Research Summary: EduGuard: A Safe RAG-Based LLM Tutor for Programming Education

Research introduces EduGuard, a novel retrieval-augmented generation (RAG) framework designed to create safe and pedagogically sound large language model (LLM) tutors for programming education. It addresses critical issues such as hallucination, policy contravention, direct solution provision, and student over-reliance prevalent with unrestricted GenAI tools. EduGuard integrates multiple safeguards, including instructor-approved content retrieval and pedagogical strategy selection, to enhance the reliability and educational efficacy of AI-driven tutoring.

Why it matters

The responsible integration of AI, particularly LLMs, into educational contexts is a critical strategic imperative to ensure academic integrity and effective learning outcomes. Frameworks like EduGuard demonstrate a pathway to mitigate inherent risks while leveraging AI's potential for personalized instruction, thereby shaping future educational strategies and technology adoption policies.

Key insights

• Unrestricted generative AI (GenAI) tutors pose significant risks in programming education, including hallucination and contradiction of course policies.
• Such tools can reveal complete solutions to assignments, potentially undermining learning objectives and fostering passive dependence among students.
• EduGuard is presented as a safe RAG-based LLM tutoring framework specifically for introductory programming.
• The framework incorporates query understanding, retrieval of instructor-approved course materials, and selection of appropriate pedagogical strategies.
• It also features rubric-aware generation, claim-level verification, and mechanisms to control student over-reliance.
• A benchmark dataset, BILearn-CS, comprising instructor-authored and TA-validated queries, was developed to explicitly evaluate the system's provenance.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2607.15738