NeurIPS 2025 Papers — Page 44
Conference on Neural Information Processing Systems · 5275 papers
Sparta Alignment: Collectively Aligning Multiple Language Models through Combat
Yuru Jiang (Zhejiang University), Yulia Tsvetkov (University of Washington)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: An algorithm named SPARTA ALIGNMENT is proposed, which collectively aligns multiple large language models (LLMs) through competition and adversarial methods. This algorithm overcomes the shortcomings of a single model in generating diversity and assessing bias by allowing multiple models to compete with each other.
SPARTAN: A Sparse Transformer World Model Attending to What Matters
Anson Lei (University of Oxford), Ingmar Posner (University of Oxford)
Autonomous DrivingTransformerWorld ModelTime Series
🎯 What it does: SPARTAN is proposed, a structured world model based on a Transformer architecture that learns local causal structures through sparse attention, used for predicting object states and achieving rapid adaptation.
Spatial Understanding from Videos: Structured Prompts Meet Simulation Data
Haoyu Zhang (Harbin Institute of Technology), Liqiang Nie (Pengcheng Laboratory)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: By introducing structured prompts (SpatialMind) and a large-scale synthetic question-answer dataset (ScanForgeQA), the ability of pre-trained visual language models to perform three-dimensional spatial reasoning based solely on scanned videos is enhanced.
Spatial-Aware Decision-Making with Ring Attractors in Reinforcement Learning Systems
Marcos Negre Saura, Theodore Papamarkou (PolyShape)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper proposes embedding the Ring Attractor into the reinforcement learning decision-making process to explicitly encode spatial relationships into the action space, achieving more efficient action selection.
Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence
Diankun Wu (Tsinghua University), Yueqi Duan (Tsinghua University)
RecognitionOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodality
🎯 What it does: This paper proposes Spatial-MLLM, which utilizes a dual encoder (2D semantic encoder + 3D spatial encoder) and spatially aware frame sampling to achieve strong spatial understanding and reasoning with only 2D video data.
SpatialLM: Training Large Language Models for Structured Indoor Modeling
Yongsen Mao (Manycore Tech Inc), Zihan Zhou (Manycore Tech Inc)
Object DetectionGenerationTransformerLarge Language ModelSupervised Fine-TuningPoint Cloud
🎯 What it does: Train and fine-tune the large language model SPATIALLM to accept 3D point clouds and generate structured scene descriptions containing walls, doors, windows, and 3D object bounding boxes (output in the form of Python scripts).
Spatially-aware Weights Tokenization for NeRF-Language Models
Andrea Amaduzzi (University of Bologna), Luigi Di Stefano (University of Bologna)
TransformerLarge Language ModelNeural Radiance FieldTextMultimodality
🎯 What it does: This paper proposes a method to directly extract spatially aware token sequences from NeRF MLP weights, and based on this, constructs a new multimodal large language model called Spatial LLaNA for understanding and interacting with NeRF in natural language.
SpatialReasoner: Towards Explicit and Generalizable 3D Spatial Reasoning
Wufei Ma (Johns Hopkins University), Alan Yuille (Johns Hopkins University)
Large Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelPoint CloudBenchmarkChain-of-Thought
🎯 What it does: This paper presents SpatialReasoner, a visual-language large model that utilizes explicit 3D representations and achieves higher accuracy and improved generalization in 3D spatial reasoning tasks through a two-stage post-training approach (SFT + RL).
Spatiotemporal Consensus with Scene Prior for Unsupervised Domain Adaptive Person Search
Yimin Jiang (Dalian Maritime University), Jinjia Peng (Hebei University)
Object DetectionRetrievalDomain AdaptationConvolutional Neural NetworkContrastive LearningVideo
🎯 What it does: This paper proposes an unsupervised domain adaptation framework for person retrieval called STCSP, which can eliminate noise in pseudo-labels and gradually bridge the gap between the source domain and the target domain.
SPAZER: Spatial-Semantic Progressive Reasoning Agent for Zero-shot 3D Visual Grounding
Zhao Jin (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
Object DetectionTransformerVision Language ModelPoint CloudRetrieval-Augmented Generation
🎯 What it does: This paper presents SPAzer, a zero-shot 3D visual localization agent that achieves the localization of target objects in 3D scenes through spatial-semantic progressive reasoning on a global 3D view rendered from multiple perspectives.
SPC: Evolving Self-Play Critic via Adversarial Games for LLM Reasoning
Jiaqi Chen (University of Hong Kong), Kwan-Yee K. Wong (University of Hong Kong)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Designed and implemented a self-play Critic (SPC) for detecting erroneous steps in the reasoning process of LLMs without human step-level annotations;
SpecEdge: Scalable Edge-Assisted Serving Framework for Interactive LLMs
Jinwoo Park, Dongsu Han
OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes SpecEdge, a distributed framework that splits LLM inference into candidate token drafting on edge GPUs and validation on server GPUs; it achieves low-cost, high-throughput interactive LLM services through proactive edge drafting and pipeline-aware scheduling.
