ICLR 2026 Papers — Page 42
International Conference on Learning Representations · 5356 papers
Scaling Agent Learning via Experience Synthesis
Zhaorun Chen (Meta Superintelligence Labs), Dat Huynh (Meta Superintelligence Labs)
Large Language ModelReinforcement LearningAgentic AIWorld ModelTextSequentialChain-of-Thought
🎯 What it does: Built the DREAMGYM framework, which leverages an inferential experience model to synthesize scalable interactive experiences in an abstract text space, supporting reinforcement learning (RL) training for large language model (LLM) agents, and achieved complete online experience synthesis, replay buffer, and adaptive task generation.
Scaling Agents via Continual Pre-training
Liangcai Su (The University of Hong Kong), Jingren Zhou (Alibaba Tongyi Lab)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: Propose Agentic Continual Pre-Training and build the AgentFounder research agent model based on Qwen3.
Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute
Kieran Didi (NVIDIA), Karsten Kreis (NVIDIA)
Drug DiscoveryProtein Structure PredictionTransformerReinforcement LearningFlow-based ModelBiomedical Data
🎯 What it does: Proposes Prote'ina-Complexa, a full-atom protein ligand generation framework that combines generative models with computational expansion during inference, enabling efficient ligand design without sequence redesign.
Scaling Attention via Feature Sparsity
Yan Xie (Xidian University), Yifei Wang (Amazon AGI SF Lab)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose sparse feature attention (SFA) and implement the FlashSFA kernel, enabling Transformers to maintain high-dimensional representation capabilities while reducing self-attention computation from O(n d²) to O(n k / d²), significantly lowering KV-cache and FLOPs.
Scaling Behavior of Discrete Diffusion Language Models
Dimitri von Rütte, Antonio Orvieto (ETH Zurich)
GenerationTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Study and quantify the scaling behavior of discrete diffusion language models (DLMs) under different noise types (masking, uniform, hybrid), focusing on key hyperparameters such as batch size and learning rate;
Scaling Direct Feedback Learning with Jacobian Alignment Guarantees
Paul Caillon (Université Paris Dauphine-PSL), Alexandre Allauzen (Université Paris Dauphine-PSL)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: Propose a novel feedback alignment method called GrAPE, which estimates the Jacobian using the forward mode JVP and combines it with sparse BP calibration to achieve layer-parallel training.
Scaling Generalist Data-Analytic Agents
Shuofei Qiao (Zhejiang University), Huajun Chen (Zhejiang University)
Data SynthesisLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextTabular
🎯 What it does: Developed DATAMIND, a scalable data synthesis and training framework for building general-purpose data analysis agents, generating the DATAMIND-12K dataset, and subsequently training DATAMIND-7B/14B models;
Scaling Goal-conditioned Reinforcement Learning with Multistep Quasimetric Distances
Bill Zheng (University Of California Berkeley), Sergey Levine (University Of California Berkeley)
Reinforcement Learning
🎯 What it does: Proposes the Multistep Quasimetric Estimation (MQE) method, which jointly learns the target distance using multi-step Monte-Carlo returns and a quasimetric network, achieving long-term goal accomplishment and task concatenation in offline goal-conditioned reinforcement learning.
Scaling Group Inference for Diverse and High-Quality Generation
Gaurav Parmar (Carnegie Mellon University), Jun-Yan Zhu (Carnegie Mellon University)
GenerationOptimizationVision Language ModelImageText
🎯 What it does: Propose a scalable group inference method that selects diverse and high-quality K outputs from a large set of candidate images via quadratic integer programming, enhancing the diversity and quality of generative models.
Scaling Knowledge Editing in LLMs to 100,000 Facts with Neural KV Database
Weizhi Fei (Tsinghua University), Xueyan Niu (Huawei Technologies Co., Ltd.)
TransformerLarge Language ModelContrastive LearningText
🎯 What it does: Built NeuralDB, a knowledge editing framework based on neural key-value (KV) databases, which can efficiently edit tens of thousands or even hundreds of thousands of facts within large language models (LLMs) while maintaining the model's original general capabilities.
Scaling Knowledge Graph Construction through Synthetic Data Generation and Distillation
Prafulla Kumar Choubey (Salesforce Research), Chien-Sheng Wu (Salesforce Research)
Data SynthesisKnowledge DistillationLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper proposes the SynthKG data synthesis pipeline and the Distill-SynthKG model for high-quality document-level knowledge graph (KG) construction, and builds a KG evaluation and retrieval framework based on this.
Scaling Large Vision-Language Model RL Training via Efficient Load Balancing
Zerui Wang (Shanghai Jiao Tong University), Dahua Lin (Shanghai Jiao Tong University)
Computational EfficiencyReinforcement LearningVision Language ModelImageVideoTextMultimodality
🎯 What it does: Propose the FlexRL system, addressing data loading and computation/memory load imbalance issues in VLM RL training through two components: ShadowLoader and FlexUlysses.
