Is social coherence necessary for functional Theory of Mind in AI agents?

Investigate whether maintaining social coherence—defined as consistent representations of self, others, and communicative context over time—is a necessary substrate for functional Theory of Mind in artificial systems such as language-model–powered agents with tool use and multi-party interactions.

Background

The authors document repeated failures of social coherence in deployed agents, including misattribution of authority, inconsistent tracking of what others know, and contradictions between reported and actual actions. They propose viewing these as failures of social coherence rather than isolated model shortcomings.

They note that several behaviors could be interpreted as Theory of Mind deficits and explicitly raise the question of whether social coherence is a prerequisite for functional Theory of Mind in artificial systems.