ICML 2026 Papers — Page 48
International Conference on Machine Learning · 6554 papers
Recovering Policy-Induced Errors: Benchmarking and Trajectory Synthesis for Robust GUI Agents
Tianpeng Bu (Alibaba Cloud Computing), Minying Zhang (Alibaba Cloud Computing)
Data SynthesisRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a robustness evaluation and enhancement scheme for GUI agents in strategy-induced errors: first, a GUI-RobustEval test set consisting of 1,216 executable error recovery scenarios is constructed, followed by the design of a tree-based online sampling framework, RoTS, for automatically synthesizing diverse, long-delay error-recovery trajectories, which are then used to train RoTS-7B/32B models.
Rectified LpJEPA: Joint-Embedding Predictive Architectures with Sparse and Maximum-Entropy Representations
Yilun Kuang (New York University), Yann LeCun (New York University)
Representation LearningTransformerDiffusion modelScore-based ModelRectified FlowAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes a new self-supervised learning framework called Rectified LpJEPA, which aligns in the feature space to the Rectified Generalized Gaussian (RGG) distribution by utilizing the Joint Embedding Prediction Architecture (JEPA) and a novel sparse regularization method called RDMReg, thus achieving controllable sparse, non-negative, and high-entropy representation learning.
RECTOR: Masked Region-Channel-Temporal Modeling for Affective and Cognitive Representation Learning
Jinhan Liu (EPFL), Mahsa Shoaran (EPFL)
Representation LearningTransformerAuto EncoderContrastive LearningBiomedical DataElectrocardiogram
🎯 What it does: Propose an end-to-end self-supervised framework called RECTOR for masked learning of regional-channel-temporal features in EEG/sEEG, aiming to capture distributed brain network dynamics related to emotional and cognitive states.
Recurrent Equivariant Constraint Modulation: Learning Per-Layer Symmetry Relaxation from Data
Stefanos Pertigkiozoglou (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
Drug DiscoveryRecurrent Neural NetworkGraph Neural NetworkDiffusion modelContrastive LearningGraphBiomedical Data
🎯 What it does: A mechanism named Recursive Equivariant Constraint Modulation (RECM) is proposed, which learns an appropriate relaxation level by leveraging the symmetry properties of the training signal and the input-target distribution of each layer, without the need to pre-specify task-related relaxation levels.
Recurrent Structural Policy Gradient for Partially Observable Mean Field Games
Clarisse Wibault (University of Oxford), Jakob Nicolaus Foerster
OptimizationComputational EfficiencyRecurrent Neural NetworkTransformerReinforcement LearningAuto EncoderTabularTime SeriesSequentialBenchmark
🎯 What it does: Proposes RSPG (Recurrent Structural Policy Gradient), a hybrid structural approach capable of using memory-based policies in partially observable Mean Field Games (MFG) with public noise; simultaneously introduces the MFAX framework, which supports white-box/black-box transfer, partial observability, public noise, and multi-initial equilibrium MFG environments.
Recursive Binding on a Budget: Subspace Carving in Order-$p$ Tensor Memories
Travis Pence (University of WisconsinMadison), Vikas Singh (University of WisconsinMadison)
RetrievalComputational EfficiencyRepresentation LearningContrastive LearningTextTabularBenchmark
🎯 What it does: Propose a new tensor memory architecture called Orthogonal Subspace Carving (OSC), which achieves deep recursive binding while maintaining low memory usage by projecting filling vectors into the orthogonal complement of the role subspace and then tensorizing the result.
Recursive Models for Long-Horizon Reasoning
Chenxiao Yang (Toyota Technological Institute at Chicago), Zhiyuan Li (Toyota Technological Institute at Chicago)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose a recursive model (through call/return mechanism) to achieve long-range reasoning, and conduct experiments on SAT to verify its effectiveness.
Recursive Monte-Carlo Tree Search
Benjamin Howard (Institute for Defense Analyses, Center for Communications Research), Keith Frankston (Institute for Defense Analyses, Center for Communications Research)
OptimizationConvolutional Neural NetworkTransformerReinforcement LearningMixture of ExpertsTabularSequential
🎯 What it does: A recursive AlphaZero-style MCTS algorithm called RMCTS was studied, which achieves high parallelism and significant speed improvement during single-threaded search by recursively computing and optimizing the posterior policy for each node.
RED-HDP-HMM: Observation-Dependent Durations for Bayesian Nonparametric Sequential Models
Mikołaj Słupiński (University of Wrocław), Piotr Lipinski (University of Wrocław)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkContrastive LearningTime SeriesSequentialBiomedical Data
🎯 What it does: Proposes RED-HDP-HMM, a Bayesian nonparametric HMM that allows state duration to depend on previous observations, and provides a weak limit Gibbs sampling inference method.
RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates
Ali Asad (Queen's University), Xiaodan Zhu (Queen's University)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This work proposes the RedDebate framework, which utilizes multi-agent LLMs to engage in debates within red team adversarial scenarios, and achieves fully automated security discovery and improvement through an evaluator and feedback generator; the framework integrates multiple memory modules to support continuous learning and defense.
Reducing Per-Sample Harm in Stochastic Optimization
Apostolos Avranas (Amadeus)
OptimizationConvolutional Neural NetworkTransformerContrastive LearningImageStochastic Differential Equation
🎯 What it does: Proposes a framework to explicitly minimize the loss increase (harm) of individual samples during stochastic gradient descent, integrating it into standard optimizers (SGD+momentum and AdamW), making each update step more balanced across all samples.
