NeurIPS 2025 Papers — Page 39
Conference on Neural Information Processing Systems · 5275 papers
RESAnything: Attribute Prompting for Arbitrary Referring Segmentation
Ruiqi Wang (Simon Fraser University), Hao Zhang (Simon Fraser University)
SegmentationTransformerLarge Language ModelPrompt EngineeringImageTextChain-of-Thought
🎯 What it does: This paper proposes a zero-shot, open-vocabulary arbitrary reference segmentation method called RESAnything, which achieves precise segmentation of complex implicit expressions using attribute prompting and multi-metric selection.
Rescaled Influence Functions: Accurate Data Attribution in High Dimension
Ittai Rubinstein (Massachusetts Institute of Technology), Samuel B. Hopkins (Massachusetts Institute of Technology)
OptimizationExplainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: This study addresses the accuracy issue of data attribution in high dimensions and proposes the Rescaled Influence Function (RIF) as an improvement over the traditional Influence Function (IF);
ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning
Mingyang Chen (Baichuan Inc.), Weipeng Chen (Baichuan Inc.)
TransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Train large language models to actively use search engines during the reasoning process, forming a complete 'think-search-retrieve-think-answer' cycle;
ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains
Guillaume Vray (École Polytechnique Fédérale de Lausanne), Behzad Bozorgtabar (École Polytechnique Fédérale de Lausanne)
ClassificationSegmentationDomain AdaptationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A plugin framework named ReservoirTTA is proposed for long-term adaptation during testing, which detects and routes samples from different domains by maintaining a set of domain-specific models (model pool) and performing style-based online clustering, achieving domain adaptation and preventing catastrophic forgetting.
Residual Stream Analysis of Overfitting And Structural Disruptions
Quan Liu (Beijing University of Posts and Telecommunications), Sen Su (Beijing University of Posts and Telecommunications)
OptimizationSafty and PrivacySupervised Fine-TuningText
🎯 What it does: To address the issue of false refusals caused by safety fine-tuning, this paper proposes Variance Concentration Loss (VCL) through the analysis of the geometric structure of residual flows to suppress excessive variance concentration in mid-level residuals, thereby reducing the false refusal rate while maintaining or improving the model's general performance.
ReSim: Reliable World Simulation for Autonomous Driving
Jiazhi Yang (Chinese University of Hong Kong), Li Chen
Autonomous DrivingTransformerReinforcement LearningDiffusion modelVideo
🎯 What it does: ReSim is proposed, a controllable driving world model that can reliably predict future driving processes in different scenarios, including dangerous non-expert behaviors, and estimate rewards from the predicted videos through the Video2Reward module.
Resolution of Simpson's paradox via the common cause principle
A. Hovhannisyan, Armen Allahverdyan (Alikhanyan National Laboratory)
Tabular
🎯 What it does: The essence and solutions of Simpson's paradox were studied, proposing that by assuming a minimal common cause (or confounding) variable C, the paradox can be eliminated in the case of binary or minimal Gaussian continuous variables;
Resounding Acoustic Fields with Reciprocity
Zitong Lan (University of Pennsylvania), Mingmin Zhao (University of Pennsylvania)
Neural Radiance FieldContrastive LearningMeshAudio
🎯 What it does: A method for estimating room impulse responses under sparse source locations using the principle of acoustic reciprocity is proposed, called Resounding, which enhances the generalization ability of the sound field model through data augmentation and self-supervised learning.
Resource-Constrained Federated Continual Learning: What Does Matter?
Yichen Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Federated LearningComputational EfficiencyKnowledge DistillationImageBenchmark
🎯 What it does: This paper systematically evaluates and analyzes Federated Continual Learning (FCL) methods under resource-constrained conditions (memory buffering, computational budget, label rate), constructs large-scale benchmark experiments, and explores the robustness of four typical techniques.
RespoDiff: Dual-Module Bottleneck Transformation for Responsible & Faithful T2I Generation
Silpa Vadakkeeveetil Sreelatha (University of Surrey), Anjan Dutta (University of Surrey)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes RespoDiff, a responsible text-to-image generation framework that uses dual-module transformations in the bottleneck layer of diffusion models.
ResponseRank: Data-Efficient Reward Modeling through Preference Strength Learning
Timo Kaufmann (LMU Munich), Eyke Hüllermeier (LMU Munich)
Recommendation SystemReinforcement Learning from Human FeedbackReinforcement LearningTextSequential
🎯 What it does: A method called ResponseRank is proposed, which utilizes local relative intensity signals (such as response time, annotator consistency, etc.) to learn preference intensity and applies it to RLHF, language models, and control tasks;
Restage4D: Reanimating Deformable 3D Reconstruction from a Single Video
Jixuan He (Cornell Tech), Ming-Hsuan Yang (University of California)
GenerationData SynthesisOptimizationDiffusion modelGaussian SplattingVideo
🎯 What it does: Based on a single real video, new motion sequences are generated using a video generation model, and through video rewind joint training and occlusion-aware regularization, deformable 3D scenes are re-animated to maintain geometric consistency under new motions.
