ICML 2025 Papers — Page 32
International Conference on Machine Learning · 3257 papers
Understanding the Unfairness in Network Quantization
Bing Liu (Huazhong University of Science and Technology), Xianjun Deng (Huazhong University of Science and Technology)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the impact of network quantization (PTQ and QAT) on the fairness (accuracy disparity) of different data subgroups and proposes a solution to mitigate this unfairness through data augmentation.
UnHiPPO: Uncertainty-aware Initialization for State Space Models
Marten Lienen (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
ClassificationRecognitionAudio
🎯 What it does: A method for uncertainty-aware initialization of state space models (SSM) called UnHiPPO is proposed, which can perform posterior inference in the presence of measurement noise.
UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control
Kaizhen Zhu (ShanghaiTech University), Ye Shi (ShanghaiTech University)
RestorationSuper ResolutionDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A unified Diffusion Bridge framework, UniDB, is proposed and modeled as an optimal control problem based on SOC, deriving a closed-form optimal controller and corresponding forward/backward SDEs;
Unifews: You Need Fewer Operations for Efficient Graph Neural Networks
Ningyi Liao (Nanyang Technological University), Siqiang Luo (Nanyang Technological University)
Computational EfficiencyGraph Neural NetworkGraph
🎯 What it does: The UNIFEWS framework is proposed, which unifies the item-wise sparsification of graph structures and model weights to significantly reduce the computational load of GNNs;
Unified Analysis of Continuous Weak Features Learning with Applications to Learning from Missing Data
Kosuke Sugiyama (Waseda University), Masato Uchida (Waseda University)
OptimizationData-Centric LearningTabular
🎯 What it does: This paper proposes a unified Continuous Weak Features Learning (cWFL) theoretical framework and provides a generalization error analysis.
Unified Breakdown Analysis for Byzantine Robust Gossip
Renaud Gaucher (Institut Polytechnique de Paris), Hadrien Hendrikx (CNRS)
Image
🎯 What it does: This paper proposes a unified robust decentralized gossip framework to address the issues of average consensus and distributed SGD convergence under Byzantine attacks in sparse communication networks.
Unified K-Means Clustering with Label-Guided Manifold Learning
Qianqian Wang (Xidian University), Quanxue Gao (Xidian University)
OptimizationTabularBenchmark
🎯 What it does: A unified, decentralized, balanced, label-guided manifold learning version of the K-Means clustering framework is proposed.
Unified Screening for Multiple Diseases
Yiğit Narter (Bilkent University), Cem Tekin (Bilkent University)
OptimizationTabularTime SeriesBiomedical Data
🎯 What it does: A unified disease screening model has been developed, considering competing risks, budget constraints, and diagnostic errors, aimed at optimizing multi-disease screening decisions under limited resources.
Uniform Mean Estimation for Heavy-Tailed Distributions via Median-of-Means
Mikael Møller Høgsgaard (Aarhus University), Andrea Paudice (Aarhus University)
🎯 What it does: Analyzed the sample complexity problem of uniformly estimating a class of function means using the median of means (MoM) estimator in the presence of heavy-tailed distributions, and provided a general upper bound on sample complexity.
Unifying 2D and 3D Vision-Language Understanding
Ayush Jain (Carnegie Mellon University), Katerina Fragkiadaki (Carnegie Mellon University)
Object DetectionSegmentationTransformerVision Language ModelImagePoint Cloud
🎯 What it does: Proposes the UniVLG integrated 2D/3D visual language model, which jointly trains 2D and 3D data and achieves 3D object localization and question answering through a language-conditioned mask decoder.
Unifying Knowledge from Diverse Datasets to Enhance Spatial-Temporal Modeling: A Granularity-Adaptive Geographical Embedding Approach
Zhigaoyuan Wang, Hengshu Zhu (Chinese Academy of Sciences)
Graph Neural NetworkTime Series
🎯 What it does: Proposes the Segment Quadtree Geographical Embedding Framework (SQGEF), which utilizes segment quadtrees to unify the representation of geographic data with different granularities and structures, addressing the issue of data scarcity in scientific research.
Unifying Specialized Visual Encoders for Video Language Models
Jihoon Chung (Princeton University), Olga Russakovsky (Princeton University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: Designed and implemented MERV (Multi-Encoder Video Representation), which first performs spatiotemporal alignment and projection of features from four visual encoders: DINOv2, ViViT, SigLIP, and LanguageBind, and then uses cross-attention for dynamic weighted fusion to obtain a unified video representation for inference by the LLaMA-2 7B language model.
