ICLR 2026 Papers — Page 45
International Conference on Learning Representations · 5356 papers
Splat and Distill: Augmenting Teachers with Feed-Forward 3D Reconstruction For 3D-Aware Distillation
David Shavin (Hebrew University of Jerusalem), Sagie Benaim (Hebrew University of Jerusalem)
SegmentationDepth EstimationKnowledge DistillationNeural Radiance FieldGaussian SplattingImagePoint Cloud
🎯 What it does: Propose the Splat and Distill framework, which utilizes feed-forward 3D reconstruction to elevate the features of the 2D VFM teacher to 3D, then generates new view features through projection, and subsequently distills them to the student network;
Splat Feature Solver
Butian Xiong (University of Southern California), Andrew Feng (University of Southern California)
SegmentationOptimizationGaussian SplattingImagePoint CloudBenchmark
🎯 What it does: Propose a unified sparse linear inverse problem framework that lifts 2D dense features (e.g., CLIP, DINO) to 3D splat representations using a closed-form solver, while enhancing robustness through two regularizations: Tikhonov Guidance and Post-Lifting Aggregation.
Splat Regression Models
Mara Daniels (Massachusetts Institute of Technology), Philippe Rigollet (Massachusetts Institute of Technology)
OptimizationExplainability and InterpretabilityRepresentation Learning
🎯 What it does: A new class of function approximators, called the Splat regression model, is proposed. The model's output is a mixture of heterogeneous and anisotropic bump functions, with each bump weighted by an output vector.
Splat the Net: Radiance Fields with Splattable Neural Primitives
xilong zhou, Christian Theobalt (Max Planck Institute for Informatics)
OptimizationComputational EfficiencyNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Proposed splattable neural primitives, achieving the unification of high expressiveness and real-time rendering.
SplitLoRA: Balancing Stability and Plasticity in Continual Learning Through Gradient Space Splitting
Haomiao Qiu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Meta LearningTransformerImageBenchmark
🎯 What it does: Propose the SplitLoRA method, which balances stability and plasticity by leveraging low-rank adapters and gradient subspace partitioning to achieve continuous learning.
SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports
Haotian Xia (Rice University), Hanjie Chen (Rice University)
TransformerSupervised Fine-TuningReinforcement LearningImageVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Created SportR, a cross-sport multimodal question-answering benchmark containing 4,789 images and 2,052 videos, along with 6,841 human-annotated chain-of-thought (CoT) examples serving as training and evaluation standards.
SpotIt: Evaluating Text-to-SQL Evaluation with Formal Verification
Rocky Klopfenstein (Amherst College), Haoze Wu (Amherst College)
AI Code AssistantTextBenchmark
🎯 What it does: Proposed and implemented SPOTIT, a text-to-SQL evaluation pipeline based on formal equivalence verification, and extended VERIEQL to support common SQL operations such as date and string handling.
Spotlight on Token Perception for Multimodal Reinforcement Learning
Siyuan Huang (Shanghai AI Laboratory), Yu Cheng (Nanjing University)
Reinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose the VPPO algorithm by integrating visual perception into multi-modal reinforcement learning, achieving hierarchical optimization of vision-dependent tokens to enhance the reasoning performance of large vision-language models.
SPR$^2$Q: Static Priority-based Rectifier Routing Quantization for Image Super-Resolution
Jingwei Xin (Xidian University), Xinbo Gao (Xidian University)
Super ResolutionImage
🎯 What it does: Propose a static priority-based compensation matrix routing quantization method, SPR²Q, specifically designed for extremely low-bit Mamba architecture image super-resolution models. The method injects a learnable low-rank compensation matrix before quantization and determines the optimal compensation for each layer through offline evaluation, significantly reducing information loss and enhancing reconstruction quality.
SPREAD: Sampling-based Pareto front Refinement via Efficient Adaptive Diffusion
Sedjro Salomon Hotegni (TU Dortmund University), Sebastian Peitz (TU Dortmund University)
OptimizationTransformerDiffusion modelBenchmark
🎯 What it does: Propose a multi-objective optimization framework SPREAD based on diffusion models, which can generate and refine approximate Pareto front points in the decision space.
SPRIG: Improving Large Language Model Performance by System Prompt Optimization
Lechen Zhang (University of Illinois Urbana-Champaign), David Jurgens (University of Michigan)
OptimizationLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Developed the SPRIG system prompt optimization framework, which iteratively constructs system prompts using genetic algorithms to enhance the performance of large language models across multiple tasks, languages, and scales.
SPRINT: Sparse-Dense Residual Fusion for Efficient Diffusion Transformers
Dogyun Park (Snap Inc.), Anil Kag (Snap Inc.)
GenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: Propose a sparse-dense residual fusion framework called SPRINT for training sparse Diffusion Transformers, significantly reducing training and inference costs.
Spurious Correlation-Aware Embedding Regularization for Worst-Group Robustness
Subeen Park (Yonsei University), Kyungwoo Song (Yonsei University)
Domain AdaptationRepresentation LearningImageText
🎯 What it does: Proposed a SCER method that regularizes feature representations in the embedding space to suppress domain-related pseudo-correlated features, thereby improving the worst-group performance under subpopulation distribution shifts.
