Leading models of human categorization posit that an observed stimulus is classified according to its similarity to storedreference points. In the present study, we investigate a category structure that elicits human generalization behavior notpredicted by the reference point framework. In a supervised classification learning task, participants were presented withsimple continuous-valued stimuli (one- or two-dimensional) based on an underlying category structure with a strict patternof alternating regions assigned to each class (e.g., A A B B A A B B ? ?). The participants were then tested on newstimuli with dimension values beyond the range seen in training. A large portion of participants classified new items byextrapolating the alternation sequence they did not classify based on similarity to the nearby reference points. Theseresults pose a challenge to reference point models and raise important issues about concept formation and generalization.