ICML 2025 Papers — Page 25
International Conference on Machine Learning · 3257 papers
ReFrame: Layer Caching for Accelerated Inference in Real-Time Rendering
Lufei Liu (University of British Columbia), Tor M. Aamodt (University of British Columbia)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a layer caching technique called ReFrame, which leverages the temporal consistency between frames in real-time rendering to cache intermediate layer features of encoder-decoder networks such as U-Net or U-Net++, avoiding redundant computations for each frame and thus accelerating neural network inference.
REG: Rectified Gradient Guidance for Conditional Diffusion Models
Zhengqi Gao (Massachusetts Institute of Technology), Duane S Boning
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper reveals the contradiction between the theoretical roots and practical implementation of existing conditional diffusion model guidance methods by reconstructing the theoretical framework of joint distribution scale. Based on this framework, a new guidance strategy called Rectified Gradient Guidance (REG) is proposed to enhance the generation quality of existing guidance methods.
Regress, Don't Guess: A Regression-like Loss on Number Tokens for Language Models
Jonas Zausinger (Technical University of Munich), Jannis Born (IBM Research Europe)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed and implemented Number Token Loss (NTL), allowing the language model to perform regression loss directly on numeric tokens during training, rather than using the traditional cross-entropy.
Regression for the Mean: Auto-Evaluation and Inference with Few Labels through Post-hoc Regression
Benjamin Eyre (Columbia University), David Madras (Google DeepMind)
Tabular
🎯 What it does: This study investigates the use of the Prediction Powered Inference (PPI) framework for mean estimation in scenarios with very few labels, and proposes regression-based improvements (Ridge-PPI and Sigmoid-PPI) to reduce estimation variance.
Regret-Free Reinforcement Learning for Temporal Logic Specifications
R Majumdar, Sadegh Soudjani (University of Birmingham)
OptimizationReinforcement LearningSequential
🎯 What it does: A no-regret reinforcement learning algorithm for LTL (Linear Temporal Logic) specifications of unknown dynamic systems is proposed, which can learn control policies online and ensure convergence to the optimal satisfaction probability.
Regularized Langevin Dynamics for Combinatorial Optimization
Shengyu Feng (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Proposes Regularized Langevin Dynamics (RLD) and implements two CO solvers based on it: RLSA (simulated annealing version) and RLNN (neural network version).
Reidentify: Context-Aware Identity Generation for Contextual Multi-Agent Reinforcement Learning
Zhiwei Xu (Shandong University), Jiangjin Yin (Huazhong Agricultural University)
TransformerReinforcement LearningSequential
🎯 What it does: This study proposes the Context-Aware Identity Generation (CAID) framework, which utilizes Transformers to dynamically generate agent identities and context encoding, enhancing the performance of Contextual Multi-Agent Reinforcement Learning (CMARL).
ReinboT: Amplifying Robot Visual-Language Manipulation with Reinforcement Learning
Hongyin Zhang (Zhejiang University), Donglin Wang (Westlake University)
Robotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelMultimodality
🎯 What it does: This paper proposes an end-to-end visual-language-action model named ReinboT, which utilizes the principle of cumulative reward maximization in reinforcement learning to enhance the decision-making quality of robots in long-horizon operation tasks.
REINFORCE Adversarial Attacks on Large Language Models: An Adaptive, Distributional, and Semantic Objective
Simon Geisler (Technical University of Munich), Stephan Günnemann
OptimizationAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: An adaptive, distributed, and semantic attack target based on REINFORCE is proposed to automate the construction of jailbreak prompts that can bypass LLM alignment.
Reinforce LLM Reasoning through Multi-Agent Reflection
Yurun Yuan (University of Wisconsin Madison), Tengyang Xie (University of Wisconsin Madison)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A reinforcement learning framework based on multi-agent feedback, DPSDP, is proposed to train LLMs to improve answer quality in reasoning tasks through iterative reflection and correction.
Reinforced Learning Explicit Circuit Representations for Quantum State Characterization from Local Measurements
Manwen Liao (University of Hong Kong), Yuxiang Yang (University of Hong Kong)
TransformerReinforcement LearningSequentialPhysics Related
🎯 What it does: This paper proposes an explicit circuit representation method based on reinforcement learning, which can learn approximate preparation circuits of quantum states using only local measurement data, and achieve accurate prediction and experimental reconstruction of quantum state properties.
Reinforced Lifelong Editing for Language Models
Zherui Li (Beijing University of Posts and Telecommunications), Xiang Wang (University of Science and Technology of China)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the RLEdit method, which models the lifelong editing problem as a reinforcement learning task, utilizing a supernetwork for offline training across the entire knowledge sequence to achieve efficient and stable continuous knowledge updates.
