Marvista: Exploring the Design of a Human-AI Collaborative News Reading Tool

1 INTRODUCTION

• Online media often comes in small bits encouraging “rapid, distracted sampling” and weakening ability to focus, read, and think deeply, especially when information such as long articles goes well beyond most social media posts.Reading can be challenging even for people who comprehend text, who may struggle to truly understand and appreciate what it says.Source: §1 · pp. 1–2
• Prior reading techniques and devices lack guidance tailored to the specific text, reader interests, and time constraints.The paper identifies text-specific guidance as an unmet need for news reading and notes NLP models as a possible resource.Source: §1 · p. 2
• Marvista supports planning with time and question filters, paragraph-level reflection questions, and an explainable human-AI summary after reading.The summary combines AI text processing with reading behavior and user-generated data; users can hover over a summary sentence to see its source and rationale.Source: §1 · p. 2, Figure 1 · p. 3
• In a study probing interaction rather than comprehension performance, participants reported greater awareness of both the text and how they read it.The study also found the features better suited to information-seeking than casual reading; participants faced limited context and some identified model issues.Source: §1 · p. 3

2.1 Human-AI Collaborative Systems

• Marvista’s reading support follows a complementary human-AI approach: AI processes text across reading stages, while readers’ behavior and generated data personalize support.The summary can reflect a reader’s paragraph-level interest through dwelling time.Source: §2.1 · p. 5

2.2 Tool Support for Reading

• Prior digital reading work includes tools and devices for annotation, navigation, organizing excerpts, and supporting active reading.Examples include XLibris, LiquidText, touch-based techniques, multi-device environments, and pen-and-touch support.Source: §2.2 · p. 5
• Unlike prior news-reading support focused on source credibility or opinion diversity, Marvista generates guidance from article content.The related work also describes tools that support active reading without specifying a particular text type.Source: §2.2 · pp. 5–6

3 DESIGN & IMPLEMENTATION

• Marvista’s implementation reuses existing NLP models as a proof of concept rather than introducing new techniques or individually re-evaluating the models.Source: §3 · p. 6, §3.3 · p. 10

3.1 Before Reading

• Readers set a time limit, and Marvista uses an estimated reading time and extractive sentence ranking to select a must-read subset within that budget.The estimate assumes an average reading speed of 150 words per minute; the time limit sets the summarization compression rate.Source: §3.1 · p. 6
• Readers select AI-generated questions, and Marvista highlights answer-bearing text and reveals the matching question on hover, supporting SQ3R’s question step.Source: §3.1 · pp. 6–7

3.2 During Reading

• Focus mode isolates one paragraph with arrow-key navigation; time spent in focus informs paragraph importance but includes time using the accompanying tools.Source: §3.2 · p. 8
• For a focused paragraph, Marvista generates a reflective question from its extractive summary and withholds the answer, quizzing the reader rather than answering a question for them.Source: §3.2 · pp. 8–9

3.3 After Reading

• The human-AI summary combines abstractive article processing with paragraph dwell time and inserted notes or highlights, whose impact readers can set to none, low, or high.Source: §3.3 · pp. 10–11
• Dwell-time weights multiply paragraph encoder representations; a small experiment using three articles set their range to [0.6, 1.4].The implementation maps time per word to a weight, with greater dwell time increasing a paragraph’s weight.Source: §3.3 · p. 12
• Hover explanations use ROUGE similarity to connect summary sentences to article text or user data, supporting attribution and access to reading history.Source: §3.3 · p. 12

4 USER STUDY: PROBING HOW USERS OPEN-ENDEDLY INTERACT WITH MARVISTA

• The authors chose an open-ended study of tool use, leaving task-specific reading-performance evaluation for future work.Source: §4 · p. 13

4.1 Participants, Tasks, & Procedure

• Seventeen university participants freely explored science and technology articles with Marvista for about 25 minutes, then completed a 30-minute interview.Sessions were in person, used a laptop and Chrome, and participants chose which articles to read.Source: §4.1 · p. 13

4.2 Measurement & Analysis

• Researchers analyzed usage logs through overall and phase-specific quantitative analyses, and used thematic analysis to code interview feedback.Source: §4.2 · pp. 13–14

5 FINDINGS: READING BEHAVIOR USING MARVISTA

• Participants read an average of 3.8 articles, each averaging 928 words, and read 64.0% of article text on average while using focus mode.Articles were read 1.9 times on average; time spent averaged 5 minutes 4 seconds per article and 2 minutes 39 seconds per session.Source: §5 · p. 14, Figure 7 · p. 15

5.1 Usage of Before-Reading Features

• Participants filtered articles by time in 41.1% of cases, by question in 35.5%, and with neither filter in 23.4%.Question-filter users selected 39.2% of generated questions on average; time-filter users set budgets averaging 57.6% of estimated full-article reading time, against a 50.0% default.Source: §5.1 · p. 14, Figure 8 · p. 15
• With the time filter, participants spent 37.5% less time on articles than without a filter, averaging 2 minutes 30 seconds versus 4 minutes per 1,000 words.Participants also tended to spend less time than the budget they had set.Source: §5.1 · p. 14, Figure 8 · p. 15

5.2 Usage of During-Reading Features

• Participants averaged 1.2 highlights, 0.9 notes, and 1.2 reflections per article; two participants used reflection much more often, at 11 and 17 times per article.Focus mode was used in 92.0% of cases, although the authors note its default status may have contributed and requires further study.Source: §5.2 · p. 15, Figure 9 · p. 16

