ICML 2026 Papers — Page 50
International Conference on Machine Learning · 6554 papers
RGGT: A Generative-Prior-Guided Transformer for Unified Rigid and Non-Rigid Point Cloud Registration
Chengyu Zheng (Nanjing University of Aeronautics and Astronautics), Mingqiang Wei (Nanjing University of Aeronautics and Astronautics)
Pose EstimationOptimizationTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningPoint Cloud
🎯 What it does: Proposed a model that can simultaneously perform rigid and non-rigid point cloud registration within the same framework
RGMem: Renormalization Group–inspired Memory Evolution for Language Agents
Ao Tian (Beihang University), Yanfang Liu (Beihang University)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the RGMem framework, modeling dialogue memory as a multi-scale self-evolving system, achieving long-term personalization through hierarchical coarsening, threshold updating, and separation of slow/fast variables.
Rh-3DGS: Robust Open-Vocabulary Scene Understanding via Riemannian Huber Distillation and Manifold-Aware Sampling
Xinpeng Zhao (National University of Defense Technology), Guangzhen Yao (National University of Defense Technology)
SegmentationKnowledge DistillationRepresentation LearningTransformerContrastive LearningGaussian SplattingImageMultimodalityPoint Cloud
🎯 What it does: Built the Rh-3DGS framework, utilizing 3D Gaussian Splatting to achieve open-vocabulary 3D semantic understanding.
RiboSphere: Learning Unified and Efficient Representations of RNA Structures
Zhou Zhang (Nanjing University), Tianfan Fu (Nanjing University)
GenerationData SynthesisRepresentation LearningProtein Structure PredictionTransformerSupervised Fine-TuningDiffusion modelFlow-based ModelAuto EncoderContrastive LearningGraphBiomedical Data
🎯 What it does: Developed a RNA structure representation and generation framework called RiboSphere based on vector quantization and flow matching.
Richer Bayesian Last Layers with Subsampled NTK Features
Sergio Calvo Ordoñez, Kamil Ciosek (Spotify)
ClassificationAnomaly DetectionOptimizationExplainability and InterpretabilityComputational EfficiencyNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingImageTabularTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose an scalable 'Rich-BLL' method, which improves the confidence estimation of traditional Bayesian last layer by projecting full NTK features into the last layer feature space, and supports uniform subsampling to significantly reduce computational cost.
Riemannian Diffusion Models on General Manifolds via Physics-Informed Neural Networks
Gyeonghoon Ko (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisGraph Neural NetworkDiffusion modelScore-based ModelImagePoint CloudMeshGraphTabularPhysics Related
🎯 What it does: This paper proposes to approximate the heat kernel on a Riemannian manifold using physics-informed neural networks, thereby achieving a general drift-free Riemannian diffusion model.
Riemannian Dueling Optimization
Yuxuan Ren (Rice University), Shiqian Ma (Texas A&M University)
OptimizationAdversarial AttackImageTabular
🎯 What it does: Proposes a framework for optimization on Riemannian manifolds using only pairwise feedback, and presents three algorithms: RDNGD (Riemannian Dueling Normalized Gradient Descent), RRDNGD (cyclic linearly convergent version), and RDFW (projection-free Frank-Wolfe), along with theoretical upper bounds on the number of iterations and comparisons.
Riemannian MeanFlow
Dongyeop Woo (Korea Advanced Institute of Science and Technology), Sungsoo Ahn (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisProtein Structure PredictionPrompt EngineeringDiffusion modelFlow-based ModelRectified FlowAuto EncoderGraphTabularTime SeriesBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the Riemannian MeanFlow (RMF) framework, which learns flow graphs on Riemannian manifolds, achieving first-order or few-step generation, significantly reducing the number of inference steps;
Riemannian MeanFlow for One-Step Generation on Manifolds
Zichen Zhong (Shandong University), Yilong Yin (Shandong University)
GenerationData SynthesisDiffusion modelScore-based ModelFlow-based ModelPoint CloudGraphTabularTime SeriesBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes Riemannian MeanFlow (RMF), a generative model that performs single-sample generation on Riemannian manifolds by defining an average velocity field and using parallel transport to achieve trajectory-free simulation training.
Riemannian Metric Matching for Scalable Geometric Modeling of Distributions
Jacob Bamberger (University of Oxford), Iolo Jones (University of Oxford)
OptimizationComputational EfficiencyRepresentation LearningTransformerDiffusion modelScore-based ModelContrastive LearningImagePoint CloudMeshGraphTabular
🎯 What it does: Designed an image-free, scalable Riemannian metric learning framework that directly regresses the carré du champ (CDC) of data via neural networks, achieving an unbiased estimation of the intrinsic geometry of high-dimensional data;
Riemannian Networks over Full-Rank Correlation Matrices
Ziheng Chen (University of Trento), Nicu Sebe (University of Trento)
ClassificationRecognitionConvolutional Neural NetworkRecurrent Neural NetworkTransformerImageVideoAudio
🎯 What it does: This paper systematically migrates multi-class logistic regression, fully connected layers, and convolutional layers to the Riemannian geometry of full-rank correlation matrices, and constructs corresponding Correlation Networks (CorNets);
Riemannian Neural Optimal Transport
Alessandro Micheli (Imperial College London), Samir Bhatt (University of Copenhagen)
OptimizationSpiking Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningPoint CloudMeshGraphPhysics Related
🎯 What it does: This paper proposes Riemannian Neural Optimal Transport (RNOT), a framework for achieving continuous neural network optimal transport on Riemannian manifolds, avoiding the dimensionality curse caused by traditional discretization methods.
