ICML 2026 Papers — Page 51
International Conference on Machine Learning · 6554 papers
Safe Reinforcement Learning with Preference-based Constraint Inference
Chenglin Li (Tsinghua University), Hua Geng (Massachusetts Institute of Technology)
Autonomous DrivingOptimizationReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningTextTabular
🎯 What it does: Proposes a safety reinforcement learning framework called PbCRL based on inferring human preferences to derive safety constraints, which can learn a cost function that meets real safety requirements from trajectory comparison data and drive policy learning.
SafeCompass: Dynamic Chain-of-Thought Steering via Inference-Time Safety Signals
Zeyang Zhang (Zhejiang University), Cheng Zhuo (Zhejiang University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: Propose SafeCompass, a method that dynamically monitors the internal hidden states of LRM during inference and inserts safety prompts when necessary based on real-time safety scores, thereby preventing jailbreak attacks.
SafeDec: Constrained Decoding for Safe Autoregressive Generalist Robot Navigation Policies
Parv kapoor, Eunsuk Kang (Carnegie Mellon University)
Autonomous DrivingRobotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelContrastive LearningImageSequential
🎯 What it does: Propose SafeDec, a method that enforces Signal Temporal Logic (STL) safety specifications during the reasoning phase of transformer-based robot navigation by using constrained decoding (Hard Constrained Decoding and Robustness Constrained Decoding).
Safeguarded Stochastic Polyak Step Sizes for Non-smooth Optimization: Robust Performance Without Small (Sub)Gradients
Dimitris Oikonomou (Johns Hopkins University), Nicolas Loizou (Johns Hopkins University)
OptimizationContrastive LearningImageTabularStochastic Differential Equation
🎯 What it does: Proposed Safe Polyak Step Size (Safeguarded SPS) for non-smooth stochastic subgradient methods and momentum iterations, and provided theoretical convergence guarantees;
SafeHarbor: Defining Precise Decision Boundaries via Hierarchical Memory-Augmented Guardrail for LLM Agent Safety
Zhe Liu (Beihang University), Hao Peng (Beihang University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the SAFEHARBOR framework, achieving precise safety decision boundaries for LLM agents through automated adversarial rule generation, hierarchical memory retrieval, and contrastive learning projection;
SafeLab: An Interactive High-Fidelity Benchmark for Embodied Safety in Scientific Robotics
Fengshuo Bai (Shanghai Jiao Tong University), Yuanpei Chen (PKU-PsiBot Joint Lab)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageVideoTextSequentialBenchmark
🎯 What it does: Proposed SafeLab, a high-fidelity interactive benchmark for safety in scientific laboratory robots, containing 64 tasks, 63 calibrated lab assets, and 6,400 expert demonstrations without manual operation;
SafeSci: Safety Evaluation of Large Language Models in Science Domains and Beyond
Xiangyang Zhu (Shanghai Artificial Intelligence Lab), Guangtao Zhai (Shanghai Artificial Intelligence Lab)
Safty and PrivacyDrug DiscoveryProtein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the SafeSci framework, which includes a large-scale interdisciplinary evaluation benchmark SafeSciBench with 0.25M samples and a training set SafeSciTrain with 1.5M samples, to systematically assess and enhance the safety of LLMs in the scientific domain.
SafeSearch: Automated Red-Teaming of LLM-Based Search Agents
Jianshuo Dong (Tsinghua University), Han Qiu (Tsinghua University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes and implements the SAFESEARCH framework, which automates the security evaluation of red team LLM search agents by generating test cases, injecting unreliable search results, and evaluating agent responses.
SafeSeek: Universal Attribution of Safety Circuits in Language Models
Miao Yu (University of Science and Technology of China), Qingsong Wen (Squirrel Ai Learning)
OptimizationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose the SafeSeek framework, which locates safe circuits in LLMs through differentiable binary masks and gradient optimization, and performs safety fine-tuning via SaCirT;
SafeSpec: Fast and Safe LLM via Dynamic Reflective Sampling
HAOTIAN XU, Cheng Zhuo (Zhejiang University)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed SafeSpec, a framework that integrates security protection into the inference-time acceleration process by adding a lightweight safety header to the target model and executing rollback-reflection multi-sampling when unsafe generations are detected, achieving dual optimization of safety and acceleration.
Safety Alignment of LMs via Non-cooperative Games
Anselm Paulus (Meta), Arman Zharmagambetov (Meta)
Safty and PrivacyAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelText
🎯 What it does: Propose to jointly train the attacker and defender language models through a non-zero-sum non-cooperative game to achieve safe alignment;
Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks
Guoxin Lu (Nanjing University of Posts and Telecommunications), Fu Xiao (Nanjing University of Posts and Telecommunications)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The study focuses on preventing harmful fine-tuning attacks in large language models, proposing a geometric bottleneck regularization method by setting safety anchors in the decoding layer (unembedding layer);
Safety Game: Inference-Time Alignment of Black-Box LLMs via Constrained Optimization
Tuan Nguyen (University of Warwick), Long Tran-Thanh (University of Warwick)
OptimizationSafty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes a safety alignment framework during inference based on a black-box LLM, achieving a balance between safety and helpfulness through self-probing scoring and linear programming within a limited candidate answer set;
Safety Generalization Under Distribution Shift in Safe Reinforcement Learning: A Diabetes Testbed
Minjae Kwon (University of Virginia), Lu Feng (University of Virginia)
Drug DiscoveryTransformerReinforcement LearningContrastive LearningTabularTime SeriesBiomedical DataBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Developed a unified diabetes simulation platform and a safety evaluation benchmark under distribution shift, and proposed a predictive shielding mechanism based on BA‑NODE.
Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away
Soumya Suvra Ghosal (University of Maryland), Amrit Singh Bedi (University of Central Florida)
Safty and PrivacyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodality
🎯 What it does: Propose a lightweight inference-time safety recovery method called SAFETHINK for multi-modal large reasoning models, which prevents malicious jailbreak by inserting safety prompts in early inference steps.
Safety-Efficacy Trade Off: Robustness against Data-Poisoning
Diego Marco Granziol, Ulugbek Abdimanabov (Purestrength AI)
Safty and PrivacyAdversarial AttackTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: This paper studies the impact of backdoors and data poisoning attacks on models, and establishes a theoretical framework based on kernel ridge regression to explain the geometric mechanism of the attacks;
SAGE-NAS: Synergizing LLM-Based Semantic Agent with Graph-Based Evaluator for Neural Architecture Search
Kaiqi Lin (Shenzhen University), Jianping Luo (Shenzhen University)
Neural Architecture SearchGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningAgentic AIGraph
🎯 What it does: This paper proposes the SAGE-NAS framework, which combines LLM semantic agents, bimodal graph evaluators, and behavioral maps to achieve closed-loop neural network architecture search.
SAGE: A Dataflow-Native Framework for Modular, Controllable, and Transparent LLM-Augmented Reasoning
Jun Liu (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A full-stack dataflow system called SAGE has been built for the inference pipeline of large language models (LLMs), which can compile multi-stage inference workflows (retrieval, embedding, vector search, state memory, context refinement, generation, etc.) into a physically executable graph that can be distributed. It achieves controllable optimization of tail latency and resource contention through explicit queue backpressure, pluggable scheduling/placement strategies, and a unified observation interface.
SAGE: Shaping Anchors for Guided Exploration in RLVR of LLMs
Chanuk Lee (KAIST), Sung Ju Hwang (KAIST)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a framework called SAGE that achieves guided exploration in RLVR by adjusting the anchor points of the inverse KL limiter.
SALAAD: Sparse And Low-Rank Adaptation via ADMM for Large Language Model Inference
Hao Ma (ETH Zurich), Michael Muehlebach (Max Planck Institute for Intelligent Systems)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: During the pre-training of large language models, a framework named SALAAD is proposed, which dynamically introduces sparse + low-rank structures during training via ADMM, enabling the model to achieve elastic adjustable capacity with a single training session.
SALE : Low-bit Estimation for Efficient Sparse Attention in Long-context LLM Prefilling
Xiaodong Ji (Peking University), Bin CUI
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes SALE (Low-bit Estimation for Efficient Sparse Attention), a fine-grained sparse attention method that accelerates self-attention during the pre-filling stage of long-context LLMs.
Saliency-Aware Model Merging
Jungin Park (Yonsei University), Kwanghoon Sohn (Yonsei University)
ClassificationRecommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerMixture of ExpertsAuto EncoderContrastive LearningImageText
🎯 What it does: Proposes a strictly data-free model merging method (SA-Merging), which selects and fuses parameters from multiple task experts based on significance metrics derived from network connectivity, and performs equivalent rank-level significance pruning on LoRA (Low-Rank Adapter).
SALSA-V: Shortcut-Augmented Long-form Synchronized Audio from Videos
Amir Dellali (ETH Zürich), Roger Wattenhofer (ETH Zürich)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: Proposes SALSA-V, a multimodal generative model capable of generating high-fidelity, long-duration, and synchronized audio from silent videos.
Salus: Strategic Diagnostic Testing for Complex Diagnosis via Multi-Agent Reinforcement Learning
Shuohao Gao (Tsinghua University), Ting Chen (Tsinghua University)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackProtein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Design a multi-agent framework Salus in multi-round diagnosis, separating differential diagnosis, strategy control, and worksheet proposal, and using RL to optimize the diagnostic process.
SAM Audio: Segment Anything in Audio
Bowen Shi (Meta SuperIntelligence Labs), Ann Lee (Meta SuperIntelligence Labs)
SegmentationTransformerPrompt EngineeringDiffusion modelFlow-based ModelVideoTextMultimodalityAudio
🎯 What it does: Proposed a general audio source separation foundation model, SAM Audio, which can freely separate any audio source through three types of prompts: text, visual masks, and time spans.
Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors
Neelam Akula (University of Texas at Dallas), Baris Coskunuzer (University of Texas at Dallas)
ClassificationDomain AdaptationRecommendation SystemRepresentation LearningGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraphBenchmark
🎯 What it does: Proposes a standardized evaluation protocol for cross-task transfer (node classification ↔ link prediction) on the same graph, and systematically evaluates the performance of various transfer strategies under different graph structures.
