NeurIPS 2025 Papers — Page 40
Conference on Neural Information Processing Systems · 5275 papers
RNNs perform task computations by dynamically warping neural representations
Arthur Pellegrino (University College London), Angus Chadwick (University of Edinburgh)
Recurrent Neural NetworkSequential
🎯 What it does: This paper proposes and validates a unified analysis framework using Riemannian geometry to represent and compute dynamic systems (especially RNNs) with varying input acceptance times, revealing the mechanism by which the network performs task computations through dynamic distortion of its neural representations.
RobIA: Robust Instance-aware Continual Test-time Adaptation for Deep Stereo
Jueun Ko (Ewha Womans University), Dongbo Min (Ewha Womans University)
Depth EstimationDomain AdaptationAutonomous DrivingMixture of ExpertsImage
🎯 What it does: A robust instance-aware continuous testing adaptation framework for stereo depth estimation (RobIA) is proposed, which achieves online adaptation to dynamic domain shifts by introducing an input-aware mixture of experts module and a robust AdaptBN teacher.
RoboRefer: Towards Spatial Referring with Reasoning in Vision-Language Models for Robotics
Enshen Zhou (Beihang University), Shanghang Zhang (Peking University)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Developed RoboRefer, a 3D perception visual-language model that can complete spatial reference tasks through single-step precise spatial understanding and multi-step spatial reasoning.
RoboScape: Physics-informed Embodied World Model
Yu Shang (Tsinghua University), Yong Li (Tsinghua University)
GenerationDepth EstimationRobotic IntelligenceTransformerWorld ModelVideo
🎯 What it does: A physical information-based world model for RoboScape is proposed, integrating RGB video generation, temporal depth prediction, and adaptive keypoint dynamics to enhance the physical realism and action controllability of robot scene videos.
Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics
Dongyoung Kim (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
🎯 What it does: The ROBOT-R1 framework is proposed, which trains large visual language models through reinforcement learning to enhance the embodied reasoning capabilities in robot control.
RobotSmith: Generative Robotic Tool Design for Acquisition of Complex Manipulation Skills
Chunru Lin (University of Massachusetts Amherst), Chuang Gan (Massachusetts Institute of Technology)
OptimizationRobotic IntelligenceVision Language ModelMesh
🎯 What it does: This work proposes RobotSmith, an end-to-end framework that automatically generates tools that meet task requirements using vision-language models (VLM) and physical simulation, enabling the entire process from no tools to completing complex manipulation tasks.
Robust and Computation-Aware Gaussian Processes
Marshal Arijona Sinaga, Samuel Kaski (Aalto University)
Anomaly DetectionOptimizationComputational EfficiencyTabularTime SeriesFinance Related
🎯 What it does: A robust computationally aware Gaussian process model (RCaGP) has been developed, which can simultaneously address the uncertainty caused by approximation in large-scale data and the robustness issues brought by outliers.
Robust and Diverse Multi-Agent Learning via Rational Policy Gradient
Niklas Lauffer (University of California Berkeley), Michael D Dennis
Adversarial AttackRobotic IntelligenceReinforcement Learning
🎯 What it does: Designed and implemented the Rational Policy Gradient (RPG), which addresses the self-destructive problem in multi-agent non-zero-sum environments through Rational Policy Optimization (RPO) constraints, achieving robustness improvement, adversarial sample discovery, and diversity learning in similar tasks.
Robust and Scalable Autonomous Reinforcement Learning in Irreversible Environments
Sang-Hyun Lee (Ajou University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A robust and scalable autonomous reinforcement learning algorithm RSA is proposed, which can reduce manual resets and improve learning efficiency in both reversible and irreversible environments.
Robust Contextual Pricing
Anupam Gupta, Jon Schneider
Optimization
🎯 What it does: In the context pricing problem, this paper proposes an improved algorithm for the case of corrupted feedback and provides corresponding upper and lower bounds;
Robust Cross-modal Alignment Learning for Cross-Scene Spatial Reasoning and Grounding
Yanglin Feng (Sichuan University), Peng Hu (Sichuan National Innovation New Vision UHD Video Technology Co., Ltd.)
Object DetectionRetrievalTransformerContrastive LearningTextPoint Cloud
🎯 What it does: This paper proposes the Cross-Scene Spatial Reasoning and Localization (CSSRG) task and introduces the matching-localization two-stage CoRe framework to address the challenges of large-scale scene retrieval and fine-grained text-object alignment.
Robust Distortion-Free Watermark for Autoregressive Audio Generation Models
Yihan Wu (University of Maryland), Heng Huang (University of Maryland)
GenerationData SynthesisRecurrent Neural NetworkAudio
🎯 What it does: A lossless watermarking framework ALIGNED-IS for autoregressive audio generation models is proposed, addressing the issue of label mismatch caused by re-encoding;
Robust Distributed Estimation: Extending Gossip Algorithms to Ranking and Trimmed Means
Anna van Elst (Telecom Paris), Stephan Clémençon (Telecom Paris)
TabularTime Series
🎯 What it does: Two robust distributed estimation algorithms based on gossip are proposed: GORANK for global ranking estimation and GOTRIM for α-trimmed mean estimation, which can converge at a rate of O(1/t) on any communication graph and resist data outliers.
