TL;DR: This paper introduces a framework for few-shot learning from demonstration by inferring the embodied reasoning process behind observed actions. On challenging reasoning tasks, our method achieves strong generalization to new tasks with only 1-3 demonstrations, moving closer towards human performance.