ICLR 2026 Papers — Page 43
International Conference on Learning Representations · 5356 papers
SelvaBox: A high‑resolution dataset for tropical tree crown detection
Hugo Baudchon (Mila - Quebec AI Institute), Etienne Laliberté (Université de Montréal)
Object DetectionConvolutional Neural NetworkTransformerImageBenchmarkAgriculture Related
🎯 What it does: This paper constructs the largest tropical tree crown detection dataset, SELVABOX, and achieves state-of-the-art detection performance in both tropical and non-tropical scenarios through multi-scale, transformation-agnostic training and a Transformer-based detection model.
SEMA: Simple yet Effective Learning for Multi-Turn Jailbreak Attacks
Mingqian Feng (University of Rochester), Jianfeng Gao (Microsoft Research)
Adversarial AttackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the SEMA framework, which trains a multi-round hacking model without strategy or external data using reinforcement learning with prefilling self-tuning and intent drift-aware reward mechanisms;
Semantic Parallelism: Redefining Efficient MoE Inference via Model-Data Co-Scheduling
Yan Li (Huawei Technologies), Pengfei Zheng (Huawei Technologies)
Computational EfficiencyMixture of ExpertsText
🎯 What it does: Propose Semantic Parallelism, reducing all-to-all communication in Mixture-of-Experts (MoE) inference through model and data co-scheduling, significantly improving inference throughput and latency performance.
Semantic Regexes: Auto-Interpreting LLM Features with a Structured Language
Angie Boggust (MIT), Fred Hohman (Apple)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: Proposed and implemented a structured language called semantic regexes for automatically interpreting features in large language models (LLMs), published at ICLR 2026.
Semantic Uncertainty Quantification of Hallucinations in LLMs: A Quantum Tensor Network Based Method
pragatheeswaran vipulanandan, Dilip Sarkar (University of Miami)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Propose an unsupervised method for hallucination detection and answer selection by quantifying the probability uncertainty of LLM-generated sequences, combining quantum tensor networks (QTN) with semantic Rényi entropy.
Semantic Visual Anomaly Detection and Reasoning in AI-Generated Images
Chuangchuang Tan (Beijing Jiaotong University), Yan Lu (Microsoft Research Asia)
Anomaly DetectionExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelImageBenchmark
🎯 What it does: Constructed the AnomReason benchmark dataset and the AnomAgent multi-modal multi-agent framework for detecting and explaining semantic anomalies in AI-generated images.
Semantic Voting: A Self-Evaluation-Free Approach for Efficient LLM Self-Improvement on Unverifiable Open-ended Tasks
Chunyang Jiang (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
Computational EfficiencyRepresentation LearningLarge Language ModelContrastive LearningText
🎯 What it does: Proposes a self-assessment-free semantic voting framework that uses sentence embeddings to perform soft voting on LLM-generated answers, generating pseudo-labels for self-improvement
Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment
Byeonghu Na (KAIST), Il-chul Moon
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a policy regularization framework (WPR) based on entropy-regularized Wasserstein distance for aligning large language models in reinforcement learning from human feedback (RLHF);
Semantic-Enhanced Time-Series Forecasting via Large Language Models
Hao Liu (University of Science and Technology Beijing), Xiaobin Zhu (University of Science and Technology Beijing)
Recurrent Neural NetworkTransformerLarge Language ModelAuto EncoderTime Series
🎯 What it does: Proposed a new time series forecasting framework called SE-LLM, which leverages the pre-trained knowledge of large language models (LLMs) and further enhances the understanding and prediction of time series through the TSCC module and Time-Adapter.
SemHiTok: A Unified Image Tokenizer via Semantic-Guided Hierarchical Codebook for Multimodal Understanding and Generation
Zisheng Chen (Sun Yat-sen University Huawei Noah's Ark Lab), Xiaodan Liang (Sun Yat-sen University Huawei Noah's Ark Lab)
GenerationRepresentation LearningLarge Language ModelVision Language ModelAuto EncoderImageMultimodality
🎯 What it does: Proposes SemHiTok, a unified image tokenizer that provides consistent discrete representations in multimodal understanding and generation tasks;
Semi-Parametric Contextual Pricing with General Smoothness
Yuxuan Han (New York University), Zhengyuan Zhou (New York University)
OptimizationReinforcement Learning
🎯 What it does: Study the semi-parametric contextual pricing problem, proposing a unified algorithm to achieve the optimal regret rate (˜O(T^{(β+1)/(2β+1)})) under any Holder smoothness with β≥1;
Semi-Supervised Preference Optimization with Limited Feedback
Seonggyun Lee (Yonsei University), Kyungwoo Song (Yonsei University)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: Propose a semi-supervised preference optimization framework called SSPO, which jointly trains using a small amount of labeled preference comparisons and a large amount of unlabeled SFT data.
SenseFlow: Scaling Distribution Matching for Flow-based Text-to-Image Distillation
Xingtong Ge (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
GenerationKnowledge DistillationVision Language ModelFlow-based ModelImageTextMultimodality
🎯 What it does: This paper extends Distribution Matching Distillation (DMD) to large-scale streaming text-image models, proposing SenseFlow, which includes Implicit Distribution Alignment (IDA), Intra-Segment Guidance (ISG), and a discriminator based on visual foundation models, achieving four-step high-quality generation.
