Do as We Do, Not as You Think: the Conformity of Large Language Models