
Jingchao Cao (Ocean University of China), Yutao Liu (Ocean University of China)
CodeTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper designs the UQ-Bench benchmark to systematically evaluate the capability of multimodal large language models in low-level visual perception and quality assessment of underwater images.
🎯 What it does: This work proposes UQ-ViT, a unified quantization framework designed to handle extreme activation distributions in Vision Transformers, achieving hardware-friendly and high-precision low-bit quantization.
URaG: Unified Retrieval and Generation in Multimodal LLMs for Efficient Long Document Understanding
Yongxin Shi (South China University of Technology), Lianwen Jin (South China University of Technology)
CodeGenerationRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a unified retrieval and generation framework, URaG, which realizes document retrieval and answer generation within multimodal large language models (LLMs), addressing issues of information interference and computational costs in long document understanding.
UrbanNav: Learning Language-Guided Embodied Urban Navigation from Web-Scale Human Trajectories
Yanghong Mei (Institute of Automation Chinese Academy of Sciences), Jing Liu (Institute of Automation Chinese Academy of Sciences)
CodeAutonomous DrivingRobotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelSimultaneous Localization and MappingMultimodality
🎯 What it does: Built an extensible framework called UrbanNav for unsupervised language-instructed urban navigation training using robot-compatible trajectories and natural language instructions automatically extracted from web-scale human walking videos.
UrbanPG: An Efficient Framework with Personalized Context and General Backbone Interaction for Urban Spatio-Temporal Learning
Aoyu Liu (Tongji University), Yaying Zhang (Tongji University)
CodeOptimizationComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringGraphTime Series
🎯 What it does: Propose the UrbanPG framework, decomposing urban spatiotemporal prediction into a general backbone and personalized context prompts to enhance performance in large-scale, few-shot, and continual learning scenarios.
URPO: A Unified Reward & Policy Optimization Framework for Large Language Models
Songshuo Lu (Moore Threads AI), Yaohua Tang (Moore Threads AI)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study proposes a unified reward and policy optimization framework, URPO, which employs a single large model to simultaneously assume the roles of generator (player) and evaluator (referee). It integrates rule-verified reasoning, open-ended instructions, and preference data within a single GRPO training loop, achieving an integrated reinforcement learning alignment process.
Using Certifying Constraint Solvers for Generating Step-wise Explanations
Ignace Bleukx (KU Leuven), Tias Guns (KU Leuven)
CodeOptimizationExplainability and InterpretabilityComputational Efficiency
🎯 What it does: Proposed a method to rapidly construct step-by-step explanations using proofs generated by evidence constraint solvers, significantly reducing explanation generation time;
USPR: Learning a Unified Solver for Profiled Routing
Chuanbo Hua (KAIST), Jinkyoo Park (KAIST)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningTabularBenchmark
🎯 What it does: Proposed a unified learning-based solver, USPR, for handling the Profiled Vehicle Routing Problem (PVRP) that includes vehicle-customer preferences and constraints;
UVLM: Benchmarking Video Language Model for Underwater World Understanding
Xizhe Xue (Northwestern Polytechnical University), Rong Xiao (Intellifusion Inc)
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Built the first underwater multimodal video-language benchmark UVLM, covering 419 marine species, 0.9 million frames, 20 tasks, using human-AI collaborative annotation and GPT-4o-assisted text generation.
🎯 What it does: V-Pruner first evaluates token importance using Fisher information, then applies Proximal Policy Optimization (PPO) reinforcement learning to perform globally sequential token pruning on Vision Transformers, significantly reducing FLOPs, inference latency, while maintaining high accuracy.
🎯 What it does: This paper proposes a GNSS-free vehicle-to-vehicle collaborative perception framework based on LiDAR localization, using PGC to estimate pose and confidence, and then using PASTAT to perform spatiotemporal adaptive alignment of features, achieving multi-vehicle collaborative detection;
VaccineRAG: Boosting Multimodal Large Language Models’ Immunity to Harmful RAG Samples
Qixin Sun (Beihang University), Linjiang Huang (QIYUAN LAB)
CodeRetrievalLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a multi-modal RAG dataset named VaccineRAG containing chain-of-thought annotations, and trained models to perform self-filtering under retrieval sample diversity.
VAEVQ: Enhancing Discrete Visual Tokenization Through Variational Modeling
Sicheng Yang (Houmo AI), Dawei Yang (Houmo AI)
CodeGenerationTransformerDiffusion modelAuto EncoderImageBiomedical Data
🎯 What it does: Propose the VAEVQ framework, which applies VAE to vector quantization and incorporates RCS and DCR to enhance codebook utilization and generation quality.
