跳到正文
arXiv · ML· Zeyang Li, Yunan Wang, Risheek Garrepalli, Mohammad Ghavamzadeh, Navid Azizan·· 2026-10-06AI 评分37

Score-Calibrated Flow 用于从非归一化密度采样及在线强化学习

Score-Calibrated Flow for Sampling from Unnormalized Densities with Applications to Generative Online Reinforcement Learning

AI 导读

summary_zh: 作者提出 Score-Calibrated Flow (SCF) 算法,用于从非归一化密度中采样。该方法无需重要性采样或反向传播,通过强制自一致性来训练生成模型。

来源:arXiv · ML · arxiv.org