UniMate: A Unified Model for Mechanical Metamaterial Generation, Property Prediction, and Condition Confirmation
Wangzhi Zhan (Virginia Polytechnic Institute and State University), Dawei Zhou (Virginia Polytechnic Institute and State University)
GenerationData SynthesisOptimizationTransformerDiffusion modelScore-based ModelAuto EncoderMeshBenchmarkPhysics Related
🎯 What it does: A unified model called UNIMATE is proposed, capable of simultaneously completing three major tasks: topological generation of mechanical metamaterials, mechanical property prediction, and condition confirmation.
UniMC: Taming Diffusion Transformer for Unified Keypoint-Guided Multi-Class Image Generation
Qin Guo (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisPose EstimationTransformerDiffusion modelImage
🎯 What it does: The UNIMC framework is proposed, which implements multi-category and multi-instance human and animal keypoint-guided image generation on the Diffusion Transformer using a unified keypoint encoder and temporal-aware keypoint modulator, and constructs the HAIG-2.9M large-scale keypoint dataset.
UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design
Xiangzhe Kong (Tsinghua University), Yang Liu (Tsinghua University)
GenerationDrug DiscoveryGraph Neural NetworkDiffusion modelAuto EncoderGraphBiomedical Data
🎯 What it does: A unified three-dimensional molecular generation framework called UniMoMo is proposed, which generates various types of molecular conjugates (peptides, antibodies, small molecules) using a single model.
UniSim: A Unified Simulator for Time-Coarsened Dynamics of Biomolecules
Ziyang Yu (Tsinghua University), Yang Liu (Tsinghua University)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkBiomedical Data
🎯 What it does: We propose UniSim, a deep learning simulator that can rapidly generate molecular trajectories across biomolecular domains using a time-coarsening method.
Unisolver: PDE-Conditional Transformers Towards Universal Neural PDE Solvers
Hang Zhou (Tsinghua University), Mingsheng Long (Tsinghua University)
TransformerLarge Language ModelTime SeriesPhysics Related
🎯 What it does: Developed Unisolver, a Transformer-based PDE conditioning model capable of unifying the handling of various types of partial differential equations.
Unisoma: A Unified Transformer-based Solver for Multi-Solid Systems
Shilong Tao (Peking University), Yunhuai Liu (Peking University)
TransformerPhysics Related
🎯 What it does: The Unisoma model is proposed, which addresses the physical simulation of multi-solid systems using an explicitly modeled Transformer framework.
Universal Approximation of Mean-Field Models via Transformers
Shiba Biswal (Los Alamos National Laboratory), Rishi Sonthalia (Boston College)
TransformerTime SeriesSequential
🎯 What it does: This paper studies the use of Transformer to approximate the mean field dynamics of particle systems, proposing to elevate finite-dimensional Transformers to expectation mappings of probability measures, and providing theoretical universal approximation and error bounds.
Universal Approximation Theorem of Deep Q-Networks
Qian Qi (Peking University)
Reinforcement LearningStochastic Differential Equation
🎯 What it does: This paper constructs a continuous-time framework and proves that the Deep Q-Network (DQN) can approximate the optimal action value function to any desired accuracy on any compact set under certain regularity conditions, and provides convergence guarantees for its training process.
Universal Biological Sequence Reranking for Improved De Novo Peptide Sequencing
Zijie Qiu (Fudan University), Siqi Sun (Fudan University)
Protein Structure PredictionTransformerBiomedical Data
🎯 What it does: This paper presents RankNovo—a deep learning framework based on list-based reordering to improve the accuracy of de novo protein sequencing.
Universal Length Generalization with Turing Programs
Kaiying Hou (Harvard University), Eran Malach (Harvard University)
TransformerSequential
🎯 What it does: This paper proposes a general Scratchpad strategy called Turing Programs, which can decompose any algorithmic task into a series of small copying and modification steps, thereby achieving length generalization.
Universal Neural Optimal Transport
Jonathan Geuter (Harvard University), Vaios Laschos (Weierstrass Institute)
GenerationOptimizationGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes UNOT, which can predict the regularized OT distance and transport plans between discrete measures at different resolutions.
Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment
Harrish Thasarathan (York University), Konstantinos G. Derpanis (University of Toronto)
Explainability and InterpretabilityRepresentation LearningAuto EncoderImage
🎯 What it does: A Unified Sparse Autoencoder (USAE) is designed and trained to discover and align interpretable concepts across multiple visual models.
Unlocking Post-hoc Dataset Inference with Synthetic Data
Bihe Zhao (CISPA Helmholtz Center for Information Security), Adam Dziedzic
Data SynthesisAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A method for dataset inference using synthetic data is proposed.
Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection
Peipeng Yu (Jinan University), Chip Hong Chang
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: A deepfake detection framework based on large visual language models is proposed, combining a knowledge-guided forgery detector, a forgery prompt learner, and LLM to achieve interpretable, generalizable, and multi-turn dialogue-supported detection.
Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective
Junze Deng (Ohio State University), Ness Shroff (Ohio State University)
Image
🎯 What it does: This paper proposes and theoretically analyzes two replay strategies in continual learning—concurrent replay and sequential replay—and further designs a hybrid replay method, exploring the impact of task similarity on the optimal strategy.
Unlocking the Power of SAM 2 for Few-Shot Segmentation
Qianxiong Xu (Nanyang Technological University), Rui Zhao (SenseTime Research)
SegmentationImage
🎯 What it does: This paper transfers the object matching capability of SAM 2 to few-shot semantic segmentation, proposing three main modules: pseudo prompt generation, iterative memory refinement, and support-calibrated attention to accommodate heterogeneous object matching.
unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning
Yafei YANG, Bo Yang (Hong Kong Polytechnic University)
Object DetectionSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A two-stage unsupervised multi-object segmentation framework called unMORE is proposed, which first learns a three-layer object center-boundary representation on ImageNet, and then automatically locates and segments multiple objects in a single image through a network-free inference module.
Unnatural Languages Are Not Bugs but Features for LLMs
Keyu Duan (National University of Singapore), Michael Qizhe Shieh (National University of Singapore)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study systematically verifies the comprehensibility and transferability of non-natural language (text that is not human-readable but has semantics for LLMs) for large language models, and proposes a search method based on multi-model gradient sampling to generate equivalent non-natural text.
Unpaired Point Cloud Completion via Unbalanced Optimal Transport
Taekyung Lee (Seoul National University), Myungjoo Kang (Seoul National University)
GenerationData SynthesisOptimizationGenerative Adversarial NetworkPoint Cloud
🎯 What it does: A model for unpaired point cloud completion based on Unbalanced Optimal Transport (UOT) theory, called UOT-UPC, is proposed, which models the unpaired completion task as an UOT Map, learning the mapping from incomplete point clouds to complete point clouds.
Unraveling the Interplay between Carryover Effects and Reward Autocorrelations in Switchback Experiments
Qianglin Wen (Yunnan University), Hongtu Zhu (University of North Carolina at Chapel Hill)
Reinforcement LearningTabularTime Series
🎯 What it does: This paper systematically compares different Switchback designs (including daily alternation, alternation every m steps, etc.) in a Markov environment and evaluates the estimation accuracy of various reinforcement learning (RL) estimators (model-based, LSTD, DRL) for the average treatment effect (ATE);
Unsupervised Learning for Class Distribution Mismatch
Pan Du (Renmin University of China), Yang You (National University of Singapore)
ClassificationGenerationDiffusion modelImage
🎯 What it does: A completely unsupervised framework UCDM is proposed, which utilizes diffusion models to generate positive and negative samples by adding and removing categories from unlabeled images, thus enabling training in scenarios where class distributions do not match.
Unveiling AI's Blind Spots: An Oracle for In-Domain, Out-of-Domain, and Adversarial Errors
Shuangpeng Han (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)
ClassificationKnowledge DistillationAdversarial AttackTransformerSupervised Fine-TuningImage
🎯 What it does: A framework for error prediction across different error types and models was constructed by training a mentor model to predict the errors of the trained model on ID, OOD, and adversarial samples.
Unveiling Markov heads in Pretrained Language Models for Offline Reinforcement Learning
Wenhao Zhao (Renmin University of China), Jiang Bian (Microsoft Research Asia)
TransformerLarge Language ModelReinforcement LearningSequential
🎯 What it does: This paper studies the transfer effects of pre-trained language models (PLM) in offline reinforcement learning, discovering and defining the 'Markov head', proving that it remains unchanged after pre-training and is difficult to alter through fine-tuning. Based on this, it proposes the Mixture of Attention (MoA) adaptive planning mechanism GPT2-DTMA, which significantly enhances the model's performance in both short-term and long-term environments.
UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent
Jianke Zhang (Tsinghua University), Jianyu Chen (Tsinghua University)
Robotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelGenerative Adversarial NetworkMultimodality
🎯 What it does: A unified visual-language-action model UP-VLA is proposed, which enhances the robot's understanding of semantic and spatial details through multimodal understanding and future prediction pre-training.
Upcycling Text-to-Image Diffusion Models for Multi-Task Capabilities
Ruchika Chavhan (Samsung AI Center), Sourav Bhattacharya (Samsung AI Center)
Image TranslationRestorationGenerationSuper ResolutionMixture of ExpertsDiffusion modelImageText
🎯 What it does: The pre-trained text-to-image diffusion model is 'upcycled' by splitting the FFN layer into multiple small experts and using dynamic routing, expanding it into a multi-task image generation model (image editing, super-resolution, inpainting, etc.).
Update Your Transformer to the Latest Release: Re-Basin of Task Vectors
Filippo Rinaldi (AImageLab, University of Modena and Reggio Emilia), Angelo Porrello (AImageLab, University of Modena and Reggio Emilia)
TransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a technique that allows for the transfer of task vectors generated from existing fine-tuning to a new version of the model based on the Transformer architecture, without retraining and without any task data; the TransFusion algorithm is constructed through a two-layer permutation of model weights.
Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting
Sunny Sanyal (University of Texas at Austin), Sujay Sanghavi (University of Texas at Austin)
ClassificationRecognitionSupervised Fine-TuningImageText
🎯 What it does: A sample weighting scheme based on pre-trained model loss is proposed to alleviate the phenomenon of catastrophic forgetting that occurs during fine-tuning.
Validating Mechanistic Interpretations: An Axiomatic Approach
Nils Palumbo (University of Wisconsin), Somesh Jha (University of Wisconsin)
Explainability and InterpretabilityTransformerTabular
🎯 What it does: A set of axiomatic frameworks is proposed for the formal verification of mechanistic interpretation of neural networks, and the effectiveness of this framework is tested in two cases (2-SAT and modular addition).
Value-Based Deep RL Scales Predictably
Oleh Rybkin (University of California, Berkeley), Aviral Kumar (Carnegie Mellon University)
Computational EfficiencyHyperparameter SearchReinforcement LearningSequential
🎯 What it does: This paper studies the predictability of off-policy deep reinforcement learning during scale-up, establishing the Pareto front between data requirements, computational requirements, and update-to-data ratio (UTD), and predicting optimal hyperparameters and performance under given computational/data budgets through this front.
Variance as a Catalyst: Efficient and Transferable Semantic Erasure Adversarial Attack for Customized Diffusion Models
Jiachen Yang (Sun Yat-sen University), Fangjun Huang (Sun Yat-sen University)
Adversarial AttackDiffusion modelAuto EncoderImage
🎯 What it does: Adversarial attacks for semantic erasure on LDM, utilizing VAE noise variance to achieve complete erasure of images.
Variance-Reduced Forward-Reflected-Backward Splitting Methods for Nonmonotone Generalized Equations
Quoc Tran-Dinh (University of North Carolina at Chapel Hill)
OptimizationTabular
🎯 What it does: A variance-reduced forward-reflective-backward splitting method for non-monotone generalized equations is proposed, capable of solving roots/optimal points in large-scale finite and random sample scenarios.
Variational Control for Guidance in Diffusion Models
Kushagra Pandey (University of California), Stephan Mandt (University of California)
GenerationOptimizationDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes the Diffusion Trajectory Matching (DTM) framework, treating the guidance of diffusion models as a variational control problem. Under this framework, the Non-Linear Diffusion Trajectory Matching (NDTM) algorithm is designed to guide pre-trained diffusion models to meet terminal constraints without retraining the model.
Variational Counterfactual Intervention Planning to Achieve Target Outcomes
Xin Wang (University of Science and Technology of China), Huanhuan Chen (University of Science and Technology of China)
Recurrent Neural NetworkTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Proposes Variational Counterfactual Intervention Planning (VCIP) to find intervention sequences that achieve specified target outcomes over time series.
Variational Learning of Fractional Posteriors
Kian Ming A. Chai (DSO National Laboratories), Edwin V. Bonilla (CSIRO Data61)
Auto EncoderImageTabular
🎯 What it does: A new one-parameter variational objective is proposed, which can lower bound data evidence and estimate approximate score posteriors. The framework is extended to support hierarchies and Bayesian posteriors, providing flexible tools for probabilistic modeling.
Variational Phylogenetic Inference with Products over Bipartitions
Evan Sidrow (Simon Fraser University), Lloyd T Elliott
Reinforcement LearningTabularBiomedical Data
🎯 What it does: This paper proposes a variational Bayesian inference method for ultrametric trees, VIPR, by modeling the distribution of pairwise coalescent times and using single-linkage clustering to map to tree structures, resulting in a differentiable variational distribution.
Variational Rectified Flow Matching
Pengsheng Guo (Apple), Alex Schwing
GenerationData SynthesisFlow-based ModelRectified FlowAuto EncoderImage
🎯 What it does: A Variational Rectified Flow Matching framework is proposed, utilizing latent variables and variational inference to capture the multimodality of the velocity field in flow matching, thereby addressing the issue of twisted flow trajectories caused by averaging in traditional regular flow matching.
VCT: Training Consistency Models with Variational Noise Coupling
Gianluigi Silvestri (OnePlanet Research Center), Yuki Mitsufuji (Sony Group Corporation)
GenerationFlow-based ModelAuto EncoderImage
🎯 What it does: This paper proposes a framework that introduces variational noise coupling in consistency training, utilizing an encoder to learn data-related noise distributions and performing joint training through KL regularization.
