EGS: a supplementary safety framework for educational LLMs
EGS: A Complementary Safety Framework for Educational LLMs
AI summary
The educational guardrail system EGS is a lightweight, model-agnostic runtime framework that supplements the intrinsic alignment of educational LLMs through query moderation, scenario-aware pedagogical guidance and response moderation. It routes inputs by risk level, guides gray-area queries with category-specific educational interventions, and moderates outputs before delivery. Experiments on general and education-specific benchmarks show that EGS improves moderation performance, refusal quality and jailbreak robustness across multiple LLMs, and the code and model weights will be released.