ICLR 2026 Papers — Page 41
International Conference on Learning Representations · 5356 papers
Robust Preference Alignment via Directional Neighborhood Consensus
Ruochen Mao (Hong Kong University of Science and Technology (Guangzhou)), Jiaheng Wei (Hong Kong University of Science and Technology (Guangzhou))
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed a Robust Preference Selection (RPS) method in the post-inference phase, leveraging directional neighborhood consistency to sample multiple candidate responses near the user preference vector and select the optimal one to enhance robustness in preference alignment.
Robust Reward Modeling via Causal Rubrics
Pragya Srivastava (Google DeepMind), Doina Precup (Google DeepMind)
Adversarial AttackReinforcement Learning from Human FeedbackLarge Language ModelContrastive LearningText
🎯 What it does: Proposes the CROME framework, which obtains causal evaluation metrics by querying an Oracle LLM and trains a robust reward model using two synthetic adversarial samples (causal augmentation and neutral augmentation) to reduce reward hijacking risks.
Robust Selective Activation with Randomized Temporal K-Winner-Take-All in Spiking Neural Networks for Continual Learning
Jiangrong Shen (Xi'an Jiaotong University), Badong Chen (Xi'an Jiaotong University)
ClassificationSpiking Neural NetworkImage
🎯 What it does: Proposed a randomized temporal winner-takes-all (RTK-WTA) mechanism for spiking neural networks (SNNs) to enhance robustness in continuous learning and reduce catastrophic forgetting.
Robust Spiking Neural Networks Against Adversarial Attacks
Shuai Wang, Haizhou Li (Shenzhen Loop Area Institute)
ClassificationAdversarial AttackSpiking Neural NetworkImage
🎯 What it does: To address the adversarial robustness of directly trained spiking neural networks, the Threshold Protection Optimization (TGO) method is proposed.
Robust Test-time Video-Text Retrieval: Benchmarking and Adapting for Query Shifts
Bingqing Zhang (University of Queensland), Sen Wang (University of Queensland)
RetrievalDomain AdaptationVision Language ModelContrastive LearningVideoTextMultimodalityBenchmark
🎯 What it does: Construct a multi-level video perturbation benchmark MLVP, investigate the robustness of video-text retrieval under query shift, and propose the HAT-VTR online adaptation framework.
Robust Training of Neural Networks at Arbitrary Precision and Sparsity
Chengxi Ye (Google DeepMind), Andrew G. Howard (Google DeepMind)
OptimizationComputational EfficiencyImageText
🎯 What it does: Proposed a unified quantization and sparse training framework by modeling quantization error as additive noise and solving denoising dequantization transformation via ridge regression, addressing STE gradient blind spots and achieving stable training for low-precision (e.g., A1W1, FP4) and sparse networks;
Robustify Spiking Neural Networks via Dominant Singular Deflation under Heterogeneous Training Vulnerability
Desong Zhang (University of Exeter), Geyong Min (University of Exeter)
ClassificationAdversarial AttackSpiking Neural NetworkImageVideo
🎯 What it does: Propose a hyperparameter-free gradient biasing method called DSD, which suppresses the collapse of SNNs under heterogeneous training and enhances robustness.
Robustness in Text-Attributed Graph Learning: Insights, Trade-offs, and New Defenses
Runlin Lei (Renmin University of China), Chuntao Hong (Ant Group)
Representation LearningAdversarial AttackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: A unified text-attribute graph (TAG) robustness evaluation framework is constructed, conducting systematic structural and text attack experiments on ten cross-domain datasets, and proposing an adaptive defense model based on LLM called SFT-auto.
Robustness in the Face of Partial Identifiability in Reward Learning
Filippo Lazzati (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper proposes a unified and quantitative reward learning framework, and based on this framework, designs robust strategies to counteract uncertainty caused by partial identifiability, ultimately proposing the Rob-ReL algorithm;
Robustness of Probabilistic Models to Low-Quality Data: A Multi-Perspective Analysis
Liu Peng (Westlake University), Yaochu Jin (Westlake University)
Explainability and InterpretabilityData-Centric LearningConvolutional Neural NetworkTransformerDiffusion modelImageTextSequential
🎯 What it does: This paper explores the robustness of different probabilistic models (autoregressive language models, conditional diffusion models, image classifiers) under low-quality training data (random noise or label corruption) through systematic comparative experiments and multi-perspective theoretical analysis, and proposes two core principles to explain these differences.
RobustSpring: Benchmarking Robustness to Image Corruptions for Optical Flow, Scene Flow and Stereo
Victor Oei (University of Stuttgart), Andres Bruhn
Optical FlowVideoBenchmark
🎯 What it does: Propose the RobustSpring benchmark for evaluating the robustness of optical flow, scene flow, and stereo matching algorithms against image distortions.
ROC-n-reroll: How verifier imperfection affects test-time scaling
Florian E. Dorner (ETH Zurich), Fanny Yang (ETH Zurich)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper investigates the relationship between the performance and computational cost of two test-time scaling methods, rejection sampling (RS) and best-N (BoN), when using imperfect validators, through theoretical analysis and experimental validation, revealing that it depends on the geometric properties of the validator's ROC curve.
