CHI-style related work · positioning draft

Augmenting reading with structure, not summary

A related-work map for an HCI framing: instead of summarizing a text (which automates the reader's schema-building and forfeits comprehension), an LLM computes latent structure and renders it inline over the preserved words. Compiled from seven parallel research passes. Companion to the notation / tools-for-thought prior art →.

jump ▸ why not summarize · closest prior art · in-place augmentation · active reading · argument structure · sensemaking · the gap

1 Why augment instead of summarize

the empirical engine — summaries flatten understanding; effortful, scaffolded reading builds durable models

Melumad & Yun — AI summaries vs. reading sources · PNAS Nexus 20257 experiments, >10,000 participants: LLM summaries yield shallower knowledge than reading the source. The strongest direct evidence that summarization flattens understanding.

Lee, Sarkar, Tankelevitch et al. — GenAI & critical thinking · CHI 2025319 knowledge workers: higher AI confidence predicts less critical thinking and effort — trusting the output displaces the cognition augmentation tries to preserve.

Generation effect · Slamecka & Graf 1978 · Desirable difficulties · Bjork 1994/2011Self-generated material is better retained; effortful processing beats fluent ease. Summaries forfeit the generation benefit; inline structure keeps you reading every word — a productive difficulty.

Cognitive offloading · Risko & Gilbert, TiCS 2016 · Cognitive forcing functions · Buçinca et al., CSCW 2021People offload even when they shouldn't (miscalibrated); designed-in friction reduces over-reliance on AI. Justifies augmentation-not-automation (Engelbart 1962) as the stance.

2 The closest prior art — and exactly where it stops

the “preserve the text, overlay computed structure” move is already published — for single, fixed lenses

closestGP-TSM — AI-resilient text rendering · Gu, Arawjo, Li, Kummerfeld, Glassman · CHI 2024Almost the thesis in print: recursively grays out lower-priority words while keeping all original text legible; explicitly argues against summarization (omission/hallucination). Differs in lens: it does saliency/emphasis, not logical-role / coreference / argument structure — and isn't user-steerable to arbitrary lenses.

Scim — intelligent skimming · Fok et al. · IUI 2023Toggleable highlighting of passages by rhetorical role (objective/novelty/method/result). Your model, but one fixed schema (scientific papers).

ScholarPhi · Head et al. · CHI 2021 · CiteRead · IUI 2022 · Qlarify · UIST 2024In-place definitions of terms/symbols; citation context in the margin; recursively expandable abstracts (detail-on-demand — the inverse of summarization). Each overlays one kind of structure on preserved text.

The Semantic Reader Project · Lo et al. · CACM 2024Umbrella for ~10 AI prototypes overlaying definitions/highlights/citations on legacy PDFs — the canonical “augment in place, don't replace” manifesto to anchor against.

LARF — Let AI Read First (dyslexia) · CHI 2025 (LBW) · Reading.help (EFL) · CHI 2026Mark “important” content via highlight/bold without changing wording (N=150); on-demand grammar/semantics overlays for L2 readers. Structure-preserving emphasis/explanation — adjacent lenses, accessibility framing.

3 Active reading & in-place annotation

the four-decade foundation: marking the exact words (not extracting a gist) is what makes reading active

O'Hara & Sellen — paper vs. on-line · CHI 1997Paper's advantage comes from in-place annotation & structure-extraction — the deficit digital overlays aim to restore.

Marshall — annotation studies · DL 1997 · Hypertext 1998Theorizes annotation as turning an authorial structure into a shared reading structure laid on the exact words — nearly a thesis statement for the framing.

XLibris · Schilit et al. · CHI 1998 · NB · Zyto et al. · CHI 2012 · CommentSpace · Willett et al. · CHI 2011Seminal active-reading machine; classroom-scale in-margin annotation valued over clean text; lightweight imposed structure (typed tags/links) measurably aids sensemaking — precedent for “computed structural annotation.”

Robust anchoring · Brush et al. · CHI 2001 · AnchoredAI · 2025 (preprint)Keeping marks attached to exact words across edits; LLM-anchored inline comments boost agency & reduce over-reliance vs. a chat sidebar — but for writing, not reading.

4 Argument / rhetorical structure of prose

well-served for scientific papers & your-own-draft; open for arbitrary argumentative/humanities text

Argumentative Zoning · Teufel & Moens · Comp. Linguistics 2002Classifies each sentence by rhetorical/argumentative role — the NLP backbone for inline argument highlighting.