Reduction of Probabilistic Chemical Reaction Networks
Mauricio Montes (Auburn University), Grégoire Sergeant-Perthuis (Sorbonne Universite)
Computational EfficiencyDrug DiscoveryGraph Neural NetworkGraphOrdinary Differential Equation
🎯 What it does: Propose a factor graph-based chemical reaction network (CRN) compilation and compression framework, implementing Bayesian inference (BP) as chemical kinetics and significantly reducing the scale of CRN through factor graph degeneration.
RedVisor: Reasoning-Aware Prompt Injection Defense via Zero-Copy KV Cache Reuse
Mingrui Liu (Nanyang Technological University), Kwok-Yan Lam (Nanyang Technological University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a unified framework called RedVisor, which utilizes a lightweight top adapter and two-stage inference to achieve both detection and prevention of Prompt Injection in LLMs, while retaining the original model's performance during the generation phase and offering strong interpretability.
RefChess: Training-Free Contextual Search for Zero-Shot Referring Image Segmentation
Shiyan Tong (Southeast University), Haikun Wei (Southeast University)
SegmentationTransformerPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes RefChess, a training-agnostic context search framework aimed at achieving more robust mask candidate selection in zero-shot referential image segmentation.
Reference-Free Meta-Learning for Generalized Implicit Neural Representation in Efficient MRI Reconstruction
Haonan Zhang (Shanghai Jiao Tong University), Hongjiang Wei (Shanghai Jiao Tong University)
RestorationComputational EfficiencyMeta LearningConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a no-reference meta-learning framework called IPOD, which learns general initialization parameters for implicit neural representations (INR) in magnetic resonance imaging (MRI) reconstruction tasks, enabling fast and accurate reconstruction.
Referring Multiple Regions with Large Multimodal Models via Contextual Latent Steering
Yun Xing (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose Contextual Latent Steering (CSteer), a training-free and architecture-agnostic method that enables large multimodal models to perform better in multi-region and context-aware visual referring tasks.
Refined Analysis of Entropy-Regularized Actor-Critic
Safwan Labbi (Institut Polytechnique de Paris), Eric Moulines (Mohamed bin Zayed University of Artificial Intelligence)
Reinforcement LearningTabular
🎯 What it does: This paper studies the role of the critic in variance reduction in entropy-regularized Actor-Critic (AC) algorithms, and provides non-asymptotic convergence analysis.
RefineEvo: Planning-Guided Heuristic Evolution with Bidirectional Experience
Yang Wu (Chinese Academy of Sciences), Jian Cheng (Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelReinforcement LearningGraphTabularBenchmark
🎯 What it does: Proposes RefineEvo, a multi-agent evolutionary framework based on planning and a bidirectional experience pool, which uses LLMs to automatically design heuristic algorithms.
Refining Context-Entangled Content Segmentation via Curriculum Selection and Anti-Curriculum Promotion
Chunming He (Duke University), Sina Farsiu (Duke University)
SegmentationConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Propose a two-stage curriculum learning framework called CurriSeg, which first uses dynamic difficulty screening and pixel uncertainty to stabilize training, and then improves robustness through spectral denoising, addressing segmentation tasks with intertwined contexts.
Refining Dual Spectral Sparsity in Transformed Tensor Singular Values
Andong Wang (RIKEN AIP), Qibin Zhao (RIKEN AIP)
RestorationOptimizationImageVideoBiomedical Data
🎯 What it does: Propose a new tensor ℓp-Schattenq quasi-norm to simultaneously constrain the sparsity between frequency domains and the low-rank property within frequency domains;
Reflect-then-Correct: Rebalancing Task Optimization for Generalizable Meta-Reinforcement Learning via Distributional Value Error Reduction
Min Wang (Beijing Institute of Technology), Yonggang Zhang (Jilin University)
OptimizationMeta LearningReinforcement Learning
🎯 What it does: A framework named RTC is proposed, which utilizes Sinkhorn divergence to perform distributed learning of Meta-values, and achieves task optimization rebalancing through importance weighting via a recursive Bellman error model.
Reflective Hamiltonian Monte Carlo: Mixing Analysis and Application to Sampling on Stiefel Manifold
Kwangmin Lee (Chonnam National University), Sewon Park (Sookmyung Women's University)
OptimizationComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackScore-based ModelContrastive LearningGraphTabularTime SeriesSequentialPhysics RelatedStochastic Differential Equation
🎯 What it does: Proposed and theoretically proved the non-asymptotic mixing time of the reflective Hamiltonian Monte Carlo (ReHMC) algorithm on bounded domains, and constructed a 'convex container plus sampling' framework to handle arbitrary bounded supports, which was subsequently applied to sampling on the Steifel manifold.
Reflector: Internalizing Step-wise Reflection against Indirect Jailbreaks
Jiachen Ma (Fudan University), Chao Yang (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Designed and implemented REFLECTOR, a two-stage training framework embedded with self-reflection during the LLM generation process, aimed at defending against hidden indirect jailbreak attacks.
