ICLR 2026 Papers — Page 39
International Conference on Learning Representations · 5356 papers
RefineBench: Evaluating Refinement Capability of Language Models via Checklists
Young-Jun Lee (KAIST), Ho-Jin Choi (KAIST)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes REFINEBENCH, the first multi-round evaluation benchmark for assessing the self-improvement and guided improvement capabilities of language models, covering 1,000 questions across 11 domains, and employing a checklist-based evaluation framework;
RefineStat: Efficient Exploration for Probabilistic Program Synthesis
Madhav Kanda (University of Illinois Urbana-Champaign), Sasa Misailovic (University of Illinois Urbana-Champaign)
Computational EfficiencyAI Code AssistantLarge Language ModelTabular
🎯 What it does: Propose the REFINESTAT framework, which automatically generates probabilistic programs that satisfy Bayesian workflow reliability metrics using small language models within a semantic constraint and diagnostic-driven iterative process.
Refining Hybrid Genetic Search for CVRP via Reinforcement Learning-Finetuned LLM
Rongjie Zhu (Nanjing University of Information Science and Technology), Zhiguang Cao (Singapore Management University)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringGraphBenchmark
🎯 What it does: Proposes RFTHGS, a reinforcement learning-based framework to fine-tune a small LLM (14B parameters) for generating high-performance crossover (and subpopulation) operators, thereby enhancing the solving capacity of HGS for the Capacitated Vehicle Routing Problem (CVRP).
ReFocusEraser: Refocusing for Small Object Removal with Robust Context-Shadow Repair
Qingping Zheng (Xiamen University), Jiankang Deng (Xiamen University)
RestorationSupervised Fine-TuningDiffusion modelImage
🎯 What it does: To address the problem of small object removal, a two-stage framework is proposed: the first stage utilizes camera adaptive zoom + LoRA fine-tuning to enhance texture restoration of small targets; the second stage eliminates color discrepancies and shadows through mask-based stitching and a shadow repair module that only fine-tunes the decoder, maintaining background consistency.
ReFORM: Reflected Flows for On-support Offline RL via Noise Manipulation
Songyuan Zhang (Massachusetts Institute of Technology), Chuchu Fan (Massachusetts Institute of Technology)
Knowledge DistillationReinforcement LearningFlow-based ModelBenchmark
🎯 What it does: Propose a reflection-flow-based offline reinforcement learning method called ReFORM, achieving a multi-modal action distribution without out-of-distribution (OOD) issues by constructing flow-based policies with support constraints.
ReForm: Reflective Autoformalization with Prospective Bounded Sequence Optimization
Guoxin Chen (Renmin University of China), Minpeng Liao (Alibaba Group)
OptimizationAI Code AssistantLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a Reflective Self-Formalization (REFORM) framework that can perform semantic self-verification and iterative correction simultaneously while generating formulas;
Reforming the Mechanism: Editing Reasoning Patterns in LLMs with Circuit Reshaping
Zhenyu Lei (University of Virginia), Jundong Li (University of Virginia)
Explainability and InterpretabilityTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposes the 'Reasoning Editing' paradigm, which enables selective modification of specific reasoning patterns in large language models (LLMs) while preserving other reasoning capabilities.
Reformulation for Pretraining Data Augmentation
Xintong Hao (ByteDance Seed), Chenggang Li (ByteDance Seed)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose a data augmentation framework called Massive Genre-Audience reformulation (MGA), which rephrases the original text by generating diverse genre-audience pairs, achieving large-scale, low repetition corpus expansion.
RefTool: Reference-Guided Tool Creation for Knowledge-Intensive Reasoning
Xiao Liu (Peking University), Yansong Feng (Peking University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelTextPhysics RelatedChain-of-Thought
🎯 What it does: Propose a framework called REFTOOL that automatically generates executable tools based on reference materials and uses them in reasoning.
ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding
Jia-Nan Li (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelDiffusion modelTextBenchmark
🎯 What it does: Propose a diffusion-based large language model named REFUSION, which combines sequence recombination with causal attention, achieving efficient generation through slot-level parallel selection and autoregressive filling.
RegionE: Adaptive Region-Aware Generation for Efficient Image Editing
Pengtao Chen (Fudan University), Tao Chen (Fudan University)
GenerationComputational EfficiencyTransformerDiffusion modelFlow-based ModelRectified FlowImageTextMultimodality
🎯 What it does: Proposes a training-agnostic, region-adaptive RegionE framework to accelerate instruction-based image editing (IIE) tasks.
RegionReasoner: Region-Grounded Multi-Round Visual Reasoning
Wenfang Sun (University of Amsterdam), Cees G. M. Snoek (University of Amsterdam)
Object DetectionSegmentationTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose RegionReasoner, a multi-round visual reasoning framework based on reinforcement learning, which forces the reasoning process to explicitly reference image region coordinates and incorporates a global-local semantic consistency reward, while constructing RegionDial-Bench as a benchmark for multi-round detection and segmentation.
