Proceedings of the Annual Meeting of the Cognitive Science Society

People are capable of learning with very little data. We argue that this is because people possess an symbolic inductive bias, consistent with the language of thought (LoT) hypothesis (Fodor & Pylyshn, 1988). To evaluate this claim, participants were asked to learn list functions by predicting how to transform an input list of numbers to an output list. We used participants' predictions on trials with no feedback to search for a congruent representation using a LoT model. The model was able to predict participants' heldout responses at above chance levels, even for participants who were unable to learn the function. Furthermore, LLMs and a neural network trained with a LoT inductive bias displayed similar patterns. These findings suggest that people and ML models display signatures of symbolic representation use to accomplish few-shot learning.