Two studies examined the effect of category variability on revision of stereotypical and non-stereotypical beliefs. In Experiment 1, participants learned the characteristics of various groups via exposure to numerical distributions of category attributes. These distributions had the same mean values, but either high or low levels of variability around the mean. The categories were either stereotyped categories or non-stereotyped categories. After learning category attributes, participants were exposed to disconfirming exemplars (drawn from a sample with higher or lower means). Participants were then asked to re-estimate the central tendency of the category and to rate the likelihood that the category had changed. Beliefs about categories with low variability were more influenced by disconfirming information than beliefs about highly variable categories. This relationship was the same across category domains. This same general pattern was found in Experiment 2 where participants had to request information about test instances before deciding whether category values had changed. The results are consistent with Bayesian models of belief revision.