Individuals with autism spectrum disorder (ASD) often have difficulties with multiple‐component direction following, which is likely due to the complexity of the auditory‐visual discrimination skills required to perform the instruction. In the present study, matrix training was used to establish multicomponent direction following in preschoolers with ASD. We also compared the acquisition of this skill across two different‐sized matrices (i.e., 3 × 3 and 5 × 5) to identify any possible advantages for using one size matrix over the other. Given the evidence of the efficacy of matrix training to promote recombinative generalization, we also attempted to extend this literature by assessing novel matrix generalization with a 5 × 5 matrix of verb–noun combinations when an extra verb was added (i.e., verb–noun–verb). A multiple‐probe across‐participants design was used to compare the acquisition of both trained and untrained target emergence. Results showed that the 5 × 5 matrix produced a more efficient acquisition rate for recombinative generalization targets across participants as compared to the 3 × 3 matrix. Recommendations for future research along with possible limitations are discussed.