Reflex: Real-Time Vision-Language-Action Control through Streaming Inference
Yuanchun Guo (Beijing University of Posts and Telecommunications), Bingyan Liu (Beijing University of Posts and Telecommunications)
Computational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelVideoTextMultimodality
🎯 What it does: Propose the Reflex framework to achieve streaming inference to support real-time Vision-Language-Action (VLA) control; by partitioning attention context, AdaRMSNorm normalization, asynchronous pipeline, and future state prediction, achieve 50 Hz continuous inference;
ReflFlow: Learning Geometry-Guided Ray Tracing for Dynamic Specular Reconstruction
Jiachen Tao (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
GenerationData SynthesisTransformerDiffusion modelRectified FlowNeural Radiance FieldGaussian SplattingOptical FlowVideo
🎯 What it does: Proposes the ReflFlow framework for reconstructing dynamic mirror scenes from monocular videos and achieving real-time high-quality view synthesis.
REG: In-Sample RL via Regularizing the Evaluation Gap
Hanpu Shen (University of California Irvine), Roy Fox (University of California Irvine)
OptimizationReinforcement LearningDiffusion modelContrastive LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: This paper proposes a REG framework based on robust optimization for offline reinforcement learning, achieving safe on-sample learning through regularized evaluation error;
ReGen: Hierarchical Multi-Prompt Representation Generation for Efficient Waveform Diffusion Models
Sang-Hoon Lee (Ajou University), Ha-Yeong Choi (KT Corp)
GenerationCompressionRepresentation LearningTransformerPrompt EngineeringDiffusion modelFlow-based ModelAudio
🎯 What it does: Proposed a hierarchical multi-prompt representation generation framework called ReGen, which simultaneously generates semantic, acoustic, and waveform representations in a single-stage Waveform Diffusion Transformer, thereby improving audio reconstruction quality at low bitrates.
Regime-Adaptive Bayesian Optimization via Dirichlet Process Mixtures of Gaussian Processes
Yan Zhang (Florida State University), Shibo Li (Florida State University)
OptimizationDrug DiscoveryDiffusion modelScore-based ModelContrastive LearningGaussian SplattingTabularTime SeriesBiomedical DataPhysics Related
🎯 What it does: Propose a Bayesian optimization framework based on Dirichlet process mixture of Gaussian processes (RAMBO), which can adaptively discover and model discrete 'protocol' regions in the search space.
Regression Language Models for Code
Yash Akhauri (Cornell University), Mohamed S. Abdelfattah (Cornell University)
AI Code AssistantNeural Architecture SearchTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphTabular
🎯 What it does: Propose a unified Regression Language Model (RLM) that directly predicts multiple performance metrics from code text using a frozen T5Gemma encoder.
Regret Minimization With a Crowd of Awakening Experts
Anna Lunghi (Politecnico di Milano), Matteo Castiglioni (Politecnico di Milano)
OptimizationMeta LearningReinforcement LearningMixture of ExpertsContrastive LearningTextTabularTime Series
🎯 What it does: Studied the Awake Collaborator Expert (ACE) problem, proposing sublinear-order scheduling strategies in both random and adversarial environments for an online learning framework where new experts join in each round and never leave;
Regret Pre-training: Bridging Prior and Posterior Views for Enhanced Knowledge Grounding
Mingkuan Zhao (Xi'an Jiaotong University), Jiayin Wang (Xi'an Jiaotong University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Through a dual-view pre-training framework, causal language models utilize future context as latent information during training to enhance knowledge representation.
Regret-Based Federated Causal Discovery with Unknown Interventions
Federico Baldo (Sorbonne Universite), Charles K. Assaad (Sorbonne Universite)
Federated LearningSafty and PrivacyExplainability and InterpretabilityDiffusion modelContrastive LearningGraphTabular
🎯 What it does: In a federated learning environment, an algorithm named I-PERI for causal structure learning is designed, which can recover the causal graph on the server side even when unknown interventions exist on the clients and the intervention targets are not shared.
Regularized Discriminative Alignment for Deep Representations under Label Shift
Hengchao Shi (Fudan University), Ming Zheng (Fudan University)
Domain AdaptationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes a novel label shift adaptation framework called RDALS, aiming to address the label shift problem by aligning distributions in a deep latent space.
Regularized Offline Policy Optimization with Posterior Hybrid Bayesian Belief
Hongqiang Lin (Zhejiang University), Nenggan Zheng (Zhejiang University)
OptimizationReinforcement LearningTabularTime SeriesBenchmark
🎯 What it does: This paper proposes a Bayesian posterior hybrid method called Posterior Hybrid Bayesian Belief (PhyB), which is used to uniformly handle cognitive uncertainty at the sample level and model level in offline reinforcement learning;
Regulating Anatomy-Aware Rewards via Trajectory-Integral Feedback for Volumetric Computed Tomography Analysis
Tianwei Lin (Zhejiang University), Ling Zhang (Damo Academy, Alibaba Group)
Anomaly DetectionReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextBiomedical DataComputed TomographyBenchmark
🎯 What it does: Designed and implemented the structured clinical abnormality benchmark CABS and the RL framework TIF-GRPO based on integral feedback, aiming to enhance the factual accuracy and reliability of medical vision-language models (VLMs) in 3D CT analysis.
