It tries to! But if there's no good way to avoid them, then it will end up taking them anyway.

More specifically, when 'avoid repetition' is enabled (it is by default), then the routing algorithm has a very significant penalty weight applied to any existing path already taken.