Rodrigues Network for Learning Robot Actions
Jialiang Zhang (MIT), Leonidas Guibas (Stanford University)
Robotic IntelligenceTransformerDiffusion model
🎯 What it does: Proposed a learnable Neural Rodrigues Operator and constructed the Rodrigues Network based on it to capture structured priors of joint motion, improving robot action learning.
ROGA: Scaling Generalist Agents for Office Productivity Tasks via Tool Generation
Mugeng Liu (Peking University), Yun Ma (Peking University)
TransformerLarge Language ModelAgentic AIWorld ModelTextTabularBenchmark
🎯 What it does: Proposes the ROGA framework to address the shortcomings of automated tool generation (ATG) agents in long-term, stateful office tasks, thereby improving task success rates.
Rolling Forcing: Autoregressive Long Video Diffusion in Real Time
Kunhao Liu (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
GenerationDiffusion modelVideo
🎯 What it does: Propose Rolling Forcing, an autoregressive diffusion model capable of real-time generation of long-duration videos.
RoRE: Rotary Ray Embedding for Generalised Multi-Modal Scene Understanding
Ryan Griffiths (University of Sydney), Donald G. Dansereau (University of Sydney)
TransformerNeural Radiance FieldGaussian SplattingMultimodality
🎯 What it does: Proposed Rotation-based Ray Embedding (RoRE), achieving unified camera geometry and multi-modal scene understanding
ROSETTA: Constructing Code-Based Reward from Unconstrained Language Preference
Sanjana Srivastava (Stanford University), Li Fei-Fei (Stanford University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the ROSETTA framework, which leverages foundation models to convert free, unconstrained natural language preferences into optimizable, interpretable code-based reward functions in a single step, enabling embodied agents to adapt online to dynamic human preferences;
Rote Learning Considered Useful: Generalizing over Memorized Data in LLMs
Qinyuan Wu (Max Planck Institute for Software Systems), Muhammad Bilal Zafar (Ruhr University Bochum)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Designed a two-phase 'rote learning-then-generalize' framework, first enabling LLMs to memorize fact triplets using semantically meaningless key tokens, then fine-tuning with extremely few semantic prompts to allow the model to generalize across various prompts, languages, and reasoning tasks
RouterArena: An Open Platform for Comprehensive Comparison of LLM Routers
Yifan Lu (Rice University), Jiarong Xing (Rice University)
OptimizationComputational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: Built a public LLM router evaluation platform called ROUTERARENA, providing a unified dataset, metrics, and automated evaluation framework, and released the first router leaderboard.
Routing Channel-Patch Dependencies in Time Series Forecasting with Graph Spectral Decomposition
Dongyuan Li (University of Tokyo), Renhe Jiang (University of Tokyo)
Graph Neural NetworkMixture of ExpertsTime SeriesBenchmark
🎯 What it does: Propose the xCPD plugin, which leverages graph Fourier decomposition to model frequency domain dependencies in the channel-patch layer, enhancing the performance of multivariate time series prediction.
Routing Manifold Alignment Improves Generalization of Mixture-of-Experts LLMs
Zhongyang Li (Johns Hopkins University), Tianyi Zhou (MBZUAI)
Computational EfficiencyRepresentation LearningLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose RoMA, a lightweight router post-training method for sparse Mixture-of-Experts large language models, leveraging manifold regularization to align routing weights with the task embedding space, thereby enhancing the model's generalization performance across multiple downstream tasks.
Routing Matters in MoE: Scaling Diffusion Transformers with Explicit Routing Guidance
Yujie Wei (Fudan University), Hongming Shan (Fudan University)
GenerationComputational EfficiencyTransformerMixture of ExpertsDiffusion modelRectified FlowAuto EncoderContrastive LearningImageText
🎯 What it does: While expanding the Diffusion Transformer, the Mixture-of-Experts framework ProMoE is introduced, along with a two-step router and contrastive learning to enhance expert specialization.
Routing, Cascades, and User Choice for LLMs
Rafid Mahmood (University of Ottawa)
OptimizationLarge Language Model
🎯 What it does: This paper constructs a Stackelberg game with a single supplier and a single user to study the impact of LLM routing, cascading, and user abandonment behavior on cost and user utility;
ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation
Yongyuan Liang (University of Maryland College Park), Furong Huang (University of Maryland College Park)
GenerationVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the ROVER benchmark to evaluate the capabilities of unified multimodal models in complementary cross-modal reasoning (language-guided image generation and image-assisted text reasoning), and constructs a self-annotated dataset containing over 1,300 tasks and nearly 2,000 images.
RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation
Jane Luo (Microsoft), Mao Yang (Microsoft)
AI Code AssistantTransformerLarge Language ModelTextGraphBenchmark
🎯 What it does: Designed and implemented the Repository Planning Graph (RPG) and ZeroRepo framework for generating complete software repositories from high-level requirements;
RPM: Reasoning-Level Personalization for Black-Box Large Language Models
Jieyong Kim (Yonsei University), Dongha Lee (Yonsei University)
Recommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented the RPM framework for reasoning-level personalization in black-box large language models, automatically extracting features from user history, clustering them into factors, constructing personalized reasoning paths, and retrieving matching examples through feature-based retrieval during inference;
RRNCO: Towards Real-World Routing with Neural Combinatorial Optimization
Jiwoo Son (Omelet), Jinkyoo Park
OptimizationTransformerReinforcement LearningGraphTabularBenchmark
🎯 What it does: Propose a new neural combinatorial optimization framework, RRNCO, specifically designed to address real-world complexities in vehicle routing problems, such as asymmetric distances, durations, and directional angles, and construct a VRP dataset containing asynchronous distance and duration matrices from 100 real cities.
Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains
Anisha Gunjal (Scale AI), Sean M. Hendryx
TransformerLarge Language ModelReinforcement LearningTextBiomedical DataBenchmark
🎯 What it does: Proposed and implemented the 'Rubrics as Rewards (RaR)' framework, utilizing instantiated multi-dimensional rubrics as reward signals for on-policy reinforcement learning in medical and scientific reasoning tasks;
RuleReasoner: Reinforced Rule-based Reasoning via Domain-aware Dynamic Sampling
Yang Liu (BIGAI), Zilong Zheng (BIGAI)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the RULEREASONER framework, significantly enhancing multi-domain rule reasoning performance through reinforcement learning and domain-aware dynamic sampling.
S2GO: Streaming Sparse Gaussian Occupancy
Jinhyung Park (Applied Intuition), Wei Zhan (Applied Intuition)
Autonomous DrivingTransformerSupervised Fine-TuningGaussian SplattingVideo
🎯 What it does: Propose a real-time semantic occupancy estimation framework S2GO based on sparse 3D queries, which can efficiently propagate context in video streams and achieve high-precision 3D occupancy maps through Gaussian decoding.
S2R-HDR: A Large-Scale Rendered Dataset for HDR Fusion
Yujin Wang (Shanghai AI Laboratory), Tianfan Xue (CUHK MMLab)
RestorationDomain AdaptationConvolutional Neural NetworkTransformerDiffusion modelImageBenchmark
🎯 What it does: Proposed a synthetic dataset named S2R-HDR containing 24,000 HDR images, and designed a dual-branch (shared + transfer) adaptation framework based on Adapter, named S2R-Adapter, significantly improving the performance of HDR fusion models in real-world scenarios.
S3OD: Towards Generalizable Salient Object Detection with Synthetic Data
Orest Kupyn (University of Oxford), Christian Rupprecht (University of Oxford)
Object DetectionData SynthesisTransformerVision Language ModelDiffusion modelImage
🎯 What it does: Constructed a large-scale synthetic dataset S3OD, proposed a multi-modal diffusion generation pipeline and iterative sampling strategy, and designed a multi-mask decoder to enhance cross-domain generalization for salient object detection.
SABRE-FL: Selective and Accurate Backdoor Rejection for Federated Prompt Learning
Momin Ahmad Khan (University of Massachusetts Amherst), Fatima M. Anwar (University of Massachusetts Amherst)
Anomaly DetectionFederated LearningSafty and PrivacyAdversarial AttackTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: Studied backdoor attacks in federated prompt learning and proposed a lightweight defense method called SABRE-FL, which can filter poisoned prompt updates without accessing original data and labels.
SAC Flow: Sample-Efficient Reinforcement Learning of Flow-Based Policies via Velocity-Reparameterized Sequential Modeling
Yixian Zhang (Tsinghua University), Wenbo Ding (Tsinghua University)
Recurrent Neural NetworkTransformerReinforcement LearningFlow-based ModelBenchmark
🎯 What it does: Designed and implemented an offline-online reinforcement learning framework named SAC Flow, which can directly end-to-end train streaming policies and eliminate gradient explosion by viewing streaming policies as RNNs and reparameterizing them as GRU/Transformer.
Saddle-to-Saddle Dynamics Explains A Simplicity Bias Across Neural Network Architectures
Yedi Zhang (University College London), Peter E. Latham (University College London)
Explainability and InterpretabilityConvolutional Neural NetworkTransformer
🎯 What it does: This paper proposes a unified theoretical framework to explain the relationship between the 'simplicity bias' and saddle-to-saddle learning dynamics observed in various neural network architectures (linear, ReLU, convolutional, self-attention).
Saddle-To-Saddle Dynamics in Deep ReLU Networks: Low-Rank Bias in the First Saddle Escape
Ioannis Bantzis (EPFL), Arthur Jacot (New York University)
OptimizationImage
🎯 What it does: This paper studies the dynamics of gradient descent in deep ReLU networks escaping from the origin saddle point with small weight initialization, revealing that the escape direction in deeper layers exhibits low-rank bias.
SAE as a Crystal Ball: Interpretable Features Predict Cross-domain Transferability of LLMs without Training
Qi Zhang (Peking University), Yisen Wang (Peking University)
Domain AdaptationExplainability and InterpretabilityRepresentation LearningLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes a method using sparse autoencoders (SAE) to predict the cross-domain transfer effectiveness of large models during post-training (e.g., supervised fine-tuning);
SAES-SVD: Self-Adaptive Suppression of Accumulated and Local Errors for SVD-based LLM Compression
Xing Hu (Houmo AI), Zukang Xu (Houmo AI)
CompressionLarge Language ModelTextBenchmark
🎯 What it does: Designed and implemented an adaptive error suppression low-rank SVD compression framework (SAES-SVD) for compressing LLMs without fine-tuning.
