图像生成 - arXivDaily 专题

2603.22304 2026-03-25 cs.LG 50%

Mitigating Premature Discretization with Progressive Quantization for Robust Vector Tokenization

通过渐进量化缓解早期内离散化以实现鲁棒的向量标记化

Wenhao Zhao, Qiran Zou, Zhouhan Lin, Dianbo Liu

机构 * National University of Singapore, Singapore（新加坡国立大学）； Shanghai Jiao Tong University, Shanghai, China（上海交通大学）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出渐进量化方法，通过引入量化难度动态作为VQ训练的关键维度，有效引导代码本向扩展良好的流形发展，提升了多模态模型的生成性能和生物序列建模能力。

2603.21782 2026-03-24 cs.LG 50%

Show Me What You Don't Know: Efficient Sampling from Invariant Sets for Model Validation

展示你所不知道的：用于模型验证的高效不变集采样

Armand Rousselot, Joran Wendebourg, Ullrich Köthe

机构 * CVL（计算机视觉与学习实验室）； Heidelberg University（海德堡大学）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出一种无需训练的采样方法，通过预训练扩散模型揭示特征提取器的不变性，展示模型在不同数据变化下的表现。

Comments 19 pages, 19 figures

2506.20223 2026-03-24 physics.optics 50%

Efficient first-principles inverse design of nanolasers

纳米激光器的高效原理性反向设计

Beñat Martinez de Aguirre Jokisch, Alexander Cerjan, Rasmus Ellebæk Christiansen, Jesper Mørk, Ole Sigmund, Steven G. Johnson

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出基于非线性Maxwell-Bloch方程和稳态ab initio激光理论的反向设计方法，用于优化纳米结构激光器，考虑空间孔燃烧修正、阈值效应、出射效率和增益扩散。

Journal ref Laser Photonics Reviews e01614 (2025)

2504.14145 2026-03-24 cs.DC cs.AI 50%

DIP: Efficient Large Multimodal Model Training with Dynamic Interleaved Pipeline

DIP: 一种高效的大规模多模态模型训练方法，采用动态交错流水线

Zhenliang Xue, Hanpeng Hu, Xing Chen, Yimin Jiang, Yixin Song, Zeyu Mi, Yibo Zhu, Daxin Jiang, Yubin Xia, Haibo Chen

机构 * Institute of Parallel and Distributed Systems（并行与分布式系统研究所）； Shanghai Jiao Tong University（上海交通大学）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出DIP框架，通过分离不同模态计算和动态分割输入数据，提升大规模多模态模型训练效率，实验显示吞吐量提升达97.3%。

Comments To be published in ASPLOS'26

2603.20859 2026-03-24 math.NA cs.NA 50%

Efficient Nehari manifold optimization algorithms for computing ground state solutions of nonlinear elliptic systems

高效Nehari流形优化算法用于计算非线性椭圆系统的基态解

Zhaoxing Chen, Wei Liu, Ziqing Xie, Wenfan Yi

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出了一种高效的Nehari流形优化算法，用于计算半线性椭圆系统的基态解，通过结合NMOM与Nesterov加速方法，改进了稳定性与计算效率。

Comments 26 pages, 17 figures

2509.11571 2026-03-24 eess.SP 50%

A Fine-Grained 3D Radio Map Construction Paradigm with Ultra-Low Sampling Rates by Large Generative Models

一种基于大生成模型的细粒度3D无线电图构建范式，通过超低采样率

Zhiyuan Liu, Qingyu Liu, Shuhang Zhang, Hongliang Zhang, Lingyang Song

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出RadioLAM范式，利用大生成模型解决超稀疏采样问题，通过增强、生成和选举三个模块高效构建细粒度3D无线电图，实验证明其在0.1%采样率下性能优于现有方法。

Comments Accepted by IEEE JSAC

2603.19647 2026-03-23 math.NA cs.NA 50%

On-the-Fly ROM-Based Acceleration of SI-DSA for Implicit Time Marching of the Radiative Transfer Equation

实时ROM基于SI-DSA加速隐式时间推进的辐射传输方程

Ningxin Liu, Zhichao Peng

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出基于ROM的实时加速方法，用于隐式时间推进的辐射传输方程SI-DSA，利用低秩结构提升预处理效果，实现1.4至2倍的加速效果。

2509.22592 2026-03-20 cs.LG 50%

OT-MeanFlow3D: Bridging Optimal Transport and Meanflow for Efficient 3D Point Cloud Generation

