How did it compare to v3.2-exp, I saw some impressive results with it! Additionally any chance you can walk through your methodology for this using the llm, I think that would be super interesting!