ICLR 2026 Papers — Page 44
International Conference on Learning Representations · 5356 papers
SmartDJ: Declarative Audio Editing with Audio Language Model
Zitong Lan (University of Pennsylvania), Mingmin Zhao (University of Pennsylvania)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelAudio
🎯 What it does: Proposed the SmartDJ framework, achieving automated editing from declarative audio editing instructions to stereo audio;
Smarter Not Harder: Generative Process Evaluation with Intrinsic-Signal Driving and Ability‑Adaptive Reward Shaping
Tao He (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
Computational EfficiencyLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a new generative process evaluation framework (GenPRM), combining intrinsic signal-driven evaluation and thought hierarchy, difficulty-adaptive reward mechanisms, and GRPO to form TP-GRPO, aiming to enhance the reinforcement learning training efficiency of large reasoning models (LRM).
SmellNet: A Dataset for Sensor-Based Smell Recognition and Mixture Prediction
Dewei Feng (MIT), Paul Pu Liang (MIT)
ClassificationRecognitionTransformerContrastive LearningTime Series
🎯 What it does: Proposed the SMELLNET dataset and the SCENTFORMER model, aiming to achieve real-time machine olfaction based on low-cost metal oxide sensors;
SMixer: Rethinking Efficient-Training and Event-Driven SNNs
Yijie Lu (Peking University), Guoqi Li (Shandong University)
Computational EfficiencyRepresentation LearningSpiking Neural NetworkImageTime Series
🎯 What it does: Propose an efficient event-driven SNN architecture based on Spiking-Token Mixer (SMixer) and design the Dynamic Spatial-Temporal Spiking Pruning (DSTSP) framework, significantly reducing training costs while maintaining performance;
Smooth Calibration Error: Uniform Convergence and Functional Gradient Analysis
Futoshi Futami (University of Osaka), Atsushi Nitanda (Agency for Science, Technology and Research)
ClassificationExplainability and InterpretabilityImage
🎯 What it does: This paper provides a unified theoretical framework, giving the upper bound of consistency convergence for smooth calibration error (smooth CE), and proves that the smooth CE during training can be controlled by the functional gradient of the loss function; subsequently, it analyzes commonly used learning algorithms such as gradient boosting trees, kernel boosting, and two-layer neural networks, proving that under appropriate hyperparameter settings, they can achieve high accuracy while satisfying any predefined smooth CE error.
Smooth Reading: Bridging the Gap of Recurrent LLM to Self-Attention LLM on Long-Context Understanding
Kai Liu (Tongji University), Kai Chen (Shanghai AI Laboratory)
Recurrent Neural NetworkTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the Smooth Reading method, which performs end-to-end multi-round reasoning on iterative large language models (LLMs) by combining chunk reading with context summary compression, significantly improving the performance of iterative LLMs on long-text understanding tasks.
SMOTE and Mirrors: Exposing Privacy Leakage from Synthetic Minority Oversampling
Georgi Ganev (SAS), Emiliano De Cristofaro (UC Riverside)
Data SynthesisSafty and PrivacyTabular
🎯 What it does: Systematically study the privacy leakage risks of SMOTE and first propose geometry-based distinguishability attack (DistinSMOTE) and reconstruction attack (ReconSMOTE), revealing SMOTE's privacy vulnerabilities in minority class samples.
SNAP-UQ: Self-supervised Next-Activation Prediction for Single-Pass Uncertainty in TinyML
Ismail Lamaakal (Mohammed First University), Yassine Maleh (Sultan Moulay Slimane University)
Anomaly DetectionComputational EfficiencyImageAudio
🎯 What it does: A single-channel uncertainty estimation method based on self-supervised next activation prediction is studied, which can provide reliable error detection and model confidence with only one forward inference and no labels on TinyML devices.
SNAPHARD CONTRAST LEARNING
Changpu Meng (University of Wollongong), Yi Guo (Western Sydney University)
ClassificationRecommendation SystemAnomaly DetectionGenerative Adversarial NetworkContrastive LearningImageTextGraph
🎯 What it does: Propose a contrastive learning framework called SPACL based on hard positive and negative sample screening. Theoretically analyze the optimality and convergence issues of contrastive learning, and enhance discriminative performance by expanding the convex hull of positive samples through farthest point iteration and constructing hard negative samples via an adversarial generator.
Sobolev Gradient Ascent for Optimal Transport: Barycenter Optimization and Convergence Analysis
Kaheon Kim (University of Notre Dame), Xiaohui Chen (University of Southern California)
OptimizationImageVideo
🎯 What it does: Proposed an unconstrained Sobolev gradient ascent algorithm for high-precision computation of Wasserstein barycenters.
Social Agents: Collective Intelligence Improves LLM Predictions
Aanisha Bhattacharyya (Adobe Media and Data Science Research), Balaji Krishnamurthy (Adobe Media and Data Science Research)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringImageVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Built the Social Agents framework, which uses multiple persona LLM agents to simulate human diverse perspectives and enhances performance on behavior prediction and text generation tasks through aggregation.
