We are not entirely sure, but our guess is as follows.
The model may be able to learn a switch that changes its behavior depending on the masking ratio, but such a switch would only resolve the conflict between the two tasks at the level of the MLM objective, i.e., masked token prediction. On the other hand, downstream task performance, such as GLUE and BEIR, is not directly tied to the MLM objective. It is rather a byproduct of the representations formed during pretraining.
Therefore, even if a switch emerges for MLM prediction, the conflict at the representation level may still remain independently of that switch.
