Principles of metareasoning
Summary: The paper treats computation as an action an agent can choose. Since thinking takes limited time and its outcome is uncertain, an agent can use probability and decision theory to estimate whether more computation is likely to improve its next real-world decision. The useful question is not simply how to reason, but when further reasoning is worth its cost.
Why it stands out: It connects resource-aware reasoning to action: an intelligent system should spend effort where it can make a meaningful difference to what it does.