One thing about the reasoning is that models are trained to generate a chain of thoughts, but it doesn’t have to be correct, accurate, or reflect the underlying logic of the LLM. It’s the same problem we have with the output, it is something plausible, but not that reliable