Separable Neural Networks: Approximation Theory, NTK Regime, and Preconditioned Gradient Descent
Yisi Luo (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
OptimizationComputational EfficiencyRepresentation LearningImageMeshPhysics Related
🎯 What it does: This paper provides a theoretical analysis of separable neural networks (SepNN), proving their universal approximation capability, deriving the convergence properties of their neural tangent kernel (NTK) under conditions of infinite width/rank and fixed rank, and proposing an efficient separable preconditioned gradient descent (SepPGD) algorithm to alleviate spectrum bias.
Seq vs Seq: An Open Suite of Paired Encoders and Decoders
Orion Weller, Benjamin Van Durme
ClassificationGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Proposed and made public the ETTIN kit, which includes a series of paired encoder and decoder models using the same data, architecture, and training scheme to conduct a fair comparison of two pretraining objectives.
Sequences of Logits Reveal the Low Rank Structure of Language Models
Noah Golowich (Microsoft Research), Abhishek Shetty (MIT)
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextSequential
🎯 What it does: This paper studies the low-rank structure of large language models by constructing an extended logit matrix, and proposes LINGEN, a text generation method based on linear combinations;
Sequential Information Bottleneck Fusion: Towards Robust and Generalizable Multi-Modal Brain Tumor Segmentation
TIANYI LIU, Kaizhu Huang (Duke Kunshan University)
SegmentationTransformerBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a sequence information bottleneck fusion based multimodal brain tumor segmentation network (SMSN) to address the robust segmentation problem under missing modalities.
Sequential Parallel Duality in Prefix Scannable Models
Morris Yau (MIT), Jacob Andreas (MIT)
Computational EfficiencyRepresentation LearningTransformerTextTabularSequential
🎯 What it does: The paper proposes and studies the Prefix-Scannable Model (PSM) and Transformer-PSM, unifying sequence models that can be parallelly trained but sequentially inferred, and systematically verifies their theoretical and experimental effectiveness.
SERE: Similarity-based Expert Re-routing for Efficient Batch Decoding in MoE Models
Juntong Wu (Taobao & Tmall Group of Alibaba), Li Yuan (Peking University)
Computational EfficiencyMixture of ExpertsText
🎯 What it does: Proposed a dynamic rerouting method called SERE based on expert similarity to accelerate batch decoding in MoE models, significantly reducing the number of activated experts while maintaining model performance.
SeRI: Gradient-Free Sensitive Region Identification in Decision-Based Black-Box Attacks
Feiyang Wang (Beijing University of Posts and Telecommunications), Hangwei Qian (Victoria University of Wellington)
Adversarial AttackImage
🎯 What it does: In the hard-label black-box attack scenario where only the model's top-1 predicted label is observable and the query budget is limited, a decision-benchmark method called SeRI is designed to continuously evaluate and refine the sensitivity of each pixel in the image. It is integrated as a plugin into existing attackers to improve attack efficiency.
SERQ: Saliency-Aware Low-Rank Error Reconstruction for LLM Quantization
Yeonsik Park (Kyung Hee University), Seungkyu Choi (Yonsei University)
OptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposes SERQ, a significance-aware error reconstruction method using a single low-rank matrix, enabling full low-precision inference for LLMs under W4A4/W4A8 precision;
SERUM: Simple, Efficient, Robust, and Unifying Marking for Diffusion-based Image Generation
Jan Kociszewski (CISPA Helmholtz Center for Information Security), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
GenerationDiffusion modelImage
🎯 What it does: Propose a method that injects a unique watermark noise into the initial noise of a diffusion model and trains a lightweight detector to mark generated images
SesaHand: Enhancing 3D Hand Reconstruction via Controllable Generation with Semantic and Structural Alignment
Zhuoran Zhao (Hong Kong University of Science and Technology), Anyi Rao (Hong Kong University of Science and Technology)
GenerationPose EstimationTransformerDiffusion modelImageTextMeshChain-of-Thought
🎯 What it does: Proposed a controllable hand image generation method called SesaHand to enhance the performance of 3D hand reconstruction
SESaMo: Symmetry-Enforcing Stochastic Modulation for Normalizing Flows
Janik Kreit (University of Bonn), Lena Funcke (University of Bonn)
GenerationData SynthesisReinforcement LearningFlow-based ModelPhysics Related
🎯 What it does: Proposed a symmetry-enhanced normalization flow (SESaMo) method that utilizes random modulation.