Riemannian Optimization for Fair Spectral Clustering
Minh Phu Vuong (Texas State University), Chul-Ho Lee (Texas State University)
OptimizationGraph
🎯 What it does: Propose a fair spectral clustering algorithm based on Riemannian geometry optimization, R-FairSC, which avoids expensive feature decomposition and addresses the scalability issue in fair clustering.
Riemannian stochastic optimization for sufficient dimension reduction
Thibault Pautrel (CentraleSupélec), François Portier (ENSAI)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningTabular
🎯 What it does: A new algorithm called SMAVE is proposed for sufficient dimension reduction (SDR), achieving low-dimensional projection for high-dimensional regression by performing Riemannian stochastic optimization on the Stiefel manifold.
Ripple Perturbations Through Structure: Likelihood-Constrained Adversarial Attacks on Heterogeneous Tabular Data
Zhengjie Zhou (Ant Group), Weiqiang Wang (Ant Group)
OptimizationAdversarial AttackGraph Neural NetworkTransformerGenerative Adversarial NetworkContrastive LearningTabularBenchmark
🎯 What it does: Proposed a white-box adversarial attack framework called LCSA, which utilizes structural causal models to constrain the generation of adversarial perturbations in tabular data.
Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility
Mengxuan Wang (South China University of Technology), Ming Li (Guangdong Laboratory of Artificial Intelligence and Digital Economy)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Proposes Risk Awareness Injection (RAI), a lightweight, training-free framework that enhances the VLM's perception and rejection of unsafe content by injecting unsafe signals into the initial layer of visual-text fusion.
Risk-Averse and Optimistic Advertiser Incentive Compatibility in Auto-bidding
Christopher Liaw (Google Research), Wennan Zhu (Google Research)
OptimizationFederated LearningReinforcement Learning from Human FeedbackFinance Related
🎯 What it does: This paper studies the problem of advertisers reporting their constraints in an automatic bidding environment, proposing more relaxed concepts of incentive compatibility, RAIC and OAIC, and proving that the second-price auction satisfies both types of compatibility in both non-uniform and uniform bidding scenarios.
Risk-Bounded Distribution Reconstruction: Stable Statistic Calibration for Long-Tailed Recognition
Guanliang Liu (Xidian University), Bo Chen (Xidian University)
ClassificationRecognitionDiffusion modelContrastive LearningImage
🎯 What it does: In the long-tailed classification task, we propose an offline statistical calibration framework called RBDR, which safely reconstructs the mean and covariance of tail classes and reconstructs the distribution in the feature space under frozen features;
RiskZero: Plan More to Risk Less with a Learned Model
Yousef Yassin (Carleton University), Junfeng Wen (Carleton University)
OptimizationGraph Neural NetworkTransformerReinforcement LearningMixture of ExpertsContrastive LearningWorld ModelGraphTabularTime SeriesSequential
🎯 What it does: Proposed and implemented RiskZero, a model-agnostic risk-aware decision-making framework based on MuZero, which can learn distributed returns over complete trajectories and perform risk-sensitive planning in unknown environmental dynamics.
RL with Learnable Textual Feedback: A Bilevel Approach
Utsav Singh (University of Central Florida), Amrit Singh Bedi (University of Central Florida)
OptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsText
🎯 What it does: This paper proposes a dual-layer natural language actor-critic framework (Bi-NAC), which enables the actor (LLM) to self-improve after receiving feedback by embedding learnable textual feedback into the reinforcement learning loop;
RL-SPH: Learning to Achieve Feasible Solutions for Integer Linear Programs
Tae-Hoon Lee (KAIST), Min-Soo Kim (KAIST)
OptimizationGraph Neural NetworkTransformerReinforcement LearningGraphTabularBenchmark
🎯 What it does: Proposed a reinforcement learning-based starting primal heuristic RL-SPH, capable of independently generating feasible solutions for integer linear programming (ILP).
RL4RLA: Teaching ML to Discover Randomized Linear Algebra Algorithms Through Curriculum Design and Graph-Based Search
Jinglong Xiong (Duke University), Yaoqing Yang (Dartmouth College)
OptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement LearningTabular
🎯 What it does: Automatically discover interpretable, symbolic randomized linear algebra (RLA) algorithms through a reinforcement learning framework, constructing programs composed of fundamental linear algebra primitives;
RLAnything: Forge Environment, Policy, and Reward Model in Completely Dynamic RL System
Yinjie Wang (University of Chicago), Ling Yang (Princeton University)
OptimizationTransformerLarge Language ModelReinforcement LearningWorld ModelTextMultimodality
🎯 What it does: Designed and implemented RLAnything, a reinforcement learning framework that dynamically optimizes the environment, policy, and reward model in a closed-loop.
RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks
Hanbo Huang (Shanghai Jiao Tong University), Shiyu Liang (Shanghai Jiao Tong University)
Adversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Studied the vulnerability of large language model (LLM) watermarks under worst-case scenarios (adaptive attacks), and proposed an attack method based on reinforcement learning called RLCracker, which can efficiently remove multiple watermarks without detector access and with only a small number of labeled samples.