Scaling Laws and Spectra of Shallow Neural Networks in the Feature Learning Regime
Leonardo Defilippis (Ecole Normale Superieure), Florent Krzakala (Ecole Polytechnique Federale De Lausanne)
OptimizationRepresentation Learning
🎯 What it does: Studied the scaling laws and weight spectrum distributions of shallow (two-layer) neural networks in feature learning regimes, and provided a complete phase diagram.
Scaling Laws and Symmetry, Evidence from Neural Force Fields
Khang Ngo (Mila - Quebec AI Institute), Siamak Ravanbakhsh (Mila - Quebec AI Institute)
Drug DiscoveryGraph Neural NetworkBiomedical DataPhysics Related
🎯 What it does: Conduct large-scale experiments on neural network models for molecular interatomic potential energy, systematically measuring the power-law scaling relationships in three dimensions (computational cost, data size, model scale) across different equivariance architectures (non-equivariant, low-order equivariant, high-order equivariant).
Scaling Laws for Diffusion Transformers
Zhengyang Liang (University of Toronto), Bo Dai (University of Hong Kong)
GenerationComputational EfficiencyTransformerDiffusion modelRectified FlowImageTextMultimodality
🎯 What it does: Studied and empirically verified the scaling laws of diffusion Transformers (DiT) under different computational budgets (1e17~6e18 FLOPs), derived power-law relationships between computational budget, model parameters, data volume, and pre-training loss, and used this law to predict model loss and FID under a larger budget (1.5e21 FLOPs). Further validated the scaling consistency on out-of-distribution data and compared the scaling exponents of two Transformer architectures (In-Context vs Cross-Attention).
Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs
Song Bian (University of Wisconsin Madison), Youngsuk Park (Amazon Web Services)
Computational EfficiencyNeural Architecture SearchTransformerLarge Language ModelText
🎯 What it does: Proposed a conditional scaling law that combines model architecture parameters (hidden layer size, MLP-attention ratio, grouped query attention) with the traditional Chinchilla scaling law, enabling prediction of model accuracy and inference efficiency under fixed parameter and training token budgets, and designed a search framework to identify both efficient and accurate LLM architectures.
Scaling Laws of SignSGD in Linear Regression: When Does It Outperform SGD?
Jihwan Kim (Seoul National University), Chulhee Yun (KAIST)
OptimizationComputational EfficiencyTabularOrdinary Differential Equation
🎯 What it does: Under the Power-law random feature (PLRF) model, derive the scaling law of SignSGD and solve for the optimal computational scale and learning rate given a FLOPS budget.
Scaling Laws Revisited: Modeling the Role of Data Quality in Language Model Pretraining
Anirudh Subramanyam (University of Chicago), Robert L. Grossman (University of Chicago)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a dimensionless data quality parameter Q, incorporated a quality term into the Chinchilla scaling law to derive a quality-aware scaling law L(N,D,Q)=A N^α + B (D Q)^βγ + E, and systematically validated it by injecting synthetic noise into NMT and CLM.
Scaling Linear Attention Capacity with Sparse State Expansion
Yuqi Pan (ByteDance Seed), Guoqi Li (Institute of Automation Chinese Academy of Sciences)
RetrievalComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Improve Transformer's long-sequence modeling by proposing row sparse updates and sparse state expansion (SSE)
Scaling Multi-Task Bayesian Optimization with Large Language Models
Yimeng Zeng (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
OptimizationHyperparameter SearchLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: In multi-task Bayesian optimization, the BOLT (Bayesian Optimization with LLM Transfer) strategy is proposed, leveraging large language models (LLMs) to provide high-quality candidate solutions only during the initialization phase, thereby accelerating optimization for new tasks;
Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models
Zhaoyi Li (University of Science and Technology of China), Ying Wei (Zhejiang University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Systematically analyze error patterns in LLMs during reasoning jump generalization, and propose a method to correct reasoning errors during testing by dynamically eliminating error-handling heads (TCR)
Scaling Sequence-to-Sequence Generative Neural Rendering
Shikun Liu (Meta AI), Juan-Manuel Perez-Rua (Meta AI)
GenerationData SynthesisTransformerRectified FlowImageVideo
🎯 What it does: Proposes Kaleido, a general generative model that treats 3D rendering as a sequence-to-sequence task, enabling fully free-viewpoint synthesis between any number of reference views and target views without explicit 3D representations.
Scaling Speech Tokenizers with Diffusion Autoencoders
Yuancheng Wang (Meta Superintelligence Labs), Xubo Liu (Meta Superintelligence Labs)
RecognitionCompressionTransformerDiffusion modelAuto EncoderAudio
🎯 What it does: Proposed SiTok speech tokenizer, which utilizes diffusion autoencoder to achieve end-to-end quantization and reconstruction, and enhances representation quality through semantic regularization, achieving extreme compression at low bitrates.
Scaling Synthetic Task Generation for Agents via Exploration
Ram Ramrakhya, Alexander T Toshev
Data SynthesisTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextMultimodality
🎯 What it does: Propose the AUTOPLAY pipeline, which leverages multimodal large language models to actively explore interactive environments and generate diverse, executable, and verifiable tasks, thereby training UI agents.
