Emergence of new strategies for gaming through collective interaction of LLM agents

Abstract

In the study of Generative Agents, the initial characteristics of each agent, such as personality and memory, traditionally needed to be predefined. This research shifts the focus towards the emergence of agents' personalities from scratch, particularly in LLM agents, without the need for explicitly defining their initial characteristics. The LLM agents engage in cooperative communication within a group game, sending context-based messages in natural language. Through analyzing this simulation, we gain insights into how norms and cooperation develop spontaneously and how personality traits manifest without predetermined specifications. During the presentation, we will explore the concepts of collective intelligence, individuality, and social norms as they arise naturally among Generative Agents.