Regret-Guided Search Control for Efficient Learning in AlphaZero
Yun-Jui Tsai (National Yang Ming Chiao Tung University), Ti-Rong Wu (Academia Sinica)
OptimizationConvolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: Proposes the Regret-Guided Search Control (RGSC) framework, extending AlphaZero by introducing a regret network and a priority regret buffer to restart self-play from high-regret states, thereby improving learning efficiency.
Regularized Latent Dynamics Prediction is a Strong Baseline For Behavioral Foundation Models
Pranaya Jajoo (University of Alberta), Martha White (University of Alberta)
Representation LearningReinforcement LearningWorld ModelSequential
🎯 What it does: This study proposes a state representation learning method based on Regularized Latent Dynamic Prediction (RLDP), aiming to provide more robust and generalizable features for zero-shot reinforcement learning.
Regulating Internal Alignment Flows for Robust Learning Under Spurious Correlations
Rajeev Ranjan Dwivedi (Indian Institute of Science Education and Research Bhopal), Vinod K. Kurmi
ClassificationDomain AdaptationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Propose Alignment-Gated Suppression (AGS), a lightweight regularizer that evaluates class-conditional alignment energy on internal neuron-class connections during training and suppresses them via percentile gates, enhancing robustness in the presence of pseudo-correlations.
REI-Bench: Can Embodied Agents Understand Vague Human Instructions in Task Planning?
Chenxi Jiang (Nanyang Technological University), Jianfei Yang
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmarkChain-of-Thought
🎯 What it does: The study investigates the impact of core referential ambiguity on LLM-driven robot task planning, constructs the REI-Bench benchmark, and proposes the TOCC method to enhance robustness.
ReIn: Conversational Error Recovery with Reasoning Inception
Takyoung Kim (University of Illinois Urbana Champaign), Dilek Hakkani-Tür (University of Illinois Urbana Champaign)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the REIN method, which helps LLM-based dialogue agents quickly diagnose and recover from interaction failures caused by user errors without modifying LLM parameters or system prompts, by injecting an initial reasoning block during the conversation.
Reinforced Latent Reasoning for LLM-based Recommendation
Yang Zhang (National University of Singapore), Tat-Seng Chua (National University of Singapore)
Recommendation SystemTransformerLarge Language ModelReinforcement LearningTextSequentialChain-of-Thought
🎯 What it does: Propose the LatentR3 framework in recommendation systems, replacing explicit Chain-of-Thought (CoT) reasoning with implicit latent reasoning to enhance LLM recommendation performance.
Reinforcement Learning Fine-Tuning Enhances Activation Intensity and Diversity in the Internal Circuitry of LLMs
Honglin Zhang (Tsinghua University), Yong Li (Tsinghua University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Systematic analysis of internal activation strength and diversity in large language models before and after reinforcement learning fine-tuning
Reinforcement Learning for Machine Learning Engineering Agents
Sherry Yang (New York University), Percy Liang (Stanford University)
TransformerReinforcement LearningAgentic AIPrompt EngineeringImageTextTabularBenchmark
🎯 What it does: Under the maximum likelihood estimation (MLE) environment, small language models (LMs) are updated via reinforcement learning gradients to enhance model performance on specific tasks, rather than relying solely on prompts from large models.
Reinforcement Learning via Value Gradient Flow
Haoran Xu (University of Texas at Austin), Amy Zhang (University of Texas at Austin)
Reinforcement Learning from Human FeedbackReinforcement LearningFlow-based ModelText
🎯 What it does: Propose a VGF method that frames behavior-regularized reinforcement learning as optimal transport from a reference distribution to a value-driven optimal policy distribution, and achieve policy parameterization-free implementation via particle gradient flow
Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs
Xumeng Wen (Microsoft Research Asia), Mao Yang (Microsoft Research Asia)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper explores and verifies how reinforcement learning with verifiable rewards (RLVR) enhances reasoning capabilities in large language models (LLMs) through implicit incentives for correct reasoning, and proposes the CoT-Pass@K evaluation metric along with corresponding theoretical analysis.
Reinforcement Mid-Training
Yijun Tian (Xi'an Jiaotong University), Wei Wang (University of Utah)
Large Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a Reinforcement Mid-Training (RMT) framework, aiming to leverage unlabeled pre-training data to enhance the model's reasoning and mathematical capabilities during the intermediate phase between pre-training and post-training through reinforcement learning.
Reinforcement Unlearning via Group Relative Policy Optimization
Efstratios Zaradoukas (Technical University of Munich), Gjergji Kasneci (Technical University of Munich)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposes a reinforcement learning (RL) framework called PURGE based on GRPO for achieving verifiable, targeted memory elimination in large language models (LLMs) without external rewards.