Reinforcement Learning Control of a Physical Robot Device for Assisted Human Walking without a Simulator
Junmin Zhong (Arizona State University), Jennie Si (Arizona State University)
Robotic IntelligenceReinforcement LearningTime Series
🎯 What it does: Combining offline imitation of expert strategies with online actor-critic reinforcement learning on a real soft exoskeleton device, personalized gait assistance control is achieved, significantly reducing the muscle load on users.
Reinforcement Learning for Quantum Control under Physical Constraints
Jan Ole Ernst (University of Oxford), Axel Kuhn (University of Oxford)
OptimizationReinforcement LearningPhysics Related
🎯 What it does: Using physics-constrained reinforcement learning, a control method for open quantum systems is proposed, capable of generating high-fidelity, realizable pulse sequences under noise and experimental constraints.
Reinforcement Learning with Adaptive Reward Modeling for Expensive-to-Evaluate Systems
Hongyuan Su (Tsinghua University), Yong Li (Tsinghua University)
OptimizationDrug DiscoveryGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: This paper proposes AdaReMo, which constructs a fast reward model for reinforcement learning systems with high reward evaluation costs, achieving decoupling of online decision-making and offline evaluation.
Reinforcement Learning with Random Time Horizons
Enric Ribera Borrell (Zuse Institute Berlin), Christof Schuette
Reinforcement Learning
🎯 What it does: This study investigates the case of random time termination within the reinforcement learning framework, deriving two policy gradient formulas for trajectory and state space under random deadlines, and implementing a model-based gradient for deterministic policies in discrete time.
Reinforcement Learning with Segment Feedback
Yihan Du (University of Illinois at Urbana-Champaign), R. Srikant (University of Illinois at Urbana-Champaign)
Reinforcement LearningTabular
🎯 What it does: This paper proposes the RL with segment feedback model, studying two types of segmented reward settings: binary feedback and summation feedback, designing efficient algorithms and providing theoretical upper and lower bounds.
Rejecting Hallucinated State Targets during Planning
Harry Zhao, Yoshua Bengio (University of Montreal)
Reinforcement Learning
🎯 What it does: This paper proposes an evaluator that can learn to assess the feasibility of goals and reject the evaluation of hallucinated state goals without altering the original planning agent.
Relating Misfit to Gain in Weak-to-Strong Generalization Beyond the Squared Loss
Abhijeet Mulgund (University of Illinois at Chicago), Chirag Pabbaraju (Stanford University)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper studies the 'weak-to-strong generalization' in training strong models with weakly supervised models, exploring the quantitative relationship between the misfit of strong and weak models and the final performance gain, and generalizing previous conclusions that were only applicable to squared loss to any Bregman divergence, particularly cross-entropy/KL divergence.
Relational Conformal Prediction for Correlated Time Series
Andrea Cini (Universita della Svizzera italiana), Filippo Maria Bianchi (UiT The Arctic University of Norway)
Graph Neural NetworkTime Series
🎯 What it does: This paper proposes an adaptive confidence interval generation method called CORE, based on graph deep learning, which combines graph neural networks with hierarchical quantile regression. It can automatically learn sparse relationships between time series on the basis of post-calibration and estimate the uncertainty of correlated sequences.
Relational Invariant Learning for Robust Solvation Free Energy Prediction
yeyunchen
Graph Neural NetworkAuto EncoderTabular
🎯 What it does: Proposed the RILOOD framework to achieve OOD generalization for solvent free energy prediction.
Relative Error Fair Clustering in the Weak-Strong Oracle Model
Vladimir Braverman (Google Research), Samson Zhou (Texas A&M University)
Tabular
🎯 What it does: Under the weak-strong oracle model, a coreset construction algorithm for (1+ε) approximate k-median (and k-z) clustering with relative error fairness is proposed, along with a coreset for fair k-median.
RelGNN: Composite Message Passing for Relational Deep Learning
Tianlang Chen (Stanford University), Jure Leskovec (Stanford University)
Recommendation SystemGraph Neural NetworkGraphBenchmark
🎯 What it does: A graph neural network framework specifically designed for relational database graphs, RELGNN, has been constructed, utilizing atomic routing to achieve composite information transmission.
Reliable Algorithm Selection for Machine Learning-Guided Design
Clara Fannjiang (Genentech), Ji Won Park
OptimizationProtein Structure PredictionTabular
🎯 What it does: A design algorithm configuration selection method based on predictive empowered inference and density ratio weighting is proposed, which can ensure a high probability of success within a multiple testing framework;
Reliable and Efficient Amortized Model-based Evaluation
Sang T. Truong, Sanmi Koyejo (Virtue AI)
Large Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a reliable and efficient evaluation framework based on Item Response Theory (IRT) for the assessment of large-scale language models. Based on this framework, an amortized difficulty predictor and a conditional question generator are developed to reduce calibration costs and build a diverse question bank.
