NeurIPS 2025 Papers with AI Summaries
Conference on Neural Information Processing Systems · 5275 papers
→ NeurIPS 2025 papers with code (2283)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 5275 NeurIPS 2025 papers by keyword, author or institution —
free trial on arXivSub.
\(\varepsilon\)-Optimally Solving Two-Player Zero-Sum POSGs
Erwan escudie, Jilles Steeve Dibangoye (University of Groningen)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabularBenchmark
🎯 What it does: A new framework for ε-optimal solving of two-player zero-sum partially observable stochastic games (zs-POSG) is proposed, which losslessly maps it to transfer-independent zero-sum stochastic games (zs-SG), allowing for the direct application of dynamic programming methods.
🎧MOSPA: Human Motion Generation Driven by Spatial Audio
Shuyang Xu (University of Hong Kong), Taku Komura
GenerationTransformerDiffusion modelMultimodalityAudio
🎯 What it does: This paper proposes a human motion generation task based on spatial audio, constructs the first SAM dataset containing various spatial audio and corresponding 3D human motions, and designs the MOSPA framework utilizing diffusion models for action generation.
$\boldsymbol{\lambda}$-Orthogonality Regularization for Compatible Representation Learning
Simone Ricci (University of Florence), Alberto Del Bimbo (University of Florence)
RetrievalRepresentation LearningContrastive LearningImage
🎯 What it does: Achieve compatible representations between independently trained retrieval models by proposing λ-orthogonal regularization that combines affine/orthogonal transformations with supervised contrastive loss, and provide an efficient partial filling strategy.
$\Delta \mathrm{Energy}$: Optimizing Energy Change During Vision-Language Alignment Improves both OOD Detection and OOD Generalization
Lin Zhu (Shanghai Jiao Tong University), Nanyang Ye (Shanghai Jiao Tong University)
Anomaly DetectionOptimizationTransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: A new OOD detection score ΔEnergy (based on energy change) is proposed, and based on this, an EBM binding maximization loss is designed to simultaneously enhance the generalization ability for closed-set OOD and the detection performance for open-set OOD during VLM fine-tuning.
$\epsilon$-Seg: Sparsely Supervised Semantic Segmentation of Microscopy Data
Sheida RahnamaiKordasiabi, Florian Jug (Human Technopole)
SegmentationAuto EncoderContrastive LearningImageBiomedical Data
🎯 What it does: A sparse supervised semantic segmentation method, ϵ‑Seg, was designed and evaluated. It utilizes a hierarchical variational autoencoder combined with center region mask reconstruction, GMM prior, contrastive learning, and an MLP segmentation head, achieving high-quality segmentation with only a minimal amount of pixel-level annotations.
$\mathcal{X}^2$-DFD: A framework for e$\mathcal{X}$plainable and e$\mathcal{X}$tendable Deepfake Detection
Yize Chen (Chinese University of Hong Kong), Baoyuan Wu (University at Buffalo)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningImageVideoMultimodality
🎯 What it does: A three-stage multimodal large language model (MLLM) deepfake detection framework called X2-DFD has been constructed, which can automatically evaluate, enhance, and supplement the model's understanding of deepfake features and generate interpretable detection results.
$\mu$PC: Scaling Predictive Coding to 100+ Layer Networks
Francesco Innocenti (University of Sussex), Christopher Buckley
ClassificationOptimizationConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: This paper proposes a new parameterization method µ PC, which enables the stable training of predictive coding networks (PCNs) at depths exceeding 100 layers, thereby breaking the scale limitations of traditional PC.
$\Psi$-Sampler: Initial Particle Sampling for SMC-Based Inference-Time Reward Alignment in Score Models
TaeHoon Yoon, Minhyuk Sung (KAIST)
OptimizationReinforcement Learning from Human FeedbackScore-based ModelTabularStochastic Differential Equation
🎯 What it does: Using the pCNL algorithm to sample initial particles in high-dimensional posterior space, and aligning rewards during inference with SMC.
$\text{G}^2\text{M}$: A Generalized Gaussian Mirror Method to Boost Feature Selection Power
Hongyu Shen (University of Illinois), Zhizhen Zhao (University of Illinois)
Biomedical Data
🎯 What it does: A general Gaussian mirror (G2M) method is proposed, improving the variance assumption of traditional data splitting and Gaussian mirror statistics, achieving stronger feature selection efficacy.
$\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation
Weilun Feng (Institute of Computing Technology Chinese Academy of Sciences), Yongjun Xu (Institute of Computing Technology Chinese Academy of Sciences)
GenerationCompressionKnowledge DistillationTransformerDiffusion modelVideo
🎯 What it does: A post-training quantization framework for video diffusion transformers, named S2Q-VDiT, is proposed, achieving low-bit quantization with almost no performance loss through significant data selection and sparse label distillation techniques.
$\textit{HiMaCon:}$ Discovering Hierarchical Manipulation Concepts from Unlabeled Multi-Modal Data
Ruizhe Liu (Hong Kong University), Yanchao Yang (Hong Kong University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerContrastive LearningMultimodalityBenchmark
🎯 What it does: A self-supervised framework HiMaCon is proposed to learn hierarchical mechanical operation concepts from unlabeled multimodal demonstrations and to enhance the generalization ability of the policy.