Vector Grimoire: Codebook-based Shape Generation under Raster Image Supervision
Marco Cipriano (Hasso Plattner Institute), Gerard de Melo (Hasso Plattner Institute)
GenerationData SynthesisTransformerAuto EncoderImageText
🎯 What it does: This paper presents GRIMOIRE, a two-stage SVG generation framework based on raster image supervision, capable of generating, completing, and editing vector graphics from natural language text.
VerbalTS: Generating Time Series from Texts
Shuqi Gu (ShanghaiTech University), Kan Ren (ShanghaiTech University)
GenerationData SynthesisDiffusion modelAuto EncoderTextTime Series
🎯 What it does: This paper studies the task of generating time series from natural language descriptions and proposes the VERBALTS framework.
Verification Learning: Make Unsupervised Neuro-Symbolic System Feasible
Lin-Han Jia (Nanjing University), Yu-Feng Li (Nanjing University)
OptimizationImage
🎯 What it does: Proposes the Verification Learning (VL) framework, which utilizes unlabeled data and verification functions to achieve neuro-symbolic learning, and efficiently solves constraint optimization problems through the Dynamic Combination Sorting (DCS) algorithm.
VersaPRM: Multi-Domain Process Reward Model via Synthetic Reasoning Data
Thomas Zeng (University of Wisconsin-Madison), Kangwook Lee (University of Wisconsin-Madison)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper presents VersaPRM, a multi-domain process reward model trained through synthetic reasoning data to enhance the reasoning performance of LLMs in non-mathematical domains.
Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations
Yucheng Hu (Tsinghua University), Jianyu Chen (Tsinghua University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelVideo
🎯 What it does: A general robot strategy VPP based on a video prediction model is proposed, utilizing the predicted representation of the video diffusion model directly as visual encoding and learning inverse dynamics.
Video-Enhanced Offline Reinforcement Learning: A Model-Based Approach
Minting Pan (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
Autonomous DrivingReinforcement LearningWorld ModelVideo
🎯 What it does: This paper proposes an offline reinforcement learning method for video enhancement called VeoRL, which constructs a dual-branch interactive world model by extracting discrete behavior abstractions from unlabeled internet videos to improve the performance of offline visual RL.
video-SALMONN-o1: Reasoning-enhanced Audio-visual Large Language Model
Guangzhi Sun (Tsinghua University), Chao Zhang (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodalityBenchmarkAudio
🎯 What it does: This paper proposes and implements video-SALMONN-o1, an open-source audio-visual large language model, and enhances its performance in general video understanding tasks through inference augmentation techniques.
VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models
Hila Chefer (GenAI, Meta), Shelly Sheynin (GenAI, Meta)
GenerationData SynthesisTransformerDiffusion modelOptical FlowVideoBenchmark
🎯 What it does: The VideoJAM framework is proposed, which significantly enhances the motion coherence of videos by incorporating a joint appearance-motion representation into video generation models.
VideoRoPE: What Makes for Good Video Rotary Position Embedding?
Xilin Wei (Fudan University), Dahua Lin (Shanghai AI Laboratory)
RetrievalTransformerVision Language ModelVideo
🎯 What it does: A new video position encoding scheme, VideoRoPE, is designed and evaluated to address four key attributes of RoPE in videos: 3D structure, frequency allocation, spatial symmetry, and temporal index scaling.
VinePPO: Refining Credit Assignment in RL Training of LLMs
Amirhossein Kazemnejad (Mila), Nicolas Le Roux (Mila)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The VinePPO method is proposed, utilizing the resettable feature of the language environment to perform Monte Carlo estimation on intermediate states, improving credit assignment in RL training.
Vintix: Action Model via In-Context Reinforcement Learning
Andrei Polubarov (Artificial Intelligence Research Institute), Vladislav Kurenkov (Artificial Intelligence Research Institute)
Knowledge DistillationRobotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: A general action model Vintix has been constructed, which is cross-domain and capable of self-correction and adaptation through trial-and-error learning during inference, utilizing the framework of In-context Reinforcement Learning (ICRL);
VIP: Vision Instructed Pre-training for Robotic Manipulation
Zhuoling Li (Hong Kong University), Hengshuang Zhao (Hong Kong University)
Robotic IntelligenceTransformerImage
🎯 What it does: This paper proposes Visual Instruction Pre-training (VIP) and a Transformer-based VIRT model, which trains robots to generate correct actions for unseen tasks by using future images and sparse point flows as visual instructions.
Vision Graph Prompting via Semantic Low-Rank Decomposition
Zixiang Ai (Peking University), Jiahuan Zhou (Peking University)
ClassificationGraph Neural NetworkPrompt EngineeringGraph
🎯 What it does: A parameter-efficient visual prompting framework for visual graph neural networks (ViG) called Vision Graph Prompting (VGP) is proposed, which achieves prompts at the graph, edge, and node levels through low-rank semantic decomposition.
