ICML 2024 Papers — Page 20
International Conference on Machine Learning · 2610 papers
Prometheus: Out-of-distribution Fluid Dynamics Modeling with Disentangled Graph ODE
Hao Wu (University of Science and Technology of China), Xiao Luo (University of California)
Domain AdaptationAnomaly DetectionOptimizationComputational EfficiencyGraph Neural NetworkContrastive LearningGraphTime SeriesBenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper presents a new large-scale dataset called Prometheus and designs a DGODE model based on discretized graph ODEs to address the out-of-distribution (OOD) generalization problem in fluid dynamics prediction.
Promises and Pitfalls of Generative Masked Language Modeling: Theoretical Framework and Practical Guidelines
Yuchen Li (Carnegie Mellon University), Andrej Risteski (Carnegie Mellon University)
GenerationData SynthesisOptimizationTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper systematically studies the training and inference mechanisms of Generative Masked Language Models (GMLM) from both theoretical and experimental perspectives, proposing a unified theoretical framework and providing practical guidelines for real-world engineering.
Promoting External and Internal Equities Under Ex-Ante/Ex-Post Metrics in Online Resource Allocation
Karthik Abinav Sankararaman (Meta AI), Pan Xu (New Jersey Institute of Technology)
Optimization
🎯 What it does: This study investigates the issues of external and internal fairness in online resource allocation, proposing a strategy based on linear programming to achieve optimal competitive ratios under both ex-ante and ex-post fairness metrics.
Prompt Sketching for Large Language Models
Luca Beurer-Kellner (ETH Zurich), Martin Vechev (ETH Zurich)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes a framework called Prompt Sketching, which views the generation process of large language models (LLMs) as a sequence decoding problem segmented by templates, achieving more structured reasoning through this framework.
Prompt-based Visual Alignment for Zero-shot Policy Transfer
Haihan Gao (University of Science and Technology of China), Yunji Chen (Institute of Computing Technology, Chinese Academy of Sciences)
Domain AdaptationAutonomous DrivingReinforcement LearningPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes the Prompt-based Visual Alignment (PVA) framework, which uses learnable text prompts to constrain the visual aligner, mapping images from different weather domains to a unified domain, thereby achieving zero-shot strategy transfer.
Prompt-guided Precise Audio Editing with Diffusion Models
Manjie Xu (Beijing Institute of Technology), Dong Yu (Tencent)
Prompt EngineeringDiffusion modelAuto EncoderAudio
🎯 What it does: A prompt-driven, untrained precise audio editing method based on diffusion models (PPAE) is proposed, which achieves the replacement, refinement, and weighting of local audio events by interpolating and fusing cross-attention maps during the diffusion process.
Prompt-tuning Latent Diffusion Models for Inverse Problems
Hyungjin Chung (KAIST), Mauricio Delbracio (Google Research)
RestorationSuper ResolutionPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper proposes a new inverse problem-solving method called P2L, which utilizes a Latent Diffusion Model (LDM) combined with text prompts for joint optimization to enhance image restoration quality.
Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution
Chrisantha Fernando (Google DeepMind), Tim Rocktäschel
Large Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Utilizing the large language model itself for self-improvement, we developed Promptbreeder—a system that automatically generates and evolves task prompts and mutation prompts through evolutionary algorithms, ultimately providing high-performance prompt strategies for various domains of LLMs.
Prompting a Pretrained Transformer Can Be a Universal Approximator
Aleksandar Petrov (University of Oxford), Adel Bibi (University of Oxford)
TransformerPrompt Engineering
🎯 What it does: The paper proves that by keeping the pre-trained Transformer parameters unchanged, the model can become a universal approximator simply by adjusting the prefix (prefix-tuning), capable of approximating any continuous function and sequence-to-sequence function;
Prompting is a Double-Edged Sword: Improving Worst-Group Robustness of Foundation Models
Amrith Setlur (Carnegie Mellon University), Sergey Levine (University of California Berkeley)
ClassificationTransformerPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: Utilizing a base model to accurately predict potential confounders in a zero-shot setting, the predicted confounder labels are used to group training samples, and then a robust classifier is learned using group DRO, proposing the Prompting for Robustness (PfR) method.
Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts
Zhi-Yi Chin (National Yang Ming Chiao Tung University), Wei-Chen Chiu (National Yang Ming Chiao Tung University)
GenerationSafty and PrivacyAdversarial AttackTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: This paper proposes and implements Prompting4Debugging (P4D) — an automated red team tool designed to find harmful prompts that can bypass security mechanisms in text-to-image diffusion models.
Prospective Side Information for Latent MDPs
Jeongyeol Kwon (Wisconsin Institute for Discovery), Constantine Caramanis (University of Texas at Austin)
Reinforcement Learning
🎯 What it does: This paper introduces prospective (pre-given) weakly revealed side information in interactive decision-making problems and studies the learning and sample complexity of such 'LMDPΨ' models, providing upper and lower bounds.