$\textit{Hyper-GoalNet}$: Goal-Conditioned Manipulation Policy Learning with HyperNetworks
Pei Zhou (University of Hong Kong), Yanchao Yang (University of Hong Kong)
Robotic IntelligenceReinforcement LearningImage
🎯 What it does: The Hyper-GoalNet framework is proposed, which dynamically generates policy network parameters based on target images through a hypernetwork, enabling goal-conditioned manipulation policy learning.
$\texttt{BetaConform}$: Efficient MAP Estimation of LLM Ensemble Judgment Performance with Prior Transfer
Huaizhi Qu, Tianlong Chen
ClassificationOptimizationTransformerLarge Language ModelText
🎯 What it does: The BetaConform framework is proposed for efficiently and accurately estimating the distribution of ensemble judgments from large language models (LLMs), enhancing estimation accuracy through prior transfer in cases of insufficient samples.
$\texttt{G1}$: Teaching LLMs to Reason on Graphs with Reinforcement Learning
Xiaojun Guo (Peking University), Yisen Wang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGraph
🎯 What it does: By applying reinforcement learning to a large number of synthetic graph theory tasks, the ability of large language models in graph structure reasoning is enhanced;
$\texttt{STRCMP}$: Integrating Graph Structural Priors with Language Models for Combinatorial Optimization
Xijun Li (Shanghai Jiao Tong University), Haibing Guan (Shanghai Jiao Tong University)
OptimizationAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: This paper proposes the STRCMP framework, which automatically generates algorithm code that conforms to solver syntax and problem topology by embedding combinatorial optimization structures extracted by graph neural networks into large language models, and iteratively improves it using evolutionary algorithms.
$i$MIND: Insightful Multi-subject Invariant Neural Decoding
Zixiang Yin (Tulane University), Zhengming Ding (Tulane University)
ClassificationRecognitionExplainability and InterpretabilityTransformerAuto EncoderBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed the iMIND framework, utilizing multi-subject fMRI for dual decoding (identity recognition and object recognition).
$O(\sqrt{T})$ Static Regret and Instance Dependent Constraint Violation for Constrained Online Convex Optimization
Rahul Vaze (Tata Institute of Fundamental Research), Abhishek Sinha (Tata Institute of Fundamental Research)
OptimizationTabular
🎯 What it does: A projection algorithm utilizing the geometry of nested convex constraint sets is proposed in constrained online convex optimization, providing algorithms that simultaneously satisfy O(√T) static regret and O(1) or instance-dependent constraint violation.
$Q\sharp$: Provably Optimal Distributional RL for LLM Post-Training
Jin Peng Zhou (Cornell University), Wen Sun (Databricks)
Large Language ModelReinforcement LearningText
🎯 What it does: A KL regularized value function Q♯ method based on distributed RL is proposed to improve the alignment and inference of LLM after training.
1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities
Kevin Wang (Princeton University), Benjamin Eysenbach (Princeton University)
Robotic IntelligenceReinforcement LearningContrastive Learning
🎯 What it does: This paper systematically evaluates the performance improvement in simulated walking, navigation, and manipulation tasks by extending the network depth of the self-supervised contrastive reinforcement learning (CRL) model from the traditional 2-5 layers to 1024 layers.
1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering
Yuheng Yuan (National University of Singapore), Xinchao Wang (National University of Singapore)
CompressionComputational EfficiencyGaussian SplattingVideo
🎯 What it does: Proposes 4DGS-1K, which significantly compresses the number of high-frequency points in 4D Gaussian Splatting through pruning and temporal filtering, achieving real-time rendering at over 1000 FPS.
3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs
Mehdi Makni (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: An efficient post-training method 3BASiL-TM is proposed for the sparse plus low-rank (S + LR) decomposition of large language models, aimed at reducing performance degradation compared to dense models.
3D Equivariant Visuomotor Policy Learning via Spherical Projection
Boce Hu (Northeastern University), Robin Walters (Northeastern University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImage
🎯 What it does: Proposes the Image-to-Sphere Policy (ISP), a SO(3) equivariant closed-loop visual motion control framework that uses only a monocular eye-in-hand RGB camera.
3D Gaussian Flats: Hybrid 2D/3D Photometric Scene Reconstruction
Maria Taktasheva (Simon Fraser University), Andrea Tagliasacchi (Simon Fraser University)
SegmentationData SynthesisOptimizationGaussian SplattingSimultaneous Localization and MappingPoint CloudMesh
🎯 What it does: A hybrid 2D/3D Gaussian plane representation is proposed, which jointly optimizes the in-plane constraints of 2D Gaussians and free-form 3D Gaussians for realistic view synthesis and mesh extraction in indoor scenes.
3D Gaussian Splatting based Scene-independent Relocalization with Unidirectional and Bidirectional Feature Fusion
Junyi Wang (Shandong University), Yue Qi (Beihang University)
Pose EstimationTransformerGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a scene-independent camera relocalization framework based on 3D Gaussian Splatting.
