At the 66th International Mathematical Olympiad (IMO) held in Australia, Google DeepMind’s Gemini model made history by achieving gold-level recognition. It successfully solved five out of six challenging problems, scoring 35 out of 42 points, which aligns with the gold threshold established by the IMO grading committee. This marks a significant improvement from the previous year when DeepMind’s AlphaProof and AlphaGeometry 2 only secured silver by solving four problems. Gemini’s innovative approach enabled it to generate rigorous, natural language solutions directly from official descriptions within the 4.5-hour contest timeframe. Its answers were praised by official graders for their clarity and precision. DeepMind attributes Gemini’s success to advanced reinforcement learning, a comprehensive mathematical solutions database, and improved parallel reasoning techniques. While other AI systems, like those from OpenAI, reported similar results unofficially, Gemini’s performance was officially validated, marking a pioneering achievement for autonomous mathematical systems. DeepMind plans to share the model with select mathematicians soon.
Source link
