Hi @huqi, hopefully this is still relevant in your project:
If you find two classes like this, it generally means there is a population of particles where A is more rigid/well ordered and a population where B is more rigid/well ordered.
It is definitely a good idea to refine the separately and would be considered common practice to include both maps in a publication and deposition to EMDB.
If you want to combine the particles (which might be useful in some cases to boost SNR) you can, as you suggested, use a local masked refinement. The hope is that if you mask only A or only B, and use all the particles, the particles where either subunit is “floppy” will align better when only considering A or B respectively. There is some chance that this would yield improved reconstruction for either subunit. However, there is also the possibility that in either set of particles, A or B is actually disordered, and this would not help in yielding better reconstruction.