ICML 2026 Papers — Page 56
International Conference on Machine Learning · 6554 papers
STARE: Step-wise Temporal Alignment and Red-teaming Engine for Multi-modal Toxicity Attack
Xutao Mao (City University of Hong Kong), Cong Wang (City University of Hong Kong)
Explainability and InterpretabilityAdversarial AttackReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelRectified FlowImageTextMultimodality
🎯 What it does: Propose STARE, a hierarchical reinforcement learning framework for end-to-end red teaming visual-language models (VLMs) with toxic attacks, combining prompt editing and image generation refinement, and performing step-by-step temporal alignment analysis of the generation process.
StarEmbed: Benchmarking Time Series Foundation Models on Astronomical Observations of Variable Stars
Weijian Li (Northwestern University), Han Liu (Northwestern University)
ClassificationAnomaly DetectionTransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningTime SeriesBenchmarkPhysics Related
🎯 What it does: This paper constructs the first publicly available benchmark for periodic variable star light curves called StarEmbed, compiling approximately 40,000 ZTF light curves and dividing them into seven categories. Subsequently, the performance of various time series foundation models and traditional handcrafted feature baselines was evaluated on three tasks: unsupervised clustering, supervised classification, and OOD detection.
State Space Model with Continuous Limit of HiPPO Matrix: Eigenvalue Analysis and Explicit Solution Formula
Atsushi Takabatake (Kyushu University), Takaharu Yaguchi (Kyushu University)
OptimizationComputational EfficiencyRepresentation LearningReview/Survey PaperStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes to continuize the HiPPO matrix into an operator acting on a function space — the Continuized-HiPPO Operator. By utilizing this operator, the asymptotic properties of the regular part eigenvalues of HiPPO are proved, and an explicit solution for the corresponding linear state space layer (LSSL) is derived (expressed in the form of Bessel function integrals), thereby providing a theoretical foundation and directly usable convolution kernels for HiPPO-related models;
State-Dependent Safety Failures in Multi-Turn Language Model Interaction
Pengcheng Li (University of Science and Technology of China), Wenbo Zhou (University of Science and Technology of China)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study proposes the STAR framework, which diagnoses the state evolution in multi-turn dialogues, revealing the safety failure mechanisms of safety-aligned models during continuous interactions.
Stationary MMD Points
Zonghao Chen (University College London), Chris J. Oates (Newcastle University)
Data SynthesisOptimizationContrastive LearningPoint CloudTabularBenchmarkStochastic Differential Equation
🎯 What it does: Propose and study 'Stationary MMD Points,' which are sets of points that satisfy zero gradient under the maximum mean discrepancy (MMD) objective, and prove that their integration error for functions in RKHS converges faster than MMD itself (superconvergence); compute these point sets using MMD gradient flow with noise injection and provide non-asymptotic error bounds;
Statistical Consistency and Generalization of Contrastive Representation Learning
Yuanfan Li (University of Sydney), Yiming Ying (University of Sydney)
RetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: A unified statistical learning theory framework was constructed, proving the statistical consistency of contrastive learning (Contrastive Representation Learning, CRL) during the pre-training phase for retrieval tasks, and providing a calibrated inequality, further deriving the generalization upper bounds for both supervised and self-supervised contrastive learning, explaining the improvement brought by a large number of negative samples.
Statistical Early Stopping for Reasoning Models
Yangxinyu Xie (University of Pennsylvania), Edgar Dobriban (University of Pennsylvania)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes two statistical early stopping methods based on the temporal statistics of uncertainty keywords during the reasoning process of language models, which can dynamically stop excessive thinking during inference and reduce unnecessary reasoning steps;
Statistically Calibrated Scaling for Token Merging in Transformers
Qing Zhou (Northwestern Polytechnical University), Qi Wang (University of Hong Kong)
ClassificationRetrievalCompressionComputational EfficiencyTransformerContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a statistically calibrated Token merging scaling method (√r-scaling), used in Transformer inference to maintain balance between feature energy and attention distribution when compressing sequence length.
Statistically Undetectable Backdoors in Deep Neural Networks
Andrej Bogdanov (University of Ottawa), Neekon Vafa (Massachusetts Institute of Technology)
Safty and PrivacyAdversarial AttackConvolutional Neural NetworkDiffusion modelScore-based ModelAuto EncoderContrastive LearningGaussian SplattingImageStochastic Differential Equation
🎯 What it does: Proposes implanting a statistically undetectable backdoor in a deep feedforward neural network that satisfies specific structural constraints, allowing the model trainer to generate compatibility input pairs at an exponential level without being discovered by users.
STD-Former: Image-Conditioned Texture Dictionary Encoding with Sparse Topological Supervision for Texture Recognition
Bo Peng (Shanghai Ocean University), Yurui Pan (Fudan University)
RecognitionTransformerAuto EncoderContrastive LearningImageBenchmark
🎯 What it does: Proposed the STD-Former framework, which achieves texture dictionary encoding under image conditions by adaptively extracting 'Intrinsic Textons' from each input image, addressing the feature misalignment problem between traditional training set dictionaries and test images.
