NeurIPS 2025 Papers — Page 41
Conference on Neural Information Processing Systems · 5275 papers
Sample-Conditional Coverage in Split-Conformal Prediction
John Duchi (Stanford University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper studies how to construct approximate conditional coverage prediction confidence sets within the split conformal framework, and provides theoretical guarantees for high-probability sample conditional coverage.
Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions
Hidde Fokkema (University of Amsterdam), Sara Magliacane (University of Amsterdam)
Representation LearningAuto EncoderImage
🎯 What it does: A two-step framework is proposed: first, causal representation learning (CRL) is used to obtain identifiable causal variables, and then a mapping is learned through a small number of concept labels to obtain a reliable concept bottleneck model.
Sample-Efficient Multi-Round Generative Data Augmentation for Long-Tail Instance Segmentation
Byunghyun Kim (Korea Advanced Institute of Science and Technology), Jae-Gil Lee (Korea Advanced Institute of Science and Technology)
Object DetectionSegmentationGenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: A multi-round collaborative augmentation framework (MRCA) is proposed for the long-tail instance segmentation task, utilizing feedback from the instance segmentation model to dynamically guide the diffusion model in generating synthetic objects of high uncertainty and rare classes, achieving highly efficient augmentation through budget optimization and size adjustment.
Sample-Efficient Tabular Self-Play for Offline Robust Reinforcement Learning
Na Li (Zhejiang University), Xinyu Li (Huazhong University of Science and Technology)
Reinforcement LearningTabular
🎯 What it does: A model-based algorithm RTZ-VI-LCB for offline robust two-player zero-sum Markov games (RTZMG) is designed, providing an approximate optimal sample complexity upper bound;
Sampled Estimators For Softmax Must Be Biased
Li-Chung Lin (National Taiwan University), Chih-Jen Lin (National Taiwan University)
🎯 What it does: The paper studies the bias problem when using sampled softmax in large category problems and proves that any estimator that only utilizes a sampled subset cannot unbiasedly estimate the complete softmax.
Sampling 3D Molecular Conformers with Diffusion Transformers
Thorben Frank, Stefan Chmiela (Technical University Berlin)
GenerationDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: A Diffusion Transformer framework named DiTMC is proposed for generating three-dimensional molecular conformations.
Sampling by averaging: A multiscale approach to score estimation
Paula Cordero Encinar, O. Deniz Akyildiz (Imperial College London)
GenerationData SynthesisOptimizationScore-based ModelPhysics RelatedStochastic Differential Equation
🎯 What it does: A training-free multi-scale sampling framework is proposed, designed with two samplers, MULTALMC and MULTCDIFF, for efficiently generating samples from unnormalized complex target distributions.
Sampling from multi-modal distributions with polynomial query complexity in fixed dimension via reverse diffusion
Adrien Vacher (CREST ENSAE Institut Polytechnique de Paris), Anna Korba (CREST ENSAE Institut Polytechnique de Paris)
Diffusion modelScore-based Model
🎯 What it does: A reverse diffusion sampling algorithm for low-dimensional multimodal distributions is proposed, achieving polynomial query complexity for multimodal distributions at fixed dimensions.
Sampling-Efficient Test-Time Scaling: Self-Estimating the Best-of-N Sampling in Early Decoding
Yiming Wang (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The paper proposes an ST-BoN decoding method that allows large language models to perform inference by self-estimating and truncating the least optimal samples in advance, without needing to fully generate N samples or relying on a reward model.
SAMPO: Scale-wise Autoregression with Motion Prompt for Generative World Models
Sen Wang (Xi'an Jiaotong University), Gang Hua
GenerationRobotic IntelligenceTransformerWorld ModelVideo
🎯 What it does: A scale autoregressive world model SAMPO is proposed, combining spatiotemporal multi-scale generation and trajectory-aware prompts to achieve high-quality action-conditioned video prediction and robot control.
SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation
Claudia Cuttano (Politecnico di Torino), Carlo Masone (Politecnico di Torino)
SegmentationTransformerPrompt EngineeringImage
🎯 What it does: This study investigates the potential semantic structure of SAM2 in few-shot segmentation and proposes the SANSA framework, which transforms SAM2 from visual matching to semantic matching through a lightweight adapter. It utilizes its built-in Memory Attention mechanism to achieve k-shot segmentation tasks with a single inference for various prompts.
SAO-Instruct: Free-form Audio Editing using Natural Language Instructions
Michael Ungersböck (ETH Zurich), Roger Wattenhofer (ETH Zurich)
GenerationData SynthesisOptimizationLarge Language ModelPrompt EngineeringDiffusion modelAudio
🎯 What it does: Train the SAO-Instruct model to achieve audio editing based on free-form natural language instructions, capable of making various modifications to audio while preserving the original background.
