Importance Sampling

A research note on adaptive granularity importance sampling for policy optimization, focusing on speech token dependencies and segment-level weighting strategies such as MASPO-Fixed, MASPO-LogProb, and MASPO-TokenVal.

一篇关于 policy optimization 中自适应粒度重要性采样的研究笔记，重点讨论语音 token 依赖结构以及 MASPO-Fixed、MASPO-LogProb、MASPO-TokenVal 等分段策略。