Extreme Equilibria: The Benefits of Correlation

A new theoretical economics paper by Kirill Rudov, Fedor Sandomirskiy, and Leeat Yariv delivers a striking result: for any Nash equilibrium involving three or more randomizing agents, there almost always exists a correlated equilibrium that strictly improves the outcome for the agents under a wide range of objectives. The authors prove this using a 'detail-free' criterion that does not require specific payoff structures, making the result broadly applicable. They also provide constructive methods to identify such improved equilibria, focusing on Pareto improvements and utilitarian welfare gains. The paper builds on the concept of correlated equilibria, where agents can condition their actions on a common signal (e.g., from a recommendation system), as opposed to playing independently.

The implications extend beyond economics to computer science and multi-agent AI systems. In contexts such as automated negotiation, traffic routing, or decentralized machine learning, where agents often rely on shared recommendations, this result suggests that randomly mixing independent strategies leaves significant welfare on the table. The authors show that even simple correlated strategies can yield better outcomes for all participants, and they outline how to compute them. Published on arXiv on April 29, 2026, the work is a significant advance in understanding when and how correlation can improve upon Nash equilibria, with potential applications in mechanism design, AI alignment, and any system where multiple decision-makers coordinate via a central signal.