PaperlayerView exact revision on arXiv

Source-linked AI summary

Auto-Slides: An Interactive Multi-Agent System for Creating and Customizing Research Presentations

Yuheng Yang, Wenjia Jiang, Yang Wang, Yi Song, Yiwei Wang, Chi Zhang

arXiv:2509.11062v3cs.HCcs.MA

TL;DR

LLM-assisted paper learning can lack structure and rely too heavily on text, limiting systematic understanding of complex research. Auto-Slides transforms papers into cognitively guided multimodal slides with interactive customization, verification, and retrieval. User studies report stronger structural support, learner preference, and improved narrative quality compared with LLM chat-based learning.

  • Problem

    LLM-based paper dialogue can fragment understanding through weak organization and text-dominant explanations lacking diagrams, tables, and visual summaries.

    Source: §I · pp. 1–2

  • Method

    Auto-Slides uses a multi-agent framework to parse papers, restructure them into PMRC narratives, generate multimodal slides, support interactive refinement, and verify content.

    Source: §I · p. 2, A. Content Understanding and Structuring · p. 3

  • Results

    Slide-based learning outperformed standard LLM chat in visual clarity, structural organization, and support for understanding, with most participants preferring slides.

    Source: C. User Study 2: Comparing Slides and… · p. 5

  • Takeaways & Limitations

    Auto-Slides strengthens structural support for understanding complex scientific content while preserving factual accuracy and improving learner experience.

    Source: §IV · p. 6

Abstract

from arXiv · show

The rapid progress of large language models (LLMs) has opened new opportunities for education. While learners can interact with academic papers through LLM-powered dialogue, limitations still exist: the lack of structured organization and the heavy reliance on text can impede systematic understanding and engagement with complex concepts. To address these challenges, we propose Auto-Slides, an LLM-driven system that converts research papers into pedagogically structured, multimodal slides (e.g., diagrams and tables). Drawing on cognitive science, it creates a presentation-oriented narrative and allows iterative refinement via an interactive editor to better match learners' knowledge level and goals. Auto-Slides further incorporates verification and knowledge retrieval mechanisms to ensure accuracy and contextual completeness. Through extensive user studies, Auto-Slides demonstrates strong learner acceptance, improved structural support for understanding, and expert-validated gains in narrative quality compared with conventional LLM-based reading. Our contributions lie in designing a multi-agent framework for transforming academic papers into pedagogically optimized slides and introducing interactive customization for personalized learning.

I. INTRODUCTION

AUTO-SLIDES addresses the fragmented, text-heavy nature of LLM-assisted paper learning by reorganizing research into multimodal, pedagogically structured presentations with interactive refinement and reliability safeguards.

Motivation

  • LLM dialogue can make specialized research more approachable, but isolated questions and answers obscure papers’ overall structure and systematic understanding.Text-dominant outputs also underserve concepts whose comprehension benefits from diagrams, tables, and visual summaries.Source: §I · pp. 1–2

System goal

  • AUTO-SLIDES transforms research papers into visually rich, well-structured, pedagogically optimized slide decks for more engaging learning.Source: §I · p. 2

Narrative restructuring

  • The system uses cognitive-science principles and a Problem–Method–Result–Conclusion framework to convert paper logic into a presentation-oriented narrative.Cognitive Load Theory and Mayer’s Multimedia Learning Theory motivate coherent sequencing and combined verbal-visual presentation.Source: §I · p. 2

Interactive customization

  • An interactive editor lets learners request additions or simplifications, tailoring slide decks to different knowledge backgrounds, goals, and detail requirements.The resulting materials are intended to remain accurate and coherent while aligning with individual learning needs.Source: §I · p. 2

Reliability and enrichment

  • Verification–Adjustment checks slide content against the source paper, while retrieval enriches explanations with background from cited or related works.Together, these mechanisms target accurate, comprehensive, and contextually enriched learning materials.Source: §I · p. 2

Evaluation

  • User studies evaluate slide quality, interactive customization, and comparison with direct LLM-assisted paper reading, reporting stronger structural support and improved expert-rated narrative quality.The studies also examine learner experience and comparative advantages over conventional LLM-based reading.Source: §I · p. 2

II. METHOD

AUTO-SLIDES uses a multi-agent pipeline that understands paper content, verifies and adjusts slide plans, then generates multimodal slides with human-in-the-loop editing.

II. METHOD

  • The pipeline combines content understanding, quality assurance, and generation and interaction to transform papers into editable multimodal slides.Parser and Planner agents design structure; Verification and Adjustment agents check fidelity; Generator and Editor agents produce and revise slides.Source: §II · p. 2

A. Content Understanding and Structuring

The content-understanding stage extracts complex paper elements and reorganizes them into a cognitively guided PMRC presentation plan.

