Collaborative text reconstruction tasks such as dictogloss have been suggested as effective second language (L2) learning tasks that promote meaningful interaction between learners and their awareness of L2 target grammatical structures. However, it should be noted that the effect of pair interaction on the final product may differ depending on co-participant characteristics and particularly on proficiency disparities between partners. To date, most studies conducted on the effect of the different L2 proficiency of learners on paired performance have focused on the ways in which language learners interact, and the quantity and quality of language-related episodes (LREs) produced (Kim & McDonough, 2008; Leeser, 2004), often sidelining learners' actual task performance. This study thus aims to investigate the extent to which partner L2 proficiency levels affect tangible language performance, particularly in terms of content accuracy in a dictogloss task. Results show large gains in idea units reproduced between first and second stages of the dictogloss across texts. However, while low-level students paired with high-level partners benefited most, this group also had the largest variation across the board and, overall, proficiency pairing did not systematically affect improvement in idea units. Idea unit analyses indicated that students tended to perform better on idea units from earlier parts of the text, and that some types of idea units were more discriminatory than others.