ICLR 2026 Papers — Page 51
International Conference on Learning Representations · 5356 papers
Understanding Task Vectors in In-Context Learning: Emergence, Functionality, and Limitations
Yuxin Dong (Ohio State University), Xia Ning (Ohio State University)
Explainability and InterpretabilityMeta LearningTransformerText
🎯 What it does: This paper analyzes the generation mechanism and function of task vectors in ICL, proposing and verifying the hypothesis that 'task vectors represent demonstrations';
Understanding the Dynamics of Forgetting and Generalization in Continual Learning via the Neural Tangent Kernel
Guodong Zheng (Huazhong University of Science and Technology), Li Shen (Guangdong Laboratory of Artificial Intelligence and Digital Economy)
OptimizationMeta LearningImageBenchmark
🎯 What it does: Analyze the training dynamics of forgetting and generalization in continual learning, and provide theoretical upper bounds during the intermediate phase under the neural tangent kernel (NTK) framework
Understanding the Emergence of Seemingly Useless Features in Next-Token Predictors
Mark Rofin (EPFL), Michael Hahn (Saarland University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderTextSequential
🎯 What it does: Investigate why Transformers learn features useless for direct prediction during next-token training, decomposing gradient signals into three paths: direct, pre-cached, and shared, and experimentally verifying their impact on feature emergence.
Understanding the Implicit Biases of Design Choices for Time Series Foundation Models
Annan Yu (Cornell University), Bernie Wang
Explainability and InterpretabilityTransformerTime SeriesBenchmark
🎯 What it does: Systematically analyze the three design options (patch size, embedding methods, training loss) in Time Series Foundation Models (TSFMs) that implicitly contain biases, and verify the impact of these biases on model performance through theoretical analysis and controlled experiments.
Understanding the Learning Phases in Self-Supervised Learning via Critical Periods
JangHyeon Lee (University of Minnesota), Dalton Lunga (Oak Ridge National Laboratory)
Domain AdaptationKnowledge DistillationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Investigate the impact of self-supervised learning (SSL) pretraining duration on transfer performance, finding that mid-checkpoints perform better on cross-domain (OOD) transfer, while longer pretraining mainly improves in-domain (ID) performance, and for the first time apply the critical period concept from neuroscience to SSL;
Understanding the Mechanisms of Fast Hyperparameter Transfer
Nikhil Ghosh (Flatiron Institute), Alberto Bietti (Flatiron Institute)
OptimizationExplainability and InterpretabilityHyperparameter SearchTransformerImageText
🎯 What it does: Investigate the rapid transfer mechanisms of hyperparameters (especially learning rates) in large neural networks across different widths, and propose a low-dimensional trajectory analysis framework based on EMA smoothing linearization and top-k loss decomposition.
Understanding the Mixture-of-Experts with Nadaraya-Watson Kernel
Chuanyang Zheng (Morgan Stanley), Yuriy Nevmyvaka (Morgan Stanley)
TransformerMixture of ExpertsTextBenchmark
🎯 What it does: This paper reinterprets the Mixture-of-Experts (MoE) routing as Nadaraya-Watson regression and proposes the KERN router, which uses only ReLU and ℓ2 normalization without introducing additional parameters;
Understanding the Robustness of Distributed Self-Supervised Learning Frameworks Against Non-IID Data
Xuanyu Chen (University of Sydney), Dong Yuan (University of Sydney)
Federated LearningRepresentation LearningData-Centric LearningAuto EncoderContrastive LearningImage
🎯 What it does: Investigated the robustness of distributed self-supervised learning in non-IID environments, and proposed the MAR loss to enhance the robustness of MIM models;
Understanding the Role of Training Data in Test-Time Scaling
Adel Javanmard (University of Southern California), Vahab Mirrokni (Google Research)
OptimizationComputational EfficiencyData-Centric LearningTransformerTextBenchmarkChain-of-Thought
🎯 What it does: The study tests the effectiveness of test-time scaling in transformer models, proposes using chain-of-thought (CoT) to implement a pseudo Newton method, and provides theoretical analysis for linear regression tasks; meanwhile, it defines a task difficulty metric and designs an optimal task selection strategy in multi-task training.
UNDERSTANDING TRANSFORMERS FOR TIME SERIES FORECASTING: A CASE STUDY ON MOIRAI
Dennis Wu (Northwestern University), Han Liu (Northwestern University)
TransformerTime Series
🎯 What it does: This paper theoretically explains the expressiveness and generalization ability of Transformers in time series forecasting, with a focus on analyzing the MOIRAI model.
Understanding Transformers for Time Series: Rank Structure, Flow-of-ranks, and Compressibility
Annan Yu (Center for Applied Mathematics, Cornell University), Bernie Wang
CompressionComputational EfficiencyTransformerTime Series
🎯 What it does: This paper systematically investigates the rank structures of the embedding layer, attention layer, and deep networks in time series Transformers (TSFM) through numerical rank analysis, introduces the concept of 'flow-of-ranks,' and applies it to compress large-scale TSFM models such as Chronos;
Understanding vs. Generation: Navigating Optimization Dilemma in Multimodal Models
Sen Ye (Peking University), Winston Hu
OptimizationReinforcement LearningVision Language ModelDiffusion modelFlow-based ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the Reason-Reflect-Refine (R3) framework, decomposing image generation into three stages: reasoning, reflection, and refinement, while explicitly leveraging the model's understanding capabilities to enhance both generation and comprehension performance.