3D Human Pose Estimation with Muscles
Kevin Zhu (University of Waterloo), John McPhee (University of Waterloo)
Pose EstimationTransformerVideo
🎯 What it does: MusclePose is proposed, an end-to-end learnable physics-driven 3D human pose estimation model that can simultaneously predict human kinematics, dynamics, muscle signals, and fine-grained body shape parameters.
3D Interaction Geometric Pre-training for Molecular Relational Learning
Namkyeong Lee (KAIST), Chanyoung Park (Nanjing University)
Drug DiscoveryGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A molecular relationship learning framework (3DMRL) utilizing a virtual interactive geometric environment for 3D pre-training is proposed, allowing a 2D MRL model to acquire 3D interaction information through global contrast and local relative geometric predictions.
3D Visual Illusion Depth Estimation
Chengtang Yao (Beijing Institute of Technology), Yunde Jia (Beijing Institute of Technology)
Depth EstimationRecurrent Neural NetworkTransformerVision Language ModelImage
🎯 What it does: A large-scale 3D visual illusion dataset is proposed, and its impact on depth estimation models is studied; a monocular-stereo fusion framework based on a vision-language model is designed to enhance depth inference in illusion scenarios;
3D-GSRD: 3D Molecular Graph Auto-Encoder with Selective Re-mask Decoding
Chang Wu (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
Knowledge DistillationRepresentation LearningDrug DiscoveryGraph Neural NetworkTransformerAuto EncoderGraph
🎯 What it does: A 3D molecular graph autoencoder, 3D-GSRD, was developed, which combines selective re-masking decoding and structure-agnostic decoding to perform self-supervised pre-training on molecular 3D structures, followed by fine-tuning on downstream tasks using MD17 and QM9.
3D-Prover: Diversity Driven Theorem Proving With Determinantal Point Processes
Sean Lamont (Australian National University), Michael Norrish (Australian National University)
TransformerTabular
🎯 What it does: This paper proposes a diversity-driven proof strategy filter, 3D-Prover, based on Determinantal Point Process (DPP), which can filter out semantically diverse and high-quality proof strategies before the proof search.
3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization
Yuze Hao (Zhejiang University), Yi Yang (Zhejiang University)
Autonomous DrivingOptimizationGraph Neural NetworkDiffusion modelMeshPhysics Related
🎯 What it does: This paper proposes a framework for inverse design directly in the three-dimensional physical-geometry connection space, called 3DID.
3DLLM-Mem: Long-Term Spatial-Temporal Memory for Embodied 3D Large Language Model
Wenbo Hu (University of California), Kai-Wei Chang (University of California)
Robotic IntelligenceTransformerLarge Language ModelPoint CloudBenchmark
🎯 What it does: A long-term spatiotemporal memory benchmark 3DMEM-BENCH has been constructed, and 3DLLM-MEM has been proposed, which enhances the performance of 3D language models in multi-room tasks, question answering, and descriptions by integrating working memory queries with long-term memory.
3DOT: Texture Transfer for 3DGS Objects from a Single Reference Image
Xiao Cao (National University of Singapore), Robby T. Tan (National University of Singapore)
GenerationData SynthesisPrompt EngineeringDiffusion modelGaussian SplattingImage
🎯 What it does: Achieved texture transfer from a single reference image to an existing 3D Gaussian Splatting (3DGS) object.
3DPE-Gaze:Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation
Yangshi Ge (Beihang University), Feng Lu (Beihang University)
RecognitionDomain AdaptationContrastive LearningImage
🎯 What it does: A 3DPE-Gaze framework is proposed, utilizing FLAME's three-dimensional geometric priors and CLIP's semantic alignment to achieve cross-domain gaze estimation.
4D-LRM: Large Space-Time Reconstruction Model From and To Any View at Any Time
Ziqiao Ma (Adobe Research), Hao Tan (Adobe Research)
RestorationGenerationData SynthesisTransformerVideo
🎯 What it does: This paper presents 4D-LRM, a large-scale spatiotemporal reconstruction model capable of generating high-quality renderings of dynamic objects from sparse views at any time and from any perspective.
4D-VLA: Spatiotemporal Vision-Language-Action Pretraining with Cross-Scene Calibration
Jiahui Zhang (Fudan University), Li Zhang (Fudan University)
Computational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelImageVideoMultimodalityBenchmark
🎯 What it does: Proposes a 4D-VLA framework that utilizes RGB-D sequences and 3D coordinate embeddings for cross-scene pre-training, constructing a multi-view evaluation benchmark MV-Bench to achieve a more efficient visual language action model.
4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos
Mengqi Guo (National University of Singapore), Gim Hee Lee (National University of Singapore)
SegmentationGenerationOptimizationTransformerGaussian SplattingVideo
🎯 What it does: A 4D3R framework is proposed, achieving 4D Gaussian splatting for directly reconstructing dynamic scenes from monocular video without pose estimation.
4DGCPro: Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming
Zihan Zheng (Shanghai Jiaotong University), Wenjun Zhang (Shanghai Jiaotong University)
CompressionVideo
🎯 What it does: A single model has been constructed to support multi-bitrate, real-time mobile decoding and rendering in a layer-wise 4D Gaussian compression framework called 4DGCPro, designed for progressive streaming of volumetric video.