Rényi Neural Processes
Xuesong Wang (CSIRO Data61), Edwin V. Bonilla (CSIRO Data61)
Domain AdaptationOptimizationImageTime Series
🎯 What it does: This paper proposes the R'enyi Neural Processes (RNP) framework, which utilizes the R'enyi divergence instead of the traditional KL divergence to alleviate the performance degradation caused by prior mis-specification in Neural Processes, and unifies the two objectives of variational inference and maximum likelihood.
RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers
Xuwei Xu (University of Queensland), Sen Wang (University of Queensland)
ClassificationObject DetectionSegmentationComputational EfficiencyTransformerSupervised Fine-TuningImage
🎯 What it does: A channel idle mechanism is proposed, allowing the FFN layer of ViT to achieve significant acceleration during inference through structural reparameterization.
RepLoRA: Reparameterizing Low-rank Adaptation via the Perspective of Mixture of Experts
Tuan Truong (Qualcomm), Nhat Ho (University of Texas at Austin)
TransformerSupervised Fine-TuningMixture of ExpertsImageVideoTextMultimodality
🎯 What it does: By treating LoRA as a Mixture of Experts model and utilizing low-rank reparameterization (using lightweight MLP) to enhance the estimation efficiency of low-rank matrices, RepLoRA is proposed and its effectiveness is validated in multi-task scenarios.
RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing
Jinyao Guo (Purdue University), Xiangyu Zhang (Purdue University)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: REPOAUDIT is proposed, an autonomous agent based on LLM for repository-level code auditing, capable of detecting defects through function-level path-sensitive analysis without compilation.
Representation Preserving Multiclass Agnostic to Realizable Reduction
Steve Hanneke (Purdue University), Amirreza Shaeiri (Purdue University)
ClassificationRepresentation Learning
🎯 What it does: This paper studies the multi-class classification problem within the PAC learning framework and proposes a simple and elegant theory of dimensionality reduction from unachievability to achievability, addressing an open question posed by Hopkins et al.
Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing
Kento Nishi (Harvard University), Ekdeep Singh Lubana (Harvard University)
Representation LearningTransformerLarge Language ModelGraph
🎯 What it does: This paper explores the impact of Knowledge Editing (KE) on the internal representations of Transformer models through the construction of a structured knowledge graph synthesis task, proposing the hypothesis of 'representation shattering';
Representation Surgery in Model Merging with Probabilistic Modeling
Qi Wei (Nanyang Technological University), Bo An (Skywork AI)
Domain AdaptationRepresentation LearningTransformerSupervised Fine-TuningImageText
🎯 What it does: A probabilistic representation correction module called ProbSurgery is proposed to correct representation biases that arise after model merging;
Representations Shape Weak-to-Strong Generalization: Theoretical Insights and Empirical Predictions
Yihao Xue (University of California), Baharan Mirzasoleiman (University of California)
Representation LearningTransformerSupervised Fine-TuningText
🎯 What it does: This paper characterizes the mechanism of Weak-to-Strong Generalization (W2SG) based on a theoretical framework of internal representations and proposes a metric to predict W2SG performance without the need for labels.
Representative Language Generation
Charlotte Peale (Stanford University), Omer Reingold (Stanford University)
GenerationText
🎯 What it does: The concept of representative generation is proposed, extending the generative theoretical framework introduced by Kleinberg et al. (2024) and Li et al. (2024), aimed at addressing the issues of diversity and bias in generative models.
Representative Ranking for Deliberation in the Public Sphere
Manon Revel (Meta), Maximilian Nickel
Text
🎯 What it does: This paper proposes the inclusion of a 'Justified Representation' constraint in the ranking of online reviews to ensure that while improving the quality of dialogue, diverse perspectives are adequately reflected.
ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation
Angxiao Yue (Renmin University of China), Hongteng Xu (Renmin University of China)
GenerationProtein Structure PredictionFlow-based ModelRectified FlowBiomedical Data
🎯 What it does: This paper proposes the ReQFlow model based on quaternion flow matching, aimed at efficiently generating high-quality protein backbone structures.
ResearchTown: Simulator of Human Research Community
Haofei Yu (University of Illinois Urbana-Champaign), Jiaxuan You (University of Illinois Urbana-Champaign)
Graph Neural NetworkLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper presents RESEARCHTOWN, a multi-agent LLM framework based on TextGNN, designed to simulate collaborative activities such as reading, writing, and reviewing papers in human research communities.
