The present study tests the effect of ability pairing in two instructional methods in L2 collaborative revision. Two continuous indices determine a pair: individual proficiency level, distance in proficiency between pair members (heterogeneity), and the interaction between both indices. Instructional methods tested are modelling and practising. Results show that the effect of pair composition depends on instructional strategies. In the Practising condition less proficient learners profit most from a heterogeneous ability pair, whereas more proficient learners are best paired homogeneously. In the Modelling condition no effect of pair composition factors was observed. This result illustrates that Modelling is a powerful instructional method for complex learning tasks like collaborative revision in L2 as it overrides some of the grouping effects which can be found in more traditional learning conditions.