A. Content Understanding and Structuring

  • The Parser Agent uses Marker to create structured Markdown, then extracts tables and equations into JSON to preserve complex multimodal data.This two-stage strategy reduces context truncation and structural errors during downstream processing.Source: A. Content Understanding and Structuring · p. 2
  • The resulting structured data lets later agents retrieve complex multimodal elements accurately without scanning the full paper text.Source: A. Content Understanding and Structuring · p. 3
  • The Planner Agent converts reader-centric IMRaD into a presenter-centric Problem–Motivation–Results–Conclusion narrative guided by Cognitive Load Theory, and Multimedia Learning Principles.Its JSON plan sequences concepts, bullet points, and visual aids while optimizing spatial contiguity between text and visuals.Source: A. Content Understanding and Structuring · p. 3

B. Quality Assurance and Content Adjustment

AUTO-SLIDES inserts verification and targeted adjustment between planning and generation, then compiles verified plans into visually consistent academic presentations.

B. Quality Assurance and Content Adjustment

  • LLM planning remains vulnerable to context-length omissions and hallucinated deviations from source facts, which can omit findings or introduce inaccuracies.Source: B. Quality Assurance and Content… · p. 3
  • Verification compares plans with structured source content and triggers targeted adjustment only for insufficient coverage or missing high-importance items.Adjustment retrieves source material, integrates it into existing slides or adds supplemental slides, and reformulates it into concise presentation bullets.Source: B. Quality Assurance and Content… · p. 3
  • The QA step is intended to prevent factual-error propagation and preserve high-value content, completeness, and presentation reliability.Source: B. Quality Assurance and Content… · p. 3

C. Multi-modal Generation

  • The Generator Agent converts the verified JSON plan into LaTeX Beamer slides containing text, figures, tables, and equations, with automatic compilation-error correction.A user-supplied theme controls colors and style parameters while the system reconstructs and positions multimodal elements.Source: C. Multi-modal Generation · p. 3

D. Interactive Optimization

AUTO-SLIDES supports natural-language slide refinement through an Editor Agent that retrieves content, edits LaTeX, and recompiles presentations. The broader evaluation examines interactive editing, learning formats, narrative optimization, and component-level fidelity.

D. Interactive Optimization

  • The Editor Agent orchestrates content retrieval and LaTeX modifications to compile updated presentations.Source: Figure 3 · p. 4, D. Interactive Optimization · p. 4
  • Natural-language commands are decomposed into locate, search, modify, insert, and delete actions for iterative slide editing.The ReAct-based Editor Agent dynamically edits the underlying LaTeX source and can retrieve external knowledge from cited papers.Source: D. Interactive Optimization · pp. 3–4

III. EXPERIMENTS

  • AUTO-SLIDES is evaluated through three user studies and one automated ablation covering editing, learning-format comparison, narrative optimization, and content fidelity.Source: §III · p. 4

A. Evaluation Metrics

The evaluation combines learner, expert, and LLM-judge assessments to measure interactive learning, comparative slide experiences, narrative quality, and pairwise presentation preferences. Human ratings use 7-point scales, while blind LLM comparisons use bidirectional checks for reliability.

A. Evaluation Metrics

  • Human participants rate items on 7-point Likert scales averaged into composite scores.Source: A. Evaluation Metrics · p. 4
  • Learner interactive perception measures learning enhancement and control and agency.Source: A. Evaluation Metrics · p. 4
  • Learner comparisons assess slides versus chat across visual clarity, structural organization, learning flow, understanding support, and satisfaction.Source: A. Evaluation Metrics · p. 4
  • Experts evaluate narrative quality through content accuracy, information density, and narrative structure and flow.Source: A. Evaluation Metrics · p. 4
  • LLM judges perform blind bidirectional pairwise comparisons and record Forward–Reverse Agreement to quantify aggregation reliability.Source: A. Evaluation Metrics · p. 4

B. User Study 1: Interactive Slide Functions from Learners’ Perspective

User Study 1 examines whether interactive slide editing improves learners’ perceived understanding and control. Participants edited an automatically generated deck, then rated both dimensions above the neutral midpoint, while some cautioned that interaction could distract.

B. User Study 1: Interactive Slide Functions from Learners’ Perspective

C. User Study 2: Comparing Slides and LLM Chat-based Learning from Learners’ Perspective

User Study 2 compares automatically generated slides with standard LLM chat for learning unfamiliar HCI papers. Slides outperformed chat in visual clarity, structural organization, and understanding support, leading most participants to prefer slides while still valuing chat’s interactive exploration.

