60 kindergarten and 60 first grade children were trained on an oddity problem with perceptual (P) or conceptual (C) attribute for a fixed number of trials (30). After mastering the training problem, the subjects were randomly assigned to either an intradimensional (CC or PP group) or an extra-dimensional (CP or PC group) oddity shift. The transfer performance showed CC > PC > PP = CP group. These differences were attributed to the pattern hypotheses sampled at the start of transfer. Subjects who had acquired a conceptual oddity rule in the training had a strong tendency to sample hypotheses by the prior rule, but subjects under conditions where a perceptual-oddity rule was relevant tended to make a shift to a cue from new (conceptual oddity) rule. The older children yielded more of such a sampling mode than the younger children did, and the former were easily able to abandon a dis-confirmed rule and sample hypotheses from the relevant rule domain.