RLHF further sharpens a model's output distribution to align with chatbot-like behavior