[4J3-GS-5] Agents:

[4J3-GS-5-01] Reproduction of Human Decision Making by LLM in the Ultimatum Games

〇Toshiya Murashige¹, Takayuki Ito¹ (1. Kyoto University)

Keywords:Decision making and cosensus building , Game theory, Simulation, Behavioral economics

Large Language Models (LLMs) have gained attention as alternatives to human subjects in experimental psychology and behavioral economics research. LLMs can perform various language processing tasks based on pre-trained data and can be assigned specific behaviors through prompts. This study investigated LLM simulation potential through the ultimatum game and its variants. Using AutoGen, we conducted four experiments: the ultimatum game, dictator game, anonymous dictator game, and ultimatum game with different proposer settings (program vs. agent). We analyzed both the decisions and reasoning of each agent. Results showed that LLMs could reproduce many human behavioral patterns, including the decrease in proposed amounts in the dictator game and increased acceptance rates when the proposer was identified as a program. However, in the anonymous dictator game, results varied by persona, and agents rarely referenced anonymity in their reasoning, revealing some differences from human behavior. This study suggests that LLMs could be valuable for experiments with ethical constraints and as preliminary tests for large-scale experiments.