SPWOOD: Sparse Partial Weakly-Supervised Oriented Object Detection
Wei Zhang (Shanghai Jiao Tong University), Xue Yang (Shanghai Jiao Tong University)
Object DetectionImage
🎯 What it does: Propose a sparse partial weakly supervised oriented object detection framework named SPWOOD, which can achieve high-precision detection using only a small number of sparse weak labels and a large amount of unlabeled data
Squeeze the Soaked Sponge: Efficient Off-policy RFT for Large Language Model
Jing Liang (Tianjin University), Jianye HAO
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Studied how to improve the efficiency of reinforcement fine-tuning for large language models through offline reinforcement learning, proposing a hybrid strategy RFT method called ReMix;
SR-Scientist: Scientific Equation Discovery With Agentic AI
Shijie Xia (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
OptimizationDrug DiscoveryTransformerLarge Language ModelReinforcement LearningAgentic AITextTabularBenchmarkPhysics Related
🎯 What it does: Proposes the SR-SCIENTIST framework, enabling large language models (LLMs) to evolve from mere equation generators into autonomous scientists capable of writing code, analyzing data, evaluating results, and self-optimizing through multi-round interactions to accomplish symbolic regression tasks.
SRFT: A Single-Stage Method with Supervised and Reinforcement Fine-Tuning for Reasoning
Yuqian Fu (State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA), Dongbin Zhao (State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA)
Large Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose a single-stage LLM fine-tuning framework SRFT, integrating supervised fine-tuning (SFT) with reinforcement learning (RL) through an entropy-aware weighting mechanism to achieve unified optimization of both paradigms.
SRT: Super-Resolution for Time Series via Disentangled Rectified Flow
Jufang Duan (Bytedance), Yuren Zhang (Bytedance)
Super ResolutionTransformerRectified FlowTime Series
🎯 What it does: This paper proposes a temporal super-resolution method called SRT, which decomposes low-resolution sequences into trend and seasonal components, aligns them with high-resolution using an implicit time function, and then generates high-resolution details using a decoupled rectified flow.
SSD-GS: Scattering and Shadow Decomposition for Relightable 3D Gaussian Splatting
Iris Zheng (Victoria University of Wellington), Fang-Lue Zhang (Victoria University of Wellington)
GenerationGaussian SplattingImagePoint CloudPhysics Related
🎯 What it does: Propose SSD-GS on the 3D Gaussian Splatting framework, enabling high-quality reconstruction and relighting under new lighting conditions.
SSDi8: Accurate and Efficient 8-bit Quantization for State Space Duality
Hyunwoo Kim (Chung-Ang University), Dahuin Jung (Chung-Ang University)
Computational EfficiencyText
🎯 What it does: Proposes SSDi8, a post-training INT8 quantization framework for Mamba-2 structured state space dual-mode (SSD), enabling continuous INT8 computation paths with significant latency reduction while maintaining accuracy close to FP16.
SSG: Scaled Spatial Guidance for Multi-Scale Visual Autoregressive Generation
Youngwoo Shin (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
GenerationImageText
🎯 What it does: Propose a training-agnostic Scaled Spatial Guidance (SSG) that leverages frequency domain priors to highlight semantic residuals at each step of the VAR model, thereby enhancing high-frequency details and overall coherence between coarse and fine-level generation.
ssToken: Self-modulated and Semantic-aware Token Selection for LLM Fine-tuning
Xiaohan Qin (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A framework called ssToken for token-level data selection in LLM supervised fine-tuning is studied, combining self-modulation and semantic-aware signals to filter useless tokens.
SSVPO: Effective Step-Level Credit Assignment for RL Training of Language Models
Yugu Li (Adelaide University), Siyi Hu (Curtin University)
TransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: In the post-training stage of LLM mathematical reasoning, we propose Sequential Shapley Value Policy Optimization (SSVPO), which achieves step-by-step credit assignment through insertable MDP and Sequential Shapley Value (SSV), and uses SSV as an advantage baseline for PPO optimization.
ST-HHOL: Spatio-Temporal Hierarchical Hypergraph Online Learning for Crime Prediction
Keqing Du (Xi'an Jiaotong University), Wei Shao (Data61)
Explainability and InterpretabilityGraph Neural NetworkTransformerSupervised Fine-TuningGraphTabularTime Series
🎯 What it does: Proposed an online spatiotemporal learning framework ST-HOL for urban streaming crime prediction, integrating hierarchical hypergraph convolutional networks with partially frozen LLMs, capable of capturing high-order crime patterns and adapting to concept drift.
ST-SimDiff: Balancing Spatiotemporal Similarity and Difference for Efficient Video Understanding with MLLMs
Bingjun Luo (Tsinghua University), Xinpeng Ding (Xidian University)
CompressionComputational EfficiencyGraph Neural NetworkVision Language ModelVideo
🎯 What it does: Propose ST-SimDiff, a no-training visual token compression framework based on spatial-temporal graphs, which preserves key video information through dual selection based on similarity and difference;
ST-WebAgentBench: A Benchmark for Evaluating Safety and Trustworthiness in Web Agents
Ido Levy (IBM Research), Segev Shlomov (IBM Research)
Safty and PrivacyLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Developed ST‑WEBAGENTBENCH, a comprehensive benchmark framework for evaluating the security and trustworthiness of web agents.