SpecEM: Training-Free LLM Ensembling via Iterative Drafting, Verification, and Online Feedback
Bo Lv (Peng Cheng Laboratory), Ping Luo (Peng Cheng Laboratory)
GenerationTransformerLarge Language ModelText
🎯 What it does: SpecEM is proposed, a plug-and-play LLM integration framework that achieves paragraph-level semantic collaboration between models through an iterative draft-validate-online feedback mechanism.
SpecMAS: A Multi-Agent System for Self-Verifying System Generation via Formal Model Checking
Rishabh Agrawal (University of Western Ontario), Apurva Narayan (University of Western Ontario)
Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: SpecMAS is constructed, a multi-agent framework that automatically generates from natural language SOP and performs formal verification through NuSMV, including a self-debugging loop.
SpecMER: Fast Protein Generation with K-mer Guided Speculative Decoding
Thomas Walton, Amirali Aghazadeh (Georgia Institute of Technology)
GenerationProtein Structure PredictionSequentialBiomedical Data
🎯 What it does: A framework named SpecMER has been developed, which combines k-mer guided Speculative Decoding to rapidly generate high-quality protein sequences.
SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning
Rui Pan (Princeton University), Ravi Netravali (Princeton University)
Computational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: SpecReason is proposed to infer intermediate reasoning steps through a lightweight model and validate them with a base model, significantly accelerating LRM inference.
Spectral Analysis of Diffusion Models with Application to Schedule Design
Roi Benita (Technion), Joseph Keshet (Technion)
Data SynthesisOptimizationDiffusion modelImage
🎯 What it does: This paper derives a closed-form transfer function of the diffusion model's reverse process in the frequency domain by assuming the target distribution is multivariate Gaussian, and designs a noise scheduling based on data spectral features using this analysis.
Spectral Analysis of Representational Similarity with Limited Neurons
Hyunmo Kang (Johns Hopkins University), SueYeon Chung (Harvard University)
OptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This study investigates the impact of limited neuron sampling on similarity measures (CCA, CKA) and constructs a theoretical framework for forward prediction and backward denoising based on random matrix theory.
Spectral Compressive Imaging via Chromaticity-Intensity Decomposition
Xiaodong Wang (Zhejiang University), Xin Yuan (Westlake University)
RestorationCompressionTransformerImage
🎯 What it does: The paper proposes a reconstruction framework CIDNet for spectral compressed imaging using chromatic-intensity decomposition in a dual-camera CASSI system.
Spectral Conditioning of Attention Improves Transformer Performance
Hemanth Saratchandran (Australian Institute for Machine Learning), Simon Lucey (Australian Institute for Machine Learning)
ClassificationObject DetectionSegmentationTransformerImageText
🎯 What it does: Analyzes the condition number of the Jacobian of the Transformer attention layer and proposes improving its spectral properties by adding correction terms to the query, key, and value matrices, thereby enhancing the condition number of the attention layer and model performance.
Spectral Convolutional Conditional Neural Process
Peiman Mohseni (Texas A&M University), Nick Duffield (Texas A&M University)
Convolutional Neural NetworkTime SeriesSequential
🎯 What it does: This paper proposes Spectral ConvCNP, which utilizes frequency domain convolution to replace traditional local CNNs, enhancing the global dependency modeling capability of neural process models.
Spectral Estimation with Free Decompression
Siavash Ameli (University of California), Michael W. Mahoney (University of California)
Graph
🎯 What it does: This paper proposes a method called free decompression, which uses the eigenvalues of randomly sampled submatrices to infer the spectral density of large Hermitian matrices that cannot be formed.
Spectral Graph Coarsening Using Inner Product Preservation and the Grassmann Manifold
Ido Cohen (Technion-Israel Institute of Technology), Ronen Talmon (Technion-Israel Institute of Technology)
Graph Neural NetworkGraph
🎯 What it does: A graph coarsening method is proposed with the goal of preserving the inner product of graph signals, utilizing Grassmann manifold to minimize the inner product error (IPE), thereby retaining both structural and feature relationships.
Spectral Graph Neural Networks are Incomplete on Graphs with a Simple Spectrum
Snir Hordan (Technion Israel Institute of Technology), Nadav Dym (Technion Israel Institute of Technology)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes and analyzes the expressive power of Spectral Enhanced Graph Neural Networks (SGNN), proving that common models (such as EPNN) are still incomplete on graphs with only simple spectra (different eigenvalues);
Spectral Learning for Infinite-Horizon Average-Reward POMDPs
Alessio Russo (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
Reinforcement Learning
🎯 What it does: In the infinite average reward POMDP learning problem, this paper proposes the Mixed Spectral Estimation and Mixed Spectral UCRL algorithms, which can utilize samples from different adaptive belief-based policies to achieve more efficient model estimation and policy learning.
Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy
Phuc Tran (Yale University), Nisheeth K. Vishnoi (Yale University)
Safty and PrivacyTabular
🎯 What it does: A new high-probability spectral norm perturbation bound is proposed to assess the robustness of low-rank approximations under noise or differential privacy, providing practical guarantees for differential privacy PCA.