SocialHarmBench: Revealing LLM Vulnerabilities to Socially Harmful Requests
Punya Syon Pandey (University of Toronto), Zhijing Jin (University of Toronto)
Safty and PrivacyAdversarial AttackTextBenchmark
🎯 What it does: Proposed the SOCIALHARMBENCH benchmark to evaluate the societal harms of LLMs in domains such as politics, historical revisionism, propaganda, and others.
SocialJax: An Evaluation Suite for Multi-agent Reinforcement Learning in Sequential Social Dilemmas
Zihao Guo (King's College London), Yali Du (King's College London)
Computational EfficiencyReinforcement LearningSequentialBenchmark
🎯 What it does: This paper constructs SocialJax, a JAX-based sequential social dilemma environment and algorithm suite, providing a complete training pipeline, benchmark evaluation metrics, and multiple baseline algorithm implementations.
SoFlow: Solution Flow Models for One-Step Generative Modeling
Tianze Luo (Princeton University), Zhuang Liu (Princeton University)
GenerationTransformerFlow-based ModelAuto EncoderImageOrdinary Differential Equation
🎯 What it does: Proposed a one-shot generation framework called Solution Flow Models (SoFlow), achieving non-iterative sampling by learning the solution function of velocity ODEs;
Soft Equivariance Regularization for Invariant Self-Supervised Learning
Joohyung Lee (AITRICS), Juho Lee (Korea Advanced Institute of Science and Technology)
Representation LearningTransformerContrastive LearningImage
🎯 What it does: Propose a Soft Equivariant Regularization (SER) that separates equivariance and invariance across different network layers to enhance the performance of self-supervised learning in Vision Transformers (ViT).
Soft Quality-Diversity Optimization
Saeed Hedayatian (University of Southern California), Stefanos Nikolaidis (University of Southern California)
OptimizationImageBenchmark
🎯 What it does: Propose the Soft Quality Diversity (SOFT QD) framework and differentiable algorithm SQUAD to address discretization issues and high-dimensional challenges in traditional Quality Diversity (QD) methods.
Soft Tokens, Hard Truths
Natasha Butt (University of Amsterdam), Yann Ollivier
Large Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a continuous Chain-of-Thought (CoT) training method based on reinforcement learning, which achieves continuous reasoning using noisy soft/fuzzy word embeddings without requiring discrete CoT annotations;
Soft-Di[M]O: Improving One-Step Discrete Image Generation with Soft Embeddings
Yuanzhi Zhu (Ecole Polytechnique), Vicky Kalogeiton (Ecole Polytechnique)
GenerationKnowledge DistillationReinforcement LearningDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Propose Soft-Embedding and combine it with the Di[M]O distillation framework to achieve full differentiability in single-step discrete image generation, enabling direct GAN refinement, differentiable reward fine-tuning, and test-time embedding optimization (TTEO).
Soft-Masked Diffusion Language Models
Michael Hersche (IBM Research), Abbas Rahimi (ETH Zurich)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelText
🎯 What it does: Propose the Soft-Masking method, which during the decoding process of Mask Diffusion Language Models (MDLM), combines the [MASK] token with the top-k word vectors predicted in the previous step through a convex combination, thereby preserving predictive information and providing continuous feedback.
SoftCFG: Uncertainty-guided Stable Guidance for Visual Autoregressive Model
Dongli Xu (Ku Leuven), Matthew B. Blaschko (Ku Leuven)
GenerationImageTextBenchmark
🎯 What it does: Propose SoftCFG, a stable guidance method in visual autoregressive models based on uncertainty-driven guidance, dynamically adjusting the prediction of subsequent tokens using the already generated visual content;
Softmax is not Enough (for Adaptive Conformal Classification)
Navid Akhavan Attar (University of Melbourne), Uwe Aickelin (University of Melbourne)
ClassificationImage
🎯 What it does: The study improves distribution-free uncertainty estimation in deep learning classification by embedding Helmholtz energy (computed from pre-softmax logits) into non-consistency scores.
Softmax Transformers are Turing-Complete
Hongjian Jiang (RPTU Kaiserslautern-Landau), Anthony Widjaja Lin
TransformerTextChain-of-Thought
🎯 What it does: Demonstrates that Chain-of-Thought (CoT) Transformers using soft maximum (softmax) attention can achieve Turing completeness and possess length-generalizability capabilities.
SoLoPO: Unlocking Long-Context Capabilities in LLMs via Short-to-Long Preference Optimization
Huashan Sun (Alibaba Group), Fei Huang (Alibaba Group)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a framework called SoLoPO, which decomposes long-text preference optimization into short-text preference optimization and short-to-long reward alignment, thereby improving training efficiency and model generalization.
Solving Football by Exploiting Equilibrium Structure of 2p0s Differential Games with One-Sided Information
Mukesh Ghimire (Arizona State University), Yi Ren (Arizona State University)
OptimizationReinforcement Learning
🎯 What it does: This paper investigates leveraging the structure of information asymmetry in two-player zero-sum differential games (2p0s1), proving that equilibrium strategies in continuous action spaces are atomic, thereby reducing the game tree complexity from exponential to polynomial level; based on this, three methods (CAMS, CAMS-DRL, CAMS-MPC) are proposed to achieve efficient solutions in scenarios such as value approximation, model-free multi-agent reinforcement learning, and model predictive control.
