Berlin 2024 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

SOE: Fachverband Physik sozio-ökonomischer Systeme

SOE 7: Poster

SOE 7.11: Poster

Montag, 18. März 2024, 18:00–20:30, Poster D

Opinion Dynamics of Interacting Large Language Models — •Vincent Brockers, David Ehrlich, Andreas Schneider, and Viola Priesemann — Max-Planck-Institute for Dynamics and Self-Organization

Classical models of opinion dynamics rely on purely symbolic representations of arguments, claims and reasons, potentially oversimplifying certain aspects of complex human behavior. This complexity can better be captured with recent Large Language Models (LLMs), as these possess the ability to mimic sophisticated human-like interactions with semantic meaning. However, how these new tools can be effectively utilized to improve on classical models remains an open research question.

Our research focuses on the interactions of LLM agents with different personality types and opinions in discussion scenarios as seen as on social media platforms. We use the Big Five personality traits to define the LLMs character and assign an initial opinion for one certain topic. Letting agents discuss for multiple rounds allows us to measure their individual influence-response-functions, which give insights about the process of opinion change, depending on the agents agreeableness, initial opinion and discussion topic.

In our simulations we find similar mechanisms of opinion reinforcement as seen as in symbolic models and that the strength can be controlled via the agreeableness trait. These findings and further analysis, especially how such agents behave in larger networks, are crucial to assess the applicability of LLMs in this research area.

Keywords: Opinion Dynamics; Large Language Models; Social Networks