The Computational Representation of Outlines

[AdamSobieski]

Introduction

Due to the growing popularity and utility of large language models, these systems being capable of creating outlines, generating documents from outlines, reverse outlining from documents, and co-authoring documents using outlines, the matter of computationally representing outlines should be of increasing interest to the field of artificial intelligence, as will be shown below.

Automatic Outline Writing

The STORM project creates encyclopedic articles in two stages. First, during a pre-writing stage, the system collects reference materials and generates outlines. Second, during a writing stage, the system generates full-length articles from these outlines (Shao, Jiang, Kanell, Xu, Khattab, and Lam, 2024).

Also, with respect to creating outlines, Liu, Shen, Zheng, Ma, Wu, Zhang, and Lu (2025) propose a "long-form outline generation framework named LOGIC that is based on continuous refinement through feedback, enabling the iterative refinement of outlines to make them more content-rich, logical, and coherent according to the feedback from the critical model."

Generating Documents from Outlines

As indicated, above, the STORM project's second stage, the writing stage, generates full-length encyclopedic articles from outlines (Shao, Jiang, Kanell, Xu, Khattab, and Lam, 2024). Outline-guided natural-language generation is also explored by Lee, Ka, Son, Kang, and Kang (2025).

Reverse Outlining from Documents

Reverse outlines are outlines made from existing works. Reverse-outlining is like reverse-engineering a document. Zhang, Guo, Fan, Lan, and Cheng (2019) explore the challenge of computationally obtaining outlines for documents.

Co-authoring Documents using Outlines

Dang, Benharrak, Lehmann, and Buschek (2022) show that automatic summarization can enhance computer-aided writing.

Zhang, Du, Liu, Chen, Huang, Ye, Liut, and Aspuru-Guzik (2026) introduce a hierarchical writing system which enables users to view and edit their documents in two complementary views, a tree view and linear view, while providing integrated contextual AI support.

Knowledge Representation

As shown in #14 and #15, outlines can be resources within larger sets of interrelated resources. Outlines can be used to express what AI systems should say when generating a document and related resources can express why, how, as who, and to whom to say it. As considered, these ingredients – perhaps requiring multiple formats to represent them – could be usefully combined for software components to use to create prompts or agentic workflows.

Whether outlines are created by end-users, AI systems, or are reverse-outlined from documents, AI systems can use them to generate documents. Modifying or varying such outlines can result in the generation of appropriately different documents. Modifying or varying the other ingredients indicated, above, can result in the generation of appropriately different documents as well. As envisioned, modifying or varying the to whom ingredient would enable adaptation and personalization.

Conclusion

How should outlines (and any related resources) be computationally represented for these important use cases?

Bibliography

Dang, Hai, Karim Benharrak, Florian Lehmann, and Daniel Buschek. "Beyond text generation: Supporting writers with continuous automatic text summaries." In Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology, pp. 1-13. 2022.

Lee, Yukyung, Soonwon Ka, Bokyung Son, Pilsung Kang, and Jaewook Kang. "Navigating the path of writing: Outline-guided text generation with large language models." In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track), pp. 233-250. 2025.

Liu, Hengwei, Yongliang Shen, Zhe Zheng, Haoyuan Ma, Xingyu Wu, Yin Zhang, Weiming Lu. "LOGIC: Long-form outline generation via imitative and critical self-refinement." Findings of the Association for Computational Linguistics: EMNLP 2025, pp. 18119–18144. 2025.

Shao, Yijia, Yucheng Jiang, Theodore Kanell, Peter Xu, Omar Khattab, and Monica Lam. "Assisting in writing Wikipedia-like articles from scratch with large language models." In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 6252-6278. 2024.

Zhang, Ruqing, Jiafeng Guo, Yixing Fan, Yanyan Lan, and Xueqi Cheng. "Outline generation: Understanding the inherent content structure of documents." In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 745-754. 2019.

Zhang, Zijian, Fangshi Du, Xingjian Liu, Pan Chen, Oliver Huang, Runlong Ye, Michael Liut, and Alán Aspuru-Guzik. "TreeWriter: AI-assisted hierarchical planning and writing for long-form documents." arXiv preprint arXiv:2601.12740 (2026).

[AdamSobieski]

For discussion, the following approach starts with nodes, <node>, which could provide metadata in their optional <head> elements, content in their optional <body> elements, and have child nodes in their optional <children> elements, recursively.

<node>
  <head>...</head>
  <body>...</body>
  <children>
    ...
  </children>
</node>

<node>
  <head>
    <metadata rel="stylesheet" type="text/css" href="style.css" />
    <metadata rel="objective">The objective of this section is to explain photosynthesis.</metadata>
    <metadata rel="precondition">...</metadata>
    <metadata rel="effect">...</metadata>
    <metadata rel="guiding-question-answer-pair" type="...">...</metadata>
    <metadata rel="reading-question-answer-pair" type="...">...</metadata>
    <metadata rel="comprehension-question-answer-pair" type="...">...</metadata>
    <metadata rel="content-evaluation-criterion">...</metadata>
    <permission name="adaptation-permission" value="granted" />
  </head>
  <children>
    ...
    <node class="x">
      <head>
        <metadata rel="objective">The objective of this section is to explain water photolysis.</metadata>
      </head>
      <body>
        Linear electron transport through a photosystem will leave the reaction center of that photosystem oxidized.
      </body>
    </node>
    ...
    <node class="y">
      <head>
        <metadata rel="objective">The objective of this section is to explain the Calvin cycle.</metadata>
      </head>
      <body type="application/xhtml+xml" xmlns:xhtml="http://www.w3.org/1999/xhtml">
        <xhtml:p><xhtml:span class="term essential">Carbon fixation</xhtml:span> produces the three-carbon sugar
        intermediate, which is then converted into the final carbohydrate products.</xhtml:p>
      </body>
    </node>
    ...
  </children>
</node>

Manually reading larger-sized outlines in this format may become strenuous. The indicated format, however, is intended for use by software components. As envisioned, outliner software would be utilized by end-users to manually create, view, and modify outlines in this format.

Remaining tasks include exploring: interoperability with other formats, uses in multipart MIME messages or within compressed archives, a metadata vocabulary, styling for software components performing tasks other than layout and presentation, terminology, citations and bibliographies, data and tables, multimedia, and overall expressiveness.