ICML 2026 Papers — Page 64
International Conference on Machine Learning · 6554 papers
VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics
Yichen Gong (Ant Group), Shuheng Shen (Ant Group)
TransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs VenusBench-Mobile, an online mobile GUI agent evaluation benchmark, and provides an annotation scheme based on capability diagnosis.
VEQ: Modality-Adaptive Quantization for MoE Vision-Language Models
Guangshuo Qin (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a post-training quantization framework called VEQ for hybrid expert vision-language models, which can simultaneously address modality differences and expert heterogeneity.
VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models
Qilin Liao (Purdue University), Ruqi Zhang (Purdue University)
Adversarial AttackTransformerReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the VERA-V framework, which transforms the Jailbreak task of vision-language models into joint posterior inference of text-image pairs, thereby enabling cross-modal red team attacks.
Verbalized Bayesian Persuasion
Wenhao Li (Tongji University), Baoxiang Wang (Chinese University of Hong Kong)
Recommendation SystemAutonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes the Verbalized Bayesian Persuasion (VBP) framework, mapping the traditional Bayesian persuasion problem into a mediator-enhanced game centered on text, and utilizing large language models (LLMs) for strategy generation and game solving.
Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
Jiayi Zhang (Northeastern University), Weiyan Shi (Northeastern University)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes and verifies a training-agnostic prompting strategy called Verbalized Sampling (VS), aiming to eliminate mode collapse in aligned large language models (LLMs) and improve the diversity and quality of model outputs, through theoretical analysis, empirical verification, and large-scale experiments.
Verified SHAP: Provable Bounds for Exact Shapley Values of Neural Networks
David Boetius (University of Konstanz), Tobias Sutter (University of St.Gallen)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkGraph Neural NetworkTransformerImageTabularTime Series
🎯 What it does: Developed and implemented VERISHAP, an algorithm that computes SHAP values of neural networks by leveraging neural network verification techniques (such as branch-and-bound, bound propagation), and can provide provable upper/lower bounds at arbitrary precision, ultimately recovering exact SHAP values.
Verifying Meta-Awareness via Predictive Rewards in Reasoning Models
Yoonjeon Kim (KAIST), Eunho Yang (KAIST)
Meta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the MAPR framework, which verifies and enhances meta-awareness on large language models through self-predictive rewards, enabling the model to predict problem difficulty, output length, and concepts used during reasoning, and to self-calibrate based on the prediction results.
VeriSimpl: Robust Optimization Modeling from Natural Language using Simplification-based Verification
Sumaya Abdul Rahman (Hamad Bin Khalifa University), Mohammad Raza (Qatar Computing Research Institute)
OptimizationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a robust verification framework called VERISIMPL that utilizes a solver to generate simplified diagnostic queries and combines a large language model (LLM) for natural language to optimization model verification.
VeRO: A Harness for Agents to Optimize Agents
Varun Ursekar (Scale AI), Samuel Marc Denton (Scale AI)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the VERO external harness for automatically improving LLM agents through code modifications, and constructs the VERO-BENCH evaluation suite to conduct systematic experiments on the agent optimization process.
Very Efficient Listwise Multimodal Reranking for Long Documents
Yiqun Sun (Magellan Technology Research Institute), Lawrence B. Hsieh (Magellan Technology Research Institute)
RetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: To address the visual center retrieval task for long documents, we propose an efficient list-wise multi-modal re-ranking model called ZipRerank, which can sort all candidate pages in a single LLM forward pass.
VFMF: Dense Forecasting by Generating Foundation Model Features
Gabrijel Boduljak (University of Oxford), Andrea Vedaldi (University of Oxford)
Image TranslationSegmentationGenerationDepth EstimationTransformerDiffusion modelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningImageVideoPoint Cloud
🎯 What it does: Propose a method for dense future prediction within the feature space of a visual foundation model (VFM) — VFMF — by compressing VFM features into a low-dimensional latent space using an autoencoder and employing flow matching for autoregressive generation;
VGGT-Motion: Motion-Aware Calibration-Free Monocular SLAM for Long-Range Consistency
Zhuang Xiong (Huazhong University of Science and Technology), Wenbing Tao (Huazhong University of Science and Technology)
Autonomous DrivingOptimizationSimultaneous Localization and MappingOptical FlowImageVideo
🎯 What it does: Propose an uncalibrated monocular SLAM system called VGGT-Motion, which achieves global consistency over kilometer-scale long sequences through motion-aware subgraph construction, anchor-driven direct Sim(3) registration, and lightweight subgraph pose graph optimization;
VIA-SD: Verification via Intra-Model Routing for Speculative Decoding
Yuchen Xian (Zhejiang University), Yi Yang (Zhejiang University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: This paper proposes a multi-level "Speculative Decoding" framework called VIA-SD, which inserts a slim-verifier obtained through intra-model routing between a lightweight drafter and a full-model validator, enabling hierarchical verification of tokens with different confidence levels;
VIBE: Disentangling Social Dynamics via Kinematics-Informed Variational Inference for Behavioral Emotion
Abhishek Pratap Singh (Indian Institute of Technology Roorkee), Balasubramanian Raman (Indian Institute of Technology Roorkee)
RecognitionTransformerContrastive LearningVideoText
🎯 What it does: Proposed an end-to-end framework called VIBE for group emotion recognition by removing environmental noise and leveraging group motion synchronization information.
