New Framework for Book-Scale AI Creative Writing

A new arXiv preprint (2605.17064) presents a training framework for book-scale creative writing using large language models. The authors argue that current instruction-following models are misaligned with fiction writing, which requires deception, moral ambiguity, and unreliable narration. They propose a dataset construction method that reframes supervised fine-tuning as prompt-to-book generation, using public-domain novels to create a multi-resolution planning scaffold. The scaffold summarizes books from high-level premise to chapter and scene structure, then inverts this hierarchy during training. The framework aims to produce stylistically rich, human-like fiction at book length.