NeurIPS 2025 Papers — Page 50

Unbalanced Optimal Total Variation Transport: A Theoretical Approach to Spatial Resource Allocation Problems

Nhan-Phu Chung (Institute of Applied Mathematics University of Economics Ho Chi Minh City), Zehao Li (Guanghua School of Management Peking University)

🎯 What it does: A class of unbalanced weak optimal transport models with total variation penalties is proposed and analyzed for spatial resource allocation problems.

Unbiased Prototype Consistency Learning for Multi-Modal and Multi-Task Object Re-Identification

Zhongao Zhou (Wuhan University), Mang Ye (Wuhan University)

🎯 What it does: A unified multi-modal, multi-task object re-identification (M³T-ReID) framework is proposed to address the problem of simultaneous retrieval across modalities and multi-modalities for different categories (people, vehicles).

Unbiased Sliced Wasserstein Kernels for High-Quality Audio Captioning

Manh Luong (Monash University), Lizhen Qu (Monash University)

🎯 What it does: The ACUS framework and unbiased sliced Wasserstein RBF kernel are proposed to address exposure bias and cross-modal alignment issues in audio caption generation.

Uncertain Knowledge Graph Completion via Semi-Supervised Confidence Distribution Learning

Tianxing Wu (Southeast University), Haofen Wang (Tongji University)

🎯 What it does: This paper proposes a semi-supervised confidence distribution learning-based UKG completion framework called ssCDL, which can simultaneously perform confidence prediction and link prediction; by transforming single confidence into a confidence distribution, it enhances the supervisory information of minority or unseen confidence levels, and utilizes a Pseudo Confidence Distribution Generator (PCDG) and meta-learning strategies for self-supervised training of unlabeled triples to improve embedding learning quality.

Uncertainty Estimation by Flexible Evidential Deep Learning

Taeseong Yoon (Korea Advanced Institute of Science and Technology), Heeyoung Kim (Korea Advanced Institute of Science and Technology)

🎯 What it does: A flexible evidence deep learning (F-EDL) framework is proposed, using a flexible Dirichlet distribution to infer the uncertainty of classification tasks in a single forward pass.

Uncertainty Estimation on Graphs with Structure Informed Stochastic Partial Differential Equations

Fred Xu (Block Inc University of California Los Angeles), Thomas Markovich (Block Inc)

🎯 What it does: A framework based on structural information for Stochastic Partial Differential Equations (SPDE) is proposed, which enhances the uncertainty estimation of graph data by injecting spatially correlated noise (Φ-Wiener process) into graph neural networks, particularly for out-of-distribution (OOD) detection in low label informative scenarios.

Uncertainty Quantification for Deep Regression using Contextualised Normalizing Flows

Adriel Sosa Marco (Arquimea Research Center), Simos Gerasimou (University of York)

🎯 What it does: A posterior uncertainty quantification method MCNF is proposed in deep regression models, utilizing Monte Carlo dropout to generate prior distributions and obtaining complete predictive distributions through contextual regularized neural spline flow.

Uncertainty Quantification for Physics-Informed Neural Networks with Extended Fiducial Inference

Frank Shih (Memorial Sloan Kettering Cancer Center), Faming Liang (Purdue University)

🎯 What it does: This paper proposes a method for uncertainty quantification in physics-informed neural networks (PINNs) based on extended fiducial inference (EFI), utilizing a narrow-neck supernetwork to learn PINN parameters and achieve reliable confidence set construction through random error interpolation.

Uncertainty Quantification with the Empirical Neural Tangent Kernel

Joseph Wilson (University of Queensland), Fred Roosta (University of Queensland)

🎯 What it does: A post-hoc sampling Bayesian uncertainty quantification method NUQLS is proposed, which linearizes the trained over-parameterized network and uses gradient descent/stochastic gradient descent sampling to construct an efficient deep ensemble to approximate the predictive distribution.

Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design

Lianghong Chen (Western University), Pingzhao Hu (Western University)

🎯 What it does: A 3D molecular diffusion model framework guided by multi-objective reinforcement learning based on uncertainty perception is proposed to generate molecules that meet multiple drug-related properties.

Uncertainty-aware Preference Alignment for Diffusion Policies

Runqing Miao (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)

🎯 What it does: This paper proposes an uncertainty-aware preference alignment algorithm for diffusion strategies, called Diff-UAPA, which optimizes directly on the policy without the need to learn a reward function.