Video-Based Optimal Transport for Feedback-Efficient Offline Preference-Based Reinforcement Learning
Tung Minh Luu, Chang D. Yoo (Korea Advanced Institute of Science and Technology)
OptimizationRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision-Language-Action ModelContrastive LearningImageVideo
🎯 What it does: This paper proposes a semi-supervised preference reinforcement learning framework called VOTP, which utilizes video baseline models and optimal transport to generate high-quality pseudo labels from a very small number of human-labeled data, thereby learning a reward function and training a policy.
Video-BCI: Bayesian Cognitive Integration of Self-Prior Hypotheses for Video Understanding
Xing Xi (South China University of Technology), Jiahao Liang (South China University of Technology)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: Propose the Video-BCI framework, modeling video understanding as a Bayesian cognitive process, first sampling self-prior hypotheses, then allowing the VLM to generate answers through critical integration;
Video-in-the-Loop: Span-Grounded Long Video QA with Interleaved Reasoning
Chendong Wang (University of Wisconsin Madison), Suman Banerjee (University of Wisconsin Madison)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Video-in-the-Loop (ViTL) framework, which divides long video QA into two stages: low-frame-rate preview localization and high-frame-rate focused inference, and interleaves the time span and final answer in the same output, achieving interpretable answer sources.
Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding
Yuan Xie (Hong Kong University of Science and Technology), Lionel Ni (Hong Kong University of Science and Technology)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built the Video-MTR framework, which uses a multi-round reasoning approach based purely on reinforcement learning to achieve long video understanding. The model gradually focuses on key information and provides answers through interactive retrieval and reasoning.
Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning
Xiangyu Zeng (Nanjing University), Limin Wang (Nanjing University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerReinforcement LearningPrompt EngineeringVision-Language-Action ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Video-o3 framework, which supports native interactive clue seeking and multi-hop reasoning in long videos.
Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation
Jiaze Li (Xiaomi Inc), Jian Luan (Xiaomi Inc)
Computational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningVideoTextMultimodality
🎯 What it does: Propose the Video-OPD framework for post-training of multi-modal large language models (MLLM) on temporal video grounding (TVG)
video-SALMONN S: Memory-Enhanced Streaming Audio-Visual LLM
Guangzhi Sun (Tsinghua University), Chao Zhang (Tsinghua University)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: Propose video-SALMONN, a streaming audio-visual large language model capable of processing videos longer than 3 hours at 1 FPS and 360p resolution; introduce test-time training (TTT) to enhance memory, and design TTTMEM layer, two-stage training, and modality-aware memory retrieval mechanism; also propose the ELViM benchmark to evaluate the model's learning and reuse capabilities in long videos.
Video-SVD: Efficient Video Diffusion via Orthogonal Basis Composition
Zhang Wan (Institute of Computing Technology, Chinese Academy of Sciences), Sheng Tang (Institute of Computing Technology, Chinese Academy of Sciences)
GenerationComputational EfficiencyTransformerDiffusion modelAuto EncoderContrastive LearningVideoText
🎯 What it does: Improve the inference efficiency of Video Diffusion Transformers (VDiT) by training an orthogonal basis combination method that does not modify model parameters, significantly reducing the quadratic complexity of self-attention.
Video2GUI: Synthesizing Large-Scale Interaction Trajectories for Generalized GUI Agent Pretraining
Weimin Xiong (Peking University), Hao Tian (Xiaomi)
Data SynthesisRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the Video2GUI framework, which automatically generates large-scale GUI interaction trajectories from Internet videos;
VideoBrain: Learning Adaptive Frame Sampling for Long Video Understanding
Junbo Zou (Georgia Institute of Technology), Weining Shen (University of California, Irvine)
RetrievalRepresentation LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelVideoTextRetrieval-Augmented Generation
🎯 What it does: Propose the VideoBrain framework, which utilizes a VLM to adaptively acquire keyframes through two sampling agents (CLIP retrieval and uniform sampling) to answer long video questions.
VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization
Andrei Atanov (Apple), Amir Zamir (Swiss Federal Institute of Technology Lausanne)
GenerationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerPrompt EngineeringDiffusion modelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningVideoText
🎯 What it does: Propose VideoFlexTok, a variable-length coarse-to-fine hierarchical video tokenizer equipped with a streaming decoder;
VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation
Hongyang Du (University of Southern California), Yue Wang (University of Southern California)
GenerationData SynthesisOptimizationKnowledge DistillationTransformerDiffusion modelScore-based ModelContrastive LearningImageVideoText
🎯 What it does: Through self-supervised 3D consistency evaluation, preference pairs are generated for pre-training video diffusion models using geometric foundation models, followed by post-training alignment via Direct Preference Optimization (DPO), enhancing the 3D structural consistency and temporal stability of videos.
VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding
Lin Fu (Zhejiang University), Yilun Zhao (Yale University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Construct the VideoKR training corpus, containing 315K questions, 145K professional domain CC-licensed videos, and design a human-computer interaction and skill-oriented QA generation and validation process, while creating the VideoKR-Eval benchmark.
VideoLoom: A Video Large Language Model for Joint Spatial-Temporal Understanding
Jiapeng Shi (Fudan University), Zuxuan Wu (Fudan University)
RecognitionSegmentationRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodalityBenchmark
🎯 What it does: Proposed a unified Video LLM (VideoLoom) that can simultaneously achieve spatiotemporal understanding of videos, supporting temporal localization and spatial segmentation.
VideoMAETok: Boosting Video Diffusion Models via Masked Autoencoders as Tokenizers
Zhan Tong (KU Leuven), Tinne Tuytelaars (KU Leuven)
GenerationRepresentation LearningTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningVideo
🎯 What it does: Designed and trained VideoMAETok, a video tokenizer based on ViT, which simulates the denoising task of diffusion models during training using a high proportion of masking and interpolated Gaussian noise to generate representations suitable for latent diffusion models.
VideoSEAL: Mitigating Evidence Misalignment in Agentic Long Video Understanding by Decoupling Answer Authority
Chenhao Qiu (Mango TV), Xusheng Liu (Mango TV)
RetrievalExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a decoupled planning-checker framework that separates planning from verification for long video question answering, avoiding agents from providing answers without sufficient evidence.
VideoSEG-O3: A Multi-turn Reinforcement Learning Framework for Reasoning Video Object Segmentation
Ming Dai (Southeast University), Jingdong Wang (Baidu Inc)
SegmentationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVideoTextMultimodalityChain-of-Thought
🎯 What it does: Proposed VideoSEG-O3, a multi-round reinforcement learning framework for inferential object segmentation in videos.
VideoTemp-o3: Harmonizing Temporal Grounding and Video Understanding in Agentic Thinking-with-Videos
Wenqi Liu (Shandong University), Xuemeng Song (Southern University of Science and Technology)
RecognitionComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Explored the agentic thinking-with-videos paradigm, proposing the VideoTemp-o3 framework, which achieves unified temporal localization and video question answering, and supports on-demand trimming and multi-round reflection.
VideoTrace-R1: Long Video-based Retrieval-Augmented Generation via Reinforcement Learning
Zongsheng Cao (Shanghai AI Laboratory), Zigan Wang (University of Queensland)
GenerationComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelVideoTextTime SeriesRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented VideoTrace-R1, which leverages the correspondence between video temporal sequences (Temporal Reasoning Traces, TRT) and model inference trajectories to enhance the temporal reasoning capability for long videos.
VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning
Hao Tan (University of Chinese Academy of Sciences), Zhen Lei (University of Chinese Academy of Sciences)
Anomaly DetectionReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringContrastive LearningVideoMultimodality
🎯 What it does: This paper proposes the VIDEOVERITAS framework for detecting AI-generated videos;
VidLaDA: Bidirectional Diffusion Large Language Models for Efficient Video Understanding
Zhihao He (Shanghai Jiao Tong University), Weiyao Lin (Shanghai Jiao Tong University)
GenerationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelDiffusion modelContrastive LearningVideoTextMultimodality
🎯 What it does: Construct a video LLM called VidLaDA using a bidirectional diffusion language model to achieve video understanding and generation.
ViEEG: Hierarchical Visual Neural Representation for EEG Brain Decoding
Minxu Liu (Nanjing University of Aeronautics and Astronautics), Qi Zhu (Nanjing University of Aeronautics and Astronautics)
RecognitionRetrievalRepresentation LearningConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: A three-stream EEG decoding framework called ViEEG based on the hierarchical visual brain was constructed, which can recover visual information from EEG signals and achieve zero-shot object recognition.
View Space: Learning Representation across Arbitrary Graphs
Dooho Lee (KAIST), Jaemin Yoo (Seoul National University)
ClassificationRepresentation LearningRecurrent Neural NetworkGraph Neural NetworkContrastive LearningTextGraph
🎯 What it does: This paper proposes View Space and Graph View Transformation (GVT), achieving fully inductive node representation learning for arbitrary graphs, and builds Recurrent GVT (RGVT) for node classification based on this.
ViewMask-1-to-3: Multi-View Consistent Image Generation via Multimodal Discrete Diffusion Models
Ruishu Zhu (Northwestern Polytechnical University), Xuelong Li (China Telecom)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageTextPoint CloudMesh
🎯 What it does: Propose ViewMask-1-to-3, which uses a discrete diffusion model to achieve multi-view consistent image generation, compatible with both image-to-multi-view and text-to-multi-view tasks.