Solving General-Utility Markov Decision Processes in the Single-Trial Regime with Online Planning
Pedro Pinto Santos (INESC-ID & Instituto Superior Técnico), Francisco S. Melo (INESC-ID & Instituto Superior Técnico)
OptimizationReinforcement LearningBenchmark
🎯 What it does: Proposes a solution method for single-trial infinite-horizon discounted generalized utility Markov decision processes (GUMDP), constructing an occupancy MDP and employing Monte Carlo tree search (MCTS) for online planning
Solving Parameter-Robust Avoid Problems with Unknown Feasibility using Reinforcement Learning
Oswin So (MIT), Chuchu Fan (MIT)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: Propose a Feasibility-Guided Exploration (FGE) method that combines feasibility estimation with saddle-point solving to solve parameter robust avoidance problems under unknown feasibility conditions.
Solving the 2-norm k-hyperplane clustering problem via multi-norm formulations
Stefano Coniglio (University of Bergamo)
OptimizationComputational Efficiency
🎯 What it does: This paper proposes a method that significantly improves the solving efficiency of spatial branch-and-bound (SBB) by incorporating 1-norm and infinity-norm constraints into the classical 2-norm k-hyperplane clustering (k-HC2) model through multi-norm relaxation.
Solving the Granularity Mismatch: Hierarchical Preference Learning for Long-Horizon LLM Agents
Heyang Gao (Guangdong Laboratory of Artificial Intelligence and Digital Economy), Xu Chen (Baidu Inc.)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose the Hierarchical Preference Learning (HPL) framework to address the granularity mismatch issue in large language model agents for long-sequence tasks, optimizing the agent through Direct Preference Optimization (DPO) at three levels: trajectory, step, and action group.
Some Neural Networks Inherently Preserve Subspace Clustering Structure
Karan Vikyath Veeranna Rupashree (University of Wisconsin-Madison), Daniel L. Pimentel-Alarcón (University of Wisconsin-Madison)
Representation LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerImage
🎯 What it does: This paper investigates and formally proves that neural network layers can retain the subspace clustering structure of input data without explicit induction, providing sufficient conditions and deriving theoretical bounds on the preservation degree.
SONA: Learning Conditional, Unconditional, and Matching-Aware Discriminator
Yuhta Takida (Sony AI), Yuki Mitsufuji (Sony AI)
GenerationGenerative Adversarial NetworkImage
🎯 What it does: Proposed a new discriminator design called SONA aimed at addressing the balance between realism and conditional alignment in conditional generation.
SONATA: Synergistic Coreset Informed Adaptive Temporal Tensor Factorization
Maolin Wang (City University of Hong Kong), Xiangyu Zhao (City University of Hong Kong)
OptimizationTime Series
🎯 What it does: Propose the SONATA framework for online decomposition and dynamic embedding learning of high-frequency tensor streams.
SongEcho: Towards Cover Song Generation via Instance-Adaptive Element-wise Linear Modulation
Sifei Li (MAIS Institute of Automation Chinese Academy of Sciences), Weiming Dong (MAIS Institute of Automation Chinese Academy of Sciences)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelAudio
🎯 What it does: Proposed the SongEcho framework for generating new singing and accompaniment under given original singer melody and text prompts, achieving full-song cover generation.
SONIC: Spectral Oriented Neural Invariant Convolutions
Gijs Joppe Moens (Netherlands Cancer Institute), Eduardo H P Pooch
ClassificationSegmentationConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: Proposed the SONIC module, a convolutional alternative based on continuous spectral parameterization, achieving global receptive field and resolution-invariant convolution operations through direction-aware low-rank spectral filters;
SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations
Wentao Guo (Princeton University), Tri Dao (Princeton University)
OptimizationComputational EfficiencyMixture of ExpertsText
🎯 What it does: This paper proposes SonicMoE, a comprehensive system for fine-grained and sparse MoE training, including memory-efficient forward/backward algorithms, GPU kernels with overlapping IO and arithmetic operations, and block-size-based Token Rounding routing.
SophiaVL-R1: Reinforcing MLLMs Reasoning with Thinking Reward
Kaixuan Fan (MMLab, Chinese University of Hong Kong), Xiangyu Yue (MMLab, Chinese University of Hong Kong)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: The Trust-GRPO reinforcement learning framework, which combines global thinking rewards and rule-based rewards, was used to train the multimodal large language model SophiaVL-R1 to enhance its reasoning capabilities.
SoSBench: Benchmarking Safety Alignment on Six Scientific Domains
Fengqing Jiang (University of Washington), Radha Poovendran (University of Washington)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkPhysics Related
🎯 What it does: Proposes the SOSBENCH safety benchmark to evaluate the safety alignment of large language models in six high-risk scientific domains (chemistry, biology, medicine, pharmacology, physics, psychology).
Source-Guided Flow Matching
Zifan Wang (KTH Royal Institute of Technology), Karl Henrik Johansson
GenerationFlow-based ModelImagePhysics Related
🎯 What it does: Proposed Source-Guided Flow Matching (SGFM), which achieves conditional generation by modifying the source distribution while keeping the original flow-matching vector field unchanged.
