NeurIPS 2025 Papers — Page 51
Conference on Neural Information Processing Systems · 5275 papers
Unleashing Hour-Scale Video Training for Long Video-Language Understanding
Jingyang Lin (AMD), Emad Barsoum (AMD)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: This paper designs the largest and longest long video question-answering dataset, VideoMarathon, and proposes the Hour-LLaVA model to achieve hour-level video language understanding.
Unleashing the Potential of Multimodal LLMs for Zero-Shot Spatio-Temporal Video Grounding
Zaiquan Yang (City University of Hong Kong), Rynson W. H. Lau (City University of Hong Kong)
RecognitionObject DetectionOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodality
🎯 What it does: Utilizing multi-modal large language models (MLLM) to achieve spatiotemporal video localization (STVG) under unsupervised conditions, by identifying and optimizing specific 'localization words', and enhancing localization accuracy through query decomposition and temporal augmentation methods.
Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator
Jianze Li (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: A first-order diffusion-based image super-resolution method D3SR is proposed, utilizing a large-scale diffusion discriminator and edge-aware DISTS for adversarial training, achieving high-quality, low-latency real-world image super-resolution.
Unlocker: Disentangle the Deadlock of Learning between Label-noisy and Long-tailed Data
Chen Shu, Hanzi Wang (Xiamen University)
ClassificationOptimizationData-Centric LearningImage
🎯 What it does: This paper proposes the Unlocker framework, which addresses the deadlock problem of noise label learning (NLL) and long-tail learning (LTL) on long-tail noisy label data through a dual-layer optimization.
Unlocking Dataset Distillation with Diffusion Models
Brian Bernhard Moser, Andreas Dengel (German Research Center for Artificial Intelligence)
Data SynthesisKnowledge DistillationDiffusion modelImage
🎯 What it does: This paper proposes a latent data distillation method using diffusion models, called LD3M, which achieves efficient distillation of a small number of synthetic samples by learning latent codes and conditional embeddings.
Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time
Daniel D. Richman (Stanford University), Ron O. Dror (Stanford University)
Protein Structure PredictionDiffusion modelBiomedical Data
🎯 What it does: Using ConforMix to enhance the sampling of diffusion models during inference to reveal the hidden conformations and energy landscapes of biological molecules such as proteins.
Unlocking Multimodal Mathematical Reasoning via Process Reward Model
Ruilin Luo (Tsinghua University), Yujiu Yang (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningTextMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a three-stage URSA training framework, constructs a large-scale multimodal CoT corpus MMathCoT-1M and process supervision data DualMath-1.1M, trains a reward model and combines it with PS-GRPO reinforcement learning, significantly enhancing multimodal mathematical reasoning capabilities.
Unlocking SLM Potential for Data Analysis Code Generation via Non-Parametric Knowledge Distillation
Jinyang Li (University of Hong Kong), Reynold Cheng (Microsoft Research)
OptimizationKnowledge DistillationAI Code AssistantTransformerLarge Language ModelTabularRetrieval-Augmented Generation
🎯 What it does: Achieving efficient and privacy-friendly deployment of data analysis code generation by transferring the knowledge of large language models (LLMs) to small language models (SLMs) in an unsupervised context learning manner.
Unmasking Puppeteers: Leveraging Biometric Leakage to Expose Impersonation in AI-Based Videoconferencing
Danial Samadi Vahdati (Drexel University), Matthew Stamm (NVIDIA)
RecognitionAnomaly DetectionRecurrent Neural NetworkContrastive LearningVideo
🎯 What it does: A real-time, registration-free defense mechanism is proposed, utilizing the biometric features leaked in the pose-expression embeddings transmitted during AI voice and video conferences to detect and prevent 'puppeteering' attacks.
Unraveling Metameric Dilemma for Spectral Reconstruction: A High-Fidelity Approach via Semi-Supervised Learning
Xingxing Yang (Hong Kong Baptist University), Zaifeng Yang (Agency for Science Technology and Research)
ClassificationRestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: The Diff-Spectra method is proposed, which combines physics-informed spectral estimation with unsupervised high-fidelity spectral regularization using semi-supervised learning to address the multi-color spectral paradox in the reconstruction of RGB to hyperspectral images.
Unsupervised Federated Graph Learning
Lele Fu (Sun Yat-sen University), Chuan Chen (Griffith University)
Federated LearningRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: The FedPAM framework is proposed to address the representation space alignment and global parameter adaptive aggregation issues in unsupervised federated graph learning.
