评分依据:Addresses fairness issues in multi-objective DPO alignment. Important for responsible alignment practice.
MGDA-Decoupled: Geometry-Aware Multi-Objective Optimisation for DPO-based LLM Alignment
发布
采集
行业动态 6.5 分
— Addresses fairness issues in multi-objective DPO alignment. Important for responsible alignment practice. 原文: arXiv