Reinforcing Diffusion Models by Direct Group Preference Optimization
Yihong Luo (Hong Kong University Of Science And Technology), Jing Tang (Hong Kong University Of Science And Technology)
GenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageOrdinary Differential Equation
🎯 What it does: In the post-training of diffusion models, the Direct Group Preference Optimization (DGPO) algorithm is proposed for model reinforcement learning;
Reinforcing General Reasoning Without Verifiers
Xiangxin Zhou (University of Chinese Academy of Sciences), Chao Du (Sea AI Lab)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a verifier-free reinforcement learning method called VeriFree, which directly maximizes the probability of generating reference answers, bypassing the answer verification step, making it suitable for general reasoning tasks where rules are difficult to verify.
Rejuvenating Cross-Entropy Loss in Knowledge Distillation for Recommender Systems
Zhangchi Zhu (East China Normal University), Wei Zhang (Shanghai Innovation Institute)
Recommendation SystemKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: Analyzed and improved the cross-entropy loss in knowledge distillation to better transmit ranking information in recommendation systems; proposed the RCE-KD method, which splits, samples, and adaptively fuses the teacher's top items, significantly enhancing distillation effectiveness.
ReLaSH: Reconstructing Joint Latent Spaces for Efficient Generation of Synthetic Hypergraphs with Hyperlink Attributes
Feiyan Ma (University of Michigan), Ji Zhu (University of Michigan)
GenerationData SynthesisScore-based ModelFlow-based ModelGraphTabularElectronic Health RecordsStochastic Differential Equation
🎯 What it does: Developed a general-purpose generative framework ReLaSH, achieving efficient synthesis of attributed hypergraphs through joint latent space embedding and distribution-agnostic generators.
Relational Feature Caching for Accelerating Diffusion Transformers
Byunggwan Son (Yonsei University), Bumsub Ham (Yonsei University)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideoText
🎯 What it does: Propose a Relational Feature Caching (RFC) framework that predicts and dynamically schedules feature caching by leveraging the relationship between input and output features, thereby accelerating Diffusion Transformers (DiT) inference.
Relational Graph Transformer
Vijay Prakash Dwivedi (Stanford University), Jure Leskovec (Stanford University)
ClassificationRepresentation LearningGraph Neural NetworkTransformerGraphTabularTime SeriesBenchmark
🎯 What it does: A new graph Transformer architecture is studied for converting relational databases into heterogeneous temporal graphs for end-to-end learning.
Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data
Rishabh Ranjan (Stanford University), Jure Leskovec (Stanford University)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringTabularBenchmark
🎯 What it does: Proposed Relational Transformer (RT), a foundational model capable of zero-shot prediction on new relational databases without requiring task- or dataset-specific fine-tuning.
Relationship Alignment for View-aware Multi-view Clustering
Shuangmei Peng (Anhui University of Technology), Xiaojun Wu (Jiangnan University)
OptimizationRepresentation LearningAuto EncoderContrastive LearningImageVideoBenchmark
🎯 What it does: Propose a multi-view clustering framework RAV that integrates cross-view relation alignment with perspective-adaptive weighted label contrastive learning, aiming to simultaneously preserve neighborhood structure consistency and avoid semantic conflicts caused by low-similarity views.
Relative Entropy Pathwise Policy Optimization
Claas A Voelcker (University of Toronto), Igor Gilitschenski (University of Toronto)
OptimizationReinforcement LearningBenchmark
🎯 What it does: Proposed a novel fully on-policy reinforcement learning algorithm called REPPO, which combines path gradient of Q-values with relative entropy constraints to achieve high sample efficiency and stable training.
Relative Value Learning
Marc Höftmann (Technical University of Dortmund), Stefan Harmeling (Technical University of Dortmund)
Convolutional Neural NetworkReinforcement LearningVideo
🎯 What it does: Propose a reinforcement learning framework (Relative Value Learning, RV) that uses an antisymmetric function to learn the value difference between state pairs, thereby removing the meaningless constant in absolute value;
Relatron: Automating Relational Machine Learning over Relational Databases
Zhikai Chen, Huzefa Rangwala (Amazon)
ClassificationHyperparameter SearchData-Centric LearningMeta LearningNeural Architecture SearchGraph Neural NetworkTransformerTabularBenchmark
🎯 What it does: Unify the design space of RDL and DFS in relational database prediction tasks, perform large-scale structured architecture search, and build a performance bank based on the search results; subsequently propose task embedding and the meta-selector Relatron, which automatically decides to use RDL or DFS and prunes the internal RDL search; further improve checkpoint selection using loss landscape metrics.
RelayFormer: A Unified Local-Global Attention Framework for Scalable Image and Video Manipulation Localization
Wen Huang (Tsinghua University), Shu-Tao Xia (Tsinghua University)
Anomaly DetectionTransformerImageVideo
🎯 What it does: To address the problem of visual forgery localization, the RelayFormer framework is proposed, which can uniformly process images and videos of different resolutions.
