Its the other way around too, HLE questions were selected adversarially to reduce the scores. I'd guess even if the questions were never released, and new training data was introduced, the scores would improve.