Research in category learning has been dominated by a ‘reference point’ view in which items are classified based on attention-weighted similarity to reference points (e.g., prototypes, exemplars, clusters) in a multidimensional space. Although much work has attempted to distinguish between particular types of reference point models, they share a core design principle that items will be classified as belonging to the category of the most proximal reference point(s). In this paper, we present an original experiment challenging this distance assumption. After classification training on a modified XOR category structure, we find that many learners generalize their category knowledge to novel exemplars in a manner that violates the distance assumption. This pattern of performance reveals a fundamental limitation in the reference point framework and suggests that stimulus generalization is not a reliable foundation for explaining human category learning.