Reinforced Sequential Monte Carlo for Amortised Sampling
Sanghyeok Choi (University of Edinburgh), Esmeralda S. Whitammer (University of Edinburgh)
OptimizationReinforcement LearningScore-based ModelTabularTime SeriesSequentialBiomedical Data
🎯 What it does: Propose a framework that combines reinforcement learning with Sequential Monte Carlo (SMC), using off-policy RL to train a reusable sampler, and utilizing samples generated by SMC for experience replay to improve the sampling quality for multi-modal distributions.
Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Training
Song Lai (Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences), Fei Zhu (Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences)
Computational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Compared the forgetting performance of supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RFT) in continual training, and proposed the rollout-based instance filtering algorithm RIF-RFT;
Reinforcement Learning for Reachability: Guaranteeing Asymptotic Optimality
Amogh Palasamudram (Georgia Institute of Technology), Krishnendu Chatterjee (Institute of Science and Technology)
Reinforcement LearningTabularBenchmark
🎯 What it does: Propose a phased model-based reinforcement learning algorithm specifically designed for reachability goals in unknown MDPs, achieving asymptotically optimal policies by progressively estimating the minimum transition probability and using PAC learning.
Reinforcement Learning for Tool-Calling Agents in Fast Healthcare Interoperability Resources (FHIR)
Marius Knorr, Nils Schweingruber (IDM gGmbH, University Medical Center Hamburg-Eppendorf)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextGraphBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented Generation
🎯 What it does: For clinical question answering tasks under the FHIR data structure, this paper implements a multi-round CodeAct agent, which performs graph traversal, filtering, and aggregation through a retrieval tool and a Python interpreter, and further uses GRPO for post-training, enabling the agent to execute reliable queries and reasoning on real hospital FHIR servers.
Reinforcement Learning via Self-Distillation
Jonas Hübotter (ETH Zurich), Andreas Krause (ETH Zurich)
OptimizationComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: A reinforcement learning framework, RLRF, was studied, and a self-distillation strategy optimization (SDPO) was proposed to leverage the rich textual feedback provided by the environment, enabling dense credit assignment for large language models and achieving more efficient post-training;
Reinforcement Learning with Action-Triggered Observations
Alexander Ryabchenko (University of Toronto), Wenlong Mou (University of Toronto)
Reinforcement Learning
🎯 What it does: Proposed the Action-Triggered Sporadically Traceable MDP (ATST-MDP) framework, derived the corresponding Bellman equation, proved the existence of the optimal policy, constructed action sequence feature mapping under the linear MDP assumption, and then proposed the ATST-LSVI-UCB algorithm with a non-asymptotic upper bound on the expected return.
Reinforcement Learning with Discrete Diffusion Policies for Combinatorial Action Spaces
Haitong Ma (Google Research), Guy Tennenholtz (Google Research)
TransformerReinforcement LearningDiffusion modelContrastive LearningImageTextTabular
🎯 What it does: Propose a reinforcement learning framework called RL‑D2 for handling large combinatorial action spaces, using a discrete diffusion model as the policy, and generating a stable target distribution through policy mirror descent (PMD), then training the model using forward or backward KL matching.
Reinforcement Learning with Pairwise Preferences in Long-Term Decision Problems
Jonathan Colaço Carr, Benjamin Van Roy (Stanford University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabularSequential
🎯 What it does: Propose the Markov Decision Competition (MDC) model for pairwise preference reinforcement learning in long-term decision making; prove that stable Markov policies are optimal, exact solutions can be found in polynomial time, and design the Hedged Policy Iteration (HPI) algorithm;
Reinforcement Learning with Verifiable Rewards: GRPO's Loss, Dynamics, and Success Amplification
Youssef Mroueh (IBM Research)
OptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: This paper conducts a theoretical analysis of the GRPO algorithm under verifiable rewards, deriving its loss form, policy recurrence, and fixed points of success rate (PoS), and compares the dynamic characteristics of different reward normalization and KL regularization variants.
Reinforcement-aware Knowledge Distillation for LLM Reasoning
Zhaoyang Zhang (AWS Agentic AI), Stefano Soatto (AWS Agentic AI)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the RL‑aware Knowledge Distillation (RLAD) framework, integrating teacher guidance into the reasoning of LLMs after RL fine-tuning;
Reinforcing Real-world Service Agents: Balancing Utility and Cost in Task-oriented Dialogue
Ning Gao (Meituan), Chaozheng Wang (Meituan)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringGenerative Adversarial NetworkTextRetrieval-Augmented Generation
🎯 What it does: Proposed the InteractCS-RL framework, which redefines task-oriented dialogue as a multi-granularity reinforcement learning process, and achieves high-fidelity online training through a user-centric interaction framework.
ReJump: A Tree-Jump Representation for Analyzing and Improving LLM Reasoning
Yuchen Zeng (University of WisconsinMadison), Kangwook Lee (University of WisconsinMadison)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a representation method called ReJump based on Tree-Jump, used for structuring, quantitative analysis, and improvement of multi-step reasoning processes generated by large language models (LLMs);
Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases
Jun Yin (Hong Kong Polytechnic University), Chengqi Zhang (Hong Kong Polytechnic University)
ClassificationGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraphTabularBenchmark
🎯 What it does: Proposes a solution to the class imbalance problem in entity classification for relational databases, designing a relation-based minority class synthetic oversampling GNN (Rel-MOSS), while introducing a relational gate controller and a relation-guided minority synthesizer;
ReLAM: Learning Anticipation Model for Rewarding Visual Robotic Manipulation
Nan Tang (National Key Laboratory for Novel Software Technology Nanjing University), Yang Yu (National Key Laboratory for Novel Software Technology Nanjing University)
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelContrastive LearningImageVideo
🎯 What it does: Automatically learn a keypoint prediction model using video demonstrations without actions, generate intermediate subgoals, and design continuous rewards based on keypoint distances to train vision-driven robot manipulation policies.