RLIE: Rule Generation with Logistic Regression, Iterative Refinement, and Evaluation for Large Language Models
Yang Yang (Hong Kong University of Science and Technology), Yutao Yue (Hong Kong University of Science and Technology)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Generate natural language rules using LLM and combine them with weighted logistic regression to build an interpretable neuro-symbolic reasoning framework called RLIE.
RLSF-V: Mitigating Hallucinations in MLLMs via Fuzzy Semantic Self-Feedback
Changhao He (Sichuan University), Peng Hu (Sichuan University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Propose the RLSF-V framework, which reduces hallucinations in multi-modal large language models by evaluating the fuzzy semantics of internally generated logit-based preference data.
RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments
Zhiyuan Zeng (University of Washington), Hannaneh Hajishirzi (University of Washington)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes RLVE (Reinforcement Learning with Adaptive Verifiable Environments), a reinforcement learning framework that dynamically adjusts the difficulty distribution using verifiable environments, aiming to enhance the reasoning capabilities of large language models.
RMNP: Row-Momentum Normalized Preconditioning for Scalable Matrix-Based Optimization
Shenyang Deng (Dartmouth College), Yaoqing Yang (Dartmouth College)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Proposed a new optimizer called RMNP, which replaces the Newton-Schulz iteration in MUON with ℓ2 normalization of row vectors, achieving more efficient computation for matrix-level preprocessing.
RN-D: Discretized Categorical Actors for On-Policy Reinforcement Learning
Yuexin Bian (University of California San Diego), Yuanyuan Shi (University of California San Diego)
Recurrent Neural NetworkTransformerReinforcement LearningTabularTime Series
🎯 What it does: Propose a discretized categorical actor (RN-D), which uses discrete action bins instead of traditional Gaussian distributions and combines them with regularized residual networks to achieve on-policy reinforcement learning for continuous control tasks.
RNA-FM: Flow-Matching Generative Model for Genome-wide RNA-Seq Prediction
Yaxuan Song (University of Sydney), Weidong Cai (University of Sydney)
GenerationData SynthesisExplainability and InterpretabilityTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowImageTabularBiomedical DataComputed TomographyReview/Survey Paper
🎯 What it does: Batch Expression Prediction Based on Whole Genome RNA-Seq
ROAMM: A Benchmark Dataset for Multimodal Human Attention Decoding and EEG-to-Text Modeling During Naturalistic Reading
Haorui Sun (University of Vermont), David C. Jangraw (University of Vermont)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningRecurrent Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningMultimodalityTime SeriesBiomedical DataBenchmark
🎯 What it does: This paper proposes and publicly releases the ROAMM dataset, which records EEG and eye movement synchronized data of 44 subjects during natural multi-page reading tasks, along with page comprehension scores and fine-grained flow drift (MW) annotations.
RoboFlow4D: A Lightweight Flow World Model Toward Real-Time Flow-Guided Robotic Manipulation
Sixu Lin (School of Data Science, Chinese University of Hong Kong (Shenzhen)), Guiliang Liu (School of Data Science, Chinese University of Hong Kong (Shenzhen))
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelFlow-based ModelWorld ModelOptical FlowImageVideoTextMultimodality
🎯 What it does: Developed a lightweight flow world model called RoboFlow4D, which can end-to-end predict multi-frame 3D flow and use this flow to plan and guide robots in real-time operations.
RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies
Yinpei Dai (University of Michigan), Joyce Chai (University of Michigan)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsVision Language ModelVision-Language-Action ModelImageVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built the RoboMME benchmark (containing 16 long-term, non-Markovian robotic manipulation tasks) and designed a series of VLA (vision-language-action) models based on π⁵, systematically evaluating the performance of different memory representations and integration strategies across multiple memory dimensions (temporal, spatial, object, procedural).
RoboOmni: Actions Are Just Another Modality for Vision-Language Models
Dong Wang (Tsinghua University), Huaping Liu (Tsinghua University)
Computational EfficiencyRepresentation LearningRobotic IntelligenceTransformerPrompt EngineeringMixture of ExpertsVision Language ModelVision-Language-Action ModelDiffusion modelContrastive LearningImageVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: The study treats action as another modality in Vision-Language Models (VLM), proposing a unified multimodal next-token prediction framework called RoboOmni, and achieving parallel multi-step action prediction through Multi-Token Action Prediction (MTAP).
RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation
Tianxing Chen (HKU), Yao Mu (SJTU)
Data SynthesisRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningImageVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes RoboTwin 2.0, an expandable dual-arm grasping data generation framework and benchmark that combines closed-loop expert synthesis, domain randomization, and simulation verification.
RobuQ: Pushing DiTs to W1.58A2 via Robust Activation Quantization
Kaicheng Yang (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
GenerationTransformerDiffusion modelImage
🎯 What it does: Propose a quantization framework named RobuQ, which can achieve low-bit weight and activation quantization at the W1.58A2 level on Diffusion Transformer (DiT), while maintaining image generation quality.
Robust AI Evaluation through Maximal Lotteries
Hadi Khalaf (Harvard University), Ariel D. Procaccia (Harvard University)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a language model evaluation framework based on robust lottery, designed to replace the traditional Bradley-Terry ranking, and can provide a more stable set of winners under diverse user preferences.
Robust and Consistent Ski Rental with Distributional Advice
Jihwan Kim (Seoul National University), Chenglin Fan (Seoul National University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabularTime Series
🎯 What it does: This paper proposes a unified framework that integrates complete distributed predictions into the decision-making process of the ski rental problem, designing a robust and consistent deterministic threshold policy (Clamp Policy) as well as a randomization policy based on water-filling, which balances prediction errors and optimal expected competitive ratios.