SAP: Exact Sorting in Splatting via Screen-Aligned Primitives
Zhanke Wang (Peking University), Ronggang Wang (Peking University)
Image TranslationOptimizationGaussian SplattingPoint Cloud
🎯 What it does: Proposes the Screen-Aligned Primitives (SAP) framework, which decomposes the representation of 3D scenes into a 3D consistent decoder and view-specific 2D primitives, achieving precise pixel-level sorting.
SAS: Simulated Attention Score
Chuanyang Zheng (Morgan Stanley), Jianfeng Gao (Microsoft Research)
TransformerLarge Language ModelTextFinance Related
🎯 What it does: This paper proposes the Simulated Attention Score (SAS), which simulates the attention mechanism of larger models through nonlinear mapping of the head dimension and feature dimension, thereby enhancing the expressive capability and training efficiency of the Transformer while keeping the model size unchanged.
SATURN: SAT-based Reinforcement Learning to Unleash LLMs Reasoning
Huanyu Liu (Peking University), Ge Li (Peking University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A SAT-based reinforcement learning framework called SATURN is proposed to enhance the reasoning capabilities of large language models.
SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing
Mingfei Chen (University of Washington), Eli Shlizerman (University of Washington)
Large Language ModelSimultaneous Localization and MappingVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Proposed SAVVY-Bench and the untrained SAVVY inference pipeline to evaluate and enhance the spatial reasoning capabilities of AV-LLM in dynamic 3D audio-visual environments.
Scaffolding Dexterous Manipulation with Vision-Language Models
Vincent de Bakker (Stanford University), Dorsa Sadigh (Stanford University)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelImageText
🎯 What it does: Using a visual language model (VLM) to generate 3D keypoint trajectories from language instructions and images as a coarse-grained scaffold, and then training a robotic hand in simulation to perform various complex tasks such as grasping, opening doors, and pushing objects through residual reinforcement learning, without the need for demonstrations or manual rewards.
Scalable and adaptive prediction bands with kernel sum-of-squares
Louis Allain (Safran), Brian Staber (Safran)
Hyperparameter SearchTabular
🎯 What it does: This paper proposes a scalable adaptive prediction interval construction method based on the kernel sum-of-squares framework, utilizing split conformal prediction to achieve distribution-independent coverage guarantees.
Scalable and Cost-Efficient de Novo Template-Based Molecular Generation
Piotr Gaiński (Jagiellonian University), Michał Koziarski (Hospital for Sick Children Research Institute)
GenerationDrug DiscoveryReinforcement LearningTabular
🎯 What it does: The SCENT framework is proposed, combining template-based synthesis with GFlowNet for synthesizable and low-cost molecular generation, addressing the three major challenges of synthesis cost, library scale, and small fragment sets.
Scalable Best-of-N Selection for Large Language Models via Self-Certainty
Zhewei Kang, Dawn Song
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Designed and evaluated a self-certainty metric based on the probability distribution of LLM itself for Best-of-N selection, and proposed a self-certainty weighting method based on Borda voting.
Scalable Cross-View Sample Alignment for Multi-View Clustering with View Structure Similarity
Jun Wang (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
OptimizationContrastive LearningMultimodality
🎯 What it does: A scalable cross-view sample alignment multi-view clustering method SSA-MVC is proposed to address the issue of misalignment of samples across different views.
Scalable Evaluation and Neural Models for Compositional Generalization
Giacomo Camposampiero (IBM Research), Abbas Rahimi (IBM Research)
ClassificationRecognitionData SynthesisComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes a scalable compositional generalization evaluation framework and Attribute Invariant Networks (AIN), achieving efficient compositional generalization capabilities in visual models.
Scalable Exploration via Ensemble++
Yingru Li (Chinese University of Hong Kong), Zhi-Quan Luo (Chinese University of Hong Kong)
TransformerReinforcement Learning
🎯 What it does: The Ensemble++ framework is proposed, which achieves scalable approximate Thompson Sampling using shared factors and random linear combinations;
Scalable Feature Learning on Huge Knowledge Graphs for Downstream Machine Learning
Félix Lefebvre (Inria), Gaël Varoquaux (Inria)
OptimizationRepresentation LearningGraph Neural NetworkGraphTabular
🎯 What it does: This paper studies a scalable knowledge graph embedding algorithm called SEPAL, which generates vector features suitable for downstream machine learning tasks on large-scale knowledge graphs.
Scalable Fingerprinting of Large Language Models
Anshul Nasery (University of Washington), Sewoong Oh (Sentient)
OptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study investigates the scalability of large model fingerprints, proposing the Perinucleus sampling method to generate fingerprints and inserting them into the model through regularization, verifying that tens of thousands of fingerprints can be inserted without degrading performance.
