Reward-free Alignment for Conflicting Objectives

Peter Chen, Xiaopeng Li, Xi Chen +1

Direct alignment methods are increasingly used to align large language models (LLMs) with human preferences. However, many real-world alignment problems involve multiple conflictin…