Restoring Pruned Large Language Models via Lost Component Compensation
Zijian Feng (Nanyang Technological University), Kezhi Mao (Nanyang Technological University)
RestorationTransformerLarge Language ModelText
🎯 What it does: A recovery method for pruned LLMs called RestoreLCC is proposed, which restores performance by compensating for lost components through comparative detection of key attention heads.
Restricted Global-Aware Graph Filters Bridging GNNs and Transformer for Node Classification
Jingyuan Zhang (Institute of Software, Chinese Academy of Sciences), Fengjun Zhang (Institute of Software, Chinese Academy of Sciences)
ClassificationGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes a global perception graph filtering framework G²Former based on Transformer, which utilizes multi-channel bandpass filters to constrain global attention, thereby stabilizing the attention mechanism and enhancing the generalization ability of Graph Neural Networks (GNNs);
Restricted Spectral Gap Decomposition for Simulated Tempering Targeting Mixture Distributions
Jhanvi Garg (Texas A&M University), Quan Zhou (Texas A&M University)
Tabular
🎯 What it does: This paper presents a decomposition theorem for the restricted spectral gap of discrete-time simulated tempering chains and uses this theorem to analyze the mixing time complexity of the Stochastic Walk Metropolis-Hastings (STMH) on Gaussian mixture distributions.
Rethinking Approximate Gaussian Inference in Classification
Bálint Mucsányi (Tübingen AI Center University of Tübingen), Philipp Hennig (Tübingen AI Center University of Tübingen)
ClassificationGaussian SplattingImage
🎯 What it does: A general framework is proposed that utilizes approximate Gaussian inference to obtain closed-form predictive distributions in the logit space, and achieves sampling-free predictions and second-order Dirichlet distributions through normCDF or sigmoid activation.
Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates
Hang Chen (Xi'an Jiaotong University), Wenya Wang (Nanyang Technological University)
TransformerLarge Language ModelText
🎯 What it does: Proposes three types of logic gates: AND, OR, and ADDER to explain the reliability and integrity of circuits, and constructs a circuit discovery framework that can directly separate the three types of gates based on noise and denoising interventions (Ns+Dn).
Rethinking Entropy in Test-Time Adaptation: The Missing Piece from Energy Duality
Mincheol Park (Samsung Electronics), Suhyun Kim (Kyung Hee University)
Domain AdaptationImage
🎯 What it does: ReTTA is proposed to enhance the robustness of the model under distribution shifts by simultaneously minimizing entropy and energy, while incorporating target class convergence and slice score matching.
Rethinking Fair Federated Learning from Parameter and Client View
Kaiqi Guan (Wuhan University), Mang Ye (Wuhan University)
Federated LearningImage
🎯 What it does: Designed and implemented FedPW, a federated learning framework that enhances fairness and overall performance through parameter pruning and adaptive weighted aggregation.
Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning
Feng Chen (Stanford University), Shaul Druckmann (Stanford University)
TransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper studies how to align the compute strategy (pass@N) during model training and testing, finding that traditional cross-entropy (CE) training can lead to overconfidence in the model, resulting in degraded performance when using pass@N for multiple sampling. To address this, we propose Direct Coverage Optimization (DCO) loss, which directly maximizes the probability of the correct answer appearing in N samples, and provide a stepwise DCO variant for proof search trees as well as an approximate implementation in Chain-of-Thought (CoT) reasoning.
Rethinking Gradient Step Denoiser: Towards Truly Pseudo-Contractive Operator
Shuchang Zhang (National University of Defense Technology), Hongxia Wang (National University of Defense Technology)
RestorationImage
🎯 What it does: This paper proposes a gradient step (GS) denoiser constructed using Input Convex Neural Networks (ICNN) and proves that it satisfies the property of true pseudo-convergence (d-SPC).
Rethinking Hebbian Principle: Low-Dimensional Structural Projection for Unsupervised Learning
Shikuang Deng (University of Electronic Science and Technology of China), Shi Gu (Zhejiang University)
OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A new unsupervised Hebbian learning framework SPHeRe is designed and implemented, utilizing structural projection and orthogonal constraints to learn robust features, and achieving hierarchical pre-training of deep networks through a lightweight auxiliary module.
Rethinking Joint Maximum Mean Discrepancy for Visual Domain Adaptation
Wei Wang (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao (Shenzhen Campus of Sun Yat-sen University)
Domain AdaptationImage
🎯 What it does: The paper derives a concise JMMD based on the representative theorem and combines it with HSIC to propose a new JMMD-HSIC loss for improving subspace learning in domain adaptation.
Rethinking Losses for Diffusion Bridge Samplers
Sebastian Sanokowski (Technical University Munich), Sebastian Lehner (Johannes Kepler University Linz)
Diffusion modelTabularTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper studies the loss function of the diffusion bridge sampler, proposing the use of the reverse Kullback-Leibler loss with the log-derivative trick (rKL-LD) as a replacement for the Log Variance loss, and proving its theoretical advantages and experimental performance;
Rethinking Multimodal Learning from the Perspective of Mitigating Classification Ability Disproportion
Qing-Yuan Jiang (Nanjing University of Science and Technology), Yang Yang (Nanjing University of Science and Technology)
ClassificationRecognitionConvolutional Neural NetworkVideoMultimodalityAudio
🎯 What it does: Proposes a sustained boosting algorithm and an adaptive classifier allocation strategy to dynamically enhance the classification capability of weak modalities, alleviating the modality imbalance problem in multimodal learning.