Vision-Language Model Selection and Reuse for Downstream Adaptation
Hao-Zhe Tan (Nanjing University), Lan-Zhe Guo (Nanjing University)
ClassificationDomain AdaptationTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper proposes the MLL (Model Label Learning) paradigm, which constructs a semantic graph to assign labels to pre-trained visual-language models (VLM), thereby enabling model selection and reuse for downstream tasks.
Vision-Language Models Create Cross-Modal Task Representations
Grace Luo (University of California), Amir Bar (University of California)
TransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This study investigates and verifies the shared task vectors of visual-language models (VLM) across different modalities, demonstrating that these vectors can transfer across modalities and models, significantly enhancing task completion performance.
VisionTS: Visual Masked Autoencoders Are Free-Lunch Zero-Shot Time Series Forecasters
Mouxiang Chen (Zhejiang University), Chenghao Liu (Salesforce Research Asia)
Auto EncoderTime SeriesBenchmark
🎯 What it does: Applying the pre-trained Masked Autoencoder (MAE) for visual data to time series forecasting by mapping one-dimensional sequences to two-dimensional images and masking the prediction window, achieving zero-shot inference.
Visual Abstraction: A Plug-and-Play Approach for Text-Visual Retrieval
Guofeng Ding (Sichuan University), Xi Peng (Sichuan University)
RetrievalTransformerLarge Language ModelVision Language ModelImageVideoTextChain-of-Thought
🎯 What it does: A scheme is proposed to abstract visual content into natural language descriptions during the testing phase and perform text retrieval (VISA), thereby enhancing text-visual retrieval performance.
Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning
RINA BAO, Yangming Ou (Boston Children's Hospital and Harvard Medical School)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: A benchmark dataset called HIE-Reasoning and a model named CGoT (Clinical Graph of Thought) were constructed to evaluate the performance of large visual-language models in professional-level medical reasoning.
Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models
Mingi Jung (Seoul National University), Sungroh Yoon (Seoul National University)
RecognitionGenerationTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: A training-free visual attention recalibration method called SPARC is proposed, aimed at improving the accuracy and recall of multimodal large language models when generating detailed image descriptions.
Visual Autoregressive Modeling for Image Super-Resolution
Yunpeng Qu (Tsinghua University), Chao Zhou (Kuaishou Technology)
RestorationSuper ResolutionTransformerDiffusion modelImage
🎯 What it does: A super-resolution model based on visual autoregression, VARSR, is proposed.
Visual Generation Without Guidance
Huayu Chen (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationDiffusion modelImage
🎯 What it does: A Guidance-Free Training (GFT) method is proposed to eliminate the dependence on dual model guidance (CFG) in visual generative models, directly training a single model to achieve low-temperature sampling.
Visual Graph Arena: Evaluating Visual Conceptualization of Vision and Multimodal Large Language Models
Zahra Babaiee (Vienna University of Technology), Radu Grosu
Graph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningMultimodalityGraphBenchmark
🎯 What it does: The Visual Graph Arena benchmark is proposed, which includes six graph-based visual reasoning tasks to evaluate AI's conceptual reasoning ability under different visual representations.
ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy
Kian Kenyon-Dean (Recursion), Oren Kraus (Recursion)
Representation LearningDrug DiscoveryTransformerAuto EncoderContrastive LearningImageBiomedical Data
🎯 What it does: A framework based on self-supervised learning is proposed for pre-training large-scale cell microscopy images, model expansion, and selecting the best representations through intermediate layer probing, resulting in biologically meaningful representations.
Volume Optimality in Conformal Prediction with Structured Prediction Sets
Chao Gao (University of Chicago), Aravindan Vijayaraghavan (Northwestern University)
Tabular
🎯 What it does: A method for quantile prediction sets based on dynamic programming is proposed, achieving a distribution-free optimal volume prediction set in the joint set of k intervals.
Volume-Aware Distance for Robust Similarity Learning
Shuo Chen (Nanjing University), Jian Yang (Nanjing University of Science and Technology)
ClassificationRetrievalRepresentation LearningContrastive LearningImageTextGraph
🎯 What it does: This paper studies a volume-aware distance (VAD) model that treats instances as data spheres with predictable volumes and applies it to both supervised and unsupervised similarity learning tasks, significantly enhancing the model's generalization performance.
Voronoi-grid-based Pareto Front Learning and Its Application to Collaborative Federated Learning
Mengmeng Chen (Beijing University of Posts and Telecommunications), Han Yu (Nanyang Technological University)
OptimizationFederated LearningTabularBenchmark
🎯 What it does: This paper proposes a Pareto front learning framework PHN-HVVS based on Voronoi grid partitioning and genetic algorithms for high-dimensional multi-objective optimization, applied to collaborative federated learning.
VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians
Pengchong Hu (Wayne State University), Zhizhong Han (Wayne State University)
Pose EstimationOptimizationComputational EfficiencyGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: This paper proposes an RGB-D SLAM system based on View-Tied 3D Gaussians, combined with a new tracking and mapping strategy to achieve efficient mapping and localization in large-scale scenes.
Vulnerability-Aware Alignment: Mitigating Uneven Forgetting in Harmful Fine-Tuning
Liang CHEN, Kam-Fai Wong (Chinese University of Hong Kong)
OptimizationData-Centric LearningSupervised Fine-TuningText
🎯 What it does: A data-forgetfulness-based alignment method VAA is proposed, which balances forgettable and memorable samples through group learning to enhance robustness against harmful fine-tuning.
Wait-Less Offline Tuning and Re-solving for Online Decision Making
Jingruo Sun (Stanford University), Yinyu Ye (Hong Kong University of Science and Technology)
OptimizationTabular
🎯 What it does: Combining the fundamentals of LP and first-order subgradient methods, a parallel multi-stage algorithm is proposed to achieve instantaneous decision-making without waiting in online linear programming.
Wasserstein Flow Matching: Generative Modeling Over Families of Distributions
Doron Haviv (Memorial Sloan Kettering Cancer Center), Brandon Amos (Meta AI)
GenerationData SynthesisTransformerFlow-based ModelPoint CloudBiomedical Data
🎯 What it does: Developed Wasserstein Flow Matching (WFM), elevating flow matching methods from Euclidean space to distribution space, achieving a generative model from distribution to distribution;
Wasserstein Policy Optimization
David Pfau (Google DeepMind), Hado van Hasselt (Google DeepMind)
OptimizationReinforcement LearningPhysics Related
🎯 What it does: A Wasserstein gradient flow-based actor-critic algorithm, WPO, is proposed for policy optimization in continuous action spaces.
Watch Out Your Album! On the Inadvertent Privacy Memorization in Multi-Modal Large Language Models
Tianjie Ju (Shanghai Jiao Tong University), Gongshen Liu (Shanghai Jiao Tong University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: This study investigates whether multimodal large models unintentionally memorize privacy content unrelated to tasks during the fine-tuning process and proposes a detection method based on gradient similarity and hierarchical probing.
WATCH: Adaptive Monitoring for AI Deployments via Weighted-Conformal Martingales
Drew Prinster (Johns Hopkins University), Suchi Saria (Johns Hopkins University)
Anomaly DetectionImageTabular
🎯 What it does: This paper proposes and implements Weighted Conformal Test Martingales (WCTMs) and the WATCH framework for online monitoring of AI/ML systems, which can adapt to slight distribution drifts, quickly alert in the case of severe drifts, and perform root cause diagnosis through parallel X-CTM.
WAVE: Weighted Autoregressive Varying Gate for Time Series Forecasting
Jiecheng Lu (Georgia Institute of Technology), Shihao Yang (Georgia Institute of Technology)
TransformerTime Series
🎯 What it does: A new WAVE (Weighted Autoregressive Varying Gate) attention mechanism is proposed, which integrates autoregressive (AR) and moving average (MA) structures into the Transformer for time series forecasting;
Weak-to-Strong Generalization Even in Random Feature Networks, Provably
Marko Medvedev (University of Chicago), Nathan Srebro (TTIC)
Knowledge DistillationConvolutional Neural Network
🎯 What it does: This paper proves that in two-layer random feature networks, even if the student model only uses labels generated by the teacher and is trained from scratch, the student can still significantly outperform the teacher model without pre-training or special priors, and weak-to-strong generalization can be achieved through early stopping.
Weak-to-Strong Jailbreaking on Large Language Models
Xuandong Zhao (University of California Santa Barbara), William Yang Wang (University of California Santa Barbara)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes a weak-to-strong Jailbreak attack method guided by a small model, which can induce a safely aligned LLM to generate harmful text with just one forward inference during the inference phase.
Weakly Supervised Anomaly Detection via Dual-Tailed Kernel
Walid Durani (Ludwig Maximilian University of Munich), Christian Böhm (University of Vienna)
Anomaly DetectionTabularBenchmark
🎯 What it does: A weakly supervised anomaly detection framework WSAD-DT is proposed, which distinguishes between abnormal and normal samples by learning the two class centers in the feature space using a dual-tail kernel.
Weakly-Supervised Contrastive Learning for Imprecise Class Labels
Zi-Hao Zhou (Southeast University), Min-Ling Zhang (Southeast University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningImage
🎯 What it does: A weakly supervised contrastive learning framework is proposed, which constructs continuous positive and negative samples through semantic similarity, effectively utilizing imprecise labels.
