To investigate whether trained flexibility would generalize to a novel discrimination task and to novel preferred and nonpreferred dimensions, 40 institutionalized retarded children and adults (IQ 50-77) were trained on five two-choice simultaneous-discriminations. Control subjects were trained with a preferred dimension relevant, the others (flexibility subjects), with a nonpreferred dimension relevant. Combinations were replicated factorially to assess transfer involving the same dimensions as those used in training, or transfer involving novel dimensions. Dimensions used were color, angular-orientation, form, and spatial-configuration. Flexibility acquired in the two-choice simultaneous-discriminations generalized to the matching-to-sample task and was independent of the stimulus dimensions used in training. At the beginning of training, subjects' discrimination learning performance was affected by dimension preferences; at the end of training, performance of flexibility-trained subjects was unaffected by dimension preferences. (Author/KW)