NeurIPS 2025 Papers — Page 38
Conference on Neural Information Processing Systems · 5275 papers
Reasoning Models Better Express Their Confidence
Dongkeun Yoon (Korea Advanced Institute of Science and Technology), Minjoon Seo (Korea Advanced Institute of Science and Technology)
Large Language ModelTextChain-of-Thought
🎯 What it does: This study explores how large language models can express confidence more accurately during chain-of-thought (CoT) reasoning and demonstrates that slow thinking behavior can significantly enhance confidence calibration.
Reasoning Models Hallucinate More: Factuality-Aware Reinforcement Learning for Large Reasoning Models
Junyi Li (National University of Singapore), Hwee Tou Ng (National University of Singapore)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies the hallucination problem that arises during the fine-tuning of large models in reinforcement learning due to the focus on the final answer, and proposes a reward mechanism that embeds factual verification at each step of reasoning—Factuality-aware Step-wise Policy Optimization (FSPO)—to reduce the occurrence of hallucinations and improve reasoning accuracy.
Reasoning Models Sometimes Output Illegible Chains of Thought
Arun Jose (Independent)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextPhysics RelatedChain-of-Thought
🎯 What it does: This paper evaluates the readability of chain-of-thought (CoT) texts generated by various RL-based reasoning models and finds that most models produce intermediate reasoning that is difficult to understand.
Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning
Jiwon Song (Seoul National University), Jae-Joon Kim (Seoul National University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A training-free Reasoning Path Compression (RPC) is proposed to accelerate inference by periodically compressing the KV cache.
Reasoning Planning for Language Models
Bao Nguyen (Chinese University of Hong Kong), Viet Anh Nguyen (Chinese University of Hong Kong)
OptimizationComputational EfficiencyLarge Language ModelContrastive LearningText
🎯 What it does: EPIC is proposed, an ensemble planning framework based on contrastive learning, designed to dynamically match the most suitable reasoning methods for language model queries, thereby reducing computational costs while maintaining accuracy.
Rebalancing Contrastive Alignment with Bottlenecked Semantic Increments in Text-Video Retrieval
Jian Xiao (Hefei University of Technology), Richang Hong (Hefei University of Technology)
RetrievalContrastive LearningVideoText
🎯 What it does: This paper proposes the GARE framework, which alleviates inter-modal tension and noisy negative samples in contrastive learning through a learnable incremental approach to text-video pairs.
Rebalancing Return Coverage for Conditional Sequence Modeling in Offline Reinforcement Learning
Wensong Bai (Zhejiang University), Hui Qian (Zhejiang University)
TransformerReinforcement LearningSequential
🎯 What it does: A return coverage rebalancing mechanism for offline reinforcement learning is proposed and integrated into the Decision Transformer, forming the RVDT algorithm.
ReCAP: Recursive Context-Aware Reasoning and Planning for Large Language Model Agents
Zhenyu Zhang (Stanford University), Jiaxin Pei (Stanford Institute for Human-Centered AI)
Robotic IntelligenceTransformerLarge Language ModelAgentic AIText
🎯 What it does: The ReCAP framework is proposed, utilizing recursive context-aware reasoning and planning to enhance the performance of LLMs in long-term tasks.
Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models
Ling Li (Hong Kong University of Science and Technology), Jiaheng Wei (Hong Kong University of Science and Technology)
RecognitionKnowledge DistillationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: The GLOBE system is proposed, utilizing a large visual-language model (LVLM) combined with multi-model distillation and reinforcement learning to geolocate social media images and generate interpretable reasoning paths.
ReCon-GS: Continuum-Preserved Guassian Streaming for Fast and Compact Reconstruction of Dynamic Scenes
Jiaye Fu (Peking University), Jian Zhang (Peking University)
CompressionOptimizationComputational EfficiencyGaussian SplattingVideo
🎯 What it does: The ReCon-GS framework is proposed, achieving fast and compact reconstruction and real-time rendering of online multi-view dynamic scenes.
ReCon: Region-Controllable Data Augmentation with Rectification and Alignment for Object Detection
Haowei Zhu (Tsinghua University), Bin Wang (Tsinghua University)
Object DetectionData SynthesisDiffusion modelImage
🎯 What it does: For the object detection task, a ReCon framework is proposed, which incorporates region-guided correction and region-aligned cross-attention into a frozen structure of controllable diffusion models (such as ControlNet + Stable Diffusion) in real-time, enhancing the consistency and quality of synthetic samples and annotations without training.
