Indeed, all the the EEG channels, including the references, should have their type set to "EEG".
Otherwise, the forward model is not estimated for the references, and the average reference is not properly applied.
In general, the type "EEG REF" should not been used, except for some file formats that have the reference information duplicated (in that case, it should be already defined as "EEG REF" automatically when loading the file in the Brainstorm database).
@gubei2008 Can you please confirm that after setting the type of M1 and M2 to "EEG", you don't need to apply the average reference explicitly before estimating the sources?
(so that future readers of this thread do not try to duplicate their data if they don't need to)
Thanks