Skip to main content
Back

Ben-Gurion Researchers Teach AI to Read Math Teachers' Geometric Reasoning

Researchers built a 33-skill dictionary that lets LLMs classify teachers' geometric reasoning across five Van Hiele levels, beating skills-blind baselines.

  • First automated system to infer teachers' Van Hiele geometric reasoning levels from open-ended responses using LLMs.
  • Uses a 33-skill dictionary built with math educators, decomposing each of the five Van Hiele levels.
  • Two methods tested: RAG with Gemini 2.0 Flash and multi-task fine-tuning of Gemma-3-4B-IT with LoRA.
  • Skills-aware MTL reaches 0.725 F1-macro and 0.717 QWK, beating skills-blind baselines significantly.
  • Noisy-skills ablation confirms gains come from correct skill semantics, not extra prompt context.
  • Code, dataset and skills dictionary released on GitHub.

Assessing whether a math teacher truly understands geometry has always required an expert to sit down and read through open-ended answers one by one. A new paper accepted at AIED 2026 shows that large language models can automate this diagnosis, but only if you first hand them a structured map of the reasoning skills they're supposed to look for.

The team from Ben-Gurion University and the University of Edinburgh tackles the Van Hiele model, a well-established framework in math education that describes five hierarchical levels of geometric reasoning demonstrated by learners. Figuring out which level a teacher operates at requires reading their explanations, not grading multiple choice, which has kept sample sizes small for decades.

Why correctness misses the point

The five Van Hiele levels climb from concrete to abstract. Level 1 is Visualization, where learners identify shapes by overall appearance. Level 2 is Analysis, describing properties like a square having four equal sides. Level 3 is Informal Deduction, reasoning about relationships between shape classes. Level 4 is Deduction, constructing formal proofs. Level 5 is Rigor, involving abstract reasoning in non-Euclidean or axiomatic systems.

The classic shortcut, a multiple-choice test built by Usiskin in 1982, has been criticized because Van Hiele levels comprise multiple distinct reasoning processes that surface only in how a learner explains their thinking. Consider a student asked whether a shape with four equal sides is necessarily a rhombus. Answering "No, maybe it's a square" reveals Level 2 reasoning, while "Yes, because a square is also a type of rhombus" reveals Level 3 by grasping class inclusion. Same correctness, very different cognition.