There is limited experimental evidence concerning how best to train students to perform differential diagnosis. We compared 2 different methods for training 2nd-year medical students to perform differential diagnosis (DDX) of heart failure: a traditional classroom-based lecture (control group) versus a cognitive sciences-based approach to DDX instruction implemented through a computer-based tutor (treatment group). Following random assignment to either group, students were trained for 75 minutes, and then given a 40-item examination comprised of cases that varied along a typicality gradient from prototypical (easy) to less typical (hard). The treatment group diagnosed correctly significantly more test cases than the control group (74% versus 60%, respectively). The treatment group also diagnosed correctly significantly more cases at the extremes of the typicality gradient: 81% versus 65%, respectively, for the prototypical cases; 65% versus 48%, respectively, for the most difficult cases. The ability to perform differential diagnosis is enhanced by training based upon principles of cognitive sciences.