Factive mindreading reflects the optimal use of limited cognitive resources

Predicting what other individuals will do is an important adaptive challenge for many organisms. Social prediction can be achieved by constructing a detailed model of the mental states of other agents, but this is computationally expensive. We argue that mindreaders can often bypass the need for constructing such a detailed model: they can keep track of the facts in their own world model that another agent also knows, instead of explicitly representing the content of the agent's world model. Using a simple computational approach, we find that this 'factive' mindreading strategy emerges as the optimal social prediction strategy for organisms with limited cognitive resources across a range of social ecologies. Factive mindreaders in our model behave like young human children and non-human primates: they successfully predict the behavior of knowledgable and ignorant agents, but fail to predict the behavior of agents with false and even accidentally true beliefs. Our results elucidate the computational principles underlying efficient social prediction, and provide a first-principles account for a range of empirical findings about human and non-human mindreading.