Several sequential-sampling models using racing diffusion processes for multiple-alternative decisions were evaluated, using data from two perceptual discrimination experiments. The structures of the models differed on a number of dimensions, including whether there was lateral inhibition between accumulators, whether there was decay in evidence, whether evidence could be negative, and whether there was variability in starting points. Data were collected from a letter discrimination task in which stimulus difficulty and probability of the response alternatives were varied along with number of response alternatives. Model-fitting results ruled out a large number of model classes in favor of a smaller number of specific models, most of which showed a moderate to high degree of mimicking. The best-fitting models had zero to moderate values of decay, had no inhibition, and assumed that the addition of alternatives affected the subprocesses contributing to the nondecisional time, the degree of caution, or the quality of evidence extracted from stimuli.