Same Question, Different Lies: Cross-Context Consistency (C³) for Black-Box Sandbagging Detection
Lin Yulong (MATS Research), Mary Phuong (Google DeepMind)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Propose a zero-shot black-box sandbag detection framework called Cross-Context Consistency (C³), used to determine whether large models intentionally understate their performance in dangerous capability assessments.
SAME: Stabilized Mixture-of-Experts for Multimodal Continual Instruction Tuning
Zhen-Hao Xie (Nanjing University), Da-Wei Zhou (Nanjing University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningTextMultimodality
🎯 What it does: Propose SAME (Stabilized Mixture-of-Experts), which reduces forgetting in multi-modal continuous instruction tuning by improving expert routing and expert update mechanisms.
Sample Complexity Bounds for Robust Mean Estimation with Mean-Shift Contamination
Ilias Diakonikolas (University of Wisconsin-Madison), Sihan Liu (University of California, San Diego)
🎯 What it does: Studied the mean estimation task under mean-shift contamination, addressing the sample complexity issue of mean estimation under general base distributions.
Sample from What You See: Visuomotor Policy Learning via Diffusion Bridge with Observation-Embedded Stochastic Differential Equation
Zhaoyang Liu (ShanghaiTech University), Ye Shi (ShanghaiTech University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerDiffusion modelContrastive LearningImageTextMultimodalityStochastic Differential Equation
🎯 What it does: A generative visual motion control strategy called BridgePolicy is proposed, which samples from an observation-priorized prior by directly embedding observations into the stochastic differential equations (SDEs) of the diffusion process.
Sample Margin-Aware Recalibration of Temperature Scaling
Haolan Guo (University of Sydney), Chang Xu (University of Sydney)
ClassificationComputational EfficiencyKnowledge DistillationTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose a lightweight post-processing calibration method called SMART, which directly maps the logit margin to sample-level temperature for temperature scaling;
Sample-Efficient Diffusion-based Reinforcement Learning with Critic Guidance
Shutong Ding (ShanghaiTech University), Ye Shi (ShanghaiTech University)
Convolutional Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelPoint CloudTabularTime Series
🎯 What it does: A critical guide-based diffusion strategy optimization framework, CGPO, is studied to improve sample efficiency and the exploration-exploitation balance in reinforcement learning.
Sampling and Identity-Testing Without Approximate Tensorization of Entropy
William Gay (Univerisity of Illinois Urbana-Champaign), Ryan O'Donnell (Carnegie Mellon University)
OptimizationComputational EfficiencyRepresentation LearningReview/Survey Paper
🎯 What it does: Theoretically studied the problem of fast mixing and identity testing for hybrid ATE distributions under data-driven initialization, and provided realizable sampling and testing algorithms.
Sampling from Your Language Model One Byte at a Time
Jonathan Hayase (University of Washington), Sewoong Oh (University of Washington)
GenerationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Proposes ByteSampler, an algorithm that transforms any autoregressive language model using a BPE tokenizer into a byte-level model during inference, addressing the Prompt Boundary Problem (PBP).
SAMT: Generating Structured Avatar Meshes and Textures from a Single Image
Muyu Wang (Wuhan University), Wenguan Wang (Zhejiang University)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelNeural Radiance FieldAuto EncoderGaussian SplattingImageMesh
🎯 What it does: Propose a two-stage framework called SAMT, which generates structured, detail-rich 3D avatar meshes from a single facial image and synthesizes texture consistent with the viewpoint.
SAOT: Self-Supervised Continual Graph Learning with Structure-Aware Optimal Transport
Yuting Zhang (Tiangong University), Xiao Wang (Beihang University)
ClassificationKnowledge DistillationRepresentation LearningGraph Neural NetworkContrastive LearningGraphBenchmark
🎯 What it does: Propose a self-supervised continual graph learning framework called SAOT, which can learn and retain graph structural knowledge in a multi-task sequence
SAQNN: Spectral Adaptive Quantum Neural Network as a Universal Approximator
Jialiang Tang (State Key Lab of Processors Institute of Computing Technology Chinese Academy of Sciences), Xiaoming Sun (State Key Lab of Processors Institute of Computing Technology Chinese Academy of Sciences)
OptimizationComputational EfficiencyRepresentation LearningDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularPhysics Related
🎯 What it does: Designed and proved a constructible quantum neural network SAQNN, demonstrating its universal approximation property for any multivariate square-integrable function, and providing an upper bound on the L2 error for functions in Sobolev spaces.
SARL: Structure-Aligned Reinforcement Learning for Bridging the Perception-Action Gap in Airspace
Binhao Gu (Southeast University), Hui Ding (Nanjing Les Information Technology)
Autonomous DrivingOptimizationGraph Neural NetworkTransformerReinforcement LearningMixture of ExpertsGraphTabular
🎯 What it does: Propose a structure-aligned multi-agent reinforcement learning framework, SARL, for conflict resolution in high-density airspace;
SARSteer: Safeguarding Large Audio Language Models via Safe-Ablated Refusal Steering
Weilin Lin (Hong Kong University of Science and Technology), Li Liu (Hong Kong University of Science and Technology)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextAudio
🎯 What it does: Proposes a reasoning-time security defense framework called SARSteer, aimed at enhancing the rejection behavior of large audio-language models (LALMs) and reducing false rejections.