Residual Matrix Transformers: Scaling the Size of the Residual Stream
Brian Mak (University of California Santa Cruz), Jeffrey Flanigan (University of California Santa Cruz)
TransformerText
🎯 What it does: A new variant of Transformer called Residual Matrix Transformer (RMT) is proposed, which replaces the residual flow with an outer product memory matrix to achieve scalable residual flow.
Residual TPP: A Unified Lightweight Approach for Event Stream Data Analysis
Ruoxin Yuan (Fudan University), Guanhua Fang (Fudan University)
Anomaly DetectionComputational EfficiencyRecurrent Neural NetworkTransformerTabularTime SeriesSequentialBenchmarkOrdinary Differential Equation
🎯 What it does: This paper proposes Residual TPP, a lightweight hybrid model that integrates traditional Hawkes processes with neural TPP through Residual Event Decomposition (RED), capable of capturing the periodicity, self-excitement, and the difficult-to-interpret residual components of event streams.
ResKoopNet: Learning Koopman Representations for Complex Dynamics with Spectral Residuals
Yuanchao Xu (University of Alberta), Zhongwei Shen (University of Alberta)
Time SeriesBiomedical DataPhysics Related
🎯 What it does: Proposes ResKoopNet, a new method for learning Koopman representations by minimizing spectral residuals;
Resolving Lexical Bias in Model Editing
Hammad Rizwan (Dalhousie University), Hassan Sajjad (Dalhousie University)
TransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper addresses the editing task of pre-trained language models, identifying and solving the issue of erroneous triggers caused by vocabulary bias, and proposes the PENME editing framework based on projection networks.
ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals
Utkarsh Saxena (Purdue University), Xin Wang (d-Matrix)
TransformerLarge Language ModelTextMultimodality
🎯 What it does: ResQ achieves efficient post-training quantization by performing low-rank PCA projection on weights, activations, and KV caches, retaining high-variance subspaces at 8-bit precision while quantizing the remaining parts to 4-bit.
RestoreGrad: Signal Restoration Using Conditional Denoising Diffusion Models with Jointly Learned Prior
Ching-Hua Lee (Samsung Electronics), Hongxia Jin (Samsung Electronics)
RestorationDiffusion modelAuto EncoderImageAudio
🎯 What it does: The RestoreGrad framework is proposed, which achieves the recovery of speech and image signals through joint learning of prior distributions and conditional diffusion models (DDPM).
Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach
Jiancong Xiao (University of Pennsylvania), Li Shen (University of Pennsylvania)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper investigates the reasons for the calibration drop of large models after preference alignment and proposes a method to restore calibration through calibration-aware fine-tuning.
Rethink GraphODE Generalization within Coupled Dynamical System
Guancheng Wan (University of California), Wei Wang (University of California)
Graph Neural NetworkGraphTime SeriesOrdinary Differential Equation
🎯 What it does: A framework named GREAT is proposed, aimed at enhancing the generalization ability of GraphODE in coupled dynamical systems, addressing the issues of mixed static attributes and dynamic states, as well as generalization degradation caused by context-specific coupling patterns.
Rethink the Role of Deep Learning towards Large-scale Quantum Systems
Yusheng Zhao (University of Science and Technology of China), Yuxuan Du (Nanyang Technological University)
Convolutional Neural NetworkTabularPhysics Related
🎯 What it does: This paper conducts experiments on the estimation of ground state properties and phase classification tasks of large-scale quantum systems (up to 127 qubits) through unified quantum resources (total measurement times), systematically evaluating and comparing the performance of traditional machine learning (ML) models with deep learning (DL) models.
Rethinking Addressing in Language Models via Contextualized Equivariant Positional Encoding
Jiajun Zhu (University of Texas at Austin), Zhangyang Wang (Georgia Tech)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Contextualized Equivariant Positional Encoding (TAPE) is proposed, which enhances the model's representation of positional information and long sequence reasoning ability by gradually contextualizing positional encoding with the sequence content between Transformer layers.
Rethinking Aleatoric and Epistemic Uncertainty
Freddie Bickford Smith (University of Oxford), Tom Rainforth (University of Oxford)
Image
🎯 What it does: A framework based on decision theory is proposed to unify and clarify the concepts of uncertainty in machine learning, reinterpret the traditional aleatoric and epistemic decomposition, elucidate their relationship with predictive performance and data dispersion, and provide a theoretical assessment of the commonly used information measure BALD.
Rethinking Benign Overfitting in Two-Layer Neural Networks
Ruichen Xu (Chinese University of Hong Kong), Kexin Chen (Chinese University of Hong Kong)
Convolutional Neural NetworkImage
🎯 What it does: An improved feature-noise data model is proposed, which incorporates category-related heterogeneous noise based on the original assumption of uniform noise. Under this model, a theoretical analysis of the 'benign overfitting' phenomenon of two-layer ReLU CNNs on long-tailed data distributions is conducted, explaining why modern networks can achieve better generalization by memorizing noise under over-parameterization.
