NeurIPS 2025 Papers — Page 45
Conference on Neural Information Processing Systems · 5275 papers
Statistical inference for Linear Stochastic Approximation with Markovian Noise
Sergey Samsonov (Higher School of Economics University), Alexey Naumov (Higher School of Economics University)
Reinforcement LearningTabular
🎯 What it does: This paper studies the non-asymptotic Berry-Esseen convergence rate of the Linear Stochastic Approximation (LSA) algorithm under Markov noise, and proposes a multiplier subsampling bootstrap method to achieve theoretical guarantees for confidence interval construction, which is then applied to TD learning for policy evaluation.
Statistical Inference under Performativity
Xiang Li (Independent Researcher), Zhun Deng (University of North Carolina at Chapel Hill)
Score-based ModelTabular
🎯 What it does: A complete statistical inference framework is proposed in the context of performativity, including the central limit theorem for the repeated risk minimization (RRM) process and the dynamic extension of prediction-driven inference (PPI).
Statistical Parity with Exponential Weights
Stephen Pasteris (Alan Turing Institute), Vasilios Mavroudis (Alan Turing Institute)
Optimization
🎯 What it does: This paper proposes an algorithm for effectively executing statistical parity in a contextual bandit setting, called SPEW (Statistical Parity with Exponential Weights), and demonstrates how it ensures statistical parity in each trial while maintaining strong performance guarantees.
Statistics Caching Test-Time Adaptation for Vision-Language Models
Zenghao Guan (Institute of Information Engineering Chinese Academy of Sciences), Xiaoyan Gu (Institute of Information Engineering Chinese Academy of Sciences)
Domain AdaptationComputational EfficiencyTransformerVision Language ModelImage
🎯 What it does: A statistical cache-based adaptive method SCA is designed, which utilizes feature statistics instead of raw features to achieve continuous knowledge accumulation, and enhances the performance of VLM on unlabeled test data through soft pseudo-labels and instance-level adaptive fusion.
STEAD: Robust Provably Secure Linguistic Steganography with Diffusion Language Model
Yuang Qi (University of Science and Technology of China), Kejiang Chen (University of Science and Technology of China)
Diffusion modelText
🎯 What it does: This paper proposes a robust and provably secure language steganography method called STEAD, which can locate robust embedding positions during the sequence generation process and achieve robust extraction.
Stealthy Yet Effective: Distribution-Preserving Backdoor Attacks on Graph Classification
Xiaobao Wang (Tianjin University), Di Jin (Tianjin University)
ClassificationAnomaly DetectionAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: A clean-label, distribution-preserving backdoor attack framework DPSBA is proposed for graph classification tasks.
SteerConf: Steering LLMs for Confidence Elicitation
Ziang Zhou (Hong Kong Polytechnic University), Li Qing
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The SteerConf framework is proposed, which guides the self-expression confidence of LLMs by adjusting the prompt level, and calibrates confidence using consistency of confidence and answer consistency to improve confidence calibration and error prediction.
Steering Generative Models with Experimental Data for Protein Fitness Optimization
Jason Yang (California Institute of Technology), Yisong Yue (California Institute of Technology)
OptimizationDrug DiscoveryReinforcement LearningDiffusion modelBiomedical Data
🎯 What it does: This study investigates a generative model-guided framework (SGPO) based on a small number of experimental labels for protein adaptive optimization.
Steering Information Utility in Key-Value Memory for Language Model Post-Training
Chunyuan Deng (Rice University), Hanjie Chen (Rice University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the InfoSteer method, which encourages the model to better utilize pre-trained knowledge during the post-training phase of LLM by performing forward intervention on the key-value structure of FFN and applying entropy regularization.
Steering When Necessary: Flexible Steering Large Language Models with Backtracking
Zifeng Cheng (Nanjing University), Qing Gu (Nanjing University)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a flexible activation scheduling and backtracking framework (FASB) to dynamically determine whether and how to intervene in the activation of LLM during the inference process.
StegoZip: Enhancing Linguistic Steganography Payload in Practice with Large Language Models
Jun Jiang (University of Science and Technology of China), Kejiang Chen (University of Science and Technology of China)
CompressionSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the StegoZip framework, which utilizes large language models for dynamic semantic redundancy pruning and index compression encoding of secret information, in order to enhance the payload of linguistic steganography while maintaining decodability.
StelLA: Subspace Learning in Low-rank Adaptation using Stiefel Manifold
Zhizhong Li (Sony AI), Lingjuan Lyu (Sony AI)
GenerationOptimizationSupervised Fine-TuningImageText
🎯 What it does: Based on LoRA, StelLA is proposed, which performs three-factor decomposition of USVᵀ and optimizes U and V on the Stiefel manifold, explicitly learning input/output subspaces to achieve parameter-efficient fine-tuning.
Stepsize anything: A unified learning rate schedule for budgeted-iteration training
Anda Tang (Peking University), Zhouchen Lin (Peking University)
OptimizationImageText
🎯 What it does: A unified, budget-aware learning rate scheduling method (UBA) is proposed, specifically designed to achieve optimal model performance under a fixed training iteration budget.
