How reinforcement learning teaches large language models to reason.