A new paper posted on arXiv challenges a common assumption in multi-agent systems: that a coordinator's theory of mind—its model of what teammates believe or plan—should be certified by how accurate those inferences are. The authors argue that accuracy is the wrong certificate for this task.

The abstract lays out two reasons, the first being that the accuracy needed to act depends on the context. In other words, a fixed accuracy threshold cannot meaningfully guarantee that a coordinator will make good decisions, because the same level of inference error may be harmless in one setting and costly in another.

The paper stops short of offering a full alternative in the abstract, but the argument points toward a need for more action-aware or decision-focused certification methods for multi-agent coordination.