Uncertainty-Based Smooth Policy Regularisation for Reinforcement Learning with Few Demonstrations

Yujie Zhu (University of Warwick), Giovanni Montana (University of Warwick)

🎯 What it does: The SPReD framework is proposed, which achieves smooth policy regularization of demonstration behaviors by using ensemble Q-value distributions and dynamically calculates uncertainty weights at each step to determine imitation strength.

Uncertainty-Calibrated Prediction of Randomly-Timed Biomarker Trajectories with Conformal Bands

Vasiliki Tassopoulou (University of Pennsylvania), Christos Davatzikos (University of Pennsylvania)

🎯 What it does: This paper proposes a segmented adaptive consistent prediction framework for biomarker trajectories at randomly sampled time points, and based on this, constructs a confidence interval that covers the entire trajectory; further, it designs conditionally consistent predictions based on population stratification and a lower bound of the progression rate based on the lower limit of the prediction interval to improve the risk identification of conversion from MCI to Alzheimer's disease;

Uncertainty-Guided Exploration for Efficient AlphaZero Training

Scott Cheng (Pennsylvania State University), Mahmut Kandemir

🎯 What it does: During the self-play process of AlphaZero, the Label Change Rate (LCR) is used to measure high uncertainty positions, and multiple variants are generated from these positions. Runner-up moves are reused for the games, and the final results of the multiple variants are averaged as value labels, thereby reducing label variance and enhancing exploration efficiency.

Uncertainty-Informed Meta Pseudo Labeling for Surrogate Modeling with Limited Labeled Data

Xingyu Ren (Zhejiang University), Dong Ni (Zhejiang University)

🎯 What it does: A meta pseudo-label learning framework based on uncertainty, UMPL, is proposed to enhance the surrogate modeling performance of physical systems under limited labeled data.

Uncertainty-quantified Rollout Policy Adaptation for Unlabelled Cross-domain Video Temporal Grounding

Jian Hu (Hong Kong University of Science and Technology), Kun Shao (Huawei Noah's Ark Lab)

🎯 What it does: A data-efficient cross-domain video segment localization method called URPA is proposed, which utilizes GRPO to generate diverse rollouts and estimate uncertainty, enabling testing adaptation with only a small amount of unlabeled target domain videos.

Uncertainty-Sensitive Privileged Learning

Fan-Ming Luo (Nanjing University), Yang Yu (Nanjing University)

🎯 What it does: A privilege learning framework based on uncertainty perception, USPL, is proposed, which utilizes an observation encoder to estimate the uncertainty of missing information and inputs it into the privileged policy, achieving alignment between the deployment policy and the privileged policy behavior.

UnCLe: Towards Scalable Dynamic Causal Discovery in Non-linear Temporal Systems

Tingzhu Bi (Peking University), Ping Wang (Peking University)

🎯 What it does: The UnCLe method is proposed, which uses parameter-shared TCN Uncoupler/Recoupler to semantically decompose time series and achieves scalable dynamic causal graph learning through autoregressive dependency matrices and temporal perturbations.

Uncoupled and Convergent Learning in Monotone Games under Bandit Feedback

Jing Dong (Chinese University of Hong Kong), Yaoliang Yu (University of Waterloo)

🎯 What it does: A learning algorithm based on mirror descent is proposed, utilizing double regularization (self-conjugate barrier function h and convex regularization p) to achieve convergence in multi-player monotonic smooth games with bandit feedback and strong coupling decoupling, and the algorithm is extended to time-varying games.

Uncover Governing Law of Pathology Propagation Mechanism Through A Mean-Field Game

Tingting Dan (University of North Carolina), Guorong Wu (University of North Carolina)

🎯 What it does: This paper proposes the MFG4AD framework, which combines mean field game theory, the Wasserstein-1 Lagrangian GAN with physical information, and symbolic regression to predict the temporal evolution of tau protein in Alzheimer's disease at high resolution on the cortical surface, and to explain the interaction mechanism between Aβ and tau.

Uncovering a Universal Abstract Algorithm for Modular Addition in Neural Networks

Gavin McCracken (Mila), Jonathan Love (Leiden University)

🎯 What it does: This study investigates the abstract algorithms for learning in modular addition tasks using neural networks and proposes the Approximate Chinese Remainder Theorem (aCRT) as a unified explanation.

Uncovering the Spectral Bias in Diagonal State Space Models

Ruben Solozabal (Mohammed bin Zayed University of Artificial Intelligence), Martin Takáč (Mohammed bin Zayed University of Artificial Intelligence)

🎯 What it does: This paper studies the initialization methods of diagonal state space models (Diagonal SSM) in the frequency domain, proposing a new S4D-DFouT initialization scheme, and experimentally validating its advantages in long sequence tasks.

