A language model asked to draw a geometry diagram produces something that looks fine. So does a wrong one. In GeoGenBench every prompt ships with its own key, a short list of facts a compiler can check from the coordinates. No vision model, no opinion. Try to pass the key by hand.
Drag A, B, C, and D. The key is checked live. Tolerances: one degree for angles, two pixels for incidence, the same kind of thresholds the benchmark's checker uses on real coordinates.
The point is not that the checks are hard. It is that they are exact, cheap, and generated with the prompt, so a thousand diagrams can be graded the same way a human grades one. How the real pipeline does it · The dataset · Back to the site