Prospector Heads: Generalized Feature Attribution for Large Models & Data
Gautam Machiraju (Stanford University), Parag Mallick (Stanford University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: This paper proposes Prospector Heads, a feature attribution module that can be directly integrated into any Encoder, enabling feature localization in multimodal large models with few samples and no backpropagation.
Protein Conformation Generation via Force-Guided SE(3) Diffusion Models
YanWang, Quanquan Gu (ByteDance Research)
GenerationProtein Structure PredictionDiffusion modelBiomedical Data
🎯 What it does: A force-guided SE(3) diffusion model called CONFDIFF is proposed for generating multi-conformation protein structures.
Proteus: Exploring Protein Structure Generation for Enhanced Designability and Efficiency
Chentong Wang (Zhejiang University), Longxing Cao (Westlake University)
Protein Structure PredictionGraph Neural NetworkDiffusion modelGraphStochastic Differential Equation
🎯 What it does: A protein backbone generation model named Proteus based on deep diffusion networks has been developed, capable of generating highly designable protein backbones without the need for pre-training.
ProtoGate: Prototype-based Neural Networks with Global-to-local Feature Selection for Tabular Biomedical Data
Xiangjian Jiang (University of Cambridge), Mateja Jamnik (University of Cambridge)
ClassificationOptimizationTabularBiomedical Data
🎯 What it does: ProtoGate proposes a prototype-based neural network that achieves global to local feature selection for high-dimensional low-sample medical tabular data.
Prototypical Transformer As Unified Motion Learners
Cheng Han (University of Missouri), Dongfang Liu (Rochester Institute of Technology)
Object TrackingDepth EstimationTransformerOptical FlowVideo
🎯 What it does: This paper proposes ProtoFormer, a unified prototype-based Transformer framework designed to address various motion tasks (optical flow, depth estimation, object tracking, video stabilization, etc.).
Provable Benefits of Local Steps in Heterogeneous Federated Learning for Neural Networks: A Feature Learning Perspective
Yajie Bao (Shanghai Jiao Tong University), Mingrui Liu (George Mason University)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: The study investigates the impact of the local step gradient descent algorithm on neural network feature learning and generalization performance in heterogeneous federated learning.
Provable Contrastive Continual Learning
Yichen Wen (Shanghai Jiao Tong University), Weiran Huang (Shanghai AI Laboratory)
Knowledge DistillationRepresentation LearningContrastive LearningSequential
🎯 What it does: A provable theory of contrastive learning for continual learning is proposed, and based on this theory, a continual learning algorithm with an adaptive distillation coefficient, CILA, is designed.
Provable Interactive Learning with Hindsight Instruction Feedback
Dipendra Misra (Microsoft Research), Robert E. Schapire
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningMultimodality
🎯 What it does: A theoretical framework for interactive learning from hindsight instruction feedback is proposed, and an algorithm LORIL that achieves no regret under low-rank structure is provided.
Provable Multi-Task Representation Learning by Two-Layer ReLU Neural Networks
Liam Collins (University of Texas), Sanjay Shakkottai (University of Texas)
Representation LearningContrastive Learning
🎯 What it does: This paper proves that in multi-task pre-training, a two-layer ReLU neural network can learn the true low-dimensional feature subspace through gradient descent and transfer this representation to downstream tasks;
Provable Privacy with Non-Private Pre-Processing
Yaxi Hu (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Safty and PrivacyTabular
🎯 What it does: The paper proposes a general framework for quantifying the additional privacy cost incurred by using non-private data-related preprocessing steps in differential privacy (DP) machine learning pipelines. It provides privacy safety upper bounds for various commonly used preprocessing algorithms (such as deduplication, quantization, missing value imputation, normalization, PCA) and, based on this, offers an unconditional privacy guarantee implementation through the Propose-Test-Release (PTR) mechanism.
Provable Representation with Efficient Planning for Partially Observable Reinforcement Learning
Hongming Zhang (University of Alberta), Bo Dai (Google DeepMind)
Robotic IntelligenceReinforcement LearningImage
🎯 What it does: A representation and planning method that is provably effective in partially observable Markov decision processes (POMDPs) is proposed, addressing the statistical and computational bottlenecks of traditional POMDP learning and planning;
Provable Risk-Sensitive Distributional Reinforcement Learning with General Function Approximation
Yu Chen (Tsinghua University), Longbo Huang (Tsinghua University)
Reinforcement Learning
🎯 What it does: This paper proposes a general framework to study risk-sensitive distributed reinforcement learning under static Lipschitz risk measures, and presents two meta-algorithms RS-DisRL-M and RS-DisRL-V.