Rethinking Neural Combinatorial Optimization for Vehicle Routing Problems with Different Constraint Tightness Degrees
Fu Luo (Southern University of Science and Technology), Zhenkun Wang (National University of Singapore)
OptimizationTransformerReinforcement LearningMixture of ExpertsTabular
🎯 What it does: This study investigates the generalization ability of neural combinatorial optimization models in vehicle routing problems with varying constraint tightness and proposes variable constraint training and a multi-expert module to enhance performance.
Rethinking Nighttime Image Deraining via Learnable Color Space Transformation
Qiyuan Guan (Dalian Polytechnic University), Jinshan Pan (Nanjing University of Science and Technology)
RestorationTransformerImage
🎯 What it does: This paper proposes a high-quality nighttime rainy image de-raining dataset, HQ-NightRain, and designs a two-stage color space transformation network, CST-Net, based on a learnable color space converter (CSC) and implicit illumination guidance (IIG), for de-raining in the Y channel of the YCbCr color space.
Rethinking Optimal Verification Granularity for Compute-Efficient Test-Time Scaling
Hao Mark Chen, Hongxiang Fan (Imperial College London)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper explores the granularity of verification during LLM inference and proposes a tunable granularity search algorithm called VG-Search, which combines an adaptive granularity strategy to improve accuracy and reduce computational costs.
Rethinking Out-of-Distribution Detection and Generalization with Collective Behavior Dynamics
Zhenbin Wang (Sichuan University), Zizhou Wang (Sichuan University)
Domain AdaptationAnomaly DetectionImageTextPhysics RelatedOrdinary Differential Equation
🎯 What it does: Treating high-level features as charged particles, the Vlasov-Poisson system is used to model their collective behavior under a self-consistent electric field, and OOD detection and generalization are achieved by solving the boundary of the steady-state potential field and particle distribution.
Rethinking PCA Through Duality
Jan Quan (KU Leuven), Panagiotis Patrinos (KU Leuven)
OptimizationTabular
🎯 What it does: This paper re-examines Principal Component Analysis (PCA) through the framework of Differential Convex (DC) duality, proposing three new DC dual formulations. It proves that if one function of the original problem is unitary invariant, then the corresponding dual is kernelizable and possesses out-of-sample generalization capability. Furthermore, it reveals the equivalence of the DC Algorithm (DCA) under the variance maximization objective with the simultaneous iteration/QR algorithm, and provides the corresponding kernelized robust PCA dual and its algorithm.
Rethinking Residual Distribution in Locate-then-Edit Model Editing
Xiaopeng Li (National University of Defence Technology), Jie Yu (National University of Defence Technology)
TransformerLarge Language ModelText
🎯 What it does: The BLUE strategy is proposed, which enhances the effectiveness of the localization-reediting model editing method by directly calculating the residuals through updating only the first and last layers of the key layers, avoiding weight errors caused by residual distribution.
Rethinking Scale-Aware Temporal Encoding for Event-based Object Detection
Lin Zhu (Beijing Institute of Technology), Hua Huang (Beijing Normal University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkVideo
🎯 What it does: A CNN-RNN hybrid network is proposed for real-time object detection in event cameras, emphasizing fine-grained temporal modeling at low resolution stages and fusing features through multi-scale branches.
Rethinking the Role of Verbatim Memorization in LLM Privacy
Tom Sander (Meta), Chuan Guo (Meta)
Safty and PrivacyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: In this paper, the authors explore the leakage of private information in chat interactions by performing unsupervised fine-tuning and instruction fine-tuning on synthetic resumes, comparing it with traditional verbatim extraction.
Rethinking Tokenized Graph Transformers for Node Classification
Jinsong Chen (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
ClassificationRepresentation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: This paper proposes SwapGT, a node classification model based on Tokenized Graph Transformer, which primarily generates diverse token sequences through a novel token swapping operation and combines a Transformer backbone network with a center alignment loss for node representation learning.
Rethinking Verification for LLM Code Generation: From Generation to Testing
Zihan Ma (Xi'an Jiaotong University), Kai Chen (Shanghai AI Laboratory)
GenerationAI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Proposed and implemented SAGA (a human-LLM collaborative test case generation framework), and based on this, constructed the TCGBench and CodeCompass evaluation benchmarks;
Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization
Jiaqi Wei (Zhejiang University), Siqi Sun (Fudan University)
RetrievalOptimizationTransformerSupervised Fine-TuningContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Redefine RAG as Retrieval-Augmented Reasoning, proposing ALIGNRAG which iteratively corrects inconsistencies between reasoning and retrieval evidence during testing through Critique-Driven Alignment (CDA).