Reconciling Geospatial Prediction and Retrieval via Sparse Representations
YI LI, Gao Cong (Nanyang Technological University)
RetrievalRecommendation SystemOptimizationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Proposed the UrbanSparse framework, unifying geographic spatial prediction and retrieval tasks;
Reconstruct, Inpaint, Test-Time Finetune: Dynamic Novel-view Synthesis from Monocular Videos
Kaihua Chen (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
RestorationGenerationData SynthesisDiffusion modelSimultaneous Localization and MappingVideo
🎯 What it does: A three-stage pipeline called CogNVS is proposed, which reconstructs 3D scenes from monocular videos, renders target viewpoints, and fills in missing areas using a video diffusion model. It then performs self-supervised fine-tuning on test videos to achieve new viewpoint synthesis for dynamic scenes.
Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery
Shayan Shekarforoush (University of Toronto), David J. Fleet (University of Toronto)
SegmentationGenerationOptimizationGaussian SplattingBiomedical Data
🎯 What it does: CryoSPIRE is proposed, a hierarchical Gaussian mixture model (GMM) framework that achieves three-dimensional reconstruction of heterogeneous cryo-EM data by learning separable parts.
Reconstruction and Secrecy under Approximate Distance Queries
Shay Moran (Technion), Elizaveta Nesterova (Technion)
🎯 What it does: This paper studies the reconstruction game for locating unknown target points under approximate distance queries, analyzing the rate and limits of reconstruction error.
Rectified CFG++ for Flow Based Models
Shreshth Saini (University of Texas at Austin), Alan Bovik
GenerationData SynthesisFlow-based ModelRectified FlowImageTextOrdinary Differential Equation
🎯 What it does: A prediction-correction guiding method suitable for direct current models is proposed—Rectified‑CFG++. It predicts along the conditional flow at each step and then corrects using time-scheduled linear interpolation, thereby keeping the generated trajectory on the data manifold and reducing the drift from the manifold and visual artifacts caused by traditional CFG in direct current models.
Rectified Point Flow: Generic Point Cloud Pose Estimation
TAO SUN, Iro Armeni (Stanford University)
Pose EstimationTransformerFlow-based ModelRectified FlowPoint Cloud
🎯 What it does: A unified flow-based point cloud pose estimation framework is proposed, capable of simultaneously handling point cloud registration and multi-part assembly tasks.
Rectifying Shortcut Behaviors in Preference-based Reward Learning
Wenqian Ye (University of Virginia), Aidong Zhang (University of Virginia)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The PRISM method is proposed, which eliminates shortcut behaviors (such as verbosity, agreement, tone, etc.) in RLHF by learning a group-invariant kernel of the reward model, thereby enhancing the consistency of the reward model and downstream policies on OOD tasks.
Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation
Chenyang Jiang (Harbin Institute of Technology), Jingyong Su (Harbin Institute of Technology)
Knowledge DistillationImage
🎯 What it does: This paper studies the soft label bias problem in dataset distillation under long-tail distribution and proposes a lightweight post-processing module ADSP (Adaptive Soft-Label Alignment) to correct the soft labels.
Recurrent Attention-based Token Selection for Efficient Streaming Video-LLMs
Vaggelis Dorovatas (Toyota Motor Europe), Rahaf Aljundi (Toyota Motor Europe)
RetrievalCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: A training-free, LLM attention-based visual token selection and iterative processing framework is designed to achieve efficient question answering for long video streams.
Recurrent Memory for Online Interdomain Gaussian Processes
Wenlong Chen (Imperial College London), Yingzhen Li (Imperial College London)
Time SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes an online high-order polynomial projection (HiPPO) driven sparse variational Gaussian process model (OHSVGP) to maintain long-term memory in online learning scenarios.
Recurrent Self-Attention Dynamics: An Energy-Agnostic Perspective from Jacobians
Akiyoshi Tomihari (Artificial Intelligence Research Center AIST), Ryo Karakida (Artificial Intelligence Research Center AIST)
Recurrent Neural NetworkTransformerTabularSequential
🎯 What it does: Analyzed the energy-independent Jacobian dynamics of recurrent self-attention, explored the role of normalization layers in oscillation and stability, and proposed a training method based on spectral regularization;
Recursive Inference Scaling: A Winning Path to Scalable Inference in Language and Multimodal Systems
Ibrahim Alabdulmohsin (Google Deepmind), Xiaohua Zhai (Google Deepmind)
TransformerLarge Language ModelImageTextMultimodality
🎯 What it does: Proposed and experimented with Recursive INference Scaling (RINS), which enhances the reasoning performance of language and multimodal models by recursively executing the first half of the model multiple times;
Red-Teaming Text-to-Image Systems by Rule-based Preference Modeling
Yichuan Cao (Chinese Academy of Sciences), Yinpeng Dong (Tsinghua University)
GenerationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringImageText
🎯 What it does: Designed and implemented a red team framework RPG-RT based on rule preference modeling, utilizing large language models to iteratively modify prompts and fine-tune through rule scoring and Direct Preference Optimization (DPO) to break through the diverse defense mechanisms of closed-source text-to-image systems;
Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation
Zuhair Hasan Shaik (Mohamed bin Zayed University of Artificial Intelligence), Md Shad Akhtar (Indraprastha Institute of Information Technology Delhi)
ClassificationGenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This study investigates toxicity detection and generation in large language models, proposing a new interpretable intervention method called EigenShift, which precisely suppresses toxic outputs by utilizing the eigenvalue decomposition of the final linear layer.