Unfolding Spatial Cognition: Evaluating Multimodal Models on Visual Simulations
Linjie Li (University of Washington), Ranjay Krishna (Stanford University)
TransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Proposed and implemented the STARE benchmark to evaluate the capabilities of multimodal large language models in spatial reasoning tasks requiring multi-step visual simulation.
Uni-CoT: Towards Unified Chain-of-Thought Reasoning Across Text and Vision
Luozheng Qin (Shanghai Academy Of Ai For Science), Hao Li (Fudan University)
GenerationTransformerMixture of ExpertsVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Developed a unified Chain-of-Thought framework, Uni-CoT, enabling cross-modal reasoning between vision and text, and enhancing performance in image generation and understanding tasks.
Uni-DPO: A Unified Paradigm for Dynamic Preference Optimization of LLMs
Shangpin Peng (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the Uni-DPO framework, achieving dynamic weighted learning based on direct preference optimization.
Uni-NTFM: A Unified Foundation Model for EEG Signal Representation Learning
Zhisheng Chen (Beijing Key Laboratory for Multimodal Collaboration and Advanced Application, Institute of Computing Technology, Chinese Academy of Sciences), Xinliang Zhou (Nanyang Technological University)
Representation LearningTransformerMixture of ExpertsBiomedical Data
🎯 What it does: Constructed and pre-trained a foundation model for EEG, Uni-NTFM, based on neural decoding principles, followed by linear probing and fine-tuning validation on nine downstream tasks.
Uni-X: Mitigating Modality Conflict with a Two-End-Separated Architecture for Unified Multimodal Models
Jitai Hao (Baidu Inc), Jun Yu (Baidu Inc)
GenerationTransformerVision Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: Proposed and implemented the Uni-X unified multimodal model, which employs modality-specific layers in shallow and deep layers, and shares parameters in the middle layer to alleviate gradient conflicts between vision and text.
UniCA: Unified Covariate Adaptation for Time Series Foundation Model
Lu Han (Ant Group), De-Chuan Zhan (Nanjing University)
Domain AdaptationTransformerImageTextMultimodalityTime Series
🎯 What it does: Explored how to make time series foundation models (TSFMs) compatible with general covariate (homogeneous, heterogeneous) prediction tasks, and proposed the UniCA framework to achieve unified assimilation and fusion of covariates.
UniCalli: A Unified Diffusion Framework for Column-Level Generation and Recognition of Chinese Calligraphy
Tianshuo Xu (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
RecognitionGenerationTransformerDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Proposed a unified diffusion framework called UniCalli, capable of simultaneously generating and recognizing column-level Chinese calligraphy, and constructed a large-scale annotated dataset with over 8000 calligraphy works.
UniCon: Unified Framework for Efficient Contrastive Alignment via Kernels
Hangke Sui (University Of Illinois Urbana Champaign), Minh N. Do (University Of Illinois Urbana Champaign)
RetrievalRepresentation LearningContrastive LearningImageTextMultimodality
🎯 What it does: Propose a unified contrastive learning framework called UniCon, which constructs a contrastive similarity weight matrix and uses spectral decomposition to obtain a closed-form solution, thereby replacing traditional SGD training with a single closed-form update.
UniEdit-Flow: Unleashing Inversion and Editing in the Era of Flow Models
Guanlong Jiao (University of British Columbia), Renjie Liao (Tsinghua University)
Image HarmonizationTransformerFlow-based ModelRectified FlowImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a parameter-free, model-agnostic image inversion method Uni-Inv and a region-aware text-driven editing method Uni-Edit based on flow models, addressing the issue of ineffective delayed injection caused by straight trajectories in flow models.
UniF$^2$ace: A $\underline{Uni}$fied $\underline{F}$ine-grained $\underline{Face}$ Understanding and Generation Model
Junzhe Li (Peking University), Shuicheng YAN
RecognitionGenerationTransformerMixture of ExpertsVision Language ModelDiffusion modelScore-based ModelImageTextMultimodality
🎯 What it does: Developed a unified multimodal model named UniF ace for fine-grained facial understanding and generation, proposed the dual discrete diffusion (D3Diff) loss and multi-level grouped Mixture-of-Experts (MoE) architecture, and constructed a large-scale facial text-paired and VQA dataset named UniF aceD-1M.
Unified 3D Scene Understanding Through Physical World Modeling
Wanhee Lee (Stanford University), Daniel LK Yamins
GenerationDepth EstimationTransformerWorld ModelOptical FlowImageVideoMultimodality
🎯 What it does: This paper proposes a unified physical world model (3WM) that can accomplish three 3D vision tasks—depth estimation, view synthesis, and object manipulation—within a single framework;
Unified Analyses for Hierarchical Federated Learning: Topology Selection under Data Heterogeneity
Ziyi Zhou (Beijing University of Posts and Telecommunications), Xinchen Lyu (Beijing University of Posts and Telecommunications)
OptimizationFederated LearningConvolutional Neural NetworkImageText
🎯 What it does: Proposes a unified non-convex optimization convergence framework, analyzes four HFL topologies (Star-Star, Star-Ring, Ring-Star, Ring-Ring), and provides convergence upper bounds.