Provably Better Explanations with Optimized Aggregation of Feature Attributions
Thomas Decker (Siemens AG), Florian Buettner (German Cancer Research Center)
OptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a technique to enhance the quality of explanations by convexly combining multiple attribution results for interpretable models.
Provably Efficient Exploration in Quantum Reinforcement Learning with Logarithmic Worst-Case Regret
Han Zhong (Peking University), Liwei Wang (Peking University)
Reinforcement Learning
🎯 What it does: An online exploration algorithm for quantum reinforcement learning is proposed, achieving logarithmic worst-case cumulative regret bounds for both finite tabular MDPs and linear mixture MDPs.
Provably Efficient Long-Horizon Exploration in Monte Carlo Tree Search through State Occupancy Regularization
Liam Schramm (Rutgers University), Abdeslam Boularias (Rutgers University)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: A Monte Carlo Tree Search algorithm based on state occupancy regularization, Volume-MCTS, is proposed to address the weaknesses of traditional MCTS in long-period exploration.
Provably Efficient Partially Observable Risk-sensitive Reinforcement Learning with Hindsight Observation
Tonghe Zhang (Tsinghua University), Longbo Huang (Tsinghua University)
OptimizationReinforcement Learning
🎯 What it does: This study proposes a novel risk-sensitive reinforcement learning algorithm suitable for partially observable environments and introduces the concept of post-observation to optimize cumulative rewards under entropy risk measures.
Provably Efficient Reinforcement Learning for Adversarial Restless Multi-Armed Bandits with Unknown Transitions and Bandit Feedback
GUOJUN XIONG, Jian Li (Stony Brook University)
OptimizationReinforcement LearningTabular
🎯 What it does: A learning framework for RMAB called UCMD-ARMAB is proposed, suitable for unknown transition functions, adversarial rewards, and only bandit feedback, with the goal of maximizing total rewards while satisfying immediate activation constraints.
Provably Neural Active Learning Succeeds via Prioritizing Perplexing Samples
Dake Bu (City University of Hong Kong), Hau-San Wong (City University of Hong Kong)
Representation LearningData-Centric LearningConvolutional Neural NetworkTabular
🎯 What it does: This paper theoretically analyzes the unified mechanism of two mainstream neural network active learning (NAL) query criteria—uncertainty sampling and diversity sampling—in the feature learning process. It demonstrates that both methods effectively learn weakly scarce features by prioritizing the selection of 'confusing samples' that have not yet been learned, thereby achieving approximately Bayesian optimal generalization performance with a limited number of labels.
Provably Robust DPO: Aligning Language Models with Noisy Feedback
Sayak Ray Chowdhury (Microsoft Research), Nagarajan Natarajan (Microsoft Research)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A robust direct preference optimization (rDPO) method is proposed for noisy preference data, aimed at training language models that are more stable to human preferences.
Provably Scalable Black-Box Variational Inference with Structured Variational Families
Joohwan Ko (KAIST), Jacob R. Gardner (University of Pennsylvania)
OptimizationTabular
🎯 What it does: This paper studies the scalability of structured variational families in black-box variational inference (BBVI) and provides theoretical bounds on its iterative complexity.
Pruned Pivot: Correlation Clustering Algorithm for Dynamic, Parallel, and Local Computation Models
Mina Dalirrooyfard (Morgan Stanley), Slobodan Mitrović (University of California Davis)
OptimizationComputational EfficiencyGraph
🎯 What it does: A Pruned Pivot algorithm is proposed for efficiently solving the unweighted related clustering problem under dynamic, MPC, and LCA models, achieving clustering results close to a 3-approximation.
Pruner-Zero: Evolving Symbolic Pruning Metric From Scratch for Large Language Models
Peijie Dong (Hong Kong University of Science and Technology), Xiaowen Chu (Hong Kong University of Science and Technology)
Large Language ModelText
🎯 What it does: Utilizing genetic programming to automatically evolve symbolic pruning metrics, the Pruner-Zero framework is proposed to achieve post-training sparsification of LLMs without updates;
PruNeRF: Segment-Centric Dataset Pruning via 3D Spatial Consistency
Yeonsung Jung (Korea Advanced Institute of Science and Technology), Eunho Yang (AITRICS)
SegmentationData-Centric LearningNeural Radiance FieldImage
🎯 What it does: This paper proposes PruNeRF, a segment-level dataset pruning framework based on 3D spatial consistency and influence functions.
Pseudo-Calibration: Improving Predictive Uncertainty Estimation in Unsupervised Domain Adaptation
Dapeng Hu (Centre for Frontier AI Research A STAR), Chuan-Sheng Foo (Institute for Infocomm Research A STAR)
Domain AdaptationImage
🎯 What it does: This paper proposes a post-hoc calibration framework called Pseudo-Calibration (PseudoCal), which addresses the issue of uncertainty estimation in predictions for unlabeled target domain samples by synthesizing pseudo-target data through mixup during the inference phase, followed by calibration of this pseudo data using temperature scaling.