WeGeFT: Weight‑Generative Fine‑Tuning for Multi‑Faceted Efficient Adaptation of Large Models
Chinmay Savadikar (North Carolina State University), Tianfu Wu (North Carolina State University)
Domain AdaptationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: A new parameter-efficient fine-tuning framework called WeGeFT is proposed, which utilizes frozen pre-trained weights to generate fine-tuning weight residuals through two shared low-rank linear mappings, achieving efficient adaptation of Transformer models.
Weight matrices compression based on PDB model in deep neural networks
Xiaoling Wu (Southern University of Science and Technology), Zeng Li (Southern University of Science and Technology)
CompressionImageText
🎯 What it does: This paper proposes the Population Double Bulk (PDB) model and the corresponding weight matrix compression algorithm, which can automatically determine the boundary between noise and information and compress the network without introducing hyperparameters.
Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win
Lorenz Kummer (University of Vienna), Nils Morten Kriege
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkGraph
🎯 What it does: This paper studies the Lottery Ticket Hypothesis (LTH) in sparse initialized Graph Neural Networks (GNNs), proving that it is still possible to find trainable sparse sub-networks while maintaining sufficient expressive power (i.e., comparable to 1-WL);
WGFormer: An SE(3)-Transformer Driven by Wasserstein Gradient Flows for Molecular Ground-State Conformation Prediction
Fanmeng Wang (Renmin University of China), Hongteng Xu (Renmin University of China)
OptimizationDrug DiscoveryTransformerAuto EncoderGraph
🎯 What it does: A SE(3)-Transformer WGFormer based on Wasserstein gradient flow is designed to predict the ground state conformation of molecules from low-quality conformations.
What can large language models do for sustainable food?
Anna Thomas, Kristina Gligorić (Stanford University)
Recommendation SystemOptimizationTransformerLarge Language ModelPrompt EngineeringTextAgriculture Related
🎯 What it does: This paper explores the application of large language models (LLMs) in the field of sustainable food, constructing four types of tasks (experimental design, menu design, sensory feature prediction, and recipe preference prediction), and evaluates the performance of six mainstream LLMs on these tasks.
What Do Learning Dynamics Reveal About Generalization in LLM Mathematical Reasoning?
Katie Kang (University of California Berkeley), Aviral Kumar (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the relationship between learning dynamics during LLM fine-tuning and the generalization of reasoning tasks, proposing and validating a pre-memory training accuracy metric.
What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models
Keyon Vafa (Harvard University), Sendhil Mullainathan (Massachusetts Institute of Technology)
Supervised Fine-TuningWorld ModelTime SeriesSequentialPhysics Related
🎯 What it does: This paper proposes an 'inductive bias probe' method, which evaluates whether a model has learned a world model by fine-tuning the base model on a synthetic dataset consistent with the hypothesized world model and observing its external inference behavior.
What If We Recaption Billions of Web Images with LLaMA-3?
Xianhang Li (University of California), Cihang Xie (University of California)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Using LLaMA-3 enhanced LLaVA for large-scale re-annotation of 1.3 billion image-text pairs from web scraping, generating the Recap-DataComp-1B dataset.
What Limits Bidirectional Model's Generative Capabilities? A Uni-Bi-Directional Mixture-of-Expert Method For Bidirectional Fine-tuning
Zuchao Li (Wuhan University), Liu Guoming
GenerationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: This paper studies a hybrid expert model UBMoE-LLM that combines unidirectional generation and bidirectional embedding capabilities to maintain the generative performance of large language models while improving text embedding effectiveness.
What Limits Virtual Agent Application? OmniBench: A Scalable Multi-Dimensional Benchmark for Essential Virtual Agent Capabilities
Wendong Bu (Zhejiang University), Yueting Zhuang (Zhejiang University)
Graph Neural NetworkLarge Language ModelMultimodalityGraphBenchmark
🎯 What it does: Proposes OmniBench, a controllable complexity graph structure task generation and evaluation framework for virtual agents.
What Makes a Good Feedforward Computational Graph?
Alex Vitvitskyi (Google DeepMind), Petar Veličković
RetrievalOptimizationGraph Neural NetworkGraphSequential
🎯 What it does: This paper studies the structural characteristics of feedforward computation graphs (directed graphs without feedback edges) and proposes two metrics for evaluating their information propagation quality: mixing time and minimax fidelity.
What makes an Ensemble (Un) Interpretable?
Shahaf Bassan (Hebrew University of Jerusalem), Guy Katz
Explainability and Interpretability
🎯 What it does: A formal computational complexity analysis of interpretability in ensemble models is conducted, studying the interpretability difficulty of various types of explanations under different base models and structural parameters.
What Makes In-context Learning Effective for Mathematical Reasoning
Jiayu Liu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
TransformerLarge Language ModelText
🎯 What it does: This paper quantifies the impact of examples on the mathematical reasoning performance of large language models in context learning through theoretical analysis and proposes an example selection method called LMS3 based on semantic similarity and reasoning stability.