Steady-State Behavior of Constant-Stepsize Stochastic Approximation: Gaussian Approximation and Tail Bounds
Yuyang Wang (University of Michigan), Siva Theja Maguluri (Georgia Institute of Technology)
OptimizationStochastic Differential Equation
🎯 What it does: Non-asymptotic analysis of the steady-state distribution of fixed-step-size stochastic approximation (SGD, linear SA, and convergent nonlinear SA) is conducted, providing explicit error upper bounds in Wasserstein distance and tail distribution. Unified results are presented for both i.i.d. and Markovian noise cases; further, it is proven that under non-convex/general convex objectives, the steady-state scaling factor changes from sqrt(α) to α^{1/h}, and the Gibbs limit distribution is provided.
Steal the Patch Size: Adversarially Manipulate Vision Language Models
Kai Hu (Carnegie Mellon University), Matt Fredrikson (Carnegie Mellon University)
Safty and PrivacyAdversarial AttackTransformerPrompt EngineeringVision Language ModelGenerative Adversarial NetworkContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a black-box model stealing attack, which exploits the edge information loss caused by the patchification of visual Transformers, to infer the visual patch size, preprocessing type, and target resolution of deployed VLMs.
Steer Like the LLM: Activation Steering that Mimics Prompting
Geert Heyman (Nokia Bell Labs), Frederik Vandeputte (Nokia Bell Labs)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The study proposes a framework that treats prompt steering as an activation-based guidance approach, and approximates the effect of prompts by learning token-level activation coefficients;
Steer Where It Matters: Token-Level Visual-Sensitivity Steering for LVLMs Hallucination Mitigation
Ruipeng Zhang (South China University of Technology), Tong Zhang (South China University of Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose an activation-guided method based on token-level visual sensitivity to control hallucinations during inference in large vision-language models;
Steering at the Source: Style Modulation Heads for Robust Persona Control
Yoshihiro Izawa (University of Tokyo), Kenjiro Taura (University of Tokyo)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: For activation steering in large language models, this paper proposes localized interventions on specific attention heads (Style Modulation Heads) to achieve precise control over the model's personalized behavior, avoiding coherence collapse that occurs when adjustments are made on the residual stream.
Steering Beyond the Support: Adversarial Training on Unsupervised Jailbroken Activation Simulation
LUOYU CHEN, Shui Yu (University of Technology Sydney)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposes a zero-supervised safety-driven method that enhances the robustness of LLMs against unknown attacks by simulating unlocking activation in the representation space and guiding during testing using gradient fields.
Steering Large Language Models through the DMTA Cycle: Structure-Based Drug Design via Knowledge-Driven Bi-Level Thompson Sampling
Xuanning Hu (Jilin University), Bo Yang (Jilin University)
Drug DiscoveryTransformerLarge Language ModelReinforcement LearningDiffusion modelBiomedical DataRetrieval-Augmented Generation
🎯 What it does: By constructing a K-BTS framework based on dual-layer Thompson sampling, utilizing LLM to achieve closed-loop iterative optimization for structured drug design in DMTA
Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning
Helena Casademunt (Harvard University), Neel Nanda
Domain AdaptationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningText
🎯 What it does: Proposed and implemented a Concept Ablation Fine-tuning (CAFT) method, which uses interpretability techniques to identify and eliminate hidden directions in pre-trained language models that cause poor out-of-distribution (OOD) generalization, thus guiding the model to perform better on new distributions without modifying the training data.
SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives
Vincent Siu (University of California Santa Cruz), Chenguang Wang (University of California Santa Cruz)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Designed and implemented the STEERINGSAFETY benchmark to systematically evaluate the effectiveness of representation steering in large language models (LLMs) across nine safety perspectives and their cross-perspective entanglement.
Stein Diffusion Guidance: Training-Free Posterior Correction for Sampling Beyond High-Density Regions
Van Khoa Nguyen (HES-SO Geneva), Alexandros Kalousis (HES-SO Geneva)
Image TranslationGenerationSuper ResolutionDrug DiscoveryDiffusion modelScore-based ModelImageTextBiomedical DataStochastic Differential Equation
🎯 What it does: This paper proposes the Stein Diffusion Guidance (SDG) framework, which achieves efficient guidance for diffusion models in low-density regions by combining stochastic optimal control (SOC) with Stein variational inference in an untrained posterior correction.
Stem: Rethinking Causal Information Flow in Sparse Attention
Lin Niu (Tencent), S Kevin Zhou
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes a training-agnostic sparse attention module called Stem, which optimizes the sparsification of self-attention through causal information flow during the pre-filling phase.
Step-Level Sparse Autoencoder for Reasoning Process Interpretation
Xuan Yang (City University of Hong Kong), Ning Miao (City University of Hong Kong)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderTextChain-of-Thought
🎯 What it does: Propose a Step-level Sparse Autoencoder (SSAE) to extract sparse, interpretable features from the reasoning steps of large language models (LLMs), and use these features for explaining and regulating the reasoning process.
Step-Resolved Data Attribution for Looped Transformers
Georgios Kaissis (University of Potsdam), Eleni Triantafillou (Google DeepMind)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Studies how to progressively attribute the impact of training samples in cyclic Transformers (shared blocks used in τ recursive iterations), proposing the Step-Decomposed Influence (SDI) method, which decomposes traditional TracIn influence scores into τ-step influence trajectories, and implements a streaming sketch-during-backprop scheme based on TensorSketch, making it feasible at the LLM scale.
Step-Size Stability in Stochastic Optimization: A Theoretical Perspective
Fabian Schaipp (Inria), Adrien Taylor (Inria)
OptimizationImageTabular
🎯 What it does: Studied the sensitivity of stochastic optimization methods to step sizes, and proposed a stability metric δ_t to measure how suboptimality degrades when the step size is too large.
STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction
Jinhao Li (Shanghai Jiao Tong University), Guohao Dai (Shanghai Jiao Tong University)
GenerationOptimizationComputational EfficiencyRobotic IntelligenceTransformerPrompt EngineeringDiffusion modelScore-based ModelContrastive LearningImageVideoMultimodalityBenchmarkStochastic Differential Equation
🎯 What it does: Propose a lightweight warm-start mechanism (spatiotemporal consistency prediction) and velocity-aware perturbation injection to accelerate diffusion-based audio-visual motion control strategies, enabling high-quality action generation in just 2 steps.
StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning
Hao Wang (Beihang University), Jie M. Zhang (King's College London)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose the StepCodeReasoner framework, which transforms code reasoning into a task of predicting observable step-by-step execution states, and achieves supervision of intermediate execution states by automatically inserting print anchor points.
STFlow: Data-Coupled Flow Matching for Geometric Trajectory Simulation
Kiet Bennema ten Brinke (Eindhoven University of Technology), Vlado Menkovski (Eindhoven University of Technology)
GenerationData SynthesisGraph Neural NetworkTransformerScore-based ModelFlow-based ModelGraphTime SeriesSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposes a conditional geometric trajectory generation model called STFlow, which achieves one-shot complete trajectory simulation using flow matching, graph neural networks, and hierarchical convolution.
StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning
Shiyang Li (University of Minnesota-Twin Cities), Caiwen Ding (University of Minnesota-Twin Cities)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed the StitchCUDA multi-agent framework for end-to-end automatic generation of GPU programs.
STLA: Spatiotemporal Lookahead Alignment for Post-Training Quantization
Zuqi Zhang (University of Macau), Jiawei Xu (University of Macau)
OptimizationComputational EfficiencyTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Proposes the STLA (Spatiotemporal Lookahead Alignment) framework to address the issue of spatiotemporal misalignment in PTQ (Post-Training Quantization), achieving efficient and accurate quantization of large language model (LLM) weights.
Stochastic Gradient Methods under Heavy-Tailed Noises in Weakly Convex Optimization
Tianxi Zhu (Dalian University of Technology), Xiangyang Ji (Tsinghua University)
OptimizationTextTabularBenchmark
🎯 What it does: This paper studies the convergence of stochastic gradient descent (SGD) and its gradient clipping version (Clip-SGD) under weakly convex optimization in the presence of heavy-tailed noise. It provides expected convergence rates and high probability convergence bounds, and proves that Clip-SGD can achieve high probability convergence with polynomial logarithmic (polylog) dependence on the domain under the condition of p-bounded-central-moment (p-BCM) or sub-Weibull distributed noise.
Stochastic Gradient Variational Inference with Price's Gradient Estimator from Bures-Wasserstein to Parameter Space
Kyurae Kim (University of Pennsylvania), Trevor Campbell (University of British Columbia)
OptimizationComputational EfficiencyRepresentation LearningScore-based ModelContrastive LearningTabularTime SeriesSequentialBenchmarkStochastic Differential Equation
🎯 What it does: Investigate and compare algorithms for stochastic gradient variational inference in the parameter space and the Bures-Wasserstein space, and clarify that their iteration complexity can be determined by a unified gradient estimator—the Price gradient.
Stochastic Lifting for Generating Trajectories of Stochastic Physical Systems
Jules Berman (New York University), Benjamin Peherstorfer (New York University)
GenerationConvolutional Neural NetworkRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelImageVideoTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose a generative model that uses stochastic lifting to perform one-step prediction for time-dependent stochastic physical systems;
Stochastic Linear Bandits with Parameter Noise
Daniel Ezer (Tel Aviv University), Yishay Mansour (Tel Aviv University)
Reinforcement LearningGaussian SplattingStochastic Differential Equation
🎯 What it does: This paper studies the stochastic linear bandit problem under a parameter noise model and proposes two algorithms: VASE (a progressive elimination algorithm for finite action sets) and VALEE (a simple explore-exploit algorithm for ℓ_p unit balls). Both algorithms achieve optimal or near-optimal amortized regret bounds based on variance.
Stochastic Minimum-Cost Reach-Avoid Reinforcement Learning
Jingduo Pan (Chinese Academy of Sciences), Bai Xue (Chinese Academy of Sciences)
Reinforcement LearningContrastive LearningTabularTime SeriesStochastic Differential Equation
🎯 What it does: Propose a minimum cost reach-avoid reinforcement learning framework in random environments, and achieve probabilistic guarantees for safety constraints through the 'Reach-Avoid Probability Certificate' (RAPC);
Stochastic Neural Ray Tracing for Radio Frequency Channel Modeling
Yinyan Bu (University of California San Diego), Piya Pal (University of California San Diego)
OptimizationDiffusion modelNeural Radiance FieldAuto EncoderPoint CloudTabularTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: Developed an RF channel modeling framework, SNRFT, that combines neural networks with ray tracing.
Stochastic Order Learning: An Approach to Rank Estimation Using Noisy Data
Chaewon Lee (Korea University), Chang-Su Kim (Korea University)
Representation LearningData-Centric LearningContrastive LearningImageTextBiomedical DataStochastic Differential Equation
🎯 What it does: Proposes a framework based on stochastic ordinal learning (SOL), which utilizes ordered labels under noise for rank estimation.
