Purpose: The benefit of stimulus variability for generalization of acquired skills and knowledge has been shown in motor, perceptual, and language learning but has rarely been studied in reading. We studied the effect of variable training in a novel language on reading trained and untrained words. Method: Sixty typical adults received 2 sessions of training in reading an artificial script. Participants were assigned to 1 of 3 groups: a variable training group practicing a large set of 24 words, and 2 nonvariable training groups practicing a smaller set of 12 words, with twice the number of repetitions per word. Results: Variable training resulted in higher accuracy for both trained and untrained items composed of the same graphemes, compared to the nonvariable training. Moreover, performance on untrained items was correlated with phonemic awareness only for the nonvariable training groups. Conclusions: High stimulus variability increases the reliance on small unit decoding in adults reading in a novel script, which is beneficial for both familiar and novel words. These results show that the statistical properties of the input during reading acquisition influence the type of acquired knowledge and have theoretical and practical implications for planning efficient reading instruction methods.