4DGT: Learning a 4D Gaussian Transformer Using Real-World Monocular Videos
Zhen Xu (Meta), Zhaoyang Lv (Meta)
GenerationData SynthesisOptimizationTransformerGaussian SplattingVideo
🎯 What it does: This paper proposes 4DGT, a Transformer model that utilizes monocular calibrated video to generate 4D Gaussian fields in real-time for dynamic scene reconstruction.
4KAgent: Agentic Any Image to 4K Super-Resolution
Yushen Zuo (Texas A&M University), Zhengzhong Tu (UC Merced)
RestorationSuper ResolutionAgentic AIMixture of ExpertsVision Language ModelImageBenchmark
🎯 What it does: This paper proposes 4KAgent, a universal agent architecture capable of uniformly upscaling any low-quality images from different domains and with different degradations to 4K (or higher) resolution. It automatically identifies the type of degradation, formulates a recovery plan, executes multi-step recovery, and rolls back if necessary during the upscaling process.
70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11)
Tianyi Zhang (Rice University), Anshumali Shrivastava (Rice University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A lossless compression framework DFloat11 based on entropy coding is proposed, which compresses LLM and DM weights to about 30% of their size while maintaining 100% accuracy and achieving GPU-efficient inference.
A Single-Swap Local Search Algorithm for k-Means of Lines
Ting Liang (Central South University), Qilong Feng (Central South University)
OptimizationTabular
🎯 What it does: This paper studies the k-means clustering problem for lines and proposes a single-swap local search algorithm to achieve approximate clustering.
A Bayesian Approach to Contextual Dynamic Pricing using the Proportional Hazards Model with Discrete Price Data
Dongguen Kim (Pohang University of Science and Technology), Minwoo Chae (Pohang University of Science and Technology)
Tabular
🎯 What it does: A context dynamic pricing algorithm based on the Bayesian Cox proportional hazards model (BayesCoxCP) is proposed, and it is proven that its cumulative regret under a discrete price grid can adapt to the grid sparsity, achieving a theoretically faster regret rate;
A Bayesian Fast-Slow Framework to Mitigate Interference in Non-Stationary Reinforcement Learning
Yihuan Mao (Institute for Interdisciplinary Information Sciences Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
Robotic IntelligenceMeta LearningReinforcement LearningAgentic AISequential
🎯 What it does: The Bayesian Fast-Slow Framework (BFSF) is proposed to address the interference problem in non-stationary MDPs by dynamically selecting a 'fast' strategy based on recent data and a 'slow' strategy obtained through meta-reinforcement learning;
A Beyond-Worst-Case Analysis of Greedy k-means++
Qingyun Chen (University of California Santa Cruz), Ruilong Zhang (Technical University of Munich)
Optimization
🎯 What it does: A beyond-worst-case theoretical analysis of the seed selection process of greedy k-means++ is conducted, proving that it can achieve a better approximation ratio on naturally exponentially distributed, well-separable, and size-balanced datasets;
A Black-Box Debiasing Framework for Conditional Sampling
Han Cui (University of Illinois), Jingbo Liu (University of Illinois)
Tabular
🎯 What it does: A black-box debiasing framework is proposed, which performs high-order debiasing on conditional sampling (posterior) using known likelihood and unknown prior training samples to improve sampling accuracy.
A Cautionary Tale on Integrating Studies with Disparate Outcome Measures for Causal Inference
Harsh Parikh (Yale University), Caleb H Miles
Drug DiscoveryBiomedical Data
🎯 What it does: This paper studies how to safely merge data when using different but related outcome measures in various studies to improve the efficiency of causal effect estimation.
A Circular Argument: Does RoPE need to be Equivariant for Vision?
Chase van de Geijn (University of Göttingen), Alexander S. Ecker (Max Planck Institute for Dynamics and Self-Organization)
TransformerImage
🎯 What it does: This paper explores the necessity of Rotational Position Encoding (RoPE) in visual Transformers, proposing and validating new variants of Spherical RoPE and Uniform RoPE, and comparing the performance of various RoPE and Absolute Position Encoding (APE) in ViT.
A Clean Slate for Offline Reinforcement Learning
Matthew Thomas Jackson, Jakob Nicolaus Foerster
Reinforcement LearningTabular
🎯 What it does: This paper proposes a rigorous classification and evaluation process for offline reinforcement learning (RL) problems, provides clean code in a single file with JAX implementation, and builds a unified offline RL framework called Unifloral. Based on this framework, two new algorithms are designed: TD3-AWR (model-free) and MoBRAC (model-based), achieving significant performance improvements on various offline datasets.
A Closer Look at Graph Transformers: Cross-Aggregation and Beyond
Jiaming Zhuo (Hebei University of Technology), Liang Yang
ClassificationGraph Neural NetworkTransformerGraph
🎯 What it does: A linear graph transformer UGCFormer based on a cross-aggregation mechanism is proposed for node classification tasks.
A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective
Lianghe Shi (University of Michigan), Qing Qu (University of Michigan)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The study investigates the transition of diffusion models from generalization to memorization in recursive training loops, proposing an entropy-based data subset selection method to alleviate model collapse.