Setting up for failure: automatic discovery of the neural mechanisms of cognitive errors
Puria Radmard (University of Cambridge), Máté Lengyel (University of Cambridge)
Explainability and InterpretabilityRecurrent Neural NetworkDiffusion modelSequentialBiomedical Data
🎯 What it does: This paper enables the automatic discovery of neural mechanisms by allowing recurrent neural networks (RNNs) to directly learn the behavioral distribution in human visual working memory tasks, particularly polynomial errors (swap errors);
SFBD-OMNI: Bridge models for lossy measurement restoration with limited clean samples
Haoye Lu (University of Waterloo), Darren Lo (University of Waterloo)
RestorationDiffusion modelScore-based ModelFlow-based ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose the SFBD-OMNI framework, which restores the original data distribution through alternating minimization under limited clean samples and a large number of corrupted samples.
SFT Doesn’t Always Hurt General Capabilities: Revisiting Domain-Specific Fine-Tuning in LLMs
Jiacheng Lin (University Of Illinois Urbana Champaign), Lihong Li
Domain AdaptationTransformerSupervised Fine-TuningFlow-based ModelTextBiomedical DataFinance Related
🎯 What it does: Perform domain-specific supervised fine-tuning (SFT) on LLMs, investigate its impact on general capabilities, and propose balancing domain performance and general performance through a small learning rate and Token-Adaptive Loss Reweighting (TALR).
SGD with Adaptive Preconditioning: Unified Analysis and Momentum Acceleration
Dmitry Kovalev (Yandex Research)
Optimization
🎯 What it does: This paper proposes a unified analytical framework to study stochastic gradient descent with adaptive preconditions (such as AdaGrad, Shampoo, ASGO, DASGO, etc.), and for the first time proves that adding Nesterov momentum enables AdaGrad-type algorithms to achieve accelerated convergence; meanwhile, it reveals the intrinsic connection between DASGO and Scion.
SGD-Based Knowledge Distillation with Bayesian Teachers: Theory and Guidelines
Itai Morad (Ben-Gurion University), Yonina C. Eldar (Weizmann Institute of Science)
Knowledge DistillationImage
🎯 What it does: Theoretical analysis of SGD convergence when using Bayesian teachers in knowledge distillation, along with design guidelines
ShapeGen4D: Towards High Quality 4D Shape Generation from Videos
Jiraphon Yenphraphai (Snap), Chaoyang Wang (Snap)
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderGaussian SplattingVideoMesh
🎯 What it does: Proposed ShapeGen4D, a 4D shape generation framework for video, which directly generates dynamic 3D mesh sequences end-to-end from monocular videos;
SHAPO: Sharpness-Aware Policy Optimization for Safe Exploration
Kaustubh Mani (University of Montreal), Liam Paull (University of Montreal)
OptimizationSafty and PrivacyReinforcement Learning
🎯 What it does: Proposes Sharpness-Aware Policy Optimization (SHAPO) by introducing gradient perturbations in the Fisher geometric space of policy parameters, enabling conservative (pessimistic) adjustments to policy updates for enhanced safe exploration;
Sharing State Between Prompts and Programs
Ellie Y Cheng, Michael Carbin (MIT)
AI Code AssistantLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and implemented a shared program state abstraction, enabling LLMs to directly read and write variables, objects, and control flow within programs, simplifying natural language programming.
Sharp asymptotic theory for Q-learning with \texttt{LD2Z} learning rate and its generalization
Soham Bonnerjee (University of Chicago), Wei Biao Wu (University of Chicago)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper systematically derives the theoretical properties of using linear decay to zero (LD2Z) and its generalized power-law decay to zero (PD2Z-ν) learning rates in Q-learning, providing non-asymptotic error upper bounds, the central limit theorem for tail Polyak-Ruppert averages, and a time-uniform strong regularity principle;
Sharp Monocular View Synthesis in Less Than a Second
Lars Mescheder (Apple), Vladlen Koltun (Apple)
GenerationDepth EstimationConvolutional Neural NetworkSupervised Fine-TuningGaussian SplattingImage
🎯 What it does: This paper proposes the SHARP method, which uses a single image to regress a high-resolution 3D Gaussian representation through a neural network in less than one second, enabling real-time realistic rendering of nearby views.
Sharpness-Aware Machine Unlearning
Haoran Tang (Purdue University), Rajiv Khanna (Purdue University)
OptimizationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Investigated the behavior and effectiveness of Sharpness-Aware Minimization (SAM) in machine unlearning tasks, provided theoretical analysis and experimental verification, and proposed a new hierarchical unlearning algorithm called Sharp MinMax based on these findings.
Sharpness-Aware Minimization in Logit Space Efficiently Enhances Direct Preference Optimization
Haocheng Luo (Monash University), Trung Le (Monash University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: This paper investigates the convergence and alignment issues in Direct Preference Optimization (DPO), elucidates the essence of the squeezing effect, and proposes an efficient variant called logits-SAM that applies SAM in the logits space, significantly enhancing the performance and robustness of DPO and its variants.
SHE-LoRA: Selective Homomorphic Encryption for Federated Tuning with Heterogeneous LoRA
Jianmin Liu (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
Federated LearningSafty and PrivacyLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the SHE-LoRA framework, integrating Selective Homomorphic Encryption (SHE) with Low-Rank Adaptation (LoRA), to achieve efficient privacy-preserving cross-device federated fine-tuning of large language models (LLMs).