Stitch and Tell: A Structured Data Augmentation Method for Spatial Understanding
Hang Yin (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
Data SynthesisRepresentation LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A novel unlabelled, lightweight multimodal data augmentation method (Stitch and Tell, abbreviated as SiTe) is proposed, which enhances the spatial understanding ability of visual-language models by stitching images together and explicitly describing spatial relationships in text.
STITCH-OPE: Trajectory Stitching with Guided Diffusion for Off-Policy Evaluation
Hossein Goli (University of Toronto), Florian Shkurti (University of Toronto)
Reinforcement LearningDiffusion modelSequential
🎯 What it does: In high-dimensional, long-horizon offline reinforcement learning environments, the STITCH-OPE method is proposed for offline policy evaluation (OPE).
STNet: Spectral Transformation Network for Solving Operator Eigenvalue Problem
Hong Wang (University of Science and Technology of China), huanshuo dong
OptimizationComputational EfficiencyPoint CloudPhysics Related
🎯 What it does: Proposes STNet, which solves the operator eigenvalue problem by performing spectral transformations on operators during the iterative process of neural networks.
Stochastic Forward-Forward Learning through Representational Dimensionality Compression
Zhichao Zhu (Fudan University), Jianfeng Feng (Fudan University)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A positive-positive learning method that does not require negative samples is proposed, utilizing noise to generate similar samples and achieving unsupervised feature learning through effective dimensionality reduction.
Stochastic Gradients under Nuisances
Facheng Yu (University of Washington), Zaid Harchaoui
OptimizationStochastic Differential Equation
🎯 What it does: In the risk minimization problem with unknown nuisance parameters, the non-asymptotic convergence properties of Stochastic Gradient Descent (SGD) and its improved version (Orthogonalized SGD, OSGD) are proven, and fourth-order or second-order scaling rates for the error term are provided.
Stochastic Momentum Methods for Non-smooth Non-Convex Finite-Sum Coupled Compositional Optimization
Xingyu Chen (Texas A&M University), Tianbao Yang (Texas A&M University)
OptimizationTabular
🎯 What it does: This paper studies the finite-sum coupled composite optimization (FCCO) problem that is non-convex and non-smooth, and proposes a solution method based on stochastic momentum.
Stochastic Optimization in Semi-Discrete Optimal Transport: Convergence Analysis and Minimax Rate
Ferdinand Genans (Sorbonne Université), Olivier Wintenberger (Sorbonne Université)
OptimizationStochastic Differential Equation
🎯 What it does: This study investigates the stochastic gradient descent method in semi-discrete optimal transport, providing theoretical convergence guarantees and achieving unbiased estimates of the OT mapping, cost, and potential.
Stochastic Principal-Agent Problems: Computing and Learning Optimal History-Dependent Policies
Jiarui Gan (University of Oxford), Goran Radanovic (Max Planck Institute for Software Systems)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes two types of algorithms under the framework of random master-slave agents: one type efficiently computes near-optimal history-dependent strategies when the transition probabilities are known; the other type implements a learning algorithm with sublinear regret through reward-free exploration and approximately incentive-compatible strategies when the transitions are unknown.
Stochastic Process Learning via Operator Flow Matching
Yaozhong Shi (California Institute of Technology), Kamyar Azizzadenesheli (NVIDIA Corporation)
Flow-based ModelTime Series
🎯 What it does: The Operator Flow Matching (OFM) framework is proposed to learn the prior of stochastic processes in arbitrary domains and achieve analytical density and regression in function spaces.
Stochastic Regret Guarantees for Online Zeroth- and First-Order Bilevel Optimization
Parvin Nazari (Amirkabir University of Technology), George Michailidis (University of California, Los Angeles)
OptimizationAdversarial AttackImageStochastic Differential Equation
🎯 What it does: An online two-layer optimization framework is proposed, utilizing a new search direction to achieve windowless smooth sublinear Bellman review loss guarantees;
Stochastic Shortest Path with Sparse Adversarial Costs
Emmeran Johnson (Imperial College London), Patrick Rebeschini (University of Oxford)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the sparse adversarial cost random shortest path (SSP) problem under full information feedback, proposing a new regularization method to accommodate sparsity and demonstrating its minimization regret bounds in known transition scenarios.
Stochastically Dominant Peer Prediction
Yichi Zhang (Rutgers University), David Pennock (Rutgers University)
Text
🎯 What it does: This paper proposes and analyzes the 'stochastically dominant truthful' (SD-truthful) peer prediction mechanism, designing direct rollback, partition rollback, and a new enforcement agreement (EA) mechanism, and evaluates their sensitivity and budget efficiency through theoretical derivation and experimental assessment.
Stop Summation: Min-Form Credit Assignment Is All Process Reward Model Needs for Reasoning
Jie Cheng (Institute of Automation, Chinese Academy of Sciences), Fei-Yue Wang (Institute of Automation, Chinese Academy of Sciences)
TransformerReinforcement LearningText
🎯 What it does: A reinforcement learning framework called PURE was designed and validated, utilizing a process reward model (PRM) to address the issue of reward hijacking that often occurs during PRM training.