VIP: Visual-guided Prompt Evolution for Efficient Dense Vision-Language Inference
Hao Zhu (Institute of Computing Technology, Chinese Academy of Sciences), Feng Dai (Institute of Computing Technology, Chinese Academy of Sciences)
SegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a training-free visual-language model method called VIP for efficient and accurate open-vocabulary semantic segmentation;
VIPO: Value Function Inconsistency Penalized Offline Reinforcement Learning
Xuyang Chen (National University of Singapore), Lin Zhao (National University of Singapore)
Reinforcement LearningTabularTime SeriesBenchmark
🎯 What it does: This paper proposes VIPO, a model-based offline reinforcement learning algorithm, which improves model accuracy by adding a value function inconsistency penalty during model training, thereby enhancing policy performance.
Vision Language Models Cannot Reason About Physical Transformation
Dezhi Luo (University of Michigan), Hokin Deng (Carnegie Mellon University)
Prompt EngineeringVision Language ModelVideoMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Constructed ConservationBench benchmark to evaluate the reasoning ability of vision-language models (VLMs) regarding physical conservation in videos.
Vision Transformer Finetuning Benefits from Non-Smooth Components
Ambroise Odonnat (Noah's Ark Lab), Ievgen Redko (Noah's Ark Lab)
ClassificationTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Investigate the plasticity (i.e., the average rate of change under input perturbations) of Vision Transformer components, and verify its impact on transfer learning through theoretical derivations and extensive finetuning experiments.
Vision-aligned Latent Reasoning for Multi-modal Large Language Model
Byungwoo Jeon (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a method to dynamically generate latent tokens aligned with visual information in multimodal large language models, thereby maintaining visual information without degradation during the chain-of-thought (CoT) process;
Vision-DeepResearch: Incentivizing DeepResearch Capability in Multimodal Large Language Models
Wenxuan Huang (Chinese University of Hong Kong), Wanli Ouyang (Chinese University of Hong Kong)
RetrievalRecommendation SystemExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMixture of ExpertsVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed Vision-DeepResearch, a system capable of achieving multi-round, cross-entity, and cross-scale visual and textual deep search and reasoning within multi-modal LLMs; integrated long-chain tool calling capabilities into the model through cold-start supervision and reinforcement learning; simultaneously constructed a large-scale multi-modal search trajectory and verification-based VQA dataset;
Vision-Language-Action Pretraining from Large-Scale Human Videos
Hao Luo (Peking University), Zongqing Lu (Peking University)
Robotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes a vision-language-action (VLA) pre-training framework called Being-H, based on large-scale human videos, which bridges human hand action templates with robotic execution through physical instruction tuning.
Vision2Web: A Hierarchical Benchmark for Visual Website Development with Agent Verification
Zehai He (Tsinghua University), Jie Tang (Tsinghua University)
AI Code AssistantTransformerLarge Language ModelAgentic AIVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose Vision2Web as a hierarchical visual website development benchmark for evaluating multimodal coding agents.
VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning
Hengbo Xu (Renmin University of China), Zhiwu Lu (Renmin University of China)
Computational EfficiencyTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose VisionPulse, which dynamically prunes visual tokens step-by-step during inference, retaining only a small amount of visual information needed for the current step.
Visual Implicit Autoregressive Modeling
Pengfei Jiang (China Telecom), Xuelong Li (China Telecom)
GenerationData SynthesisDepth EstimationComputational EfficiencyTransformerDiffusion modelScore-based ModelAuto EncoderImage
🎯 What it does: Propose an architecture (VIAR) that replaces the deep explicit intermediate stack in visual autoregressive models (VAR) with a single implicit balanced layer, and achieves computational control at each scale through adjustable iteration counts.
Visual Para-Thinker: Divide-and-Conquer Reasoning for Visual Comprehension
Haoran Xu (Zhejiang University), Jian Luan (Xiaomi Inc)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed Visual Para-Thinker, a parallel reasoning framework for multi-modal large language models, which generates multiple reasoning paths through visual segmentation and eventually merges them;
Visual Persuasion: What Influences Decisions of Vision-Language Models?
Manuel Cherep (Massachusetts Institute of Technology), Nikhil Singh (Dartmouth College)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a framework that iteratively optimizes visual presentation in image editing models through editable text prompts, aimed at probing and quantifying visual preferences of vision-language models (VLMs).
VisualPuzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge
Yueqi Song (Carnegie Mellon University), Xiang Yue (Carnegie Mellon University)
Supervised Fine-TuningPrompt EngineeringImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the VISUALPUZZLES benchmark, focusing on evaluating multimodal reasoning abilities while eliminating dependence on domain knowledge.
VisualScore: Learning Holistic Visual Quality Scores via Multi-Task Reasoning
Yiting Lu (University Of Science And Technology Of China), Weisi Lin (Nanyang Technological University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Developed a unified multi-task image quality assessment reward model called VisualScore, which can provide scores in dimensions such as technical quality, aesthetics, and text-image alignment, along with interpretable reasoning through structured inference.
ViSurf: Visual Supervised-and-Reinforcement Fine-Tuning for Large Vision-and-Language Models
Yuqi Liu (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
Anomaly DetectionOptimizationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose ViSurf, a unified single-stage visual supervision and reinforcement learning fine-tuning framework, to enhance the performance of large vision-language models across multiple tasks.
