How Can Memory-Augmented Neural Networks Pass a False-Belief Task?

A question-answering system needs to be able to reason about unobserved causes in order to answer questions of the sort that people face in everyday conversations. Recent neural network models that incorporate explicit memory and attention mechanisms have taken steps towards this capability. However, these models have not been tested in scenarios for which reasoning about the unobservable mental states of other agents is necessary to answer a question. We propose a new set of tasks inspired by the well-known false-belief test to examine how a recent question-answering model performs in situations that require reasoning about latent mental states. We find that the model is only successful when the training and test data bear substantial similarity, as it memorizes how to answer specific questions and cannot reason about the causal relationship between actions and latent mental states. We introduce an extension to the model that explicitly simulates the mental representations of different participants in a reasoning task, and show that this capacity increases the model's performance on our theory of mind test.