Confidence in the Reasoning of Large Language Models