EDUHints: A Human-in-the-Loop Small Language Model Hint Generation System for Cybersecurity Education

Abstract

The problem that we study is how to efficiently generate hints for students who are engaged in hands-on cybersecurity exercises. Students sometimes get stuck and can become frustrated when they are missing information that is necessary for solving a challenge. While large language models (LLMs) could help, they can be expensive to use and typically require the sharing of student data with third-party AI providers. In order to minimize computational overhead and financial costs, we chose to deploy a small language model (SLM) with retrieval-augmented generation (RAG). In addition, we use a human-in-the-loop approach, where the instructor reviews the AI-generated hints before they reach the student. This keeps the instructor involved, increases the quality of the hints presented to the student, and preserves student-instructor interaction while reducing the cognitive load on the instructor. We have tested our hint generation system “EDUHints” in the classroom, collecting qualitative responses from 15 students via three brief surveys.