ViTok-v2: Scaling Native Resolution Autoencoders to 5 Billion Parameters
Philippe Hansen-Estruch (University of Texas), Ali Thabet (Meta Superintelligence Labs)
RestorationGenerationSuper ResolutionTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: A Vision Transformer autoencoder called ViTok-v2 is constructed, which can be trained at any resolution and aspect ratio, and its scale is extended to 5B parameters, significantly improving image reconstruction and generation performance.
VividCam: Learning Unconventional Camera Motions from Virtual Synthetic Videos
Qiucheng Wu (UC Santa Barbara), Shiyu Chang (UC Santa Barbara)
GenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelVideoText
🎯 What it does: Propose VIVIDCAM, which trains diffusion models to learn non-traditional camera motions by synthesizing low-poly virtual scene videos, achieving high-precision generation of camera control.
VJEPA: Variational Joint Embedding Predictive Architectures as Probabilistic World Models
Yongchao Huang (University of Aberdeen)
OptimizationRepresentation LearningTransformerReinforcement LearningMixture of ExpertsAuto EncoderContrastive LearningWorld ModelImageVideoTime SeriesSequential
🎯 What it does: This paper proposes Variational Joint Embedding Predictive Architecture (VJEPA) and its Bayesian extension BJEPA, constructing world models based on potential space prediction without pixel reconstruction, and verifies their robustness and uncertainty estimation capabilities in self-supervised learning and control tasks.
VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models
Borong Zhang (Peking University), Yaodong Yang (Peking University)
Autonomous DrivingRobotic IntelligenceTransformerReinforcement LearningPrompt EngineeringVision Language ModelVision-Language-Action ModelDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImageVideoTextMultimodalityPoint CloudBenchmark
🎯 What it does: Proposed the VLA-Arena benchmark and open-source framework for systematically evaluating the capability boundaries, robustness, and safety of Vision-Language-Action (VLA) models.
VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model
Wenhao Li (University Of Sydney), Chang Xu (University Of Sydney)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision Language ModelVision-Language-Action ModelDiffusion modelFlow-based ModelMultimodalityTime SeriesBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Developed the VLA-ATTC framework, introducing adaptive test-time computation and relative action critic into VLA models, enabling multi-candidate action reasoning to be triggered only in uncertain situations.
VLANeXt: Recipes for Building Strong VLA Models
Xiao-Ming Wu (S-Lab, Nanyang Technological University), Chen Change Loy (S-Lab, Nanyang Technological University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelContrastive LearningMultimodalityTime Series
🎯 What it does: This paper conducts an in-depth analysis of the Vision-Language-Action (VLA) design space through a unified framework and systematic evaluation, proposing a practical construction route from basic components, perceptual inputs to action modeling, and achieving a state-of-the-art VLANeXt model.
VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model
Yanjiang Guo (Stanford University), Chelsea Finn (Stanford University)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelDiffusion modelFlow-based ModelWorld ModelImageVideoTextMultimodality
🎯 What it does: Leverage real-world online rollout data to alternately fine-tune action-conditioned world models and vision-language-action (VLA) policies. By generating a large amount of high-quality synthetic trajectories using the world model, the success rate of the VLA policy is improved in various contact-rich tasks.
VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models
Rohit Saxena (University of Edinburgh), Pasquale Minervini (University of Edinburgh)
Supervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Constructed VLM-RobustBench, an evaluation benchmark containing 49 types of visual noise, blur, weather, digital, and geometric distortions, as well as 7 binary transformations (totaling 133 configurations), and evaluated 15 vision-language models on two large-scale multimodal task sets: MMBench and MMMU-Pro.
VlogReward: Learning Multi-Dimensional Evaluation for Vlog Editing
Yexiang Liu (Institute of Automation, Chinese Academy of Sciences), Zhenfang Chen (ByteDance Intelligent Creation)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Built a multidimensional evaluation framework and a large-scale dataset to automatically evaluate vlogger editing plans and provide fine-grained scores and actionable textual feedback.
VocSim A Training-free Benchmark for Zero-shot Content Identity in Single-source Audio
Maris Basha (University of Zurich and ETH Zurich), Richard Hahnloser (University of Zurich and ETH Zurich)
RecognitionTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningBenchmarkAudio
🎯 What it does: Propose the VOCSIM benchmark for evaluating frozen audio embeddings on single-source audio for zero-shot content identity
Von Mises-Fisher Mixture Model with Dynamic Shrinkage for Realistic Test-Time Transduction
Jiazhen Huang (Tsinghua University), Xiao Luo (University of Wisconsin-Madison)
ClassificationDomain AdaptationComputational EfficiencyRepresentation LearningPrompt EngineeringMixture of ExpertsVision Language ModelDiffusion modelScore-based ModelContrastive LearningGaussian SplattingImageTextMultimodality
🎯 What it does: Propose a dynamic shrinking framework called MOON based on the von Mises-Fisher mixture model, aimed at enhancing the zero-shot performance of vision-language models during test-time adaptation in real-world scenarios.