Under the Shadow: Exploiting Opacity Variation for Fine-grained Shadow Detection

Xiaotian Qiao (Xidian University), Jiangtao Cui (Xidian University)

🎯 What it does: A fine-grained shadow detection method is proposed, which predicts continuous shadow masks using variations in shadow opacity.

Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation

Xiaoyu Yue (Shanghai AI Laboratory), Luping Zhou (Shanghai AI Laboratory)

🎯 What it does: The study investigates the bottlenecks of autoregressive image generation models in visual understanding and proposes a self-guided training framework ST-AR to enhance the model's learning of high-level visual semantics.

Understanding Adam Requires Better Rotation Dependent Assumptions

Tianyue H. Zhang (Mila), Charles Guille-Escuret (Mila)

🎯 What it does: This study investigates the sensitivity of the Adam optimizer to rotations in the parameter space, systematically evaluating the impact of random rotations of varying ranges and structured SVD rotations on the training of Transformers, and verifying the correlation between update orthogonality and Adam's performance.

Understanding and Enhancing Mask-Based Pretraining towards Universal Representations

Mingze Dong (Yale University), Yuval Kluger (Yale University)

🎯 What it does: This paper explains the effectiveness of masked pre-training through high-dimensional linear regression theory and proposes a Random Masked Autoencoder (RMAE2) scheme based on this.

Understanding and Enhancing Message Passing on Heterophilic Graphs via Compatibility Matrix

Zhuonan Zheng (Zhejiang University), Jiajun Bu (Zhejiang University)

🎯 What it does: This study investigates the effectiveness of message passing mechanisms in heterogeneous graphs and proposes the CMGNN model, which utilizes a Compatibility Matrix (CM) to enhance node distinguishability.

Understanding and Improving Adversarial Robustness of Neural Probabilistic Circuits

Weixin Chen (University of Illinois Urbana-Champaign), Han Zhao (University of Illinois Urbana-Champaign)

🎯 What it does: A robust neural probabilistic circuit (RNPC) is proposed, enhancing adversarial robustness through category-level ensemble in the concept bottleneck model.

Understanding and Improving Fast Adversarial Training against $l_0$ Bounded Perturbations

Xuyang Zhong (City University of Hong Kong), Chen Liu (City University of Hong Kong)

🎯 What it does: This paper studies fast adversarial training against l0 sparse perturbations, analyzing and addressing the catastrophic overfitting (CO) problem caused by one-step attacks.

Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference

Jiayi Yuan (Rice University), Zirui Liu (University of Minnesota Twin Cities)

🎯 What it does: A systematic study on the impact of floating-point precision (BF16, FP16, FP32) and runtime configurations (GPU type, number, batch size) on the reproducibility of large language model inference, and the proposal of the LayerCast mixed precision inference scheme.

Understanding and Rectifying Safety Perception Distortion in VLMs

Xiaohan Zou (Pennsylvania State University), Lu Lin (Pennsylvania State University)

🎯 What it does: This study investigates the failure mechanisms of visual-language models in safety alignment and proposes a method for activation offset correction during inference called ShiftDC.

Understanding Bias Terms in Neural Representations

Weixiang Zhang (Shenzhen International Graduate School Tsinghua University), Zhi Wang (Shenzhen International Graduate School Tsinghua University)

🎯 What it does: This paper studies the role of the bias term in implicit neural representations (INR) and finds that the bias term does not enhance nonlinear expression but is mainly used to eliminate spatial aliasing caused by the symmetry of coordinates and periodic activation functions. Based on this finding, the authors propose a new architecture that freezes the bias only at the input layer and introduces pretrained features (Feat-Bias), which is validated on image reconstruction and classification tasks.

Understanding challenges to the interpretation of disaggregated evaluations of algorithmic fairness

Stephen R Pfohl, Alexander Nicholas D'Amour

🎯 What it does: This study investigates the misleading nature of group discrete evaluation in algorithm fairness assessment and proposes a causal graph-based analysis of group difference stability and a controlled evaluation method.

Understanding Contrastive Learning via Gaussian Mixture Models

Parikshit Bansal (University of Texas at Austin), sujay sanghavi

🎯 What it does: This paper explores the performance of contrastive learning (InfoNCE, SimSiam) in Gaussian Mixture Models (GMM) through theoretical analysis. It proves that under the setting of noise-enhanced GMM, contrastive learning can find the Fisher subspace, achieving dimensionality reduction effects equivalent to supervised LDA, and extends this framework to multimodal CLIP-GMM.