Storyboard-guided Alignment for Fine-grained Video Action Recognition
Enqi Liu (Beijing Institute of Technology), Liu Liu (Huawei)
RecognitionTransformerLarge Language ModelContrastive LearningVideoText
🎯 What it does: A SFAR framework is proposed, which automatically generates and filters fine-grained sub-texts of actions through large language models, combining global semantics to generate coarse and fine dual video embeddings, thereby enhancing video action recognition.
Straight-Line Diffusion Model for Efficient 3D Molecular Generation
Yuyan Ni (Chinese Academy of Sciences), Yanyan Lan (Tsinghua University)
GenerationDrug DiscoveryDiffusion modelGraphStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The Straight-Line Diffusion Model (SLDM) is proposed, achieving efficient 3D molecular generation by designing a linear noise decay trajectory.
STRAP: Spatio-Temporal Pattern Retrieval for Out-of-Distribution Generalization
Haoyu Zhang (City University of Hong Kong), Yifan Zhang (City University of Hong Kong)
RetrievalDomain AdaptationGraph Neural NetworkReinforcement LearningTime Series
🎯 What it does: This paper addresses the generalization problem of spatiotemporal graph neural networks in spatiotemporal out-of-distribution (STOOD) scenarios by proposing the STRAP framework, which constructs a spatiotemporal pattern library and enhances continuous prediction performance through retrieval-augmented learning.
Strassen Attention, Split VC Dimension and Compositionality in Transformers
Alexander Kozachinskiy (CENIA), Cristobal Rojas
TransformerSequential
🎯 What it does: This study investigates the theoretical limitations of a single-layer softmax Transformer in combinatorial reasoning tasks (function composition, Match3, binary relation composition, etc.) and proposes the Strassen attention mechanism, proving that it can solve these tasks with sub-cubic complexity.
Strategic Classification with Non-Linear Classifiers
Benyamin Trachtenberg (Technion - Israel Institute of Technology), Nir Rosenfeld (Technion - Israel Institute of Technology)
ClassificationOptimizationTabular
🎯 What it does: This study investigates how nonlinear classifiers perform and their impact on learning in scenarios where users can modify features based on the classifier.
Strategic Cost Selection in Participatory Budgeting
Piotr Faliszewski (AGH University of Science and Technology), Mateusz Szwagierczak (AGH University of Science and Technology)
OptimizationTabular
🎯 What it does: This study investigates how project proposers in participatory budgeting (PB) can achieve optimal benefits by setting project costs when they are aware of the voting outcomes, and analyzes whether pure Nash equilibria exist under different PB rules.
Strategic Costs of Perceived Bias in Fair Selection
L. Elisa Celis, Nisheeth K. Vishnoi
Tabular
🎯 What it does: This paper constructs a dual-group strategy competition model to study how the perceived value differences of a group after being selected influence performance, representation, social welfare, and institutional gains through rational investment under a fair selection system. It derives a unique Nash equilibrium in the large-sample limit and provides threshold strategies and closed-form solutions.
Strategic Hypothesis Testing
Yatong Chen (Max Planck Institute for Intelligent Systems), Yiling Chen (Harvard University)
Drug DiscoveryAgentic AITabular
🎯 What it does: The paper discusses hypothesis testing within a principal-agent framework, where a strategic agent holds private beliefs about the effectiveness of a product and submits data to the principal to decide whether to approve it. The principal uses hypothesis testing rules aimed at selecting a p-value threshold to balance false positives and false negatives while considering the agent's incentives to maximize expected profits.
Strategyproof Reinforcement Learning from Human Feedback
Thomas Kleine Buening (ETH Zurich), Marta Kwiatkowska (University of Oxford)
Reinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper studies the provability of strategies in reinforcement learning with human feedback (RLHF) methods in a multi-labeler environment, proposing a Pessimistic Median of MLEs algorithm based on maximum likelihood estimation confidence sets, which achieves approximate social welfare optimality while ensuring approximate strategy provability.
Stratify or Die: Rethinking Data Splits in Image Segmentation
Naga Venkata Sai Jitin Jami (Friedrich-Alexander University Erlangen-Nuremberg), Heike Leutheuser (University of Bayreuth)
SegmentationOptimizationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: A dataset partitioning method for image segmentation is proposed, which includes pixel-level iterative partitioning (IPS) and genetic partitioning based on Wasserstein distance (WDES);
STRATUS: A Multi-agent System for Autonomous Reliability Engineering of Modern Clouds
Yinfang Chen (University of Illinois Urbana-Champaign), Tianyin Xu (University of Illinois Urbana-Champaign)
Anomaly DetectionOptimizationTransformerLarge Language ModelAgentic AITabularBenchmark
🎯 What it does: STRATUS has been constructed, a multi-agent system based on large language models, to achieve autonomous reliability engineering (SRE) for cloud services, including fault detection, localization, root cause analysis, and automatic mitigation; a complete transaction safety mechanism has been implemented on cloud platforms (such as Kubernetes);
StreamBP: Memory-Efficient Exact Backpropagation for Long Sequence Training of LLMs
Qijun Luo (Chinese University of Hong Kong), Xiao Li (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: A memory-efficient and precise backpropagation method named StreamBP has been developed for training large-scale language models on long sequences.
