Theoretical and empirical work in the field of classification learning is centered on a ‘reference point’ view, where learners are thought to represent categories in terms of stored points in psychological space (e.g., prototypes, exemplars, clusters). Reference point representations fully specify how regions of psychological space are associated with class labels, but they do not contain information about how features relate to one another (within- class or otherwise). We present a novel experiment suggesting human learners acquire knowledge of within-class feature correlations and use this knowledge during generalization. Our methods conform strictly to the traditional artificial classification learning paradigm, and our results cannot be explained by any prominent reference point model (i.e., GCM, ALCOVE). An alternative to the reference point framework (DIVA) provides a strong account of the observed performance. We additionally describe preliminary work on a novel discriminative clustering model that also explains our results.