🤖 AI Summary
This study investigates the economic decision-making capabilities of large language models (LLMs) in the Beauty Contest Game, focusing on strategic reasoning, opponent modeling, and parameter adaptability. We construct virtual agents based on state-of-the-art LLMs—including GPT, Claude, and Gemini—and conduct systematic iterative experiments in both multiplayer and two-player settings, incorporating dynamic feedback mechanisms. Our key contribution is the first empirical demonstration that most LLMs (except Llama) dynamically infer opponents’ cognitive levels and adjust strategies accordingly: in multiplayer games, they submit significantly lower numbers—closer to the Nash equilibrium—indicating greater strategic depth and environmental adaptability. However, in two-player games, they consistently fail to reliably identify and execute the dominant strategy. These results suggest that while LLMs hold promise for simulating higher-order human economic reasoning, their strategic rationality exhibits structural limitations, particularly in contexts requiring precise dominance reasoning.
📝 Abstract
A beauty contest is a wide class of games of guessing the most popular strategy among other players. In particular, guessing a fraction of a mean of numbers chosen by all players is a classic behavioral experiment designed to test iterative reasoning patterns among various groups of people. The previous literature reveals that the level of sophistication of the opponents is an important factor affecting the outcome of the game. Smarter decision makers choose strategies that are closer to theoretical Nash equilibrium and demonstrate faster convergence to equilibrium in iterated contests with information revelation. We replicate a series of classic experiments by running virtual experiments with modern large language models (LLMs) who play against various groups of virtual players. We test how advanced the LLMs' behavior is compared to the behavior of human players. We show that LLMs typically take into account the opponents' level of sophistication and adapt by changing the strategy. In various settings, most LLMs (with the exception of Llama) are more sophisticated and play lower numbers compared to human players. Our results suggest that LLMs (except Llama) are rather successful in identifying the underlying strategic environment and adopting the strategies to the changing set of parameters of the game in the same way that human players do. All LLMs still fail to play dominant strategies in a two-player game. Our results contribute to the discussion on the accuracy of modeling human economic agents by artificial intelligence.