Rethinking Causal Ranking: A Balanced Perspective on Uplift Model Evaluation
Minqin Zhu (Zhejiang University), Kun Kuang (Zhejiang University)
Recommendation SystemOptimizationTabular
🎯 What it does: This paper proposes the Principled Uplift Curve (PUC) for a fairer and unbiased evaluation of uplift models, and based on this curve, introduces the Principled Treatment and Outcome Network (PTONet) uplift model;
Rethinking Chain-of-Thought from the Perspective of Self-Training
Zongqian Wu (University of Electronic Science and Technology of China), Lei Feng (Southeast University)
Large Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Revisiting Chain-of-Thought reasoning from the perspective of self-training, a new CoT framework is proposed, consisting of task-specific prompts and adaptive reasoning iterations, aimed at improving the accuracy of large language models in complex reasoning tasks.
Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL
Harit Vishwakarma (University of Wisconsin-Madison), Frederic Sala
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A pseudo-labeling confidence learning framework based on error tolerance is designed, replacing traditional threshold and confidence selection;
Rethinking External Slow-Thinking: From Snowball Errors to Probability of Correct Reasoning
Zeyu Gan (Renmin University of China), Yong Liu (Renmin University of China)
Reinforcement LearningText
🎯 What it does: This paper discusses the external slow thinking (test-time scaling) mechanism, establishing a theoretical connection between avalanche errors and correct reasoning probabilities, proving that expanding the search space can reduce errors.
Rethinking Latent Redundancy in Behavior Cloning: An Information Bottleneck Approach for Robot Manipulation
Shuanghao Bai (Xi'an Jiaotong University), Badong Chen (Xi'an Jiaotong University)
Robotic IntelligenceTransformerContrastive LearningMultimodality
🎯 What it does: This paper studies the potential representation redundancy problem in behavior cloning (BC) models for robotic manipulation and proposes compressing redundant information through the information bottleneck (IB) principle to enhance generalization performance.
Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation
Jian Bi (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
ClassificationSegmentationPoint Cloud
🎯 What it does: The SinPoint method is proposed, which utilizes sine transformations and homotopy to maintain the topological consistency of point clouds for data augmentation.
Rethinking Score Distilling Sampling for 3D Editing and Generation
Xingyu Miao (Durham University), Jungong Han (Tsinghua University)
GenerationData SynthesisKnowledge DistillationScore-based ModelGaussian SplattingPoint CloudMesh
🎯 What it does: A new method called Unified Distillation Sampling (UDS) is proposed, aimed at simultaneously supporting the generation and editing of 3D assets.
Rethinking the Bias of Foundation Model under Long-tailed Distribution
Jiahao Chen (Renmin University of China), Bing Su (Renmin University of China)
ClassificationTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper addresses the bias problem of Foundation Models under long-tail distributions, exploring the imbalance between pre-training data and downstream data that leads to two types of biases in fine-tuning (parameter bias and data bias). It proposes a method to alleviate both types of biases simultaneously by integrating multiple foundation models through backdoor adjustment in causal graphs, thereby improving long-tail visual classification performance.
Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective
Aojun Lu (Sichuan University), Yanan Sun (Sichuan University)
Knowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes the Dual-Arch framework, which utilizes two networks with different architectures (a wide shallow stable network and a deep thin flexible network) to jointly complete continual learning tasks, leveraging knowledge distillation for new knowledge transfer.
Rethinking the Temperature for Federated Heterogeneous Distillation
Fan Qi (Tianjin University of Technology), Changsheng Xu (Chinese Academy of Sciences)
Federated LearningKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: In federated learning, a multi-level knowledge distillation is achieved through adaptive temperature adjustment, constructing the ReT-FHD framework.
Rethinking Time Encoding via Learnable Transformation Functions
Xi Chen (Fudan University), Yun Xiong (Fudan University)
Anomaly DetectionTransformerTime SeriesFinance Related
🎯 What it does: A learnable universal time encoding method, LeTE, is proposed to better capture complex and diverse temporal patterns and can be directly embedded into various temporal and dynamic graph models.
Retraining with Predicted Hard Labels Provably Increases Model Accuracy
Rudrajit Das (Google Research), Peilin Zhong (Google Research)
ClassificationOptimizationSafty and PrivacyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: The study shows that retraining with hard labels predicted by the model itself under label noise can improve accuracy, and it is theoretically proven to be effective in linear separable binary classification tasks.
Retraining-free Merging of Sparse MoE via Hierarchical Clustering
I-Chun Chen (National Tsing Hua University), Chun-Yi Lee (National Taiwan University)
CompressionMixture of ExpertsText
🎯 What it does: This paper proposes HC-SMoE, a hierarchical clustering expert merging framework that does not require retraining and is task-agnostic, aimed at compressing sparse mixture of experts models.
