The learning of rule-plus-exception categories relies on pattern integration and differentiation, but how the representations of rule-followers and exceptions develop through these two operations remains obscure. Here, we inspected the representational shifts in rule-plus-exception category learning by fitting a computational model to behavioral categorization data. We found that exceptions were differentiated from rule-followers within and between categories through learning. The distanced rule-follower and exception representations in each category formed distinct clusters that together constituted a hierarchically structured categorical representation. Moreover, exception learning increased the representational overlap between rule-followers of opposite categories, thereby blurring the category boundary. Our findings illuminate the representational dynamic underlying the acquisition of rule-plus-exception categories and highlight the roles of pattern integration and differentiation in category learning.