Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 10: Focus Session: Large Language Models, Social Dynamics, and Assessment of Complex Systems
SOE 10.1: Invited Talk
Thursday, March 20, 2025, 15:00–15:30, H45
Emergent Behaviors in LLMs-Populated Societies — •Giordano De Marzo1,2,3, Claudio Castellano4,2, Luciano Pietronero2, and David Garcia1,3 — 1Konstanz University — 2CREF Enrico Fermi Research Center — 3CSH Vienna — 4CNR-ISC Institute for Complex Systems
Applications of Large Language Models (LLMs) increasingly involve collaborative tasks where multiple agents interact, forming "LLM societies." In this context, we explore whether large groups of LLMs exhibit emergent group behaviors similar to those in human societies. First, by simulating social network formation, we observe that LLMs spontaneously form scale-free networks. Agents connect through linear preferential attachment, mirroring the Barabasi-Albert model and real-world social networks. Second, we investigate the ability of LLMs to reach consensus on arbitrary norms without external preferences, thereby self-regulating their behavior. In human societies, consensus without institutions is limited by cognitive capacities. Similarly, we find that LLMs can reach consensus, with the opinion dynamics described by a majority force coefficient that determines the likelihood of consensus. This majority force strengthens with higher language understanding but decreases with larger group sizes, resulting in a critical group size beyond which consensus becomes unattainable. For more advanced LLMs, this critical size grows exponentially with language capabilities, exceeding the typical size of informal human groups.
Keywords: Large Language Models; Opinion Dynamics; AI