Recursive Re-prompting and Revision: Enhancing Long-Range Coherence in Story Generation

Generating short stories has been a popular area of research in natural language generation. However, these short stories often fall short of the length and complexity found in human-authored narratives. To bridge this gap, we propose Re3, a framework that generates plot-coherent stories with a length of 2000–2500 words. Compared to stories directly generated from the same base model, Re3’s stories were judged by human evaluators to have a more coherent overarching plot (14% absolute increase) and greater relevance to the initial premise (20%).

The Challenge of Long-Range Coherence

Creating long, coherent stories requires a comprehensive understanding of language, world knowledge, and common sense. While recent advancements in large pretrained language models have shown impressive capabilities, long-form generation presents unique challenges. The system must maintain plot coherence over thousands of words, ensuring that the story remains relevant to the initial premise. Additionally, preserving the narration style and avoiding factual contradictions become crucial over a longer narrative horizon.

The Recursive Reprompting and Revision Framework (Re3)

Motivated by the process followed by human writers, the Re3 framework leverages a recursive reprompting and revision approach to generate longer stories. It consists of the following components:

1. Plan Module

Re3’s Plan module generates a detailed plan for the story. By prompting a general-purpose language model, such as GPT3, with a given premise, setting, characters, and outline, the system constructs a structured overarching plan.

2. Draft Module

The Draft module generates story passages by recursively reprompting the language model using a strategically crafted prompt. The prompt is dynamically reconstructed at each step, selectively incorporating contextually relevant information from the initial plan and the story written thus far. This recursive reprompting technique extends the concept of chain-of-thought prompting.

3. Rewrite Module

To ensure plot coherence and premise relevance, the Rewrite module reranks different continuations generated by the Draft module. This step emulates a full rewrite, allowing the system to select the most suitable continuation for the story.

4. Edit Module

The Edit module focuses on maintaining factual consistency within the story. It performs smaller local edits to align the selected continuation with the previous passages, enhancing the overall coherence of the narrative.

Evaluating Re3’s Performance

To evaluate the effectiveness of Re3 in generating longer stories, we compared its output with two GPT3-based “rolling-window” baselines of similar length. Human evaluators participated in pairwise comparisons and rated Re3’s stories. The results were impressive, with Re3 outperforming the baselines in terms of coherence in the overarching plot (up to 14% absolute increase) and relevance to the initial premise (up to 20%). Remarkably, evaluators predicted that up to 83% of stories generated by Re3 were authored by humans.

Conclusion

The Recursive Reprompting and Revision framework (Re3) offers a promising solution to the challenge of generating long, coherent stories. By combining recursive reprompting with strategic revision techniques, Re3 significantly improves plot coherence and premise relevance in longer story generation. The evaluation results demonstrate the effectiveness of Re3 in producing narratives that approach the quality of human-authored stories.