Researchers built an AI system that edits geometry diagrams instead of just describing them, and it works better than doing the math in words alone.
The framework, called MetaReason, gives vision-language models a scratchpad to draw on when solving plane geometry problems. Instead of reasoning only in text, the system parses a diagram into structured meta-information, then uses predefined editing tools to add auxiliary lines and generate a new, high-fidelity image of the modified figure. The model reasons over that updated image the way a student redraws a triangle to spot a hidden angle. To train and test it, the researchers built TutorGeo, a dataset with 17,000 image-to-meta conversion samples plus 60,000 text-only and 60,000 interleaved multimodal reasoning traces, and ExamGeo, a benchmark drawn from real exam questions.
Most multimodal models still solve geometry by translating a picture into words and letting the language model take it from there, which throws away the spatial precision that makes geometry solvable in the first place. Earlier attempts to keep an image in the loop stumbled because the generated diagrams were geometrically wrong or badly rendered - garbage in, garbage out. Structured editing instead of freeform image generation is the fix on offer here, and by the paper's own benchmarks it beats other open-source models and closes much of the gap with proprietary ones.
Whether that scratchpad trick generalizes past textbook geometry problems to messier real-world diagrams is the question the paper leaves open.