Unified and Efficient Multi-view Clustering from Probabilistic Perspective
Yalan Qin (Shanghai University), Guorui Feng (Shanghai University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningMultimodality
🎯 What it does: Propose a unified and efficient multi-view clustering method called UEMCP, combining anchor graphs with a probabilistic perspective to achieve end-to-end clustering
Unified Biomolecular Trajectory Generation via Pretrained Variational Bridge
Ziyang Yu (Tsinghua University), Yang Liu (Tsinghua University)
Drug DiscoveryReinforcement LearningAuto EncoderBiomedical DataStochastic Differential Equation
🎯 What it does: Propose a pre-trained variational bridge model (PVB) for cross-domain molecular trajectory generation, and introduce a reinforcement learning-based post-optimization for protein-ligand complexes.
Unified Brain Surface and Volume Registration
Mazdak Abulnaga (Massachusetts Institute of Technology), Adrian V Dalca
Convolutional Neural NetworkBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes the NeurAlign framework, which combines voxel and spherical networks to achieve single-step cortical and subcortical registration in brain MRI.
Unified Diffusion VLA: Vision-Language-Action Model via Joint Discrete Denosing Diffusion Process
Jiayi Chen (Hong Kong University of Science and Technology), Haoang Li (Hong Kong University of Science and Technology)
GenerationRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelMultimodality
🎯 What it does: Propose a unified diffusion-based vision-language-action (VLA) model, UD-VLA, which can simultaneously generate future images and actions along the same denoising trajectory.
Unified In-Context Video Editing
Zixuan Ye (Hong Kong University of Science and Technology), Wenhan Luo (Hong Kong University of Science and Technology)
GenerationTransformerVision-Language-Action ModelDiffusion modelAuto EncoderVideoMultimodality
🎯 What it does: Proposed a unified In-Context video editing framework called UNIC, which can achieve multiple video editing tasks within a single model;
Unified Multi-Modal Interactive and Reactive 3D Motion Generation via Rectified Flow
Prerit Gupta (Purdue University), Aniket Bera (Purdue University)
GenerationTransformerRectified FlowTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: This paper proposes the DualFlow framework, achieving dual-person 3D action generation based on text, music, and retrieval examples in both interactive and reactive modes.
Unified Privacy Guarantees for Decentralized Learning via Matrix Factorization
Aurélien Bellet (Inria), Francois Taiani
OptimizationFederated LearningSafty and PrivacyImageGraphTabular
🎯 What it does: Propose a unified privacy-preserving framework based on matrix decomposition, providing tighter differential privacy metrics, and design a new algorithm MAFALDA-SGD to achieve decentralized learning with local differential privacy.
Unified Vision-Language-Action Model
Yuqi Wang (Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
Autonomous DrivingRobotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelWorld ModelVideoMultimodalityBenchmark
🎯 What it does: Developed a unified vision-language-action model called UniVLA, which discretizes vision, language, and action into tokens and employs an autoregressive Transformer for unified modeling. The model significantly improves the effectiveness and efficiency of policy learning through pretraining on a large-scale robot video dataset as a world model.
Unified Vision–Language Modeling via Concept Space Alignment
Yifu QIU, Holger Schwenk (FAIR at Meta)
GenerationRetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelContrastive LearningImageVideoTextMultimodality
🎯 What it does: This paper proposes V-SONAR, which extends the original SONAR multilingual embedding space from text and speech to images and videos, achieving a unified representation across four modalities; based on this, V-LCM is designed, combining visual language instruction fine-tuning to enhance multimodal reasoning and generation capabilities.
UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation
Zhengrong Yue (Shanghai Jiao Tong University), Yali Wang (Chinese Academy Of Sciences)
RecognitionObject DetectionSegmentationGenerationKnowledge DistillationTransformerDiffusion modelFlow-based ModelContrastive LearningImageMultimodality
🎯 What it does: Propose a unified pixel flow tokenizer UniFlow, integrating hierarchical adaptive distillation with a lightweight patch-wise pixel flow decoder, achieving dual functions of visual understanding and generation.
Uniform Discrete Diffusion with Metric Path for Video Generation
Haoge Deng (National Laboratory of Pattern Recognition), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationTransformerLarge Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: In the paper, the authors propose the URSA framework, which utilizes a unified discrete diffusion model and a metric path to achieve video generation.
Unifying Formal Explanations: A Complexity-Theoretic Perspective
Shahaf Bassan (Hebrew University of Jerusalem), Guy Katz (Hebrew University of Jerusalem)
OptimizationExplainability and InterpretabilityComputational Efficiency
🎯 What it does: Propose a unified combinatorial optimization framework to study the computational complexity of sufficient and contrastive explanations in local and global, deterministic and probabilistic interpretations.
Unifying Stable Optimization and Reference Regularization in RLHF
Li He (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed a unified dual KL regularization framework and implemented the DAR algorithm to address reward hijacking and conflicts between stable optimization in RLHF, further replacing traditional RL with weighted SFT to simplify implementation and enhance learning stability.
