Interleaved Content Planning for Natural Language Generation

29 concepts19 papers

When humans write a story or an essay, they rarely formulate a rigid, word-for-word blueprint of the entire document before putting pen to paper. Instead, authorship is an inherently recursive process: we outline a general direction, write a sentence, evaluate how that sentence changes the context, and adjust our remaining plan accordingly.

Traditional natural language generation (NLG) systems, however, have long relied on a strict "plan-then-write" pipeline. In these architectures, a content plan or story outline is generated entirely upfront, and the surface realizer is tasked with translating that static plan into text.

While a static plan seems logical, it consistently fails in practice. As a model generates text, small variations accumulate—a phenomenon known as probabilistic drift or logical drift. If a model deviates even slightly from its initial blueprint, the remaining pre-baked plan loses relevance, leading to structural collapse, repetitive text, or logical contradictions.

To solve this, a powerful paradigm has emerged: interleaved content planning. Instead of treating planning and generation as sequential phases, this approach weaves them into a continuous feedback loop, dynamically updating the document's plan based on the text that has actually been produced.


The Failure of Static Blueprints

To understand why interleaved planning works, we must first look at why static planning breaks down. In long-form text generation, a static plan acts like an open-loop controller: it issues commands without checking the actual state of the output.

As Qianyue Wang et al. point out in their work on long-form storytelling, fixed outlines cannot account for the uncertainty inherent in creative text generation. When a language model writes a paragraph, its output is probabilistic. If the generated text drifts away from the initial outline, the remaining plan becomes an anchor, forcing the model to make jarring transitions to satisfy a blueprint that no longer matches the established context.

This issue manifests across multiple domains:

  • Logical Drift: In story generation, Chen Tang et al. demonstrate that relying on static, upfront event plans limits a model's ability to react to its own narrative choices, causing the plot to lose logical consistency over time.
  • Stiff Transitions: In creative writing, Xiaoyuan Yi et al. show that pre-defined keyword plans result in unnatural thematic shifts because the system cannot dynamically choose when to introduce a concept based on the evolving semantic flow.
  • Syntactic Constraints: Even in non-narrative text, Charles B. Callaway established that as text complexity grows, sentence revisions and merging require real-time updates to the underlying rhetorical structure. A static discourse plan cannot survive the fluid constraints of surface realization.

Early natural language generation frameworks attempted to use hierarchical classification structures to map out text. However, as Ehud Reiter et al. observed, while static classification works well for initial content selection, it is fundamentally unsuited for managing the dynamic, iterative flow of a narrative.


The Interleaved Feedback Loop

Interleaved planning transforms generation into a closed-loop system. The core mechanism is straightforward: the system alternates between generating a localized planning unit and executing that unit into text. Crucially, the next planning step is conditioned on the history of both the previous plans and the actual text generated so far.

┌────────────────────────────────────────────────────────┐
│                                                        │
│               ┌───► Local Planning Step ───┐           │
│               │                            ▼           │
│     Current Context                Text Generation     │
│               ▲                            │           │
│               └──── Actual Written Text ───┘           │
│                                                        │
└────────────────────────────────────────────────────────┘

By feeding the generated text back into the planner, the model can absorb the variance of probabilistic generation. If the model writes a sentence that introduces an unexpected nuance, the next planning step acknowledges this new reality and adjusts the trajectory of the document.

Ratish Puduppully et al. operationalized this in their SeqPlan framework for data-to-text generation. By treating document structure as a sequence of latent variables, their model conditions each new planning step on the specific linguistic context already established. This ensures that subsequent topics follow with strict local and global coherence.

A similar intuition drives the work of Peng Xu et al. in MEGATRON-CNTRL, which demonstrates that incremental, sentence-by-sentence planning maintains a significantly tighter link to the evolving context than any global plan generated at the start.


Granularity and Representation

How do these systems represent an intermediate "plan"? The representation must be abstract enough to guide generation but flexible enough to be generated iteratively.

Question-Answer Blueprints

One compelling approach is to use human-like inquiry as a planning skeleton. Danyang Liu et al. introduced a method for visual storytelling that interleaves the creation of question-answer pairs with text segments. The model first decides what question to ask about an image sequence, answers it, and then generates the corresponding narrative text. This per-sentence interleaving ensures the story remains tightly grounded and naturally paced.

Fine-Grained Evidence Selection

In knowledge-intensive tasks, interleaving helps maintain factual accuracy. The Retrieve-Plan-Generation (RPG) framework, proposed by Yuanjie Lyu et al., uses an iterative loop to combat focus shift in long-form generation. By planning the next fine-grained topic immediately before writing it, the model can select and retrieve only the most relevant snippets of information for that specific segment, preventing the dilution of context that occurs with global retrieval.

Hierarchical Outlining

For massive generation tasks like novels, planning cannot happen purely at the sentence level; it requires a macro-perspective. The DOME framework by Qianyue Wang et al. solves this by combining macro-level plot outlining with dynamic, segment-by-segment expansion. The system maintains a high-level outline but expands it into detailed sub-plans on the fly, allowing the narrative structure to adapt to the uncertainty of creative generation.


Serendipity and Agency in Interactive Environments

The ultimate test of an interleaved planner is an environment it cannot control: an interactive narrative or a live simulation. When a human user or player enters the loop, static scripts break entirely.

As Marc Cavazza et al. note, moving beyond rigid, pre-written branching paths toward truly adaptive narratives has been a long-standing challenge in narrative intelligence. In virtual environments, pre-scripted sequences fail because they cannot account for unpredictable player actions (Arnav Jhala et al.).

When players deviate from an authored path, the system requires real-time narrative mediation to rewrite story structures on the fly (Sasha Azad et al.). This requires highly flexible planning algorithms, such as partial order planning, that can dynamically adapt to unanticipated user choices without breaking the underlying plot logic (Xun Zhang).

In modern large language model (LLM) applications, this manifests as a need for dynamic script adaptation. Hongqiu Wu et al. demonstrate that by incorporating plot-based reflection mechanisms, agents can dynamically modify their dramatic goals in response to player behavior, preserving user agency and immersion.

This interactive flexibility mirrors the cognitive process of human authors. Decades ago, Natalie Dehn argued in her "AUTHOR" model that compelling generation requires a system to simulate an author's mind rather than a rigid world state. A key component of this is a serendipity engine—a dynamic planning mechanism that allows the generator to intentionally deviate from its own high-level goals when an unexpected, more compelling narrative opportunity emerges during the writing process.

Summary

Interleaved content planning works because it treats text generation as an evolving, interactive environment where the model's own past outputs serve as the primary context. By breaking global planning down into an iterative cycle of local prediction, execution, and reflection, these systems absorb probabilistic drift, maintain tight contextual coherence, and mimic the recursive, adaptive nature of human authorship.

Key Foundational Work

Story Generation After TALE-SPIN (1981) — contributed 2 concept elements to this direction.
A Computational Theory of Prose Style for Natural Language Generation (1985) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_sentence_planning_9368eb70,
  title = {dynamic sentence planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-sentence-planning/rdr_9368eb70bb7c4f039e3f053c4d060b62},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_sentence_planning_9368eb70_v1,
  title = {dynamic sentence planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-sentence-planning/rdr_9368eb70bb7c4f039e3f053c4d060b62/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}

See something wrong? Report a mistake