In my pretty untrained opinion, it feels like something that outputs vector images might well be trained using reinforcement learning, since it seems like the problem can be well modeled with a policy: "given what I want to produce and what I've produced so far, the next vector object I want to draw is x"