UniHand: A Unified Model for Diverse Controlled 4D Hand Motion Modeling
Zhihao Sun (Fudan University), Zuxuan Wu (Fudan University)
GenerationPose EstimationDiffusion modelAuto EncoderMultimodality
🎯 What it does: Propose a unified Diffusion framework named UniHand for simultaneously performing 4D hand estimation and generation;
UniHM: Unified Dexterous Hand Manipulation with Vision Language Model
Zhenhao Zhang (ShanghaiTech University), Jingya Wang (ShanghaiTech University)
OptimizationRepresentation LearningRobotic IntelligenceTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelAuto EncoderVideoTextMultimodality
🎯 What it does: Propose the UniHM framework to achieve executable multi-step dexterous hand manipulation sequence generation based on open-vocabulary language instructions.
UniLiP: Adapting CLIP for Unified Multimodal Understanding, Generation and Editing
Hao Tang (Peking University), Liwei Wang (Peking University)
RestorationGenerationTransformerVision Language ModelDiffusion modelMultimodality
🎯 What it does: Propose the UniLIP framework, extending CLIP into a unified model that supports visual understanding, image reconstruction, generation, and editing
UniOD: A Universal Model for Outlier Detection across Diverse Domains
Dazhi Fu (Chinese University of Hong Kong, Shenzhen), Jicong Fan (Chinese University of Hong Kong, Shenzhen)
Anomaly DetectionGraph Neural NetworkTransformerTabularBenchmarkFinance Related
🎯 What it does: Propose a unified anomaly detection framework called UniOD, which trains a single model using historical labeled data to directly detect anomalies on new datasets (cross-domain, multi-dimensional);
UniQL: Unified Quantization and Low-rank Compression for Adaptive Edge LLMs
Hung-Yueh Chiang (University of Texas at Austin), Diana Marculescu (University of Texas at Austin)
CompressionComputational EfficiencyLarge Language ModelText
🎯 What it does: Developed a unified post-training quantization and low-rank compression framework called UniQL for adaptive deployment of large language models on edge devices.
UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper Granularity
Jingbo Lin (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)
RestorationMixture of ExpertsVision Language ModelImage
🎯 What it does: Designed and implemented UniRestorer, a multi-grained hybrid expert framework for uniformly handling various image restoration tasks;
UniSplat: Unified Spatio-Temporal Fusion via 3D Latent Scaffolds for Dynamic Driving Scene Reconstruction
Chen Shi (Chinese University of Hong Kong), Li Jiang (Chinese University of Hong Kong)
Autonomous DrivingConvolutional Neural NetworkTransformerGaussian SplattingVideo
🎯 What it does: Proposed the UniSplat framework to achieve real-time dynamic driving scene 3D reconstruction and novel view synthesis from multi-camera videos.
UniSS: Unified Expressive Speech-to-Speech Translation with Your Voice
Sitong Cheng (Hong Kong University Of Science And Technology), Wei Xue (Hong Kong University Of Science And Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringTextChain-of-ThoughtAudio
🎯 What it does: Developed UniSS—a single-stage unified speech-to-speech translation framework that can complete content translation, speaker identity preservation, and emotional transmission in one inference process.
UNITE: Universal kNowledge Integration from Task-specific Experts
Shuxia Lin (Southeast University), Xin Geng (Southeast University)
Computational EfficiencyKnowledge DistillationRepresentation LearningMixture of ExpertsText
🎯 What it does: Propose the UNITE framework for extracting and reusing general knowledge from Mixture-of-Experts language models.
UniTrack: Differentiable Graph Representation Learning for Multi-Object Tracking
Bishoy Galoaa (Northeastern University), Sarah Ostadabbas (Northeastern University)
Object TrackingGraph Neural NetworkVideo
🎯 What it does: Propose a pluggable graph-theoretic loss function UniTrack, which unifies the optimization of detection accuracy, identity preservation, and spatiotemporal consistency during training, significantly reducing errors such as ID switching, temporal inconsistency, and cross-target ID swapping in multi-object tracking;
UniUGG: Unified 3D Understanding and Generation via Geometric-Semantic Encoding
Yueming Xu (Fudan University), Li Zhang (Fudan University)
GenerationData SynthesisRepresentation LearningTransformerLarge Language ModelVision Language ModelDiffusion modelAuto EncoderImageTextMultimodalityPoint Cloud
🎯 What it does: Propose the UniUGG framework to achieve 3D scene generation and spatial-level visual question answering based on reference images and viewpoint transformations, integrating 3D space understanding and generation.
Universal Beta Splatting
Rong Liu (University of Southern California), Ziyan Wu (United Imaging Intelligence)
Computational EfficiencyNeural Radiance FieldGaussian SplattingImageVideo
🎯 What it does: Propose a unified N-dimensional Beta kernel-based radiance field rendering framework (UBS), replacing fixed Gaussian distributions, supporting adaptive control in spatial, view, and temporal dimensions, compatible with existing 3DGS/6DGS/7DGS, and achieving real-time rendering.