OT-MeanFlow3D: 通过最优传输与Meanflow弥合，实现高效的3D点云生成

Elaheh Akbari, Shansita Sharma, Ping He, Ahmadreza Moradipari, Kyungtae Han, Hamed Pirsiavash, Yikun Bai, Soheil Kolouri

机构 * Department of Computer Science, Vanderbilt University（范德比尔特大学计算机科学系）； Toyota InfoTech Labs（丰田信息技术实验室）； Department of Computer Science, University of California, Davis（加州大学戴维斯分校计算机科学系）； Department of Mathematics, Purdue University（普渡大学数学系）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出OT-MeanFlow3D框架，通过整合最优传输采样，提升3D点云生成与修复的效率与准确性，实验表明其在ShapeNet上优于现有基线模型。

2603.17622 2026-03-19 eess.SP 50%

Fast Beam-Brainstorm: Few-Step Generative Site-Specific Beamforming with Flexible Probing

快速波束-脑暴：基于灵活探测的少步生成式波束成形

Zihao Zhou, Zhaolin Wang, Yuanwei Liu

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出快速波束-脑暴方法，通过两阶段蒸馏策略实现快速生成高保真通信波束，减少生成成本并适应灵活的信道探测长度。

2603.17052 2026-03-19 cs.LG cs.AI 50%

Early Quantization Shrinks Codebook: A Simple Fix for Diversity-Preserving Tokenization

早量化缩小代码表：一种维持多样性token化问题的简单修复

Wenhao Zhao, Qiran Zou, Rushi Shah, Yudi Wu, Zhouhan Lin, Dianbo Liu

机构 * National University of Singapore（国立新加坡大学）； Zhejiang University（浙江大学）； Shanghai Jiao Tong University（上海交通大学）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文研究了向量量化中的坍塌问题，发现随机初始化和编码器容量限制导致token和嵌入坍塌，并提出缓解方案，是首次系统探讨该问题。

2603.14895 2026-03-18 cs.SI 50%

FS_GPlib: Breaking the Web-Scale Barrier - A Unified Acceleration Framework for Graph Propagation Models

FS_GPlib：突破网络级障碍——图传播模型的统一加速框架

Chang Guo, Juyuan Zhang, Chang Su, Tianlong Fan, Linyuan Lü

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出FS_GPlib，通过微粒同步消息传递与宏观批量蒙特卡洛模拟结合，实现大规模图传播建模的高效高保真模拟，实验显示其在标准库上达到3.5万倍加速，可在11秒内完成亿级边图的完整蒙特卡洛模拟。

Comments Accepted at ACM Web Conference 2026 (WWW2026)

2511.15190 2026-03-18 cs.LG cs.AI 50%

Masked Auto-Regressive Variational Acceleration: Fast Inference Makes Practical Reinforcement Learning

遮蔽自回归变分加速：快速推理使强化学习实用

Yuxuan Gu, Weimin Bai, Yifei Wang, Weijian Luo, He Sun

机构 * Peking University（北京大学）； Rice University（Rice大学）； hi-lab, Xiaohongshu Inc（小红书实验室，小红书公司）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出MARVAL框架，通过压缩扩散链为单步自回归生成，提升推理效率并使强化学习实用化，实现了生成模型的快速采样和偏好对齐。

2603.15367 2026-03-17 cond-mat.dis-nn cs.LG 50%

Controlled Langevin Dynamics for Sampling of Feedforward Neural Networks Trained with Minibatches

基于受控 Langevin 动力学的 feedforward 神经网络参数空间采样

Alessandro Zambon, Francesca Caruso, Riccardo Zecchina, Guido Tiana

机构 * Department of Physics, Università degli Studi di Milano（米兰大学物理系）； Department of Computing Sciences（计算科学系）； Bocconi Institute for Data Science（博科尼数据科学研究所）； Analytics (BIDSA), Bocconi University（分析（BIDSA），博科尼大学）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出受控 minibatch Langevin 动力学方法，用于高效采样训练有大数据集的 feedforward 神经网络参数空间，通过调整虚拟质量与摩擦系数以确保高效采样，验证了其在大规模网络中的高效性与泛化性能。

2510.04826 2026-03-17 math.NA cs.NA 50%

Efficient structure-preserving scheme for chemotaxis PDE system with singular sensitivity in crime and epidemic modeling

高效保持结构的方案用于具有在犯罪和流行病建模中奇异敏感性的趋化PDE系统

Rui Wang, Yunfeng Xiong, Zhengru Zhang

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出一种高效保持正性的隐式-显式方案，用于解决具有奇异敏感性的趋化PDE系统中的数值挑战，通过Lagrange乘子修正实现二阶精度，并扩展到退化扩散的多代理流行病模型，验证了正性和质量守恒。

