Three experiments examined the perform of 4-year old children in matching geometric stimuli. Perform was developed as a simulation in which all components of the behavior were overt and directly measured. A correct match depended on the state of an instructional stimulus: the background color of the display. In the first two experiments, on nonidentity trials (signified by a green background), the next longer length, larger size or greater distance was correct. With a blue background a comparison identical to the sample was correct. In Experiment 3, red was added for which shorter, smaller or nearer was correct. Also here, on nonidentity trials, if a comparison of the correct length was not presented, the children adjusted their search target to the comparison of the next succeeding size (larger or smaller) so as to maintain a constant matching relation. Subsequently, when exposure to the instructional stimulus was reduced to a presentation only at the beginning of each trial, performance simulated matching based on instructions about abstract relations. In all experiments, accurate matching generalized across novel stimuli and across reduced exposure to the instructional stimuli.