This paper fixes a common problem in reasoning AIs called Lazy Reasoning, where the model rambles instead of making a good plan.
Big reasoning AIs think in many steps, which is slow and costly.