Strong LLM reasoning doesn't guarantee cooperation in multi-agent settings, but game-theoretic mechanisms like contracts and third-party mediation can reliably restore cooperative behavior—important for safe AI deployment.
This paper tests whether AI language models can cooperate with other agents in game theory scenarios like prisoner's dilemma. It finds that stronger LLMs actually defect more, then evaluates four mechanisms—repeated games, reputation systems, mediators, and contracts—to encourage cooperation.