Robust Ego-Exo Correspondence with Long-Term Memory
Yijun Hu (University of Chinese Academy of Sciences), Libo Zhang (Rochester Institute of Technology)
Object DetectionSegmentationTransformerMixture of ExpertsVideoBenchmark
🎯 What it does: A robust perspective correspondence framework LM‑EEC based on SAM 2 is proposed, specifically addressing the object-level correspondence and segmentation issues in self-perspective (ego) and external perspective (exo) videos synchronized over time.
Robust Egocentric Referring Video Object Segmentation via Dual-Modal Causal Intervention
Haijing Liu (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
Object DetectionSegmentationVideo
🎯 What it does: This paper proposes a causal inference-based bimodal intervention framework called CERES, aimed at improving the performance of first-person perspective referential video object segmentation.
Robust Equilibria in Continuous Games: From Strategic to Dynamic Robustness
Kyriakos Lotidis (Stanford University), Jose Blanchet (Stanford University)
Optimization
🎯 What it does: This paper studies the robustness of Nash equilibrium in continuous games, proposing a robust equilibrium concept under strategic and dynamic uncertainty, and proving a structural correspondence between the two.
Robust Estimation Under Heterogeneous Corruption Rates
Syomantak Chaudhuri (University of California), Thomas Courtade
Optimization
🎯 What it does: This paper proposes the λ-contamination model and studies the robust estimation problem under the condition that each sample has a different known contamination rate. It provides lower and upper bounds for mean estimation (including bounded distributions and normal distributions) as well as linear regression, along with corresponding optimal algorithms.
Robust Explanations of Graph Neural Networks via Graph Curvatures
Yazheng Liu (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
OptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: A method is proposed to enhance the robustness of graph neural network (GNN) explanations through graph curvature, aiming to improve the credibility of GNNs in high-risk applications.
Robust Federated Finetuning of LLMs via Alternating Optimization of LoRA
Shuangyi Chen (University of Toronto), Ashish J Khisti
OptimizationFederated LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The RoLoRA framework is proposed, which achieves efficient and robust LLM fine-tuning in federated learning by alternately optimizing the low-rank projection and high-rank projection.
Robust Graph Condensation via Classification Complexity Mitigation
Jiayi Luo (Beihang University), Philip S. Yu (University of Illinois)
ClassificationOptimizationAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper studies and proposes a robust framework for Graph Condensation (MRGC) to mitigate attacks on node features, structure, and labels.
Robust Hallucination Detection in LLMs via Adaptive Token Selection
Mengjia Niu (Imperial College London), Guansong Pang (Singapore Management University)
Anomaly DetectionTransformerLarge Language ModelText
🎯 What it does: This paper proposes a framework named HaMI, which adaptively selects the tokens that best represent hallucinations using multi-instance learning (MIL) within the internal representations of LLMs, and trains a hallucination detector based on this; it also incorporates predictive uncertainty information to enhance the representation.
Robust Hyperbolic Learning with Curvature-Aware Optimization
Ahmad Bdeir (University of Hildesheim), Niels Landwehr (University of Hildesheim)
OptimizationReinforcement LearningImageBiomedical Data
🎯 What it does: The study proposes a robust optimization framework for hypersurface learning, including techniques such as Riemannian AdamW, curvature-aware updates, and maximum distance scaling.
Robust Integrated Learning and Pauli Noise Mitigation for Parametrized Quantum Circuits
Md Mobasshir Arshed Naved (Purdue University), Ananth Grama (Purdue University)
Image
🎯 What it does: A joint gradient learning framework is proposed, which simultaneously optimizes parameterized quantum circuits (PQCs) and a learnable Pauli-Lindblad inverse noise model to achieve dynamic compensation for Pauli noise in NISQ devices.
Robust Label Proportions Learning
Jueyu Chen (Southeast University), Yuheng Jia (Southeast University)
ClassificationRepresentation LearningContrastive LearningImageTabular
🎯 What it does: A robust label proportion learning framework (RLPL) is proposed, which is divided into two stages: the first stage uses unsupervised contrastive learning to pre-train the encoder and trains the initial classifier with bag proportion; the second stage denoises based on optimal transport (LLP-OTD) and divides the pseudo-labels into high-confidence and low-confidence groups, and then uses the LLPMix approach of MixMatch to integrate high-confidence labels with bag proportion constraints in a semi-supervised environment, ultimately training the main classifier.
Robust learning of halfspaces under log-concave marginals
Jane Lange (Massachusetts Institute of Technology), Arsen Vasilyan (University of Texas at Austin)
Optimization
🎯 What it does: An algorithm for adaptive robust learning in half-spaces under sub-Gaussian isotropic log-concave distributions is proposed, which achieves robustness to Euclidean perturbations while ensuring a low error rate.
Robust LLM Alignment via Distributionally Robust Direct Preference Optimization
Zaiyan Xu (Texas A&M University), Deepak Ramachandran (Google DeepMind)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes distributionally robust direct preference optimization (WDPO/KLDPO) to address the alignment failure of LLMs when user preference distributions drift.
