ICML 2024 Papers — Page 21
International Conference on Machine Learning · 2610 papers
Relational DNN Verification With Cross Executional Bound Refinement
Debangshu Banerjee (University of Illinois Urbana-Champaign), Gagandeep Singh (VMware Research)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A new relational deep neural network (DNN) verification framework called RACoon is proposed, which can accurately compute linear approximations by leveraging cross-execution dependencies across multiple executions and perform relational verification through MILP.
Relational Learning in Pre-Trained Models: A Theory from Hypergraph Recovery Perspective
Yang Chen (Peking University), Bing Liu (University of Illinois Chicago)
Large Language ModelMultimodalityGraph
🎯 What it does: The paper abstracts the relationships of world entities as a weighted hypergraph and equates the relationship learning of pre-trained models to the recovery of this hypergraph, proving that it can be identified under sufficient samples and providing a sample complexity analysis.
Relaxed Quantile Regression: Prediction Intervals for Asymmetric Noise
Thomas Pouplin (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Tabular
🎯 What it does: This paper studies an algorithm for directly learning prediction intervals—Relaxed Quantile Regression (RQR), which eliminates the limitations of traditional quantile assumptions and can generate more flexible interval predictions.
Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering
Haoxuan Li (Peking University), Xiao-Hua Zhou (Peking University)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: This paper proposes a double robust (DR) estimator in collaborative filtering that does not require complete accuracy in filling erroneous pseudo-labels, and provides a corresponding learning framework.
ReLU Network with Width $d+\mathcal{O}(1)$ Can Achieve Optimal Approximation Rate
Chenghao Liu (City University of Hong Kong), Minghua Chen (City University of Hong Kong)
🎯 What it does: It is proven that ReLU networks with a width of d+1 (or d+O(1)) can achieve the optimal approximation rate O(ω_f(L^{-2/d})) for continuous functions under Lp (p∈[1,∞)) and uniform norms, and a constructive proof is provided.
ReLU to the Rescue: Improve Your On-Policy Actor-Critic with Positive Advantages
Andrew Jesson (University of Oxford), Yarin Gal (University of Oxford)
Reinforcement Learning
🎯 What it does: Three modifications were introduced based on A3C: ReLU truncation of advantage estimates (only updating positive advantages), spectral normalization weights for both actor and critic networks, and implementing approximate Bayesian inference through Dropout, resulting in positive advantage updates, stable value estimates, and adaptive exploration.
ReLUs Are Sufficient for Learning Implicit Neural Representations
Joseph Shenouda (University of Wisconsin Madison), Robert D Nowak
Image TranslationRestorationSuper ResolutionImageComputed Tomography
🎯 What it does: This paper constructs the BW-ReLU network by adding constraints based on second-order B-spline wavelets to ReLU neurons, achieving the learning of implicit neural representations (INR).
ReMax: A Simple, Effective, and Efficient Reinforcement Learning Method for Aligning Large Language Models
Ziniu Li (Chinese University of Hong Kong), Zhi-Quan Luo (Chinese University of Hong Kong)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The ReMax algorithm is proposed, utilizing REINFORCE with a greedy baseline for RLHF, replacing the traditional PPO to simplify implementation and reduce resource consumption.
REMEDI: Corrective Transformations for Improved Neural Entropy Estimation
Viktor Nilsson (KTH Royal Institute of Technology), Pierre Nyquist (Chalmers University of Technology)
GenerationData SynthesisOptimizationGenerative Adversarial NetworkImage
🎯 What it does: The REMEDI method is proposed, which combines a learnable base distribution with the Donsker-Varadhan loss to improve differential entropy estimation, and applies it to information bottleneck and generative model training.
Remembering to Be Fair: Non-Markovian Fairness in Sequential Decision Making
Parand A. Alamdari (University of Toronto), Sheila A. McIlraith (University of Toronto)
Reinforcement LearningSequential
🎯 What it does: This paper provides a non-Markovian theoretical definition of fairness in sequential decision-making with multiple stakeholders and proposes the FairQCM reinforcement learning algorithm based on memory augmentation and adversarial experience generation.
Removing Spurious Concepts from Neural Network Representations via Joint Subspace Estimation
Floris Holstege (University of Amsterdam), Cees Diks (University of Amsterdam)
Representation LearningText
🎯 What it does: A post-hoc concept removal method is proposed—Joint Subspace Estimation (JSE), which simultaneously identifies the orthogonal low-dimensional subspaces of the main task and pseudo-concepts to eliminate pseudo-concepts from neural network embeddings without losing information about the main task.
Rényi Pufferfish Privacy: General Additive Noise Mechanisms and Privacy Amplification by Iteration via Shift Reduction Lemmas
Clément Pierquin (Craft AI), Matthieu Boussard (University of Lille)
OptimizationSafty and PrivacyTabular
🎯 What it does: A Rényi-diversity-based Pufferfish privacy framework (RPP) is proposed, along with a General Wasserstein Mechanism (GWM) and its improved versions (GAWM, DAGWM), to achieve privacy protection for related data and various secrets; further, through iterative privacy amplification (PABI), it is proven that traditional adaptive composition can be avoided in iterative algorithms (such as stochastic gradient descent), providing a better privacy-utility balance.
