Social Dilemmas via Generative Agent-Based Models 

Conference open to outside participants. Funded by the Fondation de Sciences de la Modélisation & CY Initiative

Yamir Moreno is Professor & Director of the Institute for Biocomputation and Physics of Complex Systems (BIFI), and Professor at the Department of Theoretical Physics, Faculty of Sciences, University of Zaragoza, Spain. Prof. Moreno’s field of research is in the theoretical foundations of complex systems, which he investigates using tools from mathematics, physics, and network science. His interests include disease dynamics, diffusion processes, computational social sciences, nonlinear dynamical processes, and the structure and dynamics of complex systems.

Current Large Language Models (LLMs) have opened new avenues for modeling complex social dynamics. In particular, LLM-driven agents provide a unique opportunity to explore several phenomena in artificial societies. Admittedly, recent advancements have demonstrated that LLMs can exhibit human-like behaviors, including cooperation, fairness, and adherence to social norms. However, they also present significant challenges, such as sensitivity to prompt design, hallucinations, and inconsistencies in decision-making. In this talk, we discuss the capacity of GABMs to reproduce behavioral experiments and compare findings on cooperation and reputation dynamics in human groups with those obtained by implementing a reputation-based game in which LLM-driven agents played the Prisoner’s Dilemma on dynamics networks. Our results indicate that LLM-based agents can partially reproduce human cooperative behavior and network dynamics, though important limitations remain. These models will become increasingly integrated into decision-making processes, thus, understanding their constraints and interaction patterns with humans and artificial agents is crucial.