[3J5-GS-5] Agents:

[3J5-GS-5-05] Continuous individuality emergence of LLM agents through interaction with the environment

〇Masatoshi Fujiyama¹, Ryohei Orihara¹, Yasuyuki Tahara¹, Akihiko Ohsuga¹, Yuichi Sei¹ (1. The University of Electro-Communications)

Keywords:LLM, agent, personality

In recent years, many social simulations have been conducted using LLM agents, including Generative Agents. Many of these simulated social behaviors by giving agents personalities such as predefined roles and relationships with other agents. However, in interactions based on predefined roles and relationships manually assigned to agents, only limited social behaviors can be observed. This approach makes it challenging to simulate relationships between agents with more diverse individuality. The purpose of this study is to analyze whether diverse social behaviors and relationships are formed among LLM agents by simulating their lives in an environment without setting tasks in advance, and whether individuality emerges and transforms continuously through the simulation. In the simulation, many agents tended to form clusters, but some agents remained isolated. The analysis of these results showed differences in the survival rate of agents, changes in the direction of their messages, and changes in their individualities depending on whether they formed clusters or not.