Neuro-Symbolic AI Framework for Automated Concept Reinforcement in Skill-Oriented STEM Education Environments

STEM education increasingly leverages AI-based tutoring systems, yet most rely on opaque neural models that lack interpretability, while symbolic systems struggle to handle unstructured learner responses. Existing approaches [10], [14] often achieve predictive accuracy but fail to deliver conceptlevel, pedagogically aligned, and explainable feedback, limiting skill reinforcement. This study proposes a hybrid NeuroSymbolic AI framework integrating RoBERTa-based contextual embeddings with a symbolic reasoning engine to detect conceptual gaps and provide targeted reinforcement. Implemented in Python using the MathE Mathematics Learning and Assessment dataset, the model processes normalized mathematical notation, maps concepts to a knowledge graph, and applies rule-based verification via SymPy. Results demonstrate a 94.5% accuracy, representing a 3.3% improvement over the best baseline, with notable gains in precision (93.8%) and recall (94.1%). The approach ensures interpretable, adaptive learning feedback, bridging the gap between prediction and pedagogy, and offers a scalable, highperformance solution to enhance personalized STEM education outcomes.