<p>Current behavior decision-making methods for autonomous vehicles often struggle to generalize to complex scenarios. Large language models (LLMs) offer a promising approach to address these limitations. This study investigates the integration of LLMs into decision-making modules for autonomous vehicles (AVs), leveraging their reasoning capabilities to emulate intricate human-like driving behaviors and overcoming the limitations inherent in traditional methods. First, LLMs are incorporated into decision-making modules, together with a rule-based planner to complete closed-loop tasks. A scoring module is integrated to evaluate decisions, thereby reducing uncertainty associated with LLMs. Second, a memory module is introduced to facilitate a reflection mechanism, prompting LLMs to perform error correction to refine decision accuracy, assisted by human experts feedback. Subsequently, an extended naturalistic driving dataset highlighting complex scenarios is used for more rigorous evaluation. Experimental results on the extended nuPlan dataset demonstrate that the proposed methods effectively address complex scenarios, outperforming existing methods and improving the driving score by 13.6% to 22.7%. Furthermore, the reflection mechanism enhances contextual reasoning ability, resulting in more stable route completion and fewer task failures in complex scenarios.</p>