the loss of the path L->ab convergences much more slower than the loss of ab->L?

[qyxqyx]

In my reimplementation of the split brain experiment, I found the loss of the path L->ab convergences very slowly, but the loss of the path ab->L convergences quickly, is it normal? I used the regression loss in my reimplementation.