“The GRPO algorithm gained prominence not because it was a major theoretical leap, but because DeepSeek successfully scaled it and released a high-performing model trained with it.”