Sheaves Reloaded: A Direction Awakening
Stefano Fiorini (Istituto Italiano di Tecnologia), Stefano Coniglio
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes Directed Cellular Sheaf and the corresponding Directed Sheaf Laplacian L˜F, and builds the Directed Sheaf Neural Network (DSNN), achieving a graph neural network that explicitly captures edge direction information in directed graphs;
SHIELD: Suppressing Hallucinations In LVLM Encoders via Bias and Vulnerability Defense
Yiyang Huang (Northeastern University), Yun Fu (Northeastern University)
Vision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Investigates the object hallucination problem in vision-language models, identifying it as stemming from statistical bias, inherent bias, and vulnerability in visual encoders, and proposes the SHIELD zero-training framework to mitigate hallucinations.
ShieldedCode: Learning Robust Representations for Virtual Machine Protected Code
Mingqiao Mo (University of Chinese Academy of Sciences), Yangfan He (University of Minnesota Twin Cities)
Representation LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: This work constructs a large-scale paired dataset of source code and virtual machine (VMP) protected code, trains large language models to generate, compare, and reason about protected code, and learns its robust representations.
Shift-and-Sum Quantization for Visual Autoregressive Models
Jaehyeon Moon (Yonsei University), Bumsub Ham (Yonsei University)
RestorationGenerationCompressionComputational EfficiencyTransformerImageMultimodality
🎯 What it does: Propose a post-training quantization (PTQ) method for visual autoregressive models (VAR), mainly including shift-and-sum quantization and calibration data resampling techniques.
ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution
Robert Tjarko Lange (Sakana AI), Edoardo Cetin (Sakana AI)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Develop the SHINKAEVOLVE framework, leveraging LLM-driven evolutionary processes to achieve efficient sample utilization.
Shoot First, Ask Questions Later? Building Rational Agents that Explore and Act Like People
Gabriel Grand (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
Reinforcement Learning from Human FeedbackLarge Language ModelWorld ModelTextChain-of-Thought
🎯 What it does: Built and studied a two-player collaborative 'Battleship' game to evaluate and enhance language models' performance in information search and decision-making;
Shop-R1: Rewarding LLMs to Simulate Human Behavior in Online Shopping via Reinforcement Learning
Yimeng Zhang, Dakuo Wang
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextSequential
🎯 What it does: Propose Shop-R1, a reinforcement learning framework for simulating online shopping behavior, divided into two stages: reasoning (rationale generation) and action prediction;
Short Window Attention Enables Long-Term Memorization
Loïc Cabannes (Ecole Normale Superieure Paris Saclay, Paris Cite University, Inria Paris), Herve Jegou (Ecole Normale Superieure Paris Saclay, Paris Cite University, Inria Paris)
RetrievalRecurrent Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose the SWAX hybrid architecture, alternating sliding window attention (SWA) with xLSTM linear RNN to address the trade-off between long-context memory and short-context reasoning;
Shortcut Diffusion Training with Cumulative Consistency Loss: An Optimal Control View
Paribesh Regmi (Rochester Institute of Technology), Rui Li (Rochester Institute of Technology)
GenerationOptimizationReinforcement LearningDiffusion modelFlow-based ModelImage
🎯 What it does: The paper proposes Cumulative Self-Consistency Loss (CSL), improving diffusion/flow matching generative models with a control-theoretic perspective, enhancing few-step generation quality, and linking it to reinforcement learning.
Should We Still Pretrain Encoders with Masked Language Modeling?
Hippolyte Gisserot-Boukhlef (Artefact Research Center), Pierre Colombo (Cohere)
ClassificationRetrievalComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Compared the effectiveness of using masked language models (MLM) and causal language models (CLM) in encoder pre-training, and proposed two-stage (CLM+MLM) and continuous pre-training (CPT) strategies
Shrinking Proteins with Diffusion
Ethan Baron (New York University), Andrew Gordon Wilson
CompressionTransformerLarge Language ModelDiffusion modelSequentialBiomedical Data
🎯 What it does: Propose a new discrete diffusion model called SCISOR, specifically designed to learn compression of protein sequences by deleting letters while preserving their natural sequence features.
Shuffle-R1: Efficient RL framework for Multimodal Large Language Models via Data-centric Dynamic Shuffle
Linghao Zhu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Data-Centric LearningLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose the Shuffle-R1 framework, which leverages a dynamic data prioritization strategy to improve the efficiency of reinforcement learning fine-tuning for multimodal large language models.
Shuffling the Data, Extrapolating the Step: Sharper Bias In Constant Step-Size SGD
Konstantinos Emmanouilidis (University of Wisconsin-Madison), Rene Vidal
GenerationOptimizationReinforcement Learning
🎯 What it does: This paper proposes a new algorithmic framework that combines stochastic reweighting with Richardson-Romberg extrapolation to improve convergence and reduce bias in non-monotone variational inequalities (VIPs).