A Closer Look at NTK Alignment: Linking Phase Transitions in Deep Image Regression
Giuseppe Castiglione (University of Sussex), Ivor J A Simpson
RestorationSuper ResolutionImageStochastic Differential Equation
🎯 What it does: This paper analyzes and unifies three learning phases of the deep SIREN network in the image regression (super-resolution) training process through the Neural Tangent Kernel (NTK) theory: residual diffusion ripples, loss rate collapse, and the spontaneous alignment of NTK to image edges.
A Closer Look at TabPFN v2: Understanding Its Strengths and Extending Its Capabilities
Han-Jia Ye (Nanjing University), Wei-Lun Chao (Ohio State University)
ClassificationExplainability and InterpretabilityTransformerTabular
🎯 What it does: A thorough analysis of the working mechanism of TabPFN v2 is conducted, and a partitioning and combination strategy for inference without training is proposed to extend its applicability to high-dimensional, multi-class, and large-scale datasets.
A Closer Look to Positive-Unlabeled Learning from Fine-grained Perspectives: An Empirical Study
Yuanchao Dai (Jilin University), Ximing Li (Jilin University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper systematically evaluates the two main families of methods in Positive-Unlabeled Learning (PU Learning), proposes the Set-aware Empirical Risk (SAPU) based on ensemble supervision, and combines it with pseudo-label strategies to form a new GPU framework, completing comprehensive experimental validation.
A CLT for Polynomial GNNs on Community-Based Graphs
Luciano Vinas (University of California Los Angeles), Arash A. Amini (University of California Los Angeles)
Graph Neural NetworkGraph
🎯 What it does: This paper conducts an asymptotic distribution analysis of the embedding of multi-layer polynomial graph neural networks (Poly-GNN) on community structure graphs, proving its central limit theorem and providing the limit Gaussian mixture distribution of the embedding and labels.
A compressive-expressive communication framework for compositional representations
Rafael Elberg (Pontificia Universidad Catolica), Denis Parra (Pontificia Universidad Catolica)
GenerationCompressionRecurrent Neural NetworkAuto EncoderPoint Cloud
🎯 What it does: A self-supervised communication framework CELEBI based on Compression-Expression-Iterative Learning is proposed, utilizing three mechanisms: advanced decoding, final state imitation, and message diversity maximization to promote the compressibility, expressiveness, and efficiency of language.
A Computationally Viable Numerical Gradient-based Technique for Optimal Covering Problems
Gokul Rajaraman (Indian Institute of Technology Bombay), Debasish Chatterjee (Indian Institute of Technology Bombay)
OptimizationStochastic Differential Equation
🎯 What it does: A numerical algorithm based on zeroth-order gradients is proposed to solve the optimal covering (Chebyshev n-net) problem, and its feasibility is proven.
A Counterfactual Semantics for Hybrid Dynamical Systems
Andy Zane (Basis Research Institute), Sam Witty (Sorbus AI)
Time SeriesPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A counterfactual semantics for hybrid systems is proposed, which can perform causal inference on dynamically triggered instantaneous interventions.
A Cramér–von Mises Approach to Incentivizing Truthful Data Sharing
Alex Clinton (University of Wisconsin Madison), Kirthevasan Kandasamy (University of Wisconsin Madison)
Data SynthesisFederated LearningSafty and PrivacyReinforcement LearningImageText
🎯 What it does: Design a reward mechanism based on the Cramér-von Mises (CvM) two-sample test to incentivize participants to submit data truthfully and encourage the submission of more data.
A data and task-constrained mechanistic model of the mouse outer retina shows robustness to contrast variations
Kyra L. Kadhim (University of Tübingen), Philipp Berens (University of Tübingen)
ClassificationOptical FlowImage
🎯 What it does: A fully differentiable biophysical model containing 200 cones, horizontal cells, and bipolar cells was constructed and trained to classify MNIST images with varying contrast and brightness.
A Data-Driven Prism: Multi-View Source Separation with Diffusion Model Priors
Sebastian Wagner-Carena (Flatiron Institute), Sydney Erickson (Stanford University)
Diffusion modelScore-based ModelImage
🎯 What it does: This paper proposes a multi-view source separation framework (DDPRISM) that utilizes diffusion models as priors, capable of learning and inferring the prior and posterior distributions of independent sources from multi-view observations based solely on a known linear mixing matrix and noise covariance.
A Difference-of-Convex Functions Approach to Energy-Based Iterative Reasoning
Daniel Tschernutter (Infermedica), Maciej Kasiński (Infermedica)
OptimizationComputational EfficiencyText
🎯 What it does: This paper proposes an energy model inference algorithm called DCAReasoner based on Differential Convex (DC) functions, aimed at efficiently solving energy minimization problems in continuous inference tasks.
A Differential and Pointwise Control Approach to Reinforcement Learning
Minh Phuong Nguyen (University of Texas at Austin), Chandrajit L. Bajaj (University of Texas at Austin)
OptimizationReinforcement LearningPhysics Related
🎯 What it does: A Differential Reinforcement Learning (Differential RL) framework is proposed, along with the corresponding Differential Policy Optimization (dfPO) algorithm, for reinforcement learning in continuous state-action spaces. Experiments are conducted on three types of scientific computing tasks: surface modeling, multiscale grid control, and molecular dynamics.
