This is a general question. I am trying to understand the reasoning behind sequence_length = 1 during prediction of a trained RNN model. If you trained your sequence classifier e.g. some language model using sequence_length = 20, why can’t you use same sequence length for testing it during prediction? Setting a larger sequence length leads to lower prediction accuracy and I am struggling to find the reason why. It is also quite common to set batch_size = 1 during validation but that is somewhat understandable since we want higher error due to validation dataset being generally quite small. Setting larger sequence length on the other hard would allow us to be faster during prediction. Can somebody elaborate on this?