Scalable In-context Ranking with Generative Models
Nilesh Gupta (University of Texas at Austin), Felix X. Yu
RetrievalOptimizationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes BlockRank, an efficient In-Context Ranking method for large language models, which utilizes structured sparse attention and contrastive auxiliary attention loss to achieve relevance ranking of candidate lists in a single forward pass.
Scalable inference of functional neural connectivity at submillisecond timescales
Arina Medvedeva (Flatiron Institute), Stephen L Keeley (Fordham University)
Time SeriesStochastic Differential Equation
🎯 What it does: A continuous-time Poisson GLM is proposed, which fits large-scale neural spike data using two methods: Monte Carlo sampling and polynomial approximation, capable of extracting functional connectivity filters at sub-millisecond levels.
Scalable Neural Incentive Design with Parameterized Mean-Field Approximation
Nathan Corecco (ETH Zurich), Niao He (ETH Zurich)
OptimizationReinforcement LearningTabular
🎯 What it does: To address the incentive design problem in multi-agent systems, the authors propose using Parameterized Mean Field Games (PMFG) to approximate multi-player games, thereby obtaining feasible incentive parameters in large-scale scenarios (N→∞).
Scalable Neural Network Geometric Robustness Validation via Hölder Optimisation
Yanghao Zhang (Safe Intelligence), Alessio Lomuscio (Safe Intelligence)
OptimizationConvolutional Neural NetworkTransformerImageVideoAudio
🎯 What it does: Proposes the H₂V method to verify the local robustness of neural networks under geometric perturbations.
Scalable Policy-Based RL Algorithms for POMDPs
Ameya Anjarlekar (University of Illinois Urbana-Champaign), R. Srikant (University of Illinois Urbana-Champaign)
OptimizationReinforcement Learning
🎯 What it does: An approximate model is proposed that maps POMDP to a finite history MDP (Superstate MDP), and standard policy optimization algorithms are used to learn an approximately optimal policy on this model.
Scalable Signature Kernel Computations via Local Neumann Series Expansions
Matthew Tamayo-Rios (ETH Zurich), Rima Alaifari (RWTH Aachen University)
OptimizationComputational EfficiencyTime SeriesSequentialFinance Related
🎯 What it does: A scalable signature kernel computation method based on local Neumann series expansion (PowerSig) is proposed, capable of handling high-dimensional long sequences with millions of points on a single GPU;
Scalable Valuation of Human Feedback through Provably Robust Model Alignment
Masahiro Fujisawa (University of Osaka), Michael A Osborne
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper proposes Hölder-DPO, a variant of Direct Preference Optimization (DPO) based on Hölder divergence, which achieves provable redescending robustness under label noise and automatically estimates and locates mislabeling.
Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching
Zhong Li (Leiden University), Matthijs van Leeuwen (Leiden University)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyFlow-based ModelTabularBenchmark
🎯 What it does: A time-conditioned contraction matching (TCCM) algorithm based on flow matching is proposed for semi-supervised tabular data anomaly detection.
Scale-invariant attention
Ben Anson (University of Bristol), Laurence Aitchison (University of Bristol)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes the concept of scale-invariant attention and presents two types of scale-invariant properties: total attention remains unchanged and attention sparsity remains unchanged. Based on this, a transformation is proposed that applies position-dependent scaling and bias to the logarithm of attention, with theoretical proof that it satisfies the scale-invariant condition, and it is compared with common long-sequence attention methods in experiments.
ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion
Sungho Koh (Hanyang University), Dong-Jin Kim (Hanyang University)
GenerationData SynthesisTransformerDiffusion modelImageStochastic Differential Equation
🎯 What it does: A framework named ScaleDiff is proposed, which can extend pre-trained diffusion models to generate higher resolution images without additional training.
Scaling and context steer LLMs along the same computational path as the human brain
Joséphine Raugel (Meta AI), Jean-Remi King
TransformerLarge Language ModelTextAudio
🎯 What it does: This paper explores whether the sequence of representation generation in large language models (LLMs) is similar to that of the human brain in natural speech processing by comparing the brain responses of three subjects during 10 hours of audiobook listening in MEG with the hierarchical activations of 17 LLMs.
Scaling can lead to compositional generalization
Florian Redhardt (ETH Zurich), Simon Schug (Princeton University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The study achieves generalization over a family of composable tasks by expanding the data volume and model scale (mainly multi-layer perceptrons) and demonstrates that under a training distribution that sufficiently covers the task space, the network can perform well on new combinatorial tasks.
Scaling Code-Assisted Chain-of-Thoughts and Instructions for Model Reasoning
Honglin Lin (Shanghai Jiao Tong University), Lijun Wu (Shanghai Artificial Intelligence Laboratory)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: A Caco framework was constructed to automatically generate verifiable, scalable, and diverse reasoning data for training large language models through code-assisted chain reasoning.
