ConWriter: Transition-Constrained Stateful Long-Form Story Generation with Lightweight Neuro-Symbolic Consistency Control

The recently released ConWriter from arXiv is a training-free long-story generation framework that improves narrative quality through lightweight neural symbolic consistency control techniques. This framework performs incremental writing at the scene level, combining static story requirements, dynamic narrative memory, and symbolic state reasoning to implement consistency control and local repairs using uncertainty risk signals before error propagation. Evaluation shows that ConWriter outperforms direct generation and recent training-free baseline DOME in four long-story tasks of ConStory-Bench for various base large language models. The related code is now open-source.