Scaling up Memory for Robotic Control via Experience Retrieval
Ajay Sridhar (Stanford University), Chelsea Finn (Stanford University)
RetrievalRobotic IntelligenceRecurrent Neural NetworkLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelVideoMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes the MemER framework, enabling robots to memorize and retrieve key information through hierarchical strategies in long-sequence tasks, achieving real-time control with multi-minute memory.
Scaling Up, Speeding Up: A Benchmark of Speculative Decoding for Efficient LLM Test-Time Scaling
Shengyin Sun (City University of Hong Kong), Chen Ma (Huawei Technologies)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Built and evaluated the first speculative decoding benchmark specifically designed for scaling scenarios during LLM testing (BoN and multi-round reasoning).
Scaling with Collapse: Efficient and Predictable Training of LLM Families
Shane Bergsma (Cerebras Systems), Joel Hestness (Cerebras Systems)
Computational EfficiencyHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: This paper investigates whether the normalized training loss curve (TLC) during large language model training converges to the same curve across different scales, and proves that normalized TLC can achieve convergence within the 100M-3.9B parameter range when maintaining AdamW time scale τ, tokens-per-parameter (TPP), and learning rate scheduling consistency.
ScalingCache: Extreme Acceleration of DiTs through Difference Scaling and Dynamic Interval Caching
Lihui Gu (Zhejiang University), Yuliang Liu (KlingAI Research)
GenerationTransformerDiffusion modelImageVideoTextBenchmark
🎯 What it does: Proposed a no-training acceleration framework called ScalingCache for Diffusion Transformers, achieving extreme acceleration through differential scaling and dynamic caching.
scDFM: Distributional Flow Matching Model for Robust Single-Cell Perturbation Prediction
Chenglei Yu (Zhejiang University), Tailin Wu (Westlake University)
Drug DiscoveryTransformerFlow-based ModelBiomedical Data
🎯 What it does: scDFM proposes a distributed generative framework based on conditional flow matching and MMD regularization for predicting transcriptomes after single-cell perturbation.
SceneCOT: Eliciting Grounded Chain-of-Thought Reasoning in 3D Scenes
Xiongkun Linghu (State Key Laboratory of General Artificial Intelligence), Siyuan Huang (State Key Laboratory of General Artificial Intelligence)
Explainability and InterpretabilityTransformerSupervised Fine-TuningPrompt EngineeringMixture of ExpertsVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the SCENECOT framework, which employs chain-of-thought (CoT) based on 3D scenes to achieve step-by-step traceable question-answering;
SceneStreamer: Continuous Scenario Generation as Next Token Group Prediction
Zhenghao Peng (University of California Los Angeles), Bolei Zhou (University of California Los Angeles)
GenerationAutonomous DrivingTransformerLarge Language Model
🎯 What it does: Propose SceneStreamer, a unified autoregressive framework that represents the entire traffic scene with a discrete token sequence, enabling continuous scene generation and dynamic agent injection;
Scenethesis: A Language and Vision Agentic Framework for 3D Scene Generation
Lu Ling (NVIDIA Research), Max Li
GenerationTransformerLarge Language ModelAgentic AIVision Language ModelImageTextMeshRetrieval-Augmented Generation
🎯 What it does: Proposes the Scenethesis framework, achieving interactive 3D scene generation based on text, balancing layout reasonableness and physical feasibility.
SceneTransporter: Optimal Transport-Guided Compositional Latent Diffusion for Single-Image Structured 3D Scene Generation
Ling Wang (Xi'an Research Institute of Hi-Tech), Yikai Wang (Beijing Normal University)
Image TranslationGenerationData SynthesisTransformerDiffusion modelAuto EncoderImageMesh
🎯 What it does: Proposed the SceneTransporter framework to generate structured 3D scenes from a single image.
Scheduling Your LLM Reinforcement Learning with Reasoning Trees
Hong Wang (Tencent), Jiawei Chen (Zhejiang University)
Large Language ModelReinforcement LearningText
🎯 What it does: Propose an r-score learning difficulty evaluation metric based on the reasoning tree structure, and design the Re-Schedule data scheduling algorithm to improve the inference performance of LLMs in RLVR.
SCI-Verifier: Scientific Verifier with Thinking
Shenghe Zheng (Shanghai Artificial Intelligence Laboratory), Peng Ye (Shanghai Artificial Intelligence Laboratory)
Supervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Constructed an interdisciplinary scientific verification benchmark SCI-VerifyBench and designed a reasoning-based scientific verifier SCI-Verifier;
Sci2Pol: Evaluating and Fine-tuning LLMs on Scientific-to-Policy Brief Generation
Weimin Wu (Northwestern University), Han Liu (Northwestern University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed Sci2Pol-Bench and Sci2Pol-Corpus, which are evaluation benchmarks and training corpora for generating policy briefs from scientific papers, respectively, and conducted fine-grained evaluations of LLMs through a five-stage writing process.
ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows
Qiushi Sun (The University of Hong Kong), Zhiyong Wu (Shanghai AI Laboratory)
AI Code AssistantLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes SCIENCEBOARD, an interactive environment based on real scientific software and a scientific task benchmark with 169 tasks, to evaluate the performance of computer-using agents in scientific research workflows.
