Spontaneous transfer of learning is often difficult to elicit. This finding may be widespread partly because pretests proactively interfere with transfer. To test this hypothesis, 7-year-olds' transfer was examined across 2 numerical tasks (number line estimation and categorization) in which similar representational changes have been observed. As predicted, children given feedback on numerical estimates learned to use a linear representation of numerical quantity instead of a logarithmic one, but providing practice on a categorization pretest led children to continue using a logarithmic representation on the same task, which they otherwise abandoned with surprising frequency. These findings imply unsupervised practice of inappropriate representations impedes transfer, and studies of learning can greatly underestimate children's potential for transfer if pretest effects are uncontrolled.