The opinions that it is or isn't "reasoning" are widely varied and depend heavily on interpretation of the interactions, many of which are hersey.
My own testing with OpenAI calls + using Weaviate for storing historical data of exchanges indicates that such a beast appears to have the ability to learn as it goes. I've been able to teach such a system to write valid SQL from plain text feedback and from mistakes it makes by writing errors from the database back into Weaviate (which is then used to modify the prompt next time it runs).