Unsupervised Learning for Optimal Transport plan prediction between unbalanced graphs
Sonia Mazelet (Ecole Polytechnique), Bertrand Thirion (Inria)
OptimizationGraph Neural NetworkGraphBiomedical DataMagnetic Resonance Imaging
🎯 What it does: An unsupervised deep learning framework ULOT is proposed for predicting the optimal transport plan of FUGW between graphs, and directly minimizing the FUGW loss through amortized optimization;
Unsupervised Trajectory Optimization for 3D Registration in Serial Section Electron Microscopy using Neural ODEs
Zhenbang Zhang (Mohamed Bin Zayed University of Artificial Intelligence), Renmin Han (Shandong University)
OptimizationBiomedical DataOrdinary Differential Equation
🎯 What it does: A continuous trajectory optimization-based ssEM 3D registration framework called NeuroTrajReg is proposed, treating pixel trajectories as a dynamic process and smoothing them;
Unveiling Chain of Step Reasoning for Vision-Language Models with Fine-grained Rewards
Honghao Chen (Institute of Automation, Chinese Academy of Sciences), Xinlong Wang (Beijing Academy of Artificial Intelligence)
TransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposes a Chain of Steps Reasoning (CoS) framework for visual language models, utilizing structured steps (Name-Thought-Reflection) and achieving more precise reinforcement learning and reasoning time scaling through fine-grained rewards.
Unveiling Concept Attribution in Diffusion Models
Quang H Nguyen, Khoa D Doan
GenerationData SynthesisExplainability and InterpretabilityDiffusion modelImage
🎯 What it does: A framework called CAD (Component Attribution for Diffusion Model) is proposed to quantitatively attribute the contribution of each parameter in the diffusion model, and based on this, two lightweight editing algorithms for inference, CAD-Erase and CAD-Amplify, are designed to achieve the elimination and amplification of target concepts, respectively.
Unveiling Environmental Sensitivity of Individual Gains in Influence Maximization
Xinyan Su (Chinese Academy of Sciences), Jun Li (Chinese Academy of Sciences)
Recommendation SystemOptimizationGraph Neural NetworkDiffusion modelGraph
🎯 What it does: This paper proposes the CauIM framework, which utilizes causal inference to estimate individual treatment effects (ITE) and implements impact maximization in hypergraphs.
Unveiling Extraneous Sampling Bias with Data Missing-Not-At-Random
Chunyuan Zheng (Peking University), Mengyue Yang (University of Bristol)
Recommendation SystemTabular
🎯 What it does: A dual robust estimation method (SuperDR) is proposed under the condition of incomplete consistency between the training and testing set units, which eliminates additional covariance bias by correcting error estimates.
Unveiling m-Sharpness Through the Structure of Stochastic Gradient Noise
Haocheng Luo (Monash University), Trung Le (Monash University)
OptimizationTransformerImageStochastic Differential Equation
🎯 What it does: Theoretical analysis of SAM and its variants reveals the role of the structure of stochastic gradient noise in implicit sharpness regularization, and based on this, the Reweighted SAM (RW-SAM) algorithm is proposed.
Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model
Tianle Li (Chinese University of Hong Kong), Yu Cheng (Chinese University of Hong Kong)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This study investigates the combinatorial reasoning capabilities of visual-language models (VLM) in reinforcement learning (RL) fine-tuning and proposes the ComPABench benchmark to evaluate combinatorial generalization across modalities, tasks, and out-of-distribution (OOD) conditions.
Unveiling the Learning Mind of Language Models: A Cognitive Framework and Empirical Study
Zhengyu Hu (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
Large Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: A three-dimensional learning framework based on cognitive psychology (teacher-style learning, conceptual learning, experiential learning) is proposed, and large-scale experiments and benchmark construction are carried out under this framework;
Unveiling the Power of Multiple Gossip Steps: A Stability-Based Generalization Analysis in Decentralized Training
Qinglun Li (National University of Defense Technology), Li Shen (Sun Yat-sen University)
OptimizationFederated LearningImage
🎯 What it does: This paper studies the generalization error and excess error of multi-step gossip (MGS) in decentralized training, providing stability analysis and upper bounds under non-convex, gradient-free bounded assumptions.
Unveiling the Spatial-temporal Effective Receptive Fields of Spiking Neural Networks
Jieyuan Zhang (University of Electronic Science and Technology of China), Haizhou Li (University of Electronic Science and Technology of China)
Object DetectionSegmentationSpiking Neural NetworkTransformerSupervised Fine-TuningImageVideo
🎯 What it does: The ST-ERF framework is proposed to evaluate the spatiotemporal receptive field and finds that the Transformer-SNN lacks a global receptive field in visual long-sequence tasks, leading to the design of MLPixer and SRB channel mixers to enhance global spatial accessibility.
Unveiling the Uncertainty in Embodied and Operational Carbon of Large AI Models through a Probabilistic Carbon Accounting Model
Xiaoyang Zhang (Hong Kong Polytechnic University), Dan Wang (Hong Kong University of Science and Technology)
Time Series
🎯 What it does: A probabilistic carbon accounting model (PCAM) has been developed to systematically quantify the embedded carbon and operational carbon uncertainties of large AI models, providing distributed estimation results.