ReLi3D: Relightable Multi-view 3D Reconstruction with Disentangled Illumination
Jan-Niklas Dihlmann, Varun Jampani
GenerationData SynthesisTransformerNeural Radiance FieldImageMeshBenchmark
🎯 What it does: Rapidly generate complete relightable 3D meshes from sparse multi-view images, containing spatially varying PBR materials and HDR environmental lighting, with inference time under 0.3 seconds.
Reliability-Adjusted Prioritized Experience Replay
Leonard S. Pleiss (Technical University Munich), Maximilian Schiffer (Technical University Munich)
Reinforcement LearningBenchmark
🎯 What it does: Propose an improved experience replay strategy called ReaPER, which integrates the reliability assessment of Temporal Difference Error (TDE);
Reliable Evaluation of MRI Motion Correction: Dataset and Insights
Kun Wang (Technical University of Munich), Reinhard Heckel (Technical University of Munich)
RestorationOptimizationData-Centric LearningConvolutional Neural NetworkVision Language ModelBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: Propose the PMoC3D dataset and systematically evaluate three medical image motion correction assessment methods
Reliable Fine-Grained Evaluation of Natural Language Math Proofs
Wenjie Ma (University Of California Berkeley), Sewon Min (University Of California Berkeley)
Large Language ModelTextBenchmark
🎯 What it does: Built an evaluation system for natural language mathematical proofs
Reliable Poisoned Sample Detection against Backdoor Attacks Enhanced by Sharpness Aware Minimization
Mingda Zhang (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
Anomaly DetectionAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Investigated the issue of decreased effectiveness of Poisoned Sample Detection (PSD) under weak backdoor attacks (low contamination rates or weak triggers), and proposed enhancing the separability of samples and detection performance by amplifying the backdoor effect through Sharpness-Aware Minimization (SAM) training.
Reliable Probabilistic Forecasting of Irregular Time Series through Marginalization-Consistent Flows
Vijaya Krishna Yalavarthi (University of Hildesheim), Lars Schmidt-Thieme (University of Hildesheim)
TransformerFlow-based ModelTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Proposed a hybrid separable flow model called MOSES for reliable probabilistic prediction of irregular time series, ensuring marginal consistency while also capturing joint distribution expressions.
Reliable Weak-to-Strong Monitoring of LLM Agents
Neil Kale (Carnegie Mellon University), Zifan Wang
Safty and PrivacyAdversarial AttackPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the 'Monitor Red Teaming (MRT)' framework, designed threat models, adversarial strategies, and evaluation metrics, and proposed a new CUA-SHADE-Arena computer usage attack benchmark. Evaluated the robustness of multiple monitoring structures (baseline, hierarchical, sequential, hybrid) on SHADE-Arena and CUA-SHADE-Arena.
Remaining-data-free Machine Unlearning by Suppressing Sample Contribution
Xinwen Cheng (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a novel Machine Unlearning method called MU-Mis, which directly suppresses the contribution of samples during training by minimizing the input sensitivity gap between the target class logit and irrelevant class logit of the forgotten samples, thereby achieving model unlearning after training;
REMem: Reasoning with Episodic Memory in Language Agent
Yiheng Shu (Ohio State University), Yu Su (Ohio State University)
TransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the REMem two-stage framework for constructing time-aware event memory graphs and achieving recall and reasoning in language agents through tool-driven multi-step retrieval.
Remotely Detectable Robot Policy Watermarking
Michael Amir (University of Cambridge), Amanda Prorok (University of Cambridge)
Robotic IntelligenceTime SeriesSequential
🎯 What it does: Designed and implemented a detectable watermarking technique for robot control strategies under remote observation, addressing the physical observation gap problem.
Rényi Sharpness: A Novel Sharpness that Strongly Correlates with Generalization
Qiaozhe Zhang (Huazhong University of Science and Technology), Yingzhuang Liu (Huazhong University of Science and Technology)
OptimizationImage
🎯 What it does: Proposed Rényi sharpness—a Hessian spectrum measure based on negative Rényi entropy—and developed the Rényi Sharpness-Aware Minimization (RSAM) training method based on this measure;
RepIt: Steering Language Models with Concept-Specific Refusal Vectors
Vincent Siu (University of California, Santa Cruz), Chenguang Wang (University of California, Santa Cruz)
Safty and PrivacyExplainability and InterpretabilityRepresentation LearningTransformerTextBenchmark
🎯 What it does: Propose the REPIT framework to precisely isolate and edit rejection behaviors for specific concepts in language models, allowing 'unbanning' of specific categories such as weapon information while maintaining rejection of other harmful concepts.
Replicable Reinforcement Learning with Linear Function Approximation
Eric Eaton (University of Pennsylvania), Jessica Sorrell (Johns Hopkins University)
Reinforcement Learning
🎯 What it does: Proposed a replicable linear function approximation reinforcement learning algorithm, first presenting replicable ridge regression and centerless covariance estimation methods, then implementing generator models and replicable algorithms for linear Markov Decision Processes (MDP) under scenario exploration based on these tools.
