💡Easy explanation

A simpler alternative to RLHF. Instead of training a separate reward model, you directly train on 'answer A is better than B' data.