local vs global attention in further MLM pre-training.

I am currently looking into the [conversion script](https://colab.research.google.com/github/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) to create a longformer from an already pre-trained Roberta implementation. I noticed that the script further pre-trains the Longformer on MLM with only local attention.

- What exactly is the motive behind not having tokens (e.g. bos and eos) attend globally as well?
- Would it make sense to alter this if the goal you have in mind is focussed on tasks that require embeddings of the entire sequence? (e.g. unsupervised document similarity)