Research 12 min read • Published October 2026

From Re3 to DOC: The Science of Detailed Outline Control

How Kevin Yang, Dan Klein, Nanyun Peng, and Yuandong Tian proved that long-form AI coherence is a planning problem, not a writing problem.

In late 2022 and 2023, researchers from UC Berkeley, UCLA, and Meta AI published a sequence of breakthrough papers that fundamentally transformed automated storytelling: first Re3 (Recursive Reprompting and Revision), followed by DOC (Detailed Outline Control, alphaXiv:2212.10077).

1. The Limitation of Re3

Re3 demonstrated that long stories could not be generated in one shot. It decomposed story generation into Planning, Drafting, Rewriting, and Reranking. While a massive leap over baseline GPT models, Re3 suffered from a critical weakness: high-level outlines are too vague to prevent narrative drift.

If the outline simply says: "Chapter 3: The hero travels across the mountains", the drafting LLM has millions of degrees of freedom. It often introduces random side-quests, encounters unnecessary strangers, or solves the main dilemma prematurely, leaving the next chapter stranded with nowhere to go.

2. The Breakthrough: Detailed Outline Control (DOC)

The core insight of Yang et al. in DOC is profound:

"The creative burden of plot continuity should NOT be placed on the generation step. It must be resolved recursively in the hierarchical planning stage."

DOC introduced two complementary components:

  1. Detailed Outliner: Recursively expands macro beats into fine-grained chapter contracts specifying dramatic goals, entry conditions, exit conditions, and active character rosters.
  2. Detailed Controller: Verifies that drafted narrative passages strictly adhere to the fine-grained beat contract before committing them.

3. The Empirical Results

In rigorous human blind evaluations comparing stories generated by DOC against the Re3 baseline, DOC achieved astonishing improvements:

+22.5%

Plot Coherence Gain

+28.2%

Outline Relevance Gain

+20.7%

Reader Engagement Gain

4. How MythMaker Implements DOC

MythMaker's 10-Chapter Novel Engine is an industrial implementation of DOC principles. Before writing Chapter 1, Phase 1 creates a complete 3-Act / 10-Beat Matrix where every chapter has a locked Entry State, Exit State, and Dramatic Goal. Pass A writes strictly inside these boundaries, eliminating aimless filler.

MythMaker AI Engineering Team
Architecting narrative state machines and multi-chapter literary engines for EmeraldSoft.
View Research Methodology