Stability Under Scrutiny: Benchmarking Representation Paradigms for Online HD Mapping
Hao Shan (State Key Lab Of Intelligent Transportation System), Haiyang Yu (State Key Lab Of Intelligent Transportation System)
Autonomous DrivingRepresentation LearningSimultaneous Localization and MappingPoint CloudBenchmark
🎯 What it does: Introduce a temporal stability evaluation framework, propose the mAS metric, and construct a stability benchmark for online high-definition maps.
Stabilizing Policy Gradients for Sample-Efficient Reinforcement Learning in LLM Reasoning
Luckeciano Carvalho Melo (University of Oxford), Yarin Gal (University of Oxford)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposes the Curvature-Aware Policy Optimization (CAPO) method, which improves sample efficiency in reinforcement learning for large language model inference tasks by monitoring and limiting instability caused by gradients and curvature.
Stable and Scalable Deep Predictive Coding Networks with Meta-Prediction Errors
Myoung Hoon Ha (KAIST), Sang Wan Lee (KAIST)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Investigated the training instability of deep predictive coding networks (PCN) and proposed the Meta-PCN framework to achieve scalable training of deep PCN.
Stable coresets: Unleashing the power of uniform sampling
Amir Carmel (Weizmann Institute of Science), Robert Krauthgamer (Weizmann Institute of Science)
OptimizationComputational Efficiency
🎯 What it does: A new concept of stable core set is proposed, combining the advantages of strong core sets and weak core sets, utilizing uniform sampling to construct stable core sets to improve the efficiency of clustering problems.
Stable Video Infinity: Infinite-Length Video Generation with Error Recycling
Wuyang Li (École Polytechnique Fédérale de Lausanne), Alexandre Alahi (École Polytechnique Fédérale de Lausanne)
GenerationTransformerSupervised Fine-TuningDiffusion modelVideoTextMultimodalityAudio
🎯 What it does: Through Error-Recycling Fine-Tuning, the self-generated errors from the Diffusion Transformer are utilized as supervisory signals, enabling closed-loop learning to correct drift, thereby generating infinitely long, non-cyclic, visually stable long videos without increasing inference costs.
Stable-LoRA: Stabilizing Feature Learning of Low-Rank Adaptation
Yize Wu (University of Chinese Academy of Sciences), Yanjun Wu (University of Chinese Academy of Sciences)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Propose the Stable-LoRA weight decay strategy to address the instability of feature learning during LoRA training, and verify its effectiveness theoretically and experimentally.
STABLE: Shift-Tolerant Allocation via Black-Litterman Using Conditional Diffusion Estimates
YEJUN SOUN (Seoul National University), U Kang (Seoul National University)
Diffusion modelTime SeriesFinance Related
🎯 What it does: Propose the STABLE method, integrating conditional diffusion generation, two-tier guidance, and Black-Litterman optimization to achieve robust asset allocation under macroeconomic cycle changes.
StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs
Yuhan Song (Peking University), Zhou Xiao
Representation LearningAudio
🎯 What it does: Propose a speech tokenizer called StableToken based on a multi-branch voting mechanism to improve tokenization stability in noisy environments
Stacked from One: Multi-Scale Self-Injection for Context Window Extension
Wei Han (Singapore University of Technology and Design), Shuicheng YAN
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the SHAREDLLM framework, which expands the LLM context window to 128K tokens by sharing KV between a lower-level compressor and an upper-level decoder.
Stackelberg Coupling of Online Representation Learning and Reinforcement Learning
Fernando Martinez (Fordham University), Juntao Chen (Fordham University)
Reinforcement Learning
🎯 What it does: Proposed and implemented the SCORER framework, transforming representation learning and value function learning in deep Q learning into a Stackelberg leader-follower game, achieving stable co-adaptation through two-time-scale updates.
Stackelberg Learning from Human Feedback: Preference Optimization as a Sequential Game
Barna Pásztor (ETH Zurich), Andreas Krause (ETH Zurich)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the Stackelberg Learning from Human Feedback (SLHF) framework, modeling preference optimization as a sequential Stackelberg game and designing the STACKELBERGGDA two-time-scale gradient descent ascent algorithm to achieve alignment and in-context improvement for large language models (LLMs).
Stage-wise Dynamics of Classifier-Free Guidance in Diffusion Models
Cheng Jin (Tsinghua University), Yuantao Gu (Tsinghua University)
GenerationDiffusion modelMultimodalityOrdinary Differential Equation
🎯 What it does: Conduct a systematic theoretical analysis of classifier-free guidance (CFG) in diffusion models under multimodal conditional distributions, revealing that the sampling dynamics consist of three stages: directional offset, mode separation, and concentration, while explaining the fundamental mechanisms behind diversity degradation, followed by the design and validation of a time-varying guidance strength scheduling scheme.
