True solution to the problem stated in the article? They're talking about fine-grained control over model outputs, how would multimodality help there?