Giving feedback at multiple steps during reasoning, not just at the final answer, to guide the model's thinking process.