Robust Bayes-Assisted Conformal Prediction
Kianoosh Ashouritaklimi (University of Oxford), Francois Caron
Anomaly DetectionOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningRobotic IntelligenceReinforcement Learning from Human FeedbackScore-based ModelContrastive LearningGaussian SplattingImageTabularTime SeriesSequentialBiomedical DataBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose the RoBAS (Robust Bayes-Assisted Shrinkage) framework, which constructs Bayesian-assisted inconsistency scores in the residual layer. This framework can leverage accurate prior information to generate efficient prediction intervals, and automatically degenerate into robust DTA scores when the prior conflicts with the data.
Robust Bayesian Optimisation with Unbounded Corruptions
Abdelhamid Ezzerg (University College London), Jeremias Knoblauch (University College London)
OptimizationReinforcement LearningImageTabularTime Series
🎯 What it does: Propose a robust Bayesian optimization algorithm (RCGP-UCB), which can achieve sublinear cumulative reward under noise pollution with frequency limitations and unlimited amplitude, and maintains the same performance as traditional GP-UCB when there is no pollution.
Robust Causal Discovery in Real-World Time Series with Power-Laws
Matteo Tusoni (Sapienza University of Rome), Novella Bartolini (Sapienza University of Rome)
Time Series
🎯 What it does: We propose a causal discovery method called PLaCy, which restores causal structures in real-world time series data with non-stationary and noisy characteristics by fitting power-law to the time series spectrum and applying Granger causality tests on frequency domain features.
Robust Contextual Optimization with Missing Covariates
Qingyuan Xu (University of Michigan), Ruiwei Jiang (University of Michigan)
OptimizationData-Centric LearningTabularFinance Related
🎯 What it does: This paper proposes a distributionally robust contextual optimization framework under covariate missingness scenarios, directly utilizing partially observed covariates and the missing mechanism for decision-making, without performing missing value imputation.
Robust Cross-Modal Retrieval via Generative Semantic Refinement and Exclusion-Guided Adaptation
Qin Yang (Xidian University), Xinbo Gao (Xidian University)
RetrievalDomain AdaptationTransformerLarge Language ModelPrompt EngineeringGenerative Adversarial NetworkContrastive LearningImageTextMultimodality
🎯 What it does: Propose the ReEx framework, which achieves robust handling of structural noise during test time adaptation in cross-modal retrieval by using text generation for semantic correction and exclusion-guided proxy contrastive learning.
Robust Federated Learning Against Adaptive Compression
Wenjing Yan (Chinese University of Hong Kong), Ying-Jun Angela Zhang (Chinese University of Hong Kong)
OptimizationFederated LearningComputational EfficiencyData-Centric LearningImage
🎯 What it does: Propose two communication-efficient federated learning algorithms without parameter tuning—ParFreFL and ComParFreFL—which achieve robust training on heterogeneous data and partial client participation in both uncompressed and compressed scenarios.
Robust Filter Attention: Self-Attention as Precision-Weighted State Estimation
Peter Racioppo (Independent Researcher)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerTextBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose Robust Filter Attention (RFA), treating self-attention as robust state estimation under linear SDE.
Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models
Yanchen Yin (Beijing University of Posts and Telecommunications), Linghui Li (Beijing University of Posts and Telecommunications)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This work reveals that the success of jailbreak attacks mainly occurs through the suppression of specific attention heads, while the safety features (Robust Harmful Features) in the model's intermediate layers remain active, by conducting a fine-grained analysis of the internal mechanisms of large language models; based on this, a training-free detection method is proposed.
Robust Human-AI Complementarity under Uncertainty
Yewon Byun (Carnegie Mellon University), Bryan Wilder (Carnegie Mellon University)
OptimizationFederated LearningExplainability and InterpretabilityRobotic IntelligenceReinforcement Learning from Human FeedbackLarge Language ModelTextTime SeriesBenchmark
🎯 What it does: The study investigates whether human-AI collaboration can achieve complementary gains under uncertain AI quality, and proposes corresponding robust decision rules.
Robust In-Context Reinforcement Learning Under Reward Poisoning Attacks
Paulius Sasnauskas (University of Alberta), Goran Radanovic (MPI-SWS)
Adversarial AttackMeta LearningTransformerReinforcement LearningSequential
🎯 What it does: This study investigates the robustness of context learning reinforcement learning (ICRL) based on Transformer models against reward poisoning attacks, and proposes an adversarial training Decision-Pretrained Transformer (AT-DPT) framework that enables the model to learn near-optimal actions even when facing attacked reward information during testing;
Robust Inter-Series Dependency Modeling for Time Series Forecasting via Information-Theoretic Alignment
Wuqing Yu (Beijing Normal University), Jiacai Zhang (Beijing Normal University)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchGraph Neural NetworkTransformerGraphTabularTime Series
🎯 What it does: Proposed the CGTFra framework, achieving deep consistency modeling between self-attention and graph neural networks in multivariate time series forecasting, addressing the insufficient attention to deep IVD by traditional Variate Transformer.
Robust Learning via Nested Distributionally Robust Optimization
Jinyi Huang (Tongji University), Guodong Shi (University of Sydney)
Domain AdaptationAnomaly DetectionOptimizationImage
🎯 What it does: Proposes a nested distributionally robust optimization (Nested DRO) framework to simultaneously address geometric perturbations and statistical anomalies.
