Initial random assignment is actually somewhat smart about the references you provided.
In each iteration, the posterior assignment probabilities for each particle against all references are computed. In the first 5 (by default) iterations, assignment is randomized only within sets of particles that have their max probability amongst a set of repeated identical references. So if you provide 3 identical references, all particles will be randomized for the first 5 iters. If you provide 3 distinct references, particles will not be randomized (though for the first 5 iterations, classification will be binary rather than soft). If you provide 2 identical and 1 distinct reference, particles that fall into the first 2 identical classes will be randomized within those 2 classes for the first 5 iters.
Hope that helps!