SciNav: A General Agent Framework for Scientific Coding Tasks
TIANSHU ZHANG, Huan Sun (Ohio State University)
AI Code AssistantLarge Language ModelAgentic AIText
🎯 What it does: Proposed SciNav, an autonomous agent framework for scientific programming tasks, which efficiently explores the solution space and generates executable code under limited search budgets by leveraging relative judgment-driven Top-K tree search (TKCTS).
SciTS: Scientific Time Series Understanding and Generation with LLMs
Wen Wu (Shanghai Artificial Intelligence Laboratory), Chao Zhang (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsTime SeriesBenchmark
🎯 What it does: This paper constructs a scientific time series benchmark SciTS spanning 12 scientific fields, 7 task types, and 54,023 instances, performs zero-shot evaluation on 17 models (text LLMs, multimodal LLMs, unified time series models), and proposes a time series processing framework TimeOmni compatible with general-purpose LLMs.
SCOPED: Score–Curvature Out-of-distribution Proximity Evaluator for Diffusion
Brett Barkley (University of Texas at Austin), David Fridovich-Keil (University of Texas at Austin)
Anomaly DetectionReinforcement LearningDiffusion modelScore-based ModelImage
🎯 What it does: Proposes a fast OOD detection method called SCOPED that leverages the score-curvature ratio of diffusion models.
Score Distillation Beyond Acceleration: Generative Modeling from Corrupted Data
Yasi Zhang (University of California Los Angeles), Oscar Leong (University of Texas at Austin)
RestorationGenerationKnowledge DistillationDiffusion modelScore-based ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed the Restoration Score Distillation (RSD) framework, which learns high-fidelity generative models from only corrupted observations through a two-stage training process (pre-training with a corruption-aware diffusion model, followed by distilling it into a first-order generator).
Score-Based Density Estimation from Pairwise Comparisons
Petrus Mikkola (University of Helsinki), Arto Klami (University of Helsinki)
Diffusion modelScore-based ModelContrastive LearningMultimodalityTabular
🎯 What it does: This paper proposes a method to infer the subjective probability density of a space using contrastive learning from experts or large language models.
Score-based Greedy Search for Structure Identification of Partially Observed Causal Models
Xinshuai Dong (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Score-based ModelTabular
🎯 What it does: Propose a greedy search algorithm called LGES based on maximum likelihood scores for identifying structures in partially observed linear causal models with latent variables, and prove its asymptotic consistency under the Generalized N Factor Model condition.
SCoT: Teaching 3D-LLMs to Think Spatially with Million-scale CoT Annotations
Jinpeng Li (LIESMARS, Wuhan University), Bisheng Yang (LIESMARS, Wuhan University)
Explainability and InterpretabilityData-Centric LearningVision Language ModelMultimodalityPoint CloudChain-of-Thought
🎯 What it does: Proposed and constructed a three-tier spatial Chain-of-Thought (SCoT) dataset with a scale of 1.1M, aimed at training and evaluating the spatial reasoning capabilities of 3D large language models in three categories of tasks: perception, analysis, and planning.
SCRAPL: Scattering Transform with Random Paths for Machine Learning
Christopher Mitcheltree (Queen Mary University of London), Mathieu Lagrange (Ecole Centrale Nantes)
RestorationGenerationData SynthesisAudio
🎯 What it does: Developed a random path scattering transform (Scattering Transform) approximation algorithm called SCRAPL, which utilizes random sampling and path-specific optimization methods to achieve a differentiable, computationally efficient scattering transform loss for unsupervised training in audio inverse problems.
SCRIBES: Web-Scale Script-Based Semi-Structured Data Extraction with Reinforcement Learning
Shicheng Liu (Stanford University), Xin Luna Dong
Data-Centric LearningAI Code AssistantLarge Language ModelReinforcement LearningText
🎯 What it does: Developed a reinforcement learning-based framework called SCRIBES to generate generalizable extraction scripts for efficiently extracting semi-structured data at web scale.
SCUBA: Salesforce Computer Use Benchmark
Yutong Dai (Salesforce Research), Ran Xu (Salesforce Research)
TransformerLarge Language ModelAgentic AIVision Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the SCUBA benchmark for evaluating computer usage agents in CRM workflows on the Salesforce platform.
Sculpting Subspaces: Constrained Full Fine-Tuning in LLMs for Continual Learning
Nikhil Shivakumar Nayak (Red Hat AI Innovation), Akash Srivastava (Core AI, IBM)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose an Orthogonal Subspace Fine-Tuning (OSFT) method, which in continuous learning of large language models identifies and freezes critical high-rank subspaces via adaptive SVD, performing orthogonal gradient updates only in low-rank subspaces, balancing knowledge preservation and adaptability.
Sculptor: Empowering LLMs with Cognitive Agency via Active Context Management
Mo Li (Tsinghua University), Yunxin Liu (Peking University)
Computational EfficiencyLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Proposed the Active Context Management (ACM) framework Sculptor, which utilizes a toolchain to help large language models proactively manage working memory to mitigate interference issues in long-term contexts.