Robust Linear Dueling Bandits with Post-serving Context under Unknown Delays and Adversarial Corruptions
Youngmin Oh (InfiniTree)
Reinforcement LearningAuto EncoderContrastive LearningGaussian SplattingOptical FlowTabularTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes the RCDP-UCB algorithm, which addresses the triple challenges of posterior context, unknown delay, and adversarial corruption present in the linear adversarial dueling bandit problem.
Robust Multi-View Fusion via Prototype-Anchored Unbalanced Optimal Transport
Zhang Han, HUI LI
RecognitionContrastive LearningVideoMultimodalityBenchmark
🎯 What it does: This paper proposes a prototype-anchored unbalanced optimal transport (UOT) framework for robust multi-view fusion, which can automatically trim unreliable information when views are missing or damaged.
Robust Parallel Diffusion Sampling via Dynamic Jacobian Bandwidth
Zile Huang (University of Central Florida), Ser-Nam Lim (University of Central Florida)
GenerationComputational EfficiencyDiffusion modelScore-based ModelImageVideoText
🎯 What it does: Proposes ROPA (Robust Parallel Diffusion Sampling), achieving stable parallel diffusion sampling through adaptive Jacobian bandwidth, damped least squares updates, and score-aligned low-rank preconditioning.
Robust Self-reflective Hashing for Cross-modal Retrieval with Noisy Label
Hao Sun (Qufu Normal University), Lei Huang (Ocean University of China)
RetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the Robust Self-reflective Hashing (RSH) framework, which is used to learn robust hash codes in cross-modal retrieval tasks with noisy labels.
Robust Sequential Experimental Design for A/B Testing
Qianglin Wen (Yunnan University), Hongtu Zhu (University of North Carolina at Chapel Hill)
OptimizationReinforcement LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: Propose a robust sequential experimental design framework under model misspecification, compatible with context bandit and dynamic settings with delayed effects.
Robust Signal Enhancement via Fractional Detail Views and Knowledge Guided Multi-view Fusion
Zikun Jin (Shanxi University), Haijun Geng (Shanxi University)
RestorationConvolutional Neural NetworkDiffusion modelScore-based ModelFlow-based ModelContrastive LearningTime SeriesAudio
🎯 What it does: A novel signal enhancement framework called FracKGMF is proposed, integrating Fractional Distance Decay Convolution (FracConv) and Wiener-guided multi-view fusion (KGMF). It captures detailed information through structured distance decay convolution and adaptively weights two views using reliability priors, achieving robust voice and electromagnetic signal denoising under extremely low SNR conditions.
Robust Stochastic Gradient Posterior Sampling with Lattice Based Discretisation
Zier Mensch (University of Amsterdam), Miranda C. N. Cheng
OptimizationData-Centric LearningScore-based ModelTextTabularStochastic Differential Equation
🎯 What it does: Propose a stochastic gradient Markov chain sampling method called SGLRW based on lattice random walks, to address the instability of traditional SGLD under gradient noise from small batches.
Robust Strategic Classification under Decision-Dependent Cost Uncertainty
Sura Alhanouti (Ohio State University), Parinaz Naghizadeh (University of California)
ClassificationOptimizationTabular
🎯 What it does: Proposed a two-stage robust optimization framework to account for decision-related cost uncertainty in strategic classification.
Robust Vision-Language Models via Manifold-Adversarial Adapters
Hao Li (Hefei University of Technology), Wei Jia (Hefei University of Technology)
RestorationDomain AdaptationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerVision Language ModelGenerative Adversarial NetworkContrastive LearningImageTextMultimodality
🎯 What it does: Propose the Manifold-Adversarial Adapter (MAA), which introduces a parameter-efficient hierarchical adapter on a frozen visual encoder. It utilizes feature distillation and adversarial manifold constraints to correct corrupted features, thereby enhancing the robustness of Vision-Language Models (VLMs) under natural image degradation.
Robust-U1: Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding?
Jiaqi Tang (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
RecognitionImage TranslationRestorationAnomaly DetectionTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelFlow-based ModelRectified FlowImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose the Robust-U1 framework, enabling multi-modal large language models to self-recover corrupted visual content and enhance robustness through multi-modal reasoning.
Robustifying Vision-Language Models via Test-Time Prompt Adaptation
Xingyu Zhu (National University of Singapore), Long Chen (Hong Kong University of Science and Technology)
ClassificationRepresentation LearningAdversarial AttackTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Under adversarial attacks, prompt adaptation is utilized with pre-trained Vision-Language models (CLIP) during testing to enhance the robustness of zero-shot reasoning.
Robustness of Mixtures of Experts to Feature Noise
Dong Sun (CISPA Helmholtz Center for Information Security), Rebekka Burkholz (CISPA Helmholtz Center for Information Security)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningMixture of ExpertsImageTextTabularSequential
🎯 What it does: This paper studies the robustness and learning efficiency of Mixture of Experts (MoE) in feature noise environments, combining theoretical analysis with experiments based on linear models.
RoCA: Robust Cross-Domain End-to-End Autonomous Driving
Rajeev Yasarla (Qualcomm AI Research, an initiative of Qualcomm Technologies, Inc), Fatih Porikli (Qualcomm AI Research, an initiative of Qualcomm Technologies, Inc)
Domain AdaptationAutonomous DrivingTransformerReinforcement LearningAuto EncoderContrastive LearningVideoPoint Cloud
🎯 What it does: Propose the RoCA framework, which uses Gaussian processes to perform probabilistic modeling of tokens from the ego vehicle and surrounding vehicles, thereby achieving robustness and unsupervised adaptation for end-to-end cross-domain driving.