SAFA-SNN: Sparsity-Aware On-Device Few-Shot Class-Incremental Learning with Fast-Adaptive Structure of Spiking Neural Network
Huijing Zhang (Zhejiang University), Shuiguang Deng (Zhejiang University)
Computational EfficiencyMeta LearningSpiking Neural NetworkImage
🎯 What it does: Propose a SAFA-SNN based on spiking neural networks for few-shot class incremental learning on edge devices.
Safe Continuous-time Multi-Agent Reinforcement Learning via Epigraph Form
Xuefeng Wang (Purdue University), Ahmed H Qureshi (Purdue University)
OptimizationReinforcement LearningStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a continuous-time multi-agent reinforcement learning framework that maps safety constraints through epigraph deformation into continuous values, utilizing physics-informed neural networks (PINNs) for policy learning.
Safe Exploration via Policy Priors
Manuel Wendl (ETH Zurich), Andreas Krause (ETH Zurich)
Autonomous DrivingOptimizationSafty and PrivacyRobotic IntelligenceReinforcement LearningWorld ModelBenchmark
🎯 What it does: Propose the SOOPER algorithm, which utilizes a conservative prior policy and a probabilistic world model to achieve safe exploration, ensuring constraint satisfaction during the learning process and converging to an optimal policy.
SafeDialBench: A Fine-Grained Safety Evaluation Benchmark for Large Language Models in Multi-Turn Dialogues with Diverse Jailbreak Attacks
Hongye Cao (National Key Laboratory for Novel Software Technology, Nanjing University), Junlan Feng (China Mobile Research Institute China Mobile (Suzhou) Software Technology Co., Ltd.)
Safty and PrivacyAdversarial AttackLarge Language ModelTextBenchmark
🎯 What it does: Built a multi-turn dialogue safety evaluation benchmark called SafeDialBench, covering two-layer safety classification, over 4,000 bilingual (Chinese-English) multi-turn dialogues, 22 diverse scenarios, and seven jailbreak attack methods to assess the safety performance of large language models (LLMs).
SafeDPO: A Simple Approach to Direct Preference Optimization with Enhanced Safety
Geon-Hyeong Kim (LG AI Research), Moontae Lee (LG AI Research)
Reinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: Propose SafeDPO, a single-stage training method that directly aligns with safety using preference data, eliminating the need for reward models and cost models.
SafeFlowMatcher: Safe and Fast Planning using Flow Matching with Control Barrier Functions
Jeongyong Yang (Korea Advanced Institute of Science and Technology), SooJean Han (Korea Advanced Institute of Science and Technology)
OptimizationRobotic IntelligenceFlow-based ModelSequentialOrdinary Differential Equation
🎯 What it does: Propose SafeFlowMatcher, which combines flow matching and control barrier functions to achieve real-time safe planning
Safeguarding Multimodal Knowledge Copyright in the RAG-as-a-Service Environment
Tianyu Chen (ShanghaiTech University), Wenjie Wang (ShanghaiTech University)
RetrievalSafty and PrivacyTransformerVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes the AQUA framework for copyright watermarking of image knowledge in multimodal RAG systems, supporting both black-box and white-box usage.
SafeMoE: Safe Fine-Tuning for MoE LLMs by Aligning Harmful Input Routing
Jaehan Kim (KAIST), Sooel Son (KAIST)
Safty and PrivacyLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: This paper proposes SAFEMOE, a safe fine-tuning method for Mixture-of-Experts (MoE) large language models, aimed at preventing safety degradation caused by routing drift during fine-tuning.
SafeMPO: Constrained Reinforcement Learning with Probabilistic Incremental Improvement
Alexander Mattick, Christopher Mutschler (University of Technology Nuremberg)
OptimizationReinforcement LearningBenchmark
🎯 What it does: Propose a safety reinforcement learning framework called SafeMPO based on relative improvement constraints, achieving incremental convergence to the feasible set through the logarithmic barrier of a non-parametric backup solution, and achieving high reward and low cost in safety-constrained environments.
SAFER: Risk-Constrained Sample-then-Filter in Large Language Models
Qingni Wang (University of California Santa Cruz), Xin Eric Wang (University Of California Santa Barbara)
Safty and PrivacyLarge Language ModelText
🎯 What it does: Proposed a two-stage risk control framework called SAFER, aiming to ensure the coverage of acceptable answers in open-ended question answering tasks.
Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance
Jiawen Zhang (Zhejiang University), Ruoxi Jia (Virginia Tech)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a method to restore safety alignment in large language models compromised by harmful fine-tuning using only a single safe instance;
Safety Instincts: LLMs Learn to Trust Their Internal Compass for Self-Defense
Guobin Shen (Beijing Institute of AI Safety and Governance), Yi Zeng (Beijing Institute of AI Safety and Governance)
Safty and PrivacyLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose Safety Instincts Reinforcement Learning (SIRL), leveraging model internal response entropy as a reward signal to enhance defense against jailbreak attacks without requiring external labels or human feedback.
Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-Tuning and Can Be Mitigated by Machine Unlearning
Yiwei Chen (Michigan State University), Sijia Liu (Michigan State University)
Safty and PrivacySupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: This paper investigates the 'safety hallucination' phenomenon in vision-language models (VLMs) during safe fine-tuning, where models form pseudo-correlations with text surface features through supervised learning, leading to vulnerability to word substitution attacks and excessive caution. Subsequently, the paper proposes using machine unlearning (NPO, RMU) methods to eliminate these pseudo-correlations without labeled data, enhancing model security.
Safety Subspaces are Not Linearly Distinct: A Fine-Tuning Case Study
Kaustubh Ponkshe (Mohamed bin Zayed University of Artificial Intelligence), Praneeth Vepakomma (Mohamed bin Zayed University of Artificial Intelligence)
Safty and PrivacyTransformerSupervised Fine-TuningText
🎯 What it does: Empirical investigation on the safety alignment of large language models, examining whether safety behaviors can be linearly separated in weight or activation spaces.
SAFETY-GUIDED FLOW (SGF): A UNIFIED FRAMEWORK FOR NEGATIVE GUIDANCE IN SAFE GENERATION
Mingyu Kim (Kookmin University), Mijung Park (University of British Columbia)
GenerationDiffusion modelFlow-based ModelImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes Safety-Guided Flow (SGF), a negative guidance framework based on MMD potential, achieving safe generation during the sampling process of diffusion and flow models. It demonstrates through control barrier functions that negative guidance is most critical in the early decoding stage and should gradually diminish later.
SAGA: Structural Aggregation Guided Alignment with Dynamic View and Neighborhood Order Selection for Multiview Graph Domain Adaptation
Ruiyi Fang (Western University), Boyu Wang (Western University)
Domain AdaptationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper addresses the multi-view graph domain adaptation (MGDA) problem, proposing the SAGA framework, which achieves cross-domain node classification when the source graph has labels and the target graph lacks labels.
SAGE: Spatial-visual Adaptive Graph Exploration for Efficient Visual Place Recognition
Shunpeng Chen (Beijing University of Posts and Telecommunications), Shibiao Xu (Beijing University of Posts and Telecommunications)
RetrievalGraph Neural NetworkTransformerContrastive LearningImageGraphBenchmark
🎯 What it does: SAGE proposes a unified VPR training framework that leverages a frozen DINOv2 backbone and lightweight modules to construct an online-updatable geo-visual graph, and dynamically selects the most challenging samples for training through greedy weighted sampling.
SAIL: Self-Amplified Iterative Learning for Diffusion Model Alignment with Minimal Human Feedback
Xiaoxuan He (ZheJiang University), Bo Zhang (ZheJiang University)
GenerationReinforcement Learning from Human FeedbackDiffusion modelImage
🎯 What it does: Proposes the SAIL framework, achieving self-iterative alignment for diffusion models through self-generation and self-evaluation using minimal human preference data.
SAIR: Enabling Deep Learning for Protein-Ligand Interactions with a Synthetic Structural Dataset
Pablo Lemos, Martin Ganahl
Drug DiscoveryProtein Structure PredictionConvolutional Neural NetworkGraph Neural NetworkGraphBiomedical DataBenchmark
🎯 What it does: Constructed the SAIR dataset by collecting and folding over 10 million protein-ligand complex 3D structures paired with experimental IC50 activities; performed PoseBusters quality assessment on generated structures and compared the performance of multiple affinity prediction models;
Salient Object Ranking via Cyclical Perception-Viewing Interaction Modeling
Rongjin Guo, Rynson W. H. Lau
Object DetectionSegmentationTransformerVision Language ModelVision-Language-Action ModelImageMultimodality
🎯 What it does: Proposes a framework based on cyclical perception-view interaction for object ranking in images, incorporating a Story Prediction (SP) module and a Guided Ranking (GR) module to predict the priority of salient objects;
SAM 3: Segment Anything with Concepts
Nicolas Carion (Meta Superintelligence Labs), Christoph Feichtenhofer (Meta Superintelligence Labs)
Object DetectionObject TrackingSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: Proposes SAM 3, a unified model that supports concept-level detection, segmentation, and tracking using short noun phrases or image examples, and enables interactive refinement in both images and videos.
SAM-Veteran: An MLLM-Based Human-like SAM Agent for Reasoning Segmentation
Tianyuan Du (ByteDance), Yang Zhang (ByteDance)
SegmentationLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose SAM-Veteran, a SAM interactive inference segmentation agent based on a multimodal large language model (MLLM), capable of completing the full workflow including target bounding box generation, mask iterative refinement, and adaptive termination.
Same Content, Different Representations: A Controlled Study for Table QA
Yue Zhang (University of Texas at Dallas), Nikita Bhutani (Megagon Labs)
Data SynthesisRepresentation LearningTransformerLarge Language ModelTabularBenchmark
🎯 What it does: Conduct control experiments on table QA, compare the effects of structured and semi-structured tables, and propose the REPAIRTQA diagnostic benchmark.