StreamBridge: Turning Your Offline Video Large Language Model into a Proactive Streaming Assistant
Haibo Wang (Apple), Ping Huang (Apple)
GenerationRecommendation SystemTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: Transforming offline Video-LLM into a real-time streaming video assistant that supports multi-turn interaction and proactive responses.
StreamForest: Efficient Online Video Understanding with Persistent Event Memory
Xiangyu Zeng (Zhejiang University), Limin Wang (Noah's Ark Lab, Huawei)
RecognitionAutonomous DrivingComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVideoMultimodalityBenchmark
🎯 What it does: The StreamForest framework is proposed for real-time video understanding, consisting of two main modules: Fine-grained Spatiotemporal Window and Persistent Event Memory Forest, achieving efficient long-term memory and immediate perception.
Streaming Attention Approximation via Discrepancy Theory
Ekaterina Kochetkova (École Polytechnique Fédérale de Lausanne), Michael Kapralov
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes BalanceKV, a streaming attention approximation algorithm based on divergence theory, aimed at compressing key-value caches in LLMs and accelerating long-context generation.
Streaming Audio Generation from Discrete Tokens via Streaming Flow Matching
Ha-Yeong Choi (KT Corp), Sang-Hoon Lee (Ajou University)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelAudio
🎯 What it does: The StreamFlow model for real-time streaming audio generation is proposed.
Streaming Federated Learning with Markovian Data
Tan-Khiem HUYNH, Jean-Marie Gorce (Inria)
Federated LearningTime Series
🎯 What it does: This paper studies how to achieve effective collaborative learning in a federated learning scenario where client data is Markovian (non-i.i.d.);
Streaming Stochastic Submodular Maximization with On-Demand User Requests
Honglian Wang (KTH Royal Institute of Technology), Aristides Gionis (KTH Royal Institute of Technology)
Recommendation SystemOptimizationTabular
🎯 What it does: This paper proposes a new streaming stochastic submodular maximization problem—S3MOR, which addresses the issue of users accessing and only being able to display k articles at once in the news recommendation scenario, and designs various low-memory algorithms.
STree: Speculative Tree Decoding for Hybrid State Space Models
Yangchao Wu (University of California Los Angeles), Stefano Soatto (University of California Los Angeles)
TransformerLarge Language ModelText
🎯 What it does: This paper introduces STree, the first scalable tree-based speculative decoding algorithm for state space models (SSM) and their hybrid architecture with Transformers, enabling parallel generation of multi-step outputs and tree-based validation in a single forward pass.
STRIDER: Navigation via Instruction-Aligned Structural Decision Space Optimization
Diqi He (Northwestern Polytechnical University), Dingwen Zhang (Northwestern Polytechnical University)
OptimizationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes a zero-shot continuous visual and language navigation framework called STRIDER, which achieves efficient navigation in unseen 3D environments through structured path planning and task alignment adjustment.
Struct2D: A Perception-Guided Framework for Spatial Reasoning in MLLMs
Fangrui Zhu (Northeastern University), Huaizu Jiang (Northeastern University)
RecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringMultimodalityPoint CloudBenchmark
🎯 What it does: This paper proposes a structured 2D prompt framework, Struct2D, which utilizes BEV images, object tags, and metadata to enable multimodal large language models (MLLMs) to perform 3D spatial reasoning with only 2D inputs, and constructs a large-scale instruction tuning dataset, Struct2D-Set.
Structural Causal Bandits under Markov Equivalence
Min Woo Park (Seoul National University), Sanghack Lee (Seoul National University)
🎯 What it does: A structural causal Bandit framework is proposed under the condition of only obtaining the Markov equivalence class (PAG), along with graphical criteria and efficient enumeration algorithms for determining the minimal intervention set and the possibly optimal minimal intervention set.
Structural Entropy Guided Agent for Detecting and Repairing Knowledge Deficiencies in LLMs
Yifan Wei (Beihang University), Li Du (Beijing Academy of Artificial Intelligence)
Data SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringBiomedical Data
🎯 What it does: The SENATOR framework is proposed, which uses structure entropy-guided Monte Carlo tree search to locate knowledge gaps of LLMs on knowledge graphs and generate targeted synthetic data for self-repair.
Structural Information-based Hierarchical Diffusion for Offline Reinforcement Learning
Xianghua Zeng (Beihang University), Guanlin Wu (National University of Defense Technology)
Reinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: This paper proposes SIHD, an adaptive multi-scale hierarchical diffusion framework that utilizes offline trajectory structural information to address long-term sparse reward tasks in offline reinforcement learning (RL).
Structure Matters: Dynamic Policy Gradient
Sara Klein (August-Wilhelm Scheer Institute for Digital Products and Processes), Leif Döring (Institut for Mathematics University of Mannheim)
OptimizationReinforcement Learning
🎯 What it does: This paper addresses the γ-discount infinite Markov Decision Process (MDP) and proposes a Dynamic Policy Gradient (DynPG) framework, which combines the Bellman update of dynamic programming with policy gradient dynamics. This allows for the gradual expansion of the problem's time domain by training a contextual bandit at each step, thereby converging to the optimal policy.
