Game-theory behaviour of large language models: The case of Keynesian beauty contests

The growing adoption of large language models (LLMs) pre-sents potential for deeper understanding of human behav-iours within game theory frameworks. This paper examinesstrategic interactions among multiple types of LLM-basedagents in a classical beauty contest game. LLM-based agentsdemonstrate vary...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Economics and business review Jg. 11; H. 2; S. 119 - 148
1. Verfasser: Estee Lu, Siting
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Poznan Poznań University of Economics and Business Press 01.06.2025
Wydawnictwo Uniwersytetu Ekonomicznego w Poznaniu
Sciendo
Poznan University of Economics
Schlagworte:
Algorithms
Beauty contests
Behavior
C63
C70
C90
Competition
Decision making
economic games
Experiments
Game theory
Human subjects
ICT Information and Communications Technologies
Keynesian theory
Large language models
Simulation
Social research
Social sciences
Socio-Economic Research
strategic interactions
strategic interactions
economic games
large language models
ISSN:2392-1641, 2450-0097
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:The growing adoption of large language models (LLMs) pre-sents potential for deeper understanding of human behav-iours within game theory frameworks. This paper examinesstrategic interactions among multiple types of LLM-basedagents in a classical beauty contest game. LLM-based agentsdemonstrate varying depth of reasoning that fall withina range of level-0 to 1, which are lower than experimentalresults conducted with human subjects in previous studies.However, they do display a similar convergence pattern to-wards Nash Equilibrium choice in repeated settings. Throughsimulations that vary the group composition of agent types,I found that environments with a lower strategic uncertaintyenhance convergence for LLM-based agents, and environ-ments with mixed strategic types accelerate convergencefor all. Results with simulated agents not only convey in-sights into potential human behaviours in competitive set-tings, but also prove valuable for understanding strategicinteractions among algorithms.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2392-1641
2450-0097
DOI:10.18559/ebr.2025.2.2182