Universal Inverse Distillation for Matching Models with Real-Data Supervision (No GANs)
Nikita Maksimovich Kornilov, Alexander Korotin (Applied AI Institute)
GenerationKnowledge DistillationDiffusion modelScore-based ModelFlow-based ModelImage
🎯 What it does: Proposed a general GAN-free reverse distillation framework RealUID for directly utilizing real data to accelerate generation in matching models (diffusion, flow matching, bridge matching, etc.).
Universal Model Routing for Efficient LLM Inference
Wittawat Jitkrittum, Sanjiv Kumar (Google)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes UniRoute, a method for dynamic model routing that represents LLMs as the average error vector of each cluster, enabling the introduction of new models without retraining the router;
Universal Multi-Domain Translation via Diffusion Routers
Duc Kieu (Deakin University), Thin Nguyen (Deakin University)
Image TranslationDiffusion modelRectified FlowAuto EncoderImage
🎯 What it does: This paper proposes a unified multi-domain translation framework called Diffusion Router, which can learn mappings between any two domains, including indirect and direct translations, by utilizing only K-1 central domain paired data.
Universal Properties of Activation Sparsity in Modern Large Language Models
Filip Szatkowski, Bartosz Wójcik
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Systematically evaluate the universality of activation sparsity in modern large language models, propose a top-p sparsification framework without training or threshold calibration, and define 'critical sparsity.' Conduct zero-shot evaluations on models such as Gemma, LLaMA, and Qwen, revealing that larger models exhibit higher sparsity, and input layer sparsity is comparable to or superior to gating layer sparsity.
Universal Value-Function Uncertainties
Moritz Akiya Zanger, Matthijs T. J. Spaan (Delft University of Technology)
Reinforcement LearningImage
🎯 What it does: Proposed the Universal Value-Function Uncertainties (UVU) method, which estimates the deviation uncertainty of the value function by calculating prediction errors between a randomly initialized target network and an online learner.
UniVideo: Unified Understanding, Generation, and Editing for Videos
Cong Wei (University of Waterloo), Wenhu Chen (University of Waterloo)
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: A unified video multimodal model called UniVideo was constructed, capable of accomplishing video understanding, text/image-to-video generation, and editing under multimodal instructions within the same framework.
Unlearning during Training: Domain-Specific Gradient Ascent for Domain Generalization
Di Zhao (University Of Auckland), Yun Sing Koh (University Of Auckland)
Domain AdaptationImageBenchmark
🎯 What it does: Proposed a model-agnostic adaptive unlearning module (IU) that enhances domain generalization by identifying and eliminating domain-specific features in each epoch after training.
Unlearning Evaluation through Subset Statistical Independence
Chenhao Zhang (University of Queensland), Miao Xu (University of Queensland)
Safty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: Proposes a subset-level machine forgetting evaluation method called Split-Half Dependence Evaluation (SDE), which uses HSIC to measure the model's output dependency on subsets, enabling subset-level forgetting effect assessment without retraining or auxiliary classifiers.
Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs
Yiwei Chen (Michigan State University), Sijia Liu (Michigan State University)
Anomaly DetectionSafty and PrivacyExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigate whether large language models (LLMs) leave detectable traces after performing a 'forget' operation, and verify that it is possible to determine whether a model has been forgotten by analyzing only model outputs (text or pre-logit activations).
Unleashing Guidance Without Classifiers for Human-Object Interaction Animation
Ziyin Wang (University Of Illinois Urbana Champaign), Liangyan Gui (University Of Illinois Urbana Champaign)
GenerationTransformerDiffusion modelTextPoint Cloud
🎯 What it does: Propose the LIGHT framework, which generates high-quality human-object interaction animations based on text by utilizing classifier-free, asynchronous denoising guidance. It eliminates manual contact priors and enhances generalization capability through shape spectrum augmentation.
Unleashing LLMs in Bayesian Optimization: Preference-Guided Framework for Scientific Discovery
Xinzhe Yuan (Harbin Institute Of Technology), Qinying Gu (Shanghai Artificial Intelligence Laboratory)
OptimizationTransformerLarge Language ModelPrompt EngineeringBenchmarkPhysics Related
🎯 What it does: Proposed an LLM-Guided Bayesian Optimization (LGBO) framework that integrates the preferences of large language models (LLMs) continuously into the Bayesian optimization loop, achieving efficient scientific experiment search.
Unleashing Perception-Time Scaling to Multimodal Reasoning Models
Yifan Li (Renmin University of China), Minghui Qiu (ByteDance)
Computational EfficiencySupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Investigate the impact of reasoning time expansion on the perception capabilities of large vision-language models (LVLMs), proposing and validating the Perception-Time Scaling (PTS) framework.