RetrievalAttention: Accelerating Long-Context LLM Inference via Vector Retrieval
Di Liu (Shanghai Jiao Tong University), Lili Qiu (Microsoft Research)
RetrievalComputational EfficiencyTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Construct RetrievalAttention, which pre-builds a vector index for fixed context KV cache and places it on the CPU. By using Attention-aware ANNS, it dynamically retrieves key KV during decoding, achieving acceleration of long-context LLM inference and saving GPU memory.
RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis
Robin Yadav (University of British Columbia), Renjie Liao (University of British Columbia)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: This study focuses on single-step retro-synthesis tasks and proposes a method for generating reactants through discrete flow matching.
Retro-R1: LLM-based Agentic Retrosynthesis
Wei Liu (Shanghai Jiao Tong University), Hao Zhou (Institute for AI Industry Research at Tsinghua University)
Drug DiscoveryTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Designed and trained an agent RETRO-R1 based on a large language model (LLM), which interacts with a single-step model through reinforcement learning in multi-step retro-synthesis planning to dynamically construct synthesis pathways.
Retrospective In-Context Learning for Temporal Credit Assignment with Large Language Models
Wentse Chen, Jeff Schneider (Carnegie Mellon University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningSequential
🎯 What it does: A review-based context learning (RICL) method is proposed, based on pre-trained large language models, for time credit assignment with sparse rewards, and an online learning framework RICOL is constructed to iteratively optimize the policy.
Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured MDPs
Mianchu Wang (University of Warwick), Giovanni Montana (University of Warwick)
OptimizationDrug DiscoveryGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A worst-path optimization framework based on tree-structured MDP is proposed, and a search-free multi-step inverse synthesis planning method called InterRetro is implemented.
Retrv-R1: A Reasoning-Driven MLLM Framework for Universal and Efficient Multimodal Retrieval
Lanyun Zhu (Zhejiang University), Shiqi Wang (City University of Hong Kong)
RetrievalRecommendation SystemTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodalityChain-of-Thought
🎯 What it does: We propose Retrv-R1, a multimodal large language model based on the R1 thinking framework, designed for general multimodal retrieval, which significantly improves retrieval accuracy through step-by-step reasoning.
Return of ChebNet: Understanding and Improving an Overlooked GNN on Long Range Tasks
Ali Hariri (École Polytechnique Fédérale de Lausanne), Pierre Vandergheynst (École Polytechnique Fédérale de Lausanne)
Graph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: Reassess and improve ChebNet, proposing Stable-ChebNet for stable propagation of long-range dependencies.
REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects
Yassine El Ouahidi (IMT Atlantique), Giulia Lioi (IMT Atlantique)
TransformerTime SeriesBiomedical Data
🎯 What it does: A foundational model called REVE has been constructed, which is adaptable to any EEG acquisition scheme, achieving new highs on 10 downstream tasks.
Revealing Multimodal Causality with Large Language Models
Jin Li (University of Technology Sydney), Fang Chen (University of Technology Sydney)
TransformerLarge Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: A framework named MLLM-CD is proposed for the automatic identification of causal variables and inference of causal structures from multimodal unstructured data.
Reverse Diffusion Sequential Monte Carlo Samplers
Luhuan Wu (Columbia University), John Patrick Cunningham
Diffusion modelScore-based ModelTabular
🎯 What it does: A training-free reverse diffusion sequence Monte Carlo (RDSMC) sampler is proposed for sampling from unnormalized target distributions.
Reverse Engineering Human Preferences with Reinforcement Learning
Lisa Alazraki (Imperial College London), Max Bartolo (Cohere)
Adversarial AttackLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Utilize reinforcement learning to generate prefixes for frozen LLMs to induce LLM-as-a-judge to give higher scores.
Reverse-Annealed Sequential Monte Carlo for Efficient Bayesian Optimal Experiment Design
Jake Callahan (University of Arizona), Tommie Catanach (Sandia National Laboratories)
OptimizationComputational EfficiencyReinforcement LearningTabular
🎯 What it does: A single-particle sequential Monte Carlo method based on reverse annealing is proposed for efficiently estimating the expected information gain in Bayesian optimal experimental design.
Revising and Falsifying Sparse Autoencoder Feature Explanations
George Ma (University of California), Somayeh Sojoudi (University of California)
Explainability and InterpretabilityLarge Language ModelAuto EncoderTextChain-of-Thought
🎯 What it does: This study investigates how to improve the automatic interpretation of features from Sparse Autoencoders (SAE), proposing more refined structured explanations and a tree-based iterative generation method.
Revisiting 1-peer exponential graph for enhancing decentralized learning efficiency
Kenta Niwa (NTT Communication Science Laboratories), W. Bastiaan Kleijn (Victoria University of Wellington)
Federated LearningComputational EfficiencyGraph Neural NetworkImage
🎯 What it does: This paper re-examines the 1-peer index graph, proposes the k-peer index graph and null-cascade graph, and constructs a dynamic communication topology that can achieve finite-time convergence under any number of nodes while maintaining the exchangeability of the mixing matrix, aimed at enhancing the efficiency of decentralized learning.
