Olympiad-Level LLM Reasoning

30B-A3B gold IPhO/IMO/USAMO via test-time self-verification. Lean formal verification + evolutionary search solves open Erdős problems; OpenAI reasoning model disproves 1946 Erdős conjecture. First large-scale empirical study of formal theorem proving by LLM agents flagged (Dan Roy). Aligns with RLVR/self-distillation threads.