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

Created: 2026-08-03 | Tags: intelligent-tutoringstem-educationllmgenerative-aiautomated-gradingfeedback-loopscaffolding

Yicheng Sun, Yihan Liao & Xiaoxue Ma (2026) — Smart Learning Environments (Springer), Article in Press. Open Access, CC BY 4.0. doi:10.1186/s40561-026-00459-9.

📄 Full text (Springer, OA) · StructRAG code & prompt templates (GitHub)

Summary

Presents StructRAG, a pattern-aware framework that improves how AI tutoring systems interpret complex engineering diagrams (circuit schematics, network topologies, block flowcharts) in STEM. Current AI systems fail on diagrammatic questions because they cannot reliably extract spatial relationships and structural connectivity from noisy real-world diagrams (curved lines, overlapping elements, occluded labels). StructRAG shifts from surface-level visual recognition to pattern-level structural reasoning: convert the diagram to a graph, retrieve structurally similar canonical patterns, and prompt an LLM to validate plausibility and completeness — enabling detection of missing connections and structure-aware feedback.^[raw/papers/sle-2026-structrag-diagram-reasoning.md]

Architecture (four components)

1. Diagram-to-Graph Conversion — OCR (Tesseract, multi-scale 1.0/1.5/2.0, majority voting) + classical CV (Hough Line Transform, contour detection, Zhang–Suen skeletonisation for curved paths). Edge confidence Cij = 0.45·scont + 0.25·sprox + 0.20·salign + 0.10·snode; edges ≥ 0.65 accepted, 0.40–0.65 routed to an uncertain-edge set U for pattern-level validation, < 0.40 discarded as noise. 2. Structural Pattern Retrieval — abstracts the recognized graph into topology patterns (star, ring, chain/bus, bridge/mesh, tree, hybrid, cross-layer) and retrieves similar templates from a curated library using graph-edit-distance (GED) matching. 3. Pattern-Aware Prompt Construction — feeds the LLM the graph G, uncertain edges U, retrieved templates, and candidate corrections ΔE. 4. LLM-Guided Structural Reasoning and Correction — GPT-4 jointly interprets G, U, T*, and ΔE to decide which uncertain edges to add and which structures to correct, with output ensembling.^[raw/papers/sle-2026-structrag-diagram-reasoning.md]

Results (1,650 STEM diagram-based questions)

Interpretation

Related Pages

Citation

APA: Sun, Y., Liao, Y., & Ma, X. (2026). Advancing diagram-based reasoning in AI tutoring systems: A structural approach for STEM education. Smart Learning Environments. Advance online publication. https://doi.org/10.1186/s40561-026-00459-9