I am interested in this too, as I think it could be these models are overly optimized for coding. Let me know if you figure it out!