High coherence among training exemplars promotes broad generalization of face families.

How do we tailor learning experiences to promote the formation and generalization of conceptual knowledge? Exposing learners to a highly variable set of examples has been postulated to benefit generalization, but evidence is conflicting. In the present study, we manipulated training set variability in terms of both the typicality of training examples (high vs. low coherence) and the number of unique examples (small vs. large set size) while controlling the total number of training trials. The face family category structure was designed to allow participants to learn by picking up on shared features across category members and/or by attending to unique features of individual category members. We found relatively little effect of set size but a clear benefit of high-coherence (lower variability) training both in terms of category learning and generalization. Moreover, high-coherence training biased participants to make judgments based on shared features in both categorization and recognition. Using an exploratory model fitting procedure, we tested the hypothesis that high-coherence training facilitates prototype abstraction. Instead, we found an exemplar model advantage across training conditions. However, there was also systematic misfit for all models for some trial types, including underestimating the influence of shared features in categorization responses. Overall, we show that high-variability training is not necessarily beneficial for concept learning when the total length of training is controlled. Instead, training on typical examples promotes fast learning and broad category knowledge by helping learners extract shared category features. (PsycInfo Database Record (c) 2025 APA, all rights reserved).