Role-Level Inductive Bias for Cross-Task Generalization in Multi-Agent Reinforcement Learning
Chang Yao (Beijing Jiaotong University), Kai Lv (Beijing Jiaotong University)
Reinforcement Learning
🎯 What it does: Propose the GTR framework based on role-based middle-layer inductive bias, which automatically discovers transferable roles using a Gaussian Mixture Model and achieves cross-task generalization
Romberg-Extrapolated Zeroth-Order Gradient Estimator: Higher-Order Bias Reduction with Preserved Leading Directional Variance
Hongcheng Dong (Shenzhen International Center for Industrial and Applied Mathematics), Feng Yin (Chinese University of Hong Kong)
OptimizationFederated LearningHyperparameter SearchData-Centric LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextMultimodalityTabularTime SeriesSequential
🎯 What it does: Propose a zeroth-order gradient estimator using Romberg extrapolation (Romberg-ZOGE), which achieves high-order truncation error elimination by using the same perturbation direction under geometrically decreasing radii.
Root Cause Analysis of Failures in Microservices via Bayesian Root Cause Discovery
Kenneth Lee (Purdue University), Murat Kocaoglu (Johns Hopkins University)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTabularTime Series
🎯 What it does: Propose a Bayesian inference-based root cause discovery algorithm for microservices (BRCD), which uses the partial causal structure obtained during the observation period to perform posterior updates on the data after failure, thereby identifying the root cause.
Rooted Absorbed Prefix Trajectory Balance with Submodular Replay for GFlowNet Training
Xi Wang (New York University), Shenji Wan
GenerationDrug DiscoveryAI Code AssistantTransformerLarge Language ModelReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelTextSequentialBiomedical DataBenchmark
🎯 What it does: Proposed two mechanisms, Rooted Absorbed Prefix Trajectory Balance (RapTB) and Submodular Replay (SubM), to address the mode collapse, prefix collapse, and length bias issues of GFlowNet in terminable prefix trees (such as LLM generation tasks).
Rotary Position Encodings for Graphs
Isaac Reid (University of Cambridge), Petar Veličković (Google DeepMind)
Representation LearningGraph Neural NetworkTransformerContrastive LearningPoint CloudGraph
🎯 What it does: Proposes Wave-Induced Rotary Encodings (WIRE), a method that generalizes rotary position encodings (RoPE) to graph data;
Rotation-Invariant Spherical Watermarking via Third-Order SO(3) Representation Coupling
Pengzhen Chen (Institute of Information Engineering Chinese Academy of Sciences), Weiping Wang (Institute of Information Engineering Chinese Academy of Sciences)
Image TranslationRestorationRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkDiffusion modelAuto EncoderContrastive LearningImagePhysics Related
🎯 What it does: Proposes the TRIAD framework to achieve rotation-invariant digital watermark embedding and extraction for 360° panoramic images, utilizing the third-order coupling (bispectrum) of SO(3) to ensure watermark recovery under arbitrary three-dimensional rotations;
RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing
Ilya Lasy (TU Wien), Kola Ayonrinde (UK AI Security Institute)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelMixture of ExpertsAuto EncoderText
🎯 What it does: Propose the RouterInterp method, which uses sparse autoencoders (SAE) to extract expert routing-related features and generates readable routing explanations through natural language interpretation and aggregation;
Routing and Reasoned Evaluation with Large Language Models
Guiyao Tie (Huazhong University of Science and Technology), Lichao Sun (Lehigh University)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the R Eval 2 framework, constructed a multi-task, difficulty-tiered comparative evaluation set, and implemented LLM and LRM discriminator routing based on budget constraints.
Routing by Reaching: Composition of Pre-trained GFlowNets for Multi-Objective Generation
Seokwon Yoon (POSTECH), Dongwoo Kim (POSTECH)
GenerationDrug DiscoveryGraph Neural NetworkReinforcement LearningFlow-based ModelGraphTabular
🎯 What it does: Propose a method that mixes the forward strategies of pre-trained GFlowNet during inference, achieving multi-objective generation without joint training;
Row-Stochastic Matrices Can Provably Outperform Doubly Stochastic Matrices in Decentralized Learning
Bing Liu (Zhejiang University), Chengcheng Zhao (Zhejiang University)
OptimizationFederated LearningImageTabular
🎯 What it does: This paper studies that row-stochastic matrices converge faster than doubly-stochastic matrices in decentralized learning with non-uniform node weights.
RQ-MoE: Residual Quantization via Mixture of Experts for Efficient Input-Dependent Vector Compression
Zhengjia Zhong (Xiamen University), Hui Li (Xiamen University)
CompressionTransformerMixture of ExpertsAuto EncoderTabularSequentialBenchmark
🎯 What it does: Propose a residual quantization framework named RQ-MoE, combining two-layer Mixture of Experts and dual-stream quantization, achieving vector compression through input-adaptive dynamic codebooks.
RSA-CP: Efficient Conformal Prediction in Small-Sample Regimes via Random Score Alignment
Pankaj Bhagwat (University of Alberta), Linglong Kong (University of Alberta)
ClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningScore-based ModelContrastive LearningImageTabular
🎯 What it does: Propose RSA-CP, which improves compliance prediction under small sample conditions by aligning the ranks of random reference scores with actual scores, significantly reducing the prediction set width while maintaining coverage.
