Evaluating the Robustness to Instructions of Large Language Models