Relational In-Context Learning via Synthetic Pre-training with Structural Prior
Yanbo Wang (Peking University), Muhan Zhang (Peking University)
ClassificationData SynthesisComputational EfficiencyMeta LearningTransformerPrompt EngineeringMixture of ExpertsGenerative Adversarial NetworkContrastive LearningTabularBenchmark
🎯 What it does: Built a relation database foundation model RDB‑PFN pre-trained only on synthetic data, and achieved zero-gradient context learning.
Relational Structural Causal Models
Adiba Ejaz (Columbia University), Elias Bareinboim (Columbia University)
Autonomous DrivingOptimizationExplainability and InterpretabilityGraph Neural NetworkReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsContrastive LearningGraphTabular
🎯 What it does: Propose the Relational Structural Causal Model (RSCM) and the Graph-based Relational Neural Causal Model (G-RNCM), and provide a method for determining the identifiability of graphical and symbolic representations across skeletons. Subsequently, verify their identifiability and performance in a simulated traffic scenario.
Relative Entropy Estimation in Function Space: Theory and Applications to Trajectory Inference
CHAO WANG, Pietro Michiardi (EURECOM)
OptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTabularTime SeriesBiomedical DataBenchmarkStochastic Differential Equation
🎯 What it does: Proposed a general framework for estimating KL divergence in function space, and applied it to the evaluation of trajectory inference
RelaxFlow: Text-Driven Amodal 3D Generation
Jiayin Zhu (National University Of Singapore), Angela Yao (National University Of Singapore)
RestorationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelFlow-based ModelTextPoint CloudMesh
🎯 What it does: Proposes a text-driven modality 3D generation task, completing unobserved regions under occlusion conditions through text prompts while strictly maintaining the visible parts unchanged.
RelayCaching: Accelerating LLM Collaboration via Decoding KV Cache Reuse
Yingsheng Geng (Beijing University of Posts and Telecommunications), Jiang liu
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a no-training inference method called RelayCaching, which directly reuses the KV cache from the decoding stage of the previous agent in multi-agent LLM collaboration to accelerate the subsequent prefill stage, significantly reducing redundant prefix computations;
Relevance-Based Embeddings: Lightweight Candidate Retrieval via Heavy-Ranker Calls
Kirill Sergeevich Shevkunov (Yandex), Liudmila Prokhorenkova (Yandex)
RetrievalRecommendation SystemTransformerAuto EncoderContrastive LearningTextMultimodality
🎯 What it does: Propose a lightweight candidate retrieval method (RBE) based on query/item relevance vectors, which approximates any complex similarity model by representing each query or item as a relevance vector with respect to a set of supporting items/queries, and directly using these representations in candidate retrieval.
Reliability-Aware LLM Alignment from Inconsistent Human Feedback
Jingyi Huang (Miami University), Yang Zhang (Miami University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes a Reliability-Guided Preference Optimization (RGPO) framework to alleviate noise and improve alignment of large language models (LLMs) in multi-annotator human feedback settings.
Reliable Confidence Alignment for Generalized Category Discovery
Jiawei Yu (National University of Defense Technology), Kele Xu (National University of Defense Technology)
ClassificationDomain AdaptationRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes a framework based on Reliable Confidence Alignment (RCA) for classifying and discovering mixed known and unknown categories in unlabeled data under sparse supervision.
Reliable Neighborhood-Aware Multi-View Outlier Detection
Huijie Ma (Harbin Institute of Technology), Guoqing Chao (Harbin Institute of Technology)
Anomaly DetectionAuto EncoderContrastive LearningMultimodality
🎯 What it does: Propose a reliable neighborhood-aware multi-view anomaly detection framework called RNAMOD, aiming to suppress anomaly propagation and eliminate perspective scale differences;
Reliable Thinking with Images
Haobin Li (Sichuan University), Xi Peng (Sichuan University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a reliable image thinking framework, RTWI, which eliminates noise thinking in multi-modal chain-of-thought reasoning by utilizing text-centric reliability estimation, two-stage filtering, and reliable voting.
Relighting as a Probe of Visual Priors via Augmented Latent Intrinsics
Xiaoyan Xing (University of Amsterdam), Anand Bhattad (Johns Hopkins University)
Image TranslationTransformerSupervised Fine-TuningDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes using generative image illumination transfer as a means to evaluate visual priors, and constructs the Augmented Latent Intrinsics (ALI) framework, which integrates a pre-trained visual encoder with latent intrinsic features to achieve high-quality image-to-image illumination transfer.
