The evidence concerning the level at which learned CP effects occur is complex. The goal of this study was to use a different approach to this question by manipulating the abstractness of the information available for distinguishing pairs of items in an XAB task, and the presence or absence of a short task-filled delay between X and AB. Participants engaged in XAB trials containing a mixture of trials with and without the delay task before and after standard training to classify visual texture stimuli into two categories. Training improved discrimination of pairs differing on the category-relevant dimension whether within- or between-category, but not on pairs differing only on non-category relevant low level features. In addition, only successful learners in the post-training trials avoided decreased discrimination accuracy due to the delay task, suggesting that they formed more stable representations. However, this effect was not limited to pairs varying in category-relevant ways.