Unveiling Transformer Perception by Exploring Input Manifolds
Alessandro Benfenati (Università degli Studi di Milano), Elisabetta Rocchetti (Università degli Studi di Milano)
Explainability and InterpretabilityRepresentation LearningTransformerImageText
🎯 What it does: This paper proposes a method for exploring equivalence classes of Transformer input spaces based on singular Riemannian geometry, and designs two algorithms, SiMEC and SiMExp, to find similar inputs within the same equivalence class and explore different predictions across equivalence classes.
URDF-Anything: Constructing Articulated Objects with 3D Multimodal Language Model
Zhe Li (Peking University), Shanghang Zhang (Peking University)
SegmentationGenerationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningMultimodalityPoint CloudMesh
🎯 What it does: An end-to-end 3D multimodal large language model framework called URDF-Anything is proposed, which generates URDF digital twin models for physical simulation directly from visual input.
URLs Help, Topics Guide: Understanding Metadata Utility in LLM Training
Dongyang Fan (Ecole Polytechnique Federale de Lausanne), Martin Jaggi (Ecole Polytechnique Federale de Lausanne)
GenerationOptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the impact of incorporating contextual metadata (such as URLs, quality scores, and topic/format domains) into the pre-training of large language models on training speed and downstream performance, and explores the possibility of achieving more controllable generation through context-aware pre-training.
User-Instructed Disparity-aware Defocus Control
Yudong Han (Beijing Institute of Technology), Liyuan Pan (Beijing Institute of Technology)
RestorationSegmentationConvolutional Neural NetworkTransformerImageMultimodality
🎯 What it does: This paper proposes a user-instruction controllable depth-of-field re-focusing framework UiD based on dual-pixel (DP) cameras, which allows for arbitrary selection of the focus area and generates corresponding depth-of-field effects through text, boxes, or point prompts.
UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation
Jiyu Guo (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
ClassificationGenerationData SynthesisConvolutional Neural NetworkPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper proposes UTILGEN, a generative data augmentation framework that utilizes task feedback for adaptive optimization, enhancing the task-specific utility of synthetic data.
Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs
Mantas Mazeika (Center for AI Safety), Dan Hendrycks (Center for AI Safety)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper systematically analyzes and controls the internal value system of large language models (LLMs) by constructing a Utility Engineering framework.
V-CECE: Visual Counterfactual Explanations via Conceptual Edits
Nikolaos Spanos (National Technical University of Athens), Giorgos Stamou (National Technical University of Athens)
Autonomous DrivingExplainability and InterpretabilityVision Language ModelDiffusion modelImage
🎯 What it does: A fully black-box visual counterfactual explanation framework, V-CECE, is proposed, which can generate interpretable counterfactual images through concept editing and reveal the semantic gap between models and humans.
V2V: Scaling Event-Based Vision through Efficient Video-to-Voxel Simulation
Hanyue Lou (Peking University), Boxin Shi (Peking University)
Data SynthesisAutonomous DrivingComputational EfficiencyOptical FlowVideo
🎯 What it does: The Video-to-Voxel (V2V) method is proposed, which directly converts traditional video frames into event voxel representations, avoiding explicit event stream generation and significantly reducing storage and I/O burdens.
VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment
Qing Li (Southwest Jiaotong University), Yu-Shen Liu (Tsinghua University)
Gaussian SplattingPoint Cloud
🎯 What it does: A 3D Gaussian point cloud geometry enhancement method based on view alignment is proposed to achieve more accurate surface reconstruction and high-quality novel view synthesis.
Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought
Chao Huang (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodalityChain-of-Thought
🎯 What it does: This paper proposes the Video Anomaly Reasoning (VAR) task and constructs an end-to-end multimodal large language model framework called Vad-R1, which guides the model to reason and describe video anomalies step by step through the Perception-to-Cognition Chain-of-Thought (P2C-CoT);
VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree
Wenlong Li (Xi'an Jiaotong University), Shuiguang Deng (Zhejiang University)
Anomaly DetectionLarge Language ModelVision Language ModelVideo
🎯 What it does: The VADTree framework is proposed, which achieves training-independent video anomaly detection by constructing a hierarchical granularity tree based on general event boundary detection.
VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents
Kangrui Wang (Northwestern University), Manling Li (Northwestern University)
Reinforcement LearningVision Language ModelWorld ModelMultimodality
🎯 What it does: This paper proposes a method to train a multi-turn visual language model agent through reinforcement learning, enabling it to construct an internal world model through explicit visual state reasoning (State Estimation and Transition Modeling);
Valid Inference with Imperfect Synthetic Data
Yewon Byun (Carnegie Mellon University), Bryan Wilder (Carnegie Mellon University)
Data SynthesisLarge Language ModelText
🎯 What it does: This paper proposes a method based on Generalized Method of Moments (GMM) that combines fully synthetic samples generated by large language models with real samples to achieve statistically efficient parameter estimation in scenarios with limited labeled data.