SDErasure: Concept-Specific Trajectory Shifting for Concept Erasure via Adaptive Diffusion Classifier
Fengyuan Miao (University Of Science And Technology Of China), Hongtao Xie (University Of Science And Technology Of China)
GenerationDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: Propose a concept elimination framework named SDErasure, which utilizes adaptive step size selection, score rematching, and quality regulation to achieve precise, minimal-intervention elimination of specific concepts in diffusion models while maintaining generation quality.
SE-Diff: Simulator and Experience Enhanced Diffusion Model for Comprehensive ECG Generation
Xiaoda Wang (Emory University), Carl Yang (Emory University)
GenerationData SynthesisDiffusion modelTime SeriesBiomedical DataElectronic Health RecordsElectrocardiogramRetrieval-Augmented GenerationOrdinary Differential Equation
🎯 What it does: Propose a conditional latent diffusion model called SE-Diff, which generates complete 10-second, 12-lead ECG signals based on natural language descriptions, and enhances the generation quality through a physiological simulator and clinical experience retrieval.
SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models
Thinh Pham (Virginia Tech), Tu Vu (Virginia Tech)
TextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the SEALQA benchmark to evaluate the reasoning capabilities of search-augmented large language models (Search-Augmented LLMs) in factual question-answering scenarios involving conflicting, noisy, or irrelevant search results. It also introduces three difficulty levels: SEAL-0 (extremely difficult), SEAL-HARD (broader challenging problems), and LONGSEAL (long-context multi-document 'needle-in-a-haystack' scenarios).
Search Arena: Analyzing Search-Augmented LLMs
Mihran Miroyan (University of California Berkeley), Joseph E. Gonzalez (University of California Berkeley)
Data-Centric LearningTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed Search Arena — a large-scale, multilingual, multi-turn search-enhanced LLM human preference dataset, and conducted a systematic analysis of user-model interactions.
Search Self-Play: Pushing the Frontier of Agent Capability without Supervision
Hongliang Lu (Alibaba), guanjunjiang
TransformerLarge Language ModelReinforcement LearningAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented an unsupervised deep search agent self-learning method called Search Self-Play (SSP), enabling LLMs to alternate between the roles of questioner and solver within the same model, and verifying the correctness of questions through Retrieval-Augmented Generation (RAG) to achieve co-evolution of agent capabilities.
Searching for Privacy Risks in LLM Agents via Simulation
Yanzhe Zhang (Georgia Tech), Diyi Yang (Stanford University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This paper systematically identifies and quantifies privacy leakage risks in multi-round dialogues between LLM agents through a framework based on simulation and alternating search, while iteratively generating robust defense strategies.
Secondary Motion-Aware 3D Clothed Gaussian Avatars from Monocular Videos
Seungeun Lee (Korea University), Gyeong-Moon Park (Korea University)
GenerationPose EstimationGraph Neural NetworkNeural Radiance FieldVideo
🎯 What it does: Reconstruct an animatable 3D Gaussian Avatar from monocular video, focusing on secondary motion modeling for loose clothing.
SecP-Tuning: Efficient Privacy-Preserving Prompt Tuning for Large Language Models via MPC
Jinglong Luo (Pengcheng Laboratory), Zenglin Xu (Fudan University)
Domain AdaptationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextFinance Related
🎯 What it does: In this paper, the authors propose SecP-Tuning, a privacy-preserving prompt tuning framework based on secure multiparty computation (MPC), specifically designed for efficiently adapting large language models (LLMs) to sensitive domains such as healthcare and finance;
Secret-Protected Evolution for Differentially Private Synthetic Text Generation
Tianze Wang (TikTok), Qiang Yan (TikTok)
Data SynthesisSafty and PrivacyLarge Language ModelText
🎯 What it does: Proposed SecPE, an evolutionary framework based on secret protection, for generating high-quality synthetic text under differential privacy levels.
Secure Inference for Diffusion Models via Unconditional Scores
Jaeyun Song (KRAFTON), Eunho Yang (Korea Advanced Institute of Science and Technology)
GenerationSafty and PrivacyDiffusion modelScore-based ModelImageText
🎯 What it does: In privacy-preserving diffusion model inference, unconditional generated scores are utilized to correct score shifts caused by relaxed polynomial approximations, significantly improving generation quality and reducing latency.
Secure Outlier-Aware Large Language Model Inference
Lifan Zhao (Shanghai Qi Zhi Institute), Zhixuan Fang (Tsinghua University)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Designed and implemented the SOAL framework, leveraging the outlier characteristics of LLM nonlinear layers to enhance MPC inference speed.
SEED-SET: Scalable Evolving Experimental Design for System-level Ethical Testing
Anjali Parashar (Laboratory for Information and Decision Systems (LIDS) MIT), Chuchu Fan (Laboratory for Information and Decision Systems (LIDS) MIT)
Large Language ModelPhysics Related
🎯 What it does: Propose the SEED-SET framework, which utilizes sample-limited Bayesian experimental design to perform system-level ethical testing on autonomous systems, integrating objective system metrics with subjective value judgments.