ReDi: Rectified Discrete Flow
Jaehoon Yoo, Seunghoon Hong
GenerationKnowledge DistillationFlow-based ModelRectified FlowImageText
🎯 What it does: By iteratively correcting the coupling in the discrete flow model, the total correlation caused by decomposition approximations is reduced, enabling high-quality generation in just a few steps or even a single step.
ReDit: Reward Dithering for Improved LLM Policy Optimization
Chenxing Wei (Shenzhen University), Fei Yu
OptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: For the strategy optimization of LLMs, a method called Reward Dithering (ReDit) is proposed, which enhances gradient stability and convergence speed by injecting zero-mean noise into discrete rewards.
REDOUBT: Duo Safety Validation for Autonomous Vehicle Motion Planning
Shuguang Wang (City University of Hong Kong), Jianping Wang (City University of Hong Kong)
Autonomous DrivingSafty and PrivacyFlow-based ModelTabular
🎯 What it does: The REDOUBT framework is proposed for dual safety verification of autonomous driving motion planning, which detects whether the input scene is out of training distribution (OOV) and assesses the uncertainty of planning decisions.
Reducing the Probability of Undesirable Outputs in Language Models Using Probabilistic Inference
Stephen Zhao (University of Toronto and Vector Institute), Roger Baker Grosse
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Designed and evaluated a reinforcement learning alignment method called RePULSe, which actively samples low-reward outputs using a learned proposal distribution and reduces their probability, thereby decreasing the likelihood of the language model producing unacceptable outputs.
Reduction-based Pseudo-label Generation for Instance-dependent Partial Label Learning
Congyu Qiao (Southeast University), Xin Geng (Southeast University)
ClassificationMeta LearningImage
🎯 What it does: This paper proposes a method that utilizes 'reduction-based pseudo-labels' generated by a multi-branch auxiliary model to address the instance-dependent partial label learning (ID-PLL) problem, and trains the main classifier using these pseudo-labels.
Redundancy-Aware Test-Time Graph Out-of-Distribution Detection
Yue Hou (Beihang University), Ke Xu (Beihang University)
Anomaly DetectionGraph Neural NetworkContrastive LearningGraph
🎯 What it does: We propose an unsupervised framework for detecting graphical anomaly distributions at test time, called RedOUT, which can extract key structural information from test graphs without modifying the parameters of the pre-trained model, thereby better distinguishing between ID and OOD graphs.
Refinement Methods for Distributed Distribution Estimation under $\ell^p$-Losses
Deheng Yuan (Tsinghua University), Zhongyi Huang (Tsinghua University)
Optimization
🎯 What it does: The study addresses the problem of discrete distribution estimation using α_p loss in a distributed environment with communication constraints, proposing an interactive estimation protocol and providing almost optimal upper and lower error bounds for each terminal with multiple samples.
Refining Norms: A Post-hoc Framework for OOD Detection in Graph Neural Networks
Jiawei Gu (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: This paper proposes RAGNOR, a post-hoc method that utilizes the norm of graph neural network (GNN) embedding vectors to achieve node-level and graph-level out-of-distribution (OOD) detection.
RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models
Yilang Zhang (University of Minnesota), Georgios B. Giannakis (University of Minnesota)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningImageTextBenchmark
🎯 What it does: This paper proposes an improved low-rank adaptation method called RefLoRA, which enhances the fine-tuning efficiency and stability of large models by optimally reconstructing low-rank factors at each step.
Reframing Gaussian Splatting Densification with Complexity-Density Consistency of Primitives
Zhemeng Dong (Harbin Institute of Technology), Xianming Liu (Harbin Institute of Technology)
RestorationData SynthesisGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes CDC-GS, a method for allocating 3D Gaussian atoms using visual complexity and geometric density consistency, improving the reconstruction quality of 3DGS.
Refusal Direction is Universal Across Safety-Aligned Languages
Xinpeng Wang (Ludwig Maximilian University of Munich), Barbara Plank (Ludwig Maximilian University of Munich)
Safty and PrivacyTransformerLarge Language ModelTextMultimodality
🎯 What it does: This study investigates the generalizability of refusal mechanisms in large language models within multilingual environments, proposing and validating the transferability of the Refusal Direction across different languages.
REGen: Multimodal Retrieval-Embedded Generation for Long-to-Short Video Editing
Weihan Xu (Duke University), Hao-Wen Dong (University of Michigan)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: A hybrid framework based on large language models and retrieval (REGen) has been developed, capable of generating short videos that contain coherent narratives and embedded short video clips from long videos.
