Teaching a Large Language Model Tutor to Withhold the Answer: A Supervisor Architecture and an Evidence-Driven Method for Tuning Socratic Behavior
2026-08-12 • Computers and Society
Computers and Society
AI summaryⓘ
The authors built a tutoring chatbot that sometimes needs to refuse giving direct answers to help students learn better. They created a system that controls and checks each chatbot reply to make sure it doesn't give away solutions too easily. To improve the system without using real students, they tested it with fake student behaviors and used a stronger model to judge if the responses were appropriate. This process helped them identify different levels of giving too much help and fix each issue step-by-step, resulting in a tutor that reliably withholds answers when needed. They suggest their method as a general way to make AI agents refuse certain actions responsibly.
large language modelchatbot tutoranswer withholdingSocratic methodautomated evaluationhelp laddermachine-checkable contractscripted student personasLLM judgedeterministic detector
Authors
Yusuf Pisan
Abstract
An effective large language model (LLM) tutor must often decline to give an answer it could easily produce. In a randomized study, students who used an unguarded chatbot scored higher while practicing but lower on a later test taken without it, whereas a Socratically guarded version of the same model kept the practice gain and removed the later loss [4]. Reliable answer-withholding is therefore central to a tutor's value, yet a capable model pressed by a frustrated student does not withhold reliably on a prompt alone. We report a deployed tutoring system that enforces answer-withholding as a per-turn, machine-checkable contract, and a method for tuning that withholding against evidence. A non-LLM policy core, reading only trusted learner state, sets a per-turn ceiling on an eight-rung help ladder; a deterministic detector strips solution code; and a separate LLM judge checks each risky reply against the contract. We tune the behavior with an automated evaluation that uses no human subjects: scripted student personas are driven through the live pipeline and re-scored by a stronger model, and we record each rejection's stated reason so failures are fixed by cause. Doing so revealed an interpretable "over-help ladder," from blatant solution leaks, to naming the exact bug, to over-citing general facts, with each fix exposing the next. The tutor reached full compliance on all four acceptance criteria. We offer the measure, diagnose, and fix loop as a reusable recipe for any LLM agent that must refuse a capability it has.