The Coordination Gap: Alternation Metrics for Temporal Dynamics in Multi-Agent Battle of the Exes

A new research paper titled 'The Coordination Gap: Alternation Metrics for Temporal Dynamics in Multi-Agent Battle of the Exes' exposes a critical measurement failure in how we evaluate AI coordination. Computer scientists Nikolaos Papadopoulos and Konstantinos Psannis from the University of Macedonia developed six novel Alternation (ALT) metrics that specifically measure temporal coordination—how well AI agents take turns and synchronize actions over time. Their work demonstrates that conventional metrics like efficiency and fairness ratios are 'temporally blind,' unable to distinguish between structured cooperation and chaotic or monopolistic behavior.

Using Q-learning agents in a multi-agent variant of the 'Battle of the Exes' game as a diagnostic testbed, the researchers made a startling discovery: despite AI agents achieving deceptively high scores on traditional metrics (often exceeding 0.9 fairness), they performed catastrophically poorly on the new ALT metrics. The learned policies scored up to 81% worse than completely random baselines when evaluated on temporal coordination quality. This 'coordination gap' worsens as more agents are added to the system, revealing that current AI training methods may be optimizing for the wrong objectives.

The findings have profound implications for deploying multi-agent AI systems in real-world applications like autonomous vehicles, robotic teams, or economic simulations. The paper argues that without temporally-aware metrics like ALT, we risk deploying AI systems that appear cooperative on paper but fail at basic coordination in practice. The researchers emphasize that random-policy baselines should become standard null tests for evaluating coordination, providing essential context for whether learned behavior represents genuine cooperation or just chance-level performance.