Valid Selection among Conformal Sets
Mahmoud Hegazy (École polytechnique), Aymeric Dieuleveut (École polytechnique)
Tabular
🎯 What it does: This paper proposes a framework based on algorithm stability for pointwise selection of the optimal prediction set among multiple conformal predictors, while ensuring that coverage is not compromised.
Validating LLM-as-a-Judge Systems under Rating Indeterminacy
Luke Guerdan (Carnegie Mellon University), Alexandra Chouldechova (Microsoft Research)
Large Language ModelText
🎯 What it does: In response to the validation issues of the LLM-as-a-judge system under rating indeterminacy, a comprehensive theoretical framework is proposed, and large-scale experiments are conducted with 11 real evaluation tasks and 9 commercial LLMs.
Value Diffusion Reinforcement Learning
Xiaoliang Hu (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)
Reinforcement LearningDiffusion modelSequential
🎯 What it does: A value distribution learning framework based on diffusion models, VDRL, is proposed to improve the accuracy of value estimation in reinforcement learning.
Value Gradient Guidance for Flow Matching Alignment
Zhen Liu (Chinese University of Hong Kong), Dinghuai Zhang (Microsoft Research)
GenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningFlow-based ModelImage
🎯 What it does: Aligning the flow matching model and fine-tuning it using value gradient guidance to make the generated samples more in line with the given reward function while maintaining the prior distribution without distortion.
Value Improved Actor Critic Algorithms
Yaniv Oren (Delft University of Technology), Matthijs T. J. Spaan (Delft University of Technology)
Reinforcement LearningSequential
🎯 What it does: This paper proposes the Value-Improved Actor Critic (VIAC) framework, which decouples the behavior policy from the evaluation policy and performs additional greedy improvements on the value function to accelerate Actor-Critic learning.
Value-Guided Decision Transformer: A Unified Reinforcement Learning Framework for Online and Offline Settings
Hongling Zheng (Wuhan University), Dacheng Tao (Nanyang Technological University)
TransformerReinforcement LearningSequential
🎯 What it does: This paper proposes the Value-Guided Decision Transformer (VDT), which unifies the use of value functions to guide decision-making and achieve trajectory stitching in both offline and online RL scenarios.
Value-Guided KV Compression for LLMs via Approximated CUR Decomposition
Ayan Sengupta (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
RetrievalCompressionTransformerLarge Language ModelText
🎯 What it does: A KV cache compression method based on CUR decomposition, CurDKV (and its adaptive variant AdaCurDKV), is proposed. It selects the key-value pairs to retain by calculating the leverage scores of the key-value matrix, thereby reducing memory usage and inference latency of LLMs under long contexts.
Value-Guided Search for Efficient Chain-of-Thought Reasoning
Kaiwen Wang (Cornell University), Wen Sun (Cornell University)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper studies a block-level search method based on a token-level value model (Value-Guided Search, VGS) to enhance the computational efficiency and performance of Chain-of-Thought (CoT) large language models during testing.
VaMP: Variational Multi-Modal Prompt Learning for Vision-Language Models
Silin Cheng (Hong Kong University), Kai Han (Hong Kong University)
ClassificationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: A new variational multimodal prompt learning framework (VaMP) is proposed for sample-specific and uncertainty-aware prompt tuning in visual-language models.
Vanish into Thin Air: Cross-prompt Universal Adversarial Attacks for SAM2
Ziqi Zhou (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
SegmentationAdversarial AttackContrastive LearningImageVideo
🎯 What it does: A cross-prompt universal adversarial attack UAP-SAM2 has been designed and implemented for the SAM2 video segmentation model.
VaporTok: RL-Driven Adaptive Video Tokenizer with Prior & Task Awareness
Yang Minghao, Alex Jinpeng Wang (Central South University)
GenerationOptimizationReinforcement LearningVideo
🎯 What it does: This paper presents VaporTok, an adaptive video tokenizer based on probabilistic tail truncation, and achieves task-aware token optimization through GRPO reinforcement learning.
VarFlow: Proper Scoring-Rule Diffusion Distillation via Energy Matching
Huiyang Shao (Tsinghua University), Xuefeng Xiao (ByteDance Seed)
GenerationKnowledge DistillationDiffusion modelImage
🎯 What it does: A single-step generator distillation method based on energy distance, VarFlow, has been developed to compress the diffusion teacher model into an efficient single-step student generator.
Variance-Aware Feel-Good Thompson Sampling for Contextual Bandits
Xuheng Li (University of California Los Angeles), Quanquan Gu (University of California Los Angeles)
OptimizationReinforcement LearningTabular
🎯 What it does: A variance-aware Feel-Good Thompson Sampling (FGTS-VA) algorithm for heterogeneous noise is proposed, which achieves an optimal regret upper bound without knowing the noise variance distribution.
