Review

This file reads the literature on showing a nested tree of text to a reader. Strands run from existing reviews to the seminal work, then one strand per sub-question, then practice.

Terms are in frame; markings and scores, in the evidence README.

0. Existing reviews

• Cockburn 2009 (read in full): reviews overview plus detail, zooming, focus plus context and cue-based interfaces.
  • Its summary (section 8): none of the four is ideal, and research "fails to provide clear guidelines" on which fits which task.
  • It holds one reading study with comprehension measured, a web-browser study, a document-navigation study and a few on program text and trees; strands 2 and 4 cite them through it.
• Chen 1996 (abstract, secondary): a meta-analysis of 23 hypertext experiments, 56 effect sizes. Hypertext users were more effective and less efficient; maps of the organisation raised usefulness; results varied strongly.
• Dillon 1998 (abstract): hypermedia helps learning only for tasks of repeated searching and manipulation; method flaws limit every conclusion.
• Brusilovsky 2001 (read in the relevant sections): a survey of adaptive hypermedia. It lists stretchtext, text that expands in place, among the canned-text adaptation technologies, beside inserting, altering and dimming fragments.
• DeStefano 2007 is held by learning-science (strand 6) and not read again.
• Located by title only: Knutov 2009 (adaptive hypermedia survey), Lam 2011 (multi-level interfaces from study evidence).
• How this review differs: text nested to any depth, not maps or images; it adds layered documents, outliners and model readers (position.md).

1. Seminal work

• Stretchtext is Nelson's term, coined in 1970 (Wikipedia 2026c; Nelson's own text not read).
• Furnas 1986 (read): the fisheye view shows a point when its a priori importance minus its distance from the focus passes a threshold.
  • On a tree, importance is minus the depth and distance is the path length; the view keeps the ancestral line, then siblings, then cousins (Figure 2).
  • In a regular tree the view is a logarithmically compressed display, computed in time proportional to the view, not the tree (section 3).
  • In its experiment, 20 subjects judged the order of two parts of an unfamiliar botanical hierarchy, 16 pairs each (pages 20 to 21).
  • Correct answers: 52% with two flat views, 64% with one fisheye and one flat, 75% with two fisheyes.
  • A footnote links fisheye views of outlines to the outline processors then reaching the market.
• Early systems for a global and a detailed view: Engelbart 1968 (cited so by Furnas 1986), Pad++ for zooming (Bederson 1994), the document lens for continuous distortion (Robertson 1993).
• None of these three was read beyond its title and those citations.
• Carroll 1984 (abstract, secondary): blocking error-prone function for new users gave faster learning; the control group spent almost a quarter of its time recovering from errors.
• Nielsen 2006b names deferring advanced or rare features to a second level progressive disclosure, and dates the idea to over 30 years before 2006.

2. Q1: folded and expandable text

Source	What was compared	Result	Read
Furnas 1986	flat window against fisheye of a hierarchy	52% correct with two flat views, 64% with one of each, 75% with two fisheyes, on order judgments; 20 subjects, 16 pairs each	full text
Egan 1989	SuperBook, a hypertext book whose contents expands chapters and sections (Chimera 1994), against print	better search and open-book essays; MiteyBook, its PC version, searched about 25% faster and 25% more accurately than print	abstract (secondary)
Chimera 1994	contents of 1,296 items, three levels deep, 48 screens when fully expanded: stable list, expand/contract, multi-pane	both dynamic interfaces significantly faster for many of nine tasks (41 novices); on the easiest task, the stable list and multi-pane beat expand/contract (p < .01)	technical report version
Hornbæk 2003	linear text, overview plus detail, fisheye text keeping headings and first and last paragraphs	essays graded best with overview plus detail, 2.47 against 2.00 linear and 1.95 fisheye, F(2, 32) = 4.16, p < .05 (p. 140); fisheye essays fastest, 37.4 min against 44.4 and 44.5 (pp. 141 to 142), with the lowest incidental learning, 3.42 against 4.20 and 4.58; answers to questions did not differ; all but one of 20 subjects preferred overview plus detail	full text, after Cockburn 2009 section 7.2.4
Höök 1997	adaptive stretchtext that hides what the inferred task does not need, against the same system fixed	within-page actions about halved, navigation between pages about the same (Figure 3, p. 183); 19 of 27 answers chose an item the system had opened; preferred, mean 5 on a scale of 1 to 7 (Table 1, p. 184); 9 subjects, no statistical test	full text
McDonald 2009	layered privacy policy (short layer, full text behind) against full text and a standard format, 749 participants	layered was faster but less accurate on topics not in the short layer; cookie question 78% correct layered	authors' version, section 4
Heinonen 2022	maintenance instructions in DITA layered by skill level, the reader choosing the depth; user studies in three countries	well received across skill levels; observation, interviews and questionnaires, no comparison condition	abstract (secondary)