Understanding Data Influence in Reinforcement Finetuning

Haoru Tan (University of Hong Kong), XIAOJUAN QI

🎯 What it does: Proposes RFT-Inf, which is used to evaluate the impact of each sample on the final reward during reinforcement fine-tuning and to perform data selection accordingly.

Understanding Differential Transformer Unchains Pretrained Self-Attentions

Chaerin Kong (Twelve Labs), Nojun Kwak (Seoul National University)

🎯 What it does: A systematic analysis of the internal mechanisms of DIFF Transformer is conducted, and a lightweight method called DEX is proposed to transfer the advantages of DIFF to pre-trained LLMs.

Understanding Fairness and Prediction Error through Subspace Decomposition and Influence Analysis

Enze Shi (University of Alberta), Bei Jiang (University of Alberta)

🎯 What it does: This paper proposes a direct adjustment of data representation at the feature representation level through subspace decomposition and influence function analysis, achieving a controllable trade-off between fairness and predictive performance.

Understanding Generalization in Physics Informed Models through Affine Variety Dimensions

Takeshi Koshizuka (University of Tokyo), Issei Sato (University of Tokyo)

🎯 What it does: This paper proposes a unified residual form and studies the generalization performance of physics-informed regression from a geometric perspective, proving that it is determined by the affine family dimension generated by equation constraints;

Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws

Zhixuan Pan (Institute for Interdisciplinary Information Sciences Tsinghua University), Jian Li (Institute for Interdisciplinary Information Sciences Tsinghua University)

🎯 What it does: This paper proposes a 'grammar-knowledge' hierarchical generative model by viewing the training of large language models as two parts of encoding in the Kolmogorov structure function. It derives theoretical scaling laws relating data and model size, explains phenomena such as the order of knowledge acquisition in LLMs and low-frequency knowledge hallucinations, and finally validates the theoretical predictions on synthetic data.

Understanding outer learning rates in Local SGD

Ahmed Khaled (Princeton University), Manzil Zaheer (Meta)

🎯 What it does: This study investigates the impact of external optimizers (learning rate, momentum, acceleration) on convergence in Local SGD, proposes a new convergence theorem, and validates theoretical predictions through experiments.

Understanding Parametric and Contextual Knowledge Reconciliation within Large Language Models

Jun Zhao (Fudan University), Xuanjing Huang (Fudan University)

🎯 What it does: This paper proposes an entity-aware probing framework that dynamically tracks entity flows in retrieval-augmented generative LLMs, revealing the fusion and competition mechanisms of parametric knowledge and contextual knowledge within Transformer layers.

Understanding Prompt Tuning and In-Context Learning via Meta-Learning

Tim Genewein (Google DeepMind), Marcus Hutter

🎯 What it does: Explores the Bayesian perspective of prompt tuning and meta-learning, and empirically validates the effectiveness of soft prefixes on a simplified coin flipping sequence.

Understanding protein function with a multimodal retrieval-augmented foundation model

Timothy Fei Truong Jr, Tristan Bepler (OpenProtein.AI)

🎯 What it does: We propose PoET-2, a multimodal retrieval-augmented protein language model that can accurately predict the effects of mutations and generate new protein sequences under zero-shot and few-shot supervision.

Understanding Representation Dynamics of Diffusion Models via Low-Dimensional Modeling

Xiao Li (University of Michigan), Qing Qu (University of Michigan)

🎯 What it does: The study characterizes the dynamic of diffusion models under different noise levels, revealing why the performance exhibits a unimodal trend and associates it with the model's generalization ability.

Understanding Softmax Attention Layers:\\ Exact Mean-Field Analysis on a Toy Problem

Elvis Dohmatob (Concordia University)

🎯 What it does: This paper proposes and theoretically analyzes the performance of softmax attention in the single-location regression problem, providing analytical formulas for risk, gradient flow, and convergence properties.

Understanding the Evolution of the Neural Tangent Kernel at the Edge of Stability

Kaiqi Jiang (Princeton University), Yuanzhi Li (Carnegie Mellon University)

🎯 What it does: This study investigates the evolution of the neural tangent kernel (NTK) feature vectors of neural networks under the edge of stability (EoS) during gradient descent, as well as their alignment behavior with the target vectors over time.

Understanding the Gain from Data Filtering in Multimodal Contrastive Learning

Divyansh Pareek (University of Washington), Simon Shaolei Du

🎯 What it does: The theoretical advantages of filtering based on teacher models in multimodal contrastive learning are studied, proving that the error after filtering depends better on η.