Stochastic Sparse Attention for Memory-Bound Inference
Kyle Lee (University of California Santa Barbara), Kerem Yunus Camsari
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: In long-context autoregressive decoding, SANTA (Stochastic Additive No-mulT Attention) is introduced, which significantly reduces the KV cache read bandwidth by sampling a very small number of value rows from the softmax distribution and aggregating them. Additionally, S²ANTA provides hierarchical/systematic sampling variants to reduce variance, and implements GPU-friendly sampling and addition kernels. Furthermore, Bernoulli qKᵗ sampling is proposed to achieve sparse feature access during the score phase.
Stop the Flip-Flop: Context-Preserving Verification for Fast Revocable Diffusion Decoding
Yanzheng Xiang (Kings College London), Yulan He (Kings College London)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextBenchmark
🎯 What it does: Propose the COVER method to solve the flip-flop oscillation in revocable diffusion decoding, improving the efficiency of parallel decoding.
Stop Training for the Worst: Progressive Unmasking Accelerates Masked Diffusion Training
Jaeyeon Kim (Harvard University), Sitan Chen (Harvard University)
GenerationComputational EfficiencyData-Centric LearningTransformerPrompt EngineeringDiffusion modelScore-based ModelTextChain-of-Thought
🎯 What it does: Propose Progressive Unmasking (PUMA), which improves the forward masking process of MDM, making the masking pattern during training consistent with that during inference, thereby accelerating the training of MDM.
Stop When Further Reasoning Won’t Help: Attention-State Adaptive Generation in Reasoning Models
Jiakai Li (University of Electronic Science and Technology of China), Shuang Liang (University of Electronic Science and Technology of China)
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a no-training, plug-and-play attention state adaptive generation method called ASAG to address the overthinking problem in large inference models.
STORM: Segment, Track, and Object Re-Localization from a Single Image
Yu Deng (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
Object TrackingSegmentationPose EstimationTransformerVision Language ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageVideo
🎯 What it does: Propose the STORM framework, which achieves automated 6D pose estimation and tracking with only a single reference image, and can automatically relocalize when the target is missing;
StormInsight: Hierarchical Environmental Forcing and Vertical Coupling for Weather System Evolution
Jun CHEN, Kaishun Wu (Hong Kong University of Science and Technology)
GenerationData SynthesisOptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerMixture of ExpertsFlow-based ModelAuto EncoderImageVideoTabularTime SeriesSequentialBenchmarkOrdinary Differential Equation
🎯 What it does: Propose the StormInsight framework, which constructs a three-dimensional, environment-driven convective system evolution model based on multi-source observations and reanalysis data;
Strat-Reasoner: Reinforcing Strategic Reasoning of LLMs in Multi-Agent Games
Yidong He (South China University of Technology), Mengchen Zhao (South China University of Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the Strat-Reasoner framework, which enhances the strategic reasoning capabilities of LLMs in two-player games by utilizing recursive reasoning, centralized CoT comparison, and hybrid advantage estimation.
Strategic Candidacy in Generative AI Arenas
Chris Hays (Massachusetts Institute of Technology), Manish Raghavan
Recommendation SystemLarge Language ModelScore-based ModelContrastive LearningTextBenchmark
🎯 What it does: This paper investigates the issue of ranking manipulation caused by model cloning in generative AI arenas, proving that the traditional Bradley-Terry maximum likelihood ranking mechanism is not robust against cloning, and proposes a You-Rank-We-Rank (YRWR) mechanism that utilizes producer self-evaluation rankings to achieve approximate cloning robustness.
Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections
Łukasz Borchmann (Snowflake AI Research), Anupam Datta
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Introduces the MADQA benchmark, where multimodal agents complete 2,250 human-written question-answering tasks across 800 PDF documents, evaluating the agents' strategic reasoning and pure random search capabilities.
Strategy Executability in Mathematical Reasoning: Leveraging Human–Model Differences for Effective Guidance
Weida Liang (National University of Singapore), Kenji Kawaguchi (National University of Singapore)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper conducts a strategy-level comparison between human-generated and model-generated mathematical solutions, revealing that the frequency of strategy usage does not equate to its executability in the target model. Based on this finding, the paper proposes the Selective Strategy Retrieval (SSR) framework, which leverages multi-path retrieval and executability prediction to select and combine the most executable strategies, thereby enhancing model reasoning performance.
Strategy-Aware Optimization Modeling with Reasoning LLMs
Ruiqing Zhao (Beihang University), Junlan Feng (JIUTIAN Research)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the SAGE framework, which improves automated optimization modeling through explicit modeling strategies combined with reinforcement learning
Stratified GRPO: Handling Structural Heterogeneity in Reinforcement Learning of LLM Search Agents
Mingkang Zhu (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
TransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: This paper proposes Stratified GRPO, which addresses the bias in advantage estimation caused by structural heterogeneity by structuring trajectories hierarchically in reinforcement learning for the LLM search agent.
Stream RAG: Instant and Accurate Spoken Dialogue Systems with Streaming Tool Usage
Siddhant Arora (Meta AI), Zhaojiang Lin (Meta AI)
RetrievalComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningTextRetrieval-Augmented GenerationAudio
🎯 What it does: Proposed the Stream RAG framework, achieving end-to-end speech-to-speech dialogue systems that can parallelly initiate retrieval tool calls while the user is speaking, significantly improving answer accuracy and reducing first-word latency.