Regional Explanations: Bridging Local and Global Variable Importance
Salim I. Amoukou (J.P. Morgan), Nicolas J-B. Brunel (Capgemini Invent)
Explainability and InterpretabilitySupervised Fine-TuningTabular
🎯 What it does: This paper studies the limitations of commonly used local attribution methods (Local Shapley Values and LIME) and proposes a new regional explanation method called R-LOCO.
Register and [CLS] tokens induce a decoupling of local and global features in large ViTs
Alexander Lappe (Hertie Institute), Martin A. Giese (Hertie Institute)
Explainability and InterpretabilityRepresentation LearningTransformerImage
🎯 What it does: This study investigates the local-global feature decoupling problem caused by the register and CLS token in Vision Transformers.
Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection
Yuyang Yu (South China University of Technology), Shengfeng He (Singapore Management University)
Anomaly DetectionPoint CloudBenchmark
🎯 What it does: The Reg2Inv framework is proposed, which jointly learns the point cloud registration process and anomaly detection tasks to obtain rotation-invariant and locally discriminative features; during the inference phase, the registration matrix is used to align the test point cloud with the prototype, and then feature comparison is performed to achieve object-level and point-level anomaly detection.
Regression Trees Know Calculus
Nathan Wycoff (University of Massachusetts)
OptimizationExplainability and InterpretabilityTabularBiomedical Data
🎯 What it does: This paper proposes a gradient estimation method based on regression trees and uses this estimation to compute integral gradients and active subspaces.
Regression-adjusted Monte Carlo Estimators for Shapley Values and Probabilistic Values
R. Teal Witter (Claremont McKenna College), Christopher Musco (New York University)
Tabular
🎯 What it does: A new method called 'Regression MSR' is proposed, which combines Monte Carlo and regression techniques to unbiasedly estimate any probability values (including Shapley, Banzhaf, etc.) and achieve maximum sample reuse.
Regret Analysis of Average-Reward Unichain MDPs via an Actor-Critic Approach
Swetha Ganesh (Purdue University), Vaneet Aggarwal (Purdue University)
Reinforcement Learning
🎯 What it does: A Natural Actor-Critic algorithm (NAC-B) is proposed, achieving an order-optimal logarithmic order harmonic return cumulative regret upper bound in infinite average reward Markov Decision Processes (MDP) through batching.
Regret Bounds for Adversarial Contextual Bandits with General Function Approximation and Delayed Feedback
Orin Levy (Tel Aviv University), Yishay Mansour (Tel Aviv University)
OptimizationReinforcement Learning from Human Feedback
🎯 What it does: An optimal/nearly optimal regret algorithm for adversarial contextual multi-armed bandits (CMAB) with arbitrary delayed feedback is proposed, covering two major scenarios: policy-based learning and general function approximation.
Regret Lower Bounds for Decentralized Multi-Agent Stochastic Shortest Path Problems
Utkarsh U. Chavan (Indian Institute of Technology Bombay), Nandyala Hemachandra (Indian Institute of Technology Bombay)
Reinforcement Learning
🎯 What it does: In the decentralized multi-agent stochastic shortest path (Dec-MASSP) problem, a class of hard-to-learn linear function approximation instances is constructed, and it is proven that any decentralized learning algorithm has an expected cumulative regret of at least Ω(√K) after K rounds of experiments, thus providing a lower bound for the problem.
Regret-Optimal Q-Learning with Low Cost for Single-Agent and Federated Reinforcement Learning
Haochen Zhang (Pennsylvania State University), Lingzhou Xue (Pennsylvania State University)
Federated LearningReinforcement Learning
🎯 What it does: This paper studies the online reinforcement learning problem in single-agent and federated reinforcement learning, proposing two novel model-free reinforcement learning algorithms aimed at simultaneously achieving near-optimal regret, low burn-in costs, and logarithmic levels of switching costs for single agents or communication costs for federated learning.
Regularized least squares learning with heavy-tailed noise is minimax optimal
Mattes Mollenhauer (Merantix Momentum), Arthur Gretton (University College London)
🎯 What it does: This paper studies the learning performance of kernel ridge regression (regularized least squares) in RKHS under heavy-tailed noise with finite higher-order moments;
ReID5o: Achieving Omni Multi-modal Person Re-identification in a Single Model
Jialong Zuo (Huazhong University of Science and Technology), Changxin Gao (Huazhong University of Science and Technology)
RecognitionRetrievalTransformerMixture of ExpertsImageTextMultimodality
🎯 What it does: Proposes the OM-ReID problem, constructs a high-quality dataset ORBench across five modalities (RGB, infrared, colored pencil, sketch, text), and designs a unified model ReID5o that can handle any combination of modalities;
ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning
Tonghe Zhang (Carnegie Mellon University), Yu Wang (Tsinghua University)
Robotic IntelligenceReinforcement LearningFlow-based ModelRectified FlowMultimodality
🎯 What it does: The ReinFlow framework is proposed, which utilizes online reinforcement learning to fine-tune pre-trained flow matching strategies to enhance the continuous control performance of robots.
