ICLR 2025 Papers — Page 31
International Conference on Learning Representations · 3704 papers
SFS: Smarter Code Space Search improves LLM Inference Scaling
Jonathan Light (Rensselaer Polytechnic Institute), Wei Cheng (NEC Laboratories America)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Treating code generation as a black-box optimization problem, we propose SCATTERED FOREST SEARCH (SFS), which enhances exploration and exploitation through three techniques: SCATTERING, FORESTING, and SCOUTING, thereby accelerating the scaling of LLM inference.
SG-I2V: Self-Guided Trajectory Control in Image-to-Video Generation
Koichi Namekata (University of Toronto), David B. Lindell (University of Toronto)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: We propose SG-I2V, a zero-shot self-guided image-to-video generation framework that can control the motion of objects and the camera in the video solely based on a pre-trained image-video diffusion model and user-provided bounding box trajectories.
SGD with memory: fundamental properties and stochastic acceleration
Dmitry Yarotsky (Skoltech), Maksim Velikanov (Technology Innovation Institute)
OptimizationTabular
🎯 What it does: A stochastic gradient descent (SGD) framework with limited memory (M vectors) is proposed, and a general expansion for its loss convergence is provided. The convergence phase under power-law spectra is then studied, and an accelerated algorithm with memory-1 (AM1) is designed.
Shallow diffusion networks provably learn hidden low-dimensional structure
Nicholas Matthew Boffi, Ingvar Ziemann (University of Pennsylvania)
Diffusion modelStochastic Differential Equation
🎯 What it does: This paper proves that single hidden layer neural networks (Barron space) can adaptively capture low-dimensional linear subspaces and hidden independent component structures when learning diffusion models, thereby avoiding the curse of dimensionality, and provides a sample complexity upper bound primarily based on latent dimensions.
Shape as Line Segments: Accurate and Flexible Implicit Surface Representation
Siyu Ren (City University of Hong Kong), Junhui Hou (City University of Hong Kong)
SegmentationGenerationData SynthesisAutonomous DrivingPoint CloudMesh
🎯 What it does: This paper proposes an implicit surface representation method based on Line Segment Field (LSF), referred to as SALS, and implements the corresponding neural network and point cloud reconstruction pipeline.
Shapley-Guided Utility Learning for Effective Graph Inference Data Valuation
Hongliang Chi (Rensselaer Polytechnic Institute), Yao Ma (AT&T)
OptimizationData-Centric LearningGraph Neural NetworkGraph
🎯 What it does: A Shapley-Guided Utility Learning (SGUL) framework is proposed to evaluate the value of neighboring nodes during the graph reasoning phase in the absence of test labels.
Shared-AE: Automatic Identification of Shared Subspaces in High-dimensional Neural and Behavioral Activity
Daiyao Yi (Yale University), Shreya Saxena (University of California Los Angeles)
Representation LearningAuto EncoderVideoMultimodality
🎯 What it does: An autoencoder framework (Shared‑AE) is proposed to identify the common subspace between neural activity and behavioral records by separating shared and private latent subspaces.
Sharper Guarantees for Learning Neural Network Classifiers with Gradient Methods
Hossein Taheri (University of California), Arya Mazumdar (University of California)
ClassificationOptimizationImage
🎯 What it does: This study investigates the convergence and generalization of smooth activation neural networks under gradient methods, providing a theoretically guaranteed width scalability.
Sharpness-Aware Black-Box Optimization
Feiyang Ye (Southern University of Science and Technology), Ivor Tsang
OptimizationTextStochastic Differential Equation
🎯 What it does: A Sharpness-Aware Black-Box Optimization (SABO) algorithm is proposed to enhance model generalization performance through flat minimization in black-box optimization.
Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training
Zhanpeng Zhou (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the behavior of Sharpness-Aware Minimization (SAM) in the later stages of training, finding that using SAM for just a few rounds in the later training phase can achieve the same generalization performance and flatter minima as the full SAM, while early use of SAM has limited impact on the final results.
Sharpness-Aware Minimization: General Analysis and Improved Rates
Dimitris Oikonomou (Johns Hopkins University), Nicolas Loizou (Johns Hopkins University)
OptimizationImage
🎯 What it does: A unified Sharpness-Aware Minimization (Unified SAM) algorithm is proposed, combining the update rules of traditional SAM and USAM, and providing convergence analysis for arbitrary sampling (including importance sampling and τ-nice sampling);
Shedding Light on Time Series Classification using Interpretability Gated Networks
Yunshi Wen (Rensselaer Polytechnic Institute), Anak Agung Julius (Rensselaer Polytechnic Institute)
ClassificationExplainability and InterpretabilityMixture of ExpertsTime SeriesBiomedical Data
🎯 What it does: A hybrid model called InterpGN is proposed, which combines interpretable experts and deep neural networks to construct logical predicates using shapelets, achieving interpretable time series classification.
ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design
Keir Adams (Massachusetts Institute of Technology), Connor W. Coley (Massachusetts Institute of Technology)
Drug DiscoveryDiffusion modelPoint Cloud
🎯 What it does: A SE(3)-equivariant diffusion model named ShEPhERD has been developed for the joint generation of 3D molecular structures along with their shapes, electrostatics, and pharmacophores (interactive features).
