The use of abstract higher-level knowledge (overhypotheses) allows humans to learn quickly from sparse data, and make predictions in new situations. Previous research has suggested that humans may be the only species capable of abstract knowledge formation, but this remains controversial, and there is also mixed evidence for when this ability emerges over human development. Kemp et al. (2007) proposed a computational model of overhypothesis formation from sparse data. We provide the first direct test of this model: an ecologically valid paradigm for testing two species, capuchin monkeys (Sapajus spp.) and 4-5-year-old human children. We compared performance to predictions made by models with and without the capacity to learn overhypotheses. Children’s choices were consistent with the overhypothesis model predictions, whereas monkeys performed at chance level.