RELO: Reinforcement Learning to Localize for Visual Object Tracking
Xin Chen (City University of Hong Kong), Kede Ma (City University of Hong Kong)
Object TrackingReinforcement Learning from Human FeedbackRecurrent Neural NetworkTransformerReinforcement LearningAuto EncoderContrastive LearningOptical FlowImageVideo
🎯 What it does: This paper proposes a visual object tracking method called RELO based on reinforcement learning, which directly models target localization as a Markov decision process on feature maps, replacing traditional handcrafted spatial priors (center heatmap or corner distribution), and achieves localization through reward-driven policy learning;
ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference
Xiongwei Zhu (Beihang University), Limin Xiao (Beihang University)
Computational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposes the ReMoE framework, which fine-tunes the router of the MoE model to enhance expert reuse in memory-constrained scenarios, reducing the number of times experts are read from storage.
Remove the Ambiguity: Few-shot Multimodal Anomaly Detection Using Crossmodal Feature Replacers
Yuan Guo (Sichuan University), Xu Wang (Sichuan University)
Anomaly DetectionConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageMultimodalityPoint CloudRetrieval-Augmented Generation
🎯 What it does: Propose a self-supervised cross-modal feature replacement framework (CFR), which learns rough cross-modal reconstruction through cyclic mapping, and identifies and replaces unreliable features during inference to address the blurring reconstruction problem caused by the one-to-many mapping from RGB-3D, thus achieving few-shot multi-modal anomaly detection.
Removing Noise, not Finding Gold: Quality Filtering for Large-Scale Pretraining
Thiziri Nait Saada (University of Oxford), Pierre Ablin (Apple)
OptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper deeply analyzes the mechanism of classifier-based quality filtering (CQF) in large-scale pretraining and proposes an optimized 'data regulation' quality definition.
Removing Sandbagging in LLMs by Training with Weak Supervision
Emil Ryd (MATS), Vivek Hebbar (Redwood Research)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Train large language models under weak supervision, and study how to eliminate their sandbagging behavior.
ReNF: Rethinking the Design of Neural Long-Term Time Series Forecasters
Yihang Lu (HFIPS, Chinese Academy of Sciences), Enhong Chen (University of Science and Technology of China)
OptimizationComputational EfficiencyRecurrent Neural NetworkTransformerContrastive LearningTabularTime SeriesBenchmark
🎯 What it does: The paper proposes the Boosted Direct Output (BDO) framework, combining the AR structure with the DO method, using recursive sub-predictions and achieving implicit prediction combination within the network, while employing EMA to smooth the training process to address the validation-test generalization gap;
Rényi Diffusion Models
Yirong Shen (Imperial College London), Cong Ling (Imperial College London)
GenerationDiffusion modelScore-based ModelImage
🎯 What it does: A new generative framework, R'enyi diffusion model, is proposed, which improves the balance between sample quality and distribution coverage by formulating the training objective using R'enyi divergence.
RePack then Refine: Efficient Diffusion Transformers with Vision Foundation Models
Guanfang Dong (Huawei Technologies Canada Ltd), Chao Gao (Huawei Technologies Canada Ltd)
GenerationTransformerDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose the RePack-Refine framework, which compresses Vision Foundation Model (VFM) features and uses them as input for the Diffusion Transformer (DiT), and adds a detail enhancement Refiner after generation.
Reparameterization Flow Policy Optimization
Hai Zhong (Tsinghua University), Longbo Huang (Tsinghua University)
OptimizationRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelImageVideoPoint CloudTabularTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a new reinforcement learning framework based on flow models, called Reparameterization Flow Policy Optimization (RFO), which achieves efficient policy gradient updates by jointly backpropagating the flow generation process with environmental dynamics;
Reparameterization Proximal Policy Optimization
Hai Zhong (Tsinghua University), Longbo Huang (Tsinghua University)
Reinforcement LearningPoint CloudMeshTabularTime SeriesSequential
🎯 What it does: Proposed a reparameterized policy gradient algorithm called RPO that combines differentiable simulation with Proximal Policy Optimization.
RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress
Ruixuan Huang (Hong Kong University of Science and Technology), Wei Wang (Hong Kong University of Science and Technology)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: Proposed and verified RepetitionCurse, a black-box attack method that exploits repeated tokens to induce routing imbalance in Mixture-of-Experts (MoE) LLMs, and conducted systematic evaluation on 139 mainstream MoE models.
rePIRL: Learn PRM with Inverse RL for LLM Reasoning
Xian Wu (Meta AI), Wenbo Guo (University of California, Santa Barbara)
OptimizationComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Learn and train a process reward model (PRM) to enhance the performance of large language models (LLMs) in reasoning tasks, proposing the rePIRL framework that alternates between inverse reinforcement learning and policy updates for bidirectional learning.
Replay Failures as Successes: Sample-Efficient Reinforcement Learning for Instruction Following
Kongcheng Zhang (Zhejiang University), Mingli Song (Zhejiang University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the HiR framework, which achieves efficient RL training using only binary rewards by employing a hindsight replay approach of 'select-then-rewrite' that rewrites failure samples into successful ones.
RePo: Language Models with Context Re-Positioning
Huayang Li (Nara Institute of Science and Technology), Richard Sproat (Sakana AI)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a context-aware repositioning mechanism called REPO, which allows LLMs to assign context-related continuous positions to each token through a differentiable module within self-attention, thereby reducing the external burden on the attention layer.
Representation Drift Compensation: A Near-Zero Inference Cost Enhancement for LLM Decomposition
Xinhao Huang (Hong Kong University of Science and Technology), Zeyi Wen (Hong Kong University of Science and Technology)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper addresses the representation drift problem caused by low-rank decomposition, proposing the Decomper mechanism, which effectively suppresses representation drift and restores model performance by learning compensation vectors without increasing inference costs.