Sample Complexity and Representation Ability of Test-time Scaling Paradigms
Baihe Huang (University of California Berkeley), Jiantao Jiao (University of California Berkeley)
Representation LearningTransformerReinforcement LearningText
🎯 What it does: This paper investigates the sample complexity and representational capacity of the scaling paradigm during testing, proving that the sample complexity of self-consistency and optimal n-sampling can be separated, and proposes a generic Transformer architecture that achieves online learning through self-correction, enabling a single model to adaptively handle multi-task inference;
Sample Efficient Offline RL via T-Symmetry Enforced Latent State-Stitching
Peng Cheng (Beijing Jiaotong University), Xianyuan Zhan (Beijing Jiaotong University)
OptimizationReinforcement LearningAuto EncoderTabularTime SeriesOrdinary Differential Equation
🎯 What it does: Propose a sample-efficient offline reinforcement learning algorithm called TELS, which achieves state concatenation in a potential space that satisfies time-reversal symmetry (T-symmetry);
Sample Lottery: Unsupervised Discovery of Critical Instances for LLM Reasoning
Zhiping Xiao (University of Washington), Ming Zhang (Peking University)
OptimizationData-Centric LearningTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes an unsupervised framework named CONST for identifying critical samples in RLVR training, leveraging program volatility and result volatility combined with conformal prediction to select a small amount of high-value data for annotation and optimization.
Sample More to Think Less: Group Filtered Policy Optimization for Concise Reasoning
Vaishnavi Shrivastava (Microsoft Research), Dimitris Papailiopoulos (Microsoft Research)
OptimizationReinforcement LearningText
🎯 What it does: Proposed and implemented Group Filtered Policy Optimization (GFPO), which, within the reinforcement learning framework, significantly compresses the length of the generated chain by sampling more candidate answers for each problem and only updating the top k answers that meet metrics such as length or token-efficiency, while maintaining or improving reasoning accuracy;
Sample Reward Soups: Query-efficient Multi-Reward Guidance for Text-to-Image Diffusion Models
Yinghua Yao (Agency for Science, Technology and Research), Ivor Tsang (Agency for Science, Technology and Research)
GenerationDiffusion modelImageText
🎯 What it does: Proposed a no-training multi-reward alignment method called Sample Reward Soups (SRSoup), achieving Pareto optimal sampling during the inference phase of diffusion models through gradient interpolation.
Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs
Xueyan Li (ETH Zurich), Jonas Geiping (ETH Zurich)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated shifting sampling strategies from exploration to 'correctness-priority' in reasoning tasks, proposing rank-wise calibrated truncation sampling methods such as Greedy-Threshold, Calibrated-TopK, and Calibratedε.
Sample-efficient and Scalable Exploration in Continuous-Time RL
Klemens Iten (ETH Zürich), Andreas Krause (ETH Zürich)
OptimizationReinforcement LearningBenchmark
🎯 What it does: Proposed and implemented COMBRL—a model-based continuous-time reinforcement learning algorithm that leverages uncertainty-driven exploration;
Sample-Efficient Distributionally Robust Multi-Agent Reinforcement Learning via Online Interaction
Zain Ulabedeen Farhat (University of Central Florida), Yue Wang (University of Central Florida)
Reinforcement Learning
🎯 What it does: This paper studies algorithms for online learning in distributionally robust Markov games (DRMGs), proposing the multi-player optimistic robust Nash value iteration (MORNAVI) algorithm, which aims to enhance the robustness of multi-agent systems through direct interaction with the environment.
Sample-efficient evidence estimation of score based priors for model selection
Frederic Wang (Caltech), Katherine Bouman (Caltech)
OptimizationComputational EfficiencyDiffusion modelScore-based ModelImagePhysics Related
🎯 What it does: Propose a method called DiME, which leverages the posterior time marginal of diffusion models to estimate model evidence, thereby enabling model selection and validation for diffusion priors.
Samples Are Not Equal: A Sample Selection Approach for Deep Clustering
Zhengxing Jiao (Southeast University), Junhui Hou (City University of Hong Kong)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Designed a plugin that enhances the performance and efficiency of deep clustering through density-aware clustering head initialization and dynamic sample screening.
Sampling Complexity of TD and PPO in RKHS
LU ZOU, Shuang Li (Chinese University of Hong Kong)
Computational EfficiencyReinforcement LearningSequential
🎯 What it does: This paper proposes kernelized temporal difference (TD) estimators and KL-regularized natural gradient (NPG) policy improvement steps within the reproducing kernel Hilbert space (RKHS) framework, providing sampling complexity and non-asymptotic convergence rates for any RKHS (e.g., Sobolev, Gaussian, NTK, etc.).
Sampling-aware Adversarial Attacks Against Large Language Models
Tim Beyer (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
Computational EfficiencyAdversarial AttackTransformerText
🎯 What it does: Proposed and verified a sampling-aware adversarial attack framework for large language models, treating sampling as a core component of the attack resource, significantly improving attack success rates within limited computational budgets.
SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer
Junsong Chen (The University Of Hong Kong), Enze Xie
GenerationData SynthesisTransformerDiffusion modelAuto EncoderVideoTextMultimodality
🎯 What it does: Built a small diffusion model called SANA-Video, which can efficiently generate 720p 5-second videos and minute-level long videos, and supports running on edge devices such as RTX 5090.
Sapiens2
Rawal Khirodkar (Meta Reality Labs), Shunsuke Saito (Meta Reality Labs)
SegmentationPose EstimationSuper ResolutionTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Developed SAPIENS2, a series of high-resolution vision Transformers focused on human-centric visual tasks, with resolutions ranging from 1K to 4K, pre-trained and fine-tuned on a variety of dense prediction tasks using 1B portrait images.