SpectraLDS: Provable Distillation for Linear Dynamical Systems
Devan Shah (Princeton University), Elad Hazan (Google DeepMind)
Computational EfficiencyKnowledge DistillationTime SeriesSequential
🎯 What it does: The spectral filter in the Spectral Transform Unit (STU) is transformed into an explicit Linear Dynamical System (LDS) through a provable sparse approximation method, enabling low-order recursive inference for long sequences, maintaining training stability and significantly reducing inference costs.
Speculate Deep and Accurate: Lossless and Training-Free Acceleration for Offloaded LLMs via Substitute Speculative Decoding
Pei-Shuo Wang (National Yang Ming Chiao Tung University), Kai-Chiang Wu (National Yang Ming Chiao Tung University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a training-free, lossless acceleration scheme called SubSpec, which utilizes low-bit quantization alternative layers, shared GPU layers, and KV-Cache to construct a draft model that is highly aligned with the target model, thereby achieving efficient Speculative Decoding in parameter offloading scenarios.
Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation
Yao Teng (University of Hong Kong), Xihui Liu
GenerationData SynthesisTransformerImageTextStochastic Differential Equation
🎯 What it does: Proposes the Speculative Jacobi-Denoising Decoding (SJD2) framework, which combines a continuous denoising process with Jacobi iteration to significantly accelerate autoregressive text-image generation.
Spend Wisely: Maximizing Post-Training Gains in Iterative Synthetic Data Bootstrapping
Pu Yang (Peking University), Zhuoyuan Li (National University of Singapore)
Data SynthesisOptimizationLarge Language ModelDiffusion modelImageTabular
🎯 What it does: The study investigates how to allocate generation and training costs in iterative synthetic data self-training under a limited budget to maximize the final model performance.
SpEx: A Spectral Approach to Explainable Clustering
Tal Argov (Tel Aviv University), Tal Wagner (Tel Aviv University)
Explainability and InterpretabilityTabular
🎯 What it does: A general interpretable clustering method SPEX based on spectral graph partitioning is proposed, which can generate axis-aligned decision tree explanations for any reference clustering (regardless of whether it has centroids) or directly construct adjacency graphs from the data.
SPFL: Sequential updates with Parallel aggregation for Enhanced Federated Learning under Category and Domain Shifts
Haoyuan Liang (Sun Yat-sen University), Juepeng Zheng (Sun Yat-sen University)
Domain AdaptationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: The SPFL framework is proposed, which combines sequential updates with parallel aggregation to address the issues of update order sensitivity and catastrophic forgetting caused by category and domain shifts in federated learning.
SPICED: A Synaptic Homeostasis-Inspired Framework for Unsupervised Continual EEG Decoding
Yangxuan Zhou (Zhejiang University), Gang Pan (Zhejiang University)
ClassificationAnomaly DetectionSpiking Neural NetworkTime SeriesBiomedical Data
🎯 What it does: The SPICED framework is proposed for unsupervised continuous EEG decoding, enabling the model to dynamically expand and retain prior knowledge as new individuals continuously emerge.
SpiderSolver: A Geometry-Aware Transformer for Solving PDEs on Complex Geometries
Kai Qi (Xi'an Jiaotong University), Jian Sun (Xi'an Jiaotong University)
TransformerMeshPhysics Related
🎯 What it does: We propose SpiderSolver, a Transformer that utilizes a spider web-style tokenization and coarse-fine attention to quickly solve partial differential equations under complex geometric boundaries.
SPiDR: A Simple Approach for Zero-Shot Safety in Sim-to-Real Transfer
Yarden As (ETH Zurich), Andreas Krause (ETH Zurich)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A zero-shot safe simulation-to-real transfer method SPiDR based on pessimistic domain randomization is proposed, which uses model uncertainty to penalize costs to ensure the satisfaction of constraints in the real environment.
Spik-NeRF: Spiking Neural Networks for Neural Radiance Fields
Gang Wan (Space Engineering University), Yufei Guo (Peking University)
GenerationData SynthesisOptimizationKnowledge DistillationSpiking Neural NetworkNeural Radiance FieldImage
🎯 What it does: A Spiking Neural Network (SNN) version of the NeRF model, called Spik-NeRF, was constructed and trained. By introducing ternary spiking neurons, it enhances information carrying capacity and achieves rendering quality comparable to traditional ANN NeRF in just 2 time steps.
Spike-RetinexFormer: Rethinking Low-light Image Enhancement with Spiking Neural Networks
Hongzhi Wang (Zhejiang University), Weidong Geng (Zhejiang University)
RestorationSpiking Neural NetworkTransformerImage
🎯 What it does: Proposes a pulse neural network based on Retinex and Transformer for low-light image enhancement.
Spike-timing-dependent Hebbian learning as noisy gradient descent
Niklas Dexheimer (University of Twente), Johannes Schmidt-Hieber (University of Twente)
Spiking Neural NetworkSequentialOrdinary Differential Equation
🎯 What it does: Analyzed and proved that the Hebbian STDP learning rule based on spike timing is equivalent to performing noisy gradient descent on a non-convex loss function over a probability simplex, and provided a proof of exponential convergence in the presence of noise; subsequently, further connected this dynamics with mirror descent (entropy mirror descent) and proposed a generalized algorithm for multi-output neurons.