SaTeen: Learning Structural Alignment for Continual Test-Time Adaptation
Chang Liu (Northwestern Polytechnical University), Yupei Zhang (Northwestern Polytechnical University)
Domain AdaptationContrastive LearningImage
🎯 What it does: Propose the SaTeen method, which achieves continual test-time adaptation by aligning the internal structure of samples and the inter-sample structure during testing.
Saving Foundation Flow-Matching Priors for Inverse Problems
Yuxiang Wan (University of Minnesota), Ju Sun (University of Minnesota)
RestorationOptimizationTransformerDiffusion modelScore-based ModelFlow-based ModelGaussian SplattingImageBiomedical DataMagnetic Resonance ImagingComputed TomographyBenchmark
🎯 What it does: Propose the FMPlug framework, which utilizes a pre-trained fundamental Flow-Matching (FM) model, solving various inverse problems through instance-guided time-dependent warm-up and sharpened Gaussian regularization.
SAW-Bench: Learning Situated Awareness in the Real World
Chuhan Li (University of California, Santa Barbara), Xin Eric Wang (University of California, Santa Barbara)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the SAW-BENCH benchmark to evaluate the situation-awareness capabilities of multimodal foundation models in real-world scenarios.
SC-FAGC: Size Constrained Fast Anchor-based Graph Clustering
Liu Jiachen
OptimizationComputational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Proposes SC-FAGC, a fast spectral clustering method on anchor graph that achieves size-limited clustering.
SC$^{2}$-WM: A Self-Correcting World Model with Closed-Loop Feedback for Vision-and-Language Navigation in Continuous Environments
Xuan Yao (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
Autonomous DrivingComputational EfficiencyRepresentation LearningRobotic IntelligenceTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelContrastive LearningWorld ModelImageTextMultimodality
🎯 What it does: Proposed a self-correcting world model, SC-WM, which achieves closed-loop decision-making through internal prediction feedback, thereby improving the performance of vision-language navigation in continuous environments.
Scalable and Differentiable Point-Cloud Registration Using Maximum Mean Discrepancy
Rixon Crane (Data61 CSIRO), Russell Tsuchida (Monash University)
Pose EstimationOptimizationComputational EfficiencyRepresentation LearningTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningOptical FlowPoint CloudStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a differentiable point cloud registration method based on maximum mean discrepancy (MMD), named MMD-Reg, which supports large-scale registration without correspondence and can be embedded into deep networks as a differentiable optimization layer
Scalable and General Whole-Body Control for Cross-Humanoid Locomotion
Yufei Xue (Shanghai Jiao Tong University), Weinan Zhang (Shanghai Jiao Tong University)
Robotic IntelligenceMeta LearningGraph Neural NetworkTransformerSupervised Fine-TuningReinforcement LearningContrastive LearningSimultaneous Localization and MappingGraphTabularTime SeriesSequential
🎯 What it does: Proposes a whole-body control strategy called XHugWBC that can achieve zero-shot transfer across multiple types of humanoid robots.
Scalable and Interpretable Representation Alignment with Ordinal Similarity
Diogo Soares (Institute of AI for Health, Helmholtz Munich), Ewa Szczurek (Institute of AI for Health, Helmholtz Munich)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive LearningImageMultimodalityTabular
🎯 What it does: This paper proposes ordinal similarity indices (TSI, QSI) based on triplets and quadruples to measure the geometric alignment between two representation spaces.
Scalable and Stable Estimation of Amari $\alpha$-Divergence using Random Fourier Features
Jiaolong Wang (Southwest University of Finance and Economics), Lingrui Wang (Southwest University of Finance and Economics)
GenerationOptimizationComputational EfficiencyGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a scalable and stable Amari α-divergence estimation framework based on random Fourier features (RFF), which restricts the kernel function within the RKHS sphere and transforms the infinite-dimensional problem into a linear model that can be optimized using mini-batch SGD through RFF.
Scalable Bayesian Inference for Nonlinear Conservation Laws
Tim Weiland (University of Tübingen), Philipp Hennig (University of Tübingen)
OptimizationComputational EfficiencyTabularReview/Survey PaperBenchmarkPhysics Related
🎯 What it does: Propose a numerical simulation and inverse problem solving method for nonlinear conservation laws by transforming the finite volume method (FVM) into a Bayesian inference framework, achieving scalability through sparse approximation.
Scalable Event Cloud Network for Event-based Classification
Hongwei Ren (Harbin Institute of Technology), Bojun Cheng (Hong Kong University of Science and Technology)
ClassificationPose EstimationComputational EfficiencySpiking Neural NetworkContrastive LearningPoint CloudTime Series
🎯 What it does: Proposed a lightweight and scalable network called SECNet based on the Event Cloud representation for classification tasks in event cameras;
Scalable GANs with Transformers
Sangeek Hyun (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Proposed and implemented an scalable Transformer-based GAN framework called GAT, which uses pure Transformer generator and discriminator in the VAE latent space, and addresses the problems of early layer death and unstable training through intermediate layer supervision and width-aware learning rate scheduling.
