Abstract

Integrated Science, Technology, Engineering, and Mathematics (STEM) education demands personalized, adaptive instruction that addresses diverse student backgrounds and accommodates varying paces of concept mastery. Recent advances in Large Language Models (LLMs) provide promising avenues for automating pedagogical guidance and curriculum sequencing. However, standard LLM-based recommendations can suffer from overconfidence and hallucination, propagating suboptimal learning trajectories without transparent reliability metrics. In this study, we propose an uncertainty-aware LLM-based learning path recommendation framework designed for integrated STEM education. The system incorporates lightweight uncertainty quantification to assess the confidence of generative recommendations, triggering threshold-based verification and refinement when confidence drops. By coupling student knowledge representations with uncertainty-guided curriculum graphs, the framework delivers adaptive, reliable learning pathways tailored to individual learners. This work provides critical insights into deploying trustworthy, AI-enhanced educational tools to strengthen student engagement and learning outcomes across STEM curricula.