Does the multi-task training include predicting relevant paragraphs?

Hi,
The page mention in section 6.1
> We use a two-stage model that first selects the most relevant paragraphs then passes them to a second stage for answer extraction. Both stages concatenate question and context into one sequence, run it through Longformer, then use task-specific prediction layers. We train the models in a multi-task way to predict relevant paragraphs, evidence sentences, answer spans and question types (yes/no/span) jointly.

My understanding is that the first stage concatenates question and the whole context as input to Longformer, and only predict relevant paragraphs as a binary classification task for each paragraph. The second stage concatenates question and the predict relevant paragraphs as input, train another Longformer model in a multi-task way to predict **relevant paragraphs,** evidence sentences, answer spans and question types (yes/no/span) jointly.

Does the multi-task training in stage2 still need to include predicting relevant paragraphs?

Thank you