Purify Unlearnable Examples via Rate-Constrained Variational Autoencoders
Yi Yu (Nanyang Technological University), Alex Kot
Anomaly DetectionAuto EncoderImage
🎯 What it does: A pre-training decontamination method based on rate-constrained Variational Autoencoders (VAE) is proposed, which effectively eliminates unlearnable sample (UE) disturbances in the training data and restores the learnability of the model.
Purifying Quantization-conditioned Backdoors via Layer-wise Activation Correction with Distribution Approximation
Boheng Li (Wuhan University), Tianwei Zhang (Nanyang Technological University)
OptimizationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a method for eliminating backdoors in quantized conditional backdoor attacks (QCB).
Pursuing Overall Welfare in Federated Learning through Sequential Decision Making
Seok-Ju Hahn (Ulsan National Institute of Science and Technology), Junghye Lee (Seoul National University)
OptimizationFederated LearningImageTabular
🎯 What it does: The AAggFF framework is proposed, treating fair aggregation as an online convex optimization (OCO) problem. AAggFF-S (based on ONS) and AAggFF-D (based on EG and double robust estimation) are designed for cross-silo and cross-device scenarios, respectively, to enhance client-level fairness in federated learning (FL).
Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels
Haoning Wu (Nanyang Technological University), Weisi Lin (Nanyang Technological University)
ClassificationRecognitionRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageVideoMultimodality
🎯 What it does: A teaching method based on discrete text levels, Q-ALIGN, is proposed, which uses large-scale multimodal models (LMM) to score image quality, aesthetics, and video quality, and builds a unified model, ONEALIGN, covering three visual scoring tasks.
Q-Probe: A Lightweight Approach to Reward Maximization for Language Models
Kenneth Li (Harvard University), David Brandfonbrener (Harvard University)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes Q-probe, a method for training a linear probe on a frozen pre-trained language model to reweight candidate generation results for reward maximization;
Q-Star Meets Scalable Posterior Sampling: Bridging Theory and Practice via HyperAgent
Yingru Li (Chinese University of Hong Kong), Zhi-Quan Luo (Chinese University of Hong Kong)
Reinforcement LearningTabularSequential
🎯 What it does: This paper presents HyperAgent, a reinforcement learning algorithm based on a hypermodel framework, which achieves efficient exploration by incrementally approximating the posterior of Q⋆ and implementing a greedy strategy based on sampled posteriors, balancing data and computational efficiency.
Q-value Regularized Transformer for Offline Reinforcement Learning
Shengchao Hu (Shanghai Jiao Tong University), Dacheng Tao (Nanyang Technological University)
TransformerReinforcement LearningSequentialBenchmark
🎯 What it does: A Q-value regularized Transformer (QT) is proposed, which combines the conditional sequence modeling of Transformers with dynamic programming (Q-learning) for offline reinforcement learning.
QBMK: Quantum-based Matching Kernels for Un-attributed Graphs
Lu Bai (Beijing Normal University), Edwin Hancock
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a kernel for graph matching without attributes, called QBMK, based on Continuous-Time Quantum Random Walks (CTQW), to measure the similarity between graphs.
QORA: Zero-Shot Transfer via Interpretable Object-Relational Model Learning
Gabriel Stella (Texas A&M University), Dmitri Loguinov (Texas A&M University)
Domain AdaptationExplainability and Interpretability
🎯 What it does: Learned the object relation transfer model and implemented the QORA algorithm without any domain knowledge.
Quality Diversity through Human Feedback: Towards Open-Ended Diversity-Driven Optimization
Li Ding (University of Massachusetts Amherst), Joel Lehman
OptimizationRobotic IntelligenceDiffusion modelContrastive LearningText
🎯 What it does: This study proposes the QDHF method, which is based on human feedback learning diversity metrics and integrates quality-diversity (QD) algorithms to achieve open optimization.
Quality-Diversity Actor-Critic: Learning High-Performing and Diverse Behaviors via Value and Successor Features Critics
Luca Grillotti (Imperial College London), Antoine Cully (Iconic AI)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper proposes a deep reinforcement learning algorithm called QDAC, which can simultaneously learn high performance and diverse behaviors.
Quality-Diversity with Limited Resources
Ren-Jian Wang (Nanjing University), Chao Qian (Nanjing University)
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: The study investigates how to efficiently train Quality-Diversity (QD) algorithms under resource constraints and proposes and implements the RefQD method.