SP-VLA: A Joint Model Scheduling and Token Pruning Approach for VLA Model Acceleration
Ye Li, Wenwu Zhu (Tsinghua University)
Computational EfficiencyRobotic IntelligenceVision-Language-Action ModelMultimodality
🎯 What it does: Achieve real-time acceleration for vision-language-action (VLA) models through joint model scheduling and spatial semantic dual perception token pruning.
SpaCE-10: A Comprehensive Benchmark for Multimodal Large Language Models in Compositional Spatial Intelligence
Ziyang Gong (Shanghai Jiao Tong University), Rongrong Ji (Xiamen University)
Large Language ModelVision Language ModelMultimodalityPoint CloudBenchmark
🎯 What it does: This paper introduces SpaCE-10, a benchmark dataset for evaluating the spatial reasoning capabilities of multimodal large language models.
SpaCE-Eval: A Benchmark for Real-World Multi-Modal Reasoning
Xuyou Yang (Singapore University of Technology and Design), Immanuel Koh (Singapore University of Technology and Design)
Large Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes the SpaCE-Eval benchmark to evaluate the capabilities of multimodal large language models in real-world spatial reasoning, common sense knowledge, and environmental interaction.
SpaceControl: Introducing Test-Time Spatial Control to 3D Generative Modeling
Elisabetta Fedele (ETH Zurich), Leonidas Guibas (Stanford University)
GenerationData SynthesisVision Language ModelFlow-based ModelRectified FlowImageTextMultimodalityMesh
🎯 What it does: Propose a training-free spatial control method called SPACECONTROL, which directly uses 3D geometry (e.g., super tetrahedrons or full meshes) to guide existing 3D generative models to produce high-quality, shape-matching assets.
SPACeR: Self-Play Anchoring with Centralized Reference Models
Wei-Jer Chang (Applied Intuition), Wei Zhan (Applied Intuition)
Autonomous DrivingReinforcement LearningPoint CloudSequential
🎯 What it does: Propose a framework named SPACER that combines self-play reinforcement learning with a pre-trained tokenized motion model, utilizing the model's likelihood reward and KL divergence to guide learning, thereby obtaining driving strategies that exhibit human-like behavior and can be efficiently scaled in multi-agent environments.
SpareTrain: Fault-Tolerant LLM Training via Low-Cost Dual Modular Redundancy
Rihae Park (Seoul National University), Jae W. Lee (Seoul National University)
Computational EfficiencyTransformerText
🎯 What it does: This paper proposes SpareTrain, which combines activation checkpoints and GPU idle time to achieve full-coverage DMR, reducing the SDC detection overhead in LLM training.
Sparkle: A Robust and Versatile Representation for Point Cloud-based Human Motion Capture
Yiming Ren (ShanghaiTech University), Yuexin Ma (Nanyang Technological University)
Pose EstimationConvolutional Neural NetworkRecurrent Neural NetworkPoint Cloud
🎯 What it does: Propose the SparkleMotion framework, which realizes point cloud-based human motion capture using Sparkle representation (skeleton joints + surface anchors);
Sparling: End-to-End Spatial Concept Learning via Extremely Sparse Activations
Kavi Gupta (Massachusetts Institute of Technology), Armando Solar-Lezama (Massachusetts Institute of Technology)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerImageTextAudio
🎯 What it does: Proposed an end-to-end spatial concept learning framework named SPARLING with extreme sparse activation, demonstrating that intermediate sparse variables (motifs) can be identified under assumptions of locality, sparsity, and non-overlapping.
Sparse Attention Adaptation for Long Reasoning
Yizhao Gao (The University of Hong Kong), Mao Yang (Microsoft Research)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: For long-sequence decoding in inference models, this paper proposes SeerAttention-R, a sparse attention framework that achieves dynamic block-level sparse decoding through a lightweight gating module while keeping the original model parameters unchanged.
Sparse Autoencoders Trained on the Same Data Learn Different Features
Gonçalo Paulo (EleutherAI), Nora Belrose (EleutherAI)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningLarge Language ModelAuto EncoderText
🎯 What it does: This study explores the feature differences caused by different initializations by training sparse autoencoders (SAE) on the same large language model and data with different random seeds, quantifying the proportion of shared features.
Sparse but Critical: A Token-Level Analysis of Distributional Shifts in RLVR Fine-Tuning of LLMs
Haoming Meng, Jingren Zhou
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper conducts a fine-grained, token-level analysis of distribution changes during the fine-tuning process of reinforcement learning with verifiable rewards (RLVR) on large language models, revealing that RLVR primarily enhances reasoning capabilities through sparse, targeted probability redistribution;
Sparse CLIP: Co-Optimizing Interpretability and Performance in Contrastive Learning
Chuan Qin (University Of Oxford), Stefan Scherer (University Of Oxford)
Explainability and InterpretabilityRepresentation LearningTransformerContrastive LearningMultimodality
🎯 What it does: By incorporating non-negative constraints and dimension expansion into CLIP pre-training, self-consistent learning of sparse features is achieved, resulting in interpretable multi-modal representations;
Sparse Imagination for Efficient Visual World Model Planning
Junha Chun (Seoul National University), Taesup Kim (Seoul National University)
Computational EfficiencyRobotic IntelligenceTransformerReinforcement LearningVision Language ModelWorld ModelImage
🎯 What it does: Propose the Sparse Imagination method, which uses only a randomly dropped subset of visual patches in visual world models for forward prediction, significantly accelerating planning while maintaining control accuracy.
