Does bucketing generally degrade text classification accuracy?

Question

I have a number of binary text classifiers (RNN). Recently, I decided to implement bucketing to reduce the amount of padding sent through the system. To my amazement, I obse