CoopEval: LLMs with Stronger Reasoning Consistently Defect in Social Dilemmas — First Benchmark for Cooperation Mechanisms

arXiv·medium signalXBlueskyLinkedInCopy link

CoopEval (arXiv 2604.15267) presents the first comprehensive benchmark showing that recent LLMs — with or without reasoning enabled — consistently defect in single-shot social dilemmas like prisoner's dilemma and public goods games. Critically, stronger reasoning capabilities make models LESS cooperative, not more. The benchmark evaluates cooperation-sustaining mechanisms to address this safety concern. For multi-agent system builders, this confirms that scaling reasoning doesn't produce cooperative behavior by default — explicit mechanism design is required.