This study investigates the effectiveness of immediately modelling the correct solution to a task on which children were making errors. The technique is based on proposals by Saxton (1997) who, in his contrast theory of negative input, claims that corrective speech input is particularly effective when it immediately follows a child's error, such as an overgeneralisation of a verb ending. Our study concerns a very different domain, that of children learning to balance beams on a fulcrum, but one in which children also tend to overgeneralise a particular strategy. On a pre-test on the balance beam task, we identified a number of children (N = 79, mean age = 74.82 months) who were making errors. These children were randomly assigned to two groups and either (a) watched the correct solution being modelled by an adult, or (b) saw the correct solution being modelled by an adult immediately after their own error. The latter, contrast modelling, condition produced a significantly higher number of children who had improved at the task at post-test. The implications of these findings for general models of development are discussed. (C) 2002 Elsevier Science Inc. All rights reserved.