A Diffusion Model for Regular Time Series Generation from Irregular Data with Completion and Masking
Gal Fadlon (Ben-Gurion University of Negev), Omri Azencot (Ben-Gurion University of Negev)
GenerationData SynthesisTransformerDiffusion modelTime SeriesSequentialFinance Related
🎯 What it does: A two-step framework is proposed, first using a time series Transformer to complete irregularly sampled sequences, and then using a visual diffusion model (ImagenTime) combined with masking to generate regular time series.
A Driving-Style-Adaptive Framework for Vehicle Trajectory Prediction
Di Wen (Sun Yat-sen University), Zheng Qingfang (Pengcheng Laboratory)
Autonomous DrivingOptimizationExplainability and InterpretabilityMixture of ExpertsTime Series
🎯 What it does: A driving style adaptive vehicle trajectory prediction framework is proposed, utilizing polynomial basis functions that match different driving styles for trajectory modeling and prediction.
A Dynamic Learning Strategy for Dempster-Shafer Theory with Applications in Classification and Enhancement
Linlin Fan (Chongqing University), Weijia Jia (Beijing Normal University)
ClassificationRestorationConvolutional Neural NetworkImage
🎯 What it does: This study investigates a dynamic learning framework based on Dempster-Shafer evidence theory, utilizing non-uniform partitioning and Hilbert space mapping to integrate evidence, and applies it to pattern classification and low-light image enhancement.
A Fair Federated Learning Method for Handling Client Participation Probability Inconsistencies in Heterogeneous Environments
Siyuan Wu (Nanjing University), Wanchun Dou (Nanjing University)
Federated LearningKnowledge DistillationImage
🎯 What it does: A heterogeneous federated learning method (PHP-FL) is proposed to handle inconsistencies in client participation probabilities.
A faster training algorithm for regression trees with linear leaves, and an analysis of its complexity
Kuat Gazizov (University of California), Miguel Á. Carreira-Perpiñán (University of California)
OptimizationComputational EfficiencyTabularComputed Tomography
🎯 What it does: The improved TAO algorithm is studied, utilizing the Sherman–Morrison–Woodbury formula to accelerate the training of linear leaf nodes in regression trees, significantly reducing training time.
A Few Moments Please: Scalable Graphon Learning via Moment Matching
Reza Ramezanpour (Rice University), Santiago Segarra (Rice University)
ClassificationData SynthesisGraph Neural NetworkGenerative Adversarial NetworkGraph
🎯 What it does: MomentNet is proposed to directly estimate graph structures through subgraph counting (moments) matching and implicit neural representations, and based on this, MomentMixup is introduced for graph data augmentation.
A Finite Sample Analysis of Distributional TD Learning with Linear Function Approximation
Yang Peng (Peking University), Zhihua Zhang (Peking University)
Reinforcement LearningSequential
🎯 What it does: This paper studies the finite sample statistical rate of distributed temporal difference (TD) learning with linear function approximation, aiming to estimate the return distribution of a discounted Markov decision process under a given policy.
A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1
Zhaoyi Li (Mohammed Bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohammed Bin Zayed University of Artificial Intelligence)
Adversarial AttackVision Language ModelImage
🎯 What it does: A strong attack framework M-Attack based on local cropping and model ensemble is designed to target closed-source large visual language models.
A General-Purpose Theorem for High-Probability Bounds of Stochastic Approximation with Polyak Averaging
Sajad Khodadadian (Virginia Polytechnic Institute and State University), Martin Zubeldia (University of Minnesota)
Reinforcement Learning
🎯 What it does: A general framework is proposed to obtain high-probability convergence posteriors for given non-averaging stochastic approximation (SA) iterations, leading to non-asymptotic high-probability bounds for Polyak-Ruppert averaged iterations.
A Generalist Intracortical Motor Decoder
Joel Ye (Carnegie Mellon University), Robert Gaunt (University of Pittsburgh)
TransformerSupervised Fine-TuningMultimodalityTime Series
🎯 What it does: A large-scale Transformer model (NDT3) was trained using 2000 hours of cortical potential and motor behavior data from monkeys and humans for pre-training, and fine-tuned on various downstream motor decoding tasks;
A Generalized Binary Tree Mechanism for Private Approximation of All-Pair Shortest Distances
Zongrui Zou (Nanjing University), Jalaj Upadhyay (Rutgers University)
OptimizationSafty and PrivacyGraph Neural NetworkGaussian SplattingGraph
🎯 What it does: Under the framework of differential privacy, a recursive partitioning binary tree mechanism is proposed, which can efficiently approximate the all-pairs shortest distance of weighted undirected graphs.
A Generalized Bisimulation Metric of State Similarity between Markov Decision Processes: From Theoretical Propositions to Applications
Zhenyu Tao (Southeast University), Xiaohu You (Southeast University)
Reinforcement LearningTabular
🎯 What it does: A general bisimulation metric (GBSM) for different Markov Decision Processes (MDPs) is proposed, and its metric properties such as symmetry, triangle inequality, and distance bounds under the same state space are theoretically proven.
