The DPO debate: Do we need RL for RLHF?

Direct vs. RL methods for preferences, more RLHF models, and hard truths in open RLHF work. We have more questions than answers.

questions rlhf work