Scalable Kronecker-Factored Fisher Approximation for Neural Network Parameter Sensitivity
Viktoriia A. Chekalina (Fusion Brain), Evgeny Frolov (AXXX)
CompressionOptimizationKnowledge DistillationRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningGaussian SplattingTextTabular
🎯 What it does: Proposed a matrix-agnostic Fisher matrix factorization (MFF) algorithm, which approximates the Fisher information matrix using Kronecker factorization, and derived an optimal second-order weighted SVD (GFWSVD) for low-rank compression after training.
Scalable Medical Multimodal Fusion via Symmetric Consistency Modeling
Xiaowen Sun (Shandong University of Finance and Economics), Ning Mao (Qingdao University)
ClassificationRepresentation LearningGraph Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningGaussian SplattingMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyAlzheimer's DiseaseElectronic Health Records
🎯 What it does: Propose a modality-agnostic multi-modal fusion framework for medical diagnosis tasks.
Scalable Option Learning in High-Throughput Environments
Mikael Henaff (Meta Superintelligence Labs), Michael Rabbat (Meta Superintelligence Labs)
Recurrent Neural NetworkTransformerReinforcement LearningTextSequential
🎯 What it does: This paper proposes the Scalable Option Learning (SOL) framework, which can train hierarchical reinforcement learning agents in high-throughput environments with billions of frames.
Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening
Xiaotong Ji (Huawei Noah's Ark Lab), Haitham Bou Ammar (Huawei Noah's Ark Lab)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Propose a scalable power sampling method that is independent of training and validation, leveraging adaptive scaling of the base model's generated distribution to sharpen its inference capability.
Scalable Reinforcement Learning via Adaptive Batch Scaling
Jongchan Park (Hyundai Motor Company)
Reinforcement LearningImageVideoTabular
🎯 What it does: Propose Adaptive Batch Scaling (ABS), which dynamically adjusts the batch size in RL training by measuring behavioral diversity, to balance early rapid adaptation and later fine convergence.
Scalable RF Simulation in Generative 4D Worlds
Zhiwei Zheng (University of Pennsylvania), Mingmin Zhao (University of Pennsylvania)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkImageVideoTextPoint CloudMeshPhysics Related
🎯 What it does: Proposes WAVEVERSE, a scalable framework based on LLM that generates dynamic 4D indoor scenes from text prompts and simulates realistic RF signals through phase-consistent ray tracing.
Scalable Simulation-Based Model Inference with Test-Time Complexity Control
Manuel Gloeckler (University of Tübingen), Jakob H. Macke (University of Tübingen)
OptimizationFederated LearningComputational EfficiencyRepresentation LearningData-Centric LearningTransformerDiffusion modelScore-based ModelAuto EncoderTabularBiomedical DataMagnetic Resonance ImagingDiffusion Tensor Imaging
🎯 What it does: Propose the PRISM framework to achieve joint posterior inference of large-scale discrete model structures and continuous parameters, and allow control of model complexity during inference by adjusting the prior λ.
Scalable Single-Cell Gene Expression Generation with Latent Diffusion Models
Giovanni Palla (Biohub), Jakub M. Tomczak (Biohub)
GenerationData SynthesisRepresentation LearningTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderContrastive LearningTabularBiomedical Data
🎯 What it does: A single-cell gene expression generation framework called scLDM based on variational autoencoder (VAE) and latent diffusion model (LDM) was constructed, which can generate high-quality gene expression profiles without being restricted by gene order.
Scalable Topology-Preserving Graph Coarsening: Concepts and Algorithms
Xiang Wu (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
CompressionComputational EfficiencyRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Proposes an scalable topology-preserving graph compression method called STPGC, which achieves complete preservation of topological features (connected components, cycles, holes) through graph strong compression, graph edge compression, and neighborhood coning.
Scalable Traffic Signal Control with Shared Policy Framework
Haolun MA (Hong Kong University of Science and Technology), Lei Li (Hong Kong University of Science and Technology)
Autonomous DrivingOptimizationRecurrent Neural NetworkReinforcement LearningAuto EncoderGraphTime Series
🎯 What it does: Proposed an expandable shared policy framework called SLight for multi-agent traffic signal control.
SCalDA: Semantics-Calibrated and Diffusion-Enhanced Data Augmentation
Shibo Lv (Shenzhen University), Jianmin Jiang (Shenzhen University)
ClassificationData SynthesisTransformerDiffusion modelContrastive LearningImage
🎯 What it does: Propose a three-domain consistent augmented framework called SCalDA, which utilizes significance-guided label alignment, diffusion model-based local image inpainting, and metric learning-based feature alignment to achieve precise matching between visual semantics and label semantics.