STAIRS-Former: Spatio-Temporal Attention with Interleaved Recursive Structure TransFormer for Offline Mulit-task Multi-agent Reinforcement Learning
Jiwon Jeon (Korea Advanced Institute of Science and Technology), Youngchul Sung (Korea Advanced Institute of Science and Technology)
Recurrent Neural NetworkTransformerReinforcement LearningSequentialBenchmark
🎯 What it does: Propose STAIRS-Former, a spatial-temporal hierarchical Transformer structure for offline multi-task multi-agent reinforcement learning, capable of adapting to varying numbers of agents and leveraging historical information;
STaMP: Sequence Transformation and Mixed Precision for Low-Precision Activation Quantization
Marco Federici (Qualcomm AI Research), Markus Nagel (Qualcomm AI Research)
Computational EfficiencyLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the Sequence Transformation with Mixed Precision (STaMP) quantization method, which concentrates activation energy on a few tokens through linear transformation on the sequence dimension without retraining, and assigns higher bitwidths to these high-energy tokens to reduce activation quantization error.
STAR-Bench: Probing Deep Spatio-Temporal Reasoning as Audio 4D Intelligence
Zihan Liu (Beihang University), Jiaqi Wang (Shanghai AI Laboratory)
ClassificationData SynthesisLarge Language ModelBenchmarkAudio
🎯 What it does: Propose and implement the STAR-Bench evaluation framework to systematically assess audio 4D intelligence (spatiotemporal reasoning)
STAR: Similarity-guided Teacher-Assisted Refinement for Super-Tiny Function Calling Models
Jiliang Ni (Alibaba), Conggang Hu (Alibaba)
Knowledge DistillationAI Code AssistantSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the STAR framework, achieving distillation and fine-tuning of function calling capabilities through a teacher-guided ultra-small model.
STAR: Strategy-driven Automatic Jailbreak Red-teaming For Large Language Model
Jianing Liu (Zhejiang University), Shouling Ji (Zhejiang University)
Adversarial AttackLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes the STAR framework, divided into two modules: strategy generation and prompt generation, systematically producing diverse and effective jailbreak prompts;
STARK: Strategic Team of Agents for Refining Kernels
Juncheng Dong (Meta Ranking AI Research), Shuang Yang (Meta Ranking AI Research)
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringBenchmark
🎯 What it does: Proposed a multi-agent collaborative framework STARK based on LLM for automated GPU kernel optimization.
STAT: Skill-Targeted Adaptive Training
Yinghui He (Princeton University), Sanjeev Arora (Princeton University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Developed an adaptive training framework named STAT based on skill gaps, leveraging the metacognitive capabilities of state-of-the-art large language models to analyze skill deficiencies in student models on validation sets. The framework enhances model performance in mathematical reasoning tasks by reweighting existing training samples or synthesizing problems related to missing skills.
Station2Radar: Query‑Conditioned Gaussian Splatting for Precipitation Field
Doyi Kim (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
GenerationSuper ResolutionGraph Neural NetworkGaussian SplattingImageTabular
🎯 What it does: By fusing satellite imagery with automatic weather station observations, a high-resolution, continuously analyzable precipitation field is generated using Query-Conditioned Gaussian Splatting (QCGS).
Statistical Advantage of Softmax Attention: Insights from Single-Location Regression
O Duranthon (Ecole Polytechnique Federale De Lausanne), Lenka Zdeborová (Ecole Polytechnique Federale De Lausanne)
OptimizationRepresentation LearningTransformerSequentialPhysics Related
🎯 What it does: Studied the Single-Location Regression (SLR) task, theoretically analyzed the performance of softmax attention compared to linear/other activation functions in this task, and explored its statistical advantages and learning error under finite samples.
Statistical and structural identifiability in representation learning
Walter Nelson (Institute of Science and Technology Austria), Francesco Locatello (Institute of Science and Technology Austria)
Representation LearningTransformerAuto EncoderImageBiomedical Data
🎯 What it does: This paper investigates the identifiability of representations in representation learning models, proposing definitions of statistical approximate identifiability and structural approximate identifiability. It proves that intermediate layer representations in nonlinear decoder models are statistically approximately identifiable under local bi-Lipschitz conditions. Subsequently, ICA is introduced to eliminate linear uncertainty, and structural approximate identifiability theory is provided based on this. Finally, experiments verify that unsupervised disentanglement and batch effect removal can be achieved in models such as autoencoders, masked autoencoders, and GPT.
Statistical Guarantees for Offline Domain Randomization
Arnaud Fickinger (University of California Berkeley), Stuart Russell (University of California Berkeley)
Domain AdaptationSequential
🎯 What it does: This paper studies the use of offline data for domain randomization (ODR) to improve the transfer effectiveness from simulation to real-world environments.
Statistical Guarantees in the Search for Less Discriminatory Algorithms
Chris Hays, Manish Raghavan
ClassificationTabular
🎯 What it does: Propose an adaptive stopping algorithm to help enterprises prove that they have sufficiently searched for a less discriminatory algorithm (LDA) during model retraining under U.S. anti-discrimination laws.
