Proceedings of the Annual Meeting of the Cognitive Science Society

What do "signal," "heat," "contact," and "director" have in common? Terms in electrical engineering? Things needed on a movie set? Appreciating what these concepts have in common requires coming up with a situation that provides a coherent grouping for them. These "ad-hoc'' categories are constructed on the spot, but what makes some better than others? Humans and large language models (LLMs) generated categories (e.g. "Things with corners'') for randomly grouped sets of nouns (e.g., "eye," "road," "table," "paper") and rated how difficult it was to link them. We then asked additional participants to rate applicability and specificity of these categories. LLMs produced more specific and applicable categories and the model and human performance gap increased for less semantically similar categories. Models, without human cognitive constraints, may be able to generate more candidates before selecting the best one, while people settle on what's "good enough."