Tag
This paper proposes SoVA, a framework using GraphRAG to align LLM-based agents with human social values by converting psychological theories into prescriptive instructions. Experiments on the DAILYDILEMMAS benchmark show significant improvements over prompt-based baselines.