Retrieval Augmented Time Series Forecasting
Sungwon Han (Gwangju Institute of Science and Technology), Jinsung Yoon (Google Cloud AI)
RetrievalAnomaly DetectionTime SeriesRetrieval-Augmented Generation
🎯 What it does: A retrieval-enhanced time series forecasting method called RAFT is proposed, which retrieves historical segments from the training set that are similar to the current input and uses their subsequent values in conjunction with the model input to predict the future.
Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate
Jiahe Du (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
GenerationRetrievalDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelGraphRetrieval-Augmented Generation
🎯 What it does: A zero-shot substrate-specific enzyme generation framework is proposed, utilizing existing enzyme-substrate data for retrieval enhancement, and based on this, new enzymes are generated using a discrete diffusion model.
Retrieval-Augmented Language Model for Knowledge-aware Protein Encoding
Jiasheng Zhang (University of Electronic Science and Technology of China), Jie Shao (University of Electronic Science and Technology of China)
Protein Structure PredictionTransformerBiomedical DataRetrieval-Augmented Generation
🎯 What it does: A retrieval-enhanced protein language model called Kara is proposed, which directly integrates protein knowledge graphs with protein language models.
Retrieval-Augmented Perception: High-resolution Image Perception Meets Visual RAG
Wenbin Wang (Wuhan University), Dacheng Tao (Nanyang Technological University)
RecognitionRetrievalComputational EfficiencyLarge Language ModelImageMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: A training-free Retrieval-Augmented Perception (RAP) framework is proposed, which enhances the perception ability of multimodal large language models (MLLM) for high-resolution (HR) images by retrieving and fusing high-resolution image patches related to the query.
Return Capping: Sample Efficient CVaR Policy Gradient Optimisation
Harry Mead (University of Oxford), Nick Hawes (University of Oxford)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a policy gradient method to optimize CVaR through Return Capping, addressing the sample efficiency issues of traditional methods.
Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces
Henry Moss, Tom Diethe (AstraZeneca)
OptimizationDrug DiscoveryAuto EncoderTabular
🎯 What it does: A Bayesian optimization framework called COWBOYS is proposed, which completely separates the Variational Autoencoder (VAE) from the Gaussian Process (GP). It utilizes GP to directly predict the target in the structural space and restricts sampling to areas where the VAE prior probability is high and the GP predicted values are also high through Bayesian updating.
Revealing Weaknesses in Text Watermarking Through Self-Information Rewrite Attacks
Yixin Cheng (University of British Columbia), Leonid Sigal (University of British Columbia)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: A watermark removal method based on self-information rewriting (SIRA) is proposed, which can effectively eliminate watermarks in LLM-generated text under black-box conditions, without relying on watermark algorithms or model access.
ReverB-SNN: Reversing Bit of the Weight and Activation for Spiking Neural Networks
Yufei Guo (Intelligent Science and Technology Academy of CASIC), Zhe Ma (Intelligent Science and Technology Academy of CASIC)
Spiking Neural NetworkReinforcement LearningImage
🎯 What it does: Proposes the ReverB-SNN method, which combines real-valued synaptic activation with binary weights to enhance the information capacity of SNNs while maintaining the advantages of event-driven and no multiplication.
ReVISE: Learning to Refine at Test-Time via Intrinsic Self-Verification
Hyunseok Lee (KAIST), Jihoon Tack (KAIST)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A framework called ReVISE has been developed, which utilizes LLM self-verification to achieve reasoning self-correction.
Revisiting Chain-of-Thought in Code Generation: Do Language Models Need to Learn Reasoning before Coding?
Ren-Biao Liu (Nanjing University), Ming Li (Nanjing University)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This study systematically evaluates the impact of different CoT positions (CoT + code vs. code + CoT) on model performance in code generation tasks by constructing a high-quality synthetic dataset that includes natural language descriptions, reasoning steps (CoT), and code solutions.
Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning
Shuai Yi (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Domain AdaptationTransformerImageAgriculture Related
🎯 What it does: This paper studies the impact of disrupting the continuity of image tokens in Vision Transformers on cross-domain few-shot learning and proposes the ReCIT method, which reduces domain gaps and enhances model generalization by first performing spatial layer block shuffling followed by frequency domain amplitude balancing shuffling.
Revisiting Convergence: Shuffling Complexity Beyond Lipschitz Smoothness
Qi He (University of Maryland), Heng Huang (University of Texas Arlington)
OptimizationImageTabular
🎯 What it does: This paper studies the convergence of the stochastic shuffling gradient method for solving finite-sum optimization problems under the condition of not satisfying Lipschitz smoothness, providing convergence rates for non-convex, strongly convex, and non-strongly convex cases.
