Why shared mistakes can change whether teamwork helps

When a coordinator acts on a teammate’s inferred beliefs, a wrong guess can make the team worse off than if it had acted without that guess. But is accuracy enough to decide whether to trust the module?

The answer matters because mistakes can have different costs and can cluster in ways that either overlap or remove backup.

The Blind-Spot Gate estimates whether acting on inferred beliefs helps by combining error costs, pairwise interactions and measured dependence.

The authors ask whether accuracy can tell a coordinator when to act on inferred beliefs. Their theory says the answer also depends on what wrong actions cost and whether errors overlap or remove teammates’ backup.

They test this with task-assignment classifiers, Hanabi modules and language-model copies, then compare the Blind-Spot Gate with an accuracy rule on held-out episodes.

The gate accounts for error costs and dependence; its held-out results show fewer significant false approvals in the tested cells, while out-of-sample tests find lower regret on classifiers and fewer false-approval splits for a Bayesian module.

These results cover the tested task-assignment and Hanabi settings. The gate omits interactions among three or more wrong beliefs, requires a simulator that accepts injected beliefs, and assumes a module’s errors are unrelated to the episode state that sets their cost.