Source-linked AI summary
Scim: Intelligent Faceted Highlights for Interactive, Multi-Pass Skimming of Scientific Papers
Raymond Fok, Andrew Head, Jonathan Bragg, Kyle Lo, Marti A. Hearst, Daniel S. Weld
TL;DR
Researchers struggle to efficiently sift through scientific literature, motivating Scim, an AI-augmented interface that organizes automatically identified salient sentences into rhetorical facets and a highlight browser. In a 13-participant design-probe study, faceted highlights complemented skimming through rereading and guided reading, and participants were eager to use them in future reading interfaces.
Problem
Researchers need efficient ways to sift through rapidly evolving scientific literature, but information overload remains a major challenge.
Method
Scim is an augmented reading interface that automatically localizes, classifies, ranks, and organizes salient paper sentences into rhetorical facets and a highlight browser.
Results
In a 13-participant study, faceted highlights and the highlight browser complemented skimming by facilitating rereading and guided reading, while participants expressed interest in future use.
Takeaways & Limitations
Faceted highlights and highlight browsers can support researchers’ existing skimming processes within augmented reading interfaces.
Takeaways & Limitations
The study examined only one category of human-computer interaction papers in computer science, so Scim’s generalizability across papers and disciplines remains to be evaluated.
Abstract
from arXiv · showhide
Researchers are expected to keep up with an immense literature, yet often find it prohibitively time-consuming to do so. This paper explores how intelligent agents can help scaffold in-situ information seeking across scientific papers. Specifically, we present Scim, an AI-augmented reading interface designed to help researchers skim papers by automatically identifying, classifying, and highlighting salient sentences, organized into rhetorical facets rooted in common information needs. Using Scim as a design probe, we explore the benefits and drawbacks of imperfect AI assistance within an augmented reading interface. We found researchers used Scim in several different ways: from reading primarily in the `highlight browser' (side panel) to making multiple passes through the paper with different facets activated (e.g., focusing solely on objective and novelty in their first pass). From our study, we identify six key design recommendations and avenues for future research in augmented reading interfaces.
1 INTRODUCTION
Researchers use skimming to keep pace with expanding scientific literature, but the practice remains cognitively demanding and difficult to learn. Scim explores AI-augmented skimming by highlighting and classifying salient sentences while studying how researchers use and assess imperfect assistance.
- 1 INTRODUCTION: Researchers skim papers to extract important information quickly, helping them keep pace with growing literature despite skimming’s cognitive demands and learning difficulty.Skimming is faster than deep reading, but requires deliberate scanning, selective attention, and navigation through text.
- 1 INTRODUCTION: Scim automatically identifies salient sentences, classifies them into four rhetorical facets, links them to paper context, and provides interactive controls for skimming.Its interface integrates document highlights, a linked highlight browser, and facet-based filtering.
- 1 INTRODUCTION: In a study with 13 participants, Scim served as a design probe for opportunities, benefits, risks, and design implications of AI-augmented scientific-paper skimming.The study focused on integrating potentially imprecise AI support into reading interfaces.
- 1 INTRODUCTION: Participants used Scim’s highlight browser as a rereading mechanism to efficiently verify their comprehension.Researchers also treated highlights as importance cues from an opaque intelligent agent.
- 1 INTRODUCTION: Participants largely expressed interest in Scim’s core features and excitement about AI-assisted reading despite potential errors, while forming first impressions of its reliability.The paper identifies future opportunities in collaborative highlights and personalization within intelligent reading interfaces.
2 RELATED WORK
Related work characterizes skimming as goal-directed but difficult and error-prone, motivating interfaces that surface relevant information during rapid document exploration. Existing augmentations include scrolling behavior, typographical cues, scientific-paper-specific links, and summaries, although summaries can be unreliable and cannot support interactive exploration of the full paper.
- Skimming: Skimming helps scholars rapidly form a general idea by focusing on goal-relevant content, but readers may miss important information because gaze, attention, and cognition limit accurate selection.Readers may instead use satisficing, moving on when perceived information gain falls below a threshold.
- Skimming: Scientific-paper skimming is often hasty and incomplete, with attention drawn to visual content and section headers, leaving information-dense text vulnerable to being skipped.This motivates automated assistance for discovering relevant information units buried in plain text.
- Augmented Reading Interfaces: Augmented reading interfaces support exploration through contextual supplemental information, annotations, content-aware scrolling, pseudo-haptic feedback, heading resizing, and typographical highlighting.Scientific-paper systems additionally connect text with citations, charts, tables, and generated or animated visualizations [37, 35, 4, 26].
- Summarization: Summarization offers a shortened alternative to reading a full paper, using abstracts or automated extractive and abstractive methods for long-form documents [3] [63].Some systems generate extreme single-sentence summaries.
- Summarization: Summaries are unsuitable as standalone replacements because they remain error-prone and susceptible to hallucination [76], while preventing readers from interactively exploring the full paper as goals change.Readers may need to investigate particular sections or information beyond what the summary contains.
