[X]score: 0.36
RT by @huggingface: Paper of the day! https://huggingface.co/papers/2605.13301
May 15, 2026
Researchers released a 30B-active-3B MoE reasoning model achieving gold-medal performance on IPhO physics and IMO/USAMO math Olympiads using test-time self-verification and proof search scaling. The unified scaling recipe for reasoning models at this efficiency level is notable for practitioners building science and math agents.