Akata, E., Schulz, L., Coda-Forno, J. et al. Playing repeated games with large language models. Nat Hum Behav (2025). https://doi.org/10.1038/s41562-025-02172-y
This research investigates how Large Language Models (LLMs) perform in strategic interactions using behavioral game theory, specifically examining cooperation and coordination in repeated 2x2 games like the Prisoner's Dilemma and the Battle of the Sexes. The study found that LLMs, particularly GPT-4, excel in self-interested games where defection can be a dominant strategy, showing an unforgiving response to betrayal. However, they struggle significantly with coordination games requiring turn-taking or varied responses, indicating limitations in understanding and adapting to other agents' strategies. Interventions like "social chain-of-thought" prompting, where LLMs predict the opponent's move before acting, improved coordination and led human participants to perceive LLMs as more human-like, suggesting that explicit prompting can enhance LLMs' social reasoning and interaction capabilities.
Share this post