A reinforcement learning framework that trains models to predict both answers and the reasoning skill used at each step.