Structure-Aware Cooperative Ensemble Evolutionary Optimization on Combinatorial Problems with Multimodal Large Language Models
Jie Zhao (Nanyang Technological University), Kang Hao Cheong (Nanyang Technological University)
OptimizationGraph Neural NetworkLarge Language ModelMultimodalityGraph
🎯 What it does: This paper proposes a structure-aware collaborative and integrated evolutionary optimization framework, using multimodal large language models (MLLM) as evolutionary operators to achieve better search results in graph-structured combinatorial optimization problems.
Structure-Aware Fusion with Progressive Injection for Multimodal Molecular Representation Learning
Zihao Jing (Western University), Pingzhao Hu (Western University)
Representation LearningDrug DiscoveryGraph Neural NetworkMultimodalityGraph
🎯 What it does: The MuMo framework is proposed to achieve multi-modal molecular representation learning through structured fusion and progressive injection.
Structure-Aware Spectral Sparsification via Uniform Edge Sampling
Kaiwen He (Purdue University), Rajiv Khanna (Purdue University)
Graph Neural NetworkGraph
🎯 What it does: The study uses uniform edge sampling for spectral sparsification on well-structured graphs (with a large structural ratio Υ(k)), providing theoretical guarantees and experimental validation, demonstrating that uniform sampling can preserve spectral subspaces in such graphs, thereby ensuring the effectiveness of spectral clustering.
Structured Initialization for Vision Transformers
Jianqiao Zheng (Australian Institute for Machine Learning), Simon Lucey (Australian Institute for Machine Learning)
TransformerImage
🎯 What it does: This paper proposes embedding convolutional pulse filters into attention maps in Vision Transformer (ViT) solely through structured initialization to enhance performance on small-scale datasets while maintaining performance on large-scale datasets.
Structured Linear CDEs: Maximally Expressive and Parallel-in-Time Sequence Models
Benjamin Walker (Mathematical Institute, University of Oxford), Terry Lyons (Mathematical Institute, University of Oxford)
ClassificationOptimizationComputational EfficiencyTime SeriesSequentialBenchmarkOrdinary Differential Equation
🎯 What it does: This paper proposes the Structured Linear Controlled Differential Equations (SLiCE) framework, which unifies various structured, input-dependent state transition matrix sequence models. It theoretically proves that block, sparse, and Walsh-Hadamard structures can achieve maximum probabilistic expressiveness. Additionally, experiments validate the superior performance of SLiCE in state tracking, regular languages, and multivariate time series classification tasks.
Structured Reinforcement Learning for Combinatorial Decision-Making
Heiko Hoppe (Technical University of Munich), Maximilian Schiffer (Technical University of Munich)
OptimizationReinforcement LearningGaussian Splatting
🎯 What it does: This paper studies an actor-critic structure embedded with a combinatorial optimization layer to address the reinforcement learning problem of combinatorial MDPs.
Structured Sparse Transition Matrices to Enable State Tracking in State-Space Models
Aleksandar Terzic, Abbas Rahimi (IBM Research)
TransformerTime SeriesSequential
🎯 What it does: A structurally sparse PD-SSM (P×D) state space model is proposed, which can efficiently simulate any N-state finite automaton in a single-layer, N-dimensional state space.
Structured Spectral Reasoning for Frequency-Adaptive Multimodal Recommendation
Wei Yang (Kuaishou Technology), Peng Jiang (Kuaishou Technology)
Recommendation SystemGraph Neural NetworkContrastive LearningMultimodalityGraph
🎯 What it does: This paper addresses issues such as modal noise, semantic inconsistency, and unstable graph propagation in multimodal recommendation systems by proposing the Structured Spectral Reasoning (SSR) framework. It employs a four-stage processing flow involving frequency domain decomposition, spectral band masking, hyperspectral fusion, and spectral contrastive regularization to achieve structured reasoning and adaptive modulation of multimodal signals.
Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection
Dongchan Cho (SimPlatform Co. Ltd.), Namsoon Jung (SimPlatform Co. Ltd.)
Anomaly DetectionExplainability and InterpretabilityRecurrent Neural NetworkAuto EncoderTime Series
🎯 What it does: This paper presents OracleAD, an unsupervised multivariate time series anomaly detection framework that captures temporal causality using causal embeddings for each variable. It maps the embeddings to a shared latent space through self-attention and subsequently compares them with a Stable Latent Structure (SLS) to achieve anomaly detection and root cause localization through dual scoring of prediction error and structural deviation.
StruDiCO: Structured Denoising Diffusion with Gradient-free Inference-stage Boosting for Memory and Time Efficient Combinatorial Optimization
Yu Wang (Jilin University), Yi Chang (Jilin University)
OptimizationDiffusion modelGraph
🎯 What it does: A structured denoising diffusion framework, StruDiCO, is proposed to gradually construct interpretable combinatorial optimization solutions during the inference process.
Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles
Peng Wang (Zhejiang University), Peidong Liu (Westlake University)
GenerationData SynthesisTransformerGaussian SplattingImage
🎯 What it does: An end-to-end feedforward network named Styl3R is proposed, capable of generating multi-view consistent 3D high-quality stylized reconstructions from sparse, pose-free view images and arbitrary style images in less than 0.15 seconds.
