Assigning category labels to examples varying along a continuous dimension exaggerates perceived differences between members on opposite sides of category boundaries. Using social categories, we investigated how contrast may guide representation of not only features that differentiate between groups but also features that are neither diagnostic nor correlated across trained examples. In a classification task, participants learned which residence hall to correctly assign students varying along three psychological traits (academic, athletic, social). The same target category was learned alongside one of two contrast categories with either higher or lower values along one diagnostic dimension. After learning, participants provided estimates of average trait values for each dorm. Predicted contrast effects were found along diagnostic dimensions but contrast also influenced memory for non-diagnostic and uncorrelated dimensions, presumably based on assumptions about general co-occurrence of features. These findings have implications for how stereotypes are learned and applied and how illusory correlations are perpetuated.