Revisiting Cooperative Off-Policy Multi-Agent Reinforcement Learning
Yueheng Li (Peking University), Zongqing Lu (Peking University)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper addresses the issue of TD target estimation error (TEE) caused by the exponential growth of joint action space in offline/offline collaborative multi-agent reinforcement learning (MARL). It proposes three techniques: annealed multi-step bootstrapping, averaged TD targets, and restricted action representation, and applies them to various existing off-policy MARL methods (QMIX, FACMAC, MADDPG, etc.).
Revisiting Differentially Private Algorithms for Decentralized Online Learning
Xiaoyu Wang (Zhejiang University), Yuanyu Wan (Zhejiang University)
OptimizationFederated LearningSafty and PrivacyTabular
🎯 What it does: This paper proposes an algorithm that can achieve (ε,0)-differential privacy throughout decentralized online learning, and provides corresponding variants that are independent of scheduling and projection.
Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation
Bowen Zheng (Institute of Neuroscience), Tianming Yang (Institute of Neuroscience)
GenerationData SynthesisKnowledge DistillationConvolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: View the diffusion model as generative pre-training, transforming it into an efficient one-step generative model through a single GAN objective, and validating that the pre-trained model possesses generative capability by freezing most parameters.
Revisiting Instance-Optimal Cluster Recovery in the Labeled Stochastic Block Model
Kaito Ariu (CyberAgent), Se-Young Yun (KAIST)
OptimizationComputational Efficiency
🎯 What it does: This paper proposes an algorithm for the random block model with limited labels (LSBM) that achieves instance-optimal clustering under any linear scale cluster size, proving that its misclassification count matches the lower bound of model-specific information.
Revisiting Neural Networks for Few-Shot Learning: A Zero-Cost NAS Perspective
Haidong Kang (Northeastern University)
Meta LearningNeural Architecture SearchConvolutional Neural NetworkImageBenchmark
🎯 What it does: This paper studies a zero-cost neural architecture search method called IBFS, which selects the best architecture for few-shot learning using information bottleneck theory without training.
Revisiting Noise Resilience Strategies in Gesture Recognition: Short-Term Enhancement in sEMG Analysis
Weiyu Guo (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
RecognitionPose EstimationTransformerSupervised Fine-TuningTime Series
🎯 What it does: A learnable short-term enhancement module (STEM) is proposed and embedded into the Transformer architecture to form STET, improving the noise robustness of sEMG posture recognition.
Revisiting Non-Acyclic GFlowNets in Discrete Environments
Nikita Morozov (Higher School of Economics), Sergey Samsonov (Higher School of Economics)
Reinforcement Learning
🎯 What it does: A GFlowNet theory suitable for discrete non-cyclic environments is proposed, simplifying the framework presented by Brunswic et al., and proving that forward/backward policies and flows can be uniquely determined in non-cyclic graphs.
Revisiting the Predictability of Performative, Social Events
Juan Carlos Perdomo (Harvard University)
🎯 What it does: This paper studies the predictability of social forecasting, proving that under the executable performative setting, it is possible to efficiently find predictors that can influence data distribution while satisfying multi-calibration/distinguishability.
Revisiting Unbiased Implicit Variational Inference
Tobias Pielok (Ludwig Maximilian University of Munich), David Rügamer
Flow-based ModelTabularStochastic Differential Equation
🎯 What it does: A new semi-implicit variational inference method called AISIVI is proposed, which uses importance sampling and conditional normalizing flows to approximate intractable conditional distributions, eliminating the MCMC loop of UIVI;
Revolve: Optimizing AI Systems by Tracking Response Evolution in Textual Optimization
Peiyan Zhang (Hong Kong University of Science and Technology), Haohan Wang (University of Illinois at Urbana-Champaign)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By tracking the evolution of LLM responses in multi-turn interactions, we automate the optimization of prompts, strategies, and tool combinations to enhance the reasoning, problem-solving, and code generation capabilities of LLM agents.
Reward Modeling with Ordinal Feedback: Wisdom of the Crowd
Shang Liu (Imperial College London), Xiaocheng Li (Imperial College London)
Recommendation SystemReinforcement Learning from Human FeedbackSupervised Fine-TuningTabular
🎯 What it does: Proposes the use of ordinal feedback in reward modeling to improve traditional binary preference labels.
Reward Translation via Reward Machine in Semi-Alignable MDPs
Yun Hua (Shanghai Jiao Tong University), Xiangfeng Wang (East China Normal University)
Large Language ModelReinforcement LearningSequential
🎯 What it does: A Neural Reward Translation (NRT) framework is proposed in the context of semi-aligned MDP environments, utilizing reward machines and graph matching to achieve cross-domain reward transfer.
