Why not? I mean R1 even had an explicit preference reward at the end of training; I think that's there exactly for that purpose.

→ View original post on X — @rasbt