Representation Alignment for Diffusion Transformers without External Components
Dengyang Jiang (Northwestern Polytechnical University), Jingdong Wang (Baidu Inc)
GenerationRepresentation LearningTransformerDiffusion modelImageText
🎯 What it does: Proposes a self-supervised representation alignment method called SRA, which does not require external representation components, and utilizes the step-by-step denoising process of the diffusion Transformer to achieve self-guided internal representations;
Representation-Based Exploration for Language Models: From Test-Time to Post-Training
Jens Tuyls (Princeton University), Jordan T. Ash (Microsoft Research)
Representation LearningTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies the application of reinforcement learning (RL) in language models, particularly how intentional exploration can discover new behaviors rather than merely enhancing existing ones. The study shows that representation-based exploration strategies significantly improve diversity and success rates, applicable to both post-training and inference settings.
Representational Alignment Across Model Layers and Brain Regions with Multi-Level Optimal Transport
Shaan Shah (University of California San Diego), Meenakshi Khosla (University of California San Diego)
Representation LearningTransformerImageTextBiomedical Data
🎯 What it does: Proposed and implemented a Multi-Level Optimal Transport (MOT) framework to achieve layer-level and neuron-level soft global alignment across models or brain regions of different depths, while providing alignment scores at the single-network level.
Representing local protein environments with machine learning force fields
Meital Bojan (IST Austria), Alexander Bronstein (IST Austria)
Representation LearningProtein Structure PredictionGraph Neural NetworkBiomedical Data
🎯 What it does: Construct a general representation of protein local environments using atomic embeddings from pre-trained machine learning force fields (MLFF), validated on downstream tasks such as pKa prediction, chemical shift prediction, and secondary structure identification.
RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning
Mingrui Wu (Xiamen University), Rongrong Ji (Microsoft)
GenerationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes the RePrompt framework based on reinforcement learning, which significantly improves the semantic and visual alignment in text-to-image tasks by training language models to generate enhanced prompts containing reasoning trajectories.
RepSpec: Structural Re-parameterized Draft Model Training for Speculative Decoding
Feiye Huo (Peking University), Shengli Sun (Peking University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the RepSpec method, which enhances the acceptance length and speed during inference by expanding linear branches of the draft model through structural reparameterization during training.
Repurposing Foundation Model for Generalizable Medical Time Series Classification
Nan Huang (University of North Carolina at Charlotte), Xiang Zhang (University of North Carolina at Charlotte)
ClassificationTransformerSupervised Fine-TuningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Proposed a framework (FORMED) that repurposes a generic time series foundation model for medical time series classification by freezing the pre-trained backbone and introducing adaptable channel embeddings, label queries, and shared decoding attention to flexibly handle varying channels, lengths, and class numbers;
Repurposing Synthetic Data for Fine-grained Search Agent Supervision
Yida Zhao (School of Information Science and Technology, ShanghaiTech University), Yong Jiang (Tongyi Lab, Alibaba Group)
Data-Centric LearningReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose E-GRPO, which improves the sparse reward problem of traditional GRPO by leveraging entity information in synthetic data to provide dense rewards for the search agent.
RESA: Bringing Back What Sparse Attention Ignores with Residual Estimation
Weihao Yang (Harbin Institute of Technology), Xiangyu Zou (Harbin Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes the RESA framework, which enhances the inference quality and efficiency of large language models (LLMs) by estimating residuals through a low-rank prior on top of sparse attention, compensating for the contributions of unselected key-value (KV) pairs to the attention output.
RESCHED: Rethinking Flexible Job Shop Scheduling from a Transformer-based Architecture with Simplified States
Xiangjie Xiao (Singapore Management University), Zhiguang Cao (Singapore Management University)
OptimizationTransformerReinforcement LearningBenchmark
🎯 What it does: Proposed a Transformer-based deep reinforcement learning framework named RESCHED, which employs a minimalist state representation (only four features) to address flexible job shop scheduling problems, and validated its superior performance on FJSP, JSSP, and FFSP benchmarks.
ResCP: Reservoir Conformal Prediction for Time Series Forecasting
Roberto Neglia (UiT Arctic University of Norway), Filippo Maria Bianchi (UiT Arctic University of Norway)
Recurrent Neural NetworkTime Series
🎯 What it does: Proposes a conformal prediction method (RESCP) that utilizes randomized recurrent networks (Reservoir Computing) to construct prediction intervals for time series, along with asymptotic conditional coverage guarantees under theoretical conditions.
RESCUE: Retrieval Augmented Secure Code Generation
Jiahao Shi (Purdue University), Tianyi Zhang (Purdue University)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-enhanced secure code generation framework called RESCUE, which automatically builds a hierarchical security knowledge base and enhances LLM secure code generation through multi-dimensional retrieval.