Spike4DGS: Towards High-Speed Dynamic Scene Rendering with 4D Gaussian Splatting via a Spike Camera Array
Qinghong Ye (Peking University), Peixi Peng (Peking University)
GenerationData SynthesisPose EstimationAutonomous DrivingTransformerGaussian SplattingSimultaneous Localization and MappingImageVideoPoint Cloud
🎯 What it does: The Spike4DGS framework is proposed, utilizing a multi-view peak camera array and 4D Gaussian Splatting to achieve new perspective synthesis for high-speed dynamic scenes.
Spiking Meets Attention: Efficient Remote Sensing Image Super-Resolution with Attention Spiking Neural Networks
Yi Xiao (Zhengzhou University), Liangpei Zhang (Wuhan University)
RestorationSuper ResolutionConvolutional Neural NetworkSpiking Neural NetworkImage
🎯 What it does: This paper proposes a spike neural network based on attention, SpikeSR, for efficient super-resolution reconstruction of remote sensing images (RSI).
Spiking Neural Networks Need High-Frequency Information
Yuetong Fang (Hong Kong University of Science and Technology), Renjing Xu (Hong Kong University of Science and Technology)
ClassificationRecognitionConvolutional Neural NetworkSpiking Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: This paper proposes and verifies that Spiking Neural Networks (SNN) are essentially low-pass filters in the frequency domain, leading to the attenuation of high-frequency information; based on this, Max-Former (a Transformer structure) and Max-ResNet (a CNN structure) are designed to recover high-frequency information and enhance performance.
SpikingVTG: A Spiking Detection Transformer for Video Temporal Grounding
Malyaban Bal (Pennsylvania State University), Adam D. Cobb (SRI International)
Spiking Neural NetworkTransformerVideo
🎯 What it does: A SpikingVTG model based on spiking neural networks is proposed for video temporal localization and highlight detection.
SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding
Trung Le (University of Washington), Eli Shlizerman (University of Washington)
TransformerBiomedical DataBenchmark
🎯 What it does: A Permutation-Invariant Transformer SPINT is proposed, capable of handling unordered, variable-sized collections of neurons, achieving motion decoding across multiple sessions with a small amount of unlabeled calibration data.
Spiral: Semantic-Aware Progressive LiDAR Scene Generation and Understanding
Dekai Zhu (Munich Center for Machine Learning), Slobodan Ilic (Technical University of Munich)
SegmentationGenerationData SynthesisDepth EstimationAutonomous DrivingDiffusion modelPoint Cloud
🎯 What it does: A semantic-aware range perspective LiDAR diffusion model named SPIRAL is proposed, capable of generating depth, reflectance, and semantic labels in one go.
SplashNet: Split‑and‑Share Encoders for Accurate and Efficient Typing with Surface Electromyography
Nima Ryan Hadidi (University of California Los Angeles), Jonathan Kao (University of California Los Angeles)
RecognitionComputational EfficiencyConvolutional Neural NetworkBiomedical DataBenchmark
🎯 What it does: Proposes the SplashNet network, which significantly improves zero-shot and fine-tuning performance for surface electromyography text input through three enhancements: rolling time normalization, aggressive channel masking, and Split-and-Share encoder.
Split conformal classification with unsupervised calibration
Santiago Mazuelas (Basque Center for Applied Mathematics)
ClassificationOptimizationImageTabular
🎯 What it does: A method for split conformal classification using unsupervised calibration samples is proposed, which solves for the weights that minimize the IPM on the training samples, making the calibration samples statistically indistinguishable from the training samples, thereby obtaining label weights and constructing a conformal prediction set.
Split Gibbs Discrete Diffusion Posterior Sampling
Wenda Chu (California Institute of Technology), Yisong Yue (California Institute of Technology)
GenerationData SynthesisOptimizationDiffusion modelImageAudio
🎯 What it does: A discrete posterior sampling algorithm based on Split Gibbs, called SGDD, is proposed, which can achieve reward-guided generation and inverse problem solving in discrete state spaces.
SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing
Sung-Hoon Yoon (Harvard University), Mengyu Wang (Harvard University)
Image TranslationGenerationLarge Language ModelFlow-based ModelRectified FlowImageTextBenchmark
🎯 What it does: Proposes SplitFlow, which achieves high-quality text-to-image editing by semantically decomposing the target text prompt, calculating independent editing flows, and aggregating them.
SPMDM: Enhancing Masked Diffusion Models through Simplifing Sampling Path
Yichen Zhu (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationData SynthesisTransformerDiffusion modelTextSequential
🎯 What it does: The Simple Path Masked Diffusion Model (SPMDM) is proposed, which divides sequences into fixed-length subsequences and uses different noise scales for different subsequences, encouraging the generation of simpler and more efficient sampling paths during training, thereby significantly enhancing the planning and reasoning capabilities of discrete sequence generation.
Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation
Siwei Wen (Shanghai Artificial Intelligence Laboratory), Weijia Li
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes FakeVLM, a large multimodal model specifically designed for synthetic image authenticity discrimination and anomaly trace explanation, and constructs the FakeClue dataset, which contains 100k multi-category images and fine-grained natural language anomaly descriptions.
SPOT-Trip: Dual-Preference Driven Out-of-Town Trip Recommendation
Yinghui Liu (Zhejiang University of Technology), Ivan Lee (University of South Australia)
Recommendation SystemGraph Neural NetworkTransformerTabularOrdinary Differential Equation
🎯 What it does: The SPOT-Trip framework is proposed to generate a complete itinerary (including an intermediate POI sequence) when users travel from their hometown to unknown areas, explicitly learning users' static preferences (long-term interests) and dynamic preferences (short-term interests).
SPOT: Scalable Policy Optimization with Trees for Markov Decision Processes
Xuyuan Xiong (Shanghai Jiao Tong University), Cheng Hua (Shanghai Jiao Tong University)
OptimizationReinforcement LearningTabularBenchmark
🎯 What it does: Proposes the SPOT framework, which utilizes a MILP-based decision tree policy optimization combined with policy iteration and reduced space branch and bound to solve interpretable decision tree policies in Markov decision processes.
Spotlight Attention: Towards Efficient LLM Generation via Non-linear Hashing-based KV Cache Retrieval
Wenhao Li (Xiamen University), Rongrong Ji (Xiamen University)
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes Spotlight Attention, which dynamically selects important key-value pairs from the KV cache using a nonlinear hash function, significantly accelerating the inference of LLMs.
SPRINT: Enabling Interleaved Planning and Parallelized Execution in Reasoning Models
Emil Biju (Microsoft), Amin Saberi (Stanford University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The large inference model introduces a post-training and inference framework called SPRINT, which enables the model to dynamically identify and execute parallel subtasks during the inference process, significantly reducing the number of sequential tokens.
SPRO: Improving Image Generation via Self-Play
Ritika Jha, Balaji Krishnamurthy
GenerationData SynthesisOptimizationTransformerReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: By jointly optimizing a self-play visual-language model and a diffusion model, images that align with aesthetics, engagement, and user preferences are generated.
Spurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection
Reihaneh Zohrabi (TU Darmstadt), Mohammad Hossein Rohban (Sharif University of Technology)
Anomaly DetectionImage
🎯 What it does: A post-hoc prototype refinement method called SPROD is proposed to enhance the reliability of discrete distribution detection under the condition of unknown spurious correlations.
SQL-R1: Training Natural Language to SQL Reasoning Model By Reinforcement Learning
Peixian MA, Jian Guo (International Digital Economy Academy)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes SQL-R1, a NL2SQL reasoning model trained through reinforcement learning, capable of generating interpretable SQL statements in complex multi-table queries.
SQLens: An End-to-End Framework for Error Detection and Correction in Text-to-SQL
Yue Gong (Amazon Web Services), Tim Kraska (Amazon Web Services)
GenerationAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: An end-to-end SQLENS framework is proposed for detecting and correcting semantic errors in Text-to-SQL queries generated by LLMs.
SQS: Enhancing Sparse Perception Models via Query-based Splatting in Autonomous Driving
Haiming Zhang (Chinese University of Hong Kong Shenzhen), Zhen Li (Chinese University of Hong Kong Shenzhen)
Object DetectionAutonomous DrivingTransformerGaussian SplattingImagePoint Cloud
🎯 What it does: This paper proposes SQS (Query-based Splatting Pre-training), which enhances the performance of sparse perception models in autonomous driving by using sparse Gaussian queries for self-supervised reconstruction of multi-view RGB and depth during the pre-training phase.
Squared families are useful conjugate priors
Russell Tsuchida (Monash University), Dino Sejdinovic (University of Adelaide)
Tabular
🎯 What it does: This paper is the first to use squared families as prior distributions, constructing the Generalised Squared Families (GSF) and proving its conjugacy for various likelihoods. Utilizing this conjugacy, closed-form posteriors and marginal likelihoods are obtained. Subsequently, the GSF is combined with neural network features to propose the GSF Process (GSFP) model for Bayesian regression in feature space, which can naturally express multimodal uncertainty and achieve excellent performance in few-shot learning tasks.
SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation
Ziqiang Cui (City University of Hong Kong), Chen Ma (City University of Hong Kong)
Recommendation SystemTransformerLarge Language ModelContrastive LearningSequential
🎯 What it does: A framework called SRA-CL based on semantic retrieval enhanced contrastive learning has been developed to improve sequential recommendation models.
SRHand: Super-Resolving Hand Images and 3D Shapes via View/Pose-aware Neural Image Representations and Explicit Meshes
Minje Kim (KAIST), Tae-Kyun Kim (KAIST)
GenerationSuper ResolutionGenerative Adversarial NetworkImageMesh
🎯 What it does: Reconstructing high-resolution hand models from low-resolution multi-view hand images through a geometric perception implicit image function and explicit 3D mesh collaboration.
SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning
Zhongwei Wan (Ohio State University), Shen Yan (ByteDance)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodality
🎯 What it does: This paper proposes SRPO, which combines self-reflective data generation with a reinforcement learning framework that enhances the reasoning and self-correction capabilities of multimodal LLMs.
SRSR: Enhancing Semantic Accuracy in Real-World Image Super-Resolution with Spatially Re-Focused Text-Conditioning
Chen Chen (Amazon), Pulak Purkait (Amazon)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: A pluggable spatial re-focusing image super-resolution framework SRSR is proposed for the inference stage, addressing semantic inaccuracies and artifacts under text conditions.
SSIMBaD: Sigma Scaling with SSIM-Guided Balanced Diffusion for AnimeFace Colorization
Junpyo Seo (Seoul National University), Byung-Ro Moon (Seoul National University)
Image TranslationRestorationGenerationDiffusion modelImage
🎯 What it does: This paper studies a method for colorizing anime facial line art based on a continuous time diffusion model, proposing the SSIMBaD framework, which achieves consistency between forward noise and reverse recovery through a perceptually unified Sigma space transformation, and incorporates lightweight trajectory refinement.
SSR: Enhancing Depth Perception in Vision-Language Models via Rationale-Guided Spatial Reasoning
Yang Liu (Westlake University), Donglin Wang (Westlake University)
Depth EstimationKnowledge DistillationRepresentation LearningTransformerVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the SSR framework, which enhances the spatial perception and reasoning capabilities of visual language models (VLM) by converting raw depth information into structured textual reasoning.
SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs
LiuRuyue, Weiping Wang (Institute of Information Engineering, Chinese Academy of Sciences)
Domain AdaptationKnowledge DistillationRepresentation LearningGraph Neural NetworkLarge Language ModelContrastive LearningTextGraph
🎯 What it does: This paper proposes a structure-aware self-supervised learning framework called SSTAG, specifically designed for Text Attribute Graphs (TAGs), which unifies the representation of multi-domain graph data to achieve knowledge transfer and generalization in cross-domain tasks.
ST$^2$360D: Spatial-to-Temporal Consistency for Training-free 360 Monocular Depth Estimation
Zidong Cao (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
Depth EstimationImageVideo
🎯 What it does: Convert 360-degree images into a sequence of perspective frames, utilizing the temporal consistency of the video depth estimation model (VDE) to achieve training-independent 360-degree monocular depth estimation;
Stab-SGD: Noise-Adaptivity in Smooth Optimization with Stability Ratios
David A. R. Robin (INRIA - ENS Paris PSL Research University), Kevin Scaman (INRIA - ENS Paris)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Introduced a 'stability ratio' as a local relative noise measure, and based on it, proposed a Stab-SGD algorithm that is unscheduled and can adapt to noise levels.
Stability and Oracle Inequalities for Optimal Transport Maps between General Distributions
Shubo Li (Pennsylvania State University), Runze Li (Pennsylvania State University)
Optimization
🎯 What it does: This paper proposes a unified theoretical framework to construct a semi-dual OT estimation method, and reduces the requirements for higher-order smoothness and strong convexity through sieve selection.
Stability and Sharper Risk Bounds with Convergence Rate $\tilde{O}(1/n^2)$
Bowei Zhu (Renmin University of China), Yong Liu (Renmin University of China)
Optimization
🎯 What it does: Within the framework of theoretical learning, high-probability upper bounds for overfitting risks of ERM, PGD, and SGD are provided using algorithmic stability at the gradient level;
Stabilizing LTI Systems under Partial Observability: Sample Complexity and Fundamental Limits
Ziyi Zhang (Carnegie Mellon University), Guannan Qu (Carnegie Mellon University)
🎯 What it does: A LTS-P algorithm is designed to achieve rapid stabilization after learning in LTI systems that have no initial stabilizable controller, are partially observable, and may be unstable.
Stable Cinemetrics : Structured Taxonomy and Evaluation for Professional Video Generation
Agneet Chatterjee (Stability AI), Varun Jampani
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This work proposes the SCINE (Stable Cinemetrics) framework, which constructs a four-level academic control hierarchy for film based on the needs of professional film production (Setup, Events, Lighting, Camera), with a total of 76 leaf nodes. Based on this hierarchy, prompts that align with industry workflows were generated, followed by the design of a large-scale human evaluation and automatic evaluation (VLM) process to conduct a fine-grained analysis of the performance of 13 text-to-video models on professional-level film generation tasks.
Stable Coresets via Posterior Sampling: Aligning Induced and Full Loss Landscapes
WEI-KAI CHANG, Rajiv Khanna (Purdue University)
OptimizationConvolutional Neural NetworkTransformerGaussian SplattingImageText
🎯 What it does: A stable subset selection method based on posterior sampling is proposed, which smooths the loss landscape by performing Gaussian posterior sampling on model weights, enhancing the consistency of the subset with the full data gradient and curvature.
Stable Gradients for Stable Learning at Scale in Deep Reinforcement Learning
Roger Creus Castanyer (Mila - Quebec AI Institute), Pablo Samuel Castro (Google DeepMind)
Reinforcement LearningBenchmark
🎯 What it does: This paper diagnoses the decline in scalability caused by non-stationarity and gradient pathology in deep reinforcement learning (RL) through systematic experimental analysis, and proposes multi-hop residual connections and the Kron second-order optimizer as gradient stabilization interventions, validating their performance improvement across various algorithms and environments.