Reparameterized Importance Sampling for Robust Variational Bayesian Neural Networks
Yunfei Long (Harbin Engineering University), Huosheng Xu (Harbin Engineering University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A variational inference method for Bayesian neural networks that combines reparameterization importance sampling (RIS) is proposed, significantly reducing Monte Carlo sampling variance and improving convergence speed and predictive performance.
Repeat After Me: Transformers are Better than State Space Models at Copying
Samy Jelassi (Harvard University), eran malach
TransformerText
🎯 What it does: This study investigates the differences in the capabilities of Transformer and General State Space Model (GSSM) in copying tasks and provides theoretical proofs along with large-scale experimental validation.
Replicable Learning of Large-Margin Halfspaces
Alkis Kalavasis (Yale University), Felix Zhou (Yale University)
OptimizationSupervised Fine-Tuning
🎯 What it does: This paper proposes three efficient replicable algorithms for learning large-margin half-spaces, significantly reducing sample complexity.
Repoformer: Selective Retrieval for Repository-Level Code Completion
Di Wu (University of California Los Angeles), Xiaofei Ma (AWS AI Labs)
RetrievalAI Code AssistantTransformerContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: A self-triggering retrieval-enhanced generation framework called REPOFORMER is proposed, which can determine whether to retrieve cross-file information based on context in repository-level code completion, avoiding ineffective or harmful retrievals.
Representation Surgery for Multi-Task Model Merging
Enneng Yang (Northeastern University), Dacheng Tao (Nanyang Technological University)
ClassificationRepresentation LearningTransformerImage
🎯 What it does: This paper addresses the representation bias problem that arises in multi-task model fusion and proposes a lightweight task-specific 'representation surgery' module to correct features after model fusion, aiming to approximate the representations of single-task models.
Representation Surgery: Theory and Practice of Affine Steering
Shashwat Singh (International Institute of Information Technology Hyderabad), Ponnurangam Kumaraguru (International Institute of Information Technology Hyderabad)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Two optimal affine steering functions (mean matching and mean + covariance matching) are proposed to guide language models to produce fairer and non-toxic content without changing the internal representations of the model.
Representing Molecules as Random Walks Over Interpretable Grammars
Michael Sun (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
GenerationData SynthesisExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: The research proposes a data-efficient and interpretable molecular representation and generation method for complex molecular materials.
Reprompting: Automated Chain-of-Thought Prompt Inference Through Gibbs Sampling
Weijia Xu (Microsoft Research), Nebojsa Jojic (Microsoft Research)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Automatically learn chain-of-thought (CoT) prompts through Gibbs sampling without human intervention;
Reservoir Computing for Short High-Dimensional Time Series: an Application to SARS-CoV-2 Hospitalization Forecast
Thomas Ferté (Bordeaux University), Xavier Hinaut (Bordeaux University)
OptimizationHyperparameter SearchTime SeriesSequentialBiomedical DataElectronic Health Records
🎯 What it does: Using Reservoir Computing (RC) combined with Genetic Algorithm (GA) to perform a 14-day short-term prediction of SARS-CoV-2 hospitalizations at the University Hospital of Bordeaux, with training data consisting of a high-dimensional time series of 586 days and 409 features.
Reshape and Adapt for Output Quantization (RAOQ): Quantization-aware Training for In-memory Computing Systems
Bonan Zhang (Princeton University), Naveen Verma (EnCharge AI)
ClassificationObject DetectionTransformerSupervised Fine-TuningImageText
🎯 What it does: The RAOQ framework is proposed to address ADC quantization errors in in-memory computing (IMC) by enhancing the performance of quantized networks on efficient IMC through four techniques: activation shift (A-shift), weight quantization (W-reshape), bit-width augmentation (BitAug), and ADC-LoRA.
Residual Quantization with Implicit Neural Codebooks
Iris A.M. Huijben, Jakob Verbeek (Meta)
RetrievalCompressionImageTextMultimodality
🎯 What it does: A neural network vector quantization method based on residual quantization, QINCO, is proposed, which generates data-related codebooks at each step using neural networks to improve quantization accuracy and support multi-rate compression.
Residual-Conditioned Optimal Transport: Towards Structure-Preserving Unpaired and Paired Image Restoration
Xiaole Tang (Xi'an Jiaotong University), Jian Sun (Xi'an Jiaotong University)
RestorationSuper ResolutionGenerative Adversarial NetworkImage
🎯 What it does: A framework for image restoration based on Residual Conditional Optimal Transport (RCOT) is proposed, viewing image restoration as an optimal transport problem. By introducing transport residuals as degradation-specific cues, a two-step residual conditional mapping is constructed.