ResearchRubrics: A Benchmark of Prompts and Rubrics For Evaluating Deep Research Agents
Manasi Sharma (Scale AI), Bing Liu (Scale AI)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the RESEARCHRUBRICS deep research evaluation benchmark, equipped with 101 multi-domain real-world questions and over 2500 fine-grained evaluation criteria written by human experts;
RESFL: An Uncertainty-Aware Framework for Responsible Federated Learning by Balancing Privacy, Fairness and Utility
Dawood Wasif (Virginia Tech), Jin-Hee Cho (Virginia Tech)
Object DetectionFederated LearningSafty and PrivacyImageText
🎯 What it does: Under the federated learning framework, RESFL is proposed to simultaneously enhance model privacy, fairness, and robustness, with its effectiveness verified on object detection tasks.
Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection
Xingwu Chen (University of Hong Kong), Difan Zou (University of Hong Kong)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Investigate the training dynamics of reinforcement learning (RL) in large language models through empirical analysis and theoretical modeling, clarifying how RL improves model performance by optimizing sparse key tokens and reshaping reasoning mode distributions.
Residual Feature Integration is Sufficient to Prevent Negative Transfer
Yichen Xu (University of California Berkeley), Lexin Li (University of California Berkeley)
Domain AdaptationImageTextTabularBiomedical Data
🎯 What it does: Propose and study a residual feature fusion (REFINE) method that captures residual signals by adding a trainable target-side encoder on frozen source model features, thereby preventing negative transfer.
ResiliBench: Evaluating Agentic Workflow Adaptation in Stochastic Environments
Ruicheng Ao (Massachusetts Institute of Technology), Xinshang Wang (Massachusetts Institute of Technology)
TransformerReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the ResiliBench benchmark to evaluate the workflow execution capability of LLMs in real-world environments with fluctuating tool reliability and instruction quality.
Resisting Contextual Interference in RAG via Parametric-Knowledge Reinforcement
Chenyu Lin (Baidu Inc), Zhonghou Lv
GenerationRetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the Knowledgeable-R1 framework, training LLMs in retrieval-augmented generation (RAG) tasks by dynamically balancing parameter knowledge (PK) and retrieved context (CK) through joint sampling and reinforcement learning, thereby mitigating context interference.
Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis
Pengfei ZHANG, Li Liu (Hong Kong University of Science and Technology)
ClassificationGenerationData SynthesisAnomaly DetectionTransformerLarge Language ModelAgentic AIFlow-based ModelMultimodalityBiomedical DataAudio
🎯 What it does: Proposed a closed-loop multimodal system, Resp-Agent, integrating breath sound synthesis and diagnosis to achieve controllable breath sound generation and disease classification;
ReSplat: Degradation-agnostic Feed-forward Gaussian Splatting via Self-guided Residual Diffusion
Youngho Yoon (KAIST), Kuk-Jin Yoon (KAIST)
RestorationGenerationDiffusion modelGaussian SplattingImage
🎯 What it does: Propose a generic denoising and view synthesis framework called ReSplat, which can achieve multi-view image restoration and novel view rendering under any degraded input;
ReST-KV: Robust KV Cache Eviction with Layer-wise Output Reconstruction and Spatial-Temporal Smoothing
Yongqi An (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)
Computational EfficiencyTransformerText
🎯 What it does: Propose a KV cache eviction method called ReST-KV to reduce memory usage and improve speed during long context inference in LLMs.
ResT: Reshaping Token-Level Policy Gradients for Tool-Use Large Language Models
Zihan Lin (MAIS and NLPR, Institute of Automation, Chinese Academy of Sciences), Ran He (Meituan)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a tool calling method based on reinforcement learning called ResT, significantly improving the performance of large language models in multi-round tool usage tasks through entropy-aware reshaping of token-level policy gradients and lightweight curriculum learning.
RestoreVAR: Visual Autoregressive Generation for All-in-One Image Restoration
Sudarshan Rajagopalan (Johns Hopkins University), Vishal M. Patel (Johns Hopkins University)
RestorationTransformerAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Propose a versatile image restoration framework called RestoreVAR based on Visual Autoregressive (VAR), achieving high-quality restoration for multiple degradation types
RESTRAIN: From Spurious Votes to Signals — Self-Training RL with Self-Penalization
ZHAONING YU, Jing Xu (FAIR at Meta SuperIntelligence Lab)
TransformerReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a self-penalizing reinforcement learning framework called RESTRAIN, enabling language models to self-generate learning signals through their own prediction distributions and achieve self-improvement without gold-standard data.
Resurfacing the Instance-only Dependent Label Noise Model through Loss Correction
Mustafa Enes Aydın, Alexander Bertrand (KU Leuven)
ClassificationImageTabularAudio
🎯 What it does: In binary classification tasks, the authors propose an instance-aware loss correction method based on risk equivalence to address label noise.
