Symmetric Machine Theory of Mind

Theory of mind (ToM), the ability to understand others' thoughts and desires, is a cornerstone of human intelligence. Because of this, a number of previous works have attempted to measure the ability of machines to develop a theory of mind, with one agent attempting to understand anothers' internal "mental state''. However, ToM agents are often tested as passive observers or in tasks with specific predefined roles, such as speaker-listener scenarios. In this work, we propose to model machine theory of mind in a more flexible and symmetric scenario; a multi-agent environment SymmToM where all agents can speak, listen, see other agents, and move freely through a grid world. An effective strategy to solve SymmToM requires developing theory of mind to maximize each agent's rewards. We show that multi-agent deep-reinforcement learning models that model the mental states of other agents achieve significant performance improvements over agents with no such ToM model. At the same time, our best agents fail to achieve performance comparable to agents with access to the gold-standard mental state of other agents, demonstrating that the modeling of theory of mind in multi-agent scenarios is very much an open challenge.