Unleashing Scientific Reasoning for Bio-experimental Protocol Generation via Structured Component-based Reward Mechanism
Haoran Sun (Shanghai Artificial Intelligence Laboratory), Xiaosong Wang (Shanghai Artificial Intelligence Laboratory)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Automatically generate executable and logically rigorous biological experiment protocols by training large language models;
UnLoc: Leveraging Depth Uncertainties for Floorplan Localization
Matthias Wüest, Daniel Barath (ETH Zurich)
Pose EstimationDepth EstimationConvolutional Neural NetworkSimultaneous Localization and MappingImageVideo
🎯 What it does: Propose UnLoc, a sequential visual localization method that leverages depth uncertainty for floor plan localization, utilizing a pre-trained monocular depth network and modeling the uncertainty of depth prediction through probability distributions, followed by fusing multi-frame observations with a histogram filter to complete pose estimation;
Unlocking Full Efficiency of Token Filtering in Large Language Model Training
Di Chai (Shanghai University of Finance and Economics), Kai Chen (Hong Kong University of Science and Technology)
Computational EfficiencyTransformerText
🎯 What it does: Proposed and implemented the CENTRIFUGE system, which fully unleashes the efficiency potential of token filtering in LLM training by further filtering activations during backpropagation and converting sparse GEMM into dimension-compressed dense GEMM.
Unlocking Long-Horizon Agentic Search with Large-Scale End-to-End RL
Jiaxuan Gao (Tsinghua University), Yi Wu (Ant Group)
Data SynthesisTransformerLarge Language ModelReinforcement LearningAgentic AITextSequential
🎯 What it does: Proposed ASearcher, a single-model search agent that uses only a search tool and is end-to-end trained via reinforcement learning (RL);
Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization
Boyang Liu (Fudan University), Xuanjing Huang (Fudan University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageMultimodalityChain-of-Thought
🎯 What it does: Propose the Aes-R1 framework to achieve bidirectional reasoning and scoring for image aesthetic assessment. The framework includes automatically generating multi-dimensional aesthetic reasoning data (AesCoT) and optimizing with reinforcement learning based on relative-absolute rewards (RAPO), enhancing model interpretability and scoring accuracy through a two-phase training process (SFT + RL).
Unlocking the Potential of Weighting Methods in Federated Learning Through Communication Compression
Valery Parfenov (Basic Research of Artificial Intelligence Laboratory), Aleksandr Beznosikov (Basic Research of Artificial Intelligence Laboratory)
ClassificationFederated LearningComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an accelerated optimization algorithm called ADI that simultaneously addresses data heterogeneity and communication compression in federated learning.
Unlocking the Power of Co-Occurrence in CLIP: A DualPrompt-Driven Method for Training-Free Zero-Shot Multi-Label Classification
Ming-Kun Xie (RIKEN Center for Advanced Intelligence Project), Masashi Sugiyama (RIKEN Center for Advanced Intelligence Project)
ClassificationConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: Proposed the DualPrompt method, which combines CLIP's discriminative prompts and relevance prompts to achieve zero-shot multi-label classification without training;
Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation
Zhiwei Zhang (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Propose a multi-agent meta-reasoning framework named Dr.MAMR, addressing the issue of lazy agents in multi-agent reasoning.
Unlocking the Value of Text: Event-Driven Reasoning and Multi-Level Alignment for Time Series Forecasting
Siyuan Wang (East China Normal University), Yang Shu (East China Normal University)
TransformerLarge Language ModelContrastive LearningTextMultimodalityTime SeriesAgriculture RelatedFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Developed a dual-branch multi-modal time series prediction framework called VoT, which effectively integrates external text and numerical sequences using event-driven reasoning and multi-level alignment.
Unmasking Backdoors: An Explainable Defense via Gradient-Attention Anomaly Scoring for Pre-trained Language Models
Anindya Sundar Das (Umea University), Monowar Bhuyan (Umea University)
Anomaly DetectionExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose an inference-time anti-backdoor defense framework named X-GRAAD, which detects and neutralizes backdoor triggers by leveraging the abnormal effects of trigger words on attention and gradients in pre-trained language models.
Unmute the Patch Tokens: Rethinking Probing in Multi-Label Audio Classification
Lukas Rauch (University of Kassel), Christoph Scholz (University of Kassel)
ClassificationTransformerAudio
🎯 What it does: This paper addresses multi-label audio classification by proposing a Binarized Prototypical Probe to replace traditional [cls]-token or single-vector aggregation methods, improving the probing performance of frozen models.
Unpacking Human Preference for LLMs: Demographically Aware Evaluation with the HUMAINE Framework
Nora Petrova (Prolific), Enzo Blindow (Prolific)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextSequentialBenchmark
🎯 What it does: This paper proposes the HUMAINE framework, which constructs a dynamic leaderboard by collecting 119,890 multiround natural dialogues between 23,404 respondents and 28 LLMs based on a large-scale, demographically stratified, multidimensional evaluation.
Unraveling the Complexity of Memory in RL Agents: an Approach for Classification and Evaluation
Egor Cherepanov (AXXX), Aleksandr Panov
Recurrent Neural NetworkTransformerLarge Language ModelReinforcement LearningSequential
🎯 What it does: Proposed a unified framework for memory classification and evaluation, distinguishing between short-term and long-term declarative memory as well as procedural memory;
Unsupervised Learning of Efficient Exploration: Pre-training Adaptive Policies via Self-Imposed Goals
Octavio Pappalardo (University College London)
Meta LearningTransformerReinforcement LearningTabular
🎯 What it does: In a reward-free environment, the ULEE method is proposed, which enhances the exploration, adaptation, fine-tuning, and meta-learning performance of RL agents on new tasks through self-set goal meta-learning pre-training.