SAQ: Stabilizer-Aware Quantum Error Correction Decoder
David Zenati (Ben-Gurion University of Negev), Eliya Nachmani (Ben-Gurion University of Negev)
TransformerPhysics Related
🎯 What it does: Proposed SAQ-Decoder, a transformer-based quantum error correction decoder with a dual-stream architecture combined with constraint projection post-processing, achieving near ML accuracy and linear scalability.
SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
Qianzhong Chen (Stanford University), Philipp Wu (xdof.ai)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision-Language-Action ModelVideoText
🎯 What it does: Propose a stage-aware reward model (SARM) and reward-aligned behavioral cloning (RA-BC) to enhance the performance of long-term robotic manipulation.
SASFT: Sparse Autoencoder-guided Supervised Finetuning to Mitigate Unexpected Code-Switching in LLMs
Boyi Deng (Tongyi Lab, Alibaba Group Inc), Fuli Feng (Anhui Provincial Hospital)
Representation LearningLarge Language ModelSupervised Fine-TuningAuto EncoderText
🎯 What it does: This paper addresses the issue of unexpected code switching in multilingual large models by proposing a supervised fine-tuning method called SASFT based on sparse autoencoders (SAE), which guides the model to suppress the activation values of irrelevant language features during training, thus significantly reducing the proportion of code switching.
Sat3DGen: Comprehensive Street-Level 3D Scene Generation from Single Satellite Image
Ming Qian (Wuhan University), Gui-Song Xia (Wuhan University)
GenerationData SynthesisNeural Radiance FieldImage
🎯 What it does: Generate street-level high-quality 3D scenes from a single satellite image, achieving a renderable complete urban space;
SatDreamer360: Multiview-Consistent Generation of Ground-Level Scenes from Satellite Imagery
Xianghui Ze (Nanjing University of Science and Technology), Yujiao Shi (ShanghaiTech University)
GenerationDiffusion modelImageVideoBenchmark
🎯 What it does: Propose the SatDreamer360 framework, which achieves the generation of continuous multi-view ground panoramas from a single satellite image and a predefined trajectory.
SAVE: A Generalizable Framework for Multi-Condition Single-Cell Generation with Gene Block Attention
Jiahao Li (Fudan University), Fei Wang (Fudan University)
GenerationData SynthesisTransformerLarge Language ModelFlow-based ModelAuto EncoderBiomedical Data
🎯 What it does: Propose a unified single-cell RNA-seq conditional generation framework SAVE, integrating Transformer gene block attention, VAE variational inference, and Flow Matching, which can achieve controllable generation of cell expression, batch correction, and perturbation prediction under various experimental conditions;
SC-Arena: A Natural Language Benchmark for Single-Cell Reasoning with Knowledge-Augmented Evaluation
Jiahao Zhao (Northeastern University), Min Yang (Northeastern University)
GenerationLarge Language ModelBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the SC-ARENA benchmark, which designs five open-ended natural language tasks based on virtual cell abstraction (cell type annotation, cell description, cell generation, perturbation prediction, scientific question answering), and develops a knowledge-enhanced LLM-as-a-judge evaluation framework, providing interpretable, biology-knowledge-based scores and reasoning.
SCAD: Super-Class-Aware Debiasing for Long-Tailed Semi-Supervised Learning
Sunguk Jang (AITRICS), Byung-Jun Lee
ClassificationRepresentation LearningData-Centric LearningConvolutional Neural NetworkVision Language ModelImage
🎯 What it does: Proposes the Super-Class-Aware Debiasing (SCAD) framework to address the imbalance issue within the same super-class in long-tailed semi-supervised learning;
Scaf-GRPO: Scaffolded Group Relative Policy Optimization for Enhancing LLM Reasoning
Xichen Zhang (Hong Kong University of Science and Technology), Jiaya Jia (Hong Kong University of Science and Technology)
Reinforcement LearningPrompt EngineeringText
🎯 What it does: Introduce hierarchical scaffolding within prompts in RLVR training, progressively providing minimal prompts to overcome the model's learning cliff;
Scalable and Adaptive Trust-Region Learning via Projection Convex Hull
Hongyang Jia (Tsinghua University), Chongqing Kang (Tsinghua University)
OptimizationExplainability and InterpretabilityTabularBiomedical Data
🎯 What it does: This paper proposes the Projection Convex Hull (PCH) framework for learning compact, interpretable polyhedral trust regions from labeled data.
Scalable Chain of Thoughts via Elastic Reasoning
Yuhui Xu (Salesforce Ai Research), Caiming Xiong (Salesforce Ai Research)
Computational EfficiencyReinforcement LearningTextSequentialChain-of-Thought
🎯 What it does: Propose the Elastic Reasoning framework, which splits the reasoning process into a thinking phase and an answering phase, and employs budget constraints for truncation during inference.
Scalable Energy-Based Models via Adversarial Training: Unifying Discrimination and Generation
Xuwang Yin (Independent), Tony T. Wang (Massachusetts Institute of Technology)
ClassificationGenerationConvolutional Neural NetworkImage
🎯 What it does: Propose Dual Adversarial Training (DAT), replacing SGLD learning with PGD+BCE to achieve robust classification and high-quality generation simultaneously within a single framework.