SparseD: Sparse Attention for Diffusion Language Models
Zeqing Wang (National University of Singapore), Xinchao Wang (National University of Singapore)
GenerationComputational EfficiencyTransformerLarge Language ModelDiffusion modelTextMultimodalityBenchmark
🎯 What it does: Propose the SparseD sparse attention scheme, specifically designed for diffusion language models (DLM), aiming to significantly improve inference speed while maintaining original performance, particularly suitable for long-context scenarios.
SparseEval: Efficient Evaluation of Large Language Models by Sparse Optimization
Taolin Zhang (Shenzhen University), Jindong Wang (Tsinghua University)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose an efficient large language model evaluation framework called SparseEval based on sparse optimization, which can approximate the complete evaluation results by selecting a small number of representative samples (anchors) and learning weights for them.
Sparsity Forcing: Reinforcing Token Sparsity of MLLMs
Feng Chen (University of Adelaide), Qi Wu (University of Adelaide)
Computational EfficiencyLarge Language ModelReinforcement LearningVision Language ModelImageVideoMultimodality
🎯 What it does: Propose a post-training reinforcement learning framework called Sparsity Forcing, which directly learns dynamic token sparsification in multimodal large language models to achieve efficient inference;
SPARTA: Scalable and Principled Benchmark of Tree-Structured Multi-hop QA over Text and Tables
Sungho Park (Pohang University of Science and Technology), Wook-Shin Han (Pohang University of Science and Technology)
Large Language ModelTextTabularBenchmark
🎯 What it does: Built a scalable Table-Text QA benchmark SPARTA and provided an end-to-end auto-generation framework.
Spatial CAPTCHA: Generatively Benchmarking Spatial Reasoning for Human-Machine Differentiation
Arina Kharlamova (Mohamed bin Zayed University of Artificial Intelligence), Xue Liu (City University of Hong Kong)
GenerationData SynthesisImageBenchmark
🎯 What it does: Proposed a spatial reasoning-based CAPTCHA generation framework (Spatial CAPTCHA) and its corresponding benchmark dataset (Spatial-CAPTCHA-Bench), achieving programmable, adjustable difficulty CAPTCHA generation.
Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model
Fuhao Li (Hong Kong University of Science and Technology (Guangzhou)), Haoang Li (Hong Kong University of Science and Technology (Guangzhou))
Representation LearningRobotic IntelligenceVision-Language-Action ModelMultimodality
🎯 What it does: Propose the Spatial Forcing (SF) approach, implicitly enhancing the VLA's 3D perception and action accuracy by aligning the intermediate visual embeddings of VLA with the spatial representations from the 3D foundation model output.
Spatial Mental Modeling from Limited Views
Qineng Wang (Northwestern University), Manling Li (Northwestern University)
Representation LearningReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the MINDCUBE benchmark to evaluate the spatial reasoning ability of Vision-Language Models (VLMs) under limited perspectives, and enhance spatial understanding through three cognitive scaffolds (perspective interpolation, cognitive maps, and free-form reasoning) combined with SFT and RL training strategies;
Spatial Reasoning with Vision-Language Models in Ego-Centric Multi-View Scenes
Mohsen Gholami (Huawei Technologies Canada), Mohammad Akbari (Huawei Technologies Canada)
Depth EstimationAutonomous DrivingVision Language ModelImageVideoTextBenchmark
🎯 What it does: Introduce Ego3D-Bench to evaluate VLM in ego-centric multi-view environments and propose Ego3D-VLM as a training-free framework to enhance VLM's 3D understanding capabilities.
Spatial Structure and Selective Text Jointly Facilitate Image Clustering
Zizheng Jiu (Shanxi University), Lu Chen (Shanxi University)
Knowledge DistillationConvolutional Neural NetworkGraph Neural NetworkVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes a joint image clustering framework SATC that leverages spatial structure and selective textual information.
Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models
Xinmiao Huang (University of Liverpool), Xiaowei Huang (University of Liverpool)
Data SynthesisLarge Language ModelSupervised Fine-TuningMultimodalityBenchmark
🎯 What it does: Investigated and proposed the Spatial-DISE unified spatial reasoning benchmark along with a 12K-scale verifiable dataset to evaluate and diagnose the capabilities of Vision-Language Models (VLMs) in spatial reasoning.
SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild?
Azmine Toushik Wasi (Computational Intelligence and Operations Laboratory), Md Rizwan Parvez (Qatar Computing Research Institute)
Supervised Fine-TuningAgentic AIVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Designed and released the SPATIALAB benchmark, containing 1400 real-world visual question answering (VQA) samples spanning six categories (relative position, depth and occlusion, direction, size and scale, spatial navigation, 3D geometry) and 30 subtasks, supporting both multiple-choice and open-ended evaluation modes.
