Towards Multi-Value Alignment via Value Localization in LLMs

Xueqi Ma, Yanbei Jiang, Xingjun Ma +2

Aligning large language models (LLMs) with multiple, potentially conflicting human values remains challenging, as existing multi-objective alignment methods often require costly fi…