Si-GT: Fast Interconnect Signal Integrity Analysis for Integrated Circuit Design via Graph Transformers
Yuting Hu (University at Buffalo), Jinjun Xiong (University at Buffalo)
Computational EfficiencyGraph Neural NetworkTransformerMeshGraphPhysics Related
🎯 What it does: Proposed a Si-GT model based on graph transformer for fast and accurate prediction of crosstalk delay and noise in integrated circuit (IC) interconnects;
SigLIP-HD by Fine-to-Coarse Supervision
Lihe Yang (University of Hong Kong), Hengshuang Zhao (Shanghai AI Laboratory)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Through 'fine-to-coarse' supervision, fine-grained features from high-resolution images are migrated to standard-resolution visual encoders, enhancing the visual perception capabilities of multimodal LLMs.
SIGMA-Gen: Structure and Identity Guided Multi-Subject Assembly for Image Generation
Oindrila Saha (University of Massachusetts Amherst), Matheus Gadelha (Adobe Research)
SegmentationGenerationData SynthesisDepth EstimationTransformerSupervised Fine-TuningDiffusion modelImage
🎯 What it does: Developed a unified framework SIGMA-GEN that can generate multi-agent images in a single inference while maintaining the identity and layout of each agent through structural and identity instructions.
SigmaDock: Untwisting Molecular Docking with Fragment-Based SE(3) Diffusion
Alvaro Prat (University of Oxford), Garrett M Morris (University of Oxford)
Drug DiscoveryGraph Neural NetworkDiffusion modelScore-based ModelBiomedical Data
🎯 What it does: Proposes SIGMADOCK, a molecular docking method based on SE(3) m Riemannian diffusion, which predicts ligand binding poses by segmenting the ligand into rigid fragments and learning transformations to reassemble them in fragment space.
SIGMark: Scalable In-Generation Watermark with Blind Extraction for Video Diffusion
Xinjie zhu, Weifeng Zhang (Lenovo Research)
GenerationDiffusion modelOptical FlowVideo
🎯 What it does: Proposed a scalable blind watermark generation framework called SIGMark for modern video diffusion models, which can achieve information embedding and extraction without affecting video quality;
Sign-SGD via Parameter-Free Optimization
Daniil Medyakov (Basic Research of Artificial Intelligence Laboratory), Aleksandr Beznosikov (Basic Research of Artificial Intelligence Laboratory)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: Designed and implemented a SIGN-SGD optimizer (ALIAS) without manual learning rate tuning, along with its memory-efficient and momentum-enabled variants, and validated its effectiveness in training large-scale language and vision models.
Signal in the Noise: Polysemantic Interference Transfers and Predicts Cross-Model Influence
Bofan Gong (Independent Scholar), Dawn Song (UC Berkeley)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: Investigate the vulnerability of large language models under multi-semantic structures, and achieve controllable manipulation of model outputs through methods such as feature direction, gradient vectors, and prompt injection.
Signal Structure-Aware Gaussian Splatting for Large-Scale Scene Reconstruction
Weiyi Xue (Tongji University), Guang Chen (Tongji University)
Gaussian SplattingPoint Cloud
🎯 What it does: Proposes an adaptive training framework named SIG based on signal structure recovery, and introduces Sphere-Constrained Gaussians to constrain the spatial positions of Gaussian primitives, addressing reconstruction quality issues caused by insufficient floating-point, redundant, and low-frequency observations in large-scale scenes.
Silent Leaks: Implicit Knowledge Extraction Attack on RAG Systems
Yuhao Wang (National University Of Singapore), Jiaheng Zhang (National University Of Singapore)
Adversarial AttackTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: In Retrieval-Augmented Generation (RAG) systems, the authors propose Implicit Knowledge Extraction Attack (IKEA), which steals internal knowledge from models by constructing harmless queries without using malicious prompts or jailbreaking.
SIM-CoT: Supervised Implicit Chain-of-Thought
Xilin Wei (Fudan University), Dahua Lin (Shanghai AI Laboratory)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose SIM-CoT, a training module that introduces step-level supervision into implicit chain-of-thought, utilizing an auxiliary decoder to align each implicit token with corresponding explicit reasoning steps, thereby enhancing the stability and diversity of implicit CoT.
Sim2Real VLA: Zero-Shot Generalization of Synthesized Skills to Realistic Manipulation
Runyi Zhao (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
Data SynthesisDomain AdaptationRobotic IntelligenceSupervised Fine-TuningVision-Language-Action ModelImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: Proposed a dual-system Vision-Language-Action (VLA) framework called Sim2Real-VLA, which can achieve zero-shot transfer to real-world robotic manipulation tasks after training solely on synthetic data.
SimBench: Benchmarking the Ability of Large Language Models to Simulate Human Behaviors
Tiancheng Hu, Paul Röttger (University Of Oxford)
TransformerLarge Language ModelTextTabularBenchmark
🎯 What it does: Proposed SIMBENCH, the first large-scale, standardized benchmark to evaluate large language models (LLMs) in simulating human behavior at the group level, conducting zero-shot evaluation of 45 LLMs across 20 diverse datasets.