SpatialHand: Generative Object Manipulation from 3D Prespective
Zehan Wang (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationPose EstimationDepth EstimationTransformerVision Language ModelDiffusion modelAuto EncoderImageTextMeshRetrieval-Augmented Generation
🎯 What it does: Proposes SpatialHand, a generative framework for object insertion and movement from a 3D perspective, supporting full 6DoF (position + orientation) control.
SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models
Hongxing Li (Zhejiang University), Yueting Zhuang (Zhejiang University)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: Proposed the SpatialLadder-26k multimodal dataset and a three-stage progressive training framework, systematically building the spatial reasoning capabilities of vision-language models from object localization → spatial understanding → spatial reasoning.
Spatially Guided Training for Vision-Language-Action Model
Jinhui Ye (Hong Kong University of Science and Technology), Jiangmiao Pang (Shanghai AI Laboratory)
Robotic IntelligenceTransformerVision-Language-Action ModelImageMultimodality
🎯 What it does: Proposed the ST4VLA dual-system visual-language-action framework, integrating spatial priors into robot control through spatially guided training;
Spatially Informed Autoencoders for Interpretable Visual Representation Learning
Dominik Sturm (Dresden University of Technology), Ivo F. Sbalzarini (Dresden University of Technology)
Explainability and InterpretabilityRepresentation LearningAuto EncoderImageBiomedical Data
🎯 What it does: This paper proposes the Spatial Information Variational Autoencoder (SI-VAE), a self-supervised deep model that embeds spatial distribution information from images into latent vectors using a point process likelihood, thereby learning interpretable spatial representations and achieving zero-shot conditional simulation.
SpatialViz-Bench: A Cognitively-Grounded Benchmark for Diagnosing Spatial Visualization in MLLMs
Siting Wang (Chinese Academy of Sciences), Jun Wang (University College London)
Data SynthesisLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes SpatialViz-Bench, a multi-modal benchmark based on cognitive science for evaluating the spatial visualization capabilities of large language models.
SpeakerVid-5M: A Large-Scale High-Quality Dataset for Audio-Visual Dyadic Interactive Human Generation
Youliang Zhang (Tsinghua University), Xiu Li (Tsinghua University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelAuto EncoderVideoMultimodalityBenchmarkAudio
🎯 What it does: Proposed the SpeakerVid-5M large-scale high-quality two-person audio-visual interactive virtual human dataset, and designed the corresponding benchmark (VidChatBench) and autoregressive generative model;
SpecBranch: Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism
Yuhao Shen (Zhejiang University), Cong Wang (Zhejiang University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose SpecBranch, a framework that achieves parallel speculative decoding through branch resampling and hybrid rollback-aware draft length prediction, to accelerate large language model inference.
Special Unitary Parameterized Estimators of Rotation
Akshay Chandrasekhar
Pose EstimationPoint CloudPhysics Related
🎯 What it does: This paper reformulates the Wahba problem using the special unitary matrix SU(2) and proposes new rotation estimation algorithms and two high-dimensional representations for neural networks to learn rotations;
Specialization after Generalization: Towards Understanding Test-Time Training in Foundation Models
Jonas Hübotter (ETH Zürich), Gil Kur (ETH Zürich)
ClassificationRecognitionRepresentation LearningTransformerAuto EncoderImageText
🎯 What it does: This paper explores the mechanism by which testing-time training (TTT) improves performance on in-distribution data in foundational models, proposing that under-parameterization during global training enables specific tasks to transition 'from generalization to specialization' through TTT.
SPECS: Decoupling Multimodal Learning via Self-distilled Preference-based Cold Start
Kun Chen (University of Chinese Academy of Sciences), Lin Ma (Meituan)
Knowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Proposes a three-stage self-distillation preference cold start framework (SPECS), which significantly improves model reasoning performance in multi-modal reasoning tasks by generating preference data through self-distillation, using DPO for pre-alignment, and then applying GRPO reinforcement learning.
Spectral Attention Steering for Prompt Highlighting
Weixian Waylon Li (University of Edinburgh), Shay B Cohen (University of Edinburgh)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Proposes two training-free attention modulation methods, SEKA and AdaSEKA, achieving Prompt Highlighting by directly editing key embeddings before attention computation.
Spectral Bellman Method: Unifying Representation and Exploration in RL
Ofir Nabati (Technion), Guy Tennenholtz (Google Research)
Representation LearningReinforcement LearningImage
🎯 What it does: Propose the Spectral Bellman Method (SBM), which learns feature representations consistent with Bellman updates under the zero Inherent Bellman Error (IBE) condition, and integrates Thompson Sampling to achieve structured exploration.
Spectral-guided Physical Dynamics Distillation
Youjin Kim (Chung-Ang University), Junseok Kwon (Chung-Ang University)
Knowledge DistillationRepresentation LearningDrug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerGraphTime SeriesBiomedical DataPhysics Related
🎯 What it does: Propose a knowledge distillation framework named SGDD based on future trajectory privileged information, which learns spatiotemporal frequency-aware dynamics representations and achieves long-term trajectory prediction even when privileged information is absent.
