a research study was carried out to evaluate the effect of an observational training with and without descriptions on an extradimensional transfer test. two domains were used, one for training and other for testing counterbalancing their order of presentation. second-order matching-to-sample arrays were built using paintings as instances considering one of two domains: by style (how the painting is represented) or by theme (what it is represented on the painting). during the transfer test session, participants had to build a matching-to-sample array using the same instances but considering the untrained domain. performance on test was evaluated qualitatively (reconstructing real time performance of participants) and quantitatively (estimating the number of arrays that were congruent with the new matching criterion). participants who had to describe the matching criteria during training had a better performance on the extradimensional transfer test than those who did not explicitly describe matching criteria. building matching-to-sample arrays using a new matching criterion seems to be a categorical adjustment task useful when studying referential substitutive behavior.