CoT seems to improve performance at manipulation-style tasks

Impression gleaned from a Zvi roundup circa 24/1/2025. Will come back and link later. Basic impression: CoT works kind of like working memory, and is good for kind-of the same things - creating and arranging conceptual chunks of finite number in many different candidate configurations, then scoring these by their fit against the prompt. You can't directly configure more than a certain number of things, but you can certainly indirectly search a large space by choosing and configuring handles that help you navigate it. This is a pretty reasonable way to define what I call manipulation.