Revisiting Agnostic Boosting
Arthur da Cunha, Yuxin Sun
Supervised Fine-Tuning
🎯 What it does: A new boosting algorithm is proposed that elevates weak learners to strong learners under a general no-hypothesis framework, significantly reducing sample complexity;
Revisiting Bi-Linear State Transitions in Recurrent Neural Networks
MohammadReza Ebrahimi, Roland Memisevic (Qualcomm AI Research)
Recurrent Neural NetworkSequential
🎯 What it does: This paper studies the bilinear hidden state transition of RNNs, proving that it can effectively learn state tracking tasks, and experimentally validates its superiority.
Revisiting Consensus Error: A Fine-grained Analysis of Local SGD under Second-order Data Heterogeneity
Kumar Kshitij Patel (Institute for Foundations of Data Science Yale University), Lingxiao Wang (New Jersey Institute of Technology)
OptimizationFederated LearningTabular
🎯 What it does: A fine-grained convergence analysis is proposed for Local SGD (Federated Averaging) in distributed optimization, focusing on the impact of second-order data heterogeneity on communication complexity, and providing new upper and lower bounds.
Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology
Wenhao Tang (Nankai University), Ming-Ming Cheng (Nankai University)
ClassificationOptimizationConvolutional Neural NetworkBiomedical Data
🎯 What it does: This paper proposes an end-to-end slide-level supervised learning framework that combines multi-scale random sampling with a novel MIL model ABMILX, achieving end-to-end training using a ResNet encoder for computational pathology (CPath) tasks and significantly improving diagnostic and prognostic performance.
Revisiting Follow-the-Perturbed-Leader with Unbounded Perturbations in Bandit Problems
Jongyeong Lee (Korea Institute of Science and Technology), Min-hwan Oh (Seoul National University)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: In the multi-armed bandit problem, the perturbation distribution of the FTPL strategy is studied, and it is proposed and proven that BOBW performance can still be achieved when using asymmetric, unbounded Fréchet-type perturbations; numerical verification of the feasibility of symmetric Fréchet perturbations in the two-armed case is also provided.
Revisiting Frank-Wolfe for Structured Nonconvex Optimization
Hoomaan Maskan (Umeå University), Alp Yurtsever (Umeå University)
OptimizationTabular
🎯 What it does: A DC decomposition-based Frank-Wolfe (DC-FW) framework is proposed to solve structured non-convex optimization problems, featuring projection independence and the ability to solve gradient/linearized subproblems.
Revisiting Generative Infrared and Visible Image Fusion Based on Human Cognitive Laws
Lin Guo (Jiangnan University), Xiaoning Song (Jiangnan University)
SegmentationGenerationDiffusion modelAuto EncoderImage
🎯 What it does: A novel infrared-visible image fusion method called HCLFuse is proposed based on human cognitive principles, utilizing a multi-scale variational bottleneck encoder and a physics-guided diffusion model.
Revisiting Glorot Initialization for Long-Range Linear Recurrences
Noga Bar (Tel Aviv University), Raja Giryes (Tel Aviv University)
Recurrent Neural NetworkSequentialBenchmark
🎯 What it does: This paper studies the stability issues of using Glorot initialization in linear RNNs under long sequences and proposes and validates a dimension-based recalibration method to suppress the explosion of hidden states.
Revisiting Logit Distributions for Reliable Out-of-Distribution Detection
Jiachen Liang (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A post-processing OOD detection method based on logit gap (LogitGap) is proposed, which distinguishes ID and OOD samples by utilizing the difference between the maximum logit and the average of the other logits, and further enhances performance by automatically selecting top-N logits.
Revisiting LRP: Positional Attribution as the Missing Ingredient for Transformer Explainability
Yarden Bakish, Lior Wolf
Explainability and InterpretabilityTransformerImageText
🎯 What it does: A Layer-wise Relevance Propagation method considering position encoding (PA-LRP) is proposed for the interpretability of Transformers.
Revisiting Multi-Agent World Modeling from a Diffusion-Inspired Perspective
Yang Zhang (Tsinghua University), Chenjia Bai (China Telecom)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelScore-based ModelSequentialOrdinary Differential Equation
🎯 What it does: This paper proposes a multi-agent world model DIMA based on diffusion models, using sequential agent modeling to reduce complexity and improve the accuracy of global state transition predictions.
Revisiting Orbital Minimization Method for Neural Operator Decomposition
Jongha Jon Ryu (Massachusetts Institute of Technology), Gregory W. Wornell (Massachusetts Institute of Technology)
OptimizationReinforcement LearningContrastive LearningImageSequential
🎯 What it does: This paper reinterprets and promotes the traditional Orbit Minimization Method (OMM), constructing a neural network spectral decomposition framework that does not require orthogonalization constraints, and experimentally validates it on multiple tasks such as reinforcement learning, solving partial differential equations, and self-supervised contrastive learning.
Revisiting Residual Connections: Orthogonal Updates for Stable and Efficient Deep Networks
Giyeong Oh (Yonsei University), Youngjae Yu (Seoul National University)
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes an orthogonal residual update mechanism that only retains the part of the module output that is orthogonal to the input stream to update the network.
