Sparse attention papers commonly embellish experimental results by applying unequal hyperparameter tuning (such as window size and block size) to baselines versus their own methods.