VPD-100K: Towards Generalizable and Fine-grained Visual Privacy Protection
Xiaobin Hu (National University Of Singapore), xiaoyu sun
Object DetectionSafty and PrivacyConvolutional Neural NetworkTransformerDiffusion modelContrastive LearningImageVideo
🎯 What it does: Proposed a large-scale, fine-grained visual privacy protection dataset called VPD-100K, and designed a lightweight detection module based on the frequency domain to improve real-time privacy recognition performance.
VR-Thinker: Boosting Multimodal Reward Models through Think with Image Reasoning
Qunzhong Wang (Chinese University of Hong Kong), Jiaheng Liu (Nanjing University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose VR-THINKER, a reward model with visual reasoning capabilities that can dynamically retrieve and utilize frame information in video preference evaluation.
VSCD: Video-based Scene Change Detection in Unaligned Scenes
Jiae Yoon (Gwangju Institute of Science and Technology), Ue-Hwan Kim (Gwangju Institute of Science and Technology)
Object DetectionSegmentationDomain AdaptationConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningOptical FlowImageVideo
🎯 What it does: Proposes a video-based scene change detection (VSCD) framework that can predict pixel-level change masks for query frames in indoor environments with different times, perspectives, and without temporal synchronization.
VT-Bench: A Unified Benchmark for Visual-Tabular Multi-Modal Learning
Ziyi Jia, Lan-Zhe Guo (Nanjing University)
Data-Centric LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodalityTabularElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes VT-Bench, a unified visual-table multimodal learning benchmark covering two paradigms: discriminative prediction and generative reasoning;
Vulnerable Agent Identification in Large-Scale Multi-Agent Reinforcement Learning
Simin Li (Beihang University), Xianglong Liu (Beihang University)
OptimizationReinforcement LearningTabularTime SeriesSequential
🎯 What it does: This paper studies the problem of identifying vulnerable agents in large-scale multi-agent reinforcement learning, and proposes the HAD-MFC framework, decoupling the upper-level combination search from the lower-level mean-field adversarial learning through the Fenchel-Rockafellar transformation.
Wait, Wait, Wait... Why Do Reasoning Models Loop?
Charilaos Pipis (Massachusetts Institute of Technology), Dimitris Papailiopoulos (Microsoft Research)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextGraphChain-of-Thought
🎯 What it does: This paper investigates the reasons why inference models generate cyclic reasoning when producing long-chain thinking, and experiments demonstrate that learning errors are the main factor.
Walrus: A Cross-domain Foundation Model for Continuum Dynamics
Michael McCabe (Polymathic AI), Shirley Ho (Polymathic AI)
TransformerVision-Language-Action ModelDiffusion modelScore-based ModelAuto EncoderContrastive LearningTime SeriesSequentialPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Built a cross-domain fluid continuum dynamics foundation model named WALRUS, implementing long-term sequence physical simulation using the Transformer architecture.
WarmServe: Enabling One-for-Many GPU Prewarming for Multi-LLM Serving
Chiheng Lou (Peking University), Xin Jin (Peking University)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes WarmServe, a multi-model GPU preheating system based on workload prediction, which can pre-load multi-model parameters before request peaks, significantly reducing inference time and improving throughput.
Wasserstein Geometry-Aware Adaptive Control via Meta-Learning
Xingyu Yang (Xi'an Jiaotong University), Jianfu Cao (Xi'an Jiaotong University)
OptimizationMeta LearningReinforcement LearningContrastive LearningTabularTime Series
🎯 What it does: Propose an adaptive control framework based on Wasserstein transformation geometry, which utilizes meta-learning to jointly optimize nonlinear features, control gain, and transport geometry, achieving closed-loop tracking under unknown disturbances.
Watch Your Step: Information Injection in Diffusion Models via Shadow Timestep Embedding
An Huang (University of Nevada Las Vegas), Zuobin Xiong (University of Nevada Las Vegas)
GenerationData SynthesisSafty and PrivacySupervised Fine-TuningDiffusion modelScore-based ModelImage
🎯 What it does: Studied the temporal embedding space of diffusion models and proposed Shadow Timestep Embedding (STE) to inject hidden information along the time axis.
WatchLog: Efficient and Interpretable Event Reasoning for Endpoint Detection and Response Logs with Multimodal LLMs
Hongyi Zhou (Tsinghua University), Xuling Zhang (360 Security Technology Inc.)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Propose the WatchLog framework, which converts endpoint security logs into video-language structured inputs to achieve efficient and interpretable threat detection and reasoning.
Watermarking Graph Neural Networks via Explanations for Ownership Protection
Jane Downer (Illinois Institute of Technology), Binghui Wang (Illinois Institute of Technology)
Safty and PrivacyExplainability and InterpretabilityGraph Neural NetworkTransformerAuto EncoderContrastive LearningGaussian SplattingGraphTabular
🎯 What it does: This paper proposes an interpretable graph neural network watermarking method to verify model ownership in black-box environments.