Understanding the Generalization of Stochastic Gradient Adam in Learning Neural Networks

Xuan Tang (University of Hong Kong), Difan Zou (University of Hong Kong)

🎯 What it does: This study investigates the impact of Adam and AdamW on the generalization of a two-layer CNN under different batch sizes and weight decay, providing both theoretical and experimental results.

Understanding while Exploring: Semantics-driven Active Mapping

Liyan Chen (Stevens Institute of Technology), Philippos Mordohai (Stevens Institute of Technology)

🎯 What it does: This paper proposes an active semantic mapping framework called ActiveSGM based on 3D Gaussian Splatting, which can actively select the most informative viewpoints and generate high-quality geometric and semantic maps in unknown environments.

Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization

Alexander Long (Pluralis Research), Sameera Ramasinghe (Pluralis Research)

🎯 What it does: This paper proposes Unextractable Protocol Models (UPMs), which allow for the processing of model shards in decentralized training and inference, making it impossible for any participant to obtain the complete weights through time-varying reversible transformations.

Unfolding the Black Box of Recurrent Neural Networks for Path Integration

Tianhao Chu (Peking University), Si Wu (Peking University)

🎯 What it does: The path integration task is completed by training a recurrent neural network (RNN), and the internal mechanisms of the model are interpreted using neuroscientific priors (neuron type discrimination, connection pruning, continuous attractor analysis). A hierarchical information pathway from band cells to grid cells is discovered, and a new neural circuit model is constructed based on this.

Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction

Yifei Wang (Peking University), He Sun (Peking University)

🎯 What it does: By constructing a unified theoretical framework called Uni-Instruct, a one-step diffusion model was integrated and trained, unifying more than ten existing distillation methods.

Uni-LoRA: One Vector is All You Need

Kaiyang Li (University of Connecticut), Shihao Ji (University of Connecticut)

🎯 What it does: A unified framework called Uni-LoRA is proposed, which maps all LoRA variants to projections in a low-dimensional subspace, allowing LLM fine-tuning with the training of just one vector.

Uni-MuMER: Unified Multi-Task Fine-Tuning of Vision-Language Model for Handwritten Mathematical Expression Recognition

Yu Li (Peking University), Liangcai Gao (Peking University)

🎯 What it does: A unified multi-task fine-tuning framework called Uni-MuMER has been constructed to enhance the performance of VLM in the handwritten mathematical expression recognition (HMER) task.

Uni-RL: Unifying Online and Offline RL via Implicit Value Regularization

Haoran Xu (University of Texas at Austin), Amy Zhang (University of Texas at Austin)

🎯 What it does: This paper proposes Uni-RL, a unified offline/online/offline-to-online reinforcement learning framework that normalizes different learning scenarios using implicit value regularization.

UniCTokens: Boosting Personalized Understanding and Generation via Unified Concept Tokens

Ruichuan An (Peking University), Wentao Zhang (Peking University)

🎯 What it does: The UniCTokens framework is proposed, achieving personalized understanding and generation of VLM through unified concept tokens.

UniDomain: Pretraining a Unified PDDL Domain from Real-World Demonstrations for Generalizable Robot Task Planning

Haoming Ye (Shanghai Jiao Tong University), Panpan Cai (Shanghai Jiao Tong University)

🎯 What it does: The UniDomain framework is proposed, utilizing large-scale robot demonstrations to pre-train a unified PDDL domain for zero-shot task planning.

Unified 2D-3D Discrete Priors for Noise-Robust and Calibration-Free Multiview 3D Human Pose Estimation

Geng Chen (Beijing University of Posts and Telecommunications), Jingyu Wang (Beijing University of Posts and Telecommunications)

🎯 What it does: This paper proposes embedding a unified discrete codebook (UniCodebook) into a calibration-free multi-view 3D human pose estimation framework to enhance noise robustness through discrete priors.

Unified all-atom molecule generation with neural fields

Matthieu Kirchmeyer (Prescient Design, Genentech), Saeed Saremi (Genentech)

🎯 What it does: A unified atomic-level molecular generation framework called FuncBind is proposed, which can generate small molecules, cyclic peptides, and antibody CDRs under the condition of a given target protein structure.

Unified Multimodal Chain-of-Thought Reward Model through Reinforcement Fine-Tuning

Yibin Wang (Fudan University), Jiaqi Wang (Shanghai AI Lab)

🎯 What it does: A unified multimodal chain reasoning reward model, UNIFIEDREWARD-THINK, is proposed, capable of performing multidimensional, step-by-step long-chain reasoning in visual understanding and generation tasks.