SpectralGCD: Spectral Concept Selection and Cross-modal Representation Learning for Generalized Category Discovery
Lorenzo Caselli (University of Florence), Andrew D. Bagdanov (University of Florence)
ClassificationKnowledge DistillationRepresentation LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose the SpectralGCD method, which utilizes CLIP's image-concept similarity as a unified cross-modal representation, trains a parameterized classifier within GCD, and enhances performance through spectral filtering and forward-backward knowledge distillation.
SpectraLLM: Uncovering the Ability of LLMs for Molecule Structure Elucidation from Multi-Spectra
Yunyue Su (New Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences), Zhaoxiang Zhang (New Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences)
Computational EfficiencyRepresentation LearningDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBiomedical Data
🎯 What it does: Using large language models to infer molecular structures from various spectra (IR, Raman, UV-Vis, NMR, MS) and uniformly process both single-modal and multi-modal inputs;
Spectrum Tuning: Post-Training for Distributional Coverage and In-Context Steerability
Taylor Sorensen (University of Washington), Yejin Choi (Stanford University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a post-training method called SPECTRUM TUNING to enhance the adjustability, output coverage, and distribution alignment capability of language models in distributed generation tasks.
Speculative Actions: A Lossless Framework for Faster AI Agents
Naimeng Ye (Columbia University), Tianyi Peng (Columbia University)
Computational EfficiencyLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed a 'Speculative Actions' framework that significantly reduces overall latency while maintaining lossless performance by using a fast model to predict future API calls in the environment interaction loop of AI agents and executing these predictions in parallel.
Speculative Speculative Decoding
Tanishq Kumar (Stanford University), Avner May (Together AI)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes the Speculative Decoding (SSD) framework and its implementation, SAGUARO, which can parallelly predict and prepare multiple validation results during verification, fully hiding the sequence latency of the draft model to achieve lossless acceleration;
Speech World Model: Causal State–Action Planning with Explicit Reasoning for Speech
Xuanru Zhou (Zhejiang University), Gopala Anumanchipalli (Zhejiang University)
RecognitionExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelWorld ModelAudio
🎯 What it does: Propose a speech world model (SWM) based on a causal graph, decomposing speech understanding into four modules: context activation (WMA), theory of mind (ToM), pragmatic actions (SA), and goals (Prag), and achieving explicit reasoning through causal relationships.
Speech-to-LaTeX: New Models and Datasets for Converting Spoken Equations and Sentences
Dmitrii Korzh (AXXX), Ivan Oseledets (AXXX)
RecognitionData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkAudio
🎯 What it does: Developed and released a large-scale bilingual (English, Russian) spoken math dataset (S2L), and proposed multiple speech-to-LaTeX methods covering two tasks: standalone equations and embedded sentences;
SpeechJudge: Towards Human-Level Judgment for Speech Naturalness
Xueyao Zhang (Chinese University of Hong Kong Shenzhen), Zhizheng Wu (Chinese University of Hong Kong Shenzhen)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningBenchmarkChain-of-ThoughtAudio
🎯 What it does: Developed the SpeechJudge series, which includes a large-scale naturalness preference dataset, evaluation benchmarks, and generation-based reward models.
SpeechOp: Inference-Time Task Composition for Generative Speech Processing
Justin Lovelace (Cornell University), Zeyu Jin (Adobe Research)
RestorationGenerationTransformerDiffusion modelAuto EncoderTextAudio
🎯 What it does: This paper proposes SpeechOp, a multi-task speech processor based on latent diffusion models, which transfers a pre-trained TTS model into a general-purpose speech-to-speech (S2S) processing tool. It achieves task combination during inference through Implicit Task Composition (ITC), leveraging Whisper ASR text guidance for enhancement.
SPEED: Scalable, Precise, and Efficient Concept Erasure for Diffusion Models
Ouxiang Li (University of Science and Technology of China), Fuli Feng (Hefei University of Technology)
GenerationComputational EfficiencyDiffusion modelImageText
🎯 What it does: Propose the SPEED method to achieve scalable, precise, and efficient concept elimination in text-to-image diffusion models;
SPELL: Self-Play Reinforcement Learning for Evolving Long-Context Language Models
Ziyi Yang (Sun Yat-sen University), Fei Huang (Tongyi Lab, Alibaba Group)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Developed the SPELL framework, a multi-agent self-play reinforcement learning system that enhances long-context reasoning capabilities through a three-role cycle of question-answering, answering, and verification using a single LLM without human-labeled data.
SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models
Chenyu Wang (Meta Superintelligence Labs), Bo Liu (Meta Superintelligence Labs)
OptimizationTransformerReinforcement LearningDiffusion modelTextBenchmark
🎯 What it does: Proposed a reinforcement learning algorithm called Sandwiched Policy Gradient (SPG) for post-training Masked Diffusion Language Models (MDLM);
Spherical Watermark: Encryption-Free, Lossless Watermarking for Diffusion Models
Xiaoxiao Hu (Fudan University), Zhenxing Qian (Fudan University)
GenerationDiffusion modelImage
🎯 What it does: Proposes a lossless watermarking framework called Spherical Watermark, which can embed fully retrievable binary watermarks without altering the diffusion model and without affecting image quality;
SPICE: Submodular Penalized Information–Conflict Selection for Efficient Large Language Model Training
Powei Chang (Bilibili Inc), Dongying kong
OptimizationComputational EfficiencyData-Centric LearningLarge Language ModelText
🎯 What it does: Propose a conflict-aware greedy data selection framework named SPICE, which quantifies the impact of gradient conflicts on information gain decay using the ε-decomposition of submodular functions, and achieves higher quality data subsets by penalizing gradient conflicts;
Spike-based Digital Brain: a novel fundamental model for brain activity analysis
Shaolong Wei, Jiashuang Huang (Nantong University)
ClassificationRepresentation LearningSpiking Neural NetworkTransformerTime SeriesBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: Proposed a spiking-based digital brain model called Spike-DB, which encodes fMRI time series into spike sequences using IIR filter spiking neurons, and achieves high-precision modeling of brain activity and downstream tasks through self-supervised Anchor-Target prediction;
SPIKE-RL: Video-LLMs meet Bayesian Surprise
Sahithya Ravi (University of British Columbia, Vector Institute for AI), Vered Shwartz (University of British Columbia, Vector Institute for AI)
Anomaly DetectionLarge Language ModelReinforcement LearningVision Language ModelVideoText
🎯 What it does: Propose the SPIKE framework, which utilizes the reasoning process of video LLMs to perform Bayesian surprise quantification on surprising events in videos, and uses this signal for query-free frame sampling; subsequently, SPIKE-RL reinforcement learning further optimizes belief generation, enhancing surprise localization and downstream task performance.
SpikeGen: Decoupled “Rods and Cones” Visual Representation Processing with Latent Generative Framework
Gaole Dai (Peking University), Shanghang Zhang (Peking University)
RestorationGenerationTransformerDiffusion modelAuto EncoderContrastive LearningMultimodality
🎯 What it does: Proposed and implemented SpikeGen, a dual-modal visual processing framework based on latent generative models, capable of fusing RGB and spiking (event) data to achieve defogging, frame reconstruction, and high-frequency perspective synthesis.
SpikePingpong: Spike Vision-based Fast-Slow Pingpong Robot System
Hao Wang (Peking University), Shanghang Zhang (Peking University)
Object DetectionRobotic IntelligenceConvolutional Neural NetworkTransformerImage
🎯 What it does: Developed a high-speed table tennis robot system integrating spiking vision and imitation learning;
SpikeStereoNet: A Brain-Inspired Framework for Stereo Depth Estimation from Spike Streams
Zhuoheng Gao (Peking University), Tiejun Huang (Peking University)
Depth EstimationConvolutional Neural NetworkSpiking Neural NetworkSequential
🎯 What it does: Proposed SpikeStereoNet, which can directly estimate stereo depth from raw pulse streams of two spike cameras, and created a large-scale synthetic and real-world pulse stream depth dataset.
Spiking Discrepancy Transformer for Point Cloud Analysis
Yijie Lu (Peking University), Zhaokun Zhou (Tencent)
ClassificationSegmentationSpiking Neural NetworkTransformerPoint Cloud
🎯 What it does: Propose a Spiking Discrepancy Transformer based on synaptic difference attention mechanisms for 3D point cloud analysis;
Spilled Energy in Large Language Models
Adrian Robert Minut (Sapienza University of Rome), Iacopo Masi (Sapienza University of Rome)
Anomaly DetectionTransformerLarge Language ModelTextBenchmark
🎯 What it does: Treat the softmax of LLMs as an energy model, defining two untrained metrics, Spilled Energy and Marginal Energy, to detect hallucinations in generated text.
Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives
Chloe Li (University College London), Daniel Tan (University College London)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Trained the model through self-report fine-tuning (SRFT) to enable it to honestly admit hidden errors or hidden goals after task completion.
SpinBench: Perspective and Rotation as a Lens on Spatial Reasoning in VLMs
Yuyou Zhang (Carnegie Mellon University), Ding Zhao (Mitsubishi Electric Research Labs)
Vision Language ModelImageVideoBenchmarkChain-of-Thought
🎯 What it does: Created and released SPINBENCH, a diagnostic spatial reasoning benchmark designed based on cognitive science, to evaluate the capabilities of vision-language models (VLMs) in spatial tasks such as perspective transformation, rotation, translation, and relative pose;
SpineBench: A Clinically Salient, Level-Aware Benchmark Powered by the SpineMed-450k Corpus
Ming Zhao, Caifeng Shan (Nanjing University)
TransformerLarge Language ModelVision Language ModelMultimodalityBiomedical DataBenchmark
🎯 What it does: Proposed the SpineMed-450k large-scale multimodal instruction dataset and the SpineBench level-aware evaluation framework, and trained SpineGPT based on these.
Spinning Straw into Gold: Relabeling LLM Agent Trajectories in Hindsight for Successful Demonstrations
Zichao Li (Mila, McGill University), Jihyung Kil (Adobe Research)
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Improve agent performance by retrospectively re-annotating the trajectories of LLM agents using the post-training method HSL, incorporating the naturally achieved language goals into the training set.
SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning
Bo Liu (National University Of Singapore), Natasha Jaques
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: In the paper, the authors propose the SPIRAL framework, which trains large language models (LLMs) through self-play in two-player zero-sum language games, thereby enhancing reasoning capabilities without requiring human-annotated rewards.