A Generalized Label Shift Perspective for Cross-Domain Gaze Estimation
Hao-Ran Yang (Sun Yat-Sen University), Chuan-Xian Ren (Sun Yat-Sen University)
RecognitionDomain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates the label and conditional shift issues in cross-domain gaze estimation and proposes a correction framework based on the GLS perspective.
A Geometric Analysis of PCA
Ayoub El Hanchi (University of Toronto and Vector Institute), Chris J. Maddison (University of Toronto and Vector Institute)
🎯 What it does: This paper studies the geometric properties of Principal Component Analysis (PCA), providing the central limit theorem for its errors and the asymptotic excess risk distribution, along with a non-asymptotic upper bound.
A geometric framework for momentum-based optimizers for low-rank training
Steffen Schotthöfer (Oak Ridge National Laboratory), Jonas Kusch (Norwegian University of Life Sciences)
OptimizationTransformerLarge Language ModelImageText
🎯 What it does: This study investigates the geometric framework of momentum optimizers in low-rank training and proposes a momentum method that combines dynamic low-rank approximation.
A Geometry-Aware Metric for Mode Collapse in Time Series Generative Models
Yassine ABBAHADDOU, Amine M. Aboussalah
GenerationData SynthesisExplainability and InterpretabilityComputational EfficiencyDiffusion modelTime Series
🎯 What it does: A geometric metric DMD-GEN based on Dynamic Mode Decomposition (DMD) and optimal transport is proposed to quantify mode collapse in time series generative models and provide interpretive analysis.
A Gradient Guidance Perspective on Stepwise Preference Optimization for Diffusion Models
Joshua Tian Jin Tee (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationOptimizationDiffusion modelImage
🎯 What it does: This paper re-theorizes Stepwise Preference Optimization (SPO) from the perspective of gradient guidance and proposes GradSPO to more efficiently achieve human preference alignment in diffusion models.
A Gradient Guided Diffusion Framework for Chance Constrained Programming
Boyang Zhang (University of Chinese Academy of Sciences), Ya-Feng Liu (Beijing University of Posts and Telecommunications)
OptimizationDiffusion modelScore-based ModelTabularStochastic Differential Equation
🎯 What it does: A framework called GGDOpt based on gradient-guided diffusion models is proposed to solve opportunity-constrained programming problems with unknown distributions.
A Hierarchy of Graphical Models for Counterfactual Inferences
Hongshuo Yang (Columbia University), Elias Bareinboim (Columbia University)
Graph Neural NetworkGraph
🎯 What it does: Two new graphical models, CBN2.25 and CBN2.5, are proposed to describe causal distributions that can be realized in the experimentally feasible 'log layers', along with their identifiability inference rules and completeness proofs.
A High-Dimensional Statistical Method for Optimizing Transfer Quantities in Multi-Source Transfer Learning
Qingyue Zhang (Tsinghua University), Shao-Lun Huang (Tsinghua University)
Domain AdaptationOptimizationTransformerImage
🎯 What it does: This paper addresses the scenario of multi-source transfer learning, investigating and solving how many samples need to be sampled from each source task to achieve optimal generalization performance on the target task. Based on a theoretical framework, the OTQMS algorithm is proposed to implement adaptive sample size selection.
A Implies B: Circuit Analysis in LLMs for Propositional Logical Reasoning
Guan Zhe Hong (Purdue University), Rina Panigrahy (Google Research)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the internal reasoning circuits of large pre-trained language models in propositional logic reasoning tasks.
A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders
David Chanin (University College London), Joseph Isaac Bloom
Explainability and InterpretabilityRepresentation LearningLarge Language ModelAuto EncoderText
🎯 What it does: This study investigates and quantifies the phenomenon of 'feature absorption' in Sparse Autoencoders (SAE), exploring the mechanisms caused by hierarchical features and sparsity loss, and proposes a metric for detecting this phenomenon.
A Latent Multilayer Graphical Model For Complex, Interdependent Systems
Martin Ondrus (University of Alberta), Yang Feng (New York University)
Graph Neural NetworkMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A multi-layer sparse + low-rank inverse covariance estimation method called multiSLICE is proposed for estimating inter-layer edges in multimodal networks with different numbers of nodes and sample sizes across layers.
A learnability analysis on neuro-symbolic learning
Hao-Yuan He (Nanjing University), Ming Li (Nanjing University)
Autonomous DrivingImage
🎯 What it does: This paper studies the learnability of neural symbolic learning (NeSy) tasks, proposing to map NeSy tasks to Derived Constraint Satisfaction Problems (DCSP), and provides theoretical upper bounds on sample complexity and concept error.
A Learning-Augmented Approach to Online Allocation Problems
Ilan Reuven Cohen (Bar-Ilan University), Debmalya Panigrahi (Duke University)
Optimization
🎯 What it does: A general learning-enhanced online allocation framework is proposed, utilizing a single d-dimensional weight vector to achieve near-optimal allocation schemes in various online allocation problems (such as routing, scheduling, minimizing maximum congestion, maximizing Nash social welfare, etc.).
