Comment by FergusArgyll

More RLVR. Give it verifiable problems, if it doesn't find a solution move on, if it does, use that as a reward signal.