StyleGuard: Preventing Text-to-Image-Model-based Style Mimicry Attacks by Style Perturbations
Yanjie Li (Hong Kong Polytechnic University), Bin Xiao (Hong Kong Polytechnic University)
GenerationAdversarial AttackMeta LearningDiffusion modelImage
🎯 What it does: This paper proposes a style imitation protection method called StyleGuard, based on latent space style perturbations, to prevent unauthorized style replication of artists' works by text-to-image diffusion models (such as DreamBooth and Textual Inversion).
Subgraph Federated Learning via Spectral Methods
Javad Aliakbari (Chalmers University of Technology AI Sweden), Alexandre Graell i Amat (Chalmers University of Technology AI Sweden)
Federated LearningSafty and PrivacyGraph Neural NetworkGraph
🎯 What it does: Design and propose two subgraph federated learning frameworks, FEDLAP and FEDLAP+, which utilize Laplacian smoothing to extract global structural information in the spectral domain, achieving efficient and privacy-preserving node classification.
Subsampled Ensemble Can Improve Generalization Tail Exponentially
Huajie Qian (Alibaba Group), Wotao Yin (Alibaba Group)
Tabular
🎯 What it does: Achieve voting-based ensemble by training multiple times on subsamples and taking the most frequently occurring model (or ε-optimal voting), thereby improving the model's generalization performance.
Subspace Networks: Scaling Decentralized Training with Communication-Efficient Model Parallelism
Sameera Ramasinghe (Pluralis Research), Alexander Long (Pluralis Research)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: A lossless compression algorithm for model parallelism is proposed, utilizing low-rank subspace constraints to compress activations and gradients during forward and backward propagation, achieving up to 100 times communication compression, supporting training of models with hundreds of billions of parameters over low-bandwidth networks such as 80 Mbps.
SubTrack++ : Gradient Subspace Tracking for Scalable LLM Training
Sahar Rajabi (University of Waterloo), Sirisha Rambhatla (University of Waterloo)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The SubTrack++ method is proposed, which improves the memory and time efficiency of LLM training by tracking gradient subspaces on the Grassmannian and combining it with a projection-aware optimizer;
Succeed or Learn Slowly: Sample Efficient Off-Policy Reinforcement Learning for Mobile App Control
Georgios Papoudakis (Huawei Noah's Ark Lab), Kun Shao (Huawei Noah's Ark Lab)
Supervised Fine-TuningReinforcement LearningSequential
🎯 What it does: In mobile application control tasks, an offline reinforcement learning algorithm called SoLS aimed at sample efficiency is proposed, trained in conjunction with Successful Transition Replay (STR).
SuffixDecoding: Extreme Speculative Decoding for Emerging AI Applications
Gabriele Oliaro (Carnegie Mellon University), Aurick Qiao (Snowflake AI Research)
Computational EfficiencyLarge Language ModelAgentic AITextBenchmark
🎯 What it does: A model-free speculative decoding method based on suffix trees, called SuffixDecoding, is proposed to accelerate inference in proxy-based AI applications.
Sum Estimation under Personalized Local Differential Privacy
Dajun Sun (Hong Kong University of Science and Technology), Graham Cormode (University of Warwick)
Safty and PrivacyGaussian SplattingImageTabular
🎯 What it does: The research addresses the problem of sum/mean estimation under personalized local differential privacy and proposes two new protocols.
SUMO: Subspace-Aware Moment-Orthogonalization for Accelerating Memory-Efficient LLM Training
Yehonathan Refael (Tel Aviv University), Ofir Lindenbaum (Bar Ilan University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: An optimizer named SUMO is proposed, which utilizes precise singular value decomposition (SVD) in low-dimensional subspaces to orthogonalize the first moment of gradients, accelerating memory-efficient training of LLMs;
SuperCLIP: CLIP with Simple Classification Supervision
Weiheng Zhao (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
ClassificationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: In the CLIP contrastive learning framework, a lightweight linear layer is added to supervise the visual encoder to use the original text subwords as classification labels, thereby recovering fine-grained semantic information from the text.
Superposition Yields Robust Neural Scaling
Yizhou Liu (Massachusetts Institute of Technology), Jeff Gore (Massachusetts Institute of Technology)
TransformerLarge Language ModelAuto EncoderText
🎯 What it does: This study investigates the impact of representation superposition on loss scaling in large models, proposing two scenarios: weak superposition and strong superposition, and provides corresponding loss scaling laws.
Support Vector Generation: Kernelizing Large Language Models for Efficient Zero‑Shot NLP
Shohei Ohsawa
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a Support Vector Generation (SVG) method that transforms frozen language models into interpretable, training-independent kernel classifiers.
Surface-Aware Feed-Forward Quadratic Gaussian for Frame Interpolation with Large Motion
Zaoming Yan (East China Normal University), Faming Fang (East China Normal University)
RestorationData SynthesisGaussian SplattingVideo
🎯 What it does: Proposes a Surface-Aware Feed-Forward Quadratic Gaussian method that maps video frames to a three-dimensional differential surface, utilizing surface normals and curvature to achieve global alignment, thereby enhancing video frame interpolation effects in large motion scenes.
SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction
Chensheng Dai (Tsinghua University), Yueqi Duan (Tsinghua University)
Depth EstimationComputational EfficiencyRepresentation LearningTransformerGaussian SplattingPoint CloudBenchmark
🎯 What it does: SurfelSplat is proposed, achieving efficient and generalizable surface reconstruction under sparse views through Gaussian surface representation guided by Nyquist sampling.
SViMo: Synchronized Diffusion for Video and Motion Generation in Hand-object Interaction Scenarios
Lingwei Dang (South China University of Technology), Qingyao Wu (South China University of Technology)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderVideoPoint Cloud
🎯 What it does: Proposes the SViMo synchronous diffusion framework, which jointly generates hand-object interaction videos and corresponding 3D motions;
SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution
Yuxiang Wei (University of Illinois Urbana-Champaign), Sida Wang
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the SWE-RL method, which utilizes rule-based rewards and real open-source software evolution data (PR) to perform reinforcement learning on large language models, thereby enhancing their reasoning and repair capabilities in the field of software engineering, particularly in solving real GitHub issues.
SWE-SQL: Illuminating LLM Pathways to Solve User SQL Issues in Real-World Applications
Jinyang Li (Hong Kong University), Reynold Cheng (Chinese University of Hong Kong)
Large Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: This paper proposes the BIRD-CRITIC SQL debugging benchmark and the SIX-GYM training environment, and based on this, constructs the BIRD-FIXER open-source model, achieving automatic diagnosis and repair of real user SQL errors.
Switchable Token-Specific Codebook Quantization For Face Image Compression
Yongbo Wang (East China Normal University), Shouhong Ding (Tencent)
RecognitionCompressionMixture of ExpertsAuto EncoderImage
🎯 What it does: A switchable, token-specific codebook quantization method for facial image compression is proposed, enhancing reconstruction and recognition performance at low bpp.
SwS: Self-aware Weakness-driven Problem Synthesis in Reinforcement Learning for LLM Reasoning
Xiao Liang (University of California), Weizhu Chen (Microsoft)
Large Language ModelReinforcement LearningText
🎯 What it does: An adaptive weakness-driven problem synthesis framework, SwS, is proposed. In the training of Reinforcement Learning-based Visual Reasoning (RLVR), weaknesses are automatically identified based on the model's continuous failures during the pre-training phase, and targeted new problems are generated accordingly. These synthesized problems are then added to the training set for reinforcement learning.
SymMaP: Improving Computational Efficiency in Linear Solvers through Symbolic Preconditioning
Hong Wang (University of Science and Technology of China), Haoyang Liu (University of Science and Technology of China)
OptimizationComputational EfficiencyRecurrent Neural NetworkReinforcement LearningTabular
🎯 What it does: The SymMaP framework is proposed to learn efficient preprocessing parameter expressions through symbolic discovery, enhancing the efficiency of solving linear systems.
Symmetry-Preserving Conformer Ensemble Networks for Molecular Representation Learning
Yanqiao Zhu (University of California), Wei Wang (University of California)
Representation LearningDrug DiscoveryGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: A new molecular representation learning framework called SPiCE is proposed, which can predict molecular properties using a conformer ensemble rather than a single static conformation.
SYMPHONY: Synergistic Multi-agent Planning with Heterogeneous Language Model Assembly
Wei Zhu (Yunnan University), Kun Yue (Yunnan University)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: A multi-agent planning framework named SYMPHONY is proposed, which integrates various large language models (LLMs) into Monte Carlo Tree Search (MCTS) to enhance search diversity and planning efficiency through model heterogeneity.
SymRTLO: Enhancing RTL Code Optimization with LLMs and Neuron-Inspired Symbolic Reasoning
Yiting Wang (University of Maryland), Ang Li (University of Maryland)
OptimizationAI Code AssistantTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the SymRTLO framework, which combines large language models with symbolic reasoning to achieve automatic rewriting and optimization of RTL code.
SynBrain: Enhancing Visual-to-fMRI Synthesis via Probabilistic Representation Learning
Weijian Mai (Shanghai Artificial Intelligence Laboratory), Chunfeng Song (Shanghai Artificial Intelligence Laboratory)
Data SynthesisRepresentation LearningTransformerAuto EncoderContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: The SynBrain framework is proposed, which models the mapping from vision to fMRI using probability distributions, simulating one-to-many neural variations while maintaining functional consistency, and supporting few-shot cross-subject adaptation.
SyncHuman: Synchronizing 2D and 3D Generative Models for Single-view Human Reconstruction
Wenyue Chen (Peking University), Yuan Liu (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelImageMesh
🎯 What it does: A system has been developed to generate high-quality full-body 3D human meshes from a single color image, combining 2D multi-view generation and 3D native generation.
SynCL: A Synergistic Training Strategy with Instance-Aware Contrastive Learning for End-to-End Multi-Camera 3D Tracking
Shubo Lin (Chinese Academy of Sciences), Jin Gao
Object TrackingAutonomous DrivingTransformerContrastive LearningPoint Cloud
🎯 What it does: To address the challenge of joint training for detection and tracking in multi-camera 3D multi-object tracking, a SynCL training strategy is proposed.