Scaling Data-Driven Probabilistic Robustness Analysis for Semantic Segmentation Neural Networks
Navid Hashemi (Vanderbilt University), Taylor T Johnson
SegmentationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Conduct a scalable, data-driven probabilistic robustness analysis of semantic segmentation networks (SSN), utilizing Conformal Inference and surrogate models to construct observable starsets and determine the robust/non-robust/unknown status of pixels.
Scaling Diffusion Transformers Efficiently via $\mu$P
Chenyu Zheng (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationComputational EfficiencyHyperparameter SearchTransformerDiffusion modelImageText
🎯 What it does: This paper extends Maximum Update Parameterization (µP) from standard Transformers to diffusion Transformers and validates its feasibility on various mainstream diffusion models.
Scaling Embedding Layers in Language Models
Da Yu (Google), Chiyuan Zhang (Google)
TransformerLarge Language ModelText
🎯 What it does: The SCONE method is proposed, which utilizes a pre-trained f-gram transformer to learn n-gram context-aware embeddings, and caches these embeddings in accelerator external storage during the inference phase to scale the input embedding layer without increasing FLOPs and memory usage during inference.
Scaling Epidemic Inference on Contact Networks: Theory and Algorithms
Guanghui Min (University of Virginia), Chen Chen (University of Virginia)
Graph
🎯 What it does: This paper proposes a fast infection probability inference method based on residual propagation, called RAPID, for large-scale contact networks under the SIR model.
Scaling Image Geo-Localization to Continent Level
Philipp Lindenberger (ETH Zurich), Eduard Trulls (Google)
RetrievalTransformerContrastive LearningImage
🎯 What it does: A hybrid method combining classification and cross-view retrieval is proposed, achieving large-scale (continental-level) geographic localization with an accuracy of approximately 100–200 meters.
Scaling Language-centric Omnimodal Representation Learning
Chenghao Xiao (DAMO Academy), Yu Rong (Alibaba Group)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: This paper proposes a language-centered full-modal embedding framework LCO-EMB, which fine-tunes a multimodal large language model (MLLM) using contrastive learning with only text to enhance cross-modal representation quality, and explores the Generative-Representation Scaling Law (GRSL).
Scaling Law with Learning Rate Annealing
Howe Tissue, Lu Wang
OptimizationLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes a full curve scaling rule that considers learning rate annealing during the training process, which can accurately predict the validation loss curve under different learning rate schedules.
Scaling Laws for Gradient Descent and Sign Descent for Linear Bigram Models under Zipf’s Law
Frederik Kunstner (Inria), Francis Bach (Inria)
OptimizationText
🎯 What it does: This study investigates the convergence scaling laws of gradient descent (GD) and symbolic descent (SD) in linear binary models (next-token prediction) under the Zipf distribution, revealing the impact of different power law exponents α on convergence speed.
Scaling Laws for Optimal Data Mixtures
Mustafa Shukor, Pierre Ablin
OptimizationData-Centric LearningTransformerLarge Language ModelTextMultimodality
🎯 What it does: A predictive scaling law for multi-domain data mixing ratios is proposed, which quickly infers the optimal data mixing ratio under a given computational budget based on a small number of small-scale experiments for the three major pre-training tasks: LLM, NMM, and LVM. Its effectiveness is then validated on large-scale models.
Scaling Laws for Robust Comparison of Open Foundation Language-Vision Models and Datasets
Marianna Nezhurina (LAION), Jenia Jitsev (LAION)
ClassificationSegmentationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This study systematically compares the language-vision models CLIP and MaMMUT and their performance on different open datasets by deriving scaling laws, aiming to determine the advantages and disadvantages of the pre-training process.
Scaling Laws For Scalable Oversight
Joshua Engels (Massachusetts Institute of Technology), Max Tegmark (Massachusetts Institute of Technology)
Large Language ModelReinforcement Learning
🎯 What it does: This paper proposes a scalable oversight method for quantifying weakly supervised strong AI, utilizing a two-player game model and the Elo rating framework. It derives the scalability law of supervision effectiveness based on the gap in general intelligence between the supervisor and the supervised, and validates this law in four supervision games (Mafia, Debate, Backdoor Code, Wargames). Furthermore, it establishes a theoretical model of Nested Scalable Oversight (NSO) and provides analytical and numerical results for the optimal number of supervision layers and success probabilities.
Scaling Off-Policy Reinforcement Learning with Batch and Weight Normalization
Daniel Palenicek (Technical University of Darmstadt), Jan Peters (University of Freiburg)
Reinforcement LearningSequential
🎯 What it does: Weight Normalization is incorporated into the CrossQ framework to stabilize training, improve sample efficiency, and achieve scalability for high UTD ratios, eliminating the dependency on network resets.