Quality-Weighted Vendi Scores And Their Application To Diverse Experimental Design
Quan Nguyen (Washington University in St. Louis), Adji Bousso Dieng (Princeton University)
Recommendation SystemOptimizationTabular
🎯 What it does: A new qVS metric is proposed, which incorporates quality weights based on the original Vendi diversity metric, and is applied to Active Search and Bayesian Optimization tasks to achieve diversified experimental design.
Quantum Algorithm for Online Exp-concave Optimization
Jianhao He (Chinese University of Hong Kong), John C.S. Lui (Chinese University of Hong Kong)
OptimizationPhysics Related
🎯 What it does: This paper explores the application of quantum advantage in zero-order feedback online convex optimization problems, proposing a quantum online quasi-Newton method and proving that this method exhibits quantum advantage in such problems.
Quantum Algorithms and Lower Bounds for Finite-Sum Optimization
Yexin Zhang (Peking University), Tongyang Li (Peking University)
OptimizationPhysics Related
🎯 What it does: A finite-sum optimization algorithm based on quantum variational techniques is proposed, providing quantum query complexity and lower bounds.
Quantum Implicit Neural Representations
Jiaming Zhao (Tianjin University), Hui Gao (Tianjin University)
GenerationSuper ResolutionImagePhysics RelatedAudio
🎯 What it does: This paper proposes an implicit neural representation network QIREN based on quantum data re-uploading circuit for efficient modeling of continuous signals.
Quantum Positional Encodings for Graph Neural Networks
Slimane Thabet (Pasqal), Loic Henriet
Graph Neural NetworkTransformerGraphPhysics Related
🎯 What it does: This paper proposes the use of quantum computing-generated graph positional encoding to enhance the representational capacity of graph neural networks (especially graph Transformers).
Quantum Theory and Application of Contextual Optimal Transport
Nicola Mariella (IBM Quantum), Jannis Born (IBM Research)
OptimizationBiomedical DataPhysics Related
🎯 What it does: A context-optimal transport model QontOT based on quantum computing is proposed, utilizing the correspondence between quantum unit matrices and double stochastic matrices to predict transportation plans influenced by context.
Quasi-Monte Carlo Features for Kernel Approximation
Zhen Huang (Columbia University), Yian Huang (Columbia University)
Tabular
🎯 What it does: This paper proposes the use of quasi-Monte Carlo (QMC) sequences (primarily Halton sequences) as random features to approximate kernel functions. It provides deterministic error upper bounds for translation-invariant kernels such as Gaussian and Cauchy kernels, and applies this method to kernel ridge regression (KRR), proving that under suitable smoothness assumptions, the required number of features can be significantly reduced.
QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference
Jiaming Tang (Shanghai Jiao Tong University), Song Han (Massachusetts Institute of Technology)
RetrievalComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper studies the query-aware algorithm Quest for KV cache sparsity in long-context LLM inference.
QuIP$\#$: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks
Albert Tseng (Cornell University), Christopher De Sa (Cornell University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A weight quantization PTQ method called QuIP# is proposed, based on Random Hadamard Transform (RHT) and E8 voxel codebook (E8P), achieving near FP16 accuracy for LLMs (Llama 1/2) under 2-4 bit quantization while maintaining inference speed.
QuRating: Selecting High-Quality Data for Training Language Models
Alexander Wettig (Princeton University), Danqi Chen (Princeton University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By comparing text pairs using GPT-3.5, a scoring model is learned and these scores are used to select 30B training samples from the 260B token SlimPajama corpus, followed by training a 1.3B parameter language model;
R2E: Turning any Github Repository into a Programming Agent Environment
Naman Jain (University of California), Ion Stoica (University of California)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a framework called R2E that converts any GitHub repository into an interactive programming agent testing environment, and based on this, constructs the first large-scale real environment evaluation benchmark R2E-Eval1.
Random Exploration in Bayesian Optimization: Order-Optimal Regret and Computational Efficiency
Sudeep Salgia (Carnegie Mellon University), Qing Zhao (Cornell University)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a strategy that combines random sampling and domain reduction in Bayesian optimization, achieving efficient optimization without adaptive exploration.
Random features models: a way to study the success of naive imputation
Alexis Ayme (Sorbonne Université), Erwan Scornet (Sorbonne Université)
🎯 What it does: This paper studies the effectiveness of naive imputation in handling missing data, particularly in high-dimensional linear predictive models, confirming that under the assumption of missing completely at random (MCAR), the bias of naive imputation is negligible, and it remains effective even in very low-dimensional cases.
Random Latent Exploration for Deep Reinforcement Learning
Srinath V. Mahankali, Pulkit Agrawal (Massachusetts Institute of Technology)
Reinforcement LearningVideo
🎯 What it does: This paper proposes a random latent reward-based exploration strategy (RLE) that directly injects random rewards into task rewards and conditions the policy and value network on latent vectors to achieve diverse trajectories.
