EduGuard: New AI tutor cuts hallucination to 4.9%, boosts learning by 13%

EduGuard, a safe retrieval-augmented generation (RAG) tutoring framework for introductory programming, integrates query understanding, instructor-approved course retrieval, pedagogical strategy selection, rubric-aware generation, claim-level verification, and overreliance control. Using Meta-Llama-3.1-8B-Instruct as the primary generator, hybrid FAISS/BM25 retrieval, and architecturally separate DeBERTa-v3-large-MNLI verifier, it significantly outperforms strong baselines including GPT-4o-mini Tutor and Llama Socratic Tutor. On the newly constructed BILearn-CS benchmark (600 instructor-authored, TA-validated queries), EduGuard achieves 90.1% correctness, 89.4% grounding, and 90.8% rubric alignment, with the lowest hallucination rate (4.9%) and direct-answer leakage (9.8%). A controlled pilot with 10 undergraduates showed post-test accuracy improving from 68.4% to 81.2%, while overreliance dropped from 38.0% to 17.0% relative to GPT-4o-mini Tutor.

The framework addresses critical GenAI risks in education: hallucination, policy violation, answer leakage, and student passivity. By enforcing pedagogical guardrails and evidence verification, EduGuard demonstrates that safe tutoring requires more than just retrieval or strong prompting. The results suggest that explicit pedagogical control, claim verification, and deployment safeguards are essential for responsible AI in education.