From the reference of accuracy score, we know that
In multilabel classification, this function computes subset accuracy: the set of labels predicted for a sample must exactly match the corresponding set of labels in y_true.
As I tested and also regarded the PB score, currently accuracy score is identical with the exact match ratio, and I guess it is a small mistake.
The original intent was to use a strict measure of accuracy, meaning a prediction is counted as correct when it exactly matches the ground truth labels for all the classes. However the sklearn accuracy_score actually computes the exact match ratio, as it computes the accuracy for all labels separately and then takes the mean across each.
We will discuss internally, and re-assess if we update the accuracy metric, or leave it as is for this round.