Comment by kraddypatties

I believe that’s _part_ of the point (or at least a side-effect) of the KL divergence loss term they have on the AV. That and training stability.