Advancing diagram-based reasoning in AI tutoring systems: a structural approach for STEM education

Abstract The rapid advancement of AI-powered tutoring systems has transformed educational environments, yet these systems still face significant challenges in accurately interpreting complex engineering diagrams commonly encountered in STEM fields, such as circuit schematics and network topologies. Current methods often struggle to recognize spatial relationships and detect missing connections, primarily due to the inherent complexity of real-world diagrams, which limits their effectiveness in diagram-based assessments. To address this, we propose StructRAG, a framework that combines visual parsing with pattern-level structural reasoning to enhance diagram interpretation. Our approach integrates OCR-based visual recognition, large language models, and graph pattern retrieval, enabling the system to identify missing connections, correct structural inaccuracies, and provide structure-aware feedback. We evaluate StructRAG on a dataset of 1,650 STEM-related questions. The results show a macro-average question-level accuracy of 89.3% and a micro-averaged edge-level F1 score of 93.0%, outperforming OCR+CV, GPT-4 graph-only, direct-image GPT-4o, and ablated StructRAG variants. Paired bootstrap tests with Holm–Bonferroni correction indicate that these improvements are statistically significant at $$p < 0.01$$ . These findings suggest that StructRAG provides more reliable diagram interpretation for complex and ambiguous diagram types, such as mesh and hybrid graphs. Overall, the framework shows potential as a structure-aware support tool for diagram-based assessment and feedback in STEM education.