StreamFlow: Theory, Algorithm, and Implementation for High-Efficiency Rectified Flow Generation
Sen Fang (Rutgers University), Dimitris N. Metaxas (Rutgers University)
GenerationComputational EfficiencyTransformerDiffusion modelFlow-based ModelRectified FlowOptical FlowImageText
🎯 What it does: Proposed the StreamFlow acceleration framework, implementing batch velocity field computation, asynchronous time-step scheduling, and dynamic TensorRT compilation for the Rectified Flow model;
Streaming Covariate Balancing via Discrepancy-Based Feature Coresets
Yixin Ren (Fudan University), Shuigeng Zhou (Fudan University)
Domain AdaptationComputational EfficiencyData-Centric LearningContrastive LearningTabularTime SeriesBenchmark
🎯 What it does: Propose a feature coreset method based on discrepancy measurement, enabling average treatment effect (ATE) estimation under online and low-memory settings, directly learning balancing weights without relying on parameterized propensity score models.
Streaming Sliced Optimal Transport
Khai Nguyen (University of Texas at Austin)
OptimizationComputational EfficiencyData-Centric LearningScore-based ModelContrastive LearningImagePoint CloudTabularTime Series
🎯 What it does: Proposed a streaming sliced Wasserstein (Stream-SW) distance estimation method based on quantized sketches (KLL Sketch), which can compute the sliced Wasserstein distance between distributions in real-time from infinite streaming data under limited memory.
StretchTime: Adaptive Time Series Forecasting via Symplectic Attention
Yubin Kim (Georgia Institute of Technology), Jiecheng Lu (Georgia Institute of Technology)
Computational EfficiencyTransformerTime SeriesBenchmark
🎯 What it does: Propose a novel Transformer architecture called StretchTime, which can adaptively model nonlinear time distortions in time series;
STRIDE: Post-Training LLMs to Reason and Refine Bio-Sequences via Edit Trajectories
Daiheng Zhang (Rutgers University), David van Dijk (Yale University)
OptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBiomedical Data
🎯 What it does: Propose the STRIDE framework, training large language models to generate executable INSERT/DELETE/REPLACE edit trajectories for optimizing biological sequences (proteins, molecules).
Stronger Benchmarks for Prediction as a Service with Constraints
Yahav Bechavod (University of Pennsylvania), Aaron Roth (University of Pennsylvania)
OptimizationFederated LearningReinforcement Learning from Human FeedbackBenchmark
🎯 What it does: This paper proposes an online learning method based on the 'forecast-as-a-service' framework, which can provide feasible predictions for multiple downstream decision-makers under long-term constraints, and allows these decision-makers to achieve low dynamic loss and constraint violation through 'constraint-optimal response' to the predictions.
StructMamPose: From Sequential Perception to Structural Reasoning for 3D Human Pose Estimation
Jiahong Jiang (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
Pose EstimationGraph Neural NetworkTransformerContrastive LearningImageVideo
🎯 What it does: Propose the StructMamPose framework, combining the Identity Anchoring Mechanism (IAM) and the View Transformation Hub (VTH) to improve monocular 3D human pose estimation.
StructMAR: Structure-Aware Masked Autoregression for Explicit Layout Alignment in Text-to-Image Generation
Gang Cao (Xidian University), Junying Zhang (Xidian University)
GenerationTransformerReinforcement LearningDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Proposes StructMAR, a structure-aware Masked Autoregressive framework, for achieving explicit instance-level layout alignment in text-to-image generation, with the core implementation achieved through 2D Rotary Positional Embedding, layout-guided attention bias, gated cross-attention, and GRPO-based reward fine-tuning;
Structurally Aligned Subtask-Level Memory for Software Engineering Agents
Kangning Shen (Kuaishou Technology), Yang Yue (Kuaishou Technology)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a structure-aligned subtask-level memory method to enhance the long-term reasoning capabilities of LLM agents in software engineering;
Structure Abstraction and Generalization in a Hippocampal-Entorhinal Inspired World Model
Tianqiu Zhang (Peking University), Si Wu (Peking University)
GenerationRepresentation LearningReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelDiffusion modelAuto EncoderContrastive LearningWorld ModelVideo
🎯 What it does: Propose a hierarchical world model inspired by the hippocampus-medial entorhinal cortex (HPC-MEC), which can unsupervisedly learn abstract structures from raw video sequences and predict the next frame.
Structure Enables Effective Self-Localization of Errors in LLMs
Ankur Samanta (Columbia University), Yonathan Efroni (Tel Aviv University)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a self-correction framework based on "Thought-ICS" (Thought-Iterative Correction Self-calibration), which generates ideas step by step through thinking, accurately identifies errors and backtracks for regeneration, thereby achieving self-correction in LLMs.
Structure-Aware Consistency Priors for Shape from Polarization in Complex Media
Kaimin Yu (Fuzhou University), Xianyu Wu (Fuzhou University)
RestorationDepth EstimationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningImagePhysics Related
🎯 What it does: Recover the surface normals of complex media such as ice from single-view polarized images.
Structure-aware Granular-Ball based Information Bottleneck for Multi-modal Clustering
Zhengzheng Lou (Zhengzhou University), Shizhe Hu (Zhengzhou University)
Representation LearningContrastive LearningImageVideoTextMultimodality
🎯 What it does: Proposes a structure-aware information bottleneck (SGB-IB) framework based on granular spheres (GB), using granular spheres to perform adaptive hierarchical clustering on multimodal data, and introduces structural weights into the information bottleneck framework to achieve a balance between information compression and retention.
