Researchers Propose Using Synthetic Critiquing to Train Models