Although a growing literature has documented the effectiveness of informal group work during class sessions, virtually no data exist on which students are collaborating. As a result, instructors rarely know whether students are self-sorting in ways that maximize learning. This article explores which undergraduate students worked together on each of five exercises scheduled throughout the term, in a large-enrollment course for majors that emphasized intensive peer interaction. Pairwise logistic regression models were used to assess the likelihood that students collaborated based on shared demographic characteristics, socioeconomic status, and academic performance. In almost all cases, students self-sorted by ethnicity and gender. In addition, students who were predicted to do well in the course, based on their academic history, worked together initially; students who actually did well in the course, based on their final grade, were working together at the end; and students who were predicted to struggle in the course began collaborating late in the term.