🎯 What it does: This paper proposes a polynomial-time algorithm for computing the variance of weighted model counting (WMC) results under the structured d-DNNF (st-d-DNNF) representation, and proves that the problem is NP-hard under looser representations (e.g., d-DNNF, FBDD). Subsequently, the algorithm is applied to Bayesian networks with constant tree width, enabling polynomial-time computation of the variance of marginal probabilities, with experiments demonstrating its efficiency on real-world networks.
Variance Reduction via Resampling and Experience Replay
Jiale Han (UCLA), Yuhua Zhu (UCLA)
CodeReinforcement LearningTabular
🎯 What it does: This paper models experience replay as resampling U- and V-statistics, providing a theoretical framework and proving its variance reduction and stability improvement in policy evaluation (LSTD, PDE-based) and supervised learning (kernel ridge regression).
🎯 What it does: This paper proposes the Variation Ratio, a novel attribute to measure the robustness of loss functions, and defines the Variation-Bounded Loss (VBL) based on this. Through theoretical analysis of the variation ratio, its advantages in noise tolerance are demonstrated. Multiple classical losses (CE, EL, SL, etc.) are rewritten in variation-bounded forms. Subsequent comparisons with existing mainstream robust losses and semi-supervised learning methods on synthetic and real-world noisy datasets show that VBL ranks among the top three or even achieves the best performance in various noise scenarios, while maintaining strong fitting capabilities on noise-free data.
🎯 What it does: Proposed the VBF++ framework, which addresses the uncertainty in fusion strategies and target mismatch in multi-modal video recommendation by combining variational Bayesian fusion with context-aware priors and recommendation-guided adversarial refinement.
VCapsBench: A Large-scale Fine-grained Benchmark for Video Caption Quality Evaluation
Shi-Xue Zhang (University of Science and Technology Beijing), Xu-Cheng Yin (University of Science and Technology Beijing)
CodeGenerationTransformerLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: Propose VCapsBench, a fine-grained video caption quality assessment benchmark containing 5,677 videos and 109,796 QA pairs, aiming to systematically evaluate the accuracy, coverage, and consistency of captions across 21-dimensional fine-grained attributes.
VEDA: Generation of 3D Molecules via Variance-Exploding Diffusion with Annealing
Peining Zhang (University of Connecticut), Minghu Song (Institute of Health and Medicine Hefei Comprehensive National Science Center)
CodeDrug DiscoveryGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: This paper proposes a unified SE(3)-equivariant scattering generation framework, VEDA, which combines variance explosion (VE) diffusion with heating denoising and preprocessing to efficiently generate geometrically accurate and chemically valid 3D molecular structures within a small number of sampling steps.
Veli: Unsupervised Method and Unified Benchmark for Low-Cost Air Quality Sensor Correction
Yahia Dalbah (University of Amsterdam), Yen-Chia Hsu (University of Amsterdam)
CodeRestorationAuto EncoderTime SeriesBenchmark
🎯 What it does: Propose a low-cost, unsupervised, no-reference site air quality sensor reading calibration method called Veli, and create the largest air quality sensor data warehouse, AQ-SDR.
Verb Mirage: Unveiling and Assessing Verb Concept Hallucinations in Multimodal Large Language Models
Zehao Wang (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Jiao Tong University)
CodeExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: This paper first constructs a benchmark for evaluating false positives in multimodal large language models (MLLM) focused on verb concepts, systematically assesses the performance of various mainstream MLLMs on verb hallucination, and verifies the inadequacy of existing hallucination mitigation methods without training interventions for this task; subsequently, it proposes a parameter-efficient fine-tuning scheme based on enriched verb knowledge (utilizing the Pangea semantic structure), significantly reducing the verb hallucination rate in MLLMs.
Versatile Vision-Language Model for 3D Computed Tomography
Jiayu Lei (University of Science and Technology of China), Yanfeng Wang (Shanghai Artificial Intelligence Laboratory)
CodeClassificationSegmentationGenerationConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBiomedical DataComputed Tomography
🎯 What it does: Developed a multi-modal model CTInstruct for 3D CT, supporting multiple tasks such as diagnosis, segmentation, and generation;
VFCionX: Bridging Large and Small Models for Robust Vulnerability-Fixing Commit Identification
Xing Cui (Institute of Software, Chinese Academy of Sciences), Xiang Ling (Institute of Software, Chinese Academy of Sciences)
CodeClassificationAI Code AssistantConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: Propose the VFCionX framework, which collaborates between large language models and small models to automatically identify vulnerability fixes in software submissions.
🎯 What it does: Proposed a VGGS method based on 3D Gaussian Splatting, which uses the multi-view depth prior VGGT to guide surface reconstruction under sparse views.