SEED: Towards More Accurate Semantic Evaluation for Visual Brain Decoding
Juhyeon Park (Seoul National University), Taesup Moon (Seoul National University)
Explainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelImageTextBiomedical DataBenchmark
🎯 What it does: Proposed the SEED (Semantic Evaluation for Visual Brain Decoding) metric to better evaluate the semantic reconstruction quality of visual brain decoding models
SeeDNorm: Self-Rescaled Dynamic Normalization
Wenrui Cai (Bytedance Seed China), Qiyang Min (Bytedance Seed China)
OptimizationTransformerImageText
🎯 What it does: Designed an adaptive normalization layer called SeeDNorm, which preserves input norm information during the forward pass and adaptively adjusts gradients during the backward pass, capable of replacing RMSNorm, LayerNorm, etc.;
SeedPrints: Fingerprints Can Even Tell Which Seed Your Large Language Model Was Trained From
Yao Tong (National University of Singapore), Tianyang Hu (Chinese University of Hong Kong)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes a persistent fingerprint called SeedPrints, generated based on the model's random initialization seed, to trace the origin and inheritance of LLMs throughout the entire training process.
SeedVR2: One-Step Video Restoration via Diffusion Adversarial Post-Training
Jianyi Wang (Nanyang Technological University), Lu Jiang (ByteDance Seed)
RestorationTransformerDiffusion modelGenerative Adversarial NetworkVideo
🎯 What it does: Proposed a first-order diffusion transformer model SeedVR2 for high-resolution video restoration, trained using adversarial post-training (APT) and achieved single-step generation;
Seeing Across Views: Benchmarking Spatial Reasoning of Vision-Language Models in Robotic Scenes
ZhiYuan Feng, Baining Guo (Microsoft Research Asia)
Robotic IntelligenceVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed and made public MV-RoboBench, a benchmark specifically designed to evaluate multi-view spatial reasoning capabilities in robotic manipulation scenarios, containing 1.7k questions, 8 subtasks, covering two domains: spatial understanding and robotic execution.
Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs
Zhining Liu (University of Illinois Urbana-Champaign), Hanghang Tong (Amazon)
Explainability and InterpretabilityTransformerVision Language ModelMultimodality
🎯 What it does: Systematically analyze the 'see but do not believe' phenomenon between visual attention and answer correctness in vision-language models (VLM), and propose a training-free visual evidence enhancement (VEA) method during inference;
Seeing Through Deception: Uncovering Misleading Creator Intent in Multimodal News with Vision-Language Models
Jiaying Wu (National University Of Singapore), Bryan Hooi (National University Of Singapore)
Data SynthesisAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper investigates the deception intent of creators in multi-modal news and proposes the DECEPTIONDECODED benchmark for generating image-text news with explicit deceptive intent.
Seeing Through the Brain: New Insights from Decoding Visual Stimuli with fMRI
Zheng Huang (Dartmouth College), Yujun Yan (UNC Charlotte)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose the PRISM framework, which maps fMRI signals to a structured text space and then generates visual stimuli images using text-guided diffusion models.
Seeing Through the PRISM: Compound & Controllable Restoration of Scientific Images
Rupa Kurinchi-Vendhan (Massachusetts Institute of Technology), Sara Beery (Massachusetts Institute of Technology)
RestorationPrompt EngineeringDiffusion modelContrastive LearningImageBenchmark
🎯 What it does: Proposes the PRISM framework, which can simultaneously handle multiple mixed degradations and supports expert-driven selective restoration on demand.
Seeing Through Words: Controlling Visual Retrieval Quality with Language Models
Jianglin Lu (Adobe Research), Yun Fu (Northeastern University)
RetrievalLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This work proposes a quality-controllable retrieval framework, QCQC, which enhances the quality and controllability of text-to-image retrieval by appending descriptions matching specified quality (relevance and aesthetics) to short queries using a large language model.
Seeing What’s Not There: Negation Understanding Needs More Than Training
Bhuvan Aggarwal (Honda R&D), S Divakar Bhat (Honda R&D)
ClassificationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a new method to improve the understanding of negation in vision-language models (VLMs), particularly by post-processing and correcting CLIP text embeddings, rather than relying on additional training datasets.
Seeing What’s Wrong: A Trajectory-Guided Approach to Caption Error Detection
Gabriel Afriat (Massachusetts Institute of Technology), Rahul Mazumder (IBM Research)
Anomaly DetectionExplainability and InterpretabilityLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose the TRACED framework, which detects subtitle errors by generating and analyzing image-caption trajectories, while providing explainability and error correction guidance.
Seeing, Listening, Remembering, and Reasoning: A Multimodal Agent with Long-Term Memory
Lin Long (Zhejiang University), Wei Li (Zhejiang University)
RetrievalGraph Neural NetworkLarge Language ModelReinforcement LearningVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: Proposed M3-Agent—a multimodal agent with long-term memory capable of real-time perception, constructing event and semantic memories, and completing tasks through multi-round reasoning and memory retrieval.
