How Can Large Language Models Understand