RL for Reasoning by Adaptively Revealing Rationales