Scaling Offline RL via Efficient and Expressive Shortcut Models
Nicolas Espinosa-Dice (Cornell University), Wen Sun (Cornell University)
Robotic IntelligenceReinforcement LearningFlow-based Model
🎯 What it does: This paper proposes an offline reinforcement learning algorithm SORL that utilizes Shortcut Models, achieving high expressiveness in a single training session and supporting arbitrary inference computation budgets.
Scaling RL to Long Videos
Yukang Chen (NVIDIA), Song Han (MIT)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: A complete framework is proposed to enable visual language models (VLM) to perform deep reasoning on long videos, covering a large-scale long video reasoning dataset, two-stage training (chain-of-thought supervised fine-tuning + reinforcement learning), and multi-modal reinforcement learning parallel techniques for long videos.
Scaling Speculative Decoding with Lookahead Reasoning
Yichao Fu (University of California San Diego), Hao Zhang (University of California San Diego)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: A new method for accelerating large-scale inference models called 'LOOKAHEAD REASONING' is proposed, which reduces inference latency by speculative generation and semantic validation at the level of thinking steps.
Scaling Unlocks Broader Generation and Deeper Functional Understanding of Proteins
Aadyot Bhatnagar (Profluent Bio), Ali Madani (Profluent Bio)
GenerationProtein Structure PredictionTransformerLarge Language ModelMixture of ExpertsBiomedical Data
🎯 What it does: Trained and evaluated the sparse generative protein language model ProGen3, with parameters ranging from 112M to 46B. The study systematically explored the impact of model size on protein generation, diversity, laboratory expression rates, and functional predictions, and validated the expressibility of the generated proteins in the laboratory.
Scaling Up Active Testing to Large Language Models
Gabrielle Berrada (University of Oxford), Tom Rainforth (University of Oxford)
ClassificationLarge Language ModelPrompt EngineeringText
🎯 What it does: A cost-effective active testing framework is proposed and validated, utilizing a fixed agent model constructed through context learning with only a few examples, which reduces the number of gradient training and predictions between the agent and target models, and introduces a bootstrap estimator for real-time evaluation of risk estimation errors.
Scaling Up Liquid-Resistance Liquid-Capacitance Networks for Efficient Sequence Modeling
Mónika Farsang (Vienna University of Technology), Radu Grosu (Vienna University of Technology)
ClassificationComputational EfficiencyTime SeriesSequential
🎯 What it does: This paper proposes LrcSSM, a nonlinear state space model that transforms liquid resistor-capacitor networks (LRC) into a parallel computable form;
Scaling Up Parameter Generation: A Recurrent Diffusion Approach
Kai Wang (National University of Singapore), Yang You (National University of Singapore)
GenerationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: A framework called RPG based on cyclic diffusion is proposed, which can generate complete neural network models with hundreds of millions of parameters on a single GPU.
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
Jonas Geiping (ELLIS Institute Tübingen), Tom Goldstein (University of Maryland)
TransformerLarge Language ModelText
🎯 What it does: Designed and trained a Transformer model that can expand its depth on demand during inference, achieving adaptive reasoning through recursive computation in the latent space, thereby improving inference performance without increasing training data or specialized examples.
SCAN: Self-Denoising Monte Carlo Annotation for Robust Process Reward Learning
Yuyang Ding (Soochow University), Min Zhang (Soochow University)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the SCAN framework, which enhances the training effectiveness of the process reward model through self-denoising Monte Carlo annotations.
ScatterAD: Temporal-Topological Scattering Mechanism for Time Series Anomaly Detection
Tao Yin (Chongqing University), Meng Yan (Chongqing University)
Anomaly DetectionGraph Neural NetworkContrastive LearningTime Series
🎯 What it does: This paper proposes the ScatterAD method, which utilizes a spatiotemporal scatter mechanism for anomaly detection in industrial IoT multivariate time series data.
SceneDecorator: Towards Scene-Oriented Story Generation with Scene Planning and Scene Consistency
Quanjian Song (Monash University), Pheng-Ann Heng (Chinese University of Hong Kong)
GenerationTransformerVision Language ModelDiffusion modelText
🎯 What it does: This study proposes SceneDecorator, a training-free scene-oriented story generation framework that can automatically plan global scenes and local sub-scenes based on user themes while maintaining scene consistency across stories.
SceneDesigner: Controllable Multi-Object Image Generation with 9-DoF Pose Manipulation
Zhenyuan Qin (Fudan University), Henghui Ding (Fudan University)
GenerationPose EstimationReinforcement LearningImage
🎯 What it does: Proposes SceneDesigner, which supports 9-DoF pose control for multiple objects to generate images;
SceneForge: Enhancing 3D-text alignment with Structured Scene Compositions
Cristian Sbrolli (Politecnico di Milano), Matteo Matteucci (Politecnico di Milano)
Data SynthesisRetrievalTransformerLarge Language ModelContrastive LearningTextPoint Cloud
🎯 What it does: This paper proposes the SCENEFORGE framework, which stitches single 3D point clouds into multi-object scenes based on explicit spatial relationships and generates corresponding multi-object textual descriptions, thereby enhancing data diversity and difficulty in 3D-text contrastive learning.
