Scaling Long-Form Story Generation via Narrative State Tracking

TL;DR

The paper presents NstAgent, a training-free framework that enables large language models to maintain narrative consistency in long-form stories by tracking structured narrative states. It evaluates story generation from 10K to 100K words using extended benchmarks, showing that both consistency and quality remain stable or improve with length. This approach demonstrates a viable path toward generating full-length novels with LLMs.

Why It Matters

Long-form story generation is a major bottleneck in creative AI applications, where consistency degrades rapidly with length. NstAgent offers a practical, scalable solution without requiring retraining, enabling more reliable and coherent novel-length content creation. This advances the feasibility of using LLMs for professional or entertainment-grade storytelling.

Key Findings

• NstAgent maintains narrative consistency across stories up to 100K words without degradation.
• Writing quality remains high and does not decline as story length increases.
• The framework operates without training, making it adaptable to various LLMs and use cases.