Robust Minimax Boosting with Performance Guarantees
Santiago Mazuelas (Basque Center of Applied Mathematics), Veronica Alvarez
OptimizationTabular
🎯 What it does: A Robust Maximum Value Minimization Boosting (RMBoost) method is proposed, which directly minimizes the error probability and provides finite sample performance guarantees.
Robust Neural Rendering in the Wild with Asymmetric Dual 3D Gaussian Splatting
Chengqi Li (McMaster University), Xiangyu Xu (Xi'an Jiaotong University)
RestorationData SynthesisGaussian SplattingImage
🎯 What it does: The paper proposes a dual-model 3D Gaussian Splatting framework called AsymGS, which utilizes parallel training of two models and consistency constraints to suppress random artifacts from low-quality training images.
Robust Policy Expansion for Offline-to-Online RL under Diverse Data Corruption
Longxiang He (Tsinghua University), Li Shen (Shenzhen Campus of Sun Yat-sen University)
Reinforcement Learning
🎯 What it does: A robust strategy expansion method RPEX has been developed to combat data corruption in offline-to-online reinforcement learning (O2O RL).
Robust Regression of General ReLUs with Queries
Ilias Diakonikolas (University of Wisconsin Madison), Mingchen Ma (University of Wisconsin Madison)
OptimizationTabular
🎯 What it does: This paper proposes a general ReLU regression approximation learning algorithm under Gaussian distribution with accessible label queries, achieving an error guarantee of O(opt)+ε, and provides the corresponding query complexity.
Robust Reinforcement Learning in Finance: Modeling Market Impact with Elliptic Uncertainty Sets
Shaocong Ma (University of Maryland), Heng Huang (University of Maryland)
OptimizationReinforcement LearningTabularTime SeriesFinance Related
🎯 What it does: Considering market shocks in financial trading, an elliptical uncertainty set is proposed for robust reinforcement learning, providing explicit/implicit closed-form solutions and forming an efficient robust TD learning algorithm;
Robust Sampling for Active Statistical Inference
Puheng Li (Stanford University), Emmanuel Candes
TextTabular
🎯 What it does: A robust sampling method is proposed for optimal trade-off between active sampling and uniform sampling in budget-constrained active statistical inference, ensuring that the variance of the estimator is not worse than any benchmark.
Robust Satisficing Gaussian Process Bandits Under Adversarial Attacks
Artun Saday (Bilkent University), Cem Tekin (Bilkent University)
OptimizationAdversarial AttackTabular
🎯 What it does: This paper proposes a framework for robust satisficing optimization of Gaussian Process (GP) bandit problems under unknown and variable adversarial perturbations, and presents two algorithms (AdveRS-1 and AdveRS-2). The theoretical analysis demonstrates the sublinear lenient and RS regret upper bounds under different adversarial assumptions.
Robust SuperAlignment: Weak-to-Strong Robustness Generalization for Vision-Language Models
Junhao Dong (Nanyang Technological University), Yew-Soon Ong (Nanyang Technological University)
Representation LearningAdversarial AttackTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an unsupervised weak-to-strong robustness generalization framework (Adv-W2S) that enhances the zero-shot robustness of large-scale vision-language models by using adversarial samples and reverse adversarial examples under the supervision of weak models.
Robustifying Learning-Augmented Caching Efficiently without Compromising 1-Consistency
Peng Chen (Zhejiang University), Shuiguang Deng (Zhejiang University)
OptimizationTabular
🎯 What it does: A lightweight robust framework named GUARD is proposed to enhance the robustness of learning-enhanced caching algorithms while maintaining 1-consistency.
Robustly Learning Monotone Single-Index Models
Puqian Wang (University of Wisconsin Madison), Jelena Diakonikolas (University of Wisconsin Madison)
OptimizationComputational EfficiencyTabularStochastic Differential Equation
🎯 What it does: This paper proposes a computationally efficient algorithm for learning single index models (SIM) under adversarial label noise with Gaussian inputs, achieving a constant factor approximation.
RobustMerge: Parameter-Efficient Model Merging for MLLMs with Direction Robustness
Fanhu Zeng (Institute of Automation, Chinese Academy of Sciences), Hao Tang (Peking University)
GenerationOptimizationTransformerLarge Language ModelMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: This paper proposes a training-free, data-free parameter-efficient model merging method called RobustMerge, specifically designed for multi-modal large language models (MLLMs), which can merge multi-task expert models into a single multi-task model.
Robustness in Both Domains: CLIP Needs a Robust Text Encoder
Elias Abad Rocamora (Ecole Polytechnique Federale de Lausanne), Volkan Cevher (University of Tubingen)
RetrievalAdversarial AttackTransformerContrastive LearningImageTextMultimodality
🎯 What it does: The LEAF method is proposed, which efficiently fine-tunes the CLIP text encoder against adversarial attacks, enhancing the robustness of the text domain, and combines it with existing image robust encoders (FARE) to achieve multimodal robustness.