Scalable Exploration for High-Dimensional Continuous Control via Value-Guided Flow
Yunyue Wei (Tsinghua University), Yanan Sui (Tsinghua University)
Reinforcement LearningFlow-based ModelBenchmarkOrdinary Differential Equation
🎯 What it does: Proposed a scalable reinforcement learning method called QFLEX, which performs targeted exploration in high-dimensional continuous control spaces through value-guided probabilistic flows.
Scalable In-Context Q-Learning
Jinmei Liu (Nanjing University), Zhi Wang (Nanjing University)
TransformerReinforcement LearningWorld Model
🎯 What it does: Proposed S-ICQL, a scalable context learning framework for offline multi-task reinforcement learning that combines dynamic programming with world models, using a multi-head Transformer to simultaneously predict policies and values;
Scalable Multi-Task Low-Rank Model Adaptation
Zichen Tian (Singapore Management University), Qianru Sun (Singapore Management University)
ClassificationRecognitionComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerImageTextBenchmark
🎯 What it does: Propose a scalable multi-task low-rank adaptation method called mtLoRA, which maintains or even improves model performance when the number of tasks increases significantly.
Scalable Multilingual Multimodal Machine Translation with Speech-Text Fusion
Yexing Du (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
GenerationTransformerLarge Language ModelTextMultimodalityAudio
🎯 What it does: Propose a Speech-guided Machine Translation (SMT) framework that integrates text and synthesized speech, leveraging multimodal large language models (MLLM) to generate higher quality translations, and introduces a self-evolution mechanism to achieve unsupervised continuous improvement.
Scalable Offline Model-Based RL with Action Chunks
Kwanyoung Park (University Of California Berkeley), Sergey Levine (University Of California Berkeley)
Reinforcement LearningFlow-based ModelWorld ModelBenchmark
🎯 What it does: Proposes an offline model-based reinforcement learning framework MAC that combines action block models and strategies, enabling model rollouts of over 100 steps in long-horizon tasks.
Scalable Random Wavelet Features: Efficient Non-Stationary Kernel Approximation with Convergence Guarantees
Sawan Kumar (Indian Institute of Technology Delhi), Souvik Chakraborty (Indian Institute of Technology Delhi)
TabularBiomedical DataBenchmarkAudio
🎯 What it does: Propose the Random Wavelet Features (RWF) framework to efficiently approximate non-stationary kernels and achieve scalable Gaussian process regression.
Scalable Second-order Riemannian Optimization for $K$-means Clustering
Peng Xu (University of Illinois Urbana Champaign), Richard Y. Zhang (University of Illinois Urbana Champaign)
OptimizationImageTabular
🎯 What it does: Reformulate the K-means clustering problem as a smooth unconstrained optimization on a submanifold, and design a linear-time second-order Riemannian cubic regularized Newton algorithm based on this reformulation;
Scalable Spatio-Temporal SE(3) Diffusion for Long-Horizon Protein Dynamics
Nima Shoghi (ByteDance Seed), Quanquan Gu (ByteDance Seed)
Drug DiscoveryTransformerDiffusion modelSequentialBiomedical Data
🎯 What it does: Proposed a scalable SE(3)-equivariant autoregressive diffusion model called STAR-MD for generating long-term (microsecond-level) protein dynamics trajectories
Scalable Training for Vector-Quantized Networks with 100% Codebook Utilization
Yifan Chang, Xingang Wang
GenerationTransformerAuto EncoderImage
🎯 What it does: Proposed the VQBridge projector and FVQ training framework, achieving 100% codebook utilization in vector quantization networks and improving reconstruction and generation quality.
Scale-wise Distillation of Diffusion Models
Nikita Starodubcev (Yandex Research), Dmitry Baranchuk (Yandex Research)
GenerationKnowledge DistillationDiffusion modelAuto EncoderImageVideo
🎯 What it does: Propose a scale-level diffusion distillation framework called SwD, which can compress pre-trained diffusion models into a few-step generator and achieve high-quality generation through progressive resolution upscaling.
ScaleCap: Scalable Image Captioning via Dual-Modality Debiasing
Long Xing (Chinese University of Hong Kong), Dahua Lin (University of Science and Technology of China)
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a scalable image description method called ScaleCap, which continuously enriches and corrects image captions through iterative questioning and sentence scoring comparison.
ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data
Zhaoyang Liu (Hong Kong University of Science and Technology), Wenhai Wang (Hong Kong University of Science and Technology)
Data-Centric LearningAgentic AIVision Language ModelImageTextMultimodalitySequentialChain-of-Thought
🎯 What it does: Developed a cross-platform general-purpose computer usage proxy called ScaleCUA and built a dual-loop interactive data pipeline to collect large-scale GUI data across six major platforms (Windows, macOS, Linux, Android, iOS, Web);
ScaleLong: A Multi-Timescale Benchmark for Long Video Understanding
David Ma (M A P), Xiaojie Jin
Large Language ModelVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Proposed and released the ScaleLong benchmark for fine-grained evaluation of multi-timescale (Clip, Shot, Event, Story) understanding within the same long video content.