Structure-Aware Riemannian Flow Matching for Registration and Fusion of Hyperspectral and Multispectral Images
Quan Zhang (National University of Defense Technology), Yuanxi Peng (National University of Defense Technology)
Image TranslationRestorationTransformerDiffusion modelScore-based ModelFlow-based ModelOptical FlowImageAgriculture Related
🎯 What it does: Propose a joint alignment and fusion framework based on structure-aware Riemannian flow matching (SARFM) for handling non-rigid mismatches in spectral and multispectral images.
Structure-Centric Graph Foundation Model via Geometric Bases
Xiaodong He (University of Electronic Science and Technology of China), zhao kang
ClassificationRepresentation LearningDrug DiscoveryGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Propose a structure-centric graph foundation model (SCGFM), which maps heterogeneous graphs into a unified structural coordinate system through learnable geometric bases, and utilizes the Gromov-Wasserstein distance to achieve structural alignment and feature re-encoding;
Structure-Induced Information for Rerooting Levin Tree Search
Jake Tuero (University of Alberta), Levi Lelis
OptimizationReinforcement LearningTabularSequential
🎯 What it does: Propose and evaluate a structure-induced rerooting technique to improve the search efficiency of Levin Tree Search (LTS) in complex single-agent planning problems;
Structure-Preserving Learning Improves Geometry Generalization in Neural PDEs
Benjamin D Shaffer, Nathaniel Trask (Sandia National Laboratories)
TransformerDiffusion modelAuto EncoderContrastive LearningMeshBenchmarkPhysics Related
🎯 What it does: Studied a neural PDE solver called Geo-NeW that can maintain physical structure on different geometric shapes.
Structured 4D Latent Predictive Model for Robot Planning
Zhiyi Li (MIT), Yilun Du (Harvard University)
Robotic IntelligenceConvolutional Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelDiffusion modelFlow-based ModelAuto EncoderGaussian SplattingImageVideoTextPoint Cloud
🎯 What it does: A structured 4D latent prediction model is proposed, which uses a sparse voxel latent space to model the spatiotemporal evolution of 3D scenes, and converts the predicted latent states into executable robot actions for high-level planning.
Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges
Eitan Kosman (Ben-Gurion University of Negev), Chaim Baskin (Ben-Gurion University of Negev)
Image TranslationGenerationData SynthesisTransformerDiffusion modelScore-based ModelContrastive LearningImageMultimodalityPoint CloudMesh
🎯 What it does: Proposes Structured Diffusion Bridges (SDB), a diffusion bridge framework that leverages structured constraints to perform cross-modal translation when paired data is missing or scarce.
Structured Expert Routing with Multi-View Task Priors for Offline Meta-Reinforcement Learning
Yisen Zhao (Peking University), Zhuojian Li (Peking University)
Meta LearningReinforcement LearningPrompt EngineeringMixture of ExpertsTabularSequentialBenchmark
🎯 What it does: Propose Task-Guided Router (TGR), an offline meta reinforcement learning framework that utilizes multi-perspective task priors (semantics, trajectories, and latent dynamics) for structured expert routing.
Structured Multi-modal Graph Disentanglement for Psychiatric Diagnosis
Hongyu Shi (Xi'an Jiaotong University), Liangjun Chen (Xi'an Jiaotong University)
ClassificationExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkTransformerContrastive LearningGraphBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose Structured Multi-modal Graph Disentanglement (SMGD), which decomposes multi-modal graph embeddings at a fine-grained level, separating four subspaces: shared diagnosis, complementary diagnosis, incidental consensus, and modality-specific noise, while using geometric regularization to suppress shortcut paths, thereby enhancing interpretability and cross-dataset generalization in mental disorder diagnosis.
Structured Multi-step Jailbreaking under a Hamiltonian Generative Formulation
Zihan Zhou (Auburn University), Dejing Dou (Fudan University)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringDiffusion modelScore-based ModelTextStochastic Differential Equation
🎯 What it does: Propose a structured multi-step jailbreak framework SJA, which can decompose harmful queries into several safe sub-questions, and then gradually request the LLM to reconstruct the original answer.
Structured Progressive Knowledge Activation for LLM-Driven Neural Architecture Search
Zhen Liu (Xi'an Jiaotong University), Jingwen Fu (Zhongguancun Academy)
Neural Architecture SearchTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes SPARK, a structured, progressive activation knowledge-based neural architecture search method based on LLMs, aiming to reduce functional coupling and improve search efficiency and reliability through explicit functional factor selection and conditional patch generation.
STT-LLM: Structural-Temporal Tokenization for Adapting LLMs to Longitudinal Clinical Profiles
Maxx Richard Rahman (German Research Center for Artificial Intelligence), Wolfgang Maass (German Research Center for Artificial Intelligence)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringContrastive LearningTime SeriesSequentialBiomedical DataElectronic Health Records
🎯 What it does: Propose the STT-LLM structural temporal tokenization framework, converting clinical longitudinal data (such as athlete steroid metabolism profiles) into tokens compatible with LLMs, enabling sequence prediction and anomaly detection;
Student-Centered Distillation Narrows the Agentic Gap Between Small and Large LLMs
Yuanjie Lyu (University of Science and Technology of China), Tong Xu (University of Science and Technology of China)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a student-centric knowledge distillation framework called SCoRe, where the small model first generates complete trajectories on its own, and the teacher only corrects the first error. Then, the corrected trajectories are used for SFT, and further capability improvements are achieved through short-term reinforcement learning.