2408.14928 2026-03-17 cond-mat.mtrl-sci cs.LG 50%

Targeting the partition function of chemically disordered materials with a generative approach based on inverse variational autoencoders

通过基于逆变分自编码器的生成方法针对化学无序材料的配分函数

Maciej J. Karcz, Luca Messina, Eiji Kawasaki, Emeric Bourasseau

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出利用生成机器学习生成配置集，以高效计算化学无序材料的原子尺度性质，通过逆变分自编码器实现无监督主动学习，用于估计配分函数。

Journal ref Scientific Reports, 2026

2603.12689 2026-03-16 cond-mat.soft cond-mat.stat-mech 50%

Memory-aware acceleration of orientational dynamics in nanoparticle suspensions

具有记忆性的纳米颗粒悬浮液取向动力学加速

Miguel Ibáñez, Raúl A. Rica-Alarcón, María L. Jiménez

专题命中效率与蒸馏：diffusion(abstract)

AI总结研究通过实验和理论分析，揭示了纳米颗粒取向动力学中的记忆效应，并提出抑制慢速松弛模式的协议，从而显著减少弛豫时间。

2603.12480 2026-03-16 cs.RO cs.AI 50%

One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies

一步流策略：用于快速视觉-运动策略的自蒸馏

Shaolong Li, Lichao Sun, Yongchao Chen

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出一步流策略，通过自蒸馏实现高保真单步动作生成，无需预训练教师模型，有效提升动作生成速度和控制精度。

2505.18017 2026-03-13 cs.LG 50%

Strictly Constrained Generative Modeling via Split Augmented Langevin Sampling

通过分裂增广 Langevin 采样实现严格约束生成建模

Matthieu Blanke, Yongquan Qu, Sara Shamekh, Pierre Gentine

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出 CASAL 算法，通过变量分裂逐步强制约束，用于严格约束生成建模，提升物理系统预测精度和守恒量保持。

Journal ref The Fourteenth International Conference on Learning Representations (ICLR 2026)

2603.10760 2026-03-12 physics.space-ph 50%

Auroral Acceleration Generates Electron Beams in Jupiter's Middle Magnetosphere

极光加速在木星中间磁层中产生电子束

June Piasecki, Joachim Saur, George Clark, Barry H. Mauk, Annika Salveter, Jamey Szalay

专题命中效率与蒸馏：diffusion(abstract)

AI总结朱诺任务发现木星中间磁层中的电子束源于极光加速过程。

2603.10730 2026-03-12 math.NA cs.NA 50%

Efficient design of continuation methods for hyperbolic transport problems in porous media

高效设计用于多孔介质中双曲输运问题的持续法

Peter von Schultzendorff, Jakub Wiktor Both, Jan Martin Nordbotten, Tor Harald Sandve

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出了一种基于熵解的同伦持续方法，用于高效求解多孔介质中复杂多相流问题。

2603.10701 2026-03-12 cs.SD cs.AI 50%

AlphaFlowTSE: One-Step Generative Target Speaker Extraction via Conditional AlphaFlow

AlphaFlowTSE：通过条件AlphaFlow实现的一步生成目标说话人提取

Duojia Li, Shuhan Zhang, Zihan Qian, Wenxuan Wu, Shuai Wang, Qingyang Hong, Lin Li, Haizhou Li

机构 * School of Electronic Science and Engineering, Xiamen University, Xiamen, China（1 厦门大学电子科学与技术学院，厦门，中国）； School of Informatics, Xiamen University, Xiamen, China（2 厦门大学信息学院，厦门，中国）； School of Intelligence Science and Technology, Nanjing University, Suzhou, China（3 南京大学智能科学与技术学院，苏州，中国）； Shenzhen Loop Area Institute, Shenzhen, China（4 深圳循环区研究所，深圳，中国）； School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen, Shenzhen, China（5 香港中文大学人工智能学院，深圳，中国）； The Chinese University of Hong Kong, Shatin, N.T., Hong Kong SAR, China（6 香港中文大学，沙田，香港特别行政区，中国）

专题命中效率与蒸馏：diffusion(abstract)

AI总结 AlphaFlowTSE通过条件AlphaFlow实现一步生成目标说话人提取，提升目标语音保真度和真实混合泛化能力。

Comments Submitted to Interspeech 2026 for review

2603.09839 2026-03-11 astro-ph.SR 50%

Multi-spacecraft constraints on relativistic solar energetic particle transport in the widespread 28 October 2021 event