Scale-Aware Domain Harmonization for Domain Adaptation Person Search
Huibing Wang (Dalian Maritime University), Jiqing Zhang (Dalian Maritime University)
RetrievalDomain AdaptationConvolutional Neural NetworkScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImage
🎯 What it does: Propose an unsupervised cross-domain person re-identification framework called SCALE, which addresses the scale inconsistency between the source and target domains, and improves the reliability of pseudo labels through bidirectional clustering regularization.
SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models
Hyeonbeom Choi (Seoul National University), Jonghyun Choi (Seoul National University)
Computational EfficiencyRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: Proposed a single forward inference, self-uncertainty-driven visual-language-action model inference strategy named SCALE, which can simultaneously regulate visual perception and action execution without additional training.
ScaleEnv: Scaling Environment Synthesis from Scratch for Generalist Interactive Tool-Use Agent Training
Dunwei Tu (Nanjing University), Xunliang Cai (Meituan)
Data SynthesisGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextTabular
🎯 What it does: Propose the ScaleEnv framework, which can synthesize high-fidelity, verifiable interactive environments and tasks from scratch, for training general-purpose tool-using agents;
ScaleErasure: Inference-Time Minimal Intervention for Precise Concept Erasure in Next-Scale Autoregressive Image Generation
Cong Wang (Nanjing University), Qing Gu (Nanjing University)
GenerationSafty and PrivacyComputational EfficiencyTransformerPrompt EngineeringDiffusion modelScore-based ModelGenerative Adversarial NetworkContrastive LearningImageText
🎯 What it does: Under the next-scale autoregressive image generation framework, a reasoning-time concept erasure method called ScaleErasure is proposed, which can precisely erase unsafe content with minimal intervention without modifying model parameters.
ScaleMoE: Mixture-of-Experts for Scalable Continuous Control in Actor-Critic Reinforcement Learning
Yi Ma (Shanxi University), Jiye Liang (Shanxi University)
Reinforcement LearningMixture of ExpertsTabularTime SeriesSequential
🎯 What it does: Propose ScaleMoE, which embeds Mixture-of-Experts (MoE) simultaneously into both the actor and critic of continuous control, achieving scalable RL models.
ScaleSim: Serving Large-Scale Multi-Agent Simulation with Invocation Distance-Based Memory Management
Zaifeng Pan (University of California San Diego), Yufei Ding (University of California San Diego)
Autonomous DrivingOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose ScaleSim, a memory management system based on 'invocation distance' specifically designed for large-scale LLM multi-agent simulations, capable of proactive prefetching, priority eviction, and unified management of heterogeneous memory modules.
Scaling Agentic Verifier for Competitive Coding
Zeyao Ma (Renmin University of China), Binyuan Hui (Alibaba Group)
Data SynthesisAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose an agentic verifier that generates high-discriminative test inputs through multi-round interactions, which can distinguish between candidate programs, thereby improving execution-based code evaluation in competitive programming.
Scaling Behavior in Model Fine-tuning for Audio DeepFake Detection
Xiang Li (Fordham University), Wenqi Wei (Fordham University)
Anomaly DetectionTransformerSupervised Fine-TuningAudio
🎯 What it does: Systematically studied the scaling behavior of audio DeepFake detection models during the fine-tuning phase, analyzing the impact of model capacity and training data scale on detection performance, robustness, and generalization.
Scaling Beyond Masked Diffusion Language Models
Subham Sekhar Sahoo (NVIDIA), Ante Jukić (NVIDIA)
GenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Conduct a unified computational matching scale study on three types of discrete diffusion language models (Masked, Uniform-State, Interpolating), and reveal their differences in the trade-off between speed and quality;
Scaling by Diversified Experience for Vision-Language-Action Models
Leiyu Wang (Shanghai Jiao Tong University), Nanyang Ye (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerReinforcement LearningPrompt EngineeringMixture of ExpertsVision Language ModelVision-Language-Action ModelImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose the SyVLA model, which integrates VLM with flow-matching action experts, and achieves collaborative execution of robot actions and high-level reasoning through three-stage training (pre-training, task fine-tuning, reinforcement learning) combined with multi-modal data.
Scaling Continual Learning to 300+ Tasks with Bi-Level Routing Mixture-of-Experts
Meng Lou (University Of Hong Kong), Yizhou Yu (University Of Hong Kong)
ClassificationComputational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningTransformerPrompt EngineeringMixture of ExpertsContrastive LearningImageBenchmark
🎯 What it does: Propose an scalable continual learning framework called CaRE, which can achieve class-incremental learning on long sequences with over 300 tasks, and dynamically retrieves and aggregates expert knowledge at each network layer based on a pre-trained model.
Scaling depth capacity via zero/one-layer model expansion
Zhiqi Bu (FAIR)
OptimizationComputational EfficiencyRepresentation LearningHyperparameter SearchConvolutional Neural NetworkTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: The study proposes a progressive training method with zero/one-layer depth scaling to significantly reduce the training cost of large-scale models.
Scaling Generative Verifiers For Natural Language Mathematical Proof Verification And Selection
Sadegh Mahdavi (NVIDIA), Igor Gitman (NVIDIA)
GenerationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A systematic study on the verification and selection of natural language mathematical proofs, and a hybrid framework combining GenSelect and LLM-as-a-Judge is proposed.