Unsupervised Representation Learning - an Invariant Risk Minimization Perspective
Yotam Norman (Technion Israel Institute of Technology), Ron Meir (Technion Israel Institute of Technology)
Domain AdaptationRepresentation LearningAuto EncoderImage
🎯 What it does: Propose a new unsupervised invariant risk minimization (IRM) framework without labels, and design two methods within this framework: PICA (Principal Invariant Component Analysis) and VIAE (Variational Invariant AutoEncoder), achieving variable decomposition and feature extraction for cross-environment data.
Unsupervised Representation Learning for 3D Mesh Parameterization with Semantic and Visibility Objectives
AmirHossein Zamani (Autodesk Research), Arianna Rampini (Autodesk Research)
SegmentationRepresentation LearningRecurrent Neural NetworkMesh
🎯 What it does: For the UV mapping task on 3D meshes, the authors propose an unsupervised differentiable framework that simultaneously achieves semantic alignment and view-friendly notch layout while preserving geometric fidelity.
Untraceable DeepFakes via Traceable Fingerprint Elimination
Jiewei Lai (University Of Science And Technology Of China), YUNHAO WANG
GenerationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented a universal black-box method based on multiplicative attacks to eliminate fingerprints left by generative models, making DeepFake images untraceable by attribute models.
Unveiling Downstream Performance Scaling of LLMs: A Clustering-Based Perspective
Chengyin Xu (Bytedance), Chenggang Li (Bytedance)
Large Language ModelTextBenchmark
🎯 What it does: Propose the COD (Clustering-On-Difficulty) framework, which predicts the downstream performance of LLMs across different scales by clustering task difficulty.
Unveiling Perceptual Artifacts: A Fine-Grained Benchmark for Interpretable AI-Generated Image Detection
Yao Xiao (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
Anomaly DetectionExplainability and InterpretabilityTransformerImageBenchmark
🎯 What it does: This paper proposes X-AIGD, a fine-grained interpretable AI-generated image detection benchmark, providing pixel-level multi-level perceptual defect annotations.
Unveiling Super Experts in Mixture-of-Experts Large Language Models
Zunhai Su (Tsinghua University), Kehong Yuan (Tsinghua University)
Explainability and InterpretabilityTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes and systematically studies 'Super Experts' (SE) in Mixture-of-Experts LLMs, a small subset of experts critical for inference.
Unveiling the Basin-Like Loss Landscape in Large Language Models
Huanran Chen, Jun Zhu (Tsinghua University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied the loss landscape of large language models (LMMs), revealing its basinal structure and theoretically and experimentally validating its relationship with alignment fragility, fine-tuning, and jailbreaking; proposed a Gaussian noise optimizer (GO) during pre-training to expand the basin and reduce capability degradation caused by subsequent fine-tuning.
Unveiling the Cognitive Compass: Theory-of-Mind–Guided Multimodal Emotion Reasoning
Meng Luo (National University of Singapore), Wynne Hsu (National University of Singapore)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the hierarchical evaluation benchmark HitEmotion based on Theory of Mind, and designed the TMPO method that uses intermediate mental states as supervision in reinforcement learning to enhance the authenticity and coherence of multimodal emotion reasoning.
Unveiling the Mechanism of Continuous Representation Full-Waveform Inversion: A Wave Based Neural Tangent Kernel Framework
Ruihua Chen (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
Representation LearningNeural Radiance FieldPhysics Related
🎯 What it does: This paper proposes a unified wave-based neural tangent kernel (NTK) framework to theoretically analyze the convergence and robustness of full-waveform inversion (FWI), and based on this, introduces three improved CR-FWI methods: low-rank tensor decomposition, hash-based multi-grid encoding, and hybrid INR-grid representations.
Unveiling the Potential of Diffusion Large Language Model in Controllable Generation
Zhen Xiong (University of Southern California), Yiwei Wang (University of California Merced)
GenerationLarge Language ModelDiffusion modelText
🎯 What it does: Proposes the Self-adaptive Schema Scaffolding (S3) method, leveraging the global attention and reverse denoising mechanisms of diffusion-based large language models to enhance controllability in structured text generation.
UP2You: Fast Reconstruction of Yourself from Unconstrained Photo Collections
Zeyu Cai (Westlake University), Yuliang Xiu (Westlake University)
GenerationTransformerImageMesh
🎯 What it does: Proposed a method called UP2You, an untuned, real-time (approximately 1.5 minutes) approach that directly generates high-quality, textured 3D clothed human models from any number of unstructured, unlabeled outdoor photos.
Urban Socio-Semantic Segmentation with Vision-Language Reasoning
Yu Wang (Wuhan University), Yansheng Li (Wuhan University)
SegmentationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningMultimodality
🎯 What it does: Propose the urban social semantic segmentation task, construct the SocioSeg dataset, and introduce the SocioReasoner framework to achieve multi-stage reasoning segmentation based on vision-language models.
UrbanFeel：A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective
Jun He (Sun Yat-sen University), Xiang Zhang (Sun Yat-sen University)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Designed and constructed a comprehensive benchmark named UrbanFeel, aimed at evaluating the capabilities of multimodal large language models (MLLMs) in urban development understanding and subjective environmental perception.