STDDN: A Physics-Guided Deep Learning Framework for Crowd Simulation
Zijin Liu (Beihang University), You Song (Beihang University)
Graph Neural NetworkSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: The paper proposes the STDDN framework, combining the fluid continuity equation with deep learning to achieve coupling between micro trajectory prediction and macro density evolution.
STEDiff: Revealing the Spatial and Temporal Redundancy of Backdoor Attacks in Text-to-Image Diffusion Models
Yu Pan (ShanghaiTech University), Wenjie Wang (ShanghaiTech University)
GenerationAdversarial AttackDiffusion modelImageText
🎯 What it does: Propose the STEDiff framework, which includes an efficient attack module STEBA and a real-time detection module STEDF, leveraging spatial and temporal redundancy in diffusion models to achieve low-cost backdoor injection and detection.
STEER AWAY FROM MODE COLLISIONS: IMPROVING COMPOSITION IN DIFFUSION MODELS
Debottam Dutta (University of Illinois at Urbana-Champaign), Romit Roy Choudhury (University of Illinois at Urbana-Champaign)
GenerationDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Propose a gradient-free, model-agnostic concept contrastive corrector named CO3 to enhance semantic alignment and concept coverage in multi-concept text-to-image diffusion models.
Steerable Adversarial Scenario Generation through Test-Time Preference Alignment
Tong Nie (Hong Kong Polytechnic University), Jian Sun (Tongji University)
Autonomous DrivingOptimizationAdversarial AttackSupervised Fine-TuningSequential
🎯 What it does: Proposed the SAGE framework, achieving adjustable adversarial scenario generation through preference alignment between adversarial and authentic preferences during inference.
Steering and Rectifying Latent Representation Manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection
Zhaolin Cai, Guangtao Zhai (Xi'an Jiao Tong University)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Proposed a fully frozen multimodal large language model (MLLM) active geometric intervention framework called SteerVAD for unsupervised video anomaly detection;
Steering Autoregressive Music Generation with Recursive Feature Machines
Daniel Zhao (University of California, San Diego), Zachary Novack (University of California, San Diego)
GenerationData SynthesisExplainability and InterpretabilityTransformerAudio
🎯 What it does: By migrating Recursive Feature Machines (RFM) to the audio domain, injecting interpretable musical concepts into the frozen MUSICGEN-Large model, achieving real-time, fine-grained controllable music generation.
Steering Diffusion Models Towards Credible Content Recommendation
Zhuo Cai (University Of Technology Sydney), Charu C. Aggarwal (Ibm T. J. Watson Research Center)
Recommendation SystemTransformerDiffusion modelContrastive LearningText
🎯 What it does: Proposed the Disco model, achieving trustworthy content recommendation by decoupling diffusion models and trustworthy subspace projection.
Steering Evaluation-Aware Language Models To Act Like They Are Deployed
Tim Tian Hua, Neel Nanda
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Trained and tested an evaluation-aware model, using activation-driven mechanisms to ensure consistent performance in evaluation and deployment environments.
Steering Language Models with Weight Arithmetic
Constanza Fierro (University of Copenhagen), Fabien Roger (Anthropic)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose a post-training contrastive weight steering method that controls the behavior of large language models by leveraging weight differences obtained from fine-tuning on positive and negative behavior data;
Steering MoE LLMs via Expert (De)Activation
Mohsen Fayyaz (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: By comparing the differences in expert activation rates within adversarial behavior pairs, experts closely associated with specific behaviors (such as safety and authenticity) are detected. During inference, these experts are activated or suppressed by adjusting the router's logits, achieving weight-unadjusted control over the behavior of large language models (MoE).
Steering the Herd: A Framework for LLM-based Control of Social Learning
Raghu Arghal (University of Pennsylvania), Saswati Sarkar (University of Pennsylvania)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a strategy framework for studying large language models (LLMs) as information intermediaries in controlling social learning, and presents optimal signal precision strategies for altruistic and biased planners. Subsequently, the theoretical predictions are validated through LLM-based simulation experiments, and the interactions between LLM planners' strategies and non-Bayesian behaviors are analyzed.
SteinsGate: Adding Causality to Diffusions for Long Video Generation via Path Integral
Yufei Huang (Zhejiang University), Stan Z. Li (Westlake University)
GenerationTransformerLarge Language ModelDiffusion modelFlow-based ModelVideoTextMultimodalityBenchmarkOrdinary Differential Equation
🎯 What it does: Propose the Instruct-Video-Continuation framework, leveraging Temporal Action Binding and Causal Video Continuation to generate multi-action long videos, and implement the SteinsGate plugin-based inference scheme;
STEM: SCALING TRANSFORMERS WITH EMBEDDING MODULES
Ranajoy Sadhukhan, Beidi Chen (Carnegie Mellon University)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: Proposes the STEM (Scaling Transformers with Embedding Modules) approach, replacing the up projection in Transformer's FFN with a layer-local token-indexed embedding table to achieve fine-grained sparsity;
Step-Aware Residual-Guided Diffusion for EEG Spatial Super-Resolution
Hongjun Liu (University of Science and Technology Beijing), Chao Yao (University of Science and Technology Beijing)
Super ResolutionDiffusion modelAuto EncoderBiomedical Data
🎯 What it does: Propose a dynamic residual-guided diffusion model, SRGDiff, which generates high-density EEG signals from low-density EEG signals to achieve spatial super-resolution.