多航天器约束在2021年10月28日广泛事件中相对论太阳高能粒子传输

E. Lavasa, J. T. Lang, A. Papaioannou, R. D. Strauss, S. A. Mallios, A. Hillaris, A. Kouloumvakos, A. Anastasiadis, I. A. Daglis

专题命中效率与蒸馏：diffusion(abstract)

AI总结研究利用多航天器观测和数值模拟，分析2021年10月28日事件中高能粒子的传输机制，揭示注入区域狭窄及跨场扩散的重要性。

Journal ref A&A, 707 (2026) A12

2603.09597 2026-03-11 cs.NE cs.SC 50%

Symbolic Discovery of Stochastic Differential Equations with Genetic Programming

利用遗传编程发现随机微分方程的符号方法

Sigur de Vries, Sander W. Keemink, Marcel A. J. van Gerven

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出基于遗传编程的随机微分方程符号发现方法，通过最大似然估计联合优化漂移和扩散函数，实现对控制方程的准确恢复和高维系统的高效扩展。

Comments 18 pages, 5 figures, 4 tables

2603.09574 2026-03-11 cs.RO cs.LG 50%

SCDP: Learning Humanoid Locomotion from Partial Observations via Mixed-Observation Distillation

SCDP：通过混合观测蒸馏从部分观测中学习人形机器人的运动

Milo Carroll, Tianhu Peng, Lingfan Bao, Chengxu Zhou, Zhibin Li

专题命中效率与蒸馏：diffusion(abstract)

AI总结 SCDP通过混合观测蒸馏从部分观测中学习人形机器人运动，利用机载传感器实现高效可靠的真实机器人运动控制。

Comments 6 pages, 8 figures, 5 tables, iRos

2603.09415 2026-03-11 cs.RO cs.AI 50%

From Flow to One Step: Real-Time Multi-Modal Trajectory Policies via Implicit Maximum Likelihood Estimation-based Distribution Distillation

从流到一步：通过隐式最大似然估计基于的分布蒸馏实现实时多模轨迹策略

Ju Dong, Liding Zhang, Lei Zhang, Yu Fu, Kaixin Bai, Zoltan-Csaba Marton, Zhenshan Bing, Zhaopeng Chen, Alois Christian Knoll, Jianwei Zhang

机构 * TAMS (Technical Aspects of Multimodal Systems), Department of Informatics, University of Hamburg（汉堡大学信息学院TAMS（多模态系统技术方面））； Technical University of Munich（慕尼黑技术大学）； Agile Robots SE（敏捷机器人公司）

专题命中效率与蒸馏：diffusion(abstract)

AI总结通过隐式最大似然估计基于的分布蒸馏，实现实时多模轨迹策略，提升机器人操作的高频闭环控制与鲁棒性。

Comments https://sites.google.com/view/flow2one, 8 pages

2603.06517 2026-03-09 astro-ph.SR 50%

Constraining turbulent solar flare acceleration regions by connecting kinetic modeling and X-ray observations

通过连接动能建模和X射线观测来约束太阳耀斑加速区域

Morgan Stores, Natasha Jeffrey, Ewan Dickson, James McLaughlin, Eduard Kontar

专题命中效率与蒸馏：diffusion(abstract)

AI总结通过结合动能建模和X射线观测，研究太阳耀斑中湍流对电子加速的影响，确定加速区域和时间尺度。

2603.06287 2026-03-09 cs.CE cs.AI cs.LG math.AP 50%

Learning Where the Physics Is: Probabilistic Adaptive Sampling for Stiff PDEs

学习物理所在的位置：用于刚性PDEs的概率自适应采样

Akshay Govind Srinivasan, Balaji Srinivasan

机构 * Indian Institute of Technology, Madras（印度理工学院马德拉斯学院）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出GMM-PIELM方法，通过概率模型自适应采样，有效解决刚性PDEs中的刚性问题，实现高精度与高速度的平衡。

Comments Accepted at AI&PDE Workshop at the Fourteenth International Conference on Learning Representations

2603.05800 2026-03-09 cs.DC cs.AI 50%

StreamWise: Serving Multi-Modal Generation in Real-Time at Scale

StreamWise: 实时大规模多模态生成服务

Haoran Qiu, Gohar Irfan Chaudhry, Chaojie Zhang, Íñigo Goiri, Esha Choukse, Rodrigo Fonseca, Ricardo Bianchini

机构 * Microsoft Azure Research（微软Azure研究院）； MIT CSAIL（麻省理工学院计算机科学与人工智能实验室）

专题命中效率与蒸馏：image generation(abstract)

