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

ConWriter tackles the common issue of drifting consistency in long-form stories, where errors in temporality, facts, character traits, commonsense, and style accumulate. Unlike prompting-based methods that struggle as context grows, ConWriter uses a stateful, transition-constrained approach that incrementally generates scenes. It relies on static story requirements and dynamic narrative memory, combining symbolic constraints with neural generation. This neuro-symbolic consistency control is designed to be lightweight and training-free, making it easy to integrate with existing LLMs. The framework could improve AI-assisted creative writing tools by maintaining coherent storylines over extended passages, and the paper likely includes experiments demonstrating reduced error rates compared to baselines.