Unified Reinforcement and Imitation Learning for Vision-Language Models

Byung-Kwan Lee (NVIDIA, KAIST), Yueh-Hua Wu (NVIDIA)

🎯 What it does: This paper proposes a unified framework for reinforcement learning and imitation learning, RIL, which significantly improves the generation quality and performance of small visual language models by learning from the outputs of large models without changing the model structure.

Unified Scaling Laws for Compressed Representations

Andrei Panferov (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)

🎯 What it does: This paper studies the unified scaling law of compressed representations, explores the interaction between the scaling law and compression formats, verifies the applicability of the universal scaling law, and proposes a simple metric based on representational capacity to predict the parameter efficiency of different compressed representations.

Unified Transferability Metrics for Time Series Foundation Models

Weiyang Zhang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)

🎯 What it does: The TEMPLATE framework is proposed for quickly evaluating the transfer performance of time series pre-trained models without fine-tuning. The framework includes three complementary metrics: dependency learning score, pattern learning score, and task adaptation score.

Uniform Wrappers: Bridging Concave to Quadratizable Functions in Online Optimization

Mohammad Pedramfar (Mila Quebec AI Institute McGill University), Vaneet Aggarwal (Purdue University)

🎯 What it does: A framework called Uniform Wrappers is proposed to convert convex optimization algorithms into online algorithms that can handle more general 'quadratizable' functions, and guidelines are provided for the transition from regret proofs in convex optimization to proofs in quadratizable optimization.

Unifying and Enhancing Graph Transformers via a Hierarchical Mask Framework

Yujie Xing (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)

🎯 What it does: A unified hierarchical masking framework is proposed, and the M3Dphormer Graph Transformer is designed to efficiently model multi-level interactions in graphs through multi-level masks (local, clustering, global) and a two-layer expert routing.

Unifying Appearance Codes and Bilateral Grids for Driving Scene Gaussian Splatting

Nan Wang (BAAI), Hao Zhao (AIR)

🎯 What it does: A multi-scale bilateral grid is proposed, unifying global appearance codes and pixel-level bilateral grids for high-quality Gaussian splatting reconstruction in driving scenes.

Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning

Haolin Yang, Naoya Inoue

🎯 What it does: A unified framework based on the geometry of hidden states is constructed to explain the roles of attention heads and task vectors in ICL.

Unifying Proportional Fairness in Centroid and Non-Centroid Clustering

Benjamin Cookson (University of Toronto), Ziqi Yu (University of Toronto)

🎯 What it does: This paper studies a new clustering framework - semi-centroid clustering, which integrates both centroid and non-centroid losses, and explores its algorithm and approximation guarantees under proportional fairness (core and fully justified representation).

Unifying Re-Identification, Attribute Inference, and Data Reconstruction Risks in Differential Privacy

Bogdan Kulynych (Lausanne University Hospital), Jean Louis Raisaro (Lausanne University Hospital)

🎯 What it does: A unified f-DP framework is proposed to quantify privacy risks in re-identification, attribute inference, and data reconstruction.

Unifying Reconstruction and Density Estimation via Invertible Contraction Mapping in One-Class Classification

Xiaolei Wang (Xi'an Jiaotong-Liverpool University), Jimin XIAO

🎯 What it does: A reversible contraction mapping framework is proposed, unifying reconstruction and density estimation under one-class classification (OCC). An iterative contraction mapping is used to approximate abnormal samples to the normal manifold, and the likelihood of the input is estimated through the inverse mapping, using a combination of reconstruction error and log-likelihood as the anomaly score.

Unifying Symbolic Music Arrangement: Track-Aware Reconstruction and Structured Tokenization

Longshen Ou (National University of Singapore), Ye Wang (Mohamed bin Zayed University of Artificial Intelligence)

🎯 What it does: This paper proposes a unified symbolic music multi-track layout framework, achieving reconstruction and addition between arbitrary tracks through a self-supervised objective based on segment reconstruction, supporting various layout scenarios such as rewriting, simplification, and addition.

Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with Large Language Models

Siwei Zhang (Fudan University), Jiawei Zhang (University of California)

🎯 What it does: This paper proposes a framework called CROSS for handling Temporal Text‑Attributed Graphs (TTAG) that contain dynamic text attributes. It first performs time-aware dynamic semantic extraction of node text context, and then hierarchically fuses the extracted semantics with graph structural information to generate a unified, semantically-structurally coordinated representation.

UniGen: Enhanced Training & Test-Time Strategies for Unified Multimodal Understanding and Generation

Rui Tian (Fudan University), Afshin Dehghan (Apple)

🎯 What it does: This paper presents UniGen, a unified multimodal large language model capable of simultaneous image understanding and generation.