Revisiting Semi-Supervised Learning in the Era of Foundation Models
Ping Zhang (Ohio State University), Wei-Lun Chao (Ohio State University)
ClassificationObject DetectionTransformerSupervised Fine-TuningContrastive LearningImageBenchmark
🎯 What it does: In the era of Visual Foundation Models (VFM), systematic experiments on semi-supervised learning (SSL) are conducted, establishing a multi-task benchmark based on VTAB, and proposing a self-training framework V-PET achieved through efficient fine-tuning of multi-model parameters (PEFT) and the integration of VFM's pseudo-labels.
Revitalizing SVD for Global Covariance Pooling: Halley’s Method to Overcome Over-Flattening
Jiawei Gu (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
RecognitionOptimizationTransformerImage
🎯 What it does: A new global covariance pooling method called Halley-SVD is proposed for visual recognition tasks, utilizing higher-order iterations to achieve matrix square roots and addressing the excessive flattening issue of traditional iSQRT-COV.
Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models
Chenrui Cao (Microsoft Research), Fan Yang (Microsoft Research)
Large Language ModelReinforcement LearningText
🎯 What it does: An improved Draft-Sketch-Prove (DSP+) framework is proposed, achieving efficient theorem proving through fine-grained neural-symbolic collaboration.
Revolutionizing Graph Aggregation: From Suppression to Amplification via BoostGCN
Jiaxin Wu (Guangdong University of Technology), Jie Zhao (Guangdong University of Technology)
Recommendation SystemGraph Neural NetworkGraph
🎯 What it does: A novel linear GCN model called BoostGCN is proposed, which replaces the traditional Laplacian suppression aggregation by amplifying the important interactions of first-order neighbors to enhance recommendation effectiveness and training efficiency.
Revolutionizing Training-Free NAS: Towards Efficient Automatic Proxy Discovery via Large Language Models
Haidong Kang (Northeastern University), Hanling Wang
OptimizationNeural Architecture SearchTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageChain-of-Thought
🎯 What it does: This paper proposes a training-free unsupervised NAS framework called APD that utilizes large language models to automatically discover zero-cost proxies, significantly improving the search efficiency and accuracy of NAS.
Reward Reasoning Models
Jiaxin Guo, Furu Wei
TransformerReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes Reward Reasoning Models (RRM), transforming the reward model into a reasoning task that first performs chain-of-thought before providing rewards; it evolves reasoning capabilities through reinforcement learning and employs strategies such as ELO scoring or knockout tournaments in multi-response scenarios.
Reward-Aware Proto-Representations in Reinforcement Learning
Hon Tik Tse (University of Alberta), Marlos C. Machado (University of Alberta)
Reinforcement Learning
🎯 What it does: This paper proposes and deeply studies the prototype representation of reward perception—Default Representation (DR)—and compares it with the traditional reward-agnostic Inheritance Representation (SR);
Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation
Yihong Luo (Hong Kong University of Science and Technology), Jing Tang (Hong Kong University of Science and Technology)
GenerationReinforcement LearningDiffusion modelImage
🎯 What it does: Proposes the Reward-Instruct method, which directly transforms pre-trained diffusion models into few-step, reward-driven image generators without the need for diffusion distillation loss; and presents an improved version, Reward-Instruct+, which incorporates supervision at intermediate steps.
Reward-oriented Causal Representation Learning
Zirui Yan (Rensselaer Polytechnic Institute), Ali Tajer (Rensselaer Polytechnic Institute)
Representation LearningReinforcement LearningGraph
🎯 What it does: This paper proposes the framework of 'Reward-Oriented Causal Representation Learning (RO-CRL)' and designs a corresponding adaptive exploration algorithm to learn low-dimensional causal variables and graph structures from incomplete observational data through limited samples, ultimately optimizing downstream rewards within the available intervention space.
Rewind-to-Delete: Certified Machine Unlearning for Nonconvex Functions
Siqiao Mu (Northwestern University), Diego Klabjan (Northwestern University)
OptimizationTabularBiomedical DataElectronic Health RecordsOrdinary Differential Equation
🎯 What it does: A first-order black-box algorithm based on 'rewind-to-delete' is proposed to achieve machine unlearning for non-convex loss functions.
RF-Agent: Automated Reward Function Design via Language Agent Tree Search
Ning Gao (Beihang University), Yue Deng (Beihang University)
Robotic IntelligenceLarge Language ModelReinforcement LearningAgentic AI
🎯 What it does: This paper proposes the RF-Agent framework, which combines language models with Monte Carlo Tree Search to automatically generate reward functions for high-performance low-level control tasks.
RFMPose: Generative Category-level Object Pose Estimation via Riemannian Flow Matching
Wenzhe Ouyang (Zhejiang University), Jiming Chen (Zhejiang University)
Pose EstimationFlow-based ModelMultimodalityPoint Cloud
🎯 What it does: This paper proposes RFMPose, a category-level 6D object pose estimation framework based on Riemannian Flow Matching, which generates pose distributions on SE(3) using continuous probability flows.