Watermarking LLM Agent Trajectories
Wenlong Meng (Zhejiang University), Wenzhi CHEN
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Proposes a behavior-level watermarking method called ACTHOOK, which inserts hook actions into the LLM agent trajectory and activates them with a key, achieving detectable marking of the dataset without altering the original task results.
WaterSIC: Information-Theoretically (Near) Optimal \\Linear Layer Quantization
Egor Lifar (Massachusetts Institute of Technology), Yury Polyanskiy (Massachusetts Institute of Technology)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningText
🎯 What it does: Propose a post-training quantization method called WaterSIC, which is information-theoretically approximately optimal, and implement quantization on various large language models (e.g., Llama-3.2-1B, Qwen3-8B, Llama-2-7B).
WAVE: Window-Aware Vocabulary-Efficient Early-Exit for Training-Free LLM Acceleration
Seonggeun Kim (Seoul National University of Science and Technology), Hyun Kim (Seoul National University of Science and Technology)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes WAVE, a training-agnostic early exit framework that utilizes windowed exit and vocabulary-efficient proxy LM heads, significantly accelerating LLM inference.
WaveSSM: Multiscale State-Space Models for Non-stationary Signal Attention
Ruben Solozabal (MBZUAI), Martin Takáč (MBZUAI)
ClassificationRecognitionTransformerTime SeriesBiomedical DataElectrocardiogramStochastic Differential EquationOrdinary Differential EquationAudio
🎯 What it does: Propose WaveSSM, a multi-scale state space model based on the wavelet framework, designed to capture local transient features in non-stationary signals.
WBMM: Windowed Batch Matrix Multiplication for Efficient Large Receptive Field Convolution
Wan Song (Hefei University of Technology), Shu Zhan (Hefei University of Technology)
ClassificationObject DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Propose a windowed batch matrix multiplication (WBMM) method, which divides the input into windows and constructs the weight matrix using a relative position information table, thereby achieving efficient computation for large kernel convolutions.
Weak Diffusion Priors Can Still Achieve Strong Inverse-Problem Performance
Jing Jia (Rutgers University), Guanyang Wang (Rutgers University)
RestorationSuper ResolutionOptimizationDiffusion modelScore-based ModelImage
🎯 What it does: The study investigates the performance of weak (few-step or cross-domain) diffusion models in inverse problems, and provides theoretical and experimental evidence that they can reach the level of strong models when there is sufficient information.
Weak-to-Strong Generalization via Bregman Bias–Variance Decomposition
Gengze Xu (Renmin University of China), Yong Liu (Renmin University of China)
ClassificationKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextReview/Survey Paper
🎯 What it does: This paper theoretically analyzes the weak-to-strong generalization (W2SG) phenomenon through Bregman bias-variance decomposition, and provides error inequalities that do not rely on convexity and realizability assumptions. It further explores the sufficient conditions for student models approaching the posterior mean teacher, as well as their impacts on student capacity, cross-entropy, and inverse cross-entropy training objectives.
Weakly Supervised Cross-Modal Learning for 4D Radar Scene Flow Estimation
Jingyun Fu (Zhejiang University), Na Zhao (Singapore University of Technology and Design)
Autonomous DrivingOptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkVision Language ModelDiffusion modelScore-based ModelContrastive LearningSimultaneous Localization and MappingOptical FlowImageVideoPoint Cloud
🎯 What it does: Propose a weakly supervised cross-modal learning framework called IterFlow, which provides auxiliary supervision for 4D radar scene flow estimation during training using only RGB images and odometry;
Weasel: Out-of-Domain Generalization for Web Agents via Importance-Diversity Data Selection
Fatemeh Pesaran Zadeh (Seoul National University), Gunhee Kim (Seoul National University)
Domain AdaptationComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelContrastive LearningTextMultimodalitySequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the WEASEL method, which selects important and diverse data for trajectories in offline web proxy training, combined with goal-centered state pruning and self-generated inference trajectories to improve cross-domain generalization and training efficiency.
WeatherSyn: An Instruction Tuning MLLM For Weather Forecasting Report Generation
Zinan Zheng (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringImageTextMultimodalityTime SeriesRetrieval-Augmented Generation
🎯 What it does: Construct a weather forecast report generation task and the corresponding dataset, and train a specialized multimodal large language model called WeatherSyn
Weaving Graph over Tokens: Contextualizing Structured Sequences for LLMs
Jiaxuan Chen (Chinese University of Hong Kong Shenzhen), Fangxin Wang (Chinese University of Hong Kong Shenzhen)
ClassificationRecommendation SystemComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmark
🎯 What it does: Proposes Weaver, a framework without an encoder that directly embeds graph structures into the LLM attention mechanism, achieving the unification of graph reasoning and text generation through graph causal masking, GoRoPE, and the local priority principle.
Weaving in the Clouds: Achieving Synergistic Collaboration among LLM Agents via Federated Learning
Jiaxing Zhao (Jilin University), Haoran Zhang (Peking University)
OptimizationFederated LearningSafty and PrivacyComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextTabularBenchmark
🎯 What it does: Propose the FedWave framework, achieving multi-agent LLM collaboration based on federated learning, supporting workflow-style tasks.
