Proceedings of the Annual Meeting of the Cognitive Science Society

The evolution of Large Language Models (LLMs) introduced a new paradigm for investigating human behavior emulation. Recent research employs LLM Agents in sociological environments, where they exhibit behavior based on unfiltered LLM characteristics. However, these studies overlook iterative development in human-like settings. Human preference and personality are complex, constantly shaped by environmental and subjective influences. We propose the Agent Framework for Shaping Preference and Personality (AFSPP), exploring the impact of social networks and subjective consciousness on Agents' preference and personality formation. AFSPP demonstrates trends consistent with key findings from human personality experiments. AFSPP-based results indicate that planning, perception, and subjective social networks have the most pronounced influence on preference shaping. AFSPP shows potential to enhance the efficiency of psychological experiments and provides insights into preventing undesirable preference and personality development for trustworthy AI.