Notes

Short entries for formulas, small derivations, design observations, and research workflow fragments that do not need a full essay.

POMDP belief-state control

POMDP 把状态不可见的序列决策写成 belief state 上的控制问题,适合需要边观察、边更新、边行动的场景。

AdaWorld latent actions

A short note on action-aware pretraining as a reusable control interface for adaptable world models.