ResWorld: Temporal Residual World Model for End-to-End Autonomous Driving
Jinqing Zhang (Beihang University), Yunhong Wang (Beihang University)
Autonomous DrivingWorld ModelImageBenchmark
🎯 What it does: Propose a temporal residual-based world model called ResWorld, which focuses on dynamic object prediction using residuals to improve end-to-end autonomous driving planning;
ReTabAD: A Benchmark for Restoring Semantic Context in Tabular Anomaly Detection
Sanghyu Yoon (LG AI Research), Woohyung Lim (LG AI Research)
Anomaly DetectionLarge Language ModelPrompt EngineeringTextTabularBenchmarkFinance Related
🎯 What it does: Constructed the ReTabAD benchmark, collected and annotated 20 table datasets with structured text metadata, and provided implementations of 20 unsupervised anomaly detection algorithms along with a zero-shot LLM baseline.
Retain and Adapt: Auto-Balanced Model Editing for Open-Vocabulary Object Detection under Domain Shifts
Zixuan Duan (Nanjing University), Qi Fan (Nanjing University)
Object DetectionDomain AdaptationSupervised Fine-TuningImageText
🎯 What it does: Studied methods for quickly adapting open-vocabulary object detection (OVOD) under domain drift environments using model editing, and proposed the Auto-Balanced Model Editing (ABME) framework.
Retaining Suboptimal Actions to Follow Shifting Optima in Multi-Agent Reinforcement Learning
Yonghyeon Jo (Ulsan National Institute of Science and Technology), Seungyul Han (Ulsan National Institute of Science and Technology)
Reinforcement LearningBenchmark
🎯 What it does: Proposed the Successive Sub-value Q-learning (S2Q) framework, which utilizes multiple sub-value functions to continuously track sub-optimal actions, addressing the issue of value functions dynamically changing during training.
Rethinking Benign Relearning: Syntax as the Hidden Driver of Unlearning Failures
Sangyeon Yoon (Yonsei University), Albert No (Yonsei University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The paper investigates the 'goodwill relearning' phenomenon when machine models forget specific data, finding that syntactic similarity is the primary factor leading to the recovery of forgotten information, and proposes enhancing the model's forgetting effect by generating diverse syntactic variants (syntactic diversification).
Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models
Yi Ding (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a multi-image safety dataset called MIS, and the MIRage method trained on this dataset;
Rethinking Causal Mask Attention for Vision-Language Inference
Xiaohuan Pei (University of Sydney), Chang Xu (University of Sydney)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Conduct a systematic analysis of causal attention mechanisms in VLMs, and propose future-aware causal masks (full vision, vision-vision, vision-text) as well as a lightweight future attention compression method;
Rethinking Code Similarity for Automated Algorithm Design with LLMs
Rui Zhang (City University of Hong Kong), Zhichao Lu (City University of Hong Kong)
OptimizationAI Code AssistantTransformerLarge Language ModelSequential
🎯 What it does: Proposed a method called BehaveSim for algorithm similarity measurement based on problem-solving behavior trajectory (PSTrajectory), and applied it to the scenarios of large language model automatic algorithm design (LLMAAD) and algorithm analysis.
Rethinking Consistent Multi-Label Classification Under Inexact Supervision
Wei Wang (RIKEN), Masashi Sugiyama (University of Tokyo)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: Proposes a unified COMES framework to address weakly supervised problems in both partial multi-label learning and complementary multi-label learning, and designs a risk estimator that does not require estimating the label generation process.
Rethinking Continual Learning with Progressive Neural Collapse
Zheng Wang (University of Houston), Sen Lin (University of Houston)
Knowledge DistillationRepresentation LearningConvolutional Neural NetworkImageBenchmark
🎯 What it does: Propose a Progressive Neural Collapse (ProNC) framework for continuous learning tasks, which enhances feature representations and reduces catastrophic forgetting by dynamically generating and aligning simple ETF objectives.
Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods
Wanru Zhao (University of Cambridge), Nicholas D. Lane (University of Cambridge)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the ADAPT framework, which dynamically adjusts the importance of training samples through online sample reweighting, replacing traditional offline screening/mixing processes, to achieve adaptive data processing for large language model (LLM) training and instruction tuning;
Rethinking Driving World Model as Synthetic Data Generator for Perception Tasks
Kai Zeng (Peking University), Wentao Zhang (Peking University)
GenerationData SynthesisAutonomous DrivingTransformerDiffusion modelWorld ModelVideoPoint CloudMesh
🎯 What it does: Propose the Dream4Drive framework, using 3D-aware guidance maps to edit videos and generate multi-view synthetic videos for downstream perception tasks; and constructing the DriveObj3D large-scale 3D asset library.
Rethinking Expressivity and Degradation-Awareness in Attention for All-in-One Blind Image Restoration
Bin Ren (Mohamed bin Zayed University of Artificial Intelligence), Nicu Sebe (University of Trento)
RestorationConvolutional Neural NetworkTransformerImageBiomedical DataBenchmark
🎯 What it does: Proposed a novel full-blind image restoration model called ExDA, capable of uniformly handling various unknown mixed degradations such as noise reduction, blurring, and haze.
