Initially, a mentally retarded subject demonstrated, by matching-to-sample, appropriate classification of common stimuli (i.e., cars, bowls, hats, dolls, shoes, and books). Next he was taught to produce the correct verbal label (e.g., ''bowl'') for one member of each class. For four of the six stimulus classes, a single training example was not sufficient to produce generalization of the label to nontraining examples in the same class. However, each successive addition of one training example generally produced a systematic increase in the level of generalization. The specific number of training examples sufficient to produce generalization varied across the six experimental classes. In addition, receptive labelling probes, conducted after each successful demonstration of generalization in the productive modality, demonstrated no systematic changes in the accuracy of responses as productive training progressed. It was concluded that automatic generalization within or across response modalities is not necessarily an inevitable result, and therefore may require explicit programming.