SceneWeaver: All-in-One 3D Scene Synthesis with an Extensible and Self-Reflective Agent
Yandan Yang, Siyuan Huang
GenerationData SynthesisOptimizationTransformerLarge Language ModelAgentic AIVision Language ModelMultimodalityPoint CloudMesh
🎯 What it does: This paper presents SCENEWEAVER, a self-reflective agent framework that selects and invokes various 3D scene synthesis tools through an LLM planner, iteratively generating physically feasible, visually realistic, and user-instructed indoor 3D scenes in a reason-act-reflect cycle.
Scent of Knowledge: Optimizing Search-Enhanced Reasoning with Information Foraging
Hongjin Qian (Beijing Academy of Artificial Intelligence), Zheng Liu (Hong Kong Polytechnic University)
OptimizationTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: A reinforcement learning framework based on information foraging theory, InForage, is proposed to enable dynamic retrieval and information integration during the reasoning process of LLMs, encouraging the generation of high-quality intermediate retrieval results.
Schrödinger Bridge Matching for Tree-Structured Costs and Entropic Wasserstein Barycentres
Samuel Howard (University of Oxford), George Deligiannidis (University of Oxford)
OptimizationFlow-based ModelTabular
🎯 What it does: The TreeDSBM algorithm is proposed, utilizing Iterative Markov Fitting (IMF) to solve the Schrödinger Bridge (SB) problem under tree structures, and it is applied to the computation of discrete and continuous Wasserstein 2 barycenters.
scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration
Jianle Sun (Shanghai Artificial Intelligence Laboratory), Peng Ye (Guangzhou Laboratory)
OptimizationData-Centric LearningAuto EncoderGenerative Adversarial NetworkMultimodalityBiomedical Data
🎯 What it does: A framework for unpaired single-cell multi-omics data integration named scMRDR is proposed, which achieves the decoupling of shared and specific subspaces through β-VAE, and incorporates isometric regularization, adversarial alignment, and masked reconstruction loss.
SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodel LLMs
Jinhong Deng (University of Electronic Science and Technology of China), Yang He (Agency for Science, Technology and Research)
Computational EfficiencyTransformerVision Language ModelImageVideoMultimodality
🎯 What it does: A visual token pruning method named SCOPE is proposed, which preserves semantic integrity by considering both the significance of tokens and their semantic coverage.
Score-Based Diffusion Modeling for Nonparametric Empirical Bayes in Heteroscedastic Gaussian Mixtures
Gongyu Chen (University of California Berkeley), Ying Cui (University of California Berkeley)
Diffusion modelScore-based ModelStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes transforming the nonparametric Empirical Bayes problem of high-dimensional heteroscedastic Gaussian mixture models into a score learning framework based on diffusion processes, and utilizes multi-step reverse ODE for denoising.
Score-informed Neural Operator for Enhancing Ordering-based Causal Discovery
Jiyeon Kang (LG AI Research), Sungbin Lim (Korea University)
Diffusion modelScore-based ModelGraphTabularPhysics Related
🎯 What it does: This paper proposes SciNO, a functional diffusion model based on neural operators and probabilistic generative models, aimed at stabilizing the approximation of the Hessian diagonal of the log density, thereby enhancing causal ordering methods based on score matching.
SCoT: Unifying Consistency Models and Rectified Flows via Straight-Consistent Trajectories
zhangkai wu, Longbing Cao (Macquarie University)
GenerationKnowledge DistillationDiffusion modelRectified FlowImage
🎯 What it does: Distill the student model using noise-clear image pairs generated by a pre-trained diffusion model, learning both direct and consistent trajectories for fast sampling.
SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought
Guanghao Li (Tsinghua University), Chun Yuan (Tsinghua University)
Explainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: The Flow Chain-of-Thought (Flow CoT) framework and SCOUT fine-tuning method are proposed, achieving a progressively deepening cognitive trajectory through recursive reasoning on pre-trained LLMs.
scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery
Yiming Gao (Texas A and M University), Eric P. Xing (Carnegie Mellon University)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringBiomedical DataBenchmark
🎯 What it does: The SCPILOT framework is proposed to achieve 'omics-native reasoning' of large language models in single-cell RNA sequencing analysis, automating cell type annotation, developmental trajectory reconstruction, and transcription factor target prediction, while providing transparent reasoning pathways.
scSplit: Bringing Severity Cognizance to Image Decomposition in Fluorescence Microscopy
Ashesh, Florian Jug (Human Technopole)
RestorationConvolutional Neural NetworkImage
🎯 What it does: The scSPLIT method is proposed for the decomposition and removal of bleed-through in fluorescence microscopy images, capable of simultaneously handling structural separation of single-channel images.