Resisting Stochastic Risks in Diffusion Planners with the Trajectory Aggregation Tree
Lang Feng (Zhejiang University), Gang Pan (Zhejiang University)
OptimizationRobotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: A training-free, directly deployable method called Trajectory Aggregation Tree (TAT) is proposed to mitigate the risk of invalid trajectories caused by the randomness of diffusion models in diffusion planners.
REST: Efficient and Accelerated EEG Seizure Analysis through Residual State Updates
Arshia Afzal (École Polytechnique Fédérale de Lausanne), Mahsa Shoaran (École Polytechnique Fédérale de Lausanne)
ClassificationAnomaly DetectionComputational EfficiencyRecurrent Neural NetworkGraph Neural NetworkTime SeriesBiomedical Data
🎯 What it does: This paper proposes a residual state update mechanism based on graph convolution (REST) for efficient real-time epilepsy seizure detection and classification.
Restoring balance: principled under/oversampling of data for optimal classification
Emanuele Loffredo (Laboratoire de physique de l'Ecole normale superieure), Remi Monasson
ClassificationOptimizationConvolutional Neural NetworkGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: This paper provides an exact analytical expression for the generalization performance of linear support vector machines in binary classification tasks using the replica method from statistical mechanics in the high-dimensional limit. Based on this, it evaluates and optimizes under-sampling/over-sampling strategies to achieve class balance.
Rethinking Adversarial Robustness in the Context of the Right to be Forgotten
Chenxu Zhao (Iowa State University), Mengdi Huai (Iowa State University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an attack method that utilizes the machine unlearning process to enhance the weaknesses in adversarial robustness—AdvUA.
Rethinking Data Shapley for Data Selection Tasks: Misleads and Merits
Jiachen T. Wang (Princeton University), Ruoxi Jia (Virginia Tech)
Data-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: This study investigates the effectiveness of Data Shapley in data selection tasks, proposes a theoretical framework to explain its varying performance, and provides empirical evidence.
Rethinking Decision Transformer via Hierarchical Reinforcement Learning
Yi Ma (Tianjin University), Chenjun Xiao (Chinese University of Hongkong)
TransformerReinforcement LearningSequential
🎯 What it does: This paper proposes an Automatic Tuning Decision Transformer (ADT), which achieves the ability to concatenate sub-optimal trajectories in offline data by hierarchically optimizing high-level prompt generation and low-level action decision-making.
Rethinking DP-SGD in Discrete Domain: Exploring Logistic Distribution in the Realm of signSGD
Jonggyu Jang (Pohang University of Science and Technology), Hyun Jong Yang (Pohang University of Science and Technology)
OptimizationFederated LearningSafty and PrivacyConvolutional Neural NetworkTransformerImage
🎯 What it does: In response to DP-SIGNSGD in federated/distributed training, the authors propose using Logistic noise instead of traditional Gaussian noise, and provide corresponding theoretical proofs and algorithm implementation (DP-SIGNLOSGD).
Rethinking Generative Large Language Model Evaluation for Semantic Comprehension
Fangyun Wei (Microsoft Research Asia), Lin Luo (Microsoft Research Asia)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the RWQ-Elo system, which improves traditional MCQA evaluation by incorporating two-person competitions, GPT-4 as a judge, and open-ended responses to better align with real user queries; it also reevaluates 24 LLMs using various evaluation strategies across 11 benchmarks.
Rethinking Guidance Information to Utilize Unlabeled Samples: A Label Encoding Perspective
Yulong Zhang (Southern University of Science and Technology), Jiangang Lu (Zhejiang University)
ClassificationDomain AdaptationImageText
🎯 What it does: A method called Label Encoding Risk Minimization (LERM) is proposed, which extends traditional ERM to utilize unlabeled samples.
Rethinking Independent Cross-Entropy Loss For Graph-Structured Data
Rui Miao (Jilin University), Xin Wang (Jilin University)
ClassificationAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: A joint-cluster supervised learning framework is proposed, improving the traditional node-independent cross-entropy loss by considering the joint distribution of graph nodes and clusters to enhance classification accuracy and robustness.
Rethinking Momentum Knowledge Distillation in Online Continual Learning
Nicolas Michel (Gustave Eiffel University), Toshihiko Yamasaki (University of Tokyo)
Knowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: The study uses Momentum Knowledge Distillation (MKD) in Online Continual Learning (OCL) to enhance model performance and address issues such as teacher quality, quantity, and unknown task boundaries.
Rethinking Optimization and Architecture for Tiny Language Models
Yehui Tang (Huawei Noah's Ark Lab), Yunhe Wang (Huawei)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This study investigates the construction of efficient micro language models through improved neural architecture, parameter initialization, and optimization strategies, and introduces PanGuπ‑1B Pro and 1.5B Pro;
Rethinking Specificity in SBDD: Leveraging Delta Score and Energy-Guided Diffusion
Bowen Gao (Institute for AI Industry Research), Yanyan Lan (Institute for AI Industry Research)
Drug DiscoveryGraph Neural NetworkDiffusion modelContrastive LearningBiomedical Data
🎯 What it does: A Delta Score evaluation metric is proposed, along with an energy-guided diffusion model SBE-Diff based on specific binding energy, to improve the specificity of molecular generation in structure-based drug design (SBDD).
