This article investigates the conditions under which diagrammatic representations support collaborative argumentation-based learning in a computer environment. Thirty dyads of 15- to 18-year-old students participated in a writing task consisting of 3 phases. Students prepared by constructing a representation (text or diagram) individually. Then they discussed the topic and wrote a text in dyads. They consolidated their knowledge by revising their individual representation. There were 3 conditions: Students could use either (a) the individual texts they wrote, (b) the individual diagrams they constructed, or (c) a diagram that was constructed for them based on the text they wrote. Results showed that students who constructed a diagram themselves explored the topic more than students in the other conditions. We also found differences in the way collaborating dyads used their representations. Dyads who engaged in deep discussion used their representations as a basis for knowledge construction. In contrast, dyads who engaged in only shallow discussion used their representations solely to copy information to their collaborative text. We conclude that diagrammatic representations can improve collaborative learning, but only when they are used in a co-constructive way. (Contains 10 figures, 4 tables and 2 appendices.)