3 DESIGN MOTIVATIONS
A formative study found that researchers commonly begin skimming with the Abstract and Introduction, while broader strategies vary with goals and experience. These findings motivated Scim’s focus on document-wide discovery, contextual access, attention guidance, and recovery from imperfect automated judgments.
- Formative study findings: Researchers typically began skimming with the Abstract and Introduction, while strategies beyond these sections varied according to skimming goals and research experience.Researchers used typographical cues, structural cues, and visual media to identify regions likely to provide information gain.
- Formative study findings: The formative study suggested that readers often search first for cues indicating a paper’s significance before pursuing further detail.This motivated using automated tools to accelerate discovery of significance-defining cues within papers.
- Design motivations: DM1 sought to scaffold information discovery throughout an entire paper, including relevant content in middle sections that conventional skimming often overlooks.The design explored natural language processing for document-level understanding of text-heavy sections.
- Design motivations: DM2 sought to connect readers to salient content in context by pairing condensed views with on-demand access to additional context.The motivation addresses the risk that static condensed representations may require less processing but lack sufficient context.
- Design motivations: DM3 sought to direct reader attention while minimizing distractions, with prototype evaluations finding highlighting more familiar than underlining or masking.The evaluation considered three cueing mechanisms: underlining, highlighting, and masking.
- Design motivations: DM4 sought to support error recovery because automated systems can make systematic and reader-dependent relevance errors when identifying and classifying salient content.Content considered important by one reader may be irrelevant to another reader in a particular context.
4 THE SCIM SYSTEM
Scim is an augmented reading interface that organizes and filters salient content to support skimming of scientific papers through faceted highlights. An end-to-end document-processing pipeline localizes, classifies, and ranks important sentences.
- Scim organizes and filters salient content within scientific papers through an augmented reading interface.
- An end-to-end document-processing pipeline localizes, classifies, and ranks important sentences.
- The interface uses faceted highlights to support skimming of scientific papers.
4.1 User Interface
Scim augments papers with in-context highlights, rhetorical facets, a highlight browser, and navigational annotations while preserving the document’s original structure. Readers can switch between standard skimming and guided reading, view highlights efficiently, and tailor which facets appear.
- 4.1 User Interface: Scim preserves the paper’s original structure while adding overlays and side panels that support switching between standard skimming and guided reading.The web interface was implemented atop pdf.js.
- 4.1 User Interface: Scim uses unobtrusive colored in-context highlights to identify salient sentences while preserving surrounding context and conveying the paper’s highlighted structure.Highlight opacity supports attention to both highlighted text and its context.
- 4.1 User Interface: Scim organizes highlights into a minimal taxonomy of four rhetorical facets combining coarse-grained abstract classification with the NOV_ADV category from Argumentative Zoning.The taxonomy was selected to correspond closely to readers’ shared information needs during skimming.
- 4.1 User Interface: The highlight browser presents a condensed, location-ordered list of highlights in a hideable right-side panel, reducing the need to scroll through the entire paper.Browser highlights are also demarcated by document section.
- 4.1 User Interface: Readers can inspect highlight quantity and distribution through colored scrollbar annotations and filter document, browser, and scrollbar highlights by selecting facets from the palette.All facets are enabled by default, and “Everything” restores the complete view.
4.2 Document Processing Pipeline
Scim’s document-processing pipeline converts PDF content into sentence-level objects, scores and classifies sentences into rhetorical facets, refines their rankings with structural heuristics, and exports the results for the reading interface.
- Document Processing Pipeline: MMDA processes PDF tokens, mathematical symbols, headers, and metadata, while merged token and row boxes form sentence-level bounding boxes linked to section headers.The pipeline segments paper text into sentences before constructing these sentence representations.
- Document Processing Pipeline: Sentence salience is computed from the sum of nonnegative cosine similarities between Universal Sentence Encoder embeddings and normalized to [0, 1].The measure favors sentences similar to many others because they are presumed central to the document’s ideas.
- Document Processing Pipeline: A SciBERT-initialized BERT classifier assigns previously defined facets, while Novelty uses lexical and discourse heuristics because a separate annotated training set was unavailable.Novelty matching checks terms such as inconsistent and however alongside author-intent and prior-work indicators.
- Document Processing Pipeline: Structural heuristics refine rankings by favoring paragraph beginnings and endings and facet-appropriate expected sections, combining these signals with classifier probabilities.Examples include Objective in an Introduction and Novelty in a Related Work section.
- Document Processing Pipeline: The pipeline outputs ranked sentence objects containing original text, classified facet, PDF bounding box, and enclosing section headers, stored in JSON for the interface.These files were stored on a remote server and processed by the user interface.
5 USABILITY STUDY
The usability study used Scim as a design probe with 13 scientifically experienced participants completing structured paper-skimming tasks remotely. Researchers combined interaction logs, recordings, surveys, feature-quality feedback, and qualitative interviews to examine usage and improvement opportunities.
- 5 USABILITY STUDY: Thirteen participants with prior scientific-paper experience evaluated Scim through purposive and snowball sampling.The sample included 11 computer science graduate students, one undergraduate, and one senior research programmer.
