I think it's exploring in-context. Bringing up related ideas and not getting confused by them is pivotal to these models eventually being able to contribute as productive reasoners. These traces will be immediately helpful in a real world iterative loop where you don't already know the answers or how to correctly phrase the questions.