This study contrasted a rule-instantiation model and a similarity-to- exemplars model on their predictions of typicality judgments and speeded classifications for members of logically-defined categories. In Experiment 1, subjects learned a unidimensional rule based on the size of objects. It was assumed that items that maximally instantiated the rule were those furthest from the category boundary separating small and large stimuli. In Experiment 2, subjects learned a disjunctive rule of the form "x or y or both." It was assumed that items that maximally instantiated the rule were those with both positive values (x and y). In both experiments, the frequency with which different exemplars were presented during classification learning was manipulated across conditions. These frequency manipulations exerted a major impact on subjects' postacquisition goodness-of-example judgments, and also influenced reaction times in a speeded classification task. Furthermore, the goodness judgments and reaction times were predicted better by principles of exemplar-based generalization than by the degree to which the logical rules were instantiated. It was concluded that even for logically-defined concepts, stored exemplars may form a major component of the category representation.