More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration

A new research paper titled 'More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration' reveals a surprising finding about AI cooperation. Researchers Advait Yadav, Sid Black, and Oliver Sourbut built a multi-agent system designed to study cooperative behavior in frictionless environments where helping others carries no personal cost. They discovered that more capable models don't necessarily cooperate better—OpenAI's advanced o3 model achieved only 17% of optimal collective performance, while the smaller o3-mini reached 50%, despite identical instructions to maximize group revenue.

Through causal decomposition that automated one side of agent communication, the researchers separated cooperation failures from competence failures, tracing their origins through agent reasoning analysis. The study tested targeted interventions and found that explicit protocols doubled performance for low-competence models, while tiny sharing incentives improved models with weak cooperation. These findings suggest that scaling intelligence alone won't solve coordination problems in multi-agent systems, requiring deliberate cooperative design even when helping others costs nothing.

The research, accepted at ICLR 2026 Workshop on Agents in the Wild, has significant implications for real-world applications where AI agents must coordinate, from knowledge sharing in organizations to code documentation. The 24-page study with 5 figures demonstrates that without specific cooperative design, even the most capable AI models may fail at basic coordination tasks that should be trivial in zero-cost scenarios.