Rethinking the Flat Minima Searching in Federated Learning
Taehwan Lee (Ulsan National Institute of Science and Technology), Sung Whan Yoon (Ulsan National Institute of Science and Technology)
OptimizationFederated LearningImage
🎯 What it does: This paper proposes a new federated learning method called FedGF, which seeks global flat extrema by considering both global and local perturbations during local training.
Rethinking Transformers in Solving POMDPs
Chenhao Lu (Tsinghua University), Huazhe Xu (Tsinghua University)
Recurrent Neural NetworkTransformerReinforcement LearningSequential
🎯 What it does: This paper explores the effectiveness of Transformers in partially observable Markov decision processes (POMDPs) and theoretically and experimentally demonstrates their fundamental limitations in learning tasks related to regular languages, proposing Linear Recursive Units (LRU) as a more suitable sequence model.
Retrieval Across Any Domains via Large-scale Pre-trained Model
Jiexi Yan (Xidian University), Heng Huang (University of Maryland)
RetrievalDomain AdaptationTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: A data unsupervised cross-domain retrieval method TKI is proposed, which learns domain-invariant projections under the condition of no training images by utilizing the text representations of large-scale visual-language pre-trained models.
Retrieval-Augmented Score Distillation for Text-to-3D Generation
Junyoung Seo (Korea University), Seungryong Kim (Korea University)
GenerationRetrievalKnowledge DistillationDiffusion modelScore-based ModelNeural Radiance FieldPoint CloudMeshRetrieval-Augmented Generation
🎯 What it does: This paper proposes a retrieval-enhanced score distillation framework called ReDream, which utilizes 3D assets semantically similar to the text as geometric priors and perspective adaptation of a 2D diffusion model, thereby improving the geometric consistency and detail fidelity of text-to-3D generation.
Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness
Honghao Chen (Chinese Academy of Sciences), Kaiqi Huang (Chinese Academy of Sciences)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper studies the performance of large convolution kernel CNNs in terms of robustness and systematically evaluates their comparison with traditional small convolution CNNs and Vision Transformers;
Revealing Vision-Language Integration in the Brain with Multimodal Networks
Vighnesh Subramaniam (Massachusetts Institute of Technology), Andrei Barbu (Massachusetts Institute of Technology)
Vision Language ModelMultimodalityTime Series
🎯 What it does: Using a pre-trained multimodal deep network to predict human SEEG brain recordings, thereby locating the brain regions involved in visual-language integration.
Revisit the Essence of Distilling Knowledge through Calibration
Wen-Shu Fan (Nanjing University), Le Gan (Nanjing University)
ClassificationKnowledge DistillationImage
🎯 What it does: This study investigates the phenomenon of capacity mismatch in knowledge distillation and provides a unified explanation from the perspective of teacher model calibration.
Revisiting Character-level Adversarial Attacks for Language Models
Elias Abad Rocamora (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes a character-level black-box adversarial attack method named Charmer, which can efficiently generate robust attack samples that maintain semantics and are extremely difficult for humans to detect.
Revisiting Context Aggregation for Image Matting
Qinglin Liu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
RestorationSegmentationTransformerImage
🎯 What it does: This paper re-evaluates the context aggregation mechanism in image matting networks, finding that a basic encoder-decoder network can learn general context aggregation, and based on this, proposes the AEMatter network.
Revisiting Inexact Fixed-Point Iterations for Min-Max Problems: Stochasticity and Structured Nonconvexity
Ahmet Alacaoglu (University of British Columbia), Stephen Wright
OptimizationReinforcement LearningGenerative Adversarial Network
🎯 What it does: This paper studies the theoretical complexity of forward-backward-forward, Halpern, and Krasnosel'skii–Mann iterations for constrained non-convex-concave min-max problems under common supermonotonicity or weak Minty VI conditions, and extends the upper limit of the parameter ρ from the traditional 1/(2L) to 1/L.
Revisiting Scalable Hessian Diagonal Approximations for Applications in Reinforcement Learning
Mohamed Elsayed (University of Alberta), A. Rupam Mahmood (University of Alberta)
OptimizationReinforcement LearningImageSequential
🎯 What it does: A scalable Hessian diagonal approximation method called HesScale is proposed, and it is applied to the second-order optimizer AdaHesScale in supervised learning and a step size scaling scheme in reinforcement learning.
Revisiting the Power of Prompt for Visual Tuning
Yuzhu Wang (Zhejiang Lab), Meng Wang (Hefei University of Technology)
ClassificationRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringImage
🎯 What it does: Based on the pre-trained visual Transformer, a Self-Prompt Tuning (SPT) scheme is proposed, initializing the embedding prototypes of downstream tasks as learnable prompt words, and only updating the prompt words and task head during the training process.
