AlphaGeometry: DeepMind solves olympiad-level geometry
In one sentence Google DeepMind publishes AlphaGeometry in Nature, a neuro-symbolic system that solves 25 of 30 IMO geometry problems on the IMO-AG-30 benchmark, close to the average gold medalist's 25.9, trained without human demonstrations.
DeepMind unveils an AI that can solve geometry problems at the level of the International Mathematical Olympiad (IMO).
It was not trained on human-written solutions: the system generated a huge number of random geometric figures on its own, derived 100 million synthetic examples of problems with proofs, and learned from those.
The system combines two parts: a language model that suggests which constructions to add to the figure (a point, a line, a circle) and a symbolic engine that carries out the proof rigorously. On 30 olympiad geometry problems it solves 25 within competition time; the average gold medalist solves 25.9.
Companies
Google DeepMind
Tools
AlphaGeometry
Tags
Sources