Stable Matching with Ties: Approximation Ratios and Learning
Shiyun Lin (Peking University), Vianney Perchet (CREST)
🎯 What it does: This paper studies one-sided stable matching with indifference preferences, proposing the OSS (Optimal Stable Share) ratio as a measure of fairness, and provides lower and upper bounds for different matching categories.
Stable Minima of ReLU Neural Networks Suffer from the Curse of Dimensionality: The Neural Shattering Phenomenon
Tongtong Liang, Rahul Parhi
Recurrent Neural Network
🎯 What it does: This study investigates the implicit bias of flatness/low curvature in over-parameterized two-layer ReLU neural networks with multivariate inputs and its impact on generalization, particularly the performance of stable minima in high-dimensional and non-interpolating cases.
Stable Part Diffusion 4D: Multi-View RGB and Kinematic Parts Video Generation
Hao Zhang (Stability AI), Varun Jampani (Stability AI)
SegmentationGenerationData SynthesisDiffusion modelContrastive LearningVideoMesh
🎯 What it does: Generate multi-view, temporally consistent RGB videos and corresponding kinematic part segmentation videos from monocular videos or images; then map 2D parts to 3D to obtain animatable meshes.
Stable Port-Hamiltonian Neural Networks
Fabian J. Roth (Technical University of Darmstadt), Oliver Weeger (Technical University of Darmstadt)
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A stable end-to-end physics-guided neural network (sPHNN) is proposed, which can learn globally asymptotically stable nonlinear dynamics from sparse data.
StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models
Haoxin Yang (South China University of Technology), Shengfeng He (Singapore Management University)
GenerationData SynthesisAnomaly DetectionSafty and PrivacyTransformerMixture of ExpertsDiffusion modelAuto EncoderImage
🎯 What it does: A global binary watermark is embedded in the generation process of the latent diffusion model (LDM), and a unified copyright protection and tampering localization framework (StableGuard) is constructed through self-supervised joint training.
STACI: Spatio-Temporal Aleatoric Conformal Inference
Brandon R. Feng (North Carolina State University), Brian Reich
Time Series
🎯 What it does: The STACI method is proposed, which uses variational Bayesian neural networks to approximate non-stationary spatiotemporal Gaussian processes and infers statistically valid confidence intervals through local spatiotemporal consistency, thereby achieving large-scale spatiotemporal interpolation and reliable uncertainty quantification.
Stackelberg Learning with Outcome-based Payment
Tom Yan (Carnegie Mellon University), Chicheng Zhang (University of Arizona)
OptimizationReinforcement Learning
🎯 What it does: This paper studies how to incentivize cooperation among agents with different interests in a decentralized multi-agent environment through outcome-based payment schemes. Specifically, we model this problem as a Stackelberg Markov game, where the leader can commit to a strategy and specify a set of outcome-based payments.
Stackelberg Self-Annotation: A Robust Approach to Data-Efficient LLM Alignment
Xu Chu (Peking University), Yujie Jin (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A robust alignment framework SGPO based on Stackelberg game is proposed, and a self-labeling, data-efficient SSAPO algorithm is implemented, achieving LLM alignment with very few human labels (about 2K).
Staggered Environment Resets Improve Massively Parallel On-Policy Reinforcement Learning
Sid Bharthulwar (Harvard University), Hao Su (University of California San Diego)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A method for staggered resets of on-policy RL in large-scale GPU parallel environments is proposed, addressing the periodic non-stationarity caused by synchronous resets, significantly improving sample efficiency and convergence speed.
STAIR: Addressing Stage Misalignment through Temporal-Aligned Preference Reinforcement Learning
Yao Luan (Tsinghua University), Qing-Shan Jia (Chinese Academy of Sciences)
Robotic IntelligenceReinforcement LearningContrastive LearningSequential
🎯 What it does: The STAIR method is proposed, which approximates phase differences using temporal distance and selects only phase-aligned segments in preference learning, thereby completing multi-stage tasks more efficiently.
STAR-Bets: Sequential TArget-Recalculating Bets for Tighter Confidence Intervals
Vaclav Voracek, Francesco Orabona (King Abdullah University of Science and Technology)
Reinforcement LearningTabularFinance Related
🎯 What it does: A betting strategy based on sequential target-recalculating, STaR-Bets, is proposed to construct mean confidence intervals.
STAR: Efficient Preference-based Reinforcement Learning via Dual Regularization
Fengshuo Bai (Shanghai Jiao Tong University), Yaodong Yang (Chinese Academy of Sciences)
Reinforcement LearningSequential
🎯 What it does: The STAR method is proposed to improve the feedback efficiency of preference reinforcement learning through preference marginal regularization and conservative Q-value estimation policy regularization, addressing the issues of reward model overfitting and Q-value overestimation.
