In Experiment 1, "more" and "less" relations were trained for arbitrary Stimuli A1 and A2 with 3 children with autism. The following conditional discriminations were then trained: A1-B1, A2-B2, B1-C1, B2-C2. In subsequent tests, participants showed derived more-less mands (mand with C1 for more and mand with C2 for less). A training procedure reversed the B-C conditional discriminations, and participants then showed derived reversed more-less mands (mand with C1 for less, C2 for more). Baseline B-C relations were retrained, and participants subsequently demonstrated a return to the original derived manding. A second experiment with 1 prior participant and 1 naive participant removed a possible confounding effect. Establishing derived manding may be an advantageous component when teaching a mand repertoire in applied settings. (Contains 8 figures.)