I can't really comment on variants, but if this is only happening for one particular model/family of models, then maybe it doesn't capture the relevant dynamics well without the stability of 1-o upwind.