StyleDistillation: A New Insight of Image Style Enables Personalized Aesthetic Manipulation
Yuxin Wang, Zheng Wang (Wuhan University)
Image TranslationGenerationKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes the StyleDistillation framework, which can distill pure style representations from a single reference image and combine them with text prompts to generate images in text-guided diffusion models that are both consistent with the textual content and highly faithful to the reference style; it also supports text-driven style editing and multi-style fusion.
Subgroup Discovery with the Cox Model
Zachary Izzo (NEC Labs America), Iain Melvin (NEC Labs America)
TabularTime SeriesBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: To address the problem of subgroup discovery in the Cox model for survival analysis, two new evaluation metrics and corresponding algorithms are proposed and verified.
Subliminal Effects in Your Data: A General Mechanism via Log-Linearity
Ishaq Aden-Ali (University of California Berkeley), Nika Haghtalab (University of California Berkeley)
Explainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the LOGIT-LINEAR SELECTION (LLS) method, which achieves subconscious learning for large language models by performing subset selection on preference datasets;
Submodular Optimization for Minimal Augmentation in Robust Language Model Alignment
Ching-Chia Kao (National Taiwan University), Chu-song Chen
OptimizationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose an incremental data augmentation framework based on submodular functions, utilizing cyclic augmentation and submodular coverage theory to achieve minimal augmentation for secure alignment.
Subspace-Aware Feature Reshaping for Open-Set Graph Class-Incremental Learning
Weichao Zhang (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
ClassificationFederated LearningRepresentation LearningGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraph
🎯 What it does: Propose the SAFER framework to address the open-set (unknown class) problem in graph-class incremental learning (GCIL), using subspace-aware feature reshaping and task routing under the condition of no replay.
SubspacePath Pruner: Inference-time Pruning via Probe-based Representation–Parameter Coupling
Gong Zhiren (Nanyang Technological University), Wei Yang Bryan Lim (Nanyang Technological University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose a subspace path pruning method achieved through detector and domain axis coupling during inference
Success-Conditioning as Policy Improvement: The Optimization Problem Solved by Imitating Success
Daniel Russo (Columbia University)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningSequential
🎯 What it does: This paper studies success conditioning — supervised imitation learning on successfully executed trajectories, and proves its equivalence to performing a specific χ²-constrained trust region optimization on the behavior policy.
SuCo: Sufficiency-guided Continuous Adaptive Reasoning
Jiahao Wang (Harbin Institute of Technology), Xuelong Li (TeleAI of China Telecom)
OptimizationComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialChain-of-Thought
🎯 What it does: Proposed and trained the SuCo framework, enabling large models to continuously adjust the length of chain-of-thought reasoning based on problem difficulty.
SuperHype: Hypergraph Generation via Graph-Superposition Decomposition
Lucas Gantes (École Polytechnique), Lydia Chen (University of Neuchâtel)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGraph
🎯 What it does: This paper proposes a new hypergraph generation method called SuperHype, which uses graph-stacking decomposition to precisely map hypergraphs to multi-layer regular graphs, and builds a graph-stacking Transformer to achieve reversible diffusion generation.
Supervise Less, See More: Training-free Nuclear Instance Segmentation with Prototype-Guided Prompting
Wen Zhang (Stony Brook University), Chenyu You (Stony Brook University)
SegmentationData-Centric LearningTransformerPrompt EngineeringDiffusion modelContrastive LearningOptical FlowImageBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission TomographyUltrasoundFibre Orientation DistributionDiffusion Tensor ImagingAlzheimer's DiseaseElectronic Health RecordsElectrocardiogram
🎯 What it does: Developed a fully unsupervised, no-training nuclear cell instance segmentation framework called SPROUT, which utilizes H&E color priors for self-retrieved prototypes and generates point prompts via partial optimal transport, guiding SAM to produce nuclear segmentation.
Supervised Classification Heads as Semantic Prototypes: Unlocking Vision-Language Alignment via Weight Recycling
David Méndez (University of Granada), Natalia Díaz-Rodríguez (University of Granada)
ClassificationRetrievalRepresentation LearningData-Centric LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: By reusing the classification head weights of pre-trained visual models as semantic prototypes, achieving post-hoc visual-language alignment without requiring image-text paired data, and using it as a data augmentation method to enhance the performance of existing alignment methods.
Supervised Graph Contrastive Learning for Gene Regulatory Networks
Sho Oshima (Kyoto University), Ryosuke Kojima (Kyoto University)
Representation LearningDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningGraphBiomedical Data
🎯 What it does: Proposed a supervised graph contrastive learning method called SupGCL, which uses real perturbations generated from gene knockdown experiments as supervision to learn representations of patient-specific gene regulatory networks.
Supervised Guidance Training for Infinite-Dimensional Diffusion Models
Elizabeth Louise Baker (Technical University Of Denmark), Jes Frellsen (Technical University Of Denmark)
GenerationData SynthesisOptimizationComputational EfficiencyRepresentation LearningTransformerDiffusion modelScore-based ModelImageTabularTime SeriesSequentialBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: A conditional sampling theory is proposed in the function space for score-based diffusion models, and a supervised guided training (SGT) method is introduced to learn the guidance terms that are not directly accessible, thereby achieving efficient and stable posterior sampling.
Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization
Yonghan Yang (Mohamed bin Zayed University of Artificial Intelligence), Xue Liu (Mohamed bin Zayed University of Artificial Intelligence)
OptimizationConvolutional Neural NetworkRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelTabularBenchmark
🎯 What it does: Propose SPADE, a conditional diffusion model combined with support proximity regularization, for offline black-box optimization;
Suppress and Diversify: Refining Robust Pathways for Corruption Robustness
Jiangang Yang (Institute of Microelectronics Chinese Academy of Sciences), Jian Liu (Institute of Microelectronics Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationContrastive LearningImage
🎯 What it does: This paper proposes a parameter-free, zero-inference-cost S&D method during the training phase, which improves the model's robustness to natural image corruption by suppressing non-robust paths and diversifying robust paths.
SURF: Separation via Unsupervised Remixing Flow
Henry Li (Google), John R. Hershey (Google DeepMind)
RestorationGenerationRepresentation LearningTransformerScore-based ModelFlow-based ModelAuto EncoderContrastive LearningImageAudio
🎯 What it does: Propose a fully unsupervised source separation method called SURF, which learns source separation from only mixed signals by combining flow matching with self-supervised remixing.
SURGE: Approximation and Training Free Particle Filter for Diffusion Surrogate
Lifu Wei (Northwestern University), Yiping Lu (Northwestern University)
Data SynthesisOptimizationDiffusion modelImageTabularTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose a particle filter framework (SURGE) that utilizes diffusion models as digital twins in data assimilation, without requiring training and approximately unbiased, by correcting the bias caused by guidance through Girsanov reweighting and stepwise resampling of complete trajectories to strictly integrate observational information;
SURGE: Surrogate Gradient Adaptation in Binary Neural Networks
Haoyu Huang (Beihang University), Baochang Zhang (Beihang University)
ClassificationImage TranslationObject DetectionSegmentationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageTextBenchmark
🎯 What it does: Propose the SURGE framework, improving the training of binary neural networks through a dual-path gradient compensator and adaptive gradient scaling.
Surgery: Mitigating Harmful Fine-Tuning for Large Language Models via Attention Sink
Guozhi Liu (South China University of Technology), Li Shen (Sun Yat-sen University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This work proposes the Surgery method, which suppresses the model from learning harmful patterns during the fine-tuning phase by utilizing the conflicting information from attention sinks, thereby enhancing safety.
SurrogateSHAP: Training-Free Contributor Attribution for Text-to-Image (T2I) Models
MingYu Lu, Su-In Lee (University of Washington)
GenerationExplainability and InterpretabilityTransformerDiffusion modelImageText
🎯 What it does: Propose an untrained contributor attribution framework called SurrogateSHAP, which can compute group-level Shapley values for text-to-image diffusion models.
SurvDiff: A Diffusion Model for Generating Synthetic Data in Survival Analysis
Marie Brockschmidt (LMU Munich), Stefan Feuerriegel (LMU Munich)
Data SynthesisTransformerDiffusion modelScore-based ModelTabularBiomedical DataElectronic Health RecordsStochastic Differential Equation
🎯 What it does: Developed SURVDIFF, an end-to-end diffusion model for generating synthetic survival data that includes covariates, event times, and right-censoring information.
SVD as a Fast Interpretability Method for Transformers
Min Xue (Heidelberg University), Artur Andrzejak (Heidelberg University)
Explainability and InterpretabilityComputational EfficiencyTransformerAuto EncoderText
🎯 What it does: By performing SVD decomposition on the Transformer MLP weights, we propose Detector-Effector Units (DEU) and Subspace Contribution Analysis (SCA) to achieve a training-free interpretable framework called NaNA.
SVL: Empowering Spiking Neural Networks for Efficient 3D Open-World Understanding
Xuerui Qiu (Chinese Academy of Sciences), Guoqi Li (Chinese Academy of Sciences)
ClassificationObject DetectionSegmentationAutonomous DrivingComputational EfficiencyRepresentation LearningSpiking Neural NetworkTransformerVision Language ModelDiffusion modelScore-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextMultimodalityPoint Cloud
🎯 What it does: Propose the SVL framework, enabling SNNs to achieve zero-shot classification, segmentation, detection, and action recognition in 3D open-world tasks, and support 3D object description and question answering;
SVL: Goal-Conditioned Reinforcement Learning as Survival Learning
Franki NGUIMATSIA TIOFACK, Justin Carpentier (Inria and Ecole Normale Supérieure, PSL Research University)
Reinforcement LearningGraphTabularTime SeriesBenchmark
🎯 What it does: Proposes a probabilistic framework that treats goal-conditioned reinforcement learning as survival learning, estimating the value function by maximum likelihood learning of the hazard function.
SVRG and Beyond via Posterior Correction
Nico Daheim (Technical University of Darmstadt), Mohammad Emtiyaz Khan (RIKEN)
OptimizationComputational EfficiencyRepresentation LearningImageTextTabular
🎯 What it does: Studied the relationship between SVRG and the Bayesian method Posterior Correction (PoCo), and based on this relationship, proposed new variants of SVRG, constructing algorithms such as Newton-like SVRG and Adam-like IVON-PoCo.
SWE-ABS: Adversarial Benchmark Strengthening Exposes Inflated Success Rates on Test-based Benchmark
Boxi Yu (University of Limerick), Lionel BRIAND
Adversarial AttackAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the SWE-ABS framework, which enhances the SWE-Bench test set adversarially to reveal and correct semantic errors in model-generated patches.