AI总结 StreamWise通过实时播客视频生成，整合LLM、文本转语音和视频音频生成，实现高效多模态实时服务，兼顾延迟、成本和质量。

2603.05527 2026-03-09 physics.chem-ph cs.NA math.NA physics.comp-ph 50%

Drifting to Boltzmann: Million-Fold Acceleration in Boltzmann Sampling with Force-Guided Drifting

漂向玻尔兹曼：在玻尔兹曼采样中实现百万倍加速的力引导漂移

Pipi Hu

专题命中效率与蒸馏：diffusion(abstract)

AI总结该研究提出力引导漂移模型，通过漂移分数恒等式实现分子构型生成，显著提升玻尔兹曼采样效率，达到百万倍加速并保持结构和分布准确性。

2510.00425 2026-03-06 cs.MA cs.RO 50%

Conflict-Based Search as a Protocol: A Multi-Agent Motion Planning Protocol for Heterogeneous Agents, Solvers, and Independent Tasks

基于冲突的搜索作为协议：一种多智能体运动规划协议用于异构智能体、求解器和独立任务

Rishi Veerapaneni, Alvin Tang, Haodong He, Sophia Zhao, Viraj Shah, Yidai Cen, Ziteng Ji, Gabriel Olin, Jon Arrizabalaga, Yorai Shaoul, Jiaoyang Li, Maxim Likhachev

机构 * Carnegie Mellon University（卡内基梅隆大学）； Tongji University（同济大学）； UC Berkeley（伯克利大学）

专题命中效率与蒸馏：diffusion(abstract)

AI总结本文提出了一种基于冲突的搜索协议，用于异构智能体的多智能体运动规划，利用多种单智能体规划算法实现无碰撞路径规划。

Comments Published at ICRA 2026, Project webpage: https://rishi-v.github.io/CBS-Protocol/

1. 效率与蒸馏 2542 篇

Mitigating Premature Discretization with Progressive Quantization for Robust Vector Tokenization

Show Me What You Don't Know: Efficient Sampling from Invariant Sets for Model Validation

Efficient first-principles inverse design of nanolasers

DIP: Efficient Large Multimodal Model Training with Dynamic Interleaved Pipeline

Efficient Nehari manifold optimization algorithms for computing ground state solutions of nonlinear elliptic systems

A Fine-Grained 3D Radio Map Construction Paradigm with Ultra-Low Sampling Rates by Large Generative Models

On-the-Fly ROM-Based Acceleration of SI-DSA for Implicit Time Marching of the Radiative Transfer Equation

OT-MeanFlow3D: Bridging Optimal Transport and Meanflow for Efficient 3D Point Cloud Generation

Fast Beam-Brainstorm: Few-Step Generative Site-Specific Beamforming with Flexible Probing

Early Quantization Shrinks Codebook: A Simple Fix for Diversity-Preserving Tokenization

FS_GPlib: Breaking the Web-Scale Barrier - A Unified Acceleration Framework for Graph Propagation Models

Masked Auto-Regressive Variational Acceleration: Fast Inference Makes Practical Reinforcement Learning

Controlled Langevin Dynamics for Sampling of Feedforward Neural Networks Trained with Minibatches

Efficient structure-preserving scheme for chemotaxis PDE system with singular sensitivity in crime and epidemic modeling

Targeting the partition function of chemically disordered materials with a generative approach based on inverse variational autoencoders

Memory-aware acceleration of orientational dynamics in nanoparticle suspensions

One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies

Strictly Constrained Generative Modeling via Split Augmented Langevin Sampling

Auroral Acceleration Generates Electron Beams in Jupiter's Middle Magnetosphere

Efficient design of continuation methods for hyperbolic transport problems in porous media

AlphaFlowTSE: One-Step Generative Target Speaker Extraction via Conditional AlphaFlow

Multi-spacecraft constraints on relativistic solar energetic particle transport in the widespread 28 October 2021 event

Symbolic Discovery of Stochastic Differential Equations with Genetic Programming

SCDP: Learning Humanoid Locomotion from Partial Observations via Mixed-Observation Distillation

From Flow to One Step: Real-Time Multi-Modal Trajectory Policies via Implicit Maximum Likelihood Estimation-based Distribution Distillation

Constraining turbulent solar flare acceleration regions by connecting kinetic modeling and X-ray observations

Learning Where the Physics Is: Probabilistic Adaptive Sampling for Stiff PDEs

StreamWise: Serving Multi-Modal Generation in Real-Time at Scale

Drifting to Boltzmann: Million-Fold Acceleration in Boltzmann Sampling with Force-Guided Drifting

Conflict-Based Search as a Protocol: A Multi-Agent Motion Planning Protocol for Heterogeneous Agents, Solvers, and Independent Tasks