RODS: Robust Optimization Inspired Diffusion Sampling for Detecting and Reducing Hallucination in Generative Models
Yiqi Tian (Massachusetts General Hospital and Harvard Medical School), Quanzheng Li (Massachusetts General Hospital)
GenerationOptimizationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A robust optimization-based diffusion sampling framework RODS is proposed, which can actively detect and correct hallucinated outputs during the generation process.
ROGR: Relightable 3D Objects using Generative Relighting
Jiapeng Tang (Google Research), Philipp Henzler (Google Research)
GenerationData SynthesisTransformerDiffusion modelNeural Radiance FieldImage
🎯 What it does: Utilize a multi-view diffusion model to synthesize view-consistent multi-lighting samples, and then train a dual-branch lighting condition NeRF that can render under any environmental lighting using these samples;
Role Bias in Diffusion Models: Diagnosing and Mitigating through Intermediate Decomposition
Sina Malakouti (University of Pittsburgh), Adriana Kovashka (University of Pittsburgh)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningDiffusion modelImageTextBenchmarkChain-of-Thought
🎯 What it does: This paper studies the role collapse in action relationship generation using text-to-image diffusion models, proposing the RoleBench benchmark to quantify the generation quality of rare action combinations (e.g., 'mouse chasing cat'), and designing the ReBind framework to fine-tune through active/passive intermediate combinations generated by LLMs, significantly reducing the model's preference for frequent combinations.
Role-aware Multi-agent Reinforcement Learning for Coordinated Emergency Traffic Control
Ming Cheng (Central South University), Senzhang Wang (Central South University)
Autonomous DrivingRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningGraphTime Series
🎯 What it does: A Role-aware Multi-agent Traffic Control (RMTC) framework is proposed, which jointly learns the collaborative decision-making of traffic lights, emergency vehicles (EMV), and regular vehicles (REV) to achieve rapid passage for EMVs while maintaining overall traffic smoothness.
Rollout Roulette: A Probabilistic Inference Approach to Inference-Time Scaling of LLMs using Particle-Based Monte Carlo Methods
Isha Puri (Massachusetts Institute of Technology), Akash Srivastava (Red Hat)
TransformerLarge Language ModelReinforcement LearningTextSequentialFinance Related
🎯 What it does: This paper proposes an extension of inference for large language models using particle filtering to avoid the early pruning problem caused by the imperfection of the reward model.
RoMa: A Robust Model Watermarking Scheme for Protecting IP in Diffusion Models
Yingsha Xie (Sun Yat-sen University), Xiaochun Cao (Sun Yat-sen University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A robust model watermarking scheme named RoMa is proposed, which enhances the persistence of the watermark against fine-tuning by embedding triggers in the diffusion model and introducing path-specific smoothness.
RoMA: Scaling up Mamba-based Foundation Models for Remote Sensing
Fengxiang Wang (National University of Defense Technology), Jing Zhang (Wuhan University)
ClassificationSegmentationTransformerImage
🎯 What it does: The RoMA framework is proposed to achieve self-supervised autoregressive pre-training of the Mamba architecture, which can efficiently handle high-resolution remote sensing images.
RoME: Domain-Robust Mixture-of-Experts for MILP Solution Prediction across Domains
Tianle Pu (National University of Defense Technology), Changjun Fan (National University of Defense Technology)
OptimizationGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: The RoME framework is proposed, utilizing Mixture-of-Experts and distributionally robust optimization to achieve training and inference of cross-domain MILP solvers.
RoomEditor: High-Fidelity Furniture Synthesis with Parameter-Sharing U-Net
Zhenyi Lin (Tianjin University), Qinghua Hu (Tianjin University)
Image TranslationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelImageBenchmark
🎯 What it does: A public furniture synthesis benchmark dataset, RoomBench, has been constructed, and a parameter-shared dual U-Net network, RoomEditor, has been proposed to achieve high-fidelity and geometrically consistent indoor furniture insertion.
Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping
Justin Lazarow (Apple), Afshin Dehghan (Apple)
Object DetectionPose EstimationTransformerSimultaneous Localization and MappingOptical FlowImage
🎯 What it does: Without relying on explicit point clouds or depth maps, this paper utilizes 3D bounding boxes as primitives to achieve 3D object detection, camera localization, and global semantic map construction from unordered RGB (or RGB-D) image sets in indoor spaces.
Root Cause Analysis of Outliers with Missing Structural Knowledge
William Roy Orchard (University of Cambridge), Dominik Janzing (Amazon)
Anomaly DetectionTabular
🎯 What it does: This paper proposes two root cause analysis methods for single-sample anomaly situations that do not require estimating conditional distributions or complete structural models, avoiding the statistical infeasibility of traditional methods when there is a lack of a large number of posterior samples.
ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge
Manh Cuong Dao (National University of Singapore), Trong Nghia Hoang (Washington State University)
OptimizationReinforcement LearningTabularBenchmark
🎯 What it does: By treating offline black-box optimization as a distribution translation task, a probabilistic bridge model is proposed, utilizing synthetic Gaussian process functions for pre-training, generating low-value to high-value design transfer paths, ultimately producing better candidate designs.
