We're not "selecting" sentences as an extractive summarizer might. The sentences are generated.

As for how does the model deal with co-reference? There's no special logic for that.