Reward-free Alignment for Conflicting Objectives 事件

Reward-free Alignment for Conflicting Objectives arXiv:2602.02495v3 Announce Type: replace Abstract: Direct alignment methods are increasingly used to align large language models (LLMs) with human preferences. However, many real-world alignment problems involve multiple conflicting objectives, where naive aggregation of preferences can lead to unstable training and poor trade-offs. In particular, weighted loss methods may fail to identify update directions that simultaneously improve all object