강화 학습 기반 후훈련

the team allocated over 10% of their compute budget to reinforcement learning-based post-training, creating a model that genuinely reasons rather than pattern-matches.