Sculpting Features from Noise: Reward-Guided Hierarchical Diffusion for Task-Optimal Feature Transformation
Nanxu Gong (Arizona State University), Yanjie Fu (Arizona State University)
Reinforcement LearningDiffusion modelAuto EncoderTabular
🎯 What it does: A reward-guided hierarchical diffusion model DIFFT is proposed to automatically generate the task-optimal feature transformation sequence.
SD-KDE: Score-Debiased Kernel Density Estimation
Elliot L Epstein, Jerry Weihong Liu (Stanford University)
Diffusion modelScore-based ModelImage
🎯 What it does: A method is proposed to debias traditional kernel density estimation (KDE) using an estimated score function, called Score-Debised Kernel Density Estimation (SD-KDE);
SD-VLM: Spatial Measuring and Understanding with Depth-Encoded Vision-Language Models
Pingyi Chen (Zhejiang University), Jieping Ye (Alibaba Cloud Computing)
Object DetectionDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: This study investigates the insufficient spatial reasoning of VLM, proposing the MSMU dataset and deep position encoding to enhance spatial measurement and understanding.
SDPGO: Efficient Self-Distillation Training Meets Proximal Gradient Optimization
Tongtong Su (Tianjin Normal University), Fengbo Zheng (Tianjin Normal University)
OptimizationKnowledge DistillationTransformerImage
🎯 What it does: A gradient-based self-distillation framework SDPGO is proposed, which utilizes proximal gradient optimization to dynamically weight features and achieves self-distillation through sequential iterative learning.
SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction
Fabian Immel (FZI Research Center for Information Technology), Christoph Stiller (Karlsruhe Institute of Technology)
Autonomous DrivingTransformerContrastive LearningGraph
🎯 What it does: A model called SDTagNet is proposed for online HD map construction using the full textual annotations from OpenStreetMap and all map elements.
SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents
Yifu Guo (Sun Yat-sen University), Mingguang Chen (University of Toronto)
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes SE-Agent, a self-evolving framework that enhances the reasoning quality of LLM-based agents in complex tasks by revising, recombining, and refining multi-step reasoning trajectories.
SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning
Xinbin Yuan (Nankai University), Bo Li (Vivo Mobile Communication Company)
Reinforcement LearningVision Language ModelImage
🎯 What it does: A training set consisting of only 3,000 high-quality samples was constructed, and reinforcement learning was used for self-evolution refinement, enhancing the visual localization capability of the GUI agent.
SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery
Zhenqi He (Visual AI Lab, University of Hong Kong), Kai Han (Visual AI Lab, University of Hong Kong)
Knowledge DistillationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: In the task of general category discovery, the SEAL framework is proposed, utilizing natural semantic hierarchies to achieve multi-granularity joint learning, cross-granularity consistency distillation, and hierarchical soft contrastive learning.
Search and Refine During Think: Facilitating Knowledge Refinement for Improved Retrieval-Augmented Reasoning
Yaorui Shi (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
RetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: The AutoRefine framework is proposed in retrieval-augmented reasoning, improving the reasoning quality of LLMs by incorporating an explicit knowledge refinement step between retrieval steps.
Searching Efficient Semantic Segmentation Architectures via Dynamic Path Selection
Yuxi Liu (Hunan University), Yaonan Wang (Hunan University)
SegmentationAutonomous DrivingNeural Architecture SearchImage
🎯 What it does: A Dynamic Path Selection (DPS) strategy is proposed for one-shot NAS semantic segmentation networks, dynamically focusing on path evaluation and training across three stages: convergence, expressiveness, and generalization.
Searching Latent Program Spaces
Matthew Macfarlane, Clément Bonnet (Ndea)
Meta LearningTransformerAuto EncoderText
🎯 What it does: Proposes the Latent Program Network (LPN), which learns an implicit program space in neural networks and adapts through gradient search during testing.
SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks
Hwiwon Lee (University of Illinois Urbana-Champaign), LINGMING ZHANG
Large Language ModelAgentic AIBenchmark
🎯 What it does: An automated evaluation framework for LLM security agents, SEC-bench, has been constructed, capable of automatically collecting, reproducing, and verifying vulnerabilities from real C/C++ CVEs.
SECA: Semantically Equivalent and Coherent Attacks for Eliciting LLM Hallucinations
Buyun Liang (University of Pennsylvania), Rene Vidal
GenerationOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a zero-order optimization method named SECA, which induces hallucinations in large language models (LLMs) by generating semantically equivalent and coherent rephrased prompts.
SeCon-RAG: A Two-Stage Semantic Filtering and Conflict-Free Framework for Trustworthy RAG
Xiaonan si, Xiaojun Jia (Nanyang Technological University)
RetrievalAdversarial AttackTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The SeCon-RAG framework is proposed, which employs a two-stage semantic filtering and conflict identification to enhance the robustness of RAG against poisoning attacks in the corpus.