C. User Study 2: Comparing Slides and LLM Chat-based Learning from Learners’ Perspective

  • The study recruited 24 HCI and computer-science researchers balanced by prior LLM proficiency.Source: C. User Study 2: Comparing Slides and… · p. 5
  • Participants studied three unfamiliar peer-reviewed HCI papers in both free-navigation slide and open-follow-up chat conditions for up to 15 minutes each.The counterbalanced within-subject procedure included absolute ratings, comparative ratings, forced-choice preference, and open-ended feedback.Source: C. User Study 2: Comparing Slides and… · p. 5
  • Slides outperformed standard LLM chat in visual clarity, structural organization, and support for understanding, and most participants preferred slides.Table II reports mean and standard-deviation ratings on a 7-point Likert scale with paired tests and Holm–Bonferroni correction.Source: C. User Study 2: Comparing Slides and… · p. 5, Table II · p. 5
  • Participants valued slides’ visual and structural grounding as a scaffold, while valuing LLM chat for interactive exploration.The findings support a hybrid workflow in which slides provide essential coverage before personalized inquiry through chat.Source: C. User Study 2: Comparing Slides and… · p. 5

D. User Study 3: Evaluating Narrative Structure Optimization from Experts’ Perspective

User Study 3 found that reorganizing paper content into a narrative structure improved experts’ assessments of slide accuracy and flow without changing information density.

Goal and Participants

  • The study recruited 8 experienced researchers to compare optimized and non-optimized slides generated from their own papers.Participants rated accuracy, information density, and narrative flow using a 9-item rubric.Source: D. User Study 3: Evaluating Narrative… · p. 5

Materials

  • The full pipeline converted IMRaD content into a pedagogical Problem–Motivation–Results–Conclusion narrative, whereas the ablation preserved the original sequence.Parsing and verification stages were identical, and no human post-editing was performed.Source: D. User Study 3: Evaluating Narrative… · p. 5

Results

  • Slides with narrative structure optimization outperformed ablations in Content Accuracy and Narrative Structure & Flow, while Information Density remained consistent.Experts linked the gains to smoother conceptual navigation and higher fidelity in presenting findings.Source: D. User Study 3: Evaluating Narrative… · p. 5, Table III · p. 6
  • Experts attributed the improvements to smoother conceptual navigation and more faithful presentation of findings.The reported interpretation is that reorganization, rather than increased content volume, supported understanding.Source: D. User Study 3: Evaluating Narrative… · pp. 5–6

E. Automatic Evaluation: LLM-as-Judge Component Ablation

An LLM-as-judge ablation evaluated whether enhanced parsing and verification improve slide fidelity and accuracy. Enhanced parsing substantially improved table fidelity, while verification provided mainly supplementary protection because the base pipeline was already accurate.

Experimental Setup

  • The study compared the Full AUTO-SLIDES pipeline with variants omitting second-stage parsing or verification across 14 diverse academic papers.Slides were normalized into a plain-text slides.json format to avoid formatting bias in pairwise judgments.Source: E. Automatic Evaluation: LLM-as-Judge… · p. 6

Evaluation Protocol

  • GPT-4o served as the primary judge, Claude-3.7-Sonnet replicated fidelity evaluations, and human re-evaluation agreed with GPT-4o at 90% with κ = 0.82.The validation supports alignment between automated judgments and domain-informed human consensus.Source: E. Automatic Evaluation: LLM-as-Judge… · p. 6

Results

  • 67.9% of Table Fidelity comparisons favored the Full system versus 17.9% losses across both judge models.This result highlights the Parser Agent’s role in preserving structured data, formulas, and precision.Source: E. Automatic Evaluation: LLM-as-Judge… · p. 6
  • Content Accuracy comparisons were dominated by ties at 85.7%, with 7.1% Full-system wins and 0% losses.The pattern indicates that verification mainly safeguards borderline cases rather than correcting frequent errors.Source: E. Automatic Evaluation: LLM-as-Judge… · p. 6

IV. CONCLUSION

AUTO-SLIDES is an LLM-driven multi-agent framework that transforms research papers into pedagogically optimized, multimodal slide decks. Its pipeline combines parsing, narrative restructuring, verification, generation, and interactive editing to support understanding while preserving factual accuracy.

Conclusion

  • AUTO-SLIDES transforms papers into pedagogically optimized slide decks through high-fidelity multimodal parsing, narrative restructuring, and verification loops.The conclusion states that this combination strengthens structural support for understanding complex scientific content while preserving factual accuracy.Source: §IV · p. 6
  • Table IV reports component-level comparisons between the Full system and ablated variants using aggregated win/tie rates from forward and reverse evaluations.Agreement measures consistency across presentation positions.Source: Table IV · p. 6
Loading 2509.11062v3…