Factive mindreading reflects the optimal use of limited cognitive resources

The capacity to represent the mental states of other individuals, known as 'mindreading' or 'theory of mind', is key to successful social prediction. We suggest that cognitive systems for mindreading are resource-rational: they are optimized for generating good predictions about the behaviour of other individuals, while not exceeding the computational capacity of the mindreader. We explore this hypothesis in a simple formal model where we derive cognitive strategies that excel at social prediction while minimizing cognitive effort. We find that it is often optimal for resource-limited mindreaders to keep track of the facts that another agent also knows, instead of explicitly representing the content of the agent's beliefs. When evaluated in mindreading tasks, simulated agents that use this 'factive' strategy tend to make mistakes in the same cases as non-human primates and young human children. Even agents that use more sophisticated strategies avoid representing beliefs unless necessary. Our results elucidate the computational principles underlying efficient social prediction and explain many of the successes and failures of human and non-human mindreading from first principles.