StepORLM: A Self-Evolving Framework With Generative Process Supervision For Operations Research Language Models
Chenyu Zhou (Shanghai Jiao Tong University), Dongdong Ge (Shanghai Jiao Tong University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposed and implemented the StepORLM framework, which enables joint training of policy models and generative process reward models (GenPRM) through a self-evolutionary loop, achieving dual supervision of both process and results for operations research (OR) problems, significantly enhancing the reasoning and modeling capabilities of large language models (LLMs) in the OR domain.
STITCH: Simultaneous Thinking and Talking with Chunked Reasoning for Spoken Language Models
Cheng-Han Chiang (National Taiwan University), Lijuan Wang (Microsoft)
GenerationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Designed the STITCH framework, enabling speech-language models (SLM) to perform silent reasoning while speaking, and generate reasoning segments in the background through long time delays in audio blocks;
Stochastic Neural Networks for Causal Inference with Missing Confounders
Yaxin Fang (Stanford University), Faming Liang (Purdue University)
Explainability and InterpretabilityTabularBiomedical DataBenchmarkStochastic Differential Equation
🎯 What it does: Propose the CI-StoNet framework, which infers missing confounding variables using sparse deep neural networks and adaptive SGHMC, and estimates causal effects based on this inference.
Stochastic Optimal Control for Continuous-Time fMRI Representation Learning
Joonhyeong Park (KAIST), Juho Lee (Yonsei University)
ClassificationComputational EfficiencyRepresentation LearningTransformerContrastive LearningBiomedical DataMagnetic Resonance ImagingStochastic Differential Equation
🎯 What it does: Propose a continuous-time self-supervised learning framework BDO based on stochastic optimal control, aimed at learning robust brain dynamics representations from fMRI data.
Stochastic Self-Guidance for Training-Free Enhancement of Diffusion Models
Chubin Chen (Tsinghua University), Xiu Li (Tsinghua University)
GenerationDiffusion modelImageVideoText
🎯 What it does: This paper proposes a self-supervised guidance method called S2-Guidance, which does not require additional training. It activates subnetworks by randomly dropping partial network blocks during the denoising process of diffusion models, thereby correcting the suboptimal results of Classifier-free Guidance (CFG) without constructing external weak models.
Stochastic Self-Organization in Multi-Agent Systems
Nurbek Tastan (Mohamed bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes SELFORG, a multi-agent system based on LLMs that constructs a directed acyclic graph through response-adaptive mechanisms to achieve self-organizing collaborative structures;
StochasTok: Improving Fine-Grained Subword Understanding in LLMs
Anya Sims (University of Oxford), Cong Lu (University of British Columbia)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed and implemented a stochastic tokenization scheme called STOCHASTOK, which enhances the fine-grained subword understanding capability of large language models by randomly splitting generated tokens during training, enabling the models to directly perceive subword structures.
Stop Guessing: Choosing the Optimization-Consistent Uncertainty Measurement for Evidential Deep Learning
Linye Li (Tongji University), Qunjie Chen (Tongji University)
Anomaly DetectionOptimizationConvolutional Neural NetworkImage
🎯 What it does: Re-examine Evidential Deep Learning (EDL) from an optimization perspective, revealing its implicit maximum margin characteristic, proposing the optimization consistency principle, and designing a new uncertainty measure MPU that is consistent with the UCE loss. Subsequently, validate its effectiveness in out-of-distribution (OOD) and misclassification detection tasks.
Stop Tracking Me! Proactive Defense Against Attribute Inference Attack in LLMs
Dong Yan (University of Chinese Academy of Sciences), Tieniu Tan (University of Chinese Academy of Sciences)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Research and propose a unified defense framework TRACE-RPS, combining fine-grained anonymization (TRACE) with optimization-based rejection induction (RPS), for actively defending against attribute inference attacks in large language models.
Stop Unnecessary Reflection: Training LRMs for Efficient Reasoning with Adaptive Reflection and Length Coordinated Penalty
Zewei Yu (Zhejiang University), Haobo Wang (Zhejiang University)
Computational EfficiencyLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: A reinforcement learning framework with adaptive reflection and length synergistic penalties is used to train large reasoning models to reduce excessive reflection and verbose outputs during chain-of-thought generation, thereby improving reasoning efficiency and accuracy.
Stop Wasting Your Tokens: Towards Efficient Runtime Multi-Agent Systems
Fulin Lin (Zhejiang University), Tao Lin (Westlake University)
Computational EfficiencyTransformerLarge Language ModelAgentic AIText
🎯 What it does: Proposes the SUPERVISORAGENT framework, leveraging real-time, adaptable supervision to enhance the robustness and token efficiency of multi-agent systems.
Stopping Computation for Converged Tokens in Masked Diffusion-LM Decoding
Daisuke Oba (Institute of Science Tokyo), Naoaki Okazaki (Institute of Science Tokyo)
Computational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: During the MDLM decoding process, the SURELOCK method is proposed, which locks the position by detecting whether the posterior distribution of a single token converges (using a local KL threshold), thereby skipping the query projection and feed-forward sublayer computations in subsequent steps, retaining only the key/value cache for use by other positions, thus gradually reducing computational costs.
