Collaborative learning has often been associated with the construction of a shared understanding of the situation at hand. The psycholinguistics mechanisms at work while establishing common grounds are the object of scientific controversy. We postulate that collaborative tasks require some level of mutual modelling, i.e. that each partner needs some model of what the other partners know/want/intend at a given time. We use the term “some model” to stress the fact that this model is not necessarily detailed or complete, but that we acquire some representations of the persons we interact with. The question we address is: Does the quality of the partner model depend upon the modeler’s ability to represent his or her partner? Upon the modelee’s ability to make his state clear to the modeler? Or rather, upon the quality of their interactions? We address this question by comparing the respective accuracies of the models built by different team members. We report on 5 experiments on collaborative problem solving or collaborative learning that vary in terms of tasks (how important it is to build an accurate model) and settings (how difficult it is to build an accurate model). In 4 studies, the accuracy of the model that A built about B was correlated with the accuracy of the model that B built about A, which seems to imply that the quality of interactions matters more than individual abilities when building mutual models. However, these findings do not rule out the fact that individual abilities also contribute to the quality of modelling process.