I have released two new arXiv preprints that examine a shared problem in trustworthy educational AI: how to place explicit, testable controls between model output and deployment. Both manuscripts are preprints and have not yet completed peer review.
The first control point is executable. “EduPluginBench: Executable Assurance for AI-Generated Educational Plugins” introduces a benchmark and staged admission process for generated educational plugins. In a study of 1,440 activation-checked first-order mutants from 30 specifications, extending the assurance process from P0–P2 to P0–P4 increased release-blocking-defect recall by 74.7 percentage points, with no observed rejections among 120 clean references. In a separate transfer study, 300 of 600 generations parsed, but none passed P0 or achieved P0–P4 conformance.
The second control point is pedagogical. “Auditable Release Control for Pedagogical Leakage in LLM Tutors” formalises a failure mode in which a tutor response is correct and helpful but reveals decisive content before the learner is authorised to receive it. The paper evaluates an authorisation-aware complete-mediation boundary. In the reported evaluations, high-assurance release reduced panel-majority leakage flags while also documenting response-replacement costs, lower helpfulness, and residual failures in an external replication.
Read the pedagogical leakage paper on arXiv
Together, the papers argue that educational AI needs governance at two layers:
- executable assurance over what generated components may do; and
- release control over what tutoring systems may disclose at a given learning state.
The broader aim is machine pedagogical intelligence that is not only capable, but also testable, attributable, and auditable. Feedback and replication are welcome.
arXiv-issued DataCite DOIs
https://doi.org/10.48550/arXiv.2608.00739
https://doi.org/10.48550/arXiv.2608.00515
Leave a Reply