Instruct-StoryMix

Instruct-SkillMix (https://arxiv.org/abs/2408.14774) shows that we could use LLMs in order to decompose tasks into skills and then use those skills to construct synthetic data that would allow us to actually make good instruction tuning data. And I wonder can we do the same thing with stories? Can we have models describe components of stories and then use that to construct synthetic data and then somehow make a model that can actually control its own story generation better. The key problem here of course being that models are naturally very very bad at writing stories but I still wonder if there's some kind of arbitrage there.