Variance-Reduced Long-Term Rehearsal Learning with Quadratic Programming Reformulation
Wen-Bo Du (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabularTime Series
🎯 What it does: This paper proposes a replay learning framework for long-term avoidance of undesirable futures (AUF) decision-making, deriving an explicit expression for the aggregated objective and transforming the originally intractable probabilistic optimization into a solvable quadratic programming problem.
Variational Inference with Mixtures of Isotropic Gaussians
Marguerite Petit Talamon, Anna Korba (CREST ENSAE Institut Polytechnique de Paris)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a variational inference framework based on an equal-weight, scale-variant isotropic Gaussian mixture model, and presents an efficient algorithm using Bures-Wasserstein gradient descent and mirror descent.
Variational Learning Finds Flatter Solutions at the Edge of Stability
Avrajit Ghosh (Michigan State University), Thomas Möllenhoff (RIKEN Center for AI Project)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: Under the Edge of Stability framework, it is analyzed and verified that Variational Learning (VL) can find flatter solutions than gradient descent.
Variational Polya Tree
Lu Xu (University of Hong Kong), Guosheng Yin (University of Hong Kong)
GenerationData SynthesisOptimizationAuto EncoderImageTabular
🎯 What it does: This paper proposes the Variational Polynomial Tree (VPT), which combines the Pólya tree prior with deep generative models to achieve scalable non-parametric density estimation.
Variational Regularized Unbalanced Optimal Transport: Single Network, Least Action
Yuhao Sun (Peking University), Peijie Zhou (Peking University)
OptimizationBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes the Variational RUOT method, which utilizes variational methods to solve the Regularized Unbalanced Optimal Transport (RUOT) problem by learning a scalar field to simultaneously obtain particle velocity and growth rate, thereby reconstructing the continuous dynamics of high-dimensional systems.
Variational Supervised Contrastive Learning
Ziwen Wang (University of Illinois Urbana-Champaign), Ge Liu (University of Illinois Urbana-Champaign)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes VarCon, a supervised contrastive learning framework based on variational inference, which controls the embedding distribution through a posterior-weighted ELBO objective and introduces confidence-adaptive temperature to finely adjust intra-class dispersion.
Variational Task Vector Composition
Boyuan Zhang (University of Chinese Academy of Sciences), Ling Shao (University of Chinese Academy of Sciences)
ClassificationImageText
🎯 What it does: Proposes Variational Task Vector Composition (VTVC), which achieves sample-level task vector composition through Bayesian inference;
Variational Uncertainty Decomposition for In-Context Learning
I. Shavindra Jayasekera (Imperial College London), Yingzhen Li (Imperial College London)
Large Language ModelPrompt EngineeringText
🎯 What it does: A Variational Uncertainty Decomposition (VUD) framework is proposed, which decomposes the uncertainty in context learning of large language models into distinguishable aleatoric and epistemic uncertainties using auxiliary queries.
VASA-3D: Lifelike Audio-Driven Gaussian Head Avatars from a Single Image
Sicheng Xu (Microsoft Research Asia), Baining Guo (Microsoft Research Asia)
GenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImageVideoAudio
🎯 What it does: Proposes VASA-3D, which can generate audio-driven, free-view 3D head avatars from a single portrait photo.
VCM: Vision Concept Modeling with Adaptive Vision Token Compression via Instruction Fine-Tuning
Run Luo (Shenzhen Institute of Advanced Technology), Xiaobo Xia (National University of Singapore)
ClassificationObject DetectionSegmentationCompressionTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: Proposes a self-supervised visual concept model (VCM) that allows large-scale visual language models (LVLM) to extract only the visual concept tokens related to instructions, rather than full image tokens;
Vector Database Watermarking
zhiwen ren, Nenghai Yu (University of Science and Technology of China)
Point Cloud
🎯 What it does: This paper proposes the first watermarking scheme for vector databases, which is further improved to a Transparent Vector Priority (TVP) strategy;
Vector Quantization in the Brain: Grid-like Codes in World Models
Xiangyuan Peng (Peking University), Si Wu (Peking University)
CompressionOptimizationExplainability and InterpretabilityWorld ModelImageSequential
🎯 What it does: This paper proposes a grid code quantization-based observation-action sequence compression method (GCQ), which serves as a unified world model for long-range prediction, goal planning, and inverse model reasoning.
Venus-MAXWELL: Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models
Yuanxi Yu (Shanghai Jiao Tong University), Mingchen Li (Shanghai Jiao Tong University)
Computational EfficiencyDrug DiscoveryTransformerSupervised Fine-TuningBiomedical Data
🎯 What it does: The Venus-MAXWELL framework is proposed, transforming the prediction of protein variant stability from sequence-to-label to sequence-to-landscape, utilizing a matrix scoring method to predict all single-point mutations' ΔΔG in one go.
VERA: Variational Inference Framework for Jailbreaking Large Language Models
Anamika Lochab (Purdue University), Ruqi Zhang (Purdue University)
Adversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper presents VERA, a black-box jailbreak framework based on variational inference that can generate diverse and efficient attack prompts without the need for manual templates or gradient information.