STAR: Spatial-Temporal Tracklet Matching for Multi-Object Tracking
Xuewei Bai (Renmin University of China), LI Chunxu (China Waterborne Transport Research Institute)
Object TrackingGraph Neural NetworkVideo
🎯 What it does: Proposes the STAR framework, which achieves multi-object tracking by constructing a Tracklet Clip graph and using graph matching;
STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis
Jiatao Gu (Apple), Shuangfei Zhai (Apple)
GenerationData SynthesisTransformerFlow-based ModelAuto EncoderImageText
🎯 What it does: A scalable latent space normal flow model STARFlow based on Transformer autoregressive flow has been developed for high-resolution image generation.
STaRFormer: Semi-Supervised Task-Informed Representation Learning via Dynamic Attention-Based Regional Masking for Sequential Data
Maximilian Forstenhäusler (BMW Group), Natascha Weber (BMW Group)
Anomaly DetectionRepresentation LearningTransformerContrastive LearningTime SeriesSequential
🎯 What it does: This paper proposes STaRFormer, a Transformer framework that combines dynamic area masking and semi-supervised contrastive learning to handle non-stationary, spatiotemporal, and irregularly sampled sequence data.
StarTrail: Concentric Ring Sequence Parallelism for Efficient Near-Infinite-Context Transformer Model Training
Ziming Liu (National University of Singapore), Yang You (National University of Singapore)
TransformerLarge Language ModelDiffusion modelTextSequential
🎯 What it does: Designed and implemented the StarTrail system, which improves the communication efficiency and scalability of Transformer long sequence training using a multi-dimensional concentric ring sequence parallel method.
State Entropy Regularization for Robust Reinforcement Learning
Yonatan Ashlag (Technion), Shie Mannor (NVIDIA Research)
Reinforcement Learning
🎯 What it does: This paper studies the impact of state entropy regularization on the robustness of policies in reinforcement learning, providing a complete theoretical analysis and experimental validation in discrete and continuous control environments.
State Size Independent Statistical Error Bound for Discrete Diffusion Models
Shintaro Wakasugi (University of Tokyo), Taiji Suzuki (University of Tokyo)
Diffusion modelScore-based Model
🎯 What it does: This paper presents a theoretical error bound for score function estimation based on neural networks in discrete diffusion models, proposes an error upper bound that is independent of the state space dimension, and provides specific examples for hypercubes and graphical diffusion processes.
State Space Prompting via Gathering and Spreading Spatio-Temporal Information for Video Understanding
Jiahuan Zhou (Wangxuan Institute of Computer Technology Peking University), Gang Hua (Amazon.com)
RecognitionPrompt EngineeringVideo
🎯 What it does: A Prompt learning framework called State Space Prompting (SSP) is proposed for the pre-trained state space model (VideoMamba), which achieves efficient video understanding by aggregating local spatial information within each frame and propagating global temporal information across frames.
State-Covering Trajectory Stitching for Diffusion Planners
Kyowoon Lee (KAIST), Jaesik Choi (KAIST)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelSequential
🎯 What it does: The SCoTS method is proposed, which generates diverse and long-term trajectories through reward-free trajectory stitching and expansion on offline datasets to enhance the performance of diffusion planners.
StateSpaceDiffuser: Bringing Long Context to Diffusion World Models
Nedko Savov (Insait Institute), Luc Van Gool
GenerationData SynthesisDiffusion modelWorld ModelImageVideo
🎯 What it does: This paper proposes a long-sequence visual world model called StateSpaceDiffuser, which combines state space models with diffusion models, achieving a unification of long-sequence context memory and high-quality generation.
Statistical Analysis of an Adversarial Bayesian Weak Supervision Method
Steven An (University of California)
OptimizationAdversarial AttackTabularBenchmark
🎯 What it does: Designed and implemented the Bayesian Balsubramani-Freund (BBF) label model, which utilizes a Bayesian framework to model the accuracy and label distribution from weakly supervised sources, and generates labels by maximizing the posterior mode.
Statistical Analysis of the Sinkhorn Iterations for Two-Sample Schr\"{o}dinger Bridge Estimation
Ibuki Maeda (Kyushu Institute of Technology), Atsushi Nitanda (Agency for Science Technology and Research)
Stochastic Differential Equation
🎯 What it does: This study analyzes the statistical performance of the Schrödinger bridge problem, particularly in the two-sample estimation setting, and investigates the statistical behavior of the Sinkhorn iteration.
Statistical Guarantees for High-Dimensional Stochastic Gradient Descent
Jiaqi Li (University of Chicago), Wei Biao Wu (University of Chicago)
OptimizationTime Series
🎯 What it does: This paper studies the statistical convergence properties of stochastic gradient descent (SGD) with a fixed learning rate and its averaged variant (ASGD) in high-dimensional parameter spaces.
Statistical Inference for Gradient Boosting Regression
Haimo Fang (Fudan University), Giles Hooker (University of Pennsylvania)
Tabular
🎯 What it does: A gradient boosting regression statistical inference framework is proposed, which incorporates dropout and parallel training based on Boulevard regularization, and provides the central limit theorem, confidence intervals, prediction intervals, and variable importance testing methods.