ViCToR: Improving Visual Comprehension via Token Reconstruction for Pretraining LMMs
Yin Xie (DeepGlint), Jiankang Deng (DeepGlint)
CodeRepresentation LearningTransformerLarge Language ModelVision Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: Propose the ViCToR framework, which introduces a visual understanding stage in multi-modal pre-training. It replaces original visual tokens with a learnable visual token pool and Hungarian matching, helping LLMs better understand images;
🎯 What it does: Proposes MiD-VMD, a new framework for video mirror detection that utilizes Motion-in-Depth (MiD) hints combining 3D motion, depth, and image features.
VideoSeg-R1:Reasoning Video Object Segmentation via Reinforcement Learning
Zishan Xu (South China Normal University), Lihua Cai (South China Normal University)
CodeSegmentationExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose VideoSeg-R1, a framework that introduces reinforcement learning into video reasoning segmentation, enabling precise segmentation and tracking of video objects given natural language descriptions.
ViDia2Std: A Parallel Corpus and Methods for Low-Resource Vietnamese Dialect-to-Standard Translation
Khoa Anh Ta (University of Information Technology), Kiet Van Nguyen (University of Information Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper constructs a nationwide parallel corpus from Vietnamese dialects to standard Vietnamese called ViDia2Std, trains and benchmarks various seq2seq models on this corpus, and evaluates the practical value of dialect normalization through downstream tasks such as machine translation and sentiment analysis.
VietCheckMed: Explainable Regulatory Compliance Checking for Medical Advertisements on Vietnamese Social Media
Nguyen Thanh Tam (University of Science), Binh T. Nguyen (University of Science)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the VietCheckMed framework to enable explainable regulatory compliance checks on Vietnamese medical advertisements, and released the first expert-verified benchmark, VietAestheticAds.
View-on-Graph: Zero-Shot 3D Visual Grounding via Vision-Language Reasoning on Scene Graphs
Yuanyuan Liu (Dalian University of Technology), Xin Yang (Dalian University of Technology)
CodeRecognitionGraph Neural NetworkVision Language ModelMultimodalityGraph
🎯 What it does: Proposes the View-on-Graph (VoG) method to achieve zero-shot 3D visual localization by leveraging multimodal multi-layer scene graphs for interactive exploration with VLM;
🎯 What it does: Designed a two-stage multi-view clustering framework named GFSAF, first using granular sphere fuzzy contrastive learning to extract mutual information and obtaining complementary information through noise stripping loss, then employing cross-view attention fusion to achieve robust clustering features.
ViG-RAG: Video-aware Graph Retrieval-Augmented Generation via Temporal and Semantic Hybrid Reasoning
Zongsheng Cao (Shanghai Artificial Intelligence Laboratory), Zigan Wang (Tsinghua University)
CodeGenerationRetrievalTransformerVision Language ModelVideoMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes the ViG-RAG framework, which for long video understanding constructs a probabilistic temporal knowledge graph and combines semantic and temporal dual retrieval to achieve retrieval-augmented generation.
VIR-Bench: Evaluating Geospatial and Temporal Understanding of MLLMs via Travel Video Itinerary Reconstruction
Hao Wang (Waseda University), Daisuke Kawahara (Waseda University)
CodeTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: Propose the VIR-Bench benchmark, construct 200 Japanese travel vlogs and generate itinerary maps, split into node prediction (province, city, POI) and edge prediction (relationship and time transition), and develop a travel planning agent based on this benchmark.
CodeLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Built VisAssist, the first large-scale first-person video question-answering dataset recorded by visually impaired individuals, and conducted benchmark evaluations on multiple state-of-the-art video question-answering models;
Vision-G1: Towards General Reasoning Vision-Language Models via Reinforcement Learning
Yuheng Zha (UC San Diego), Zhiting Hu (UC San Diego)
CodeReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Constructed a visual reasoning dataset covering 5 domains and 13 dimensions with 46 tasks, and trained the Vision-G1 VLM using reinforcement learning from verifiable rewards to enhance its general visual reasoning capabilities.
Vision-language Incremental Learning with Dual Class-individual Memory
Fuhai Chen (Fuzhou University), Xuri Ge (Shandong University)
CodeRepresentation LearningMeta LearningConvolutional Neural NetworkTransformerVision Language ModelAuto EncoderMultimodalityBenchmark
🎯 What it does: Proposes the Dual Class-Individual Memory (DCIM) framework to address category-level and scene-level forgetting issues in visual-lingual incremental learning.
CodeGenerationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningVision Language ModelImageVideo
🎯 What it does: This paper proposes VisionReward, a multidimensional fine-grained human preference learning framework for image and video generation tasks, and uses it as a reward model for reinforcement learning optimization.