RGB-Only Supervised Camera Parameter Optimization in Dynamic Scenes
Fang Li (University of Illinois), Narendra Ahuja (University of Illinois)
OptimizationVideo
🎯 What it does: For dynamic scenes in RGB videos, the ROS-Cam method is proposed, achieving optimization of camera parameters (focal length, rotation, translation) using only RGB supervision.
RGNMR: A Gauss-Newton method for robust matrix completion with theoretical guarantees
Eilon Vaknin Laufer (Weizmann Institute of Science), Boaz Nadler (Weizmann Institute of Science)
Anomaly DetectionOptimizationVideo
🎯 What it does: A new robust matrix completion method RGNMR is proposed, which achieves accurate recovery of low-rank matrices by combining Gauss-Newton iteration with outlier removal.
RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility
Haoyu He (Northeastern University), Qi Wang
TransformerLarge Language ModelPrompt EngineeringTime SeriesSequential
🎯 What it does: Proposed and implemented the RHYTHM framework, which combines a frozen pre-trained large language model (LLM) with hierarchical temporal tokenization and semantic prompts to predict human mobility trajectories.
RiboFlow: Conditional De Novo RNA Co-Design via Synergistic Flow Matching
Runze Ma (Shanghai Jiao Tong University), Shuangjia Zheng (Shanghai Jiao Tong University)
GenerationDrug DiscoveryFlow-based ModelSequentialBiomedical Data
🎯 What it does: The RiboFlow model is proposed, which can generate RNA molecules that satisfy structural and sequence consistency and have high binding affinity under the condition of given small molecule ligands.
Ridge Boosting is Both Robust and Efficient
David Bruns-Smith, Avi Feller
Domain AdaptationComputational EfficiencyTabular
🎯 What it does: This paper proposes a simple estimator called ridge boosting, which achieves both efficiency and distribution robustness by performing a boosting step on the initial predictor and combining it with kernel ridge regression.
RidgeLoRA: Matrix Ridge Enhanced Low-Rank Adaptation of Large Language Models
Junda Zhu (Beihang University), Qun Liu (Huawei Noah's Ark Lab)
TransformerLarge Language ModelTextMultimodality
🎯 What it does: In the parameter-efficient fine-tuning of large language models, the RidgeLoRA architecture is proposed, which improves the parallel structure of LoRA to a series connection and adds a diagonal ridge term to enhance representational capacity.
Riemannian Consistency Model
Chaoran Cheng (University of Illinois Urbana Champaign), Ge Liu
GenerationData SynthesisOptimizationDiffusion modelMultimodality
🎯 What it does: The Riemannian Consistency Model (RCM) is proposed, which can generate high-quality samples on Riemannian manifolds with non-zero curvature using only a small number of sampling steps.
Riemannian Flow Matching for Brain Connectivity Matrices via Pullback Geometry
Antoine Collas (Inria), Bertrand Thirion (Inria)
GenerationData SynthesisFlow-based ModelTime SeriesBiomedical DataMagnetic Resonance ImagingAlzheimer's DiseaseOrdinary Differential Equation
🎯 What it does: A method for matching Lagrangian flows based on global diffeomorphisms, called DIFFEOCFM, is proposed to generate realistic samples on brain connectivity matrices (SPD/correlation matrices).
Riemannian Proximal Sampler for High-accuracy Sampling on Manifolds
Yunrui Guan (Rice University), Shiqian Ma (Rice University)
Stochastic Differential Equation
🎯 What it does: A Riemannian Proximal Sampler is proposed to achieve high-precision sampling in Riemannian geometric spaces.
Rig3R: Rig-Aware Conditioning and Discovery for 3D Reconstruction
Samuel Li (Wayve Technologies), Matthew Brown (Wayve Technologies)
Pose EstimationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: Designed and trained Rig3R, a Transformer 3D reconstruction and pose estimation model that can utilize or learn constraints from multi-camera platforms.
RigAnyFace: Scaling Neural Facial Mesh Auto-Rigging with Unlabeled Data
Wenchao Ma (Roblox), Sharon X Huang
GenerationPose EstimationDiffusion modelOptical FlowImageMesh
🎯 What it does: A scalable facial mesh auto-binding framework called RigAnyFace (RAF) has been developed, which can automatically deform any topology (including multiple disconnected components) facial mesh from a neutral pose to a controllable FACS blendshape binding.
Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI
Luca Andolfi (University of Rome La Sapienza), Eleonora Giunchiglia (Imperial College London)
ClassificationAutonomous DrivingImageBenchmark
🎯 What it does: A Prototypical Neurosymbolic AI model is proposed, which embeds conceptual prototypes within neural networks to avoid reasoning shortcuts and reliably adhere to symbolic constraints in scenarios with very few labeled examples.
Right Question is Already Half the Answer: Fully Unsupervised LLM Reasoning Incentivization
Qingyang Zhang (Tianjin University), Yatao Bian (National University of Singapore)
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A completely unsupervised LLM inference incentive method called EMPO is proposed, which enhances the model's reasoning ability by minimizing semantic entropy.
