Source-linked AI summary
Narrative-Driven Paper-to-Slide Generation via ArcDeck
Tarik Can Ozden, Sachidanand VS, Furkan Horoz, Ozgur Kara, Junho Kim, James Matthew Rehg
TL;DR
Academic paper-to-slide generation must preserve dense technical content and the source paper’s logical flow, but existing approaches often summarize or organize content without explicit discourse modeling. ArcDeck reconstructs the narrative using discourse trees, global commitments, and iterative multi-agent outline refinement, while ArcBench provides a curated evaluation benchmark. Experiments report improved narrative coherence and structural alignment, with ArcDeck achieving the strongest narrative-flow performance across the reported comparisons.
Problem
Automatic slide generation must convey dense technical content while maintaining concise presentation, visual engagement, and clear narrative flow, and prior datasets have limited evaluation of expert-level quality.
Method
ArcDeck uses RST-based discourse parsing, global commitments, and a multi-agent critique–judge–revise loop to reconstruct and refine paper narratives into slide outlines.
Results
ArcDeck consistently outperforms baselines in narrative-flow comparisons and achieves the highest or second-highest VLM-as-Judge narrative-flow scores across generation backbones and judge models.
Takeaways & Limitations
Explicit discourse modeling and coordinated narrative refinement improve the generated decks’ narrative coherence and structural alignment.
Takeaways & Limitations
Performance may vary between open- and closed-source models, and optimal results may require model-specific prompt tuning.
Abstract
from arXiv · showhide
We introduce ArcDeck, a multi-agent framework that formulates paper-to-slide generation as a structured narrative reconstruction task. Unlike existing methods that directly summarize raw text into slides, ArcDeck explicitly models the source paper's logical flow. It first parses the input to construct a discourse tree and establish a global commitment document, ensuring the high-level intent is preserved. These structural priors then guide an iterative multi-agent refinement process, where specialized agents iteratively critique and revise the presentation outline before rendering the final visual layouts and designs. To evaluate our approach, we also introduce ArcBench, a newly curated benchmark of academic paper-slide pairs. Experimental results demonstrate that explicit discourse modeling, combined with role-specific agent coordination, significantly improves the narrative flow and logical coherence of the generated presentations.
1 Introduction
ArcDeck treats paper-to-slide generation as structured narrative reconstruction, addressing the challenge of preserving logical flow while selecting, organizing, and visually presenting dense technical content. It combines discourse modeling, global commitments, iterative refinement, and a curated benchmark to support coherent generation and evaluation.
- Academic slide generation must simultaneously select content, organize structure, design visuals, and maintain clear narrative flow.
- The central challenge is identifying how concepts are introduced, explained, and related, then mapping that structure into slides alongside technical content and visuals.
- ArcDeck formulates paper-to-slide generation as structured narrative reconstruction and uses a discourse tree to guide structural planning and preserve rhetorical relationships.
- Global commitments capture a paper’s central thesis, scope, and overall narrative structure as persistent conditioning signals for paper-level consistency.
- A closed-loop critique–judge–revise process iteratively refines slide outlines to improve coherent structural planning and narrative flow across the deck.
- ArcBench is a curated, open-source benchmark of author-prepared paper–slide pairs designed to evaluate text coverage, narrative coherence, visual quality, and content retention.
2 Related Work
Prior paper-to-slide datasets often lack academic specialization, presentation-quality curation, or public availability. ArcBench addresses these gaps by releasing 100 high-quality oral-presentation pairs curated from 994 papers at major computer vision and machine learning venues.
- Earlier systems evolved from regression-based content selection and section summarization toward LLM, multi-agent, intermediate-JSON, and code-generation approaches.
- Paper-Slide Datasets: Existing paired datasets may lack academic specialization, presentation-quality curation, or public release, limiting evaluation of expert-level technical slide generation.
- Paper-Slide Datasets: ArcBench curates 100 oral, high-quality paper–presentation pairs from 994 candidates drawn from six major computer vision and machine learning conferences between 2022 and 2025.
- Paper-Slide Datasets: Its author-prepared presentations provide a human reference standard for measuring how closely generated slides approximate expert preparation.
3 ArcDeck: Narrative-Driven Slide Generation
ArcDeck reconstructs a paper’s narrative structure before generating slides. Its discourse, commitment, refinement, and rendering stages produce a structured outline and visually refined deck, while ArcBench provides a curated evaluation set.
- Preprocessing: ArcDeck preprocesses papers into structured textual and visual assets, including extracted figures, tables, captions, sizes, and citation mappings.The preprocessing stage converts paper text into Markdown and stores visual assets for downstream agents.