REINFORCE Converges to Optimal Policies with Any Learning Rate
Samuel McLaughlin Robertson (University of Alberta), Jincheng Mei (Google DeepMind)
Reinforcement LearningTabular
🎯 What it does: Proves that the classic REINFORCE (softmax parameterization) can globally converge to the optimal policy using any constant learning rate in finite-horizon Markov Decision Processes (MDPs).
Reinforced Active Learning for Large-Scale Virtual Screening with Learnable Policy Model
Yicong Chen (Sun Yat-sen University), Yuedong Yang (Sun Yat-sen University)
Drug DiscoveryGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: The GLARE framework is proposed, which transforms the large-scale virtual screening (VS) problem into a Markov Decision Process (MDP), adaptively selecting candidate molecules for experimental labeling through reinforcement learning (RL).
Reinforced Context Order Recovery for Adaptive Reasoning and Planning
Long Ma (Peking University), Yizhou Wang (Peking University)
GenerationTransformerReinforcement LearningText
🎯 What it does: This paper proposes a reinforcement learning framework named ReCOR, which is used to self-supervise the recovery of adaptive generation order from pure text data to enhance the generation quality in reasoning and planning tasks.
Reinforcement Learning Finetunes Small Subnetworks in Large Language Models
Sagnik Mukherjee (University of Illinois Urbana-Champaign), Hao Peng (University of Illinois Urbana-Champaign)
Large Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper studies the fine-tuning process of reinforcement learning in large language models, finding that RL only updates a very small proportion (5%-30%) of the parameter subnetwork, and the fine-tuning of this subnetwork can restore the performance and parameters of the fully fine-tuned model. It further analyzes the consistency of the subnetwork and the factors leading to sparse updates.
Reinforcement learning for one-shot DAG scheduling with comparability identification and dense reward
Xumai Qi (Tongji University), Hongcheng Wang (Tongji University)
OptimizationGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: A one-shot DAG scheduling method based on policy gradient is proposed, combining comparable anti-chain recognition and dense rewards to improve node priority learning and training stability.
Reinforcement Learning for Out-of-Distribution Reasoning in LLMs: An Empirical Study on Diagnosis-Related Group Coding
Hanyin Wang (Mayo Clinic), Jimeng Sun (University of Illinois Urbana-Champaign)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes and trains the DRG-SAPPHIRE model for the medical coding task of DRG coding, utilizing large-scale reinforcement learning to achieve automated coding and generate interpretable clinical reasoning.
Reinforcement Learning for Reasoning in Large Language Models with One Training Example
Yiping Wang (University of Washington), yelong shen
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: In the paper, the authors propose the '1-shot RLVR' method, which uses only one training sample to stimulate the mathematical reasoning ability of large language models through reinforcement learning, significantly improving performance on multiple mathematical benchmarks.
Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation
Yifu Luo (Tsinghua University), Xueqian Wang (Tsinghua University)
GenerationData SynthesisReinforcement LearningImageText
🎯 What it does: This paper proposes the Mask-GRPO method, which applies reinforcement learning (GRPO) to masked generative models (MGMs) for the first time, aiming to enhance the generation performance of text-to-image (T2I) tasks.
Reinforcement Learning Teachers of Test Time Scaling
Edoardo Cetin (Sakana AI), Yujin Tang (Sakana AI)
Knowledge DistillationLarge Language ModelReinforcement LearningText
🎯 What it does: A new Reinforcement Learning Teacher (RLT) framework is proposed, which trains language models to provide explanations for known answers, used for student distillation and RL cold start.
Reinforcement Learning with Action Chunking
Qiyang Li (University of California Berkeley), Sergey Levine (University of California Berkeley)
Reinforcement Learning
🎯 What it does: Proposes Q-Chunking, which achieves more efficient value backpropagation and temporally coherent exploration in offline to online reinforcement learning through action chunking.
Reinforcement Learning with Backtracking Feedback
Bilgehan Sel (Google), Dingcheng Li (Google)
Large Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A reinforcement learning framework called RLBF is proposed, which trains large language models to dynamically self-correct during the generation process by detecting safety violations and issuing a 'rollback x tokens' instruction.
Reinforcement Learning with Imperfect Transition Predictions: A Bellman-Jensen Approach
Chenbei Lu (Tsinghua University), Adam Wierman (California Institute of Technology)
OptimizationReinforcement LearningTabularTime Series
🎯 What it does: This paper proposes a framework for utilizing multi-step predictions in reinforcement learning, designing a Bayesian value function and the BOLA algorithm to effectively leverage imperfect multi-step transition predictions.
