A general theoretical paradigm to understand learning from human preferences

The paper proposes a novel approach to aligning language models with human preferences, focusing on the use of preference optimization in reward-free RLHF.

BibTex: