New Framework to Prevent Pedagogical Leakage in AI Tutors
A recent research article presents a structured approach aimed at curbing pedagogical leakage in large language model (LLM) tutors, preventing models from revealing answers or critical reasoning without permission. The authors introduce an authorization-aware complete-mediation boundary, which includes a selector that generates one of five disclosure contracts, trusted policy gates for privileged modes, and a language-rendering component. A unified release function enforces inspectable checks, optional cumulative verification, and action-specific fallback, with replayable traces distinguishing failures in selection, generation, verification, and enforcement. In experiments involving 599 fixed Gemini 3.5 proposals, strict mediation eliminated leakage flags from 181 to 0, yielding a paired problem-cluster difference of -30.22 points (95% CI [-35.00, -25.72]), while also replacing 581 responses and diminishing helpfulness. This paper, addressing a vital concern in AI tutoring systems, can be found on arXiv under identifier 2608.00515, categorized as 'cross'.
Key facts
- The paper formalizes pedagogical leakage as a state- and action-dependent failure in LLM tutors.
- Introduces an authorization-aware complete-mediation boundary with five disclosure contracts.
- The release function includes inspectable checks, cumulative verification, and fallback mechanisms.
- Replayable traces separate selection, generation, verification, and enforcement failures.
- Matched component attribution reveals a safety-utility frontier.
- Testing on 599 Gemini 3.5 proposals reduced leakage flags from 181 to 0.
- The intervention replaced 581 responses and lowered helpfulness.
- The paper is available on arXiv with identifier 2608.00515.
Entities
Institutions
- arXiv
- Gemini