- Narrative-Driven Outline Generation: The Discourse Parser builds hierarchical RST trees whose paragraph-level leaves and rhetorical relations guide content grouping and narrative ordering.The trees model dependencies between text segments before slide planning.
- Narrative-Driven Outline Generation: The Commitment Builder combines paper content with target audience and presentation duration to define a global plan covering the snapshot, thesis, prerequisites, narrative spine, and section plan.This global commitment supplies high-level guidance for subsequent refinement.
- Narrative-Driven Outline Generation: A Narrative Refinement Loop uses a planner, critic, judge, and reviser to iteratively improve the slide outline until approval or three refinement cycles.The finalized outline is forwarded in JSON format to slide generation.
- Slide Generation: Slide generation maps outline content to figures, tables, and layouts, then the Aesthetics Refiner adjusts content density, figure matching, and formatting.The Slide Deck Constructor creates the draft deck, while the refiner polishes its presentation.
- ArcBench: ArcBench is curated from 994 candidate paper–presentation pairs into 100 pairs selected for oral presentations and articles containing at least three figures and three tables.The benchmark draws from six major computer vision and machine learning conferences between 2022 and 2025.
4 Experiments
ArcDeck is evaluated against multiple generation systems using content, narrative, textual, visual, and pairwise assessments. Across these evaluations, it generally leads in narrative flow and content coverage, competes strongly on text quality, and improves visual presentation through refinement.
- Experimental Setup: The experiments compare ArcDeck with prompt-based, multi-agent, and human-prepared presentation references under fixed slide themes.The evaluated systems include HTML, Paper2Poster, PPTAgent, and SlideGen.
- Experimental Setup: Evaluation measures content coverage, narrative coherence, textual quality, and visual quality using quizzes, VLM judges, automated metrics, and pairwise preferences.The protocols include ROUGE-L, perplexity, rubric-based scoring, and comparisons against baselines and author-prepared slides.
- Narrative Flow: ArcDeck achieves the highest overall pairwise win rate for narrative flow and the highest or second-highest Narrative Flow scores across generation backbones and judges.Its advantage is especially pronounced over systems whose outlines mirror paper headings without rhetorical restructuring.
- Coverage of Content: 6.65 on Hard and 3.43 on Depth are ArcDeck’s gains over the next-best baseline under Qwen-3-VL.ArcDeck also leads or remains competitive on Story and Visuals, indicating preservation of narrative, conceptual, detailed, and visual content.
- Textual & Visual Quality: ArcDeck attains the highest ROUGE-L under GPT-4o and Qwen-3-VL while remaining competitive under GPT-5, with fluent slide text across models.The authors note that slide writing favors concise bullet-style summaries over direct sentence reuse.
- Analysis and Discussion: ArcDeck produces clearer narrative progression than baselines in qualitative comparisons and achieves the highest average selection rate against author-prepared slides.The qualitative examples show a progression from problem context to prior work and then the proposed method.
- Ablation Study & Token Analysis: Removing the Discourse Parser yields condensed, less structured slides, while removing the Commitment Builder disorganizes slide ordering.The full pipeline instead progresses from existing methods to the method overview and detailed methodology.
5 Conclusion
ArcDeck frames paper-to-slide generation as structured narrative reconstruction, using discourse parsing, global commitments, and iterative refinement to improve coherence and alignment. ArcBench provides a paired benchmark for evaluating content, narrative coherence, and visual quality.
- ArcDeck models discourse structure through RST-based parsing to guide slide generation.
- A global commitment preserves the paper’s high-level intent during presentation generation.
- A multi-agent critique–judge–revise loop refines outlines for improved narrative coherence and structural alignment.
- ArcBench is a curated paper–slide dataset evaluating content, narrative coherence, and visual quality.
- Experimental results show explicit discourse modeling improves narrative structure and information retention while remaining competitive in textual and visual quality.
A.1 Human Evaluation of Slide Deck Content and Narrative Flow
Human evaluation compared generated slide decks on content quality and narrative flow. ArcDeck achieved the best average ranking across the evaluated papers, indicating strongest overall coherence among the compared frameworks.
- 25 undergraduate or MS/PhD students participated in the human evaluation.Participants selected the research area they knew best from three options before ranking decks.
- Participants ranked slide decks according to content quality and narrative flow.The evaluation used a custom web interface for inspecting decks and submitting rankings.
- ArcDeck achieved the best average ranking across the evaluated papers.The reported result indicates the strongest narrative flow and overall content coherence among compared baseline frameworks.
A.2 Analysis of Global Commitment Conditioning
Global Commitment conditioning adapts generated slides to presentation duration and target audience. Shorter presentations condense content, while audience-specific decks vary in detail, terminology, and abstraction.