• McDonald 2009 concludes that few participants went beyond the first page of a layered policy (section 4.1).
• Springer 2020 (abstract, secondary): in four studies of an explained system, users gained from simple feedback first and the detail on demand; it is about explanations, not documents.
• Located, not read: Zellweger 2000 (fluid documents, an observational study), Boyle 1994 (MetaDoc, adaptive stretchtext).
• Johnson 2013 (a blog, 13 of 20, scored in information-typing): one DITA topic per contents entry scatters brief tasks; he would nest them under a parent and favours collapsible sections; no data.

3. Q2: hierarchical against networked hypertext

The linked reviews hold the overview and map effects: reading-order (strand 1) and learning-science (strand 6), with DeStefano 2007 and Amadieu 2009.

• Mohageg 1992 (full text): 64 novices, 16 per structure. Hierarchical links beat network links: mean task time 109.21 s against 149.35 s, linear 155.56 s, F(3, 60) = 4.59, p < .01 (Experiment A, p. 359).
• The abstract's 49 s less per task is an average not found in the pages read. The combination, 115.78 s, did no worse than the hierarchy, with no consistent gain for novices.
• McDonald 1998 (full text): 30 readers, 15 with prior knowledge. In the first phase the hierarchy and the mixed structure were equal, both faster than the network; at the delayed phase the mixed structure was fastest and opened the fewest extra nodes (pp. 22 to 23).
• Prior knowledge helped overall; it interacted with structure only for nodes opened while reading, F(2, 24) = 8.1 (Table 1, p. 22).
• Located, abstract not readable on the day: McDonald 1996, Calisir 2003, Potelle 2003.

4. Q3: focus plus context and zooming for documents and trees

• Hornbæk 2002 (abstract, secondary): 32 subjects, zoomable maps with and without an overview. Accuracy did not differ and 80% preferred the overview; on the map built in levels with semantic zoom, subjects were faster without it.
• Through Cockburn 2009 (section 7.2, secondary):
  • Jakobsen 2006: a fisheye of source code that folds by degree of interest was significantly faster than a flat listing, and preferred.
  • Nekrasovski 2006: on large trees, pan and zoom beat a rubber-sheet fisheye; an overview did not change task time.
  • Plaisant 2002 and the hyperbolic-tree studies: results depend on the task and on how well labels predict what lies below.
  • Baudisch 2002: a focus plus context screen was up to 56% faster than overview plus detail and zooming on static graphs.
• Cockburn 2009 (section 5.2.2): folding of program blocks is standard in editors and of sections in word processors, while automatic folding by degree of interest stays in research labs.

5. Q4: levels in one view

• No study found varies the depth of a stored tree while holding the depth in view fixed.
• Measured levels: Chimera 1994 tested three levels in view at most; Furnas 1986 shows the ancestral line plus one or two rings around the focus.
• For menus, two levels beat three: file-structure (Kiger 1984; Larson 1998).
• For proofs, about 4 to 10 steps per level, a convention: mathematical-writing.
• Nielsen 2006b: designs beyond two disclosure levels typically have low usability; no data given.
• For model readers, He 2026 (strand 6) found one level of disclosure enough and a second harmful.
• A node read alone loses its place in the tree; in a corpus of topic-based documentation, relative clauses carry the context of related topics (Swarts 2022, abstract).