UniGist: Towards General and Hardware-aligned Sequence-level Long Context Compression

Chenlong Deng (Renmin University of China), Zhicheng Dou (Renmin University of China)

🎯 What it does: This paper proposes UniGist, a unified gist-token level sequence compression framework that eliminates block-level training, allowing for the processing of long sequences in one go and dynamically discarding irrelevant raw tokens during inference, significantly reducing KV cache usage and improving inference efficiency.

UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains

Duo Wang (Beihang University), Junjie Wu (Beihang University)

🎯 What it does: This paper proposes UniGTE, a unified encoder-decoder framework that achieves zero-shot graph learning across tasks and domains through instruction tuning. It utilizes a pre-trained autoregressive LLM as the encoder, incorporates learnable alignment tokens, and employs structure-aware graph-text attention. The decoder is a frozen LLM that generates task predictions and reconstructs graph prompts based on the alignment tokens.

UniLumos: Fast and Unified Image and Video Relighting with Physics-Plausible Feedback

Pengwei Liu (Zhejiang University), Fan Wang (DAMO Academy, Alibaba Group)

🎯 What it does: A unified image and video lighting reshaping framework called UniLumos is proposed, which achieves physically feasible lighting control while maintaining scene geometry and temporal consistency.

UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning

Nan Song (Fudan University), Li Zhang (Fudan University)

🎯 What it does: A unified motion framework called UniMotion is proposed, capable of simultaneously handling motion simulation, trajectory prediction, and self-driving planning tasks.

UniMRSeg: Unified Modality-Relax Segmentation via Hierarchical Self-Supervised Compensation

Xiaoqi Zhao (Yale University), Xiaofeng Liu (Yale University)

🎯 What it does: A unified multimodal segmentation framework, UniMRSeg, is proposed, achieving robust segmentation of missing modalities through hierarchical self-supervised compensation.

UniPixel: Unified Object Referring and Segmentation for Pixel-Level Visual Reasoning

Ye Liu (Hong Kong Polytechnic University), Chang Wen Chen (Hong Kong Polytechnic University)

UniRelight: Learning Joint Decomposition and Synthesis for Video Relighting

Kai He (NVIDIA), Zian Wang (NVIDIA)

🎯 What it does: A single video diffusion model is proposed, jointly estimating scene albedo and relit videos under arbitrary lighting conditions, achieving end-to-end image relighting and intrinsic decomposition.

UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection

Jigang Fan (Peking University), Liwei Wang (Peking University)

🎯 What it does: A multi-site dataset at the UniProt level, UniSite-DS, has been constructed, and an end-to-end Ligand binding site detection framework, UniSite-1D/3D, has been proposed, along with the introduction of an IoU-based AP evaluation metric.

UniteFormer: Unifying Node and Edge Modalities in Transformers for Vehicle Routing Problems

Dian Meng (Dalian University of Technology), Yaqing Hou (Dalian University of Technology)

🎯 What it does: A unified neural solver called UniteFormer is proposed, capable of handling vehicle routing problems (VRP) with only node input, only edge input, and a mixture of node and edge inputs within the same model.

UniTok: a Unified Tokenizer for Visual Generation and Understanding

Chuofan Ma (University of Hong Kong), XIAOJUAN QI

🎯 What it does: A unified visual tokenizer, UniTok, is proposed, which supports both visual generation and understanding tasks with the same tokenizer.

UniTraj: Learning a Universal Trajectory Foundation Model from Billion-Scale Worldwide Traces

Yuanshao Zhu (Southern University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)

🎯 What it does: A universal trajectory foundation model called UniTraj is proposed, along with the construction of a global trajectory dataset named WorldTrace, covering 70 countries.

UniTransfer: Video Concept Transfer via Progressive Spatio-Temporal Decomposition

guojunlei, Weiwei Xu (Zhejiang University)

🎯 What it does: A video concept transfer framework named UniTransfer based on Diffusion Transformer is proposed, which achieves controllable transfer of foreground, background, and motion through spatial and temporal decomposition.

Universal Causal Inference in a Topos

Sridhar Mahadevan (University of Massachusetts), Sridhar Mahadevan (Adobe Research)

🎯 What it does: This paper proposes Topos Causal Models (TCM) within the framework of topological categories and demonstrates their universality and interpretability in causal inference.

