Exploring the Hidden Reasoning Process of Large Language Models by Misleading Them