Shh, don't say that! Domain Certification in LLMs
Cornelius Emde (University of Oxford), Adel Bibi (University of Oxford)
Domain AdaptationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: A domain certification framework and the VALID algorithm are proposed, providing a provable upper bound on out-of-domain generation probabilities for LLMs under any input, enhancing the model's adversarial robustness.
Shifting the Paradigm: A Diffeomorphism Between Time Series Data Manifolds for Achieving Shift-Invariancy in Deep Learning
Berken Utku Demirel (ETH Zurich), Christian Holz (ETH Zurich)
ClassificationRecognitionAnomaly DetectionConvolutional Neural NetworkRecurrent Neural NetworkTime SeriesSequentialElectrocardiogram
🎯 What it does: This paper proposes a differentiable bijective transformation that maps time-series data to the same data manifold point after shifting in the time domain, thereby achieving complete invariance of deep learning models to arbitrary time shifts.
ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents
Haiyang SHEN, Yun Ma (Peking University)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: A large-scale real API benchmark called SHORTCUTSBENCH was created, and the performance of 10 LLM-driven API agents was evaluated in terms of API selection, parameter filling, and input requests.
Shot2Story: A New Benchmark for Comprehensive Understanding of Multi-shot Videos
Mingfei Han (Bytedance Inc.), Heng Wang (Data61, CSIRO)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: Proposed the Shot2Story benchmark, providing 42,958 multi-shot short videos with single-shot visuals and narrative subtitles, long video summaries, and question-answer pairs. Designed single-shot subtitles, video summaries, and multi-shot question-answer tasks, and conducted baseline experiments.
Should VLMs be Pre-trained with Image Data?
Sedrick Keh (Toyota Research Institute), Achal Dave (Toyota Research Institute)
TransformerVision Language ModelImageTextMultimodality
🎯 What it does: The study investigates the impact of introducing image data early in the pre-training phase on VLM performance, comparing two-stage and one-stage pre-training strategies.
Show-o: One Single Transformer to Unify Multimodal Understanding and Generation
Jinheng Xie (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A single Transformer model, Show-o, has been constructed to simultaneously achieve multimodal understanding (such as visual question answering and image captioning) and multimodal generation (such as text-to-image, text-guided filling, and video generation).
SigDiffusions: Score-Based Diffusion Models for Time Series via Log-Signature Embeddings
Barbora Barancikova (Imperial College London), Cristopher Salvi (Imperial College London)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelTime Series
🎯 What it does: This paper proposes a method that utilizes log-signatures as embeddings for time series data and trains a score-based diffusion model (SigDiffusion) on this embedding to generate long time series. It also provides a closed-form inverse formula for signatures, allowing for the direct recovery of the original path from the log-signature.
Signature Kernel Conditional Independence Tests in Causal Discovery for Stochastic Processes
Georg Manten (Technical University of Munich), Niki Kilbertus (Technical University of Munich)
Time SeriesSequentialFinance RelatedStochastic Differential Equation
🎯 What it does: A conditional independence constraint and constraint-based causal discovery algorithm for continuous-time stochastic processes is proposed, capable of handling both complete and partial observation data, and uniquely recovering the causal graph of SDE models through temporal directionality.
SIM: Surface-based fMRI Analysis for Inter-Subject Multimodal Decoding from Movie-Watching Experiments
Simon Dahan (King's College London), Emma Claire Robinson
GenerationRetrievalTransformerDiffusion modelAuto EncoderContrastive LearningVideoMultimodalityMagnetic Resonance ImagingAudio
🎯 What it does: A SIM framework based on Surface Vision Transformer (SiT) is proposed, utilizing self-supervised video surface masking autoencoder (vsMAE) pre-training, and aligning through tri-modal CLIP to map the brain electroencephalogram (fMRI), video, and audio representations of 3-second movie clips into a shared space, achieving cross-individual and cross-scene multimodal retrieval and video reconstruction.
SimBa: Simplicity Bias for Scaling Up Parameters in Deep Reinforcement Learning
Hojoon Lee (KAIST), Takuma Seno (Sony AI)
Computational EfficiencyReinforcement LearningSequential
🎯 What it does: A network architecture named SimBa is designed, achieving parameter scaling in deep reinforcement learning through observation normalization, residual feedforward blocks, and post-layer normalization.
SiMHand: Mining Similar Hands for Large-Scale 3D Hand Pose Pre-training
Nie Lin (University of Tokyo), Yoichi Sato (University of Tokyo)
Pose EstimationConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: The SiMHand framework is proposed, which pre-trains 3D gesture estimation using a large-scale dataset of hand images from the wild, leveraging contrastive learning on similar hand pairs mined from different videos.
SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters
Teng Xiao (Pennsylvania State University), Vasant G Honavar (Sun Yat-Sen University)
Recommendation SystemOptimizationTransformerLarge Language ModelTextBenchmark
🎯 What it does: A simple and effective hyperparameter-free preference optimization algorithm, SimPER, is proposed for aligning language models.
