On the Democratic Society of LLM Agents

The emergence of Large Language Models (LLMs) has opened new avenues for simulating complex socio-political dynamics with human-like agents. In this work, we introduce De CivAI, a platform designed to simulate a democratic society using LLM-based agents endowed with personalities, emotions, and memory. Our simulation focuses on the interplay between individual and collective interests in the face of a common threat, specifically a looming flood that requires collective action to build a protective dam. We implement a democratic policy-making process where agents engage in conversations, propose policies, and make collective decisions through voting. To evaluate the effectiveness of this framework, we design experiments to measure agreeableness, favorability towards dam-building policies, and the impact of democratic processes on cooperative behaviors. Our results demonstrate that democratic policy-making can significantly influence agents' willingness to cooperate for the common good, revealing emergent behaviors that align with socio-political theories. This work advances the understanding of multi-agent systems in socio-political contexts and highlights the potential of using LLM-based agents for exploring complex societal dynamics.