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.
The article analyzes the psychological phenomenon of users forming emotional attachments to AI agents, discussing concepts like social surrogacy and expectancy violation theory, and how this impacts user experience in professional settings.