UrbanGraph: Physics-Informed Spatio-Temporal Dynamic Heterogeneous Graphs for Urban Microclimate Prediction
Weilin Xin (National University of Singapore), Jiawei Yao (Tongji University)
Recurrent Neural NetworkGraph Neural NetworkGraphTime SeriesPhysics Related
🎯 What it does: Proposed a physics-driven dynamic graph network called UrbanGraph for urban microclimate prediction.
UrbanGS: Efficient and Scalable Architecture for Geometrically Accurate Large-Scene Reconstruction
Changbai Li (Beihang University), Baochang Zhang (Beihang University)
GenerationGaussian SplattingImage
🎯 What it does: Proposed a scalable UrbanGS framework for high-precision large-scale urban scene 3D Gaussian Splatting reconstruction;
UrbanVerse: Scaling Urban Simulation by Watching City-Tour Videos
Mingxuan Liu (University of California, Los Angeles), Bolei Zhou (University of California, Los Angeles)
GenerationData SynthesisAutonomous DrivingConvolutional Neural NetworkReinforcement LearningVision Language ModelSimultaneous Localization and MappingImageVideoMesh
🎯 What it does: This paper proposes UrbanVerse, an end-to-end system that generates interactive 3D simulation environments from urban tour videos, creating realistic city layouts, objects, and dynamic agents with physically interactive scenarios;
Use the Online Network If You Can: Towards Fast and Stable Reinforcement Learning
Ahmed Hendawy (Technical University of Darmstadt), Carlo D'Eramo (University of Würzburg)
Reinforcement LearningBenchmark
🎯 What it does: Designed and verified a new target calculation method called Minimum of Online and Target Networks (MINTO), which takes the minimum value between the online network and target network estimates to achieve fast and stable reinforcement learning.
Using maximal information auxiliary variables to improve synthetic data generation based on TabPFN foundation models
Elias Chaibub Neto (Sage Bionetworks)
Data SynthesisSafty and PrivacyTransformerTabular
🎯 What it does: Proposed and implemented a synthetic data generation method based on Maximum Information Auxiliary Variable (MIAV), leveraging TabPFN (and TabICL) within a context learning framework to generate high-fidelity and privacy-preserving synthetic tabular data.
Using Reinforcement Learning to Train Large Language Models to Explain Human Decisions
Jian-Qiao Zhu (Princeton University), Thomas L. Griffiths (Princeton University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Generate chain-of-thought (CoT) explanations during the post-training phase of large language models (LLMs) using reinforcement learning, while simultaneously predicting human risk decisions.
USTBench: Benchmarking and Dissecting Spatiotemporal Reasoning Capabilities of LLMs as Urban Agents
Siqi Lai (Hong Kong University of Science and Technology (Guangzhou)), Hao Liu (Hong Kong University of Science and Technology (Guangzhou))
Large Language ModelReinforcement LearningAgentic AIWorld ModelTextGraphTabularBenchmarkChain-of-Thought
🎯 What it does: Propose USTBench and construct UAgentEnv, an interactive urban environment, to evaluate LLMs in four processes of urban spatiotemporal reasoning: understanding, prediction, planning, and reflection, containing 62,466 structured question-answer pairs and nine real-world urban tasks.
V2P-Bench: Evaluating Video-Language Understanding with Visual Prompts for Better Human-Model Interaction
Yiming Zhao, Feng Zhao (University Of Science And Technology Of China)
RecognitionData-Centric LearningSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Constructed and evaluated V2P-Bench, a vision-language understanding benchmark based on visual prompts.
VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning
Bo Jiang (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
Autonomous DrivingTransformerImageBenchmark
🎯 What it does: Propose the VADv2 end-to-end autonomous driving model, achieving efficient and safe driving through probabilistic planning
Value Flows
Perry Dong (Stanford University), Benjamin Eysenbach (Princeton University)
Reinforcement LearningFlow-based ModelImageTabularOrdinary Differential Equation
🎯 What it does: Propose Value Flows, a reinforcement learning framework that utilizes flow-matching models to estimate the full future return distribution;
Value Matching: Scalable and Gradient-Free Reward-Guided Flow Adaptation
Cristian Perez Jensen (ETH Zürich), Andreas Krause (ETH Zürich)
GenerationData SynthesisReinforcement Learning from Human FeedbackReinforcement LearningFlow-based ModelImageBiomedical DataStochastic Differential Equation
🎯 What it does: Proposed an online Value Matching (VM) algorithm that achieves gradient-agnostic adaptation to arbitrary reward functions by learning a value function, without updating the parameters of pre-trained flow models;
VARestorer: One-Step VAR Distillation for Real-World Image Super-Resolution
Yixuan Zhu (Tsinghua University), Jie Zhou (Tsinghua University)
Super ResolutionKnowledge DistillationPrompt EngineeringAuto EncoderImageText
🎯 What it does: This paper proposes VARestorer, a single-step real image super-resolution framework based on a visual autoregressive model (VAR), which can directly generate high-quality images from low-quality images in a single forward pass.