Random Masking Finds Winning Tickets for Parameter Efficient Fine-tuning
Jing Xu (Tsinghua University), Jingzhao Zhang (Shanghai Qizhi Institute)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper studies a minimalist Random Masking method for Parameter-Efficient Fine-Tuning (PEFT) and demonstrates its competitive performance with traditional methods like LoRA across various language and vision tasks.
Random matrix theory improved Fréchet mean of symmetric positive definite matrices
Florent Bouchard (Universite Paris Saclay), Frederic Pascal
OptimizationTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: This paper studies a method for estimating the Fréchet mean of symmetric positive definite (SPD) matrices based on random matrix theory (RMT) and proposes improved algorithms for covariance estimation and distance calculation.
Random Scaling and Momentum for Non-smooth Non-convex Optimization
Qinzi Zhang (Boston University), Ashok Cutkosky (Boston University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A (c,ε)-Goldstein stability point definition is proposed, and the Exponentiated O2NC framework is constructed, combining SGDM with exponential random scaling to achieve optimal convergence rates for non-smooth non-convex optimization.
Randomized Confidence Bounds for Stochastic Partial Monitoring
Maxime Heuillet (Universite Laval), Audrey Durand (Universite Laval)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: A partial monitoring (PM) algorithm under randomized confidence bounds (RandCBP, RandCBPside⋆) is proposed, and it is proven that it maintains the same theoretical lower bounds as traditional CBP and CBPside under randomization; a practical linear context PM framework is also provided, demonstrating the application of this method in real-world error rate monitoring tasks.
Ranking-based Client Imitation Selection for Efficient Federated Learning
Chunlin Tian (University of Macau), Cheng-zhong Xu
Federated LearningConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: FedRank is proposed, a federated learning client selection framework based on ranking and imitation learning, which can adaptively and efficiently select the most contributive devices;
Rapid Learning without Catastrophic Forgetting in the Morris Water Maze
Raymond Wang, Ila R Fiete (Massachusetts Institute of Technology)
Convolutional Neural NetworkSequential
🎯 What it does: A continuous and rapid learning framework based on biological principles (vHSN) has been constructed, capable of quickly learning new platform locations in a series of continuous Morris Water Maze (sWM) environments while retaining memory of old environments, achieving zero-forgetting transfer from new to old environments.
Rate-Optimal Policy Optimization for Linear Markov Decision Processes
Uri Sherman (Tel Aviv University), Yishay Mansour (Google Research)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper studies regret minimization in online contextual linear Markov decision processes and proposes a computationally efficient policy optimization algorithm that achieves optimal regret of e O(√K) over K rounds.
RAUCA: A Novel Physical Adversarial Attack on Vehicle Detectors via Robust and Accurate Camouflage Generation
Jiawei Zhou (Harbin Institute of Technology), YU LI
Object DetectionAutonomous DrivingAdversarial AttackNeural Radiance FieldImage
🎯 What it does: A UV map-based physical adversarial camouflage attack framework RAUCA is proposed, which evades vehicle detectors by optimizing the surface texture of vehicles.
Re-Dock: Towards Flexible and Realistic Molecular Docking with Diffusion Bridge
Yufei Huang (Zhejiang University), Stan Z. Li (Westlake University)
Drug DiscoveryGraph Neural NetworkDiffusion modelBiomedical Data
🎯 What it does: A generative molecular docking framework named Re-Dock is proposed and implemented, specifically designed to address flexible docking tasks, capable of simultaneously predicting the binding poses of ligands and pocket side chains.
Realistic Unsupervised CLIP Fine-tuning with Universal Entropy Optimization
Jian Liang (Chinese Academy of Sciences), Tieniu Tan (Chinese Academy of Sciences)
ClassificationDomain AdaptationAnomaly DetectionOptimizationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: In real-world scenarios without labels and potentially containing unknown categories, unsupervised fine-tuning based on CLIP improves the recognition accuracy of known categories while enhancing the detection capability for OOD (out-of-distribution) samples.
Reason for Future, Act for Now: A Principled Architecture for Autonomous LLM Agents
Zhihan Liu (Northwestern University), Zhaoran Wang
OptimizationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: The RAFA framework is proposed, enabling LLM to achieve adaptive interaction through 'reasoning for the future first, then acting for the present', and provides a theoretical upper bound on √T regret.
Receptive Fields As Experts in Convolutional Neural Architectures
Dongze Lian (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: A Mixture of Receptive Fields (MoRF) framework is proposed and implemented, which can dynamically select experts with different convolution kernel sizes based on the input, thereby enhancing the receptive field expressiveness of CNNs.
ReconBoost: Boosting Can Achieve Modality Reconcilement
Cong Hua (Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
OptimizationMultimodality
🎯 What it does: A multi-modal learning framework called ReconBoost based on modal alternating updates is designed to alleviate the modal competition problem and enhance the utilization of each modality.