5.3 Usage of After-Reading Features

• About two thirds of finished sessions (67.7%) led participants to generate a human-AI summary, and they sought summary explanations in 59.3% of finished sessions.Participants selected dwelling time and highlights as summary factors in 64.3% of cases each, compared with notes in 26.2%.Source: §5.3 · pp. 15–16, Figure 10 · p. 16

5.4 Summary of Reading Behavior While Using Marvista

• In over three quarters of cases, participants used time or question filters; during reading, they rarely opted out of focus mode but scarcely used active-reading features.After reading, about two thirds generated a human-AI summary, often emphasizing dwelling time and highlights, and participants sought explanations in over half of cases.Source: §5.4 · p. 16

6.1 Feedbacks to Individual Features

• Time and question filters helped participants focus on relevant material and read more efficiently, especially for work, study, or long articles.Generated questions previewed key points and could direct readers to answer-bearing text, but selecting questions could be effortful or difficult without context.Source: §6.1 · pp. 17–18
• Focus mode was universally well-received, helping participants concentrate on one paragraph, though some missed surrounding context or found keyboard navigation unnatural.Participants also described it as useful in distracting environments and as making articles feel less overwhelming.Source: §6.1 · p. 18
• Seven participants found reflection useful for checking understanding or guiding rereading, though some did not want its interruption and others questioned the generated questions.Participants considered the extra effort of rereading and answering worthwhile for understanding a paragraph.Source: §6.1 · pp. 18–19
• Participants used AI-generated summaries both as alternatives to thorough reading and as checks on their own understanding, while some valued the human-AI summary’s personalization.Seven participants reported occasional gist inaccuracies, but did not discount the summary’s usefulness when they could rely on their own judgment.Source: §6.1 · p. 19
• Summary explanations helped participants connect summary content to article details and reflect on their reading, though some found navigating between summary and explanations effortful.Some participants also nearly missed the feature because it was not enabled by default.Source: §6.1 · p. 20

6.2 Recurring Themes

• Participants used Marvista to understand both article content and how they read, including reflecting on time spent and checking paragraph understanding.They also reported that its combination of features changed their usual reading process, which participants were still learning.Source: §6.2 · p. 20
• Marvista’s features fit information-seeking or thorough reading better than entertainment or quick searches, and setting a time limit could feel pressuring.Participants reported reflection as useful when reading thoroughly, rather than searching quickly for specific information.Source: §6.1 · p. 17, §6.2 · p. 21
• Lack of context, rather than extra effort, was a recurring challenge: isolated paragraphs, questions, or extracted sentences could be hard to interpret without surrounding material.The authors suggest coarser paragraph-level recommendations or summary explanations as ways to address context gaps; participants also noticed inaccurate AI outputs and did not always trust them uncritically.Source: §6.2 · p. 21

7.1 Further Evaluations of Reading Performance Using Marvista

• Participants found Marvista more valuable for specific information-seeking tasks than casual reading, motivating controlled studies of reading performance.Planned measures include quiz-like questions and report quality after reading multiple articles.Source: §7.1 · p. 21

7.2 Further Studies and Designs to Address Over-reliance on Marvista

• Current safeguards delay summaries, highlight rather than hide recommended text, and expose explanations; future designs may ask readers to summarize or reflect before seeing AI output.The proposed redesign draws on strategies of prior work to mitigate over-reliance.Source: §7.2 · p. 22

7.3 From Supporting Reading to Encouraging Reading

• A longitudinal study will test whether Marvista is associated with more reading and whether any increase reflects intrinsic motivation or extrinsic assistance.The authors also propose using Marvista as a representative case for mapping human-AI collaborative reading design dimensions.Source: §7.3 · p. 22

7.4 Implications for General Human-AI Collaboration Tools

• The authors generalize that AI can preprocess information and use generated insights to guide human navigation, with physician review of AI-highlighted medical images as an analogy.The analogy also includes summaries that account for a physician’s attention and annotations.Source: §7.4 · p. 22

7.5 Technical Limitations of Marvista’s Current Implementation

• Dwelling time includes non-reading time, post hoc source matching cannot reveal summary generation, and future designs include eye tracking, richer explanations, and smartphone interaction.The proposed smartphone interface could use gestures and reader-AI conversations about text.Source: §7.5 · pp. 22–23
• Extractive highlights can lack context, while news-trained abstractive models may not transfer across domains and can truncate long articles.The authors discuss costly multi-domain training or reference-free pretraining, training-data-dependent summary style, and heuristic feedback weighting without model fine-tuning.Source: §7.5 · p. 23
• Abstractive summaries may hallucinate; explanations support user review, while hallucination detection, reranking, post-processing, and improved training remain future options.Hallucinations and factual inconsistencies remain an open research problem.Source: §7.5 · p. 23
• Generated questions may emphasize low-level facts rather than high-level ideas because training QA corpora favor short factual answers.Future work proposes questions targeting understanding, application, analysis, evaluation, and creation.Source: §7.5 · p. 24

• Participants reported reading more efficiently, focusedly, and informedly with Marvista.Source: §8 · p. 24
• Participants also learned more about their own reading behavior through interacting with Marvista.Source: §8 · p. 24
• The authors hope continued design and model improvements will help encourage more people to read and thrive in an information-rich world.Source: §8 · p. 24