Re3: Generating Longer Stories With Recursive Reprompting and Revision (2210.06774v3)

Abstract: We consider the problem of automatically generating longer stories of over two thousand words. Compared to prior work on shorter stories, long-range plot coherence and relevance are more central challenges here. We propose the Recursive Reprompting and Revision framework (Re3) to address these challenges by (a) prompting a general-purpose LLM to construct a structured overarching plan, and (b) generating story passages by repeatedly injecting contextual information from both the plan and current story state into a LLM prompt. We then revise by (c) reranking different continuations for plot coherence and premise relevance, and finally (d) editing the best continuation for factual consistency. Compared to similar-length stories generated directly from the same base model, human evaluators judged substantially more of Re3's stories as having a coherent overarching plot (by 14% absolute increase), and relevant to the given initial premise (by 20%).

• The paper introduces a novel Re³ framework that divides story generation into Plan, Draft, Rewrite, and Edit modules to enhance narrative coherence.
• The method employs recursive reprompting to build on initial outlines, achieving a 14% improvement in coherence and a 20% increase in premise relevance.
• The approach reduces reliance on domain-specific data and produces AI-generated stories that 83.3% of human evaluators rated as nearly human-written.

Recursive Reprompting and Revision Framework for Long-Form Story Generation

The paper "Re³: Generating Longer Stories With Recursive Reprompting and Revision" addresses the challenge of generating long-form stories using AI, specifically focusing on maintaining plot coherence and relevance over texts exceeding two thousand words. The authors introduce the Recursive Reprompting and Revision framework (Re³), an innovative approach that leverages a structured plan-inference methodology to enhance narrative generation.

Method Overview

Re³ divides the story generation process into four primary modules, inspired by human writing processes: Plan, Draft, Rewrite, and Edit. This systematic division allows for dynamic interaction between modules, significantly improving narrative continuity and contextual relevance.

1. Plan Module: This module begins by using prompts to generate a structured outline, character descriptions, and setting details. The plan serves as a critical backbone, enabling the system to maintain a coherent narrative thread throughout the story's development.
2. Draft Module: The draft module performs what the authors term "recursive reprompting"—continually feeding parts of the draft, together with elements from the initial plan, back into the model. This process ensures that each generation step builds on previously established narrative elements, thus preserving coherence.
3. Rewrite Module: Utilizing coherence and relevance rerankers, this component selects the most contextually aligned continuations from multiple generated candidates. The rerankers are trained on datasets such as the WritingPrompts to recognize narrative continuity and thematic adherence.
4. Edit Module: Focusing on factual consistency, this module detects and corrects character-related inconsistencies using a precision-focused information extraction pipeline. This sophisticated editing mechanism allows the model to refine story passages without wholesale rewrites.

Numerical Results and Observations

In comparative evaluations against two baseline models—both variations of GPT3-based rolling window architectures—the Re³ framework excels. Human evaluators found Re³-generated stories significantly more coherent (14% absolute increase) and relevant to the initial premise (20% increase). Furthermore, up to 83.3% of stories were perceived as written by humans, underscoring the model's enhanced capacity to emulate human writing nuances.

Implications and Future Developments

Re³'s framework presents substantial implications for AI-assisted creative writing. The recursive reprompting and structured revision strategies represent a paradigmatic shift towards more coherent and engaging long-form narrative generation. This methodology not only enhances text generation capabilities but also reduces reliance on domain-specific datasets, opening avenues for more generalized applications in storytelling and content synthesis.

Future research could focus on scaling the approach to even more substantial narrative constructs, such as novels, or refining modules to handle thematic consistency and multi-threaded plots. Furthermore, integrating adaptive pipelines for dynamic outline expansion could facilitate even longer story continuations, potentially setting a new standard for AI narrative production systems.

In summary, Re³ introduces a robust framework specifically tailored for long-form AI narrative generation. The integration of planning, recursive feedback, and multi-layered revision culminates in a system capable of producing intricate and coherent stories. This technology advances the frontier of AI's role in creative writing, contributing meaningfully to both theoretical understanding and practical applications in the field.