ViTCoP: Accelerating Large Vision-Language Models via Visual and Textual Semantic Collaborative Pruning
Wen Luo (Huazhong University of Science and Technology), LiQun Huang (Huazhong University of Science and Technology)
CodeComputational EfficiencyTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: Proposes a vision-text collaborative pruning framework named ViTCoP, which can significantly reduce the number of visual tokens in vision-language models and lower inference costs.
ViTE: Virtual Graph Trajectory Expert Router for Pedestrian Trajectory Prediction
Ruochen Li (Durham University), Hubert P. H. Shum (Durham University)
CodeAutonomous DrivingComputational EfficiencyGraph Neural NetworkMixture of ExpertsVideo
🎯 What it does: Propose the ViTE framework, which predicts pedestrian trajectories by utilizing virtual graphs constructed with virtual nodes and an expert router based on Mixture-of-Experts.
Voices, Faces, and Feelings: Multi-modal Emotion-Cognition Captioning for Mental Health Understanding
Zhiyuan Zhou (Hefei University of Technology), Shijie Hao (Hefei University of Technology)
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Propose a multimodal emotion-cognition collaborative generation framework (ECMC) that can generate interpretable emotion-cognition descriptions from video, audio, and text multimodal data for mental health assessment.
VP-Bench: A Comprehensive Benchmark for Visual Prompting in Multimodal Large Language Models
Mingjie Xu (Hong Kong University of Science and Technology), Wenqiang Lei (Huazhong University of Science and Technology)
CodeTransformerPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes VP-Bench, a two-phase evaluation framework designed to measure the performance of multimodal large language models in visual prompt perception and practical tasks.
Shuo Feng (Nanjing University of Aeronautics and Astronautics), Shuqiang Jiang (University of Chinese Academy of Sciences)
CodeGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelImageGraph
🎯 What it does: Proposed the Visual Prompt Navigation (VPN) paradigm, which utilizes visual prompts drawn by users on a 2D panoramic map to guide agents in completing navigation tasks.
CodeOptimizationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Propose VQ-Insight, a video quality assessment framework for AIGC (AI-generated content) based on reasoning-driven vision-language models, capable of video preference comparison, multi-dimensional quality scoring, and natural video scoring, with support for joint fine-tuning with generative models;
VQAThinker: Exploring Generalizable and Explainable Video Quality Assessment via Reinforcement Learning
Linhan Cao (Shanghai Jiao Tong University), Xiongkuo Min (Shanghai Jiao Tong University)
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodalityChain-of-Thought
🎯 What it does: Proposed a no-reference video quality assessment framework called VQAThinker, which leverages a large multimodal model combined with reinforcement learning (GRPO) to achieve joint modeling of video quality understanding and scoring, and generates explainable results through reasoning trajectories;
VSPO: Validating Semantic Pitfalls in Ontology via LLM-Based CQ Generation
Hyojun Choi (Yonsei University), Kyong-Ho Lee (Yonsei University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Constructed an alignment error-injected dataset and trained LLaMA-3.1-8B-Instruct to generate CQs with verifiable semantic flaws, addressing the limitations of traditional CQs that rely solely on similarity-based evaluation.
CodeClassificationTransformerPrompt EngineeringVision Language ModelVideo
🎯 What it does: A frozen CLIP model is used to map video frames to text prototypes, generating discrete text representations. The codebook is updated through adaptive text prompts, and confidence-aware fusion is applied to enhance video classification performance.
🎯 What it does: This paper proposes the VTinker framework, combining guided flow upscaling (GFU) with texture mapping to achieve high-resolution video frame interpolation;
W2S-AlignTree: Weak-to-Strong Inference-Time Alignment for Large Language Models via Monte Carlo Tree Search
Zhenyu Ding (Xi'an Jiaotong University), Ning Ding (Xi'an Jiaotong University)
CodeClassificationGenerationTransformerLarge Language ModelText
🎯 What it does: Proposed a plug-and-play framework called W2S-AlignTree that aligns large language models (LLMs) during the reasoning phase by combining Monte Carlo Tree Search (MCTS) with the weak-to-strong (W2S) strategy;
Walking Further: Semantic-Aware Multimodal Gait Recognition Under Long-Range Conditions
Zhiyang Lu (Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University), Ming Cheng (OPPO Research Institute)
CodeRecognitionConvolutional Neural NetworkGraph Neural NetworkTransformerSupervised Fine-TuningVision Language ModelContrastive LearningVideoTextMultimodalityPoint CloudBenchmark
🎯 What it does: Proposes the LRGait multimodal long-range gait recognition benchmark and the EMGaitNet framework, directly using raw RGB videos and LiDAR point clouds for end-to-end semantic-guided multimodal fusion.