In general, it's helpful for me to separate assumptions that are characteristics of the generative model (and possibly an inference procedure used with it), and "assumptions" meaning things that could lead to poor out of sample prediction.