English(EN)The Geometric Nature and a Free Proxy for Flow-Matching Uncertainty
新研究探讨流匹配模型增强与漏洞 · 跟踪9个来源
作者PulseAugur 编辑部·[13 个来源]·
研究人员正在探索增强流匹配模型的新方法，流匹配是生成任务的一种流行范式。一篇论文引入了“去噪加速”（accel）作为估计流匹配动作不确定性的免费代理，通过识别错误输出来提高实时控制的安全性，而无需额外的计算开销。另一项研究提出了“DRIFT”，一种对抗性补丁攻击，通过靶向去噪轨迹来有效破坏流匹配的视觉-语言-动作模型，揭示了其感知鲁棒性中令人惊讶的漏洞。此外，还提出了“单侧分位数耦合流匹配”（QC-FM）和“能量引导流匹配”（EG-FM）等新方法，以提高训练效率和样本质量，其中QC-FM降低了回归方差，EG-FM对粗到精的生成轨迹进行建模以获得更好的图像生成。
AI
Pixel-space generative models bypass lossy latent compression, yet necessitate joint learning of global structure and fine-grained details in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean-image endpoint, leaving the spectral evolution to…
arXiv:2608.03207v1 Announce Type: cross Abstract: Flow-matching vision-language-action (VLA) models such as pi0 generate robot actions by integrating a learned denoising velocity field, and have been reported to resist adversarial perturbations that readily fool autoregressive VL…
arXiv:2607.27933v2 Announce Type: replace Abstract: Flow matching (FM) has become a popular action head paradigm for modern embodied models. However, as a conditional generative model, it does not explicitly expose its inherent uncertainty, producing faulty action chunks even whe…
arXiv cs.LG
TIER_1English(EN)·Jin-Young Kim, So-Yoon Cho, Hyun-Gyoon Kim·
arXiv:2608.00978v1 Announce Type: new Abstract: Flow Matching trains continuous-time generative models by regressing the velocity field of a probability path between a simple source distribution and a target data distribution. The coupling that pairs source and target samples str…
Flow-matching vision-language-action (VLA) models such as pi0 generate robot actions by integrating a learned denoising velocity field, and have been reported to resist adversarial perturbations that readily fool autoregressive VLAs. We show that this robustness is largely illuso…
arXiv cs.LG
TIER_1English(EN)·Fairoz Nower Khan, Nabuat Zaman Nahim, Peizhong Ju·
arXiv:2607.28698v1 Announce Type: new Abstract: Flow matching assumes fully observed training data, which many real-world applications rarely provide. We propose Missing-Data Flow Matching, which treats the missing coordinates of training samples as latent variables and averages …
Flow matching (FM) has become a popular action head paradigm for modern embodied models. However, as a conditional generative model, it does not explicitly expose its inherent uncertainty, producing faulty action chunks even when it misinterprets the scene or encounters out-of-di…
arXiv:2602.01591v2 Announce Type: replace Abstract: Recent advances in flow matching models, particularly with reinforcement learning (RL), have significantly enhanced human preference alignment in few-step text-to-image generators. However, existing RL-based approaches for flow …
arXiv cs.CV
TIER_1English(EN)·Haoyang Tong (MAIS & NLPR, CASIA, JD.com), Yu He (JD.com), Fang Li (JD.com), Lichen Ma (JD.com, Xi'an Jiaotong University), Jingling Fu (JD.com), Dong Chen (JD.com), Zhen Chen (JD.com), Junshi Huang (JD.com), Jie Cao (MAIS & NLPR, CASIA)·
arXiv:2608.05811v1 Announce Type: new Abstract: Pixel-space generative models bypass lossy latent compression, yet necessitate joint learning of global structure and fine-grained details in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean-i…
arXiv stat.ML
TIER_1English(EN)·Lennon J. Shikhman·
arXiv:2608.04531v1 Announce Type: cross Abstract: Functional flow matching is posed on distributions of functions but implemented from finitely many coefficients or point values. Under scattered or adaptive refinement, the resulting conditioning sigma-algebras need not be nested,…
arXiv cs.CV
TIER_1English(EN)·Adrian Urba\'nski, Gabriel della Maggiora, Artur Yakimovich·
arXiv:2608.00064v1 Announce Type: new Abstract: Generative models learn the statistical properties of their training data, so high-quality generation depends on clean and representative datasets. In scientific imaging, acquisition often yields noisy measurements, while collecting…
arXiv:2608.01990v1 Announce Type: new Abstract: Denoising diffusion transformers achieve strong generation quality but converge slowly during training. Regularizing their internal representations has emerged as an effective accelerator, yet existing methods split into two familie…
