Chain-of-thought.

prompting (or training) a model to work through its reasoning step by step before answering, which improves results on harder problems. Whether it’s genuine reasoning is contested.