Llms keep story details straight when writing very long novels
Scaling Long-Form Story Generation via Narrative State Tracking
Computation and Language
Summary
Writing very long stories like full novels is hard for language models because they often forget important details. The authors present a method called Narrative State Tracking Agent (NstAgent) that helps these models keep track of characters and events as they write. This method improves the consistency and quality of stories even when they get much longer, up to 100,000 words. It does this without needing extra training. Their tests show that stories stay coherent and well-written as they get longer with this approach.
What this means in practice
- •For creative writing software developers: Integrate narrative state tracking to improve consistency in AI tools generating long stories or novels.$Commercial implications: Enables selling AI writing assistants that handle full-length novels with fewer contradictions and higher quality.
- •For game narrative designers: Maintain complex storylines and character details over extended interactive narratives in games.
Authors
Zhennan Wan, Jianfei Chen
Abstract
LLMs have demonstrated strong capabilities in creative writing. However, scaling them to full-length novels remains challenging, as maintaining narrative consistency becomes increasingly difficult. Existing story-generation methods typically focus on stories of up to about ten thousand words, leaving their ability to scale to full-length novels underexplored. In this work, we introduce Narrative State Tracking Agent (NstAgent), a training-free agentic framework that allows LLMs to track a structured narrative state including characters, past events and future requirements. We extend an existing benchmark to compare narrative consistency across lengths, and use it together with a writing-quality benchmark to systematically evaluate stories ranging from 10K to 100K words. We show that NstAgent achieves better narrative consistency and writing quality as stories grow longer, and neither of them degrades noticeably as length increases, suggesting that it provides an effective approach to scaling story generation toward full-length novels.