Rethinking Global Text Conditioning in Diffusion Transformers
Nikita Starodubcev (Yandex Research), Dmitry Baranchuk (Yandex Research)
GenerationTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: This paper re-evaluates the global text modulation mechanism in diffusion transformers, proposing a training-free 'modulation guidance' method that utilizes pooled CLIP embeddings for controllable guidance in the modulation space;
Rethinking JEPA: Compute‑Efficient Video Self-Supervised Learning with Frozen Teachers
Xianhang Li (Apple), Etai Littwin (Apple)
Computational EfficiencyRepresentation LearningAuto EncoderVideo
🎯 What it does: Propose a two-stage video self-supervised pre-training method called SALT, which first trains a static teacher and then uses its features to train the student
Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity
Cristian Hinostroza (Pontificia Universidad Católica de Chile), Jorge F Silva
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates methods for evaluating the importance of layers in large language models, demonstrating the severe flaws of cosine similarity as a layer relevance metric, and proposes a layer importance measurement based on actual accuracy degradation. Subsequently, experiments are conducted on multiple models and tasks, and this metric is applied to structured pruning, showcasing its superiority.
Rethinking LLM Evaluation: Can We Evaluate LLMs with 200× Less Data?
Shaobo Wang, Linfeng Zhang
CompressionLarge Language ModelTextBenchmark
🎯 What it does: The research objective is to compress large language model evaluation benchmarks to just dozens of examples while maintaining the accuracy of model rankings.
Rethinking LLM Reasoning: From Explicit Trajectories to Latent Representations
Cong Jiang (Harbin Institute of Technology), Zheng Zhang (Harbin Institute of Technology)
Computational EfficiencyKnowledge DistillationRepresentation LearningSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the Latent Reasoning Tuning (LRT) framework, which uses a lightweight reasoning network to generate implicit reasoning representations, replacing traditional long-text reasoning trajectories.
Rethinking LLM-as-a-Judge: Representation-as-a-Judge with Small Language Models via Semantic Capacity Asymmetry
Zhuochun Li (Ping An Technology Co., Ltd.), Daqing He (University of Pittsburgh)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: The study investigates whether the internal hidden representations of small language models can replace large models for no-reference evaluation, proposing the Representation-as-a-Judge paradigm and implementing the INSPECTOR framework.
Rethinking LoRA for Privacy-Preserving Federated Learning in Large Models
Jin Liu (Xidian University), Junkang Liu (Xidian University)
Federated LearningSafty and PrivacyImageText
🎯 What it does: This paper proposes a privacy-preserving federated learning framework called LA-LoRA, which alternately updates the two low-rank matrices of LoRA within each local training round and applies low-pass filtering to gradients to mitigate issues of gradient coupling, noise amplification, and sharp global convergence.
Rethinking Model Calibration through Spectral Entropy Regularization in Medical Image Segmentation
Kun Cheng (Beijing University of Posts and Telecommunications), Tonggang Zhao (Beijing University of Posts and Telecommunications)
SegmentationConvolutional Neural NetworkBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper investigates the problem of model confidence misalignment in medical image segmentation and proposes an adaptive calibration framework based on frequency domain analysis.
Rethinking Pareto Frontier: On the Optimal Trade-offs in Fair Classification
Junyi Chai (Purdue University), Xiaoqian Wang (Purdue University)
ClassificationSupervised Fine-TuningImageTabularBiomedical Data
🎯 What it does: This paper proposes a model-specific (MS) Pareto-optimal fairness-accuracy trade-off and the balance between fairness and accuracy, and designs a group-dependent bias-aware last-layer fine-tuning framework to improve the equilibrium between fairness and accuracy;
Rethinking Policy Diversity in Ensemble Policy Gradient in Large-Scale Reinforcement Learning
Naoki Shitanda (University of Tokyo), Takayuki Osa (RIKEN Center for Advanced Intelligence Project)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: In a large-scale parallel reinforcement learning environment, a method called Coupled Policy Optimization (CPO) is proposed, which regulates diversity among multiple policies by incorporating KL constraints and adversarial rewards within a leader-follower framework, thereby achieving more efficient and stable sample collection and learning.
Rethinking Radiology Report Generation: From Narrative Flow to Topic-Guided Findings
Sheng Cheng (Rice University), Devika Subramanian (Rice University)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBiomedical Data
🎯 What it does: Propose the LLaVA-TA framework, decomposing radiology report generation into independent topics and achieving visual-textual alignment through anatomical region masks
Rethinking Reasoning in Document Ranking: Why Chain-of-Thought Falls Short
Xuan Lu (Shanghai Jiao Tong University), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Systematically evaluated the effectiveness of chained reasoning in document re-ranking, comparing point-wise and list-wise re-ranking models, using SFT and GRPO training.
Rethinking Residual Errors in Compensation-based LLM Quantization
Shuaiting Li (Zhejiang University), Kejie Huang (Zhejiang University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Redefined the residual error of GPTQ and GPTAQ, expanding the error source to internal errors caused by weight compensation and incorporating it into weight updates;