Large Language Models Still Face Challenges in Multi-Hop Reasoning with External Knowledge