SiMO: Single-Modality-Operable Multimodal Collaborative Perception
Jiageng Wen (Tongji University), Hao Deng (Tongji University)
Autonomous DrivingConvolutional Neural NetworkTransformerMultimodalityPoint Cloud
🎯 What it does: Proposes the SiMO framework, enabling multimodal collaborative perception to maintain normal operation even when any unimodal sensor fails.
SimpleFold: Folding Proteins is Simpler than You Think
Yuyang Wang, Miguel Ángel Bautista
Protein Structure PredictionTransformerFlow-based ModelBiomedical Data
🎯 What it does: Propose SimpleFold, a flow-matching generative model using only generic Transformer blocks for predicting full-atom protein 3D structures from amino acid sequences.
SimpleGVR: A Simple Baseline for Latent-Cascaded Generative Video Super-Resolution
Liangbin Xie (University of Macau), Chao Dong (Chinese Academy of Sciences)
GenerationSuper ResolutionConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderOptical FlowVideoStochastic Differential Equation
🎯 What it does: Proposed and implemented a lightweight latent space video super-resolution model called SimpleGVR, which directly performs super-resolution on low-resolution latent representations generated by large text-to-video (T2V) models, forming an efficient cascaded video generation pipeline.
SimpleTIR: End-to-End Reinforcement Learning for Multi-Turn Tool-Integrated Reasoning
Zhenghai Xue (Nanyang Technological University), Bo An (TikTok)
Reinforcement Learning from Human FeedbackReinforcement LearningTextBenchmark
🎯 What it does: In multi-round tool-integrated reasoning (TIR), the SimpleTIR framework is proposed, which stabilizes and enhances reward-based zero-shot RL training by detecting and filtering trajectories containing 'void turns';
SimpleToM: Exposing the Gap between Explicit ToM Inference and Implicit ToM Application in LLMs
Yuling Gu (Allen Institute for AI), Yejin Choi (Stanford University)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes and constructs the SimpleToM dataset to evaluate the capabilities of large language models in explicit theory of mind (ToM) reasoning and implicit behavior prediction/judgment under diverse everyday scenarios.
SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning
Haozhan Li (Tsinghua University), Ning Ding (Tsinghua University)
Robotic IntelligenceReinforcement LearningVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Proposed a reinforcement learning framework called SimpleVLA-RL, which utilizes sparse rewards based solely on task completion to perform online training of Vision-Language-Action models, thereby enhancing long-term action planning and generality in data-scarce scenarios;
Simplicial Embeddings Improve Sample Efficiency in Actor–Critic Agents
Johan Obando-Ceron (Mila Quebec AI Institute), Pablo Samuel Castro (Mila Quebec AI Institute)
Representation LearningReinforcement Learning
🎯 What it does: This paper introduces Simplicial Embeddings (SEM) into fast actor-critic algorithms, enforcing block-level softmax constraints on features to project embeddings onto multiple simplices, thereby achieving sparse, bounded, and structured representations;
SimuHome: A Temporal- and Environment-Aware Benchmark for Smart Home LLM Agents
Gyuhyeon Seo (Seoul National University), Yohan Jo (Seoul National University)
Large Language ModelWorld ModelTextBenchmark
🎯 What it does: Developed SimuHome—a real-time accelerated smart home simulator based on the Matter protocol and a multi-task evaluation benchmark with 600 sentences;
Simulation to Rules: A Dual-VLM Framework for Formal Visual Planning
Yilun Hao (Massachusetts Institute Of Technology), Yang Zhang (Massachusetts Institute Of Technology-Ibm Watson Ai Lab)
Robotic IntelligenceAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageSequential
🎯 What it does: Proposed a Dual-VLM framework (VLMFP) that can automatically generate PDDL domain files and problem files from visual inputs, achieving long-horizon visual planning;
SimULi: Real-Time LiDAR and Camera Simulation with Unscented Transforms
Haithem Turki (NVIDIA), Riccardo de Lutio (NVIDIA)
Autonomous DrivingNeural Radiance FieldGaussian SplattingImagePoint Cloud
🎯 What it does: Propose SimULi, which can real-time render any camera model and LiDAR, and supports time-related effects such as rolling shutter.
SiNGER: A Clearer Voice Distills Vision Transformers Further
Geunhyeok Yu (Kyung Hee University), Hyoseok Hwang (Kyung Hee University)
Knowledge DistillationTransformerImage
🎯 What it does: This paper proposes a knowledge distillation framework based on zero-space guided energy redistribution (SiNGER), addressing the high norm artifact problem in visual Transformers. It applies low-rank perturbations only in the zero space of the next layer on teacher features using a lightweight LoRA adapter to eliminate artifacts while preserving useful information.
Single Index Bandits: Generalized Linear Contextual Bandits with Unknown Reward Functions
Yue Kang (Microsoft), Yao Li (National University of Singapore)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a Single Exponential Bandit (SIB) model to study contextual bandit tasks under unknown reward functions, addressing the limitation of traditional GLB that requires a known link function.
