Paper | Playing Repeated Games With Large Language Models

While doing preparatory research for my work on Agents in Game Theory, I came across this paper and thought it was really interesting. This paper aims to perform a sociological analysis of LLMs by having LLMs play games. I am particularly interested in games of the "prisoner’s dilemma” family, which I aim to use for my research as well. This paper explores several examples in that space that I found interesting.

Goal Develop an understanding of how Large Language Models (LLMs) interact with humans and other agents in the context of behavioral games with clearly defined rules.

Problem Is it correct to assume that current generations of LLMs are benevolent assistants to humans? What makes LLM chatbots better, and more pleasant, interactive partners? And I think these are important questions to ask as we are including genAI in our social and workplaces.

Solution The research team aims to solve this by having LLMs play repeated games against each other. The games are played in finite rounds with full information. Then they analyze how the LLMs “behave” when playing against each other.

Leave a Comment