Revisiting the Role of Language Priors in Vision-Language Models
Zhiqiu Lin (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
RetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Utilizes the text generation probability of generative visual language models (VisualGPTScore) for image-text retrieval tasks and proposes a training-free language bias correction method;
Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark
Yihua Zhang (Michigan State University), Tianlong Chen (University of North Carolina at Chapel Hill)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper systematically evaluates the performance of various BP-free methods under different models, tasks, and fine-tuning strategies by constructing a zero-order optimization (Zero-Order, ZO) benchmark for large-scale language model (LLM) fine-tuning, and compares it with traditional first-order (FO) optimization.
Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling
Guoqi Yu (University of Electronic Science and Technology of China), Shujun Wang (Hong Kong Polytechnic University)
Convolutional Neural NetworkTransformerTime Series
🎯 What it does: Predicting multivariate time series through learnable convolutional decomposition and dual attention mechanisms.
Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences
Andi Nika (Max Planck Institute for Software Systems), Adish Singla
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: A systematic theoretical comparison of two mainstream paradigms for learning from human preferences—RLHF (Reward Learning followed by Reinforcement Learning) and DPO (Direct Policy Optimization)—is conducted, providing an analysis of their sample complexity, convergence rates, and dependencies on parameters such as reward/policy dimensions and temperature.
Reward Shaping for Reinforcement Learning with An Assistant Reward Agent
Haozhe Ma (National University of Singapore), Tze-Yun Leong (National University of Singapore)
Reinforcement Learning
🎯 What it does: A dual-agent reward shaping framework called ReLara is proposed, which uses a reward agent to adaptively generate auxiliary rewards to enhance learning efficiency in sparse reward environments.
Reward-Free Kernel-Based Reinforcement Learning
Sattar Vakili (MediaTek Research), Alberto Bernacchia (MediaTek Research)
Reinforcement Learning
🎯 What it does: This paper proposes a reward-free reinforcement learning framework that collects reward-free trajectories during the exploration phase, and then utilizes these trajectories to generate approximately optimal policies under any given reward function during the planning phase.
Rewards-in-Context: Multi-objective Alignment of Foundation Models with Dynamic Preference Adjustment
Rui Yang (Tencent AI Lab), Jianshu Chen (Tencent AI Lab)
Recommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageText
🎯 What it does: The Rewards-in-Context (RiC) method is proposed, which achieves multi-objective alignment of large models through supervised fine-tuning (SFT) by embedding multiple reward information in the prompts, and dynamically adjusts rewards during the inference phase to meet user preferences.
Reweighted Solutions for Weighted Low Rank Approximation
David Woodruff, Taisuke Yasuda (Carnegie Mellon University)
CompressionOptimizationImage
🎯 What it does: A weighted low-rank approximation algorithm based on reweighting a low-rank weight matrix is proposed, where the output approximate matrix does not need to be strictly low-rank, yet can be stored with a small number of parameters and provides an approximation error guarantee.
RICE: Breaking Through the Training Bottlenecks of Reinforcement Learning with Explanation
Zelei Cheng (Northwestern University), Xinyu Xing (Northwestern University)
Autonomous DrivingExplainability and InterpretabilityReinforcement LearningTabular
🎯 What it does: RICE breaks the training bottleneck of reinforcement learning by identifying key states using interpretability methods, constructing a mixed initial state distribution, and incorporating random network distillation exploration.
Rich-Observation Reinforcement Learning with Continuous Latent Dynamics
Yuda Song (Carnegie Mellon University), Akshay Krishnamurthy (Microsoft Research)
Reinforcement Learning
🎯 What it does: The RichCLD framework is proposed, along with algorithms BCRL.C and CRIEE that can achieve sparse observation reinforcement learning under continuous latent dynamics.
Riemannian Accelerated Zeroth-order Algorithm: Improved Robustness and Lower Query Complexity
Chang He (Shanghai University of Finance and Economics), Bo Jiang (Shanghai University of Finance and Economics)
Optimization
🎯 What it does: An improved Riemannian accelerated zeroth-order algorithm is proposed, aimed at enhancing robustness and reducing query complexity.
Riemannian coordinate descent algorithms on matrix manifolds
Andi Han (Riken AIP), Bamdev Mishra (Microsoft India)
OptimizationKnowledge DistillationTabular
🎯 What it does: A general Riemannian Coordinate Descent (RCD) framework has been developed, and efficient coordinate update algorithms have been implemented for various matrix manifolds (such as Stiefel, Grassmann, hyperbolic, symmetric positive definite, etc.); at the same time, a first-order approximation-based gradient reduction version (RCDlin) has been proposed.
Riemannian Preconditioned LoRA for Fine-Tuning Foundation Models
Fangzhao Zhang (Stanford University), Mert Pilanci (Stanford University)
GenerationOptimizationTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes an r×r preconditioner based on Riemannian metrics, which scales the low-rank matrix gradients during the LoRA fine-tuning process, thereby achieving stable feature learning and accelerating convergence.
RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content
Zhuowen Yuan (University of Illinois Urbana-Champaign), Bo Li (University of Chicago)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A multi-level input/output content review framework named RigorLLM has been developed, specifically designed to enhance the safety and robustness of large language models (LLMs).
RIME: Robust Preference-based Reinforcement Learning with Noisy Preferences
Jie Cheng (Chinese Academy of Sciences and University of Chinese Academy of Sciences), Fei-Yue Wang (Chinese Academy of Sciences and University of Chinese Academy of Sciences)
Robotic IntelligenceReinforcement LearningTabular
🎯 What it does: The RIME algorithm is proposed to address the robustness issue caused by noisy labels in preference-based reinforcement learning, enabling the use of imperfect human preferences for training in non-expert or crowdsourced environments.
Risk Aware Benchmarking of Large Language Models
Apoorva Nitsure (IBM Research), Jarret Ross (IBM Research)
Large Language ModelTextBenchmarkFinance Related
🎯 What it does: A distributed risk-aware LLM benchmark framework is proposed, which ranks models based on statistical significance through a combination of metrics (Copula) and stochastic dominance tests.
Risk Estimation in a Markov Cost Process: Lower and Upper Bounds
Gugan Thoppe (Indian Institute of Science), Sanjay P. Bhat (TCS Research)
OptimizationSequentialFinance Related
🎯 What it does: This paper studies the estimation problem of risk measures in Markov cost processes, mainly including the estimation of variance, Value at Risk (VaR), and Conditional Value at Risk (CVaR).
Risk-Sensitive Policy Optimization via Predictive CVaR Policy Gradient
Ju-Hyun Kim (KAIST), Seungki Min (KAIST)
OptimizationReinforcement LearningTime SeriesSequentialFinance Related
🎯 What it does: A method called Predictive CVaR Strategy Gradient (PCVAR) is proposed, which utilizes predictive tail probabilities to weight all trajectories, thereby achieving risk-sensitive policy optimization.
Risk-Sensitive Reward-Free Reinforcement Learning with CVaR
Xinyi Ni (University of California), Lifeng Lai (University of California)
Reinforcement LearningTabular
🎯 What it does: A risk-sensitive reward-free reinforcement learning framework based on CVaR is proposed, with the design of the CVaR-RF-UCRL exploration algorithm and the CVaR-VI / CVaR-VI-DISC planning algorithm, achieving single-stage exploration for arbitrary reward functions.
RL-CFR: Improving Action Abstraction for Imperfect Information Extensive-Form Games with Reinforcement Learning
Boning Li (Tsinghua University), Longbo Huang (Tsinghua University)
Reinforcement LearningSequential
🎯 What it does: This paper proposes the RL-CFR framework, which uses reinforcement learning to dynamically determine action abstraction and combines Counterfactual Regret Minimization (CFR) to solve large-scale incomplete information games.
RL-VLM-F: Reinforcement Learning from Vision Language Foundation Model Feedback
Yufei Wang (Carnegie Mellon University), Zackory Erickson (Carnegie Mellon University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Automatically generate reward functions using preference feedback from visual language foundation models, allowing RL agents to learn various tasks solely based on text task descriptions and visual observations.
RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback
Harrison Lee (Google DeepMind), Sushant Prakash (Google)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper studies the feasibility of using AI-generated preference labels (RLAIF) as a substitute for human preference labels (RLHF) for aligning large language models, and compares the two methods across three types of tasks: summarization, helpful dialogue, and harmless dialogue.
RLVF: Learning from Verbal Feedback without Overgeneralization
Moritz Pascal Stephan (Stanford University), Chelsea Finn (Stanford University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a method called C3PO, which utilizes high-level verbal feedback to customize the behavior of large language models while avoiding overgeneralization in irrelevant contexts.
RMIB: Representation Matching Information Bottleneck for Matching Text Representations
Haihui Pan (Central South University), Kun Han (Cheetah Mobile Inc.)
Domain AdaptationRepresentation LearningTransformerContrastive LearningText
🎯 What it does: The RMIB framework is proposed, which improves the performance of text matching tasks by aligning the prior distributions of heterogeneous domain text representations and combining information bottleneck theory.
RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching
Divya Nori (Massachusetts Institute of Technology), Wengong Jin (Broad Institute of MIT and Harvard)
Protein Structure PredictionGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: The RNA structure and sequence design method RNAFlow, based on flow matching, can generate RNA sequences and three-dimensional conformations that meet the target protein binding site in one go, given the protein structure.
RoboCodeX: Multimodal Code Generation for Robotic Behavior Synthesis
Yao Mu (University of Hong Kong), Ping Luo (University of Hong Kong)
Robotic IntelligenceAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodality
🎯 What it does: RoboCodeX is designed as a multimodal tree-structured code generation framework that transforms high-level instructions into executable robot control code.
