Mathematical diagrams are central to K–12 instruction, but they remain difficult for AI systems to generate reliably. Effective diagrams must be mathematically accurate, visually clear, and appropriate to the instructional context.
This post describes Coteach’s collaboration with Anthropic to improve mathematical diagram generation by creating an autonomous R&D loop. The system combines a benchmark based on real educator requests, an evaluator calibrated to human expert judgment, and AI agents that can develop hypotheses, run experiments, benchmark results, and iteratively improve the diagram-generation pipeline.
The work led to a two-layer approach: a framework layer that gives AI models structured tools for creating mathematically accurate and visually consistent diagrams, and a knowledge layer that incorporates educator-informed guidance about the pedagogical use of different mathematical representations.
On Coteach’s benchmark of 250 K–12 diagram requests, its system produced classroom-ready diagrams 88.8% of the time, compared with 60.0% for GPT Image 2 and 54.4% for Gemini 3 Pro Image. The post also discusses what the team learned about combining autonomous agents with human expertise to advance AI research and development for education.
