Abstract

In the late 1970s, the term theory of mind (ToM) was coined by psychologists studying the behavior of chimpanzees to describe the ability of individuals to attribute mental states to themselves and others. Soon after, inquiry also turned towards humans. In recent years there has been an enormous interest in studying ToM in machines (i.e., machine theory of mind)—especially in language models (LMs). Such inferences underlie many of LMs most studied behaviors: intention recognition, belief prediction, deception, emotion labeling, reading comprehension, and many more all rely on models that can generalize this reasoning to disparate contexts. I’ll talk about some approaches to machine ToM across three levels: modeling individual mental states (such as belief, hedging, and common ground); reasoning over socially shared or group beliefs (including politeness and face-management); and some attempts to find methods that generalize across ToM tasks.