Rope to Nope and Back Again: A New Hybrid Attention Strategy
Bowen Yang (Cohere), Acyr Locatelli (Cohere)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A hybrid attention architecture combining RoPE and NoPE is proposed, introducing a sliding window in the RoPE layer to enhance long context modeling capabilities.
RoPECraft: Training-Free Motion Transfer with Trajectory-Guided RoPE Optimization on Diffusion Transformers
Ahmet Berke Gökmen, Aysegul Dundar
GenerationOptimizationTransformerDiffusion modelOptical FlowVideo
🎯 What it does: RoPECraft is a training-independent motion transfer method that achieves precise control of video motion by modifying RoPE in the Diffusion Transformer.
ROSE: Remove Objects with Side Effects in Videos
Chenxuan Miao (Zhejiang University), Hengshuang Zhao (The University of Hong Kong)
RestorationGenerationData SynthesisTransformerDiffusion modelVideoBenchmark
🎯 What it does: The ROSE framework is proposed to simultaneously remove target objects and their shadows, reflections, lighting, and other side effects in videos.
Rotary Masked Autoencoders are Versatile Learners
Uros Zivanovic (University of Trieste), Roberto Trotta (Scuola Internazionale Superiore di Studi Avanzati)
ClassificationAnomaly DetectionTransformerAuto EncoderImageTime SeriesAudio
🎯 What it does: Proposes RoMAE, an architecture that embeds RoPE into a Masked Autoencoder, capable of handling multi-dimensional continuous positional information, compatible with irregular time series, images, and audio.
Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning
Haozhen Zhang (University of Illinois at Urbana-Champaign), Jiaxuan You (University of Illinois at Urbana-Champaign)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes Router-R1, a multi-round LLM routing and aggregation framework based on reinforcement learning.
Routing Mamba: Scaling State Space Models with Mixture-of-Experts Projection
Zheng Zhan (Microsoft), yelong shen
Mixture of ExpertsText
🎯 What it does: Proposes the Routing Mamba (RoM) framework, applying sparse experts to the projection layer of state space models like Mamba to achieve efficient scaling.
ROVER: Recursive Reasoning Over Videos with Vision-Language Models for Embodied Tasks
Philip Schroeder (Massachusetts Institute of Technology), James R. Glass (Massachusetts Institute of Technology)
Robotic IntelligenceTransformerVision Language ModelVideo
🎯 What it does: Proposes the ROVER framework, which recursively decomposes tasks in video embedded tasks to achieve precise temporal reasoning for long sequence videos.
RPG360: Robust 360 Depth Estimation with Perspective Foundation Models and Graph Optimization
Dongki Jung (University of Maryland), Dinesh Manocha (University of Maryland)
Depth EstimationOptimizationImage
🎯 What it does: This paper proposes a training-independent 360-degree monocular depth estimation method called RPG360. It utilizes existing perspective-based models to first project panoramic images into six-face cube perspective images, and then predicts depth and normal maps separately. By introducing graph optimization and scale parameters for each face, it ensures scale consistency when merging the six depth maps back into the panoramic view, resulting in more accurate 3D reconstruction.
RrED: Black-box Unsupervised Domain Adaptation via Rectifying-reasoning Errors of Diffusion
Yuwu Lu (South China Normal University), Chunzhi Liu (South China Normal University)
Domain AdaptationPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper proposes a black-box unsupervised domain adaptation method based on diffusion models called RrED, which employs a two-stage learning approach (DTR and SRM) to correct inference errors and enhance the capabilities of the target model.
RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards
Jingnan Zheng (National University of Singapore), Tat-Seng Chua (National University of Singapore)
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A security monitoring model based on active inference, RSafe, has been designed and implemented. It can perform step-by-step reasoning on inputs within the user-specified security policy range and provide interpretable security judgments.
RSAVQ: Riemannian Sensitivity-Aware Vector Quantization for Large Language Models
Zukang Xu (Houmo AI), Dawei Yang (Houmo AI)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: A Riemannian geometry-based vector quantization framework RSAVQ is proposed for ultra-low bit quantization of large language models.
rStar-Coder: Scaling Competitive Code Reasoning with a Large-Scale Verified Dataset
Yifei Liu (University of Science and Technology of China), Mao Yang (Microsoft Research Asia)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A large-scale, high-quality competitive programming dataset rStar-Coder has been constructed, containing 418K problems and 580K long reasoning answers, and solutions are validated through reliable test case generation and mutual verification mechanisms.
RUAGO: Effective and Practical Retain-Free Unlearning via Adversarial Attack and OOD Generator
SangYong Lee, Simon S. Woo (Sungkyunkwan University)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a framework called RUAGO for instance-level model unlearning in the absence of retained data, aiming to safely and quickly remove specified training samples from the model while maintaining overall model performance and privacy security.
RULE: Reinforcement UnLEarning Achieves Forget-retain Pareto Optimality
Chenlong Zhang (Chinese Academy of Sciences), Yubo Chen (Chinese Academy of Sciences)
OptimizationLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the RULE framework, which utilizes reinforcement learning for selective forgetting in large language models, focusing on learning rejection strategies to delineate the boundaries of forgetting and retention, thereby avoiding unnatural or crashing responses from the model when faced with forgetting queries.