6. Q6: trees of summaries for models as readers

• Wu 2021 (abstract, secondary): a book summarized section by section, then the summaries recursively; matched human quality on about 5% of books.
• Sarthi 2024 (RAPTOR, read): leaves are chunks, parents are summaries of clusters of children, about 72% shorter (Appendix C).
  • With GPT-4 it reached 82.6% on QuALITY against a previous best of 62.3% (Table 7); 18.5% to 57% of retrieved nodes were non-leaves (Appendix I.2).
  • Searching all layers at once, the tree flattened, beat descending it layer by layer on 20 QASPER stories (section 3, Figure 3).
• Laitenberger 2025 (read): passages retrieved and kept in the document's original order matched or beat RAPTOR and an agent at matched token budgets.
  • On ∞Bench at 30K tokens: 93.1% against 87.8% for the same passages ordered by similarity, and 2 to 8 points above the multi-stage methods (section 3).
  • RAPTOR without its summaries beat RAPTOR with them; its original ablation used a 3B reader limited to 400 tokens (section 4).
• He 2026 (read): agents reading books on ∞Bench, three harnesses, three model families.
  • A second level of disclosure never beat one level and sometimes collapsed accuracy: 0.9126 to 0.6398 on one cell (Table 1).
  • With 20 books staged, raw navigation fell to 0.26 on English open questions where one level held 0.46.
• Jayraj 2026 (abstract, secondary): a tree used only to navigate, emitting leaves, matched the strongest flat retriever; the gain came from navigation, not summaries.
• Chen 2023 (abstract): a model walks a tree of summary nodes to answer; better than long-context and retrieval baselines. Edge 2024 (abstract) summarizes graph communities to answer questions about a whole corpus; it is a graph, not a tree.
• Mortezaagha 2026 (abstract, secondary): on 2,359 PubMed Central articles, content-similarity retrieval had the best MRR (0.517) but drew from one section; structure-aware retrieval drew from up to 15.6 times more sections.
• Chang 2024 (abstract, secondary): on 100 books, merging summaries up a hierarchy was more coherent than updating one running summary, which kept more detail.

7. Practice

Who	Form	Evidence of use
WHATWG 2026	details and summary: one disclosure; a shared name makes an exclusive accordion; not for footnotes or tabs	the HTML standard, updated 2026-10-26
W3C 2026b	tree view pattern: nested groups, expand with Right, collapse with Left; level and position given only when nodes load lazily	no depth limit stated
GOV.UK 2026a	accordion: not for content all users need; never inside another accordion; not all users notice it	government design system; research reported without numbers
GOV.UK 2026b	details: never hide what most users need; some users avoid it, fearing it leaves the page	same
Nielsen 2006b	progressive disclosure: few options first, at most two levels	company article, no data
Wikimedia 2016	mobile Wikipedia folds every section below the lead; a test opened them	expanded: reading time 190 s against 146 s at the 90th percentile, 53 s against 38 s at the 75th
Outliners (Logseq, SiYuan, Obsidian plugins)	nested bullets, fold, zoom into one bullet as the root	counts in state-of-the-art.md
Code editors	folding of blocks and #region	standard feature (Cockburn 2009)
Tree retrieval (RAPTOR, GraphRAG, PageIndex)	trees or graphs of summaries for model readers	counts in state-of-the-art.md; PageIndex reports 98.7% on FinanceBench itself (VectifyAI 2026, example only)

• Wikimedia 2016 is not a strict A/B test: about 0.05% of mobile users per group, run in 2015-12 and 2016-01; the session chart covers 2016-01-02 to 2016-01-08.
• Wang 2023 (example only, held by typography-and-layout) says folded content may be missed.

Q/A sites (searched 2026-10-29; 12 threads, 2 under 10 of 20; strings in protocol). Practitioner opinion, no data.

• Unbounded depth is a known problem: indenting fails after a few levels of nested comments; one site moves a thread to a new page after about ten levels (UXSE 2010).
• A path of a hundred decisions needs checkpoints, not a longer breadcrumb (UXSE 2015).
• Nesting folds is advised against: an accordion inside an accordion confuses (UXSE 2012a; UXSE 2014b), as GOV.UK 2026a says.
• Do not hide answers to save scrolling; list the questions with anchor links instead (UXSE 2012b).
• A deep heading should name itself in full, since a reader may meet it without its ancestors (WSE 2013); whether every level carries text of its own divides practitioners (WSE 2018e).
• One long page or many granular ones: cohesion matters more than length, and a tree keeps overview pages near the root (WSE 2018c, scored in information-typing).
• UXSE 2018b (example only) asks the same of a three-level nest.
• Folding code hides a long block but gives it no name, scope or dependencies, which a function does (SESE 2014e; SESE 2011o).
• Depth can be shown by colour or a gutter per level instead of indentation (SESE 2011r).
• Examples only: a five-level tree on one mobile screen (UXSE 2018a); "expand all" for searching folded sections (UXSE 2016).