A Learning-Augmented Dynamic Programming Approach for Orienteering Problem with Time Windows
Guansheng Peng (Great Bay University), Pieter Vansteenwegen (KU Leuven)
OptimizationGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: A learning-enhanced dynamic programming algorithm DP-NG-ML is proposed, which uses deep learning to predict ng-sets to accelerate the solution of OPTW.
A Little Depth Goes a Long Way: The Expressive Power of Log-Depth Transformers
William Merrill (Allen Institute for AI), Ashish Sabharwal (Allen Institute for AI)
TransformerGraphSequential
🎯 What it does: This paper studies the expressive power of Transformers when the depth grows logarithmically with the input length (log-depth), and proves that even a fully unified Transformer (parameter sharing) requires only O(log n) layers to recognize regular languages and solve graph connectivity problems.
A machine learning approach that beats Rubik's cubes
Alexander Chervov (Institut Curie), Alexey M. Romanov (RTU MIREA)
OptimizationRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: This paper studies a machine learning-based multi-agent path planning method to find the shortest or approximately shortest paths on large Cayley graphs, particularly for Rubik's Cubes of dimensions 4x4x4 and 5x5x5.
A Minimalist Example of Edge-of-Stability and Progressive Sharpening
Liming Liu (Georgia Institute of Technology), Tuo Zhao (Georgia Institute of Technology)
OptimizationTabular
🎯 What it does: This study investigates the phenomena of boundary stability (EoS) and progressive sharpening (PS) that occur in gradient descent under high learning rates. It proposes a two-layer linear network model with a width of one and provides non-asymptotic theory, proving the existence of PS and self-stabilization, global convergence, and the sharpness of the entire GD trajectory in relation to loss behavior.
A Minimalistic Unified Framework for Incremental Learning across Image Restoration Tasks
Xiaoxuan Gong (Huazhong University of Science and Technology), Jie Ma (Huazhong University of Science and Technology)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A task incremental learning image restoration framework called MINI is proposed, which achieves task-level parameter isolation through the MINIconv module embedded in a pooling layer.
A Multimodal BiMamba Network with Test-Time Adaptation for Emotion Recognition Based on Physiological Signals
Ziyu Jia (Institute of Automation, Chinese Academy of Sciences), Chenyu Liu (Nanyang Technological University)
RecognitionDomain AdaptationRecurrent Neural NetworkSupervised Fine-TuningMultimodalityBiomedical Data
🎯 What it does: This paper proposes a multi-modal BiMamba network combined with Test-Time Adaptation (TTA) for emotion recognition based on physiological signals.
A multiscale analysis of mean-field transformers in the moderate interaction regime
Giuseppe Bruno (University of Bern), Andrea Agazzi (University of Bern)
TransformerOrdinary Differential Equation
🎯 What it does: This paper analyzes the dynamics of token representations during the inference phase of large-context transformers at moderate interaction strengths through mean-field particle systems, revealing three different temporal phase scales: alignment, thermal diffusion, and pairing.
A Near-Optimal Algorithm for Decentralized Convex-Concave Finite-Sum Minimax Optimization
Hongxu Chen (Fudan University), Luo Luo (Fudan University)
OptimizationTabular
🎯 What it does: This study investigates decentralized convex-concave finite and optimization problems and proposes the DIVERSE algorithm.
A Near-optimal, Scalable and Parallelizable Framework for Stochastic Bandits Robust to Adversarial Corruptions and Beyond
Zicheng Hu (East China Normal University), Cheng Chen (East China Normal University)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: The BARBAT framework is proposed and implemented to address the stochastic multi-armed bandit problem in the presence of adversarial perturbations, eliminating the non-optimal upper bound caused by the number of arms K in the original algorithm, and extending it to various scenarios such as multi-agent, batch processing, strongly observable graphs, and d-set semi-games.
A Novel General Framework for Sharp Lower Bounds in Succinct Stochastic Bandits
Guo Zeng (University of Melbourne), Jean Honorio (University of Melbourne)
🎯 What it does: A general framework is proposed to derive lower bounds for the stochastic Bandit problem under a simplified structure, and three specific lower bound results are provided (sparse vectors, group sparse matrices, low-rank matrices).
A Partition Cover Approach to Tokenization
Jia Peng Lim (Singapore Management University), Hady W. Lauw (Singapore Management University)
CompressionOptimizationTextBiomedical Data
🎯 What it does: This paper proposes a new partition cover form of tokenization optimization model and presents a greedy approximation algorithm called GREEDTOK.
A Physics-preserved Transfer Learning Method for Differential Equations
Hao-Ran Yang (Sun Yat-Sen University), Chuan-Xian Ren (Sun Yat-Sen University)
Domain AdaptationOptimizationTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: An optimal tensor transport (POTT) transfer learning method based on physical preservation is proposed to address the transfer problem of data-driven models for partial differential equations under domain shift.
A Plug-and-Play Query Synthesis Active Learning Framework for Neural PDE Solvers
Zhiyuan Wang (Texas A&M University), Xiaoning Qian (Texas A&M University)
OptimizationConvolutional Neural NetworkReinforcement LearningSequentialPhysics Related
🎯 What it does: The PaPQS framework is proposed, which synthesizes information-rich initial conditions in a continuous design space through gradient ascent, enabling active learning for neural PDE solvers.