Synergistic Tensor and Pipeline Parallelism
Mengshi Qi (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
TransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: A collaborative scheduling scheme is proposed to simultaneously reduce the communication bottleneck of tensor parallelism (TP) and the synchronization bottleneck of pipeline parallelism (PP). It interleaves fine-grained computation blocks to form woven execution blocks, thereby achieving efficient overlap between TP communication and PP computation.
Synergy Between the Strong and the Weak: Spiking Neural Networks are Inherently Self-Distillers
Yongqi Ding (University of Electronic Science and Technology of China), Tonglan Xie (University of Electronic Science and Technology of China)
Computational EfficiencyKnowledge DistillationSpiking Neural NetworkImageTime Series
🎯 What it does: Decomposing multi-temporal SNN into sub-models and self-distilling to improve performance
Synergy over Discrepancy: A Partition-Based Approach to Multi-Domain LLM Fine-Tuning
Hua Ye (Nanjing University), Xuan Zhang (Carnegie Mellon University)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a partition-based multi-stage fine-tuning framework aimed at effectively adapting large language models (LLMs) across multiple heterogeneous domains, leveraging inter-domain synergy and mitigating negative transfer.
SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond
Junteng Liu (Hong Kong University of Science and Technology), Junxian He (Hong Kong University of Science and Technology)
Data SynthesisReinforcement LearningText
🎯 What it does: A scalable logical reasoning data synthesis framework, SYNLOGIC, has been constructed, generating verifiable reasoning data with 35 tasks and adjustable difficulty; reinforcement learning (RL) training has been conducted based on this data; furthermore, mixing logical data with mathematical and coding data has improved multi-task learning efficiency; state-of-the-art performance has been achieved on multiple logical and mathematical benchmarks.
Synthesize Privacy-Preserving High-Resolution Images via Private Textual Intermediaries
Haoxiang Wang (Peking University), Huishuai Zhang (Peking University)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: By first converting private images into textual descriptions, then using a private evolutionary algorithm to generate differentially private sentences in the text domain, and finally reconstructing these sentences into high-resolution synthetic images using a text-to-image diffusion model, this approach achieves differential privacy image generation without training costs.
Synthesizing Performance Constraints for Evaluating and Improving Code Efficiency
Jun Yang (University of Chicago), Kexin Pei (University of Chicago)
OptimizationComputational EfficiencyLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes the WEDGE framework, which utilizes LLM to synthesize performance constraints and combines coverage-guided fuzz testing to automatically generate efficient stress tests for code implementations.
Synthesizing Photorealistic and Dynamic Urban Environments for Multimodal Robot Navigation and Collaboration
Yan Zhuang (University of Virginia), Tianmin Shu
Robotic IntelligenceReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper presents the SimWorld-Robotics platform for generating infinite photorealistic dynamic urban environments, and based on this platform, it creates two new benchmarks: multimodal navigation and multi-robot search, while constructing a training set of 20K steps.
Synthetic Series-Symbol Data Generation for Time Series Foundation Models
Wenxuan Wang (Xidian University), Xiaoyu Zhang (Xidian University)
Data SynthesisAnomaly DetectionTransformerContrastive LearningTime SeriesSequential
🎯 What it does: This paper proposes a method for infinitely generating time-series-symbol (S²) data pairs and based on this, the SymTime temporal pre-training model trained on large-scale synthetic data.
Synthetic-powered predictive inference
Meshi Bashari (Technion Israel Institute of Technology), Yaniv Romano (Technion Israel Institute of Technology)
Data SynthesisComputational EfficiencyDiffusion modelImageTabular
🎯 What it does: This paper proposes Synthetic-Powered Predictive Inference (SPI), a framework that utilizes synthetic data to enhance the efficiency of synthetic calibration samples while maintaining distribution-independent coverage guarantees.
System Prompt Optimization with Meta-Learning
Yumin Choi (Korea Advanced Institute of Science and Technology), Sung Ju Hwang (Korea Advanced Institute of Science and Technology)
OptimizationMeta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A bilevel system prompt optimization framework is proposed and implemented, utilizing meta-learning to automatically optimize system prompts across multiple tasks and domains, enhancing the generalization and adaptation performance of large language models on unknown tasks and user prompts.
System-1.5 Reasoning: Traversal in Language and Latent Spaces with Dynamic Shortcuts
Xiaoqiang Wang (Université de Montréal), Bang Liu (Université de Montréal)
Computational EfficiencyKnowledge DistillationTransformerTextChain-of-Thought
🎯 What it does: Proposes the System-1.5 Reasoning framework, which uses dynamic short-circuiting in the latent space to achieve efficient reasoning.
System-Embedded Diffusion Bridge Models
Bartlomiej Sobieski (University of Warsaw), Przemyslaw Biecek (Warsaw University of Technology)
RestorationSuper ResolutionDiffusion modelImageBiomedical DataMagnetic Resonance ImagingComputed TomographyStochastic Differential Equation
🎯 What it does: A system-embedded diffusion bridge model (SDB) is proposed, which achieves a generative solution to the inverse problem by directly embedding the linear measurement system into the coefficients of matrix-valued stochastic differential equations;