Towards Multi-Value Alignment via Value Localization in LLMs
Aligning large language models (LLMs) with multiple, potentially conflicting human values remains challenging, as existing multi-objective alignment methods often require costly fine-tuning and provide limited insight into the internal mechanisms underlying value-related behaviors. In this work, we investigate multi-va...