AcaWriter / AWA · Knight, Buckingham Shum et al. · JWR 2020 · Mover · 2003Highlight rhetorical moves in a text — but on the user's own academic draft for writing feedback, not on arbitrary published prose for reading.

Marvista · Chen et al. · TOCHI 2023 · fallacy / persuasion-technique highlighting · NLP venuesPer-paragraph inline questions for news; tools tagging fallacies/persuasion tactics — but locked to a fixed taxonomy, not general structure.

ITSS — structure-strategy tutoring · Meyer, Wijekumar et al. · RCTsStrong RCT evidence that teaching readers to detect/visualize text-structure patterns (cause/effect, compare/contrast) aids comprehension — but a fixed pattern set on instructional texts.

5 Sensemaking & qualitative structuring

the deepest theoretical hook — deep reading is schematization, and tools push it off to the side

Sensemaking loop · Russell, Stefik, Pirolli & Card · CHI 1993 · Pirolli & Card 2005Casts deep reading as encoding text into a schema, and names that “schematize” step as the costly, high-leverage transition to support. Your move: push schematization back onto the verbatim text instead of a side panel.

Jigsaw · Stasko et al. · VAST 2007 · ForceSPIRE · Endert et al. · CHI 2012Surface entity/relation structure & learn it from reader gestures — but in separate views, not overlaid on the text.

Cody · PaTAT · CollabCoder · CHI 2021 / CHI 2023 / CHI 2024Mark text spans into a coded structure while reading (AI-assisted) — but require a human-authored codebook; the LLM doesn't propose the schema.

brat · Stenetorp et al. · EACL 2012 · rstWeb · Zeldes · NAACL 2016 · Passages · Han et al. · CHI 2022Render entity/relation/coreference & RST structure in place — but hand-annotated, surface-level, and built as annotation editors, not reading tools. No verified HCI reading interface renders discourse/coreference structure for reading.

6 The gap → contribution

already claimed “Preserve exact words, compute & overlay structure” — GP-TSM, Scim, ScholarPhi, CiteRead, LARF — each for one fixed lens (saliency, rhetorical role, definitions, citations, importance).

already claimed Impose structure from reading — Cody/PaTAT/CollabCoder — but with a human-authored codebook, and rendered in a side view or as surface spans.

open A single LLM-driven reading-lens renderer where the structure is arbitrary & on-demand (logical roles, coreference chains, inferential/argument structure), computed on any text (esp. humanities/argumentative prose with no schema), rendered in situ over unchanged words, and evaluated on comprehension depth / model-building against a summarization baseline.

Defensible contribution claim

Concept: reframe LLM-for-reading from compression (summary) to structure rendering — and from one fixed lens to a general, declarative lens renderer (a parsing spec the LLM compiles onto arbitrary prose).
System: in-situ, exact-text-preserving overlay of latent logical/inferential structure — the hard case being implicit-structure genres (philosophy) no schema-parser handles.
Study: comprehension/transfer (not skimming speed), plain text vs. summary vs. structure-rendered — hypothesis: summaries win speed/recall, structure-rendering wins transfer & “where the argument turns,” because the reader built the model.

Caveats to stay honest about

GP-TSM already owns “augment-not-summarize via rendering”; differentiate hard on lens generality + logical/argument structure + evaluation target, not on the premise.
Color-for-comprehension evidence is thin (Sarkar 2015 is about highlighting broadly; Bionic Reading failed its evals) — don't claim “color → faster”; claim “structure-preserving scaffolding → deeper model.”
Anchoring (Brush; AnchoredAI) and “hidden state outside the text” (Potluck) are real design constraints — favor re-derivable structure over hand-maintained markup.
Several on-theme items are 2026 / preprint (AI Margin Notes, Reading.help) or non-HCI venues (education/psych) — venue-verify before formal citation.

Synthesized from seven parallel research agents (HCI augmented-reading · tools-for-thought/notation · argument-mapping · active-reading/annotation · argument/essay scaffolds · sensemaking/qualitative-coding · recent LLM reading-augmentation), each instructed to prefer primary sources and flag unverifiable claims. Venues/years are best-effort and should be checked against ACM DL / DBLP before citation.

← back to the parsed text