VeriLoC: Line-of-Code Level Prediction of Hardware Design Quality from Verilog Code
Raghu Vamshi Hemadri (New York University), Siddharth Garg (New York University)
Large Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a framework called VeriLoC, based on Verilog LLM embedding, to predict hardware design quality in real-time at the RTL level, including timing and routing congestion, and provides impact assessments for specific lines of code.
VeriThinker: Learning to Verify Makes Reasoning Model Efficient
Zigeng Chen (National University of Singapore), Xinchao Wang (National University of Singapore)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: By performing supervised fine-tuning on auxiliary verification tasks in large reasoning models, chain-of-thought (CoT) compression is achieved, significantly reducing the length of reasoning chains;
Versatile Transferable Unlearnable Example Generator
Zhihao Li (Western University), Boyu Wang (Western University)
Domain AdaptationAdversarial AttackGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A transferable zero-shot sample generator (VTG) is proposed, making data unlearnable by models in various scenarios.
Vertical Federated Feature Screening
Huajun Yin (Renmin University of China), Danyang Huang (Renmin University of China)
Federated LearningSafty and PrivacyComputational EfficiencyTabular
🎯 What it does: A two-stage vertical federated feature selection (VFS) algorithm is proposed, which significantly reduces the computational, communication, and encryption costs of high-dimensional sparse data while ensuring privacy.
VESSA: Video-based objEct-centric Self-Supervised Adaptation for Visual Foundation Models
Jesimon Barreto (Federal University of Minas Gerais), William Robson Schwartz
Domain AdaptationKnowledge DistillationContrastive LearningVideo
🎯 What it does: Self-supervised fine-tuning of visual foundation models on unlabeled object-centered short videos to achieve domain adaptation.
VETA-DiT: Variance-Equalized and Temporally Adaptive Quantization for Efficient 4-bit Diffusion Transformers
QinkaiXu, Yuxiang Fu (Nanjing University)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImageVideo
🎯 What it does: Aiming at post-training quantization (PTQ) for Diffusion Transformers (DiTs), the VETA-DiT framework is proposed to achieve efficient inference with 4-bit weights/activations while maintaining generation quality.
VFRTok: Variable Frame Rates Video Tokenizer with Duration-Proportional Information Assumption
Tianxiong Zhong (Beijing Institute of Technology), Zhiwei Zhang (Beijing Institute of Technology)
GenerationCompressionTransformerDiffusion modelVideo
🎯 What it does: This paper proposes a video tokenizer called VFRTok based on the Duration-Proportional Information Assumption, which supports variable frame rate encoding and decoding, and enhances content modeling through Partial RoPE.
Vgent: Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding
Xiaoqian Shen (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
GenerationRetrievalGraph Neural NetworkLarge Language ModelVideoMultimodalityRetrieval-Augmented Generation
🎯 What it does: A graph-structured retrieval-inference-generation framework Vgent is proposed for long video understanding.
VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold
Dominic Rosario Maggio (Massachusetts Institute of Technology), Luca Carlone (Massachusetts Institute of Technology)
OptimizationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Using an uncalibrated monocular RGB camera, combined with VGGT's feedforward scene reconstruction, a dense 3D map is incrementally constructed by subgraphs, and global consistency is achieved through 15-degree-of-freedom projection homotopy (SL(4)) optimization between subgraphs;
VIBE: Annotation-Free Video-to-Text Information Bottleneck Evaluation for TL;DR
Shenghui Chen (University of Texas at Austin), ufuk topcu
OptimizationTransformerVision Language ModelVideoText
🎯 What it does: The VIBE framework is proposed to evaluate and select video-to-text summaries in an unsupervised manner, helping humans make decisions faster and more accurately.
Vicinal Label Supervision for Reliable Aleatoric and Epistemic Uncertainty Estimation
Linye Li (Tongji University), Xiaodong Yue
ClassificationAnomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: By introducing soft labels generated by vicinal risk minimization (VRM) into the evidence-based deep learning (EDL) framework, the traditional hard labels are smoothed, thereby enhancing the model's estimation of prior and likelihood (aleatoric and epistemic) uncertainties.
Vicinity-Guided Discriminative Latent Diffusion for Privacy-Preserving Domain Adaptation
Jing Wang (University of British Columbia), Junhyug Noh (Ewha Womans University)
Domain AdaptationSafty and PrivacyDiffusion modelContrastive LearningImage
🎯 What it does: A framework for privacy-preserving source domain unsupervised domain adaptation (DVD) using implicit diffusion models is proposed.
ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs
Xiyao Wang (University of Maryland), Lijuan Wang (Microsoft)
RetrievalTransformerReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: ViCrit is designed and implemented as a verifiable agent task based on RL, training visual language models to detect and locate fine-grained visual hallucinations in long paragraph-style image descriptions.
