Preprint: AdaTutor learns when to change tutoring strategy in simulated dialogues

A preprint submitted on October 26, 2026 introduces AdaTutor, a system that learns to choose among six tutoring strategies, including hints, worked examples and requests for self-explanation. Rajeev Bhatt Ambati and colleagues call the training approach Comparative Pedagogical Supervision. It compares possible responses within a tutoring context rather than prescribing the same form of help throughout a dialogue. Across mathematics and science reasoning datasets, the authors report stronger results than their best fixed-strategy baselines after five tutoring turns.

The evaluation boundary matters. The experiments use a stateful language-model student simulator, and the paper also examines the quality of generated responses. Those checks are useful for developing a tutoring policy, but they are not a classroom trial with measured student learning. Success with a simulator does not establish retention, transfer or the appropriateness of a strategy for a particular child. The work is presented here as a preprint, with those distinctions kept visible.

For educators, the design question is concrete: what evidence would justify changing the kind of help? The following activity is an AIEDHK teaching proposal. A teacher-education group could prepare three fictional responses to the same proportional-reasoning problem. One response reveals a misunderstanding of the relationship, another shows a sound setup with an arithmetic slip, and the third reaches the answer without explaining why it works. Each case creates a different reason to intervene.

Participants could first choose a response themselves and explain their choice. They might offer a contrasting example for the misunderstanding, a small checking prompt for the calculation or an explanation request for the unexplained answer. They would then inspect an AI tutor's next turn using the same criteria. The aim would be to determine whether the chosen strategy responds to the learner evidence, rather than whether its wording sounds friendly or sophisticated.

The group could extend each fictional exchange by one turn. If a learner now explains the relationship correctly, should the assistant continue with a worked example or reduce the support? Reviewers would identify the observation that changed their decision. Keeping the earlier state visible would prevent an apparently adaptive sequence from being judged only by its final message. It would also expose cases where the assistant repeats a strategy despite new information.

A later classroom pilot could examine a small number of teacher-approved tasks with appropriate participation arrangements. Teachers would record the help offered, opportunities for independent reasoning and the learner's subsequent explanation. A separate task after assistance ends would address a different question from immediate success during tutoring. Students' descriptions of helpful or intrusive support would add another perspective without replacing the performance evidence.

AdaTutor supplies a research approach to strategy selection. The educational opportunity is to make the choice of assistance discussable and testable. Before adopting such a policy, a school would still need evidence that the system recognizes relevant learner differences and that the resulting support advances its actual teaching objectives.