Can LLMs learn cooperation in multi-agent systems?
Cultural Evolution of Cooperation among LLM Agents
This research investigates how groups of LLMs learn to cooperate (or not) over time in a simulated social situation called the Donor Game. LLMs are given resources and decide how much to donate to another LLM, with the recipient getting double the donated amount. This process repeats over generations, with successful (wealthier) LLMs' strategies influencing the next generation.
Key findings for LLM-based multi-agent systems include: different LLMs exhibit vastly different cooperation abilities (Claude being best, GPT-4 worst), initial conditions matter significantly, the ability to punish defectors improves cooperation in some LLMs, and more complex cooperation strategies emerge over time. The research suggests a need for new benchmarks to test the long-term, multi-agent behavior of LLMs, especially regarding cooperation and its potential misuse (collusion).