STORK: Faster Diffusion and Flow Matching Sampling by Resolving both Stiffness and Structure-Dependence
Zheng Tan (University of California, Los Angeles), Ernest K. Ryu (University of California, Los Angeles)
GenerationDiffusion modelFlow-based ModelImageVideo
🎯 What it does: Propose a new training-free fast sampling method called STORK, which combines robust Runge-Kutta formulas with Taylor expansion to achieve high-quality sampling of noise prediction models and flow matching models with only a small number of NFEs.
STORM: Synergistic Cross-Scale Spatio-Temporal Modeling for Weather Forecasting
Qihe Huang, Yang Wang
Convolutional Neural NetworkTransformerTime SeriesSequentialPhysics Related
🎯 What it does: Propose STORM, a cross-scale collaborative spatiotemporal modeling framework for weather forecasting;
Story-Iter: A Training-free Iterative Paradigm for Long Story Visualization
Jiawei Mao (University Of California Santa Cruz), Yuyin Zhou (University Of California Santa Cruz)
GenerationTransformerDiffusion modelImageTextMultimodality
🎯 What it does: Proposed a training-free iterative paradigm called Story-Iter, which uses the complete story frames from the previous round as a reference to iteratively refine long-form story visualization.
StoryAlign: Evaluating and Training Reward Models for Story Generation
Haotian Xia (Tsinghua University), Juanzi Li (Tsinghua University)
GenerationReinforcement Learning from Human FeedbackLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper proposes the STORYRMB benchmark and trains the STORYREWARD reward model to enhance the quality of LLMs in story generation and improve alignment with human preferences.
StPR: Spatiotemporal Preservation and Routing for Exemplar-Free Video Class-Incremental Learning
Huaijie Wang (Xidian University), Xinbo Gao (Xidian University)
ClassificationKnowledge DistillationTransformerMixture of ExpertsContrastive LearningVideo
🎯 What it does: Propose an example-free, CLIP-based continuous video classification framework StPR, integrating frame-shared semantic distillation (FSSD) and time-decomposed Mixture-of-Experts (TD-MoE), achieving stable retention of spatial semantics and adaptive routing for temporal dynamics;
Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs
Alexander Panfilov (ELLIS Institute Tübingen & MPI for Intelligent Systems), Jonas Geiping (ELLIS Institute Tübingen & MPI for Intelligent Systems)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper investigates the strategic dishonesty exhibited by cutting-edge large language models when faced with harmful requests, and verifies its universality and impact through multiple-choice questions (MCQ) experiments, output monitoring, and internal activation detection;
Strategic Obfuscation of Deceptive Reasoning in Language Models
Arun Jose (Center on Long-Term Risk), Mia Taylor (Center on Long-Term Risk)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The study investigates how to covertly perform malicious reasoning through 'alignment faking' in large language models, and evaluates the behavioral changes of Claude 3.5 Sonnet when prompted that its internal reasoning will be monitored.
Strategic Planning and Rationalizing on Trees Make LLMs Better Debaters
Danqing Wang (Carnegie Mellon University), Lei Li (Carnegie Mellon University)
Reinforcement Learning from Human FeedbackLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed the TreeDebater system for competitive debate scenarios, utilizing a Rehearsal Tree to simulate potential opponent attacks and defenses, a Debate Flow Tree to track debate status in real-time, and combining speech duration control with simulated audience feedback to generate time-constrained, more persuasive speeches.
Strategic Scaling of Test-Time Compute: A Bandit Learning Approach
Bowen Zuo (University of California, Riverside), Yinglun Zhu (University of California, Riverside)
Computational EfficiencyText
🎯 What it does: Proposes modeling the computational allocation during LLM testing as a multi-armed bandit problem and presents an adaptive elimination-based allocation algorithm.
STream3R: Scalable Sequential 3D Reconstruction with Causal Transformer
Yushi LAN, Xingang Pan (Nanyang Technological University)
Pose EstimationDepth EstimationTransformerImagePoint CloudSequential
🎯 What it does: Proposed a causal Transformer based solely on the decoder, named STREAM3R, for online and incremental prediction of dense 3D geometric point maps and camera poses from streaming or unstructured image sequences.
Streaming Autoregressive Video Generation via Diagonal Distillation
Jinxiu Liu (Johns Hopkins University), Weiyang Liu (Chinese University of Hong Kong)
GenerationKnowledge DistillationTransformerDiffusion modelOptical FlowVideo
🎯 What it does: Developed an efficient autoregressive video generation framework called Diagonal Distillation, which employs diagonal denoising and diagonal forcing training to achieve high-quality and low-latency streaming video generation.
Streaming Drag-Oriented Interactive Video Manipulation: Drag Anything, Anytime!
Junbao Zhou (Nanyang Technological University), Hanwang Zhang (Nanyang Technological University)
GenerationTransformerDiffusion modelVideo
🎯 What it does: Propose a new streaming drag-based interactive video editing task REVEL, and develop the DragStream method to achieve drag editing and animation at any position and time during video generation without requiring additional training.