Single-Loop Byzantine-Resilient Federated Bilevel Optimization
Yangnan Li (Hong Kong University of Science and Technology), Xuanyu Cao (Washington State University)
OptimizationFederated LearningImage
🎯 What it does: Proposed a class of single-loop Byzantine-robust federated bi-level optimization algorithms (BR-FedBi and its variants BR-FedBiM, BR-FedBiP) to address the Byzantine attack problem in bi-level optimization.
Single-stream Policy Optimization
Zhongwen Xu (Tencent), Zihan Ding (Tencent)
OptimizationLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose a single-stream policy optimization (SPO) method for implementing reward-based reinforcement learning in large language models (LLMs);
Singleton-Optimized Conformal Prediction
Tao Wang (University of Pennsylvania), Edgar Dobriban (University of Pennsylvania)
OptimizationExplainability and InterpretabilityComputational EfficiencyImageTextBiomedical Data
🎯 What it does: Proposed a conformal prediction method called SOCOP (Singleton-Optimized Conformal Prediction), which balances high singleton prediction frequency and small average prediction set size by optimizing a linear combination of single prediction probability and prediction set length.
SinkTrack: Attention Sink based Context Anchoring for Large Language Models
Xu Liu (Zhejiang University), Wenguan Wang (Zhejiang University)
TransformerLarge Language ModelTextMultimodalityChain-of-Thought
🎯 What it does: Leverages the inherent attention sink property of LLMs, using the first token (<BOS>) as an information anchor, and mitigates hallucinations and context forgetting by injecting critical information into its representation;
SIPDO: Closed-Loop Prompt Optimization via Synthetic Data Feedback
Yaoning Yu (University Of Illinois At Urbana Champaign), Haohan Wang (University Of Illinois At Urbana Champaign)
Data SynthesisOptimizationData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the SIPDO framework, which utilizes a synthetic data feedback loop to achieve self-improvement of prompts.
SK2Decompile: LLM-based Two-Phase Binary Decompilation from Skeleton to Skin
Hanzhuo Tan (Southern University of Science and Technology), Yuqun Zhang (Southern University of Science and Technology)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a two-stage framework based on large language models for binary decompilation, first generating intermediate representation (IR) via a structural recovery model, then producing readable source code through an identifier naming model;
SketchEvo: Leveraging Drawing Dynamics for Enhanced Image Synthesis
Zhixin Feng (Beijing University of Posts and Telecommunications), Yi-Zhe Song (Beijing University of Posts and Telecommunications)
GenerationSupervised Fine-TuningReinforcement LearningDiffusion modelImageTextSequential
🎯 What it does: Propose the SketchEvo framework, which utilizes sketch sequences during the painting process combined with preference learning to generate images that maintain sketch structure while enhancing aesthetic quality.
SketchingReality: From Freehand Scene Sketches to Photorealistic Images
Ahmed Bourouis (University of Surrey), Yulia Gryaditskaya (University of Surrey)
GenerationConvolutional Neural NetworkVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose a model based on stable diffusion that can generate high-quality photorealistic images from freehand scene sketches and text prompts.
SketchThinker-R1: Towards Efficient Sketch-Style Reasoning in Large Multimodal Models
Ruiyang Zhang (University of Macau), Zhedong Zheng (University of Macau)
Computational EfficiencyLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the SketchThinker-R1 framework, which employs a three-phase training process (Sketch-Mode Cold Start, SketchJudge Reward Model, Sketch-Thinking RL) to enable large-scale multimodal models to generate concise sketch-style reasoning, significantly reducing token consumption.
Skill Learning via Policy Diversity Yields Identifiable Representations for Reinforcement Learning
Patrik Reizinger (Max Planck Institute for Intelligent Systems), Wieland Brendel (Max Planck Institute for Intelligent Systems)
Reinforcement LearningContrastive LearningImage
🎯 What it does: This paper proves that under the Mutual Information Skill Learning (MISL) framework, Contrastive Successor Features (CSF) learned features can reconstruct the environment's true state under linear transformations, completing identifiability analysis;
SkillFactory: Self-Distillation for Learning Cognitive Behaviors
Zayne Rea Sprague (New York University), Greg Durrett (New York University)
Knowledge DistillationTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Restructure the model's self-generated reasoning trajectories into silver supervision data incorporating verification and retry behaviors. First, preheat the model using supervised fine-tuning (SFT), then further enhance reasoning capabilities through reinforcement learning (RL).
Skirting Additive Error Barriers for Private Turnstile Streams
Anders Aamand (BARC University of Copenhagen), Sandeep Silwal (University of Wisconsin-Madison)
Safty and Privacy
🎯 What it does: Proposes a differential privacy algorithm for continuously releasing discrete element counts and F2 moments in a turnstile stream, achieving polynomial logarithmic additive and multiplicative errors under both strict and general models;
SkyEvents: A Large-Scale Event-enhanced UAV Dataset for Robust 3D Scene Reconstruction
Wenzong Ma (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
Pose EstimationDepth EstimationGaussian SplattingImageMultimodalityPoint CloudBenchmark
🎯 What it does: Released SkyEvents, a large-scale, event-enhanced UAV dataset integrating synchronized RGB, event, LiDAR, and precise 6-DoF poses for city-scale 3D reconstruction.
Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy
Chris Yuhao Liu (2050 Research, Skywork AI), Yang Liu (2050 Research, Skywork AI)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed SynPref-40M with 40 million preference pairs, achieving high-quality filtering through a two-stage annotation pipeline combining human and machine collaboration; subsequently trained the Skywork-Reward-V2 series of reward models (0.6B–8B) using this data.
SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse–Linear Attention
Jintao Zhang (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationTransformerDiffusion modelImageVideo
🎯 What it does: Propose an SLA scheme that categorizes the attention of diffusion Transformers into three types: key, edge, and ignorable. FlashAttention is used to compute key blocks, linear attention processes edge blocks, and ignorable blocks are skipped, significantly reducing O(N²) computation.
SLAP: Shortcut Learning for Abstract Planning
Y. Isabel Liu (Princeton University), Tom Silver (Princeton University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: In existing task and motion planning (TAMP) frameworks, reinforcement learning is introduced to automatically discover new low-level options (shortcuts), using these shortcuts to compress paths in the abstract planning graph, thereby significantly improving the execution efficiency of long-horizon sparse-reward tasks.
sleep2vec: Unified Cross-Modal Alignment for Heterogeneous Nocturnal Biosignals
Weixuan Yuan (Five Seasons Medical), Xuesong Chen (Five Seasons Medical)
Domain AdaptationRepresentation LearningTransformerContrastive LearningMultimodalityTime SeriesBiomedical Data
🎯 What it does: Designed and trained the Sleep2vec foundation model, leveraging cross-modal contrastive learning to unify up to nine PSG signals into a shared representation space, achieving robust inference across different devices and missing sensors.
Slicing Wasserstein over Wasserstein via Functional Optimal Transport
Moritz Piening (Technische Universität Berlin), Robert Beinert (Technische Universität Berlin)
ClassificationOptimizationComputational EfficiencyRepresentation LearningImagePoint Cloud
🎯 What it does: Proposed a slice Wasserstein (DSW) metric based on Banach space to replace the computationally expensive Wasserstein over Wasserstein (WoW) distance.
SliderQuant: Accurate Post-Training Quantization for LLMs
Shigeng Wang (Intel Labs China), Anbang Yao (Intel Labs China)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes the SliderQuant framework, which achieves post-training quantization for high-precision LLMs by reducing quantization errors through adaptive sliding windows for layer-wise optimization.
SLM-MUX: Orchestrating Small Language Models for Reasoning
Chenyu Wang (Harvard University), Yilun Du (Harvard University)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes SLM-MUX, a discussion-free, multi-model collaboration architecture that selects the optimal output based on self-consistency, efficiently combining multiple small language models (SLMs) to enhance reasoning accuracy.
SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks
Seungwon Jeong (Dongguk University), Woojin Lee (Dongguk University)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Addressing jailbreaking attacks on LLMs, the study investigates the importance of insertion positions (slots) and proposes the SlotGCG method
Slow-Fast Policy Optimization: Reposition-Before-Update for LLM Reasoning
Ziyan Wang (Georgia Institute of Technology), Xiaoming Huo (Georgia Institute of Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a three-stage update framework SFPO (Slow-Fast Policy Optimization), which stabilizes the training process of RL in LLM inference tasks by performing multiple internal gradient updates (fast trajectory) on the same batch, interpolating back to the starting point (relocalization), and a single slow correction.
Small Drafts, Big Verdict: Information-Intensive Visual Reasoning via Speculation
Yuhan Liu, Shenji Wan
Computational EfficiencyLarge Language ModelMixture of ExpertsVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose a training-free Speculative Verdict framework, which leverages a lightweight Vision-Language Model (VLM) as a draft expert to generate multiple reasoning paths, followed by a large VLM synthesizing the final answer in one step;
Small Transformers Don’t Need LayerNorm at Inference Time: Scaling LayerNorm Removal to GPT-2 XL and Implications for Mechanistic Interpretability
Luca Baroni (Charles University), Stefan Heimersheim (Apollo Research)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: In the pre-trained GPT-2 series models, the authors replace LayerNorm (LN) layer-by-layer with linearized FakeLN and perform fine-tuning, resulting in a GPT-2 model without LN, and evaluate its performance in reasoning, language understanding, and interpretability.
SMAN-Bench: A Cross-System Benchmark for Mobile Agents under Single- and Multi-path, Ambiguous, and Noisy Tasks
Weikai Xu (Nanyang Technological University), Bo An (Tsinghua University)
Vision Language ModelGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the SMAN-Bench cross-system mobile agent benchmark, covering single-path, multi-path, ambiguous, and noisy tasks, and implemented offline multi-path evaluation along with noise/ambiguous subsets.
SmartChunk Retrieval: Query-Aware Chunk Compression with Planning for Efficient Document RAG
Xuechen Zhang (University of Michigan), Nedim Lipka (Adobe Research)
RetrievalCompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes SmartChunk Retrieval, combining a query-aware chunk planner and a compression encoder to dynamically select document chunk levels to improve retrieval accuracy and efficiency for long document question answering.