RoboDreamer: Learning Compositional World Models for Robot Imagination
Siyuan Zhou (Hong Kong University of Science and Technology), Chuang Gan (University of Massachusetts Amherst)
GenerationRobotic IntelligenceTransformerDiffusion modelWorld ModelVideoTextMultimodality
🎯 What it does: This paper proposes RoboDreamer, which utilizes text parsing to decompose language instructions into action and spatial relationship primitives, and constructs a composable world model for robot imagination and planning.
RoboGen: Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation
Yufei Wang (Carnegie Mellon University), Chuang Gan (UMass Amherst)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelText
🎯 What it does: This paper presents RoboGen, an automated robot learning framework based on Generative Simulation, capable of self-generating tasks, scenarios, training supervision, and learning diverse skills.
RoboMP$^2$: A Robotic Multimodal Perception-Planning Framework with Multimodal Large Language Models
Qi Lv (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: The RoboMP2 framework is proposed, which includes a goal-conditioned multimodal perceiver (GCMP) based on MLLM and a retrieval-augmented multimodal planner (RAMP), enabling robots to perceive and plan complex instructions.
Robust and Conjugate Gaussian Process Regression
Matias Altamirano (University College London), Jeremias Knoblauch (University College London)
Anomaly DetectionOptimizationScore-based ModelTabular
🎯 What it does: A robust and conjugate Gaussian process regression (RCGP) is proposed, achieving robustness to outliers through generalized Bayesian inference while retaining closed-form updates.
Robust Classification via a Single Diffusion Model
Huanran Chen (Beijing Institute of Technology), Jun Zhu (Tsinghua University)
ClassificationComputational EfficiencyAdversarial AttackDiffusion modelImage
🎯 What it does: A robust classifier RDC based on a pre-trained diffusion model is proposed, which directly calculates conditional likelihood using the diffusion model for classification.
Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings for Robust Large Vision-Language Models
Christian Schlarmann (University of Tuebingen), Matthias Hein (University of Tuebingen)
Representation LearningAdversarial AttackTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an unsupervised adversarial fine-tuning method (FARE) that improves the visual encoder of CLIP, enhancing its robustness against image attacks while maintaining its original performance.
Robust Data-driven Prescriptiveness Optimization
Mehran Poursoltani (McGill University), Angelos Georghiou (University of Cyprus)
OptimizationTabular
🎯 What it does: A distributionally robust context optimization framework is designed and solved, directly maximizing the 'prescriptiveness coefficient' to obtain decision strategies that better utilize side information.
Robust Graph Matching when Nodes are Corrupt
Taha Ameen (University of Illinois Urbana-Champaign), Bruce Hajek (University of Illinois Urbana-Champaign)
Graph Neural NetworkGraph
🎯 What it does: The paper studies the graph matching problem in the presence of tampered nodes, proposing two models: weak adversary (random tampering) and strong adversary (observable and arbitrary tampering), and provides the infeasibility, feasibility, and accuracy thresholds for matching.
Robust Inverse Constrained Reinforcement Learning under Model Misspecification
Sheng Xu (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
Safty and PrivacyKnowledge DistillationRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: The research addresses the safety issues of inverse constrained reinforcement learning in the presence of dynamic errors in training and deployment environments, and proposes a Robust Constrained Inference (RCI) framework and an Adaptive Robust Inverse Constrained Reinforcement Learning (AR-ICRL) algorithm.
Robust Inverse Graphics via Probabilistic Inference
Tuan Anh Le (Google), Rif A. Saurous (Google)
RestorationDepth EstimationDiffusion modelNeural Radiance FieldImage
🎯 What it does: This paper proposes a Robust Inverse Graphics (RIG) framework that performs full probabilistic inference on the scene and corruption parameters, utilizing a prior NeRF model to achieve 3D reconstruction and denoising from a single image.
Robust Learning-Augmented Dictionaries
Ali Zeynali (University of Massachusetts Amherst), Mohammad Hajiesmaili (York University)
Text
🎯 What it does: A new skip list structure called RobustSL is constructed, which achieves static optimality when predictions are accurate and maintains O(log n) optimal robustness when predictions are incorrect by utilizing predicted access frequencies.
Robust Multi-Task Learning with Excess Risks
Yifei He (University of Illinois Urbana-Champaign), Han Zhao (University of Illinois Urbana-Champaign)
OptimizationReinforcement LearningImageBenchmark
🎯 What it does: This paper proposes a task weight update method based on 'excess risk' called ExcessMTL, aimed at addressing the imbalance in weight allocation caused by label noise in multi-task learning.
Robust Optimization in Protein Fitness Landscapes Using Reinforcement Learning in Latent Space
Minji Lee (Korea Advanced Institute of Science and Technology), Ho Min Kim (Korea Advanced Institute of Science and Technology)
OptimizationReinforcement LearningPrompt EngineeringBiomedical Data
🎯 What it does: Under the starting point of low fitness protein sequences, a reinforcement learning-based framework called LatProtRL is proposed, which utilizes a pre-trained protein language model to learn latent representations and conducts multi-step exploration in the latent space to optimize protein function (fitness).