Recovering Labels from Local Updates in Federated Learning
Huancheng Chen (University of Texas at Austin), Haris Vikalo (University of Texas at Austin)
Federated LearningAdversarial AttackImage
🎯 What it does: A label recovery attack method based on local model updates, called RLU, has been designed;
Recovering the Pre-Fine-Tuning Weights of Generative Models
Eliahu Horwitz (Hebrew University of Jerusalem), Yedid Hoshen (Hebrew University of Jerusalem)
GenerationOptimizationAdversarial AttackTransformerSupervised Fine-TuningDiffusion modelImageBenchmark
🎯 What it does: This paper proposes a novel attack - Pre-Fine-Tuning Weight Recovery, which utilizes LoRA fine-tuned models to recover the original weights of pre-fine-tuned models through an unsupervised, data-free iterative low-rank decomposition method.
Recurrent Distance Filtering for Graph Representation Learning
Yuhui Ding (ETH Zurich), Thomas Hofmann (ETH Zurich)
Representation LearningRecurrent Neural NetworkGraph Neural NetworkGraph
🎯 What it does: A distance-filtered recursive encoding model GRED is proposed, which updates the graph node representations using multi-scale aggregation and linear RNN.
Recurrent Early Exits for Federated Learning with Heterogeneous Clients
Royson Lee (University of Cambridge), Nicholas Donald Lane
Federated LearningKnowledge DistillationTransformerSupervised Fine-TuningImageAudio
🎯 What it does: A recursive early exit mechanism named ReeFL is proposed, which utilizes a shared Transformer-based module to fuse features from different depth sub-models, supporting multi-level early stopping and feature modulation on heterogeneous clients.
ReDiffuser: Reliable Decision-Making Using a Diffuser with Confidence Estimation
Nantian He (Tsinghua University), You He (Tsinghua University)
Robotic IntelligenceReinforcement LearningDiffusion modelTabular
🎯 What it does: Proposes ReDiffuser, which incorporates confidence estimation based on Random Network Distillation (RND) into diffusion-based offline reinforcement learning to achieve reliable decision-making;
Reducing Balancing Error for Causal Inference via Optimal Transport
Yuguang Yan (Guangdong University of Technology), Zhifeng Hao (Shantou University)
OptimizationRepresentation LearningTabular
🎯 What it does: The paper proposes a new framework for estimating the Average Treatment Effect (ATE) in observational data by linking covariance error (balancing error) with Wasserstein distance. This framework learns sample weights and transport costs simultaneously within the Optimal Transport model, achieving balance between the treatment and control group distributions, ultimately reducing bias in causal inference.
Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation
Yuchen Yang (Nankai University), Jun Xu (Nankai University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: A strategy is proposed to reduce activation memory usage during the fine-tuning process of large models through approximate backpropagation and memory sharing.
Reducing Item Discrepancy via Differentially Private Robust Embedding Alignment for Privacy-Preserving Cross Domain Recommendation
Weiming Liu (Zhejiang University), Yew-Soon Ong (Nanyang Technological University)
Recommendation SystemSafty and PrivacyText
🎯 What it does: To address the cross-domain recommendation problem with non-overlapping users and items while protecting data privacy, the RidCDR model is proposed, which achieves knowledge sharing and data sparsity resolution through two modules: single-domain prediction and private robust embedding alignment.
Reducing sequential change detection to sequential estimation
Shubhanshu Shekhar (Carnegie Mellon University), Aaditya Ramdas (Carnegie Mellon University)
Anomaly DetectionTime SeriesSequential
🎯 What it does: A new sequential change point detection method (RCS-Detector) is proposed, which announces detection when the intersection of all active confidence sequences (CS) is empty at each moment.
Referee Can Play: An Alternative Approach to Conditional Generation via Model Inversion
Xuantong LIU, Yuan Yao (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerVision Language ModelScore-based ModelAuto EncoderImageText
🎯 What it does: A training-free text-to-image generation method based on a reverse Vision-Language Model (VLM) is proposed, which directly optimizes the image (latent) during the generation process to meet the text prompt.
Reference Neural Operators: Learning the Smooth Dependence of Solutions of PDEs on Geometric Deformations
Ze Cheng (Bosch China Investment Co Ltd), Hang Su (Tsinghua University)
TransformerMeshPhysics Related
🎯 What it does: This paper studies a novel neural operator—Reference Neural Operator (RNO)—to learn the smooth dependence of reference geometric solutions on geometric deformations, enabling rapid prediction of PDE solutions under geometric transformations.
Refined Coreset Selection: Towards Minimal Coreset Size under Model Performance Constraints
Xiaobo Xia (University of Sydney), Tongliang Liu (University of Sydney)
OptimizationImage
🎯 What it does: The refined core subset selection (RCS) problem is proposed, and a dictionary-based bi-level optimization algorithm is designed to minimize the size of the core subset while satisfying model performance constraints.