SIMPL: Scalable and hassle-free optimisation of neural representations from behaviour
Tom George, Claudia Clopath (Imperial College London)
OptimizationExplainability and InterpretabilityComputational EfficiencyTime Series
🎯 What it does: The SIMPL algorithm is proposed, which utilizes behavioral initial conditions to iteratively optimize latent variables and tuning curves in neural networks.
Simple Guidance Mechanisms for Discrete Diffusion Models
Yair Schiff (Cornell University), Volodymyr Kuleshov (InstaDeep)
GenerationData SynthesisDrug DiscoveryDiffusion modelImageTextBiomedical Data
🎯 What it does: This study investigates controllable generation methods for discrete diffusion models, proposing a classifier-guided/unguided mechanism and an improved Unified Diffusion Language Model (UDLM), and validating its performance on multi-domain data such as genomics, molecules, images, and text.
Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation
Mufei Li (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
GenerationRetrievalGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: This paper proposes and implements a knowledge graph-based retrieval-augmented generation framework called SubgraphRAG, which first retrieves variable-sized and flexibly structured knowledge subgraphs through efficient subgraph retrieval, and then uses a non-fine-tuned large language model (LLM) to reason over the subgraphs, providing answers along with interpretable explanations.
Simple ReFlow: Improved Techniques for Fast Flow Models
Beomsu Kim (Apple and Korea Advanced Institute of Science and Technology), James Thornton (Apple)
GenerationData SynthesisComputational EfficiencyFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Improved the training and inference of the flow matching model under the ReFlow framework to achieve faster and higher quality generation.
Simple yet Effective Incomplete Multi-view Clustering: Similarity-level Imputation and Intra-view Hybrid-group Prototype Construction
Shengju Yu (National University of Defense Technology), Yiu-ming Cheung (Hong Kong Baptist University)
OptimizationMultimodality
🎯 What it does: An end-to-end framework for incomplete multi-view clustering, SIIHPC, is proposed, which achieves more accurate clustering by completing at the similarity level and learning a set of mixed prototypes for each view.
Simple, Good, Fast: Self-Supervised World Models Free of Baggage
Jan Robine (TU Dortmund), Stefan Harmeling (TU Dortmund)
Computational EfficiencyRepresentation LearningReinforcement LearningContrastive LearningWorld ModelImageVideoBenchmark
🎯 What it does: This paper proposes SGF, a self-supervised world model that does not use RNNs, Transformers, discrete representations, or image reconstruction. It learns temporally consistent representations through frame/action stacking and data augmentation, achieving efficient training on the Atari 100k benchmark.
SimpleTM: A Simple Baseline for Multivariate Time Series Forecasting
Hui Chen (University of Wisconsin-Madison), Vikas Singh (University of Wisconsin-Madison)
TransformerTime Series
🎯 What it does: A lightweight multivariate time series forecasting model called SimpleTM is proposed, which uses Stationary Wavelet Transform (SWT) for multi-scale decomposition of time series and incorporates the geometric algebra vector product into the attention mechanism to capture high-order correlations between channels.
Simplifying Deep Temporal Difference Learning
Matteo Gallici (Universitat Politècnica de Catalunya), Mario Martin (Universitat Politècnica de Catalunya)
Recurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: A simplified deep temporal difference learning method called PQN is proposed, eliminating the target network and large replay buffer, achieving stability through regularization (LayerNorm + ℓ2).
Simplifying, Stabilizing and Scaling Continuous-time Consistency Models
Cheng Lu (OpenAI), Yang Song (OpenAI)
GenerationData SynthesisDiffusion modelFlow-based ModelImage
🎯 What it does: A simplified, stable, and scalable continuous time consistency model (sCM) training framework is proposed, capable of achieving high-quality image generation with large-scale parameters (up to 1.5B);
Simulating Human-like Daily Activities with Desire-driven Autonomy
Yiding Wang (Peking University), Yizhou Wang (Peking University)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes a Desire-Driven Autonomous Agent (D2A) that allows large language models to automatically generate and execute behaviors aligned with human daily activities without explicit task instructions, through an internal value system.
Simulating Training Dynamics to Reconstruct Training Data from Deep Neural Networks
Hanling Tian (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
Data SynthesisOptimizationConvolutional Neural NetworkImage
🎯 What it does: By simulating the training dynamics, we reverse-engineer the training data—given the initial and final parameters of a deep network, we optimize a synthetic dataset composed of random noise so that the final parameters of the network trained on this synthetic dataset are as similar as possible to the final parameters of the real network, thereby obtaining images similar to the original training data.
SimulPL: Aligning Human Preferences in Simultaneous Machine Translation
Donglei Yu (University of Chinese Academy of Sciences), Chengqing Zong (Institute of Automation, Chinese Academy of Sciences)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The SimulPL framework is proposed, which enhances model performance by aligning five major preferences of humans in synchronous machine translation (translation quality, monotonicity, key information, conciseness, and latency);
SimXRD-4M: Big Simulated X-ray Diffraction Data and Crystal Symmetry Classification Benchmark
Bin CAO, Tong-yi Zhang (Hong Kong University of Science and Technology)
ClassificationRecognitionConvolutional Neural NetworkRecurrent Neural NetworkTransformerSequentialBenchmarkPhysics Related
🎯 What it does: This paper constructs the largest simulated X-ray powder diffraction (XRD) dataset, SimXRD-4M, and conducts benchmark experiments on various sequence models to explore the impact of long-tail label distribution on crystal symmetry recognition.
SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects
Jiayi Liu (Simon Fraser University), Ali Mahdavi Amiri
GenerationTransformerDiffusion modelImageMesh
🎯 What it does: Generate complete assets of 3D movable objects from a single image, including part connection diagrams, joint parameters, and geometry.
SINGER: Stochastic Network Graph Evolving Operator for High Dimensional PDEs
Mingquan Feng (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Graph Neural NetworkGraphPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A framework named SINGER is proposed, which learns the solution operator of high-dimensional partial differential equations (PDEs) using subnetwork parameters driven by graph neural network-based stochastic differential equations, thereby directly obtaining approximate analytical solutions to PDEs without the need for spatial discretization.
Single Teacher, Multiple Perspectives: Teacher Knowledge Augmentation for Enhanced Knowledge Distillation
Md Imtiaz Hossain (Kyung Hee University), Eui-Nam Huh (Kyung Hee University)
CompressionKnowledge DistillationGaussian SplattingImage
🎯 What it does: This paper proposes a technique called TeKAP, which generates various synthetic teacher knowledge by injecting random noise into the feature maps and logits of a single teacher model, thereby enhancing the generalization ability of the student model using multiple perspectives without the need to train multiple teachers.
Single-agent Poisoning Attacks Suffice to Ruin Multi-Agent Learning
Fan Yao (University of Virginia), Haifeng Xu (University of Chicago)
OptimizationAdversarial AttackGraph
🎯 What it does: This paper studies the robustness of multi-agent learning (MAL) under strongly monotonic games, proposing the Single-Agent Utility Spoofing Attack (SUSA). It proves that even by contaminating just one agent, the attacker can mislead any MAL algorithm to a non-original Nash equilibrium within a sublinear budget, and further analyzes the trade-off between learning rate, convergence speed, and robustness.
Singular Subspace Perturbation Bounds via Rectangular Random Matrix Diffusions
Peiyao Lai (Worcester Polytechnic Institute), Oren Mangoubi (Worcester Polytechnic Institute)
Stochastic Differential Equation
🎯 What it does: This paper presents an upper bound on the Frobenius norm perturbation of the subspace spanned by the first k right singular vectors of matrix A after being disturbed by Gaussian noise.
SiReRAG: Indexing Similar and Related Information for Multihop Reasoning
Nan Zhang (Pennsylvania State University), Chien-Sheng Wu (Pennsylvania State University)
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A dual-perspective RAG indexing method called SIRERAG is proposed, which combines similarity and relevance to enhance the retrieval performance of multi-hop reasoning question answering.
Sitcom-Crafter: A Plot-Driven Human Motion Generation System in 3D Scenes
Jianqi Chen (Beihang University), Xiaodan Liang (Sun Yat-sen University)
GenerationData SynthesisRobotic IntelligenceLarge Language ModelDiffusion modelGenerative Adversarial NetworkTextPoint CloudMesh
🎯 What it does: The Sitcom-Crafter system has been developed to uniformly generate various types of human movements, such as walking, scene interaction, and multi-person interaction, based on long narrative texts in 3D scenes.
Size-Generalizable RNA Structure Evaluation by Exploring Hierarchical Geometries
Zongzhao Li (Renmin University of China), Le Song (BioMap Research)
Protein Structure PredictionGraph Neural NetworkBiomedical Data
🎯 What it does: A type of equivariant graph neural network named EquiRNA is proposed to evaluate the 3D structure of RNA and address the generalization issues caused by size differences in RNA.
Sketch2Diagram: Generating Vector Diagrams from Hand-Drawn Sketches
Itsumi Saito (Tohoku University), Keisuke Sakaguchi (Tohoku University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: A complete process for generating high-quality vector graphics (TikZ) from hand-drawn sketches is proposed, including the SKETI kZ dataset, data augmentation methods, and a dedicated image-language model IMGTI kZ.
Sketching for Convex and Nonconvex Regularized Least Squares with Sharp Guarantees
Yingzhen Yang (Arizona State University), Ping Li (VecML Inc.)
OptimizationTabular
🎯 What it does: This paper proposes a matrix sampling-based regularized least squares optimization algorithm called SRO, and further introduces an iterative version, Iterative SRO, to achieve efficient solutions while preserving the sparse structure.
Skill Expansion and Composition in Parameter Space
Tenglong Liu (National University of Defense Technology), Xianyuan Zhan (Tsinghua University)
Autonomous DrivingRobotic IntelligenceReinforcement LearningDiffusion model
🎯 What it does: A Parametric Skill Expansion and Composition (PSEC) framework is proposed, which constructs a pluggable skill library using Low-Rank Adaptation (LoRA) modules, and quickly reuses and expands existing skills in new tasks through context-aware parameter-level combinations.
SleepSMC: Ubiquitous Sleep Staging via Supervised Multimodal Coordination
Shuo Ma (Chinese Academy of Sciences), Ziyu Jia (AI Dream Intelligent Technology Co. Ltd.)
