arxiv How Easily do Irrelevant Inputs Skew the Responses of Large Language Models?