WebWorld: A Large-Scale World Model for Web Agent Training
Zikai Xiao (Zhejiang University), Zuozhu Liu (Zhejiang University)
Autonomous DrivingOptimizationSafty and PrivacyComputational EfficiencyData-Centric LearningRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringWorld ModelTextTime SeriesSequentialReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose WebWorld, a large-scale web world model trained on 1M+ real web interaction trajectories, supporting multi-format, long-sequence simulation for offline training of web agents.
WeDLM: Reconciling Diffusion Language Models with Standard Causal Attention for Fast Inference
Aiwei Liu (WeChat AI, Tencent), Jie Zhou (WeChat AI, Tencent)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Designed the WeDLM framework, achieving parallel diffusion language model decoding under standard causal attention, thereby being compatible with KV cache and improving inference efficiency.
Weight Decay Improves Language Model Plasticity
Tessa Han (Broad Institute), Sham M. Kakade (Harvard University)
OptimizationHyperparameter SearchData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Studied the impact of weight decay during the pre-training phase of large language models on model plasticity and downstream performance
Weight Updates as Activation Shifts: A Principled Framework for Steering
Dyah Adila (University of WisconsinMadison), Frederic Sala (University of WisconsinMadison)
Computational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose an activation layer intervention method on large language models that is theoretically equivalent to weight fine-tuning, and construct a first-order equivalent framework that maps weight updates to activation updates;
Weight-Space Learning for Certifiable Few-shot Transfer Learning
Fady Rezk (Samsung AI Center), Minyoung Kim (Samsung AI Center)
Federated LearningRepresentation LearningMeta LearningTransformerSupervised Fine-TuningDiffusion modelContrastive LearningImageText
🎯 What it does: Propose the STEEL method, which achieves provable few-shot transfer learning by training diffusion models in upstream tasks to learn the PEFT parameter distribution, and then performing sampling-evaluation selection in downstream few-shot tasks.
Weight-sparse transformers have interpretable circuits
Leo Gao (OpenAI), Daniel P Mossing
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelAuto EncoderContrastive LearningTextSequential
🎯 What it does: By applying L0 sparsity constraints to the majority of weights in the Transformer, a sparse-weight model is trained, and structured pruning methods are used to extract the minimal circuit for each manually designed task; this circuit is then manually parsed to verify its interpretability.
WeightCLIP: Aligning Datasets and Models for Weight Space Learning
Aron Asefaw (University of St.Gallen), Damian Borth (University of St.Gallen)
RetrievalRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes WeightCLIP, which aligns model weights and datasets through contrastive learning, making model weights cluster in the latent space according to training data, thus achieving model retrieval, generation, and fine-tuning.
Weights to Code: Extracting Interpretable Algorithms from the Discrete Transformer
Yifan Zhang (Peking University), Zhi Jin (Peking University)
OptimizationExplainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerTabularTime SeriesSequentialBenchmark
🎯 What it does: Propose the Discrete Transformer architecture and extract interpretable programs from its weights through hypothesis testing and symbolic regression.
Welfare-Optimal Classification with Accuracy Auctions
Bana Sadi (Technion), Nir Rosenfeld (Technion)
ClassificationOptimizationFederated LearningData-Centric LearningReinforcement LearningContrastive LearningTabularFinance Related
🎯 What it does: Propose a welfare-optimal classification method based on an 'accuracy bidding' mechanism, which acquires users' private values for accurate predictions through an auction during training, and uses these values as sample weights to train the classifier, ultimately achieving social welfare maximization.
Well-Posed KL-Regularized Control via Wasserstein and Kalman–Wasserstein KL Divergences
Viktor Stein (Technical University of Munich), Nihat Ay (Hamburg University of Technology)
OptimizationReinforcement Learning
🎯 What it does: Proposed a state-space aware KL divergence based on Wasserstein and Kalman-Wasserstein geometry, and applied it to KL regularized control to address the issue of KL divergence divergence under low noise conditions.
WestWorld: A Knowledge-Encoded Scalable Trajectory World Model for Diverse Robotic Systems
Yuchen Wang (William & Mary), Huajie Shao (William & Mary)
Robotic IntelligenceTransformerReinforcement LearningMixture of ExpertsWorld ModelTime SeriesSequential
🎯 What it does: A scalable trajectory world model, WestWorld, was constructed, aiming to learn the dynamics of multiple robot systems at once and achieve high-quality prediction and control in diverse environments.
WET: Mitigating World-Conditioned Knowledge Conflicts via World Entropy Tethering
Zixuan Wang (Harbin Institute of Technology), Chaomeng Chen (Great Bay University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringScore-based ModelTextRetrieval-Augmented Generation
🎯 What it does: Designed the World Entropy Tethering (WET) framework, which dynamically adjusts sparse attention heads by monitoring world entropy during inference and guiding them with geometric directions, to suppress drift and hallucination of LLMs in world conflict scenarios;