Vid-SME: Membership Inference Attacks against Large Video Understanding Models
Qi Li (National University of Singapore), Xinchao Wang (National University of Singapore)
Adversarial AttackLarge Language ModelVideo
🎯 What it does: Proposes a membership inference attack method for video understanding large language models called Vid-SME.
ViDAR: Video Diffusion-Aware 4D Reconstruction From Monocular Inputs
Michal Nazarczuk (Huawei Noah's Ark Lab), Eduardo Pérez-Pellitero (Huawei Noah's Ark Lab)
RestorationGenerationDiffusion modelGaussian SplattingVideo
🎯 What it does: By combining a personalized diffusion model with 4D Gaussian splatting, 4D reconstruction of monocular videos is achieved, enabling the generation of high-quality dynamic new perspectives from monocular videos.
VidEmo: Affective-Tree Reasoning for Emotion-Centric Video Foundation Models
Zhicheng Zhang (Nankai University), Jufeng Yang (Kuaishou Technology)
ClassificationRecognitionGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodality
🎯 What it does: A tree-structured reasoning framework called VidEmo is proposed, guided by emotional cues, achieving fine-grained attribute perception, expression analysis, and high-level emotional understanding of videos through two-stage training (curriculum emotion learning + emotional tree reinforcement learning);
Video Diffusion Models Excel at Tracking Similar-Looking Objects Without Supervision
Chenshuang Zhang (KAIST), Chengzhi Mao (Rutgers University)
Object TrackingDiffusion modelVideo
🎯 What it does: By utilizing the motion representations generated by the pre-trained video diffusion model during the high noise stage, a self-supervised video object tracking framework TED was constructed, achieving pixel-level tracking of targets with similar appearances.
Video Perception Models for 3D Scene Synthesis
Rui Huang (Tsinghua University), Francis Engelmann (Stanford University)
SegmentationGenerationData SynthesisRetrievalLarge Language ModelDiffusion modelVideoText
🎯 What it does: Proposes the VIPSCENE framework, which generates editable 3D scenes from text or image prompts using video generation, 3D reconstruction, 2D perception, and asset retrieval;
Video World Models with Long-term Spatial Memory
Tong Wu (Stanford University), Gordon Wetzstein (Stanford University)
GenerationData SynthesisTransformerDiffusion modelWorld ModelVideoPoint Cloud
🎯 What it does: A long-term spatial memory mechanism is proposed to be integrated into the video world model to enhance the long-term consistency and spatial coherence of generated videos.
Video-R1: Reinforcing Video Reasoning in MLLMs
Kaituo Feng (Chinese University of Hong Kong), Xiangyu Yue (Chinese University of Hong Kong)
Supervised Fine-TuningReinforcement LearningVision Language ModelImageVideoMultimodality
🎯 What it does: This paper presents Video-R1, a multimodal large language model based on a rule reinforcement learning (R1) framework, specifically designed to enhance video reasoning capabilities.
Video-RAG: Visually-aligned Retrieval-Augmented Long Video Comprehension
Yongdong Luo (Xiamen University), Rongrong Ji (Xiamen University)
Object DetectionRetrievalTransformerLarge Language ModelVision Language ModelVideoTextRetrieval-Augmented Generation
🎯 What it does: A no-training, pluggable retrieval-augmented long video understanding framework called Video-RAG has been developed, which enhances long video reasoning capabilities by utilizing OCR, ASR, and object detection to generate visually aligned auxiliary text fed into existing large language models.
VideoChat-R1.5: Visual Test-Time Scaling to Reinforce Multimodal Reasoning by Iterative Perception
Ziang Yan (Zhejiang University), Yi Wang (Shanghai AI Laboratory)
OptimizationLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Proposes Visual Test-Time Scaling (VTTS), which enhances the reasoning ability of multimodal large language models (MLLMs) through iterative visual perception during the inference process.
VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding
Zongxia Li (University of Maryland), Jordan Lee Boyd-Graber (University of Maryland)
Data SynthesisAnomaly DetectionTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: The VideoHallu dataset is proposed to test the physical and common sense understanding of VLM using synthetic videos, and RL fine-tuning is used to enhance its ability to recognize abnormal scenes.
VideoLucy: Deep Memory Backtracking for Long Video Understanding
Jialong Zuo (Huazhong University of Science and Technology), Changxin Gao (Huazhong University of Science and Technology)
RetrievalTransformerLarge Language ModelAgentic AIPrompt EngineeringVideoTextMultimodalityBenchmark
🎯 What it does: The VideoLucy framework is proposed, which implements long video understanding through a hierarchical memory structure and agent-based iterative backtracking, allowing for a layered retrieval of all video information related to a question from coarse to fine and providing answers.
VideoMAR: Autoregressive Video Generation with Continuous Tokens
Hu Yu (University of Science and Technology of China), Feng Zhao
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: This paper presents VideoMAR, a single-sided autoregressive image-video generation model based on continuous tokens.
VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models
Xiangdong Zhang (Shanghai Jiao Tong University), Yu Cheng (Chinese University of Hong Kong)
GenerationData SynthesisKnowledge DistillationTransformerSupervised Fine-TuningDiffusion modelVideoPhysics Related
🎯 What it does: Inject physical knowledge from the video foundation model into the text-to-video diffusion model (CogVideoX) through Token Relation Distillation (TRD) soft alignment to enhance the physical feasibility of generated videos.
VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning
Qi Wang (Beijing Institute of Technology), Tianfei Zhou (Beijing Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodalityChain-of-Thought
🎯 What it does: A framework called VIDEORFT is proposed to cultivate video reasoning capabilities by implementing Reinforcement Fine-Tuning (RFT) on a multimodal large language model (MLLM);
Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations
Xin Liu (Institute of Automation, Chinese Academy of Sciences), Dongbin Zhao (Institute of Automation, Chinese Academy of Sciences)
Representation LearningRobotic IntelligenceReinforcement LearningContrastive LearningWorld ModelVideo
🎯 What it does: A behavior cloning framework based on latent representations, BCV-LR, is proposed, which can achieve sample-efficient visual policy learning in a reward-free environment using only unlabeled videos.
VideoTitans: Scalable Video Prediction with Integrated Short- and Long-term Memory
Young-Jae Park (Gwangju Institute of Science and Technology), Hae-Gon Jeon (Yonsei University)
GenerationData SynthesisPose EstimationComputational EfficiencyTransformerVideo
🎯 What it does: A video prediction framework named VideoTitans is proposed, utilizing a gradient-driven Titans memory module, sliding window attention, and persistent memory to capture short-term motion and long-term dependencies.
VideoVLA: Video Generators Can Be Generalizable Robot Manipulators
Yichao Shen (Xi'an Jiaotong University), Baining Guo (Microsoft Research Asia)
GenerationRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelVideoTextMultimodality
🎯 What it does: This paper proposes the VideoVLA framework, which combines pre-trained video generation models with robotic action prediction. It enables the simultaneous prediction of the next action sequence and the generation of corresponding future visual content when given a language instruction and current visual observation, thereby supporting interactive control of robots.
ViewCraft3D: High-fidelity and View-Consistent 3D Vector Graphics Synthesis
Chuang Wang (Beihang University), Qian Yu (Beihang University)
GenerationData SynthesisOptimizationScore-based ModelMesh
🎯 What it does: Based on a single image, high-fidelity and viewpoint-consistent 3D vector graphics (3D Bézier curve representation) are generated using 3D prior information.
ViewPoint: Panoramic Video Generation with Pretrained Diffusion Models
Zixun Fang (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: This paper studies a new framework called ViewPoint that utilizes a pre-trained perspective video model to generate high-quality panoramic videos.
VIKING: Deep variational inference with stochastic projections
Samuel G. Fadel (Technical University of Denmark), Søren Hauberg (Technical University of Denmark)
OptimizationImage
🎯 What it does: A variational inference framework VIKING is proposed for over-parameterized deep networks, which divides the parameter space into the kernel of Fisher-Rao metric (the directions invariant to training data) and the image (the directions that change the loss), and describes the two types of uncertainty using fully correlated Gaussian posteriors.
VimoRAG: Video-based Retrieval-augmented 3D Motion Generation for Motion Language Models
Haidong Xu (Harbin Institute of Technology), Hao Fei (National University of Singapore)
GenerationRetrievalLarge Language ModelVideoTextRetrieval-Augmented Generation
🎯 What it does: Proposes the VimoRAG framework, which utilizes an unlabelled wild video database as a retrieval enhancement signal to improve the quality of text-to-3D action generation;
Vinci: Deep Thinking in Text-to-Image Generation using Unified Model with Reinforcement Learning
Wang Lin (Zhejiang University), Hanwang Zhang (Nanyang Technological University)
GenerationLarge Language ModelReinforcement LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Vinci is proposed, a unified model capable of deep thinking during the text-to-image generation process, which can perform real-time image understanding and reflection after generating images, and iteratively improve the images based on the reflection results.
VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion
Jaekyun Park (Korea Advanced Institute of Science and Technology), Hye Won Chung (Korea Advanced Institute of Science and Technology)
ClassificationRepresentation LearningTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes the VIPAMIN visual prompt initialization method, which improves the visual prompt tuning of self-supervised models through attention-guided matching and orthogonal projection.
Virtual Fitting Room: Generating Arbitrarily Long Videos of Virtual Try-On from a Single Image
Jun-Kun Chen (SpreeAI), Yu-Xiong Wang (University of Illinois Urbana-Champaign)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: Automatically generate high-resolution (720×1152) virtual try-on videos of arbitrary length (up to minutes) based on a single user image, clothing image, and motion reference video.