RIGNO: A Graph-based Framework For Robust And Accurate Operator Learning For PDEs On Arbitrary Domains
Sepehr Mousavi (ETH Zurich), Siddhartha Mishra (ETH Zurich)
Graph Neural NetworkPoint CloudMeshBenchmarkPhysics Related
🎯 What it does: This study proposes a region interaction operator RIGNO based on graph neural networks for learning PDE solution operators on arbitrary geometric domains.
RiOSWorld: Benchmarking the Risk of Multimodal Computer-Use Agents
Yang JingYi, Jing Shao (Shanghai Artificial Intelligence Laboratory)
Anomaly DetectionSafty and PrivacyTransformerLarge Language ModelAgentic AIMultimodalityBenchmark
🎯 What it does: A risk assessment benchmark for multi-modal large model (MLLM) autonomous computer usage agents has been constructed—RiOSWorld, which includes 492 real interaction scenarios covering 13 sub-tasks such as web pages, social media, multimedia, operating systems, files, programming, emails, and office software. It evaluates the risk intentions and completion of agents from two dimensions: environmental risks and user-induced risks.
Rising from Ashes: Generalized Federated Learning via Dynamic Parameter Reset
Jiahao Wu (East China Normal University), Mingsong Chen (East China Normal University)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: The FedPhoenix framework is proposed, which disrupts client-specific overfitting features by randomly resetting some parameters of the global model during each communication round, thereby guiding the global model to learn more general features.
Risk Bounds For Distributional Regression
Carlos Misael Madrid Padilla (Washington University in St Louis), Sabyasachi Chatterjee (University of Illinois)
Tabular
🎯 What it does: A unified nonparametric distribution regression framework is proposed, providing upper bounds on the continuous rank probability score (CRPS) and mean squared error (MSE) risk under convex constraints (monotonicity, trend filtering) and non-convex constraints (deep networks), along with the corresponding convergence rates.
Risk-Averse Constrained Reinforcement Learning with Optimized Certainty Equivalents
Jane H. Lee (Yale University), Dionysis Kalogerias (Yale University)
OptimizationRobotic IntelligenceReinforcement LearningSequentialStochastic Differential Equation
🎯 What it does: This paper proposes a risk-aware constrained reinforcement learning framework using Optimized Certainty Equivalence (OCE), which can simultaneously consider tail risk in rewards and constraints.
Risk-Averse Total-Reward Reinforcement Learning
Xihong Su (University of New Hampshire), Marek Petrik (University of New Hampshire)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes two model-free Q-learning algorithms for risk aversion, targeting the exponential risk measure (ERM) and exponential value at risk (EVaR) objectives under the total-reward criterion, and provides rigorous convergence proofs; it also presents experimental results on two classic discrete domains (Cliff-Walking and Gambler's Ruin).
Risk-aware Direct Preference Optimization under Nested Risk Measure
Lijun Zhang (Shanxi University), Wei Wei (Shanxi University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies the introduction of a token-level direct preference optimization method Ra-DPO using nested risk measures during the alignment process of large language models, aiming to suppress the risk of the model deviating from the reference model while maintaining alignment performance.
RiverMamba: A State Space Model for Global River Discharge and Flood Forecasting
Mohamad Hakam Shams Eddin (University of Bonn), Juergen Gall (University of Bonn)
Time Series
🎯 What it does: The RiverMamba model is proposed for generating global river flow and flood medium-term (0-7 days) forecasts on a 0.05° grid;
RL Tango: Reinforcing Generator and Verifier Together for Language Reasoning
Kaiwen Zha (Massachusetts Institute of Technology), Dina Katabi (Massachusetts Institute of Technology)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Through the alternating training of the generator and validator in the RL framework TANGO, the multi-step reasoning ability of LLM is jointly enhanced, avoiding reliance on fixed or SFT-trained reward models;
RLGF: Reinforcement Learning with Geometric Feedback for Autonomous Driving Video Generation
Tianyi Yan (University of Macau), Jianbing Shen (University of Macau)
GenerationData SynthesisAutonomous DrivingReinforcement LearningDiffusion modelVideo
🎯 What it does: A framework is proposed to enhance autonomous driving video generation through Reinforcement Learning with Geometric Feedback (RLGF), addressing the issue of geometric distortion in synthetic videos.
RLVR-World: Training World Models with Reinforcement Learning
Jialong Wu (Tsinghua University), Mingsheng Long (Tsinghua University)
Robotic IntelligenceTransformerReinforcement LearningWorld ModelVideoTextMultimodality
🎯 What it does: Train a world model for language and video, using Reinforcement Learning with Verifiable Rewards (RLVR) for post-training, allowing the model to directly optimize task-related metrics.
RLZero: Direct Policy Inference from Language Without In-Domain Supervision
Harshit Sikchi (University of Texas at Austin), Scott Niekum (University of Massachusetts Amherst)
Robotic IntelligenceReinforcement LearningVision Language ModelVideoText
🎯 What it does: Proposes the RLZero framework, which utilizes unsupervised RL priors and video language models to directly infer zero-shot control policies from natural language instructions, without the need for training or domain labels during testing.