DeepMind’s new math model gets IMO Gold without any human demonstration data

AlphaGeometry solved 25 Olympiad problems in the allotted time limit, improving on the previous state-of-the-art, which solved 10.

Importantly, it was trained on 100 million synthetic problems, which eliminates an existing data bottleneck. DeepMind has open-sourced the code and weights.

SOURCE: deepmind.google/discover/blog/alphageometry-an-olympiad-level-ai-system-for-geometry