RvLLM: LLM Runtime Verification with Domain Knowledge
Yedi Zhang (National University of Singapore), Jin Song Dong (National University of Singapore)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The RvLLM framework is proposed, which combines a lightweight domain knowledge specification language ESL to verify the consistency of LLM runtime outputs.
S-Crescendo: A Nested Transformer Weaving Framework for Scalable Nonlinear System in S-Domain Representation
Junlang Huang, Zhong Guan
TransformerTime SeriesPhysics Related
🎯 What it does: Proposes the S-Crescendo framework, which utilizes Transformer and S-domain physical priors to achieve time-domain predictions of high-order nonlinear RC networks, significantly reducing computational complexity.
S-GRPO: Early Exit via Reinforcement Learning in Reasoning Models
Mz Dai, Qingyi Si (Huawei Technologies)
Computational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: A novel reinforcement learning method S-GRPO is proposed and implemented, which can achieve early stopping in the chain reasoning process of large language models, improving reasoning efficiency and accuracy.
S'MoRE: Structural Mixture of Residual Experts for Parameter-Efficient LLM Fine-tuning
Hanqing Zeng (Meta AI), Benyu Zhang (Meta AI)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: A parameter-efficient LLM fine-tuning framework named S'MoRE is proposed and implemented, integrating the low-rank efficiency of LoRA with the plasticity of MoE.
S$^2$M-Former: Spiking Symmetric Mixing Branchformer for Brain Auditory Attention Detection
Jiaqi Wang (Harbin Institute of Technology), Zhiguo Zhang (Harbin Institute of Technology)
Spiking Neural NetworkTime SeriesBiomedical Data
🎯 What it does: Proposes S2M-Former, a dual-branch symmetric hybrid spiking neural network for EEG auditory attention detection.
S$^2$NN: Sub-bit Spiking Neural Networks
Wenjie Wei (University of Electronic Science and Technology of China), Haizhou Li (Chinese University of Hong Kong Shenzhen)
ClassificationObject DetectionSegmentationCompressionKnowledge DistillationConvolutional Neural NetworkSpiking Neural NetworkTransformerImage
🎯 What it does: This paper proposes Sub-bit Spiking Neural Networks (SNN), which greatly compresses the SNN model by quantizing the weights to less than 1 bit while maintaining high accuracy.
SAD Neural Networks: Divergent Gradient Flows and Asymptotic Optimality via o-minimal Structures
Julian Kranz (University of Münster), Arnulf Jentzen (University of Münster)
OptimizationImageOrdinary Differential Equation
🎯 What it does: This paper studies the training dynamics of fully connected feedforward neural networks under continuous time gradient flow, proving a dichotomy: the gradient flow either converges to a critical point or approaches a generalized critical value as the parameters go to infinity. It further shows that when the loss is below a certain threshold, it will necessarily converge to the optimal value. For polynomial objective functions, the authors prove that in the case of sufficiently large networks and datasets, the loss cannot reach zero (there is no global minimum), but can get arbitrarily close to zero, leading to the inference that the gradient flow must diverge. Numerical experiments validate this divergence phenomenon and extend it to more practical tasks such as PDE solving and MNIST image classification.
SAEMark: Steering Personalized Multilingual LLM Watermarks with Sparse Autoencoders
Zhuohao Yu (Peking University), Wei Ye (Peking University)
Large Language ModelAuto EncoderText
🎯 What it does: This paper proposes a feature rejection sampling framework called SAEMARK based on sparse autoencoders, which enables multi-bit watermark embedding and detection during the inference phase without modifying the model or logits. It is compatible with any black-box LLM and can be used across languages and domains.
Safe + Safe = Unsafe? Exploring How Safe Images Can Be Exploited to Jailbreak Large Vision-Language Models
Chenhang Cui (National University of Singapore), Tat-Seng Chua (National University of Singapore)
GenerationAdversarial AttackTransformerLarge Language ModelAgentic AIVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: This study investigates an attack method that uses seemingly safe images and prompts to induce large visual language models (LVLM) to generate unsafe content; it proposes a Safe Avalanche Agent (SSA) framework that utilizes the model's general reasoning capabilities and the safe avalanche effect for staged attacks;
Safe and Stable Control via Lyapunov-Guided Diffusion Models
Xiaoyuan Cheng (University College London), Yiming Yang (University College London)
OptimizationRobotic IntelligenceDiffusion modelSequential
🎯 What it does: This paper proposes a safety and stability control framework based on diffusion models—Safe and Stable Diffusion (S2 Diff)—which guides diffusion sampling by learning the Control Lyapunov Barrier Function (CLBF) to achieve global safety and stability control for nonlinear, uncontrolled affine dynamics.
Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback
Jiaming Ji (Peking University), Yaodong Yang (Peking University)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This study proposes the Safe RLHF-V framework, which achieves safety alignment through multimodal RLHF, and creates the first open dataset BeaverTails-V containing dual preferences (usefulness and safety) and multi-level safety labels, while also designing a multi-level Guardrail system Beaver-Guard-V;
Safe-Sora: Safe Text-to-Video Generation via Graphical Watermarking
Zihan Su (Tsinghua University), Fei Yu
GenerationData SynthesisSafty and PrivacyVideoText
🎯 What it does: Directly embed visible graphic watermarks in the text-to-video generation framework to achieve copyright protection for generated videos.
SAFE: Multitask Failure Detection for Vision-Language-Action Models
Qiao Gu (University of Toronto), Florian Shkurti (University of Toronto)
Anomaly DetectionRobotic IntelligenceRecurrent Neural NetworkVision-Language-Action ModelMultimodality
🎯 What it does: A multi-task failure detection method called SAFE is proposed, which can identify failures in the execution of generalized Visual Language Action models (VLA) in real-time without collecting new task data.
Safely Learning Controlled Stochastic Dynamics
Luc Brogat-Motte (Italian Institute of Technology), Riccardo Bonalli (National Center for Scientific Research)
Safty and PrivacyReinforcement LearningTime Series
🎯 What it does: A provably safe algorithm is proposed for safely learning controlled stochastic dynamics from discrete time trajectory data without knowledge of the system dynamics, while keeping the system trajectories within a predetermined safe region during training and deployment.
SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment
Wonje Jeung (Yonsei University), Albert No (Yonsei University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: To address the issue of large reasoning models (LRM) generating dangerous outputs when receiving harmful prompts, the SAFEPATH method is proposed: it inserts an 8-word Safety Primer at the beginning of the reasoning process, maintaining the integrity of subsequent reasoning while suppressing harmful content.
SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism
Beitao Chen (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China), Lianli Gao (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes an untrained SafePTR framework to address multimodal jailbreak attacks in multimodal large language models (MLLMs) by pruning harmful tokens at vulnerable layers and restoring beneficial features in subsequent layers to enhance security.
Safety Depth in Large Language Models: A Markov Chain Perspective
Ching-Chia Kao (Academia Sinica), Chu-Song Chen (National Taiwan University)
Safty and PrivacyReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the concept of 'safety depth' and conducts a theoretical analysis of the safety alignment of large language models from the perspective of Markov chains. It further introduces cyclic group data augmentation techniques and ensemble methods to enhance the model's ability to reject harmful content at any output position. Experiments on six open-source LLMs demonstrate an improvement in safety scores.
Safety Pretraining: Toward the Next Generation of Safe AI
Pratyush Maini (Carnegie Mellon University), J Zico Kolter
Safty and PrivacyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A data center-based secure pre-training framework is proposed, embedding security into the model from the pre-training stage. The framework includes four main steps: security filtering, rewriting dangerous texts into safe contexts, active rejection training, and marking potential dangerous content with special labels during pre-training.
SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning
Borong Zhang (Peking University), Yaodong Yang (Peking University)
Robotic IntelligenceReinforcement LearningVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: An integrated safety approach called ISA is proposed, which aligns safety constraints through CMDP and safe reinforcement learning in the visual-language-action model.
SAFEx: Analyzing Vulnerabilities of MoE-Based LLMs via Stable Safety-critical Expert Identification
ZhengLin Lai, Jianqiang Li (Shenzhen University)
Safty and PrivacyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: For the Mixture-of-Experts (MoE) architecture of LLMs, a systematic analysis and quantification of its 'positional vulnerability' in safety alignment is conducted, and the SAFEX framework is proposed to identify, locate, and verify safety-critical experts; through expert-level masking and LoRA fine-tuning, the model's rejection rate of harmful prompts is validated and improved.
SaFiRe: Saccade-Fixation Reiteration with Mamba for Referring Image Segmentation
Zhenjie Mao (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
SegmentationTransformerSupervised Fine-TuningImageTextBenchmark
🎯 What it does: Designed and implemented the SaFiRe framework for handling ambiguous natural language descriptions in Referring Image Segmentation.
SAGE: A Unified Framework for Generalizable Object State Recognition with State-Action Graph Embedding
Yuan Zang (Brown University), Chen Sun (Brown University)
RecognitionTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: A unified object state recognition framework SAGE is proposed, which identifies the physical states of objects in videos by decomposing object states into fine-grained visual concepts and constructing a state-action graph shared across actions and objects.
SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training
Jintao Zhang (Tsinghua University), Jianfei Chen (Tsinghua University)
GenerationOptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageVideoTextMultimodality
🎯 What it does: Designed an FP4 micro-scale attention implementation of SageAttention3 for inference acceleration, and proposed a trainable 8-bit attention for training acceleration in SageBwd.
SAINT: Sequence-Aware Integration for Spatial Transcriptomics Multi-View Clustering
Zeyu Zhu (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
ClassificationRepresentation LearningGraph Neural NetworkTransformerMultimodalityBiomedical Data
🎯 What it does: A multimodal framework named SAINT is proposed, which integrates gene DNA sequence information (encoded by a pretrained Nucleotide Transformer) with spatial location and expression profiles for clustering in spatial transcriptomics.
