Signal Teaching a large language model tutor to withhold the answer

Summary

Researchers report on a deployed LLM tutoring system designed to avoid giving students direct answers. The system enforces this answer-withholding behavior as a contract that is mechanically checkable on every conversational turn. A non-LLM policy core that reads only trusted learner state sets a ceiling on how much help can be given, a separate detector filters out solution code, and another LLM acting as a judge checks risky responses against the contract's terms. The team tuned the system's behavior by repeatedly running it against scripted student personas rather than human subjects. The tutor achieved full compliance across all four defined acceptance criteria. Along the way, the researchers also surfaced interpretable failure patterns involving the system giving excessive help.