Reinforcing Spatial Reasoning in Vision-Language Models with Interwoven Thinking and Visual Drawing
Junfei Wu (Institute of Automation, Chinese Academy of Sciences), Tieniu Tan (Institute of Automation, Chinese Academy of Sciences)
TransformerReinforcement LearningVision Language ModelImage
🎯 What it does: This paper proposes a reasoning paradigm for drawing operations in visual space, enabling visual-language models to perform spatial reasoning directly on images through drawing boxes and auxiliary lines.
Reinforcing the Diffusion Chain of Lateral Thought with Diffusion Language Models
Zemin Huang (Westlake University), Guo-Jun Qi (Westlake University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: A new reasoning framework called Diffusion Chain of Lateral Thought (DCoLT) is proposed, which treats the intermediate steps in the reverse diffusion process of diffusion language models (DLM) as 'thinking' actions and employs reinforcement learning based on the correctness of the final answer as a reward.
Reinventing Multi-Agent Collaboration through Gaussian-Image Synergy in Diffusion Policies
Ziye Wang (Sun Yat-sen University), Ruimao Zhang (Sun Yat-sen University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelGaussian SplattingImagePoint CloudBenchmark
🎯 What it does: Constructing a globally consistent 3D Gaussian field in multi-agent collaboration, allowing each agent to query key features based on its local perspective to achieve scalable, perception-driven imitation learning.
RelationAdapter: Learning and Transferring Visual Relation with Diffusion Transformers
Yan Gong (Zhejiang University), Yin Zhang (Zhejiang University)
Image TranslationGenerationTransformerDiffusion modelImage
🎯 What it does: A visual prompt-based image editing framework called RelationAdapter is proposed, which can efficiently extract and transfer editing intentions on the Diffusion Transformer;
Relaxing partition admissibility in Cluster-DAGs: a causal calculus with arbitrary variable clustering
Clément Yvernes (University of Grenoble Alpes), Eric Gaussier (University of Grenoble Alpes)
🎯 What it does: This paper proposes a Cluster-DAG (C-DAG) framework that can handle arbitrary clustering and allows cycles, and provides a provably atomic complete causal inference calculus within this framework.
ReliabilityRAG: Effective and Provably Robust Defense for RAG-based Web-Search
Zeyu Shen (Princeton University), Aleksandra Korolova (Princeton University)
RetrievalAdversarial AttackTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the ReliabilityRAG framework, which utilizes the reliability information of retrieved documents (such as ranking or weight) to enhance the robustness of RAG systems against retrieval corpus contamination, implementing defenses through the maximum independent set (MIS) from graph theory and weighted sampling aggregation methods.
Reliable Decision‑Making via Calibration‑Oriented Retrieval‑Augmented Generation
Chaeyun Jang (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A Calibrated Retrieval-Augmented Generation (CalibRAG) framework is proposed to achieve confidence calibration in the decision-making process based on retrieval-augmented generation.
Reliable Lifelong Multimodal Editing: Conflict-Aware Retrieval Meets Multi-Level Guidance
Qiang Zhang (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A retrieval-augmented framework CARML for lifelong knowledge editing in multimodal large language models is proposed.
Reliably detecting model failures in deployment without labels
Viet Nguyen (University of Toronto), Rahul Krishnan
Domain AdaptationAnomaly DetectionImageTabularBiomedical DataElectronic Health Records
🎯 What it does: A post-deployment model performance degradation detection method called D3M is proposed, which requires no labels and no training data after training, and can automatically identify model failure during distribution drift.
Relieving the Over-Aggregating Effect in Graph Transformers
Junshu Sun (Institute of Computing Technology), Shuhui Wang (Institute of Computing Technology)
Graph Neural NetworkTransformerGraph
🎯 What it does: The Wideformer method is proposed to address the over-aggregation problem in graph Transformers, significantly improving the model's efficiency in utilizing global information.
ReMA: Learning to Meta-Think for LLMs with Multi-agent Reinforcement Learning
Ziyu Wan (Shanghai Jiao Tong University), Ying Wen (Shanghai Jiao Tong University)
Meta LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposed and implemented the ReMA framework, which utilizes multi-agent reinforcement learning to decompose the reasoning process of large language models into two stages: meta-thinking and fine-grained reasoning, and further enhances reasoning capabilities through multi-round interactions.
Remarkable Robustness of LLMs: Stages of Inference?
Vedang Lad (Massachusetts Institute of Technology), Max Tegmark (Massachusetts Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the robustness of large language models (LLMs) during the reasoning process through structural interventions of layer deletion and swapping, and proposes a four-stage reasoning hypothesis (detokenization, feature engineering, prediction ensembling, residual calibration).
Remasking Discrete Diffusion Models with Inference-Time Scaling
Guanghan Wang (Cornell University), Volodymyr Kuleshov (Cornell University)
GenerationData SynthesisDrug DiscoveryDiffusion modelImageText
🎯 What it does: The Remasking Diffusion Model (ReMDM) sampler is proposed, enabling discrete mask diffusion models to remask and correct errors during the generation process, achieving iterative refinement.