Streaming Visual Geometry Transformer
Dong Zhuo (Tsinghua University), Jiwen Lu (Tsinghua University)
Pose EstimationDepth EstimationKnowledge DistillationTransformerVideoPoint Cloud
🎯 What it does: Proposed a Transformer model called StreamVGGT that can perform low-latency, incremental reconstruction of 3D geometry in streaming video.
StreamingThinker: Large Language Models Can Think While Reading
Junlong Tong (Shanghai Jiao Tong University), Xiaoyu Shen (Eastern Institute Of Technology)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed and implemented the StreamingThinker framework, enabling large language models to perform reasoning as input streams arrive and further adjust reasoning depth after reading completes.
StreamingVLM: Real-Time Understanding for Infinite Video Streams
Ruyi Xu (MIT), Song Han (MIT)
Computational EfficiencyTransformerSupervised Fine-TuningVision Language ModelVideoTextBenchmark
🎯 What it does: Proposed a unified training and inference framework called StreamingVLM, which can maintain low-latency and stable visual language understanding in real-time infinite video streams.
StreamSplat: Towards Online Dynamic 3D Reconstruction from Uncalibrated Video Streams
Zike Wu (University of British Columbia), Renjie Liao (University of British Columbia)
GenerationDepth EstimationTransformerGaussian SplattingOptical FlowVideo
🎯 What it does: Proposes a fully forward-reasoning based online dynamic 3D reconstruction framework called StreamSplat, which can real-time convert arbitrary-length video streams into dynamic 3D Gaussian Splatting (3DGS) representations without relying on camera calibration.
Stretching Beyond the Obvious: A Gradient-Free Framework to Unveil the Hidden Landscape of Visual Invariance
Lorenzo Tausani (International School for Advanced Studies), Davide Zoccolan (International School for Advanced Studies)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: This study proposes the Stretch-and-Squeeze (SnS) framework, which employs gradient-free bi-objective optimization to stretch and compress images in the representation space at arbitrary levels, systematically exploring the maximal invariance and adversarial spaces of visual units.
Strict Subgoal Execution: Reliable Long-Horizon Planning in Hierarchical Reinforcement Learning
Seungyul Han (Ulsan National Institute of Science and Technology), Jeongmo Kim (Ulsan National Institute of Science and Technology)
Reinforcement LearningTabular
🎯 What it does: Proposed the Strict Subgoal Execution (SSE) framework, which utilizes Frontier Experience Replay (FER) to enforce the completion of subgoals in a single step within graph-structured hierarchical reinforcement learning, significantly improving the learning efficiency of long-horizon tasks;
Strictly Constrained Generative Modeling via Split Augmented Langevin Sampling
Matthieu Blanke (New York University), Pierre Gentine (Columbia University)
GenerationDiffusion modelPhysics RelatedStochastic Differential Equation
🎯 What it does: Designed an algorithm called CASAL that strictly samples constraints under a zero-shot scenario through variable splitting and Lagrangian duality;
String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation
Kou Misaki (Sakana AI), Takuya Akiba (Sakana AI)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied a novel prompting method called String Seed of Thought (SSoT), enabling LLMs to select answers according to a given probability distribution and significantly enhance output diversity.
Stroke3D: Lifting 2D strokes into rigged 3D model via latent diffusion models
Ruisi Zhao (Zhejiang University), Yi Yang (Zhejiang University)
GenerationGraph Neural NetworkTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelTextMesh
🎯 What it does: By generating an animatable 3D skeleton from lines and text prompts drawn by users on a 2D canvas, and then utilizing the skeleton to generate texture-mapped, bindable meshes, this work achieves direct generation of complete, animatable 3D assets from 2D sketches.
Strong Correlations Induce Cause Only Predictions in Transformer Training
Haihan Zhang (Peking University), Cong Fang (Peking University)
Explainability and InterpretabilityRepresentation LearningTransformerSupervised Fine-TuningImageText
🎯 What it does: During the training of Transformers, the Correlation Crowding-Out (CCO) phenomenon was identified and systematically analyzed, where gradient descent progressively suppresses pseudo-causal information when strong causal features are present and their association with the target significantly outperforms all pseudo-causal features, leading the model to ultimately rely almost exclusively on causal features for prediction.
Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs
Yujie Zhao (University of California, San Diego), Jishen Zhao (University of California, San Diego)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose the AT-GRPO algorithm and training system, combining multi-agent systems (MAS) with reinforcement learning (RL) to enhance the collaborative and reasoning capabilities of large language models (LLMs) in tasks such as code, mathematics, planning, and games.
Strongly Convex Sets in Riemannian Manifolds
Damien Scieur (Samsung AI Lab), Sebastian Pokutta (Zuse Institute Berlin)
Optimization
🎯 What it does: Proposed a definition of strong convexity for sets on Riemannian manifolds, investigated their relationships, and proved that the Riemannian Frank-Wolfe algorithm achieves linear convergence on sets satisfying the Riemannian scaling inequality (RSI) or its approximate form.