RSAgent: Learning to Reason and Act via Multi-Turn Tool Invocations for Text-Guided Segmentation
Xingqi He (Fudan University), Wenqiang Zhang (Fudan University)
SegmentationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose RSAgent, a multi-modal large language model based on multi-round tool calls, capable of gradually reasoning, executing visual tools, and iteratively refining segmentation masks in interactions between images and text descriptions.
RSF-GLLM: Bridging the Semantic Gap in Multi-Hop Knowledge Graph QA via Recurrent Soft-Flow and Decoupled LLM Generation
Sambaran Bandyopadhyay (Adobe Research), Ananth Muppidi (Adobe Systems)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningFlow-based ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the RSF-GLLM framework, decoupling differentiable graph reasoning (Recurrent Soft-Flow) from answer generation (LLM fine-tuning), to address the semantic gap in multi-hop KGQA.
RSPO: Regularized Self-Play Alignment of Large Language Models
Xiaohang Tang (University College London), Ilija Bogunovic (University College London)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Studied the impact of regularizing the reference policy in self-play methods on the alignment of large language models (LLMs), and proposed a pluggable, theoretically convergent regularized self-play framework (RSPO).
RSTR: Reducing SpatioTemporal Redundancy in Diffusion Transformers
Ruitong Sun (University of Georgia), Jin Sun (University of Georgia)
GenerationData SynthesisOptimizationComputational EfficiencyHyperparameter SearchTransformerDiffusion modelImageVideo
🎯 What it does: This paper reduces the spatiotemporal redundancy of diffusion Transformers and improves inference speed by jointly optimizing the timing and scale of CFG execution and adopting adaptive rank allocation.
RT-Lynx: Putting the GEMM Sparsity In a Right Way for Diffusion Models
Xing Cong (Beihang University), chenhao xie
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelScore-based ModelImage
🎯 What it does: Propose the RT‑Lynx method, which migrates N:M structured sparsity from weights to activations of Diffusion Transformer, and combines l2 normalization compensation, lightweight LoRA fine-tuning, and CUDA core fusion to achieve inference acceleration without quality loss.
RTInfer: Real-Time Inference of Multiple DNNs on Edge GPUs
Renjie Li (Zhejiang University), Wei Dong (Zhejiang University)
Autonomous DrivingOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerReinforcement LearningPrompt EngineeringMixture of ExpertsAuto EncoderGenerative Adversarial NetworkContrastive LearningImageVideoText
🎯 What it does: Propose the RTInfer system to achieve concurrent real-time inference of multiple DNN tasks on edge GPUs.
RTPrune: Reading-Twice Inspired Token Pruning for Efficient DeepSeek-OCR Inference
Ben Wan (JD.com), Tongxuan Liu (JD.com)
RecognitionComputational EfficiencyTransformerVision Language ModelContrastive LearningImageTextBenchmark
🎯 What it does: Proposes a training-agnostic visual token pruning method called RTPrune for DeepSeek-OCR, which simulates the two-stage reading behavior during the model decoding process. It first retains tokens with the highest embedded norm, and then merges the remaining tokens using optimal transport.
Rubric Curriculum RL: Exploiting the Generation-Verification Gap in Non-Verifiable Domains
Tejas Krishnan (University of Oxford), Shital Shah (Microsoft Research)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and evaluated Rubric Curriculum RL (RcRL), a self-improvement method that leverages the generation-validation gap and gradually introduces rubric standards, significantly enhancing language model quality in non-verifyable tasks such as open-ended creative writing and health-related question answering.
RubricRobustness: Evaluating the Sensitivity of Rubrics-Based Benchmarks to Simple Perturbations
Manasi Sharma (Scale AI)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the RUBRICROBUSTNESS framework, conducting a systematic sensitivity analysis on rubric-based LLM-as-a-judge evaluation.
RuCL: Stratified Rubric-Based Curriculum Learning for Multimodal Large Language Model Reasoning
Yukun Chen (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences), Min Yang (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Proposes Stratified Rubric-Based Curriculum Learning (RuCL), which guides multi-modal large language models to progressively learn from basic perception to advanced reasoning by formulating hierarchical evaluation rubrics and dynamically adjusting weights.
Rule2DRC: Benchmarking LLM Agents for DRC Script Synthesis with Execution-Guided Test Generation
Jinuk Kim (Seoul National University), Hyun Oh Song (Seoul National University)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a large-scale benchmark called Rule2DRC for evaluating the ability of LLMs to translate natural language rules into executable design rule check (DRC) scripts, and introduced SplitTester, a testing tool that leverages execution feedback to improve script selection.
RulePlanner: All-in-One Reinforcement Learner for Unifying Design Rules in 3D Floorplanning
Ruizhe Zhong (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationConvolutional Neural NetworkGraph Neural NetworkTransformerReinforcement LearningPrompt EngineeringGraphTabularBenchmark
🎯 What it does: Proposed a deep reinforcement learning framework named RulePlanner for simultaneously satisfying multiple hardware design rules in 3D IC floorplanning;
RVAS: Referring Video Active Exploration and Segmentation
Hengrui Hu (Fudan University), Henghui Ding (Fudan University)
SegmentationRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the Referring Video Active Exploration and Segmentation (RVAS) task, which requires the agent to actively explore and segment target objects in real-time video streams based on natural language expressions.