Reward-Augmented Data Enhances Direct Preference Alignment of LLMs
Shenao Zhang (Northwestern University), Zhaoran Wang
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Improvements are made to direct preference alignment methods (such as DPO) by proposing a reward-enhanced relabeling approach, allowing the model to consider both the quality of responses and their relative merits during alignment.
Reward-free World Models for Online Imitation Learning
Shangzhe Li (South China University of Technology), Hao Su (University of California, San Diego)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningWorld ModelSequential
🎯 What it does: This paper proposes an online imitation learning framework based on reward-agnostic world models (IQ-MPC), which learns environmental dynamics in latent space and utilizes inverse soft Q-learning to achieve reward decoding, enabling reward-free imitation learning.
Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design
Masatoshi Uehara (EvolutionaryScale), Tommaso Biancalani (EvolutionaryScale)
OptimizationDrug DiscoveryReinforcement LearningDiffusion modelBiomedical Data
🎯 What it does: A new framework is proposed for reward-guided iterative optimization during the testing phase of diffusion models, specifically applied to protein and DNA design.
Reward-Guided Prompt Evolving in Reinforcement Learning for LLMs
Ziyu Ye (Google DeepMind), Yuan Liu (Google DeepMind)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A reward-guided prompt evolution framework called eva is proposed, allowing LLMs to adaptively generate and prioritize prompts during RL fine-tuning, thereby enhancing alignment performance.
Reward-Guided Speculative Decoding for Efficient LLM Reasoning
Baohao Liao (Language Technology Lab, University of Amsterdam), Caiming Xiong (Salesforce AI Research)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A Reward-Guided Speculative Decoding (RSD) framework is proposed, which dynamically mixes a lightweight draft model with a powerful target model and evaluates the quality of each step through a process reward model to enhance the inference efficiency of large language models.
Rhomboid Tiling for Geometric Graph Deep Learning
Yipeng Zhang (Nanyang Technological University), Kelin Xia (Nanyang Technological University)
ClassificationRepresentation LearningDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A hierarchical geometric clustering method based on Rhomboid Tiling (RT clustering) is proposed, and an RTPool graph pooling model is constructed based on this clustering for graph classification tasks.
Riemann Tensor Neural Networks: Learning Conservative Systems with Physics-Constrained Networks
Anas Jnini (University of Trento), Flavio Vella (University of Trento)
Time SeriesPhysics Related
🎯 What it does: A neural network RTNN has been designed and implemented that can strictly satisfy a divergence-free and symmetric tensor field (DFST) for approximating the PDE solutions of conservative physical systems.
Riemannian Diffusion Adaptation for Distributed Optimization on Manifolds
Xiuheng Wang (Universite de Lorraine), Ali H. Sayed (Ecole Polytechnique Federale de Lausanne)
OptimizationTabular
🎯 What it does: This paper proposes a diffusion adaptation algorithm based on Riemannian geometry for online distributed optimization on general Riemannian manifolds.
RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers
Min Zhao (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelVideo
🎯 What it does: A training-free method called RIFLEx is proposed, which reduces the intrinsic frequency of RoPE for video length extrapolation, addressing issues of repetition and slow motion.
Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
Seongho Son (University College London), Ilija Bogunovic (University College London)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper studies the algorithm NS-DPO, which directly optimizes preferences for time drift in large language models (LLMs), and provides theoretical analysis and experimental validation.
Right Time to Learn: Promoting Generalization via Bio-inspired Spacing Effect in Knowledge Distillation
Guanglong Sun (Tsinghua University), Yi Zhong (Tsinghua University)
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A knowledge distillation method based on the spacing effect, Spaced KD, is proposed, which improves the generalization performance of online KD and self-distillation.
Ringmaster ASGD: The First Asynchronous SGD with Optimal Time Complexity
Arto Maranjyan, Peter Richtárik (Skolkovo Institute of Science and Technology)
Optimization
🎯 What it does: This paper proposes Ringmaster ASGD, a new asynchronous stochastic gradient descent algorithm that achieves optimal time complexity under arbitrary heterogeneous computation delays.
RISE: Radius of Influence based Subgraph Extraction for 3D Molecular Graph Explanation
Jingxiang Qu (Stony Brook University), Yi Liu (Stony Brook University)
Explainability and InterpretabilityDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A subgraph extraction method based on the influence radius of each atom, RISE, is proposed to explain the predictive decisions of 3D molecular graph neural networks.
Risk and cross validation in ridge regression with correlated samples
Alexander Atanasov (Harvard University), Cengiz Pehlevan (Harvard University)
Time Series
🎯 What it does: The study investigates the risk of high-dimensional ridge regression in the presence of correlated samples and provides precise asymptotic analysis.