REMI: Reconstructing Episodic Memory During Internally Driven Path Planning
Zhaoze Wang (University of Pennsylvania), Vijay Balasubramanian (University of Pennsylvania)
Recurrent Neural NetworkImage
🎯 What it does: A framework for internal path planning based on MEC-Hippocampal networks, called REMI, is proposed, which can recall target locations through sensory cues and generate planned paths.
ReMindRAG: Low-Cost LLM-Guided Knowledge Graph Traversal for Efficient RAG
Yikuan Hu (Sichuan University), Chen Huang (National University of Singapore)
RetrievalComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: This paper proposes REMINDRAG, a KG-RAG system that utilizes LLM-guided knowledge graph traversal, combining node exploration, node utilization, and memory replay.
Removing Concepts from Text-to-Image Models with Only Negative Samples
Hanwen Liu (Peking University), Yadong MU
GenerationData SynthesisOptimizationDiffusion modelContrastive LearningImage
🎯 What it does: A method called Clipout is proposed, which achieves machine forgetting of target concepts by randomly cropping concept embeddings in a pre-trained text-image model and using negative sample contrastive loss, allowing the model to delete private, copyrighted, or harmful concepts without retraining.
REN: Fast and Efficient Region Encodings from Patch-Based Image Encoders
Savya Khosla (University of Illinois Urbana-Champaign), Derek Hoiem (University of Illinois Urbana-Champaign)
Object DetectionSegmentationRetrievalTransformerContrastive LearningImage
🎯 What it does: Proposes the Region Encoder Network (REN), which directly extracts high-quality region vectors from a frozen patch encoder using point prompts, eliminating the expensive segmentation step.
Rendering-Aware Reinforcement Learning for Vector Graphics Generation
Juan A. Rodriguez (ServiceNow Research), Marco Pedersoli (ServiceNow Research)
GenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningVision Language ModelImageText
🎯 What it does: A reinforcement learning framework based on rendering feedback, RLRF, is proposed to enhance the accuracy and code efficiency of visual language models in the SVG generation task.
REOrdering Patches Improves Vision Models
Declan Kutscher (University of Pittsburgh), Ritwik Gupta (University of California, Berkeley)
ClassificationOptimizationTransformerReinforcement LearningImage
🎯 What it does: This paper studies the sensitivity of visual long-sequence models to the order of image patches and proposes a framework called REOrder that learns the optimal patch order through information-theoretic priors and reinforcement learning.
REP: Resource-Efficient Prompting for Rehearsal-Free Continual Learning
Sungho Jeon (POSTECH), Myeongjae Jeon (POSTECH)
SegmentationComputational EfficiencyTransformerPrompt EngineeringImage
🎯 What it does: A resource-efficient prompt (REP) framework is proposed in the context of continuous learning without replay, aimed at reducing the computational and memory overhead of visual Transformers.
REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training
Ziqiao Wang (National University of Singapore), Yang You (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: A training acceleration framework named HASTE is proposed, which first integrates global features and attention alignment in the Diffusion Transformer (DiT/SiT), and then cuts off the alignment loss at preset iteration points, allowing the model to freely complete denoising learning.
Reparameterized LLM Training via Orthogonal Equivalence Transformation
Zeju Qiu (Max Planck Institute for Intelligent Systems), Weiyang Liu (Chinese University of Hong Kong)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The POET method is proposed, which reparameterizes large model training through orthogonal equivalence transformations, maintaining the weight spectrum unchanged, thereby achieving more stable and better generalization in LLM training;
RepGuard: Adaptive Feature Decoupling for Robust Backdoor Defense in Large Language Models
Chenxu Niu (Institute of Information Engineering Chinese Academy of Sciences), Yue Hu (Institute of Information Engineering Chinese Academy of Sciences)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: An adaptive feature decoupling defense framework named RepGuard has been designed and implemented, which utilizes dual-perspective analysis and mask generation of LLM hidden layer representations to eliminate the anomalous features relied upon by backdoor triggers, thereby reducing the success rate of backdoor attacks.
RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models
Yeongtak Oh (Seoul National University), Sungroh Yoon (Seoul National University)
RecognitionGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: A post-training framework based on reinforcement learning, RePIC, is designed to enhance the visual recognition and personalized generation capabilities of multimodal large language models in personalized image captioning.
ReplaceMe: Network Simplification via Depth Pruning and Transformer Block Linearization
Dmitriy Shopkhoev (ITMO University), Sergey Zagoruyko (Polynome)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: A training-independent deep pruning method called ReplaceMe is proposed, which compresses the model by replacing consecutive Transformer blocks with linear transformations.
