A new benchmark finds that today's best AI models can barely sketch their way through a math problem.
Researchers built MathVis, a set of 512 tasks that each pair a frontier math statement with a reference diagram, executable TikZ or SVG code, a gold intermediate representation, and detailed scoring rubrics. The goal was to see whether models can visualize a problem well enough to use the drawing as part of their reasoning, not just produce a pretty picture at the end. GPT-6-Astra led the pack with a pass@1 score of 31.25%, Claude Fable 5 scored 26.56%, and Gemini 3.1 Pro Preview came in at 22.85%. The researchers also noted that models tended to lean on formal symbols and graphics code rather than reasoning through an actual diagram, and even outputs that technically included every required element often looked visually confusing to a human reader.
That gap matters because diagrams in math are not decoration, they are a reasoning tool mathematicians use to catch errors and spot structure that equations alone can hide. Text-based math benchmarks have seen scores climb for years, but MathVis suggests visual reasoning, the kind a student does when sketching a proof on a whiteboard, is a separate and much weaker skill for these systems.
A diagram that compiles cleanly but still confuses the person reading it is not really a diagram. It is code wearing a diagram's clothes.