I am wondering what do they mean by "works well with networks that have dynamic structures that change for every training instance".

Can somebody explain why it could be useful and point towards some use case?

I checked the only currently given example https://github.com/clab/dynet/blob/master/doc/tutorial.md and it seems to be just a very simple neural network.

Their papers indicate that their main interest is in textual models that is a bit different from the main focus of my interest.