RepLDM: Reprogramming Pretrained Latent Diffusion Models for High-Quality, High-Efficiency, High-Resolution Image Generation
Boyuan Cao (Fudan University), Hongming Shan (Fudan University)
GenerationData SynthesisComputational EfficiencyDiffusion modelImage
🎯 What it does: Reprogramming existing latent diffusion models (LDM) without training to achieve high-quality, high-efficiency high-resolution image generation.
Replicable Distribution Testing
Ilias Diakonikolas (University of Wisconsin Madison), Christopher Ye (University of California San Diego)
🎯 What it does: This paper systematically studies distribution testing under the framework of algorithmic reproducibility, particularly focusing on proximity and independence testing for discrete distributions, proposing new reproducible algorithms and establishing lower bounds on sample complexity.
Replicable Online Learning
Saba Ahmadi (Toyota Technological Institute at Chicago), Avrim Blum (Toyota Technological Institute at Chicago)
Optimization
🎯 What it does: This paper proposes a replicable algorithm framework for online learning, primarily targeting online linear optimization and expert problems. It designs algorithms that maintain the same decision sequence under any time-varying distribution while achieving low regret.
Replicable Online pricing
Kiarash Banihashem (University of Maryland), MohammadTaghi Hajiaghayi (University of Maryland)
Optimization
🎯 What it does: This paper proposes replicable online pricing (ROP) and delegation algorithms that can output approximately optimal thresholds under limited samples while satisfying replicability constraints.
RePO: Understanding Preference Learning Through ReLU-Based Optimization
Junkang Wu (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: A ReLU-based preference learning method called RePO is proposed, which uses a threshold γ to filter training samples, thereby achieving alignment with large language models without using Sigmoid weights or regularization.
Repo2Run: Automated Building Executable Environment for Code Repository at Scale
Ruida Hu (Harbin Institute of Technology), Cuiyun Gao (Harbin Institute of Technology)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Proposes Repo2Run, an LLM-based agent for automating the construction of executable code environments, capable of generating runnable Dockerfiles for any Python repository and successfully executing unit tests.
RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Task Solving
Huacan Wang (Midea Group), Pin Lyu (Institute of Automation, Chinese Academy of Sciences)
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AITextGraphBenchmark
🎯 What it does: The RepoMaster framework is proposed, which automatically searches, analyzes, and extracts core components of GitHub repositories, and autonomously explores and executes tasks with the assistance of LLMs, thereby completing complex real-world tasks.
Representation Consistency for Accurate and Coherent LLM Answer Aggregation
Junqi Jiang (Imperial College London), Francesca Toni (Imperial College London)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: To improve answer aggregation for LLM responses generated from multiple prompts and samples, we propose using the consistency of internal model activations to enhance accuracy in reasoning tasks.
Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think
Ge Wu (JIIOV Technology), Xiang Li (Nankai University)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: Proposes the REG method, which synchronously injects noise and concatenates low-level image latent variables with high-level category tokens in diffusion training, achieving joint denoising of images and semantics;
Representation-Level Counterfactual Calibration for Debiased Zero-Shot Recognition
Pei Peng (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
RecognitionRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: This paper estimates the expectations of objects and backgrounds in the representation space through causal inference within the CLIP framework, and synthesizes adversarial embeddings during inference to correct the object-background shortcut bias in zero-shot recognition; simultaneously, it utilizes Total Direct Effect (TDE) to eliminate background false activations.
Representational Difference Explanations
Neehar Kondapaneni (California Institute of Technology), Pietro Perona (California Institute of Technology)
Explainability and InterpretabilityRepresentation LearningImage
🎯 What it does: A method called Representational Difference Explanations (RDX) is proposed, which is a non-training approach for comparing and visualizing the differences between the embedding representations of two models, helping to directly explain model differences.
Reproducing Kernel Banach Space Models for Neural Networks with Application to Rademacher Complexity Analysis
Alistair Shilton (Deakin University), Svetha Venkatesh (Deakin University)
Transformer
🎯 What it does: An accurate Reproducing Kernel Banach Space (RKBS) model is proposed to fully describe any feedforward neural network (including ResNet and Transformer), and under this model, a width-independent and depth-exponential bound for Rademacher complexity is derived.
Repurposing AlphaFold3-like Protein Folding Models for Antibody Sequence and Structure Co-design
Nianzu Yang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Protein Structure PredictionDiffusion modelBiomedical Data
🎯 What it does: Transform the AlphaFold3-like protein folding model into an antibody sequence-structure co-design model, incorporating a sequence diffusion head to achieve co-diffusion of CDR sequences and structures;
Repurposing Marigold for Zero-Shot Metric Depth Estimation via Defocus Blur Cues
Chinmay Talegaonkar (University of California San Diego), Nicholas Antipa (University of California San Diego)
Depth EstimationDiffusion modelImage
🎯 What it does: In monocular scene settings, panoramic focused images taken with different apertures and defocused images are used to infer optimized pre-trained relative depth generated by Marigold, combined with a physical degradation model to obtain zero-shot, measurable depth estimation.