Seek-CAD: A Self-refined Generative Modeling for 3D Parametric CAD Using Local Inference via DeepSeek
Xueyang Li (Fudan University), Xiangdong Zhou (Fudan University)
GenerationTransformerLarge Language ModelVision Language ModelTextMeshBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes Seek-CAD, a training-free, locally-inferenced generative CAD modeling framework that combines the DeepSeek-R1-32B-Q4 LLM with retrieval-augmented generation (RAG), and achieves self-correction through step-wise visual feedback and chain-of-thought (CoT) reasoning;
Seesaw: Accelerating Training by Balancing Batch Size and Learning Rate Scheduling
Alexandru Meterez (Harvard University), Sham M. Kakade
OptimizationComputational EfficiencyHyperparameter SearchTransformerTextStochastic Differential Equation
🎯 What it does: Proposed the Seesaw scheduler, balancing learning rate decay and batch size increase to accelerate the pretraining of large language models
Segment Any Events with Language
Seungjun Lee (National University of Singapore), Gim Hee Lee (National University of Singapore)
SegmentationVision Language ModelMultimodality
🎯 What it does: Developed the SEAL framework for open-vocabulary instance segmentation (OV-EIS) in event cameras, supporting multi-granularity semantics (instance, part, semantic) and free-text language queries;
Segment-Level Attribution for Selective Learning of Long Reasoning Traces
Siyuan Wang (University of Southern California), Xiang Ren (University of Southern California)
Explainability and InterpretabilityKnowledge DistillationTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper proposes a paragraph-level importance assessment method based on integrated gradients, and applies it to implement selective fine-tuning (Selective SFT) on long-chain reasoning (CoT), thereby improving reasoning accuracy and compressing output length.
Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens
Yunlong Deng (Mohamed bin Zayed University of Artificial Intelligence), Guangyi Chen (Mohamed bin Zayed University of Artificial Intelligence)
Computational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Proposed the SR2 framework, treating reasoning tasks as causal selection mechanisms, utilizing a three-step iterative process consisting of reflective representation learning, dependency self-refinement, and periodic alignment to enhance reasoning capabilities.
Selective Expert Guidance for Effective and Diverse Exploration in Reinforcement Learning of LLMs
Zishang Jiang (Fudan University), Yanghua Xiao (Ant Group)
Large Language ModelReinforcement LearningMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes the MENTOR framework, which achieves high-quality exploration in RLVR by providing expert guidance to LLMs at critical decision points;
Selective Rotary Position Embedding
Sajad Movahedi (ELLIS Institute), Volkan Cevher (EPFL)
Representation LearningTransformerText
🎯 What it does: Proposed Selective RoPE, an input-dependent, learnable rotary position encoding to enhance the recall capability and expressiveness of linear Transformers
Self-Aligned Reward: Towards Effective and Efficient Reasoners
Peixuan Han (University of Illinois Urbana Champaign), Luyang Kong (Amazon Web Services)
Computational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: Investigated an internal reward mechanism called Self-Aligned Reward (SAR) for reinforcement learning training of large language models, enabling them to more precisely and effectively evaluate answer quality and improve reasoning efficiency in reasoning tasks.
Self-Aug: Query and Entropy Adaptive Decoding for Large Vision-Language Models
Eun Woo Im (Arizona State University), Vivek Gupta (Arizona State University)
Prompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: To address hallucination issues in large vision-language models (LVLMs), this paper proposes a training-free adaptive decoding strategy called Self-Aug, which dynamically selects visually enhanced content matching query semantics and applies entropy-based threshold filtering in contrastive decoding, significantly improving the factual consistency of generated outputs.
Self-Consistency Improves the Trustworthiness of Self-Interpretable GNNs
Wenxin Tai (University of Electronic Science and Technology of China), Fan Zhou (University of Electronic Science and Technology of China)
Explainability and InterpretabilityGraph Neural NetworkSupervised Fine-TuningGraphBenchmark
🎯 What it does: Proposes incorporating a self-consistency (SC) fine-tuning strategy into self-explaining graph neural networks (SI-GNN) to directly optimize the credibility and consistency of model explanations during training.
Self-Destructive Language Models
Yuhui Wang (Stony Brook University), Ting Wang (Stony Brook University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Under adversarial fine-tuning attacks, the proposed SEAM method transforms LLMs into self-destructing models, retaining normal task capabilities while experiencing performance collapse when subjected to harmful fine-tuning;
Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking
Wen Wen (City University of Hong Kong), Li zhang
Reinforcement LearningVision Language ModelImageText
🎯 What it does: Developed EvoQuality, a fully self-supervised visual-language model self-evolution framework that enhances image quality assessment (IQA) capabilities by leveraging internal model voting and relative ranking.
Self-Forcing++: Towards Minute-Scale High-Quality Video Generation
Justin Cui (Ucla), Cho-Jui Hsieh (Bytedance Seed)
GenerationDiffusion modelOptical FlowVideoText
🎯 What it does: Propose the Self-Forcing++ method, which leverages a short video teacher model to guide the student model in correcting errors within self-generated long videos, significantly improving the quality of long-sequence video generation.
