Recent work on reading suggests variability in irrelevant elements benefits the learning of sound/spelling correspondences(Apfelbaum et al, 2013). However, under some conditions similarity helps, perhaps depending on the order of items duringtraining (Roembke et al., submitted). To investigate this in the laboratory, we trained adults to map abstract four-symbolstrings onto three-finger manual responses. As in reading, there were one-to-one mappings (”consonants”, where onesymbol indicates a specific finger) and two-to-one mappings (”digraph vowels” like AI where two symbols map to onefinger). Participants (N=15/condition) were trained on variable or similar consonant sets, and with vowels either blockedor interleaved. We found a similarity benefit for interleaved but not blocked training. However, for generalization, therewas a variability benefit. Surprisingly, a simple backpropagation model showed both patternsincluding the blocking effect.This suggests that blocking effectstypically thought to invoke explicit strategiesmay derive from associative principles.