Standard language models are excellent at local prediction—choosing the next word based on the immediate context. However, as the target length increases from a few sentences to thousands of words, this "flat" autoregressive approach often fails. Models lose the thread of the plot, contradict themselves, or drift into aimless repetition.
To solve this, researchers have converged on a "Plan-then-Write" paradigm. By decoupling the high-level structural planning from the low-level prose generation, models can maintain global coherence across long horizons.
The Problem of Semantic Drift
When a model generates text word-by-word, it is essentially improvising without a script. In Hierarchical Neural Story Generation, Angela Fan and Mike Lewis observed that while models are fluent, they lack a "big picture" view. Without a global anchor, the probability distribution of the next word is easily swayed by recent tokens, leading to "semantic drift"—where the narrative gradually wanders away from its original premise.
This issue is particularly acute in structured formats. In classical Chinese poetry, Hua Wu and Haifeng Wang found that models often failed to maintain thematic consistency across lines. Their solution in Chinese Poetry Generation with Planning based Neural Network was to first generate a sequence of sub-topic keywords to serve as a roadmap, ensuring that each line remained semantically tied to the overall theme.
Decoupling the What from the How
The core intuition behind this research direction is that writing is a hierarchical process. Humans rarely write a novel or a technical paper by starting at page one and writing straight through. Instead, we outline, draft, and revise.
Research has formalized this into a multi-stage pipeline:
- Planning: Generating a structured skeleton (an outline, a list of sub-topics, or a set of character profiles).
- Realization: Expanding each section of the skeleton into fluent prose.
- Revision: Refining the text to ensure it matches the plan.
In Re3: Generating Longer Stories With Recursive Reprompting and Revision, Kevin Yang and Dan Klein demonstrated that mimicking this human workflow allows models to generate stories exceeding 2,000 words. By treating the initial plan as a persistent reference, the model can "self-correct" if a specific segment begins to drift.
Recursive Expansion and Detailed Control
A simple outline is often not enough for extremely long documents. If a 10-word outline is used to generate a 1,000-word chapter, the model still has too much room to wander.
The Detailed Outline Control (DOC) framework, developed by Kevin Yang and Nanyun Peng, addresses this through recursive expansion. In DOC: Improving Long Story Coherence With Detailed Outline Control, the system takes a brief premise and recursively expands it into a dense tree of plot points. This shifts the creative burden: the model stops being an "improviser" and becomes a "translator" of detailed instructions. By tracking character attributes and settings within this hierarchical tree, the system ensures that a character introduced in chapter one doesn't mysteriously change eye color in chapter ten.
Specialized Planning for Diverse Domains
This hierarchical approach is not limited to creative fiction; it is being applied across technical and scientific domains where logical flow is paramount:
- Scientific Writing: In Generating Related Work, Regina Barzilay and Darsh J. Shah used "motivation-cognizant" skeletons to plan literature reviews. Instead of just summarizing papers, the model plans why a paper is being cited (e.g., as a baseline or a related method) before writing the text.
- Technical Documentation: AutoPatent uses a multi-agent framework to generate 17,000-token patent applications by following a strict hierarchical blueprint of legal and technical requirements.
- Data Storytelling: The DATANARRATIVE framework decomposes data analysis into reflection, outlining, and narration phases. Omitting the planning phase in these complex tasks leads to a nearly 80% decrease in quality.
Multi-Agent Collaboration as the Next Frontier
The most recent evolution of this idea moves from a single model following a plan to multiple specialized agents collaborating on a "scratchpad."
In Agents' Room: Narrative Generation through Multi-step Collaboration, Fantine Huot and Mirella Lapata describe a system where "planning agents" establish character arcs and conflict while "writing agents" focus on prose. This division of labor prevents the model from prematurely concluding a story and ensures that the narrative adheres to established structural constraints like exposition and climax.
By treating long-form generation as a structured, multi-step engineering problem rather than a simple sequence-to-sequence task, these frameworks allow AI to move beyond short-form "chat" into the realm of complex, coherent, and sustained authorship.