Self-Guided Low Light Object Detection Framework
Gwangik Shin (Hanyang University), Soonmin Hwang (Hanyang University)
Object DetectionDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: Propose a self-guided low-light object detection framework that employs a detachable auxiliary pipeline (including self-supervised enhancement, denoising, and Fourier fusion) during training, but adds no modules or parameters during inference.
SELF-HARMONY: LEARNING TO HARMONIZE SELF-SUPERVISION AND SELF-PLAY IN TEST-TIME REINFORCEMENT LEARNING
Ru Wang (University of Tokyo), Jiaxian Guo (Google Research)
Large Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposed the Self-Harmony framework, which leverages a single LLM during testing by employing two roles—self-restatement (Reframer) and problem-solving (Solver)—to generate a set of answers consisting of the original question and semantically equivalent restatements. It selects stable pseudo-labels using the harmonic mean score, thereby enhancing reasoning performance in unsupervised reinforcement learning on test sets.
Self-Improving Loops for Visual Robotic Planning
Calvin Luo (Brown University), Chen Sun (Brown University)
Knowledge DistillationRobotic IntelligenceVision Language ModelDiffusion modelScore-based ModelVideoText
🎯 What it does: Propose the SILVR framework, which leverages visual planning from video generation models to enable adaptive loops, continuously improving robot performance on unseen tasks.
Self-Improving Skill Learning for Robust Skill-based Meta-Reinforcement Learning
Seungyul Han (Ulsan National Institute of Science and Technology), Yisak Park (Ulsan National Institute of Science and Technology)
Meta LearningReinforcement LearningBenchmark
🎯 What it does: Proposed the self-improving skill learning framework (SISL), achieving robust skill primitive meta reinforcement learning in long-horizon tasks with noisy offline demonstrations.
Self-Improving Vision-Language-Action Models with Data Generation via Residual RL
Wenli Xiao (NVIDIA), Yuke Zhu (Carnegie Mellon University)
Robotic IntelligenceSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelMultimodality
🎯 What it does: Achieve self-improvement of vision-language-action models through a three-stage process (residual RL expert acquisition, distribution-based hybrid trajectory collection, supervised fine-tuning);
Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training
Zheng Xin Yong, Stephen Bach
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Discovered and systematically studied the phenomenon of self-escape in reasoning language models during chain-of-thought reasoning, where models bypass their own safety defenses through internal reasoning without external attacks;
Self-Predictive Representations for Combinatorial Generalization in Behavioral Cloning
Daniel Lawson (Mila), Khimya Khetarpal (Mila)
Representation LearningContrastive LearningImageBenchmark
🎯 What it does: This paper studies how to enhance the zero-shot performance of behavior cloning in compositional generalization tasks through self-predictive representations.
Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning
Carl Qi (University of Texas Austin), Yesh Dattatreya (Amazon Robotics)
Anomaly DetectionRobotic IntelligenceVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the ARMOR model, achieving robot fault detection and open-ended reasoning using a multi-round self-refinement vision-language model.
Self-Speculative Decoding Accelerates Lossless Inference in Any-Order and Any-Subset Autoregressive Models
Gabe Guo (Stanford University), Stefano Ermon (Stanford University)
GenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed Any Subset Autoregressive Model (AS-ARM) and Arbitrary Subset Speculative Decoding (ASSD), achieving parallel efficient filling generation in any order.
Self-Speculative Masked Diffusions
Andrew Campbell (Google DeepMind), Arnaud Doucet (Google DeepMind)
GenerationTransformerLarge Language ModelDiffusion modelTextBiomedical Data
🎯 What it does: Proposed a self-predictive masked diffusion model for efficient generation of discrete data.
Self-Supervised Evolution Operator Learning for High-Dimensional Dynamical Systems
Giacomo Turri (Italian Institute of Technology), Pietro Novelli (Italian Institute of Technology)
Representation LearningConvolutional Neural NetworkGraph Neural NetworkContrastive LearningTime SeriesSequentialBiomedical DataPhysics Related
🎯 What it does: An end-to-end self-supervised learning framework that learns the evolution operator and its spectral decomposition of high-dimensional dynamical systems.
Self-Supervised Learning from Structural Invariance
Yipeng Zhang (Mila Quebec AI Institute), Laurent Charlin (Mila Quebec AI Institute)
Knowledge DistillationRepresentation LearningContrastive LearningImageVideo
🎯 What it does: Studied the one-to-many mapping problem in natural alignment data, proposing the AdaSSL method which introduces latent variables to capture conditional uncertainty, thereby improving the positive sample similarity modeling in self-supervised learning.
SelfReflect: Can LLMs Communicate Their Internal Answer Distribution?
Michael Kirchhof (Apple), Sinead Williamson (Apple)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the SelfReflect metric to evaluate whether a single text faithfully summarizes the internal answer distribution of a large language model (LLM) under a given query, and experimentally examines whether modern LLMs can naturally generate self-reflective uncertainty expressions; further, the metric is open-sourced along with benchmark evaluation code.