Salient Concept-Aware Generative Data Augmentation
Tianchen Zhao, Yifan Xing (Amazon Web Services)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Based on a personalized image generation framework, a method called 'Significant Concept Awareness (SCA) Image Embedding' is proposed to generate diverse training samples that maintain key image concepts while aligning with text descriptions, achieving data augmentation through generation.
SALMONN-omni: A Standalone Speech LLM without Codec Injection for Full-duplex Conversation
Wenyi Yu (Tsinghua University), Chao Zhang (Tsinghua University)
Large Language ModelSupervised Fine-TuningReinforcement LearningTextAudio
🎯 What it does: This paper presents SALMONN-omni, a single LLM capable of full-duplex independent processing of speech input and output without audio stream injection.
SALoM: Structure Aware Temporal Graph Networks with Long-Short Memory Updater
Hanwen Liu (Zhejiang University), Mingli Song (Zhejiang University)
Recurrent Neural NetworkGraph Neural NetworkMixture of ExpertsGraphTime SeriesOrdinary Differential Equation
🎯 What it does: The SALoM framework is proposed, utilizing continuous time memory modules and Long Short-Term Memory Updater (LSMU), combining co-occurrence encoding and information bottleneck fusion to achieve unified modeling of long-term and short-term temporal dependencies as well as structural information in dynamic graphs.
SALS: Sparse Attention in Latent Space for KV Cache Compression
Junlin Mu (Beijing Jiaotong University), Yidong Li (ByteDance)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: In the area of KV cache compression and sparse attention for large-scale language models, the Sparse Attention in Latent Space (SALS) framework is proposed. It first projects the multi-head KV cache into a low-dimensional latent space, utilizes the low-rank features of the pre-RoPE key for token selection, reconstructs only a few key tokens, and performs sparse attention, significantly reducing KV cache usage and attention computation overhead.
SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning
Jiaqi Huang (Tsinghua University), Kehong Yuan (Tsinghua University)
SegmentationReinforcement LearningImageMultimodalityChain-of-Thought
🎯 What it does: Proposes the SAM-R1 framework, which utilizes reinforcement learning to use the Segment Anything Model as a reward signal, enhancing the performance of multimodal large models in fine-grained reasoning segmentation tasks.
SAM2Flow: Interactive Optical Flow Estimation with Dual Memory for in vivo Microcirculation Analysis
Luojie Huang (Johns Hopkins University), Nicholas J. Durr (Johns Hopkins University)
Recurrent Neural NetworkTransformerOptical FlowVideoBiomedical Data
🎯 What it does: An interactive optical flow estimation model, SAM2Flow, is proposed to extract microvascular blood flow information from OBM videos, allowing users to specify regions of interest (ROI) through point prompts.
SAMA: Towards Multi-Turn Referential Grounded Video Chat with Large Language Models
Ye Sun (Fudan University), Yu-Gang Jiang (Fudan University)
Object DetectionSegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningVideoTextBenchmark
🎯 What it does: Developed a multi-turn video dialogue system SAMA, providing unified video localization understanding, a video grounding dataset, model architecture, and benchmarks;
Same Task, Different Circuits: Disentangling Modality-Specific Mechanisms in VLMs
Yaniv Nikankin (Technion Israel Institute of Technology), Yonatan Belinkov (Technion Israel Institute of Technology)
Representation LearningTransformerVision Language ModelTextMultimodality
🎯 What it does: By comparing the computational subgraphs (circuits) of visual and text tasks, this study investigates the performance gap between visual and text in visual language models under the same task, and proposes a back-patching method during testing to improve the accuracy of visual tasks.
Sample and Map from a Single Convex Potential: Generation using Conjugate Moment Measures
Nina Vesseron (CREST-ENSAE IP Paris), marco cuturi
GenerationData SynthesisImageStochastic Differential Equation
🎯 What it does: Proposed a conjugate moment measure decomposition and designed two generation algorithms, CMFGen and CMFMA, based on this. It uses ICNN to learn the latent function w, and then generates samples through Langevin sampling and conjugate solving.
Sample complexity of data-driven tuning of model hyperparameters in neural networks with structured parameter-dependent dual function
Maria Florina Balcan, Dravyansh Sharma (Toyota Technological Institute at Chicago)
OptimizationHyperparameter SearchGraph Neural Network
🎯 What it does: A data-driven deep learning hyperparameter tuning framework is proposed, which learns a hyperparameter that performs well on average for future tasks using task distribution.
Sample Complexity of Distributionally Robust Average-Reward Reinforcement Learning
Zijun Chen (Hong Kong University of Science and Technology), Nian Si (Hong Kong University of Science and Technology)
Reinforcement Learning
🎯 What it does: This paper studies the distributionally robust average reward reinforcement learning problem and proposes two algorithms to achieve approximately optimal sample complexity.
Sample-Adaptivity Tradeoff in On-Demand Sampling
Nika Haghtalab (University of California), Mingda Qiao (University of Massachusetts Amherst)
🎯 What it does: This paper studies the trade-off between sample complexity and round complexity in Multi-Distribution Learning (MDL), providing approximately optimal solutions in both realizable and agnostic scenarios.