ClassificationRecognitionConvolutional Neural NetworkContrastive LearningMultimodalityBiomedical Data
🎯 What it does: A sleep staging framework named SleepSMC is proposed, which can achieve high-precision staging using only a single modality after multimodal training.
SLMRec: Distilling Large Language Models into Small for Sequential Recommendation
Wujiang Xu (Rutgers University), Yongfeng Zhang (Rutgers University)
Recommendation SystemComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningSequential
🎯 What it does: In this work, the authors first evaluate the role of large language models (LLMs) in sequential recommendation, finding that most intermediate layers are redundant; they then propose a hierarchical knowledge distillation-based SLMREC model that utilizes a small LLM to achieve performance on par with large LLM Rec, significantly reducing model size and inference/training costs.
SLoPe: Double-Pruned Sparse Plus Lazy Low-Rank Adapter Pretraining of LLMs
Mohammad Mozaffari (University of Toronto), Maryam Mehri Dehnavi (University of Toronto)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: SLOPE introduces dual sparsity and lazy low-rank adapters in LLM pre-training and inference, enhancing speed and memory efficiency.
Slot-Guided Adaptation of Pre-trained Diffusion Models for Object-Centric Learning and Compositional Generation
Kaan Akan, Yucel Yemez (Koc University)
SegmentationGenerationDiffusion modelImage
🎯 What it does: This paper proposes SlotAdapt, an unsupervised object-centric learning framework that combines slot attention with pre-trained diffusion models, enabling object discovery, segmentation, and compositional generation on complex real images.
SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation
Yining Hong (University of California Los Angeles), Lijuan Wang (Microsoft Research)
GenerationData SynthesisDiffusion modelWorld ModelVideo
🎯 What it does: A dual-speed learning framework (SLOWFAST-VGEN) has been constructed to achieve action-based long video generation, balancing slow learning of world models and fast learning of plot memory.
Small Models are LLM Knowledge Triggers for Medical Tabular Prediction
Jiahuan Yan (Zhejiang University), Jian Wu (University of Illinois Urbana-Champaign)
ClassificationOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTabularBiomedical Data
🎯 What it does: An unsupervised self-prompting cycle named SERSAL is proposed, utilizing collaborative learning between small models and large language models (LLMs) to enhance the performance of LLMs in numerical table prediction tasks (such as medical diagnosis).
Small-to-Large Generalization: Training Data Influences Models Consistently Across Scale
Alaa Khaddaj (Massachusetts Institute of Technology), Aleksander Madry (Massachusetts Institute of Technology)
TransformerLarge Language ModelImageText
🎯 What it does: This paper studies the impact of training data distribution on language models of different scales, exploring whether small proxy models can accurately predict the performance of large models on various downstream tasks, and applies this finding to two major scenarios: data attribution and dataset selection, evaluating the trade-off between the scale and effectiveness of proxy models.
Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling
Hritik Bansal (University of California Los Angeles), Mehran Kazemi (Google DeepMind)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper proposes and verifies that fine-tuning models using synthetic reasoning data generated by weak (small) models is more computationally efficient than using data generated by strong (large) models under a fixed computational budget. It significantly improves reasoning performance across various fine-tuning paradigms (knowledge distillation, self-improvement, weak-to-strong enhancement) and multiple tasks (MATH, GSM-8K, Functional MATH, IFEval).
SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction
Yang Zhou (SenseTime Research), Yu Liu (Shanghai Artificial Intelligence Laboratory)
Autonomous DrivingRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkTransformerContrastive LearningPoint Cloud
🎯 What it does: This paper proposes SmartPretrain, a general, model-agnostic, and dataset-agnostic self-supervised pre-training framework that combines contrastive learning and reconstruction learning to enhance motion prediction performance in autonomous driving.
SmartRAG: Jointly Learn RAG-Related Tasks From the Environment Feedback
Jingsheng Gao (Shanghai Jiao Tong University), Bin Dai (Xiaobing.AI)
RetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: An end-to-end retrieval-augmented generation framework named SmartRAG is proposed, which can simultaneously decide when to retrieve, how to construct retrieval queries, and how to utilize retrieval results to generate answers.
SMI-Editor: Edit-based SMILES Language Model with Fragment-level Supervision
Kangjie Zheng (Peking University), Ming Zhang (Peking University)
Drug DiscoveryTransformerTextGraph
🎯 What it does: A SMILES language model based on editing, called SMI-EDITOR, is designed to randomly drop molecular substructures and allow the model to recover the original SMILES through editing operations, thus addressing the rapid saturation and substructure semantic loss issues of traditional MLMs.
SMITE: Segment Me In TimE
Amirhossein Alimohammadi (Simon Fraser University), Ali Mahdavi Amiri (Google DeepMind)
Object TrackingSegmentationDiffusion modelVideo
🎯 What it does: The SMITE method is proposed, utilizing the attention mechanism of a pre-trained text-to-image diffusion model and a spatiotemporal dilated UNet, combined with tracking voting and low-pass regularization, to achieve multi-granularity and temporally consistent segmentation of unseen videos using only a small number of reference images.
Smoothing the Shift: Towards Stable Test-Time Adaptation under Complex Multimodal Noises
Zirun Guo (Zhejiang University), Tao Jin (Zhejiang University)
Domain AdaptationVideoMultimodalityAudio
🎯 What it does: Proposes the task of adaptive wild test-time adaptation (wild TTA) and designs the SuMi method to achieve stable adaptation.
SMT: Fine-Tuning Large Language Models with Sparse Matrices
Haoze He (Carnegie Mellon University), Heather Miller (University of California, Berkeley)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: For fine-tuning large language models, a Sparse Matrix Tuning (SMT) method is proposed that only updates sparse submatrices, significantly reducing memory and computational overhead;
SOAP: Improving and Stabilizing Shampoo using Adam for Language Modeling
Nikhil Vyas (Harvard University), Sham M. Kakade
OptimizationTransformerLarge Language ModelText
🎯 What it does: An optimizer named SOAP is proposed, which combines the feature vector space of the Shampoo preprocessor with the adaptive updates of AdamW;
SoftCVI: Contrastive variational inference with self-generated soft labels
Daniel Ward (Bristol University), Matteo Fasiolo (Bristol University)
OptimizationContrastive Learning
🎯 What it does: A variational inference framework based on contrastive learning, SoftCVI, is proposed, which directly trains the variational distribution by generating soft labels for each sampling, resulting in unbiased, low-variance gradients.
SoftMatcha: A Soft and Fast Pattern Matcher for Billion-Scale Corpus Searches
Hiroyuki Deguchi (NAIST), Sho Yokoi (NINJAL)
RetrievalComputational EfficiencyText
🎯 What it does: A soft pattern matching algorithm that combines word vectors and inverted indexing is proposed, capable of achieving second-level retrieval on a billion-level corpus;
Solving Differential Equations with Constrained Learning
Viggo Moro (University of Oxford), Luiz F. O. Chamon (École polytechnique)
OptimizationPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A PDE solving framework based on constraint learning, SCL, is proposed and implemented, capable of uniformly handling unsupervised, supervised, and scenarios with prior knowledge (structure, measurements, known solutions).
Solving hidden monotone variational inequalities with surrogate losses
Ryan D'Orazio (Mila Quebec AI Institute, Universite Montreal), Gauthier Gidel (Mila Quebec AI Institute, Universite Montreal)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a deep learning framework based on surrogate loss for solving variational inequality (VI) problems with hidden monotonic structures;
Solving New Tasks by Adapting Internet Video Knowledge
Calvin Luo (Brown University), Chen Sun (Brown University)
GenerationRobotic IntelligenceDiffusion modelVideo
🎯 What it does: This study explores how to combine large-scale video generation models from the internet (such as AnimateDiff) with a small number of domain-specific demonstrations, achieving text-conditioned new task generalization for robotic tasks through adaptation techniques, and proposes the Inverse Probabilistic Adaptation method.
Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model
Longrong Yang (Zhejiang University), Xi Li (Kuaishou Technology)
Mixture of ExpertsVision Language ModelMultimodality
🎯 What it does: Identifying and eliminating gradient conflicts in Mixture-of-Experts through token-level gradient analysis, thereby enhancing the performance of large visual language models.
Solving Video Inverse Problems Using Image Diffusion Models
Taesung Kwon (KAIST), Jong Chul Ye (KAIST)
RestorationOptimizationDiffusion modelVideo
🎯 What it does: This paper proposes a method to solve video inverse problems using only image diffusion models, treating the time dimension as a batch of images for reverse diffusion sampling.
SONICS: Synthetic Or Not - Identifying Counterfeit Songs
Md Awsafur Rahman (University of California Santa Barbara), Shaikh Anowarul Fattah (Bangladesh University of Engineering and Technology)
ClassificationData SynthesisComputational EfficiencyTransformerAudio
🎯 What it does: A large-scale end-to-end synthetic song detection dataset, SONICS, has been constructed, and an efficient SpecTTTra model for capturing long-term dependencies has been proposed, evaluating its performance in distinguishing between artificial and real songs.
SonicSim: A customizable simulation platform for speech processing in moving sound source scenarios
Kai Li (Tsinghua University), Xiaolin Hu (Tsinghua University)
RestorationGenerationData SynthesisAudio
🎯 What it does: A customizable mobile sound source simulation tool called SonicSim based on Habitat-sim has been developed, and it has been used to generate a large-scale mobile sound source speech separation/enhancement dataset called SonicSet.
SOO-Bench: Benchmarks for Evaluating the Stability of Offline Black-Box Optimization
Hong Qian (East China Normal University), Yang Yu (Nanjing University)
OptimizationTabularBenchmark
🎯 What it does: The SOO-Bench benchmark suite is proposed, providing customizable narrow distribution offline datasets and stability-optimality metrics to systematically evaluate the stability and optimality of offline black-box optimization algorithms under different data distributions.
SOREL: A Stochastic Algorithm for Spectral Risks Minimization
Yuze Ge (Fudan University), Rujun Jiang (Fudan University)
OptimizationTabular
🎯 What it does: The SOREL algorithm is proposed for achieving stochastic gradient optimization in the spectral risk minimization problem and provides convergence guarantees.
SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal
Tinghao Xie (Princeton University), Prateek Mittal (Princeton University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes SORRY-Bench, a benchmark for systematically evaluating large language models (LLMs) on their ability to refuse unsafe requests;
Sort-free Gaussian Splatting via Weighted Sum Rendering
Qiqi Hou (Qualcomm AI Research), Amir Said (Qualcomm AI Research)
Computational EfficiencyGaussian SplattingPoint Cloud
🎯 What it does: An efficient 3D Gaussian dispersion rendering method without deep sorting is proposed.
SoundCTM: Unifying Score-based and Consistency Models for Full-band Text-to-Sound Generation
Koichi Saito (Sony AI), Yuki Mitsufuji (Sony Group Corporation)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelScore-based ModelAudio
🎯 What it does: The SoundCTM model is proposed, achieving high-speed generation in one step and high-quality generation in multiple steps, while maintaining semantic consistency through deterministic sampling.
SPA-BENCH: A COMPREHENSIVE BENCHMARK FOR SMARTPHONE AGENT EVALUATION
Jingxuan Chen (Huawei Noah's Ark Lab), Kun Shao (Huawei Noah's Ark Lab)
Large Language ModelAgentic AITextMultimodalityBenchmark
🎯 What it does: This paper presents SPA-BENCH, a comprehensive smartphone proxy evaluation benchmark covering English, Chinese, and third-party applications.
SPA: 3D Spatial-Awareness Enables Effective Embodied Representation
Haoyi Zhu (University of Science and Technology of China), Tong He (Shanghai AI Lab)
Representation LearningRobotic IntelligenceTransformerContrastive LearningImageMultimodality
🎯 What it does: Proposes the SPA framework, utilizing differentiable neural rendering of multi-view images for pre-training 3D spatial perception representations in Vision Transformer (ViT);
SpaceGNN: Multi-Space Graph Neural Network for Node Anomaly Detection with Extremely Limited Labels
Xiangyu Dong (Chinese University of Hong Kong), Sibo Wang (Chinese University of Hong Kong)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: The research focuses on node anomaly detection (NAD) with very few labels and proposes a multi-space graph neural network framework called SpaceGNN, aimed at better utilizing graph structural information and improving detection performance in scenarios with limited labels.
SPAM: Spike-Aware Adam with Momentum Reset for Stable LLM Training
Tianjin Huang (University of Exeter), Shiwei Liu (University of Oxford)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study addresses and mitigates gradient and loss spikes during the training of large language models, proposing an optimizer called SPAM that features momentum reset and spike-aware clipping, along with a low-memory implementation of sparse momentum.
SPaR: Self-Play with Tree-Search Refinement to Improve Instruction-Following in Large Language Models
Jiale Cheng (Tsinghua University), Minlie Huang (Tsinghua University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: A SPAR self-play + tree search self-improvement framework is proposed, which enhances the instruction-following ability of LLMs through self-generated and self-evaluated preference pairs by the actor and evaluator.
Sparse Autoencoders Do Not Find Canonical Units of Analysis
Patrick Leask (Durham University), Neel Nanda
Large Language ModelAuto Encoder
🎯 What it does: Evaluate the feature integrity and atomicity of different sizes of Sparse Autoencoders (SAE), and propose methods for SAE stitching and Meta-SAE disassembly.
Sparse autoencoders reveal selective remapping of visual concepts during adaptation
Hyesu Lim (KAIST AI), Steffen Schneider (Helmholtz Munich)
Domain AdaptationExplainability and InterpretabilityRepresentation LearningTransformerPrompt EngineeringAuto EncoderImage
🎯 What it does: Train PatchSAE on CLIP ViT to extract interpretable visual concepts and achieve spatial localization, and then analyze the relationship between concepts and task categories during the prompt-based adaptation process using this model.
Sparse Autoencoders Reveal Temporal Difference Learning in Large Language Models
Can Demircan (Helmholtz Munich), Eric Schulz (Helmholtz Munich)
TransformerLarge Language ModelReinforcement LearningAuto EncoderGraphSequential
🎯 What it does: The paper demonstrates that large language models (such as Llama-70B) can achieve reinforcement learning through contextual learning, with the model showing the ability to learn reward, value function, and temporal difference error information across three tasks (Two-Step, Grid World, Graph Prediction).
Sparse components distinguish visual pathways & their alignment to neural networks
Ammar I Marvi (Massachusetts Institute of Technology), Meenakshi Khosla (University of California San Diego)
Convolutional Neural NetworkTransformerImageMagnetic Resonance Imaging
🎯 What it does: This paper uses Bayesian Non-negative Matrix Factorization (Bayesian NMF) to extract three visual pathways (dorsal, lateral, and ventral) from human NSD fMRI data without prior assumptions, and proposes Sparse Component Alignment (SCA) to measure the alignment of artificial neural networks with brain representations. It finds that standard visual DNNs align better with the ventral pathway in SCA, while the alignment with the lateral and dorsal pathways is significantly weaker.
Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Samuel Marks (Northeastern University), Aaron Mueller (Northeastern University)
Explainability and InterpretabilityTransformerAuto EncoderText
🎯 What it does: This paper proposes a method called Sparse Feature Circuits, which utilizes sparse autoencoders to extract interpretable features and calculates indirect effects through linear approximation, automatically discovering causal subgraphs related to language model behavior. Based on this circuit, feature pruning (SHIFT) is performed to eliminate unintended signals, while achieving unsupervised large-scale behavior discovery.
Sparse Learning for State Space Models on Mobile
Xuan Shen (Northeastern University), Wei Niu (University of Georgia)
OptimizationComputational EfficiencyText
🎯 What it does: This paper proposes a sparse learning framework based on the C_n^4 structure, combining compiler optimization to achieve efficient inference of the Mamba model on mobile devices.
SparsyFed: Sparse Adaptive Federated Learning
Adriano Guastella (Universita di Bologna), Nicholas Donald Lane
Federated LearningConvolutional Neural NetworkImageAudio
🎯 What it does: Proposes SparsyFed, a sparse training method suitable for cross-device federated learning, achieving a 95% sparse model with only one hyperparameter.
SPARTUN3D: Situated Spatial Understanding of 3D World in Large Language Model
Yue Zhang (Michigan State University), Lifu Huang (University of California Davis)
Data SynthesisRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityPoint Cloud
🎯 What it does: A scalable LLM-based coordinate 3D dataset, Spartun3D, has been constructed, and based on this, improvements have been made to the 3D visual language model to achieve situated spatial understanding.
Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion
Chaodong Xiao (Hong Kong Polytechnic University), Lei Zhang (Xi'an Jiaotong University)
Object DetectionSegmentationConvolutional Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: Spatial-Mamba is proposed in visual tasks, capturing spatial dependencies in images by introducing structure-aware state fusion in the state space.
SPDIM: Source-Free Unsupervised Conditional and Label Shift Adaptation in EEG
Shanglin Li (Nara Institute of Science and Technology), Reinmar J Kobler
Domain AdaptationFlow-based ModelTime SeriesBiomedical Data
🎯 What it does: A source-free unsupervised domain adaptation framework called SPDIM is proposed for cross-domain generalization of electroencephalography (EEG) data in the presence of label and conditional distribution shifts.
Specialized Foundation Models Struggle to Beat Supervised Baselines
Zongzhe Xu (Carnegie Mellon University), Mikhail Khodak (Princeton University)
Hyperparameter SearchDrug DiscoveryNeural Architecture SearchConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: Evaluated whether foundational models in specialized fields (genomics, satellite imaging, time series) can outperform traditional supervised learning, and proposed two automated workflows for systematic comparison.
Spectral Compressive Imaging via Unmixing-driven Subspace Diffusion Refinement
Haijin Zeng (Harvard University), Yong Xu (Harbin Institute of Technology)
RestorationCompressionDiffusion modelImage
🎯 What it does: A prediction-based mixed decoupled subspace diffusion framework (PSR-SCI) is proposed, which first uses a lightweight predictor to generate a coarse multispectral image, then decomposes it into low-dimensional abundance maps and spectral coefficients through a reversible spectral embedding module, and subsequently reconstructs high-frequency details in the low-dimensional subspace using a pre-trained RGB diffusion model, ultimately restoring the complete multispectral image.
Spectral-Refiner: Accurate Fine-Tuning of Spatiotemporal Fourier Neural Operator for Turbulent Flows
Shuhao Cao (University of Missouri Kansas City), Yuanzhe Xi (Emory University)
OptimizationComputational EfficiencyTime SeriesPhysics Related
🎯 What it does: A new spatiotemporal Fourier neural operator (ST-FNO) and a hybrid training paradigm have been designed to accurately solve the turbulent Navier-Stokes equations.
Spectro-Riemannian Graph Neural Networks
Karish Grover (Carnegie Mellon University), Christos Faloutsos (Carnegie Mellon University)
ClassificationRecommendation SystemGraph Neural NetworkGraph
🎯 What it does: A graph neural network that combines spectral filtering with Riemannian geometry—CUSP—is proposed for node classification and link prediction tasks.
Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling
Wenda Xu (University of California Santa Barbara), Tomas Pfister (Google Cloud AI Research)
Knowledge DistillationSupervised Fine-TuningText
🎯 What it does: Proposes Speculative Knowledge Distillation (SKD), which addresses the shortcomings of supervised KD and self-policy KD through teacher-student interleaved sampling.
Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting
Zilong Wang (University of California), Tomas Pfister (Google Cloud AI)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The SPECULATIVE RAG framework is proposed, which splits the RAG task into generating multiple answer drafts in parallel by a small specialized LLM, and then validating and integrating them through self-consistency and self-reflection scoring by a large general LLM, resulting in more accurate and low-latency answers.
Speech Robust Bench: A Robustness Benchmark For Speech Recognition
Muhammad A Shah, Nicolas Kourtellis (Telefonica Research)
RecognitionAdversarial AttackBenchmarkAudio
🎯 What it does: This paper presents the Speech Robust Bench (SRB), a comprehensive robustness benchmark covering 114 types of speech recognition challenge scenarios.