Generalization via Analogy in Young Children and Large Language Models - A Comparative Study

Watch a 48-minute lecture from UC Berkeley's Alison Gopnik at the Simons Institute exploring how young children use relational and analogical reasoning to generalize from small samples and predict novel visual transformations. Discover empirical studies comparing toddlers' and preschoolers' abilities to understand size changes, number increases, rotations, and reflections through experiments with "change machines" and other transformative scenarios. Learn about a new benchmark developed for testing visual analogical reasoning in large multimodal models, consisting of 1,400 visual transformations of everyday objects. Examine the comparative results showing how models like GPT-4V, LLaVA-1.5, and MANTIS perform well at identifying changes but struggle with quantifying and extrapolating rules to new objects, while children and adults demonstrate stronger analogical reasoning capabilities across all evaluation stages. The research, conducted in collaboration with Eunice Yiu and Mariel Goddu, provides insights into the "Unknown Futures of Generalization" and the current limitations of artificial intelligence compared to human cognitive development.