Universal Cross-Tokenizer Distillation via Approximate Likelihood Matching

Benjamin Minixhofer (University of Cambridge), Edoardo Ponti

🎯 What it does: This paper proposes a general cross-tokenizer distillation method called Approximate Likelihood Matching (ALM), which enables effective distillation when the teacher and student use completely different tokenizers (e.g., subword → byte) and supports pure distillation rather than just serving as an auxiliary objective.

Universal Few-shot Spatial Control for Diffusion Models

Kiet T Nguyen, Seunghoon Hong (Korea Advanced Institute of Science and Technology)

🎯 What it does: Proposes Universal Few-shot Control (UFC), achieving fine control over arbitrary spatial conditions with only a small amount of labeled data on pre-trained text-to-image diffusion models.

Universal Sequence Preconditioning

Annie Marsden (Google Deepmind), Elad Hazan (Princeton University)

🎯 What it does: A general sequence preprocessing method is proposed, which convolves the target sequence with Chebyshev or Legendre polynomial coefficients to enhance the learnability of linear dynamic systems.

Universal Video Temporal Grounding with Generative Multi-modal Large Language Models

Zeqian Li (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)

🎯 What it does: The UniTime model is proposed, utilizing a multimodal large language model to achieve general video temporal localization.

Universal Visuo-Tactile Video Understanding for Embodied Interaction

Yifan Xie (Tsinghua University), Wenbo Ding (Sun Yat-sen University)

🎯 What it does: This paper presents VTV-LLM, a multimodal large language model capable of understanding visual-haptic videos and performing haptic reasoning.

Universally Invariant Learning in Equivariant GNNs

Jiacheng Cen (Renmin University of China), Wenbing Huang (Renmin University of China)

🎯 What it does: A universal dynamic method Uni-EGNN is proposed, achieving the integrity of graph neural networks invariant to alignment transformations (capable of approximating any continuous equivariant function).

UniViT: Unifying Image and Video Understanding in One Vision Encoder

Feilong Tang (Monash University), Zongyuan Ge (DeepGlint)

🎯 What it does: Proposes UniViT, a unified learning framework for structured semantics of images and videos, achieving cross-modal representation through event-level and object-level clustering and contrastive learning.

UniZyme: A Unified Protein Cleavage Site Predictor Enhanced with Enzyme Active-Site Knowledge

Chenao Li (Hong Kong University of Science and Technology), Enyan Dai (Hong Kong University of Science and Technology)

🎯 What it does: A unified protease cleavage site prediction model, UniZyme, is proposed, which can generalize across different enzymes and improve prediction accuracy by utilizing knowledge of enzyme active sites.

Unlabeled Data Can Provably Enhance In-Context Learning of Transformers

Renpu Liu (University of Virginia), Jing Yang (University of Virginia)

🎯 What it does: This paper studies the enhancement of In-Context Learning (ICL) effects in Transformers using unlabeled data, proposing the 'Augmented ICL' framework;

Unlabeled Data Improves Fine-Grained Image Zero-shot Classification with Multimodal LLMs

Yunqi Hong (University of California), Cho-Jui Hsieh (University of California)

🎯 What it does: This study investigates the enhancement of fine-grained image zero-shot classification performance of multimodal large language models using self-supervised prompt learning with unlabeled data.

Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM

Xiaoyu Wu (Rice University), Steven Wu

🎯 What it does: This paper proposes an attack method that can leak forgotten data even after exact unlearning of large language models, utilizing the differences between pre-unlearning and post-unlearning models to enhance data extraction success rates through guided generation and vocabulary filtering.

Unlearning-Aware Minimization

Hoki Kim (Chung Ang University), Sangwon Yoon (Ministry of Justice Republic of Korea)

🎯 What it does: A second-order extremum optimization framework named Unlearning-Aware Minimization (UAM) is proposed and validated, which effectively removes specified training samples from the model without retraining, while maintaining performance on the remaining samples.

Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations

Chaofan Gan (Shanghai Jiao Tong University), Weiyao Lin (Shanghai Jiao Tong University)

🎯 What it does: This study investigates the phenomenon of 'massive activation' that occurs in Diffusion Transformers (DiTs) and proposes a training-free feature extraction framework called DiTF based on AdaLN, which can suppress massive activation and improve performance in visual correspondence tasks.

Unleashing Foundation Vision Models: Adaptive Transfer for Diverse Data-Limited Scientific Domains

Qiankun Li (University of Science and Technology of China), Zengfu Wang (University of Science and Technology of China)

🎯 What it does: This paper proposes the Cluster Attention Adapter (CLAdapter), which utilizes the knowledge of large pre-trained models to provide adaptive feature transfer and fine-tuning solutions for data-scarce scientific tasks.