Representation Learning for Equivariant Inference with Guarantees
Daniel Ordonez-Apraez, Massimiliano Pontil (Italian Institute of Technology)
Representation LearningRobotic IntelligenceConvolutional Neural NetworkGraph Neural NetworkSpiking Neural NetworkAuto EncoderContrastive LearningGaussian SplattingPoint CloudGraphTabular
🎯 What it does: The eNCP framework is proposed by combining symmetry priors with spectral contrastive learning for conditional probability estimation, aiming to learn representations with invariant/equivariant properties, thereby achieving efficient regression, conditional probability estimation, and uncertainty quantification.
Representation Unlearning: Forgetting through Information Compression
Antonio Almudévar (University of Zaragoza), Alfonso Ortega (University of Zaragoza)
ClassificationSafty and PrivacyComputational EfficiencyRepresentation LearningConvolutional Neural NetworkAuto EncoderContrastive LearningImage
🎯 What it does: Propose a framework for machine forgetting in the representation space—Representation Unlearning, which learns a lightweight transformation to compress representations, removing information about the forgotten data while preserving the performance on the retained data;
Representational Curvature Modulates Behavioral Uncertainty in Large Language Models
Jack G. King (Massachusetts Institute Of Technology), Eghbal A. Hosseini (Massachusetts Institute Of Technology)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This study investigates the relationship between the curvature of internal representations in autoregressive LLMs and the uncertainty in predicting the next token. It analyzes the LAMBADA and UD short-sentence datasets using two models (GPT-2 XL and Pythia-2.8B), and validates the causal effect through experiments involving trajectory-aligned perturbations and curvature regularization on internal activations.
Representational Similarity and Model Behavior in Multi-Agent Interaction
Yujin Potter (University of California Berkeley), Alane Suhr (University of California Berkeley)
Explainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextSequential
🎯 What it does: Investigate the representation similarity of large language models (LLMs) in multi-agent interactions, and explore its impact on cooperative and innovative behaviors;
RePro: Training Language Models to Faithfully Recycle the Web for Pretraining
Zichun Yu (Carnegie Mellon University), Chenyan Xiong (Carnegie Mellon University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose the REPRO method, which uses reinforcement learning to train a small language model to perform high-quality restatements of web data, thereby improving the quality of pre-training data.
ReQAT: Achieving Full-Precision Reasoning Accuracy with 4-bit Floating-Point Quantization-Aware Training
Janghwan Lee (Hanyang University), Jungwook Choi (Hanyang University)
Computational EfficiencyKnowledge DistillationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: The ReQAT framework is proposed by introducing rationale-guided strategies in 4-bit FP quantization training to achieve W4A4KV4 deployment while maintaining or exceeding full-precision performance.
Required Spine Optional Limbs: Heterogeneous Federated Learning via Backbone-sharing and Activation-guided Selection
Mingsheng Cao (University of Electronic Science and Technology of China), Junlong Zhou (Nanjing University of Science and Technology)
ClassificationFederated LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes the SpineFL framework, achieving federated learning on heterogeneous devices through backbone sharing and activation-guided submodel extraction.
Reranker Helps, but Not Enough: Towards Strong Poisoning Attacks Against Retrieval-Augmented Generation
Xiaokun Yang (Chinese Academy of Sciences), Tieniu Tan (Chinese Academy of Sciences)
RetrievalAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper designs and evaluates an attack framework named Prompt‑Perturbation Poisoning Attack (P A 3), which can inject malicious documents into retrieval-augmented generation (RAG) systems containing a reranker and induce the model to output answers specified by the attacker.
ReSeek: A Self-Correcting Framework for Search Agents with Instructive Rewards
Shiyu Li, Yang Tang (Tencent)
Recommendation SystemAutonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the ReSeek framework, training a search agent to self-correct erroneous paths during execution.
Residual Context Diffusion Language Models
Yuezhou Hu (University of California Berkeley), Chenfeng Xu (University of California Berkeley)
GenerationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelMixture of ExpertsDiffusion modelContrastive LearningText
🎯 What it does: Propose Residual Context Diffusion (RCD), a method that utilizes the representations of discarded low-confidence tokens in diffusion large language models as residual context, which is re-injected in subsequent denoising steps to improve generation quality.
Residual-Guided Multi-Resolution Refinement of Foundation Models: A Case Study in Drought Forecasting
Wentao Gao (Adelaide University), Yun Chen (CSIRO Environment)
Anomaly DetectionOptimizationData-Centric LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningTabularTime SeriesAgriculture Related
🎯 What it does: Proposed the RGMR (Residual-Guided Multi-Resolution Refinement) framework, which performs multi-scale residual correction on frozen time series base models during inference to improve regional climate (drought) prediction.
Resilient Coresets and Clustering
Ashkan Norouzi-Fard (Google Research), Morteza Monemizadeh (Eindhoven University of Technology)
OptimizationContrastive Learning
🎯 What it does: Proposed the concept of γ-resilient coresets and presented randomized construction algorithms for k-median/k-means, ensuring that the coreset maintains a 1±ε approximation error and assignment consistency even when the input point set undergoes small geometric perturbations.
