What should an AI tutor observe before intervening?

Before the next hint

A wrong answer tells a tutor that something went wrong. It may not reveal whether the learner misunderstood the task, lost track of a variable, or connected the right ideas incorrectly.

The observe-first tutor begins with an interaction question: what small observation would reduce uncertainty enough to choose a useful intervention?

Candidate observations

A prediction, a short trace, a comparison between two cases, or a learner-generated explanation could each reveal something different. Asking for more work is not automatically good; the observation has a cost for the learner.

The choice needs to be tied to a concrete pedagogical decision. A request to trace the loop is useful only if the response changes what the tutor does next.

From text to other modalities

The long-term question includes handwritten and physical work. Those settings add uncertainty about perception, sequence, and context. The note on physical-world companions keeps those challenges separate from what a text benchmark can establish.