Solving the Baby Intuitions Benchmark with a Hierarchically Bayesian Theory of Mind

TL;DR

This paper introduces a hierarchical Bayesian Theory of Mind model that effectively solves the Baby Intuitions Benchmark, outperforming deep learning methods and providing interpretable, human-like social reasoning.

Contribution

The paper presents a novel hierarchically Bayesian approach to Theory of Mind that enables few-shot learning and high accuracy on social reasoning tasks.

Findings

Achieves near-perfect accuracy on most benchmark tasks.

Outperforms deep learning and imitation learning baselines.

Produces interpretable, human-like inferences.

Abstract

To facilitate the development of new models to bridge the gap between machine and human social intelligence, the recently proposed Baby Intuitions Benchmark (arXiv:2102.11938) provides a suite of tasks designed to evaluate commonsense reasoning about agents' goals and actions that even young infants exhibit. Here we present a principled Bayesian solution to this benchmark, based on a hierarchically Bayesian Theory of Mind (HBToM). By including hierarchical priors on agent goals and dispositions, inference over our HBToM model enables few-shot learning of the efficiency and preferences of an agent, which can then be used in commonsense plausibility judgements about subsequent agent behavior. This approach achieves near-perfect accuracy on most benchmark tasks, outperforming deep learning and imitation learning baselines while producing interpretable human-like inferences, demonstrating the advantages of structured Bayesian models of human social cognition.