Theory of Mind in Prisoner’s Dilemma with Small LLMs

In this work, we host a tournament of games of iterative prisoner’s dilemma between LLMs and classic prisoner’s dilemma strategies, as well as employ Theory of Mind (ToM) prompting. While previousworks have focused primarily on the performance of large models, highlighting the capabilities of GPT4 in particular, we focus our investigation on smaller, cost-effective models and whether they demonstrate emergent social reasoning. Our results indicate that for the LLaMA and Falcon families, including ToM can cause cooperative behavior to significantly decrease, while the Qwen family tends to remain trusting of their opponents, despite the detriment to its performance and its accuracy in predicting its opponents next move.