Robust Sparse Estimation for Gaussians with Optimal Error under Huber Contamination
Ilias Diakonikolas (University of Wisconsin Madison), Thanasis Pittas (University of Wisconsin Madison)
OptimizationComputational EfficiencyTabular
🎯 What it does: Under the Huber contamination model of high-dimensional Gaussian distribution, the first set of computationally efficient algorithms for sparse mean estimation, sparse PCA, and sparse linear regression that can achieve the information-theoretic limit error O(ε) is proposed;
Robust Stable Spiking Neural Networks
Jianhao Ding (Peking University), Tiejun Huang (Peking University)
ClassificationAdversarial AttackSpiking Neural NetworkImage
🎯 What it does: A robustness metric based on membrane potential disturbance is proposed, and by minimizing its mean square error combined with dynamic LIF neurons, a deep SNN is trained to enhance robustness against adversarial disturbances.
Robust Universal Adversarial Perturbations
Changming Xu (University of Illinois), Gagandeep Singh (VMWare)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a method for generating robust universal adversarial perturbations (RobustUAP) against real-world transformations, validated on ImageNet and CIFAR-10.
Robust Yet Efficient Conformal Prediction Sets
Soroush H. Zargarbashi (CISPA Helmholtz Center for Information Security), Aleksandar Bojchevski (University of Cologne)
Computational EfficiencyAdversarial AttackGraph Neural NetworkGaussian SplattingImageGraph
🎯 What it does: This paper proposes a conformal prediction set that is robust to adversarial attacks (CAS), achieving a more compact and coverage-guaranteed set by aligning the upper bounds of distribution functions.
Robustly Learning Single-Index Models via Alignment Sharpness
Nikos Zarifis (University of Wisconsin Madison), Jelena Diakonikolas (University of Wisconsin Madison)
Optimization
🎯 What it does: An efficient learning algorithm is proposed for the Single-Index Model (SIM) under a completely agnostic label setting, which can approximate the optimal loss with a constant factor under the assumptions of 'well-behaved distribution' and 'well-approximable' activation function families.
Robustness of Deep Learning for Accelerated MRI: Benefits of Diverse Training Data
Kang Lin (Technical University of Munich), Reinhard Heckel (Technical University of Munich)
RestorationDomain AdaptationConvolutional Neural NetworkTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study explores the impact of diversified training data on the robustness of accelerated magnetic resonance imaging (MRI) reconstruction models by training deep learning models on various distributions (different anatomical structures, scanners, magnetic field strengths, image contrasts, etc.) and verifies its universality across different network architectures (U-net, ViT, VarNet).
Robustness of Nonlinear Representation Learning
Simon Buchholz (Max Planck Institute for Intelligent Systems), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Representation Learning
🎯 What it does: This study investigates the identifiability and robustness issues of nonlinear representation learning under slight errors (mixed function approximations of local isometry or mildly perturbed linear ICA), providing theoretical proofs and error bounds.
RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples
Hossein Mirzaei (Sharif University of Technology), Mohammad Hossein Rohban (Sharif University of Technology)
Anomaly DetectionDiffusion modelImageMagnetic Resonance ImagingComputed Tomography
🎯 What it does: A robust anomaly detection framework RODEO is proposed, which enhances the model's robustness by adaptively generating external samples that are close to the original distribution and diverse.
Rolling Diffusion Models
David Ruhe (Google Deepmind), Emiel Hoogeboom (Google Deepmind)
GenerationData SynthesisDiffusion modelVideoTime SeriesPhysics Related
🎯 What it does: A rolling diffusion model that adaptively adds noise in time series (Rolling Diffusion) is proposed, which denoises frame by frame using a sliding window;
Roping in Uncertainty: Robustness and Regularization in Markov Games
Jeremy McMahan (University of Wisconsin Madison), Qiaomin Xie (University of Wisconsin Madison)
OptimizationReinforcement Learning
🎯 What it does: This paper studies robust Markov games (RMG) with s-rectangular uncertainty and proves that computing its robust Nash equilibrium (RNE) is equivalent to solving the corresponding regularized Markov game (MPNE), thus providing a planning method that can directly utilize existing regularization algorithms.
RoSA: Accurate Parameter-Efficient Fine-Tuning via Robust Adaptation
Mahdi Nikdan (IST Austria), Dan Alistarh (Neural Magic)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A new parameter-efficient fine-tuning method called RoSA is proposed, which approximates full fine-tuning results by combining low-rank and sparse adapters.
Rotational Equilibrium: How Weight Decay Balances Learning Across Neural Networks
Atli Kosson (École Polytechnique Fédérale de Lausanne), Martin Jaggi (École Polytechnique Fédérale de Lausanne)
OptimizationImageText
🎯 What it does: The study investigates how weight decay affects the updates of individual neurons, introducing the concept of 'rotational balance' and explaining why AdamW outperforms Adam+ℓ2;