mentalizing

This paper investigates the emergence of situation modeling and mentalizing abilities in transformer language models across training stages, finding that false belief task performance depends on model size and training volume, emerges late in pretraining, and shows fragility with non-factive verbs.

The paper introduces the Theory of Mind Utility (ToM-U), a formal computational-level specification for inferring others' epistemic states by constructing Local Epistemic World Models (LEWMs). It differs from Bayesian ToM and simulation theory by providing a domain-agnostic mechanism for belief inference without commitment to algorithmic implementation.