The Observe-First Adaptive Tutor

Research question

What should a tutor observe before it decides to intervene, and what should it leave for the learner to do?

This project connects the counterfactual benchmark to a possible future intervention. The benchmark studies response sensitivity in a controlled setting; an intervention would need evidence about learning and interaction over time.

Current hypothesis

A small diagnostic interaction may sometimes be more useful than an immediate hint. Whether that holds depends on the task, the learner, and the cost of asking. It is a design hypothesis to test.

Proposed interaction loop

1. Observe the learner’s current work.
2. State uncertainty about what the work demonstrates.
3. Ask for a targeted prediction, trace, or explanation when needed.
4. Choose a scaffold tied to that evidence.
5. Leave a substantive step for the learner.
6. Observe what happens next and revise the interpretation.

Methods

A future evaluation should distinguish response quality, diagnostic value, learning outcomes, and the learner’s experience of control. The proposed system should also let a learner challenge its interpretation.

The note on what to observe is a place to develop the diagnostic choices. Solution leakage supplies one boundary on what an intervention might reveal.

Future modalities

Voice, multimodal interaction, and observation of handwritten or physical work are future directions. They have not been implemented or validated by this project page.

Results & artifacts

No intervention results, implementation repository, dataset, or released system are linked yet.

Open questions

How often should the tutor wait? How can it recover from an incorrect interpretation? What amount of observation is useful without becoming burdensome? How can adaptation remain accountable to the learner?