“DeepMind's AlphaProof, which won a silver medal at the 2024 International Mathematical Olympiad, was a formal methods-based model, but the company used an informal reasoning model (Gemini) for its 2025 gold medal performance.”