- Global Commitment conditioning was analyzed for target audience and presentation duration.
- Presentation duration: 5-minute presentations condense multiple detailed slides into fewer slides and substantially reduce textual content relative to 20-minute presentations.The shorter format preserves the most significant qualitative content to satisfy the time constraint.
- Target audience: General Public decks emphasize accessible high-level information, while Research Scientists decks provide more detailed method, experiment, and result content.
- Target audience: Terminology becomes less technical for the General Public and more domain-specific for Research Scientists.
A.3 Analysis of Discourse Parser
ArcDeck’s discourse analysis examines relation patterns and structural properties of generated discourse trees, while ablations and iterative comparisons assess their contribution to narrative flow. The broader evaluation materials document the benchmark and human-ranking setup used to assess these outputs.
- Discourse relation distribution: Elaboration is the most frequent discourse relation across section groups, especially in Methodology and Evaluation.Explanation is prominent in Introduction and Methodology, Context in Introduction and Related Work, and Evaluation in Evaluation and Conclusion.
- Tree properties: Generated discourse trees exhibit the near-linear relationship between EDU count and group count expected in RST theory.This relationship is presented as supporting the soundness of the discourse parser.
- Ablation study: Removing the Discourse Parser, Global Commitment, or Narrative Refinement Loop was evaluated through VLM-as-Judge and A/B testing against the full ArcDeck system.
- Narrative refinement: Across five papers, the narrative refinement loop iteratively improved outline narrative flow, with ArcDeck Final overall winning and two iterations outperforming one iteration in win rate.
- Evaluation criteria: Human evaluation ranks decks on concrete opening, logical progression, cross-slide coherence, and technical depth.
- ArcBench construction: ArcBench comprises a 994-pair pool and a curated 100-pair benchmark drawn from academic paper–slide pairs.The pool spans six major AI/CV venues, while the curated benchmark applies filtering criteria including oral presentations and visual-material requirements.
B.3 Benchmark Statistics.
ArcBench spans diverse research topics and content structures, supporting evaluation of paper-to-slide generation across varied academic domains.
- ArcBench covers different topic clusters across computer vision and machine learning domains.
- The benchmark’s per-venue and per-year breakdown is reported for both splits.
- ArcBench is compared with prior paper-to-slide datasets, including DOC2PPT, SciDuet, Zenodo10K, and SLIDESBENCH.
C.1 Generation Pipeline Implementation
ArcDeck implements narrative-driven outline generation, multi-dimensional VLM evaluation, pairwise comparison, and quiz-based coverage assessment.
- Narrative-driven outline generation: Discourse trees and Global Commitments are stored in JSON and Markdown formats and supplied to agents during outline generation.
- Slide construction: The Slide Deck Constructor matches figures and tables to slide content using captions, while the Aesthetic Refiner improves visuals, formatting, and sparse content.
- VLM evaluation: VLM judges score presentations across Text Quality, Narrative Flow, Visual Layout, and Visual Thematic dimensions using checklist-based criteria.
- Pairwise evaluation: Pairwise evaluation compares narrative flow and overall quality, with results reported as majority votes across iterations.
- Quiz evaluation: Quiz evaluation measures StoryFlow, Visual, Hard, and Depth coverage using questions answered from slide text or rendered images.
- Cross-judge analysis: Across metrics and evaluators, correlations reflect sensitivity to the underlying model’s generation strength.
D.3 Limitations
ArcDeck’s reported limitations include model-dependent variation and occasional slide-overflow failures, while its aesthetic refiner improves visual quality.
- Model dependence: Results may vary between GPT-4o and Qwen because of differences between open-source and closed-source model capabilities.
- Aesthetic refinement: The Aesthetic Refiner consistently improves Visual Layout and Visual Thematic metrics, with refined slides preferred in 75% of pairwise comparisons.
- Aesthetic refinement: Qualitative examples show added content and formatting improvements after refinement, including boldface and colored text.
- Qualitative comparisons: Compared with baselines, ArcDeck produces more comprehensive slides with improved narrative structure and stronger prioritization of technical sections.
E.4 ArcDeck Results Across Different Slide Themes
ArcDeck preserves slide organization and content placement across user-defined themes, output formats, and demonstrated Physics and Biology documents.
- Theme flexibility: Different user-defined themes preserve ArcDeck’s slide layout and content structure.
- Output flexibility: ArcDeck supports JavaScript and LaTeX Beamer rendering without modifying the core generation process.
- Domain examples: Additional examples demonstrate ArcDeck-generated slides for Physics and Biology documents.