This paper asks a simple but important question: how can a large language model (LLM) tutor that knows the answer refuse to give it? Research shows that when students get instant answers from an AI, they can do well while practicing but learn less for later tests. The authors treat answer-withholding as the core job of a tutor, not a safety add-on. They call the difficult case “refusal-under-knowledge,” meaning the system knows the solution, the student sees it could know, and yet the tutor must withhold while still being helpful.

The team built a supervisor architecture that enforces withholding for each turn of dialogue. A non-LLM policy core computes a per-turn contract using only trusted learner state (not the student’s words). Help is measured on an explicit eight-rung “help ladder,” from H0 (encourage, no new content) up to H7 (show the full, compilable solution). A deterministic detector strips or blocks solution code before anything else. When a reply looks risky, a second large language model acts as a judge to check the draft output against the contract. The whole system logs every decision so it is auditable.

To tune the system, the authors ran an automated evaluation with no human subjects. Scripted student personas were driven through the live pipeline. A stronger model re-scored each turn and every rejection was recorded with its stated reason. That loop—measure, diagnose, fix—revealed an “over-help ladder”: from blatant solution leaks, to naming the exact bug, to over-citing general facts. Fixes for one failure exposed the next, and the tutor eventually met four stated acceptance criteria for compliance.

Why this matters: schooling research says students learn best when they do the thinking themselves. By making answer-withholding a machine-checkable, per-turn contract, the system aims to keep students doing the work while still giving stepwise help. The design also favors revising a reply over flat refusal, so students stay engaged rather than abandoning the tutor for an unguided chatbot. The authors present this measure-diagnose-fix recipe as a reusable approach for any LLM agent that must refuse a capability it actually has.