Human word concept learning leverages prior knowledge to generalize from limited exemplars, yet its neural implementation remains unclear. We developed a Neural Bayesian Model (NBM) that incorporates neural representational priors to explain word concept learning. Using fMRI, we first constructed a neural prior space from activity elicited by familiar objects (with novel shapes as controls), and then examined neural responses during novel word concept learning based on these objects. The NBM integrating priors from ventral occipitotemporal cortex predicted both neural representations and behavioral generalization, outperforming control models lacking neural priors. In contrast, hippocampal activity supported learning for novel shapes without benefit from the NBM, consistent with prior-free associative mechanisms. Large language models showed weaker alignment with human generalization patterns. These findings dissociate prior-based cortical inference from hippocampal exemplar-associative learning, providing a neural instantiation of Bayesian concept acquisition and clarifying the interplay between semantic and episodic memory systems. Zhang et al. show that prior knowledge stored in visual cortex supports word concept learning through Bayesian inference, while the hippocampus enables prior-free association. Large language models fail to capture this human-like learning pattern.
Zhang et al. (Wed,) studied this question.