Second-Order Convergence in Private Stochastic Non-Convex Optimization
Youming Tao (Technical University of Berlin), Di Wang (King Abdullah University of Science and Technology)
OptimizationSafty and PrivacyStochastic Differential Equation
🎯 What it does: This paper studies the problem of finding second-order stationary points (SOSP) in random non-convex optimization under differential privacy.
Second-order Optimization under Heavy-Tailed Noise: Hessian Clipping and Sample Complexity Limits
Abdurakhmon Sadiev (King Abdullah University of Science and Technology), Ilyas Fatkhullin (ETH Zurich)
OptimizationTabular
🎯 What it does: This paper studies second-order optimization under heavy-tailed noise, proposing an algorithm based on gradient and Hessian matrix clipping, and establishes a lower bound on sample complexity.
Secure and Confidential Certificates of Online Fairness
Olive Franzese (University of Toronto), Hamed Haddadi (Imperial College London)
Safty and PrivacyComputational EfficiencyTabular
🎯 What it does: The OATH protocol is proposed, utilizing zero-knowledge proofs to achieve online group fairness certification, ensuring both model fairness and confidentiality.
Securing the Language of Life: Inheritable Watermarks from DNA Language Models to Proteins
ZAIXI ZHANG, Mengdi Wang (Princeton University)
Safty and PrivacyData-Centric LearningLarge Language ModelBiomedical Data
🎯 What it does: Two watermarking methods, DNAMark and CentralMark, were designed to enhance the security and traceability of gene sequences by tracking the output of DNA language models.
See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction
Wuyuan, Jian Yang (Nanjing University of Science and Technology)
Object DetectionSegmentationAutonomous DrivingConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A 3D occupancy prediction framework for nighttime scenes, LIAR, is proposed, which enhances visual perception performance using lighting information.
See&Trek: Training-Free Spatial Prompting for Multimodal Large Language Model
Pengteng Li, Hui Xiong
RecognitionObject DetectionAutonomous DrivingLarge Language ModelPrompt EngineeringVision Language ModelSimultaneous Localization and MappingImageVideoMultimodalityBenchmark
🎯 What it does: This paper presents SEE&Trek, a training-free and GPU-free spatial prompting framework designed to enhance the spatial understanding capabilities of multimodal large language models under purely visual conditions.
SEEA-R1: Tree-Structured Reinforcement Fine-Tuning for Self-Evolving Embodied Agents
Wanxin Tian (Beijing Innovation Center of Humanoid Robotics), Jian Tang (Beijing Innovation Center of Humanoid Robotics)
Robotic IntelligenceLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
🎯 What it does: Designed and implemented the SEEA-R1 framework, utilizing tree-based MCTS and GRPO for reinforcement fine-tuning, and constructed a multimodal generative reward model to achieve self-evolving embodied agents.
Seeds of Structure: Patch PCA Reveals Universal Compositional Cues in Diffusion Models
Qingsong Wang (Halıcıoğlu Data Science Institute University of California San Diego), Yusu Wang (Halıcıoğlu Data Science Institute University of California San Diego)
RestorationGenerationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This study investigates the relationship between noise and image structure in diffusion models, discovering that low-frequency noise dominates the generated structure, and proposes a denoiser based on patch PCA and a zero-shot noise editing method.
Seeing is Believing? Mitigating OCR Hallucinations in Multimodal Large Language Models
Zhentao he, Minghui Qiu (ByteDance)
OptimizationLarge Language ModelReinforcement LearningImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This study investigates the OCR hallucination problem in visually degraded documents using multimodal large language models, proposing the KIE-HVQA benchmark and a GRPO-based reinforcement learning framework to reduce hallucinations.
Seeing Sound, Hearing Sight: Uncovering Modality Bias and Conflict of AI models in Sound Localization
Yanhao Jia (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)
RecognitionData SynthesisContrastive LearningMultimodalityAudio
🎯 What it does: The system evaluates and improves the performance of the sound localization model under multimodal conflicts.
Seeing the Arrow of Time in Large Multimodal Models
Zihui Xue (University of Texas at Austin), Kristen Grauman (University of Texas at Austin)
Reinforcement LearningVideoMultimodalityBenchmark
🎯 What it does: This study explores how to enable large-scale multimodal models to perceive the Arrow of Time (AoT) in videos and enhance their temporal sensitivity.
Seeing the Wind from a Falling Leaf
Zhiyuan Gao (University of Southern California), Yue Wang (University of Southern California)
RestorationObject TrackingOptimizationRepresentation LearningVision Language ModelVideoPhysics Related
🎯 What it does: This paper proposes a differentiable inverse graphics framework that can recover invisible force fields from videos, such as wind fields.