Evaluating LLMs' Inherent Multi-hop Reasoning Ability