It can't reason at all. This has been shown time and time again, and is inherent to the very nature of how LLMs work.