Proceedings of the Annual Meeting of the Cognitive Science Society

Wide applications of LLM-based agents require strong alignment with human social values. However, current works still exhibit deficiencies in self-cognition, emotion determination, and proactivity in social behaviors. To remedy this, we propose an evaluation scenario in which the agent needs to align with three famous psychological theories: Maslow's Hierarchy of Needs, Plutchik's Wheel of Emotion, and Moral Foundation Theory. We then design a novel value-based framework that employs GraphRAG to extract and index the prescriptive, human-written seed principles, forming a knowledge graph of emotions, needs, and moralities. The framework further conducts online query-based summarization based on a semantic retriever, with a top-k ranking mechanism. This dynamic instruction finally steers the agent such that it behaves as expected by the descriptive theories. We define the alignment metric as the ratio of expected behaviors, as well as the similarity-based metrics, when the golden responses are available. By experimenting with our method on the DAILYDILEMMAS benchmark, we observe significant performance gains over both prompting, finetuning, and retrieval-based baselines. Our method provides a basis for the social value alignment of LLM-based agents.