Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 10: Focus Session: Large Language Models, Social Dynamics, and Assessment of Complex Systems
SOE 10.2: Vortrag
Donnerstag, 20. März 2025, 15:30–15:45, H45
A Framework for Multi-Step Discussions of LLM-Based Agents — •Vincent Brockers1,2, David Ehrlich2,3, and Viola Priesemann1,2 — 1Max Planck Institute for Dynamics and Self-Organization — 2University of Göttingen — 3Göttingen Campus Institute for Dynamics of Biological Networks
Recent studies have demonstrated the utility of Large Language Models (LLMs) across various settings and domains, particularly in their ability to approximate human behavior for simulating and studying societal phenomena. A critical component of this capability is their proficiency in generating realistic dialogues. We are interested in the collective dynamics that arise from their dialogues, in particular in the context of understanding collective opinion dynamics.
We developed a framework for simulating multi-step discussions among LLM-based agents, which is crucial for these studies. We conduct an in-depth analysis of the components of such simulations, highlighting the challenges and specifically investigating potential biases.
We find that these opinion dynamics are primarily influenced by model-inherent topic biases, which exhibit attractors in the opinion space. Additionally, the framing of opinion evaluation plays a crucial role in analyzing these dynamics. Our findings offer valuable insights for future research aimed at optimizing the design of LLM agent-based simulations and enhancing our modeling capabilities of complex systems, such as human networks.
Keywords: Opinion Dynamics; Collective Dynamics; Large Language Models (LLMs); Social Networks; Artificial Intelligence (AI)