Scaling Inference-Time Computation via Opponent Simulation: Enabling Online Strategic Adaptation in Repeated Negotiation
Xiangyu Liu (Google Research), Aranyak Mehta (Google Research)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: For repeated negotiation games, a method is designed to enable LLM online strategy adaptation by increasing computational load during the reasoning phase.
Scaling Law for Quantization-Aware Training
Mengzhao Chen (University of Hong Kong), Ping Luo (University of Hong Kong)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper conducts 4-bit quantization-aware training (QAT) for large language models and proposes a unified scaling law.
Scaling Laws and Architectural Frontiers in Metagenomic Foundation Models
Geraldene Munsamy (Basecamp Research), Francesco Farina (Basecamp Research)
Protein Structure PredictionConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningBiomedical DataReview/Survey Paper
🎯 What it does: This paper builds and systematically evaluates three architectural base models (Transformer, State-Space, and Long Convolution) based on a large-scale metagenomic corpus, ultimately introducing the EDEN series and expanding it to 28B parameters.
Scaling Laws for Precision in High-Dimensional Linear Regression
Dechen Zhang (University of Hong Kong), Difan Zou (University of Hong Kong)
OptimizationComputational EfficiencyRepresentation LearningTabular
🎯 What it does: Propose a scaling law for low-precision training, theoretically derive and distinguish the effects of multiplicative quantization and additive quantization on model capacity and data volume, and provide upper and lower bounds;
Scaling Laws of Global Weather Models
Yuejiang Yu (ETH Zurich), Torsten Hoefler (ETH Zurich)
OptimizationComputational EfficiencyGraph Neural NetworkTransformerTime SeriesBenchmarkPhysics Related
🎯 What it does: This study systematically evaluates the scalability of five global weather forecasting models in terms of model size, data volume, and computational budget, and proposes a power-law relationship between validation loss and model size, data volume, and computational budget.
Scaling Long-Horizon Agent via Context Folding
Weiwei Sun (Carnegie Mellon University), Jiecao Chen (ByteDance Seed)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the Context Folding mechanism, enabling LLM agents to actively branch out to handle subtasks and then fold the intermediate steps into a concise summary after completion, keeping the main context concise; meanwhile, introduces FoldGRPO (RL-based folding strategy optimization) to end-to-end train the folding behavior;
Scaling Multi-Agent Environment Co-Design with Diffusion Models
Hao Xiang Li (University of Cambridge), Amanda Prorok (University of Cambridge)
OptimizationTransformerReinforcement LearningDiffusion modelTabularTime SeriesSequentialBenchmark
🎯 What it does: Proposes Diffusion Co-Design (DiCoDe), a method that integrates diffusion models with multi-agent reinforcement learning to jointly optimize the environment and policy;
Scaling Prompt Synthesis for Large Language Model Reasoning
Xueliang Zhao (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
OptimizationData-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialChain-of-Thought
🎯 What it does: Achieve unsupervised prompt synthesis through an expectation maximization (EM) framework, generating more challenging and richer reasoning tasks.
Scaling Real-World Robot Policy Evaluation via Discrete Diffusion World Model
Yaxuan Li, Yichen Zhu
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelDiffusion modelWorld ModelVideoTextMultimodality
🎯 What it does: Propose dWorldEval, a discrete diffusion world model for large-scale evaluation of robot manipulation policies.
Scaling Small Agents Through Strategy Auctions
Lisa Alazraki (Imperial College London), Akhil Mathur (Meta Superintelligence Labs)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIMixture of ExpertsContrastive LearningTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a proxy allocation framework called SALE based on strategic bidding, which can dynamically allocate language models of different scales according to task complexity without training a router, thereby improving the utilization of small models and enabling adaptive enhancement.
Scaling the Prior: Size-Consistent Geometric Diffusion for 3D Molecular Generation
Wenhan Gao (Stony Brook University), Yi Liu (Stony Brook University)
Drug DiscoveryGraph Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelPoint CloudMeshGraph
🎯 What it does: Proposed the StP method, a size-consistent prior scaling approach, to address the inconsistency in denoising dynamics caused by molecular size in 3D molecular diffusion models.
Scaling the Scaling Logic: Agentic Meta-Synthesis of Logic Reasoning
Bowen LIU, Jia Li (Hong Kong University of Science and Technology)
Data SynthesisMeta LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: This work proposes an Agentic Meta-Synthesis framework called SSLogic, which utilizes large language model agents to automatically generate and continuously evolve verifiable logic task families (Generator-Validator pairs) through a closed-loop Generate-Validate-Refine process, ensuring the correctness and measurability of generated instances through multi-gate verification.
Scaling Transformers for End-to-End Discrete Audio Tokenization
Yitian Gong (Fudan University), Xipeng Qiu (Fudan University)
RecognitionGenerationCompressionTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderGenerative Adversarial NetworkContrastive LearningTextMultimodalityAudio
🎯 What it does: Propose a full end-to-end Transformer-based Audio Codec (TAC), achieving high-fidelity discrete audio tokenization through a unified causal Transformer encoder, quantizer, and decoder, and applying this tokenizer to autoregressive text-to-speech (TTS) and automatic speech recognition (ASR) tasks.
