Evolutionary games in LLM-agent societies: Early milestones and perspectives

Evolutionary game theory (EGT) provides an efficient mathematical framework for describing strategic interactions and population dynamics, traditionally based on predefined microscopic rules of strategy update. In contrast, large language model (LLM) agents perceive environmental information and generate actions through natural-language prompts, providing greater flexibility in decision-making. Integrating LLM agents into EGT has created an intriguing new research direction that is attracting increasing scholarly attention. Our paper presents a brief survey of early steps along this path and offers some promising questions for future studies. We first summarize the operating principle of LLM agents in EGT, and then examine why their introduction does not necessarily enhance cooperation, with particular emphasis on the limitations of mechanisms previously identified in conventional evolutionary games. The conditions required to sustain cooperation among LLM agents are also discussed in depth. Finally, potential research directions are considered, including prompt evolution, long-term adaptation, and human–AI interaction.