Beyond Sally-Anne: Evaluating Theory of Mind in LLMs Using Epistemic Schelling Points

The paper introduces the Epistemic Asymmetry Schelling Task (EAST), a two-player dialogue game designed to test whether LLMs can use Theory of Mind for coordination rather than reproduce familiar static test patterns. Model dyads must independently converge on semantic Schelling points under different levels of epistemic transparency. According to the abstract, only frontier models handle the varying epistemic demands reliably. Reasoning-trace analysis attributes many failures to epistemic tracking errors, especially confusing private knowledge with mutual knowledge. The work positions robust social reasoning as a distinct bottleneck despite strong traditional benchmark performance.