- 5 USABILITY STUDY: The remote study recorded participant screens and audio and logged all Scim interactions while accommodating realistic reading environments.The researchers could not control participants’ hardware, screen sizes, or external distractions.
- 5 USABILITY STUDY: Participants completed timed skimming tasks involving comprehension questions, strengths and weaknesses, or presentation-outline creation across three recent ACM UIST papers.Participants were split into two groups, with each group completing different combinations of six-minute skimming and follow-up activities.
- 5 USABILITY STUDY: After the reading tasks, participants completed the System Usability Survey and answered questions about feature usefulness and automatically generated highlight quality.The study concluded with 10–15-minute semi-structured interviews about participants’ experiences, preferred features, improvement areas, and future AI-assisted reading interfaces.
- 5 USABILITY STUDY: Two authors analyzed interview data qualitatively by transcribing recordings, developing and refining themes, and extracting relevant participant utterances.Participants were identified using pseudonyms P1–P13, and quoted utterances were edited to remove identifying information while preserving meaning.
6 RESULTS
Participants generally found Scim useful and easy to use for faceted skimming, navigation, and post-reading comprehension, especially when seeking information under time pressure. However, AI errors could reduce trust, while reliance on highlights risked producing an overly positive and biased impression of papers.
- Perceived Benefits: Most participants viewed Scim as helpful for knowledge distillation, with less experienced readers valuing guidance and participants expecting benefits under time pressure.Participants also used the highlight browser after reading to revisit filtered sentences, summarize the paper, and verify their understanding.
- Skimming Practices: Participants used Scim’s inline highlights, highlight browser, and facet palette to navigate papers, pursue specific reading goals, and make multiple skimming passes.They often filtered for Objective and Novelty first, then examined Method or Result facets, sometimes opening highlights in paper context.
- Usability: Scim received an average SUS score of 86.8, indicating excellent overall usability, while its visual and interactive simplicity supported adoption.Lower usability assessments reflected the cognitive load of verifying inconsistent AI highlights.
- Errors and Trust: Participants encountered salience-detection, facet-classification, and PDF-processing errors, and some disabled Scim after judging its highlights too inaccurate.Others tolerated errors when automated guidance seemed more beneficial than harmful, so adoption depended on individual expectations rather than a universal tolerable error level.
- Precision and Recall: Participants differed on the desired precision–recall balance, with some preferring high-confidence highlights first and the option to add more potentially noisy highlights.Too few highlights forced conventional skimming, whereas too many irrelevant highlights created another burden.
- Bias and Limitations: Heavy reliance on Scim’s highlights could bias skimming toward a paper’s positive aspects because readers focused on highlighted content and overlooked much of the rest.One participant reported reading only about 10% of non-highlighted text and developing an unnaturally positive impression.
7 DISCUSSION
The discussion proposes six design recommendations for Scim-like augmented reading interfaces, while identifying limitations, adoption barriers, and future directions for adaptive and personalized assistance.
- Design Recommendations: Reading interfaces should support details-on-demand, consistent highlight coverage, and both passive skimming and active annotation workflows.Scim linked highlights to their paper context; readers preferred highlights distributed throughout papers, while nested interactions could support deeper inspection and note-taking.
- Design Recommendations: Augmented reading systems should preserve the human-AI reading process across sessions and provide recovery paths when AI highlights are irrelevant or inadequate.Persistence may include exported faceted highlights, reader annotations, preferences, and paper interactions, while recovery can involve ignoring a highlight or reading surrounding context.
- Design Recommendations: Interfaces should reduce AI-induced reading bias because participants sometimes relied on highlighted content and read less non-highlighted text, potentially producing a positive bias.Selective highlighting may alter what readers attend to compared with a traditional skim.
- Limitations: Scim’s generalizability is limited because the study used one category of HCI papers, limited interaction time, and a small predetermined paper set.Longer-term use might improve familiarity with facets and color associations, but could also reveal adoption-limiting shortcomings.
- Adoption Challenges: Adoption may be constrained by distrust that AI highlights identify the most relevant content, with some participants favoring highlights created by fellow researchers.The discussion points to shared or popular highlights as a possible alternative source of trust.
- Future Directions: Future systems could learn from passive and active interactions and tailor assistance to readers’ experience, information needs, and expectations.Potential signals include dwell patterns, highlighting, removing AI highlights, and interactions that communicate uncertainty or recover from AI errors.
8 CONCLUSION
Scim is an augmented reading interface that uses faceted highlights to support researchers skimming scientific papers. In a 13-participant design-probe study, highlights within the paper and a rhetorical-facet browser complemented existing skimming by facilitating rereading and guided reading.
- 8 CONCLUSION: Scim provided colored, rhetorically organized highlights within papers and a highlight browser to support scientific-paper skimming.The interface was evaluated as a design probe with 13 participants.
- 8 CONCLUSION: In a 13-participant study, Scim complemented participants’ existing skimming processes by facilitating rereading and guided reading.