S-Quant: Rethinking Weight Quantization with Seed-Based Generation
Mingzi Wang (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Propose S-Quant, a post-training quantization method that only compresses weights, utilizing LFSR seed to dynamically generate basis matrices and reconstruct weight blocks through linear combinations.
S$^3$GNN: Efficient Global Mixing and Local Message Passing for Long-Range Graph Learning
Dai Shi (University of Cambridge), José Miguel Hernández-Lobato (University of Cambridge)
Computational EfficiencyRepresentation LearningGraph Neural NetworkTransformerTextGraphTabularTime Series
🎯 What it does: Proposed a new graph neural network framework called S3-GNN, combining lightweight global mixing with local message passing to address the oversquashing (information compression) problem in traditional MPNNs.
S2GS: Streaming Semantic Gaussian Splatting for Online Scene Understanding and Reconstruction
Renhe Zhang (East China Normal University), Xin Tan (East China Normal University)
RecognitionSegmentationData SynthesisComputational EfficiencyKnowledge DistillationTransformerNeural Radiance FieldContrastive LearningGaussian SplattingOptical FlowImageVideoPoint Cloud
🎯 What it does: Propose a strictly causal, no-reprocessing flow-based semantic Gaussian splatting framework called S2GS, which can real-time update 3D geometry, appearance, and instance-level semantics.
S2M-Net: Spectral-Spatial Mixing with Morphology-Aware Adaptive Loss for Medical Image Segmentation
Sanaullah Chowdhury (Independent Researcher), Lameya Sabrin (Independent Researcher)
SegmentationConvolutional Neural NetworkTransformerMixture of ExpertsAuto EncoderContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission TomographyUltrasoundDiffusion Tensor ImagingAlzheimer's DiseaseBenchmark
🎯 What it does: Proposed S2M-Net, a lightweight medical image segmentation network that combines spectral and spatial domain hybrid features;
SABER: Continual Learning with Representation Conflict Management
Xuandi Luo (South China University of Technology), Shengfeng He (Singapore Management University)
Representation LearningImageBenchmark
🎯 What it does: Proposes a new method called SABER for actively managing representation conflicts between tasks in continual learning. The core idea is to achieve a balance between stability and plasticity through subspace alignment, decomposition and recombination of task-shared and task-specific knowledge, and an energy balancing mechanism.
SAC-Opt: Semantic Anchors for Iterative Correction in Optimization Modeling
Yansen Zhang (City University of Hong Kong), Chen Ma (City University of Hong Kong)
OptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the SAC-Opt framework, which enhances the semantic consistency of the optimized model generated by LLM through backward iterative correction using semantic anchors.
SAD-Flower: Flow Matching for Safe, Admissible, and Dynamically Consistent Planning
Tzu-Yuan Huang (Technical University of Munich), Sandra Hirche (Technical University of Munich)
Autonomous DrivingOptimizationSafty and PrivacyRobotic IntelligenceDiffusion modelScore-based ModelFlow-based ModelTabularTime SeriesSequentialBenchmarkOrdinary Differential Equation
🎯 What it does: Proposes a control-enhanced flow matching framework (SAD-Flower), which generates both safe and executable trajectories without retraining by incorporating virtual control inputs into flow matching and using control barrier functions (CBF) and control Lyapunov functions (CLF).
SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders
Enrico Cassano (University of Turin), Marco Grangetto (University of Turin)
GenerationExplainability and InterpretabilityHyperparameter SearchData-Centric LearningSupervised Fine-TuningDiffusion modelAuto EncoderImage
🎯 What it does: Propose a concept forgetting framework based on supervised sparse autoencoder (SAEmnesia), which can precisely delete target concepts in diffusion models.
SAEs-BrainMap: Unveiling the Emergence of Specialized Concepts in Deep Models via Brain Alignment
Ziming Mao (Beijing Institute of Technology), Guoyuan Yang (Beijing Institute of Technology)
Explainability and InterpretabilityRepresentation LearningTransformerAuto EncoderContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose the SAEs-BrainMap framework, which uses fMRI activation in the human brain's ventral visual pathway as a functional probe to guide Sparse Autoencoders (SAEs) in selecting features and to trace the hierarchical emergence trajectory of general concepts (such as faces, scenes, text, etc.) in deep visual models.
Safe and Scalable Web Agent Learning via Recreated Websites
Hyungjoo Chae (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
Data SynthesisAutonomous DrivingSafty and PrivacyComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringWorld ModelImageVideoTextTabularRetrieval-Augmented Generation
🎯 What it does: Clone real websites automatically through language models and encoding agents, generate executable synthetic environments, and create verifiable tasks and evaluators within them, allowing web agents to self-evolve and learn in a safe and controllable environment.
Safe Autoregressive Image Generation with Iterative Self-Improving Codebooks
Yunqi Xue (Wuhan University), Jindong Gu (University of Oxford)
GenerationSafty and PrivacyTransformerSupervised Fine-TuningVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Leverage the generation and comprehension capabilities of a unified autoregressive multimodal model to construct a safe codebook, eliminating harmful content in generated images through an iterative self-improvement approach.
Safe In-Context Reinforcement Learning
Amir Moeini (University of Virginia), Shangtong Zhang (University of Virginia)
TransformerReinforcement LearningPrompt EngineeringTabularTime SeriesSequentialBenchmark
🎯 What it does: Propose the SCARED method to achieve parameter-free update safe in-context reinforcement learning;