Refining Minimax Regret for Unsupervised Environment Design
Michael Beukman (University of Oxford), Jakob Nicolaus Foerster
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper proposes the Bayesian level-perfect Minimax Regret (BLP) objective and the corresponding ReMiDi algorithm to address the issue of stagnation in MMR training in partially observable environments.
Reflected Flow Matching
Tianyu Xie (Peking University), Cheng Zhang (Peking University)
GenerationData SynthesisFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: A framework for learning continuous normalizing flows (CNF) within constrained domains is proposed, called the Reflective Flow Matching (RFM) framework.
Reflective Policy Optimization
Yaozhong Gan (QiYuan Lab), Junliang Xing (QiYuan Lab)
OptimizationReinforcement LearningSequential
🎯 What it does: A novel on-policy reinforcement learning algorithm named Reflective Policy Optimization (RPO) is proposed, which utilizes information from state-action pairs before and after to reflect and update the policy.
ReGAL: Refactoring Programs to Discover Generalizable Abstractions
Elias Stengel-Eskin (University of North Carolina), Mohit Bansal (University of North Carolina)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: By refactoring and validating existing programs, an automatically learned reusable function library (REGAL) enhances the accuracy of large language models (LLMs) in program generation tasks.
Regression Learning with Limited Observations of Multivariate Outcomes and Features
Yifan Sun (University of Western Ontario), Grace Yi
OptimizationBiomedical Data
🎯 What it does: An online learning algorithm for multivariate linear regression is proposed, which addresses both missing features and response variables. Efficient gradient estimation and iterative update schemes are provided for L2 and L1 losses, as well as ridge/lasso regularization, along with PAC error bounds.
Regression with Multi-Expert Deferral
Anqi Mao (New York University), Yutao Zhong (Google Research)
Mixture of ExpertsTabular
🎯 What it does: This paper studies a framework for using multiple experts for deferred prediction in regression tasks and proposes single-stage and two-stage surrogate losses along with corresponding H-consistency theoretical guarantees.
Regularized Q-learning through Robust Averaging
Peter Schmitt-Förster (University of Konstanz), Tobias Sutter (University of Konstanz)
Reinforcement LearningTabular
🎯 What it does: A robust average regularized Q-learning (2RA Q-learning) is proposed, which can learn optimal Q-values and control estimation bias in asynchronous environments.
Regularizing with Pseudo-Negatives for Continual Self-Supervised Learning
Sungmin Cha (New York University), Taesup Moon (Seoul National University)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A Pseudo-Negative Regularization (PNR) framework is proposed for Continual Self-Supervised Learning (CSSL), which balances plasticity and stability by incorporating pseudo-negative samples from the current and previous models into both contrastive and non-contrastive losses.
Reinforcement Learning and Regret Bounds for Admission Control
Lucas Weber (Inria), Jiamin Zhu (IFP Energies nouvelles)
OptimizationReinforcement Learning
🎯 What it does: This paper addresses the multi-class job acceptance control problem in the M/M/c/S queue and proposes a UCRL-AC algorithm based on UCRL2, utilizing the system structure to achieve optimal exploration of average rewards. It provides the expected regret upper bounds for both finite and infinite server scenarios.
Reinforcement Learning from Reachability Specifications: PAC Guarantees with Expected Conditional Distance
Jakub Svoboda (Institute of Science and Technology), Krishnendu Chatterjee (Institute of Science and Technology)
Reinforcement Learning
🎯 What it does: The paper studies PAC learning of reachability specifications in unknown environments by introducing the Expected Condition Distance (ECD).
Reinforcement Learning within Tree Search for Fast Macro Placement
Zijie Geng (University of Science and Technology of China), Feng Wu (Huawei)
OptimizationReinforcement LearningTabularBenchmark
🎯 What it does: EfficientPlace is proposed, a dual-layer framework that combines MCTS tree search with reinforcement learning for fast and high-quality macro layout.
Reinformer: Max-Return Sequence Modeling for Offline RL
Zifeng Zhuang (Zhejiang University), Donglin Wang (Westlake University)
TransformerReinforcement LearningSequential
🎯 What it does: Proposes a max-return sequence modeling framework and implements the Reinformer algorithm for sequence prediction in offline reinforcement learning.
Rejuvenating image-GPT as Strong Visual Representation Learners
Sucheng Ren (Johns Hopkins University), Cihang Xie (UCSanta Cruz)
SegmentationGenerationKnowledge DistillationRepresentation LearningTransformerImage
🎯 What it does: Transformed from iGPT to D-iGPT by changing the prediction target from original pixels to semantic tokens provided by CLIP, and adding supervision on visible tokens, a stronger visual representation learning framework has been constructed.