Resolution as a Direction: Vector-Panning Feature Alignment for Cross-Resolution Re-Identification
Zanwu Liu (Beihang University), Guanglin Niu (Beihang University)
Super ResolutionRetrievalConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose a lightweight post-processing module called VPFA, which aligns low-resolution features with high-resolution features by 'vector shifting' them, thus addressing the feature misalignment problem in cross-resolution person re-identification.
Resolving Blind Inverse Problems under Dynamic Range Compression via Structured Forward Operator Modeling
Muyu Liu (ShanghaiTech University), Yuyao Zhang (ShanghaiTech University)
RestorationDiffusion modelScore-based ModelImage
🎯 What it does: Proposed a zero-shot framework called CaMB-Diff, which utilizes cascaded monotonic Bernstein polynomials to impose hard constraints on the forward model with unknown dynamic range compression, and combines it with diffusion models for joint optimization, restoring high-quality compressed images.
Resolving the Timestep Scaling Paradox in Spiking Neural Networks with a Timestep-Scalable Neuron Model
Binghao Ye (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Kun Shang (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
ClassificationObject DetectionOptimizationComputational EfficiencySpiking Neural NetworkSupervised Fine-TuningAuto EncoderContrastive LearningImageVideoTime SeriesSequential
🎯 What it does: This paper proposes a time-step scalable synapse neuron model called TS-LIF, aimed at addressing the performance degradation of traditional LIF neurons when increasing the time step.
Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement
Yunjian Zhang (University of Chinese Academic of Sciences), Yao Zhu (Zhejiang University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Achieve resource-efficient RLVR training through Dynamic One-shot Policy Refinement (DoPR), significantly reducing sample and computational demands.
Respecting Modality Gap in Post-hoc Out-of-distribution Detection with Pre-trained Vision-Language Models
Yuanwei Hu (University of Technology Sydney), Jie Lu (University of Technology Sydney)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper studies the use of pre-trained vision-language models (such as CLIP) for zero-shot post-hoc OOD detection, challenges the common practice of using text as class prototypes, and proposes an online pseudo-supervised prototype learning framework.
ReSpinQuant: Efficient Layer-Wise LLM Quantization via Subspace Residual Rotation Approximation
Suyoung Kim (Seoul National University), Nojun Kwak (Seoul National University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelAuto EncoderContrastive LearningGaussian SplattingText
🎯 What it does: For post-training quantization of large language models, the ReSpinQuant framework is proposed, which achieves efficient low-bit inference by offline fusion of hierarchical rotation and utilizing subspace residual rotation approximation.
Responsible Text-to-Image Diffusion: Interpretable and Linearly Controllable Semantics for Fair and Safe Generation
Sayedmoslem Shokrolahi (Queen's University), Il-Min Kim (Queen's University)
GenerationSafty and PrivacyExplainability and InterpretabilityTransformerPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: This work proposes an architecture-agnostic framework that injects interpretable and linearly controllable concept vectors into frozen text-to-image diffusion models, enabling fair and safe image generation;
ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning
Zihan Lin (Chinese Academy of Sciences), Guojun Yin (Meituan)
Computational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextSequentialBenchmark
🎯 What it does: To enhance the reasoning ability of large language models (LLMs), we propose Negative Sample Projection Residual Reinforcement Learning (ResRL), which projects negative sample hidden representations onto the low-rank subspace of positive samples and reweights the negative sample gradients using projection residuals, thereby suppressing gradient conflicts and maintaining generation diversity.
REST: Diffusion-based Real-time End-to-end Streaming Talking Head Generation via ID-Context Caching and Asynchronous Streaming Distillation
Haotian Wang (University of Science and Technology of China), Qingfeng Liu (University of Science and Technology of China)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelAuto EncoderContrastive LearningVideoAudio
🎯 What it does: Propose the REST framework to achieve real-time, end-to-end streaming speaker head video generation based on diffusion models.
Resting Neurons, Active Insights: Robustifying Activation Sparsity in LLMs via Spontaneity
Haotian Xu (Stony Brook University), Tengfei Ma (Stony Brook University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose injecting spontaneous neurons into activation sparse inference of large language models to maintain representation stability and enhance inference performance.
Restoring Exploration after Post-Training: Latent Exploration Decoding for Large Reasoning Models
Wenhui Tan (Renmin University of China), Jian Luan (Xiaomi Inc)
OptimizationAI Code AssistantTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose Latent Exploration Decoding (LED), which restores the exploration capability of large inference models after RL training by aggregating intermediate layer hidden state outputs and dynamically selecting the maximum entropy depth during the DeepThink phase.
Restoring Initial Noise Sensitivity in Text-to-Image Distillation through Geometric Alignment
Huayang Huang (Wuhan University), Ye Zhu (Ecole Polytechnique)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageText
🎯 What it does: Propose Geometry-Aware Distillation (GAD), which restores the initial noise sensitivity in the text-to-image distillation process by matching the local geometric responses of the teacher and student models to input noise.
ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning
Xiaofeng Lin (University of California, Los Angeles), Guang Cheng (University of California, Los Angeles)
Data SynthesisOptimizationSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelContrastive LearningTabularBenchmark
🎯 What it does: Propose a table data generation framework based on reinforcement learning called ReTabSyn, which directly optimizes the generator's conditional distribution using unsupervised preference comparison, enabling the generated data to maintain higher downstream model performance in scenarios with low sample sizes, extreme imbalance, and distribution drift.