Scaling Unsupervised Multi-Source Federated Domain Adaptation through Group-Wise Discrepancy Minimization
Larissa Reichart (University of Tübingen), Mete Akgün
Domain AdaptationFederated LearningAgentic AIContrastive LearningImage
🎯 What it does: Proposes the GALA framework, achieving a scalable and robust target domain alignment method in federated unsupervised multi-source domain adaptation.
Scaling up Multi-Turn Off-Policy RL and Multi-Agent Tree Search for LLM Step-Provers
Ran Xin (ByteDance Seed), Xia Xiao (ByteDance Seed)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes BFS-Prover-V2, a step-level formal reasoning system that combines multi-stage expert iterative reinforcement learning with hierarchical multi-agent tree search, achieving scalability in both training and inference phases.
Scaling Vision Transformers for Functional MRI with Flat Maps
Connor Lane (MedARC), Paul Steven Scotti
Image TranslationRestorationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes and trains a self-supervised fMRI foundation model based on Vision Transformer called CortexMAE. It converts 3D fMRI data into 2D images using a brain cortex flat map projection, and then performs large-scale pre-training on 2.1K hours of publicly available HCP-YA data, followed by applying the model to multiple downstream tasks.
Scaling-Aware Adapter for Structure-Grounded LLM Reasoning
Zihao Jing (Western University), Pingzhao Hu (Western University)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraphBiomedical Data
🎯 What it does: This paper proposes Cuttlefish, a unified multimodal large language model that achieves reasoning over all atomic-level structures through scalable structural patching and geometric alignment adapters.
Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation
Heng Qu (Wuhan University), Jian Luan (Xiaomi)
Autonomous DrivingOptimizationSafty and PrivacyComputational EfficiencyTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelGaussian SplattingImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a large-scale Chinese mobile GUI task dataset called HyperTrack, developed a unified evaluation framework named GUIEvalKit, and systematically studied the training scale of VLM agents, the comparison between reinforcement learning and supervised learning, semi-online evaluation methods, and the impact of reasoning on decision diversity and stability.
ScalingAR: Scaling Confidence for Autoregressive Image Generation
Harold Haodong Chen (Hong Kong University of Science and Technology (Guangzhou)), Ying-Cong Chen (Hong Kong University of Science and Technology (Guangzhou))
GenerationTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: Proposed a test-time scaling framework called ScalingAR for next-step prediction in AR image generation, which utilizes token entropy for confidence evaluation and dynamically trims and guides during inference.
ScaLoRA: Optimally Scaled Low-Rank Adaptation for Efficient High-Rank Fine-Tuning
Yilang Zhang (Morgan Stanley), Georgios B. Giannakis (University of Minnesota)
OptimizationFederated LearningComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTextTabularSequential
🎯 What it does: Propose ScaLoRA, a parameter-efficient method for high-rank fine-tuning that achieves optimal scaling of the LoRA adapter and implements high-rank weight updates through accumulated updates at regular intervals, balancing convergence speed and effectiveness.
Scam2Prompt: A Scalable Framework for Auditing Malicious Scam Endpoints in Production LLMs
Zhiyang Chen (University of Toronto), Fan Long (University of Toronto)
Anomaly DetectionSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed the Scam2Prompt automated auditing framework, which uses known scam websites to generate developer-style prompts, evaluating production LLMs on whether they would generate code containing malicious URLs without adversarial prompts.
scCBGM: Single-Cell Editing via Concept Bottlenecks
Alma Andersson (Genentech), Hector Corrada Bravo (Genentech)
GenerationExplainability and InterpretabilityData-Centric LearningTransformerDiffusion modelFlow-based ModelAuto EncoderContrastive LearningBiomedical Data
🎯 What it does: Propose a single-cell concept bottleneck generative model (scCBGM) to achieve interpretable and accurate single-cell counterfactual editing.
scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
Jiawei Zhang (Tianjin University), Jiachen Yang (Tianjin University)
Computational EfficiencyRepresentation LearningData-Centric LearningDrug DiscoveryTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderContrastive LearningMultimodalityTabularBiomedical Data
🎯 What it does: Propose a two-stage framework called scChord for converting single-cell transcriptome (RNA) information into protein abundance;
scDataset: Scalable Data Loading for Deep Learning on Large-Scale Single-Cell Omics
Davide D'Ascenzo (University of Milan), Sebastiano Cultrera di Montesano (Broad Institute of MIT and Harvard)
Computational EfficiencyData-Centric LearningTabularBiomedical Data
🎯 What it does: This paper proposes scDataset, an IterableDataset for PyTorch, designed to efficiently load single-cell data directly from disk, enabling deep learning training on large-scale single-cell omics data.
scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning
Jieun Sung (Ewha Womans University), Wankyu Kim (Ewha Womans University)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningBiomedical Data
🎯 What it does: Built a pre-training framework called scDEBART based on BART, which specifically learns gene expression changes (logFC) under different baseline cell expression backgrounds, and predicts single-cell gene perturbations based on this.