Why double softmax？

[Ericzay]

Hi，i wander why u use double softmax in ur paper? the probability pi is already the output of the softmax function as ur paper mensioned, if use a softmax on eht pi,
for instance, if i have a prob "tensor([0.1000, 0.8000, 0.1000])", the output of double softmax is tensor([0.2491, 0.2491, 0.5017]), apparently the high pi 0.6 becomes 0.4272. This appears to lower the confident.
so why use double softmax ?Please help me understand this design,Thanks!!!