Narrative-Driven Paper-to-Slide Generation via ArcDeck

• Effective paper-to-slide generation requires selecting, organizing, and visually presenting dense technical content while preserving clear narrative flow.The central challenge is identifying how concepts are introduced, explained, and related, then mapping that structure into slides alongside the paper’s results, graphics, and other content.Source: §1 · pp. 1–2
• Prior approaches use whole-paper generation, section-level processing, or multi-agent outlining, but struggle with long contexts, fractured narratives, or fine-grained structural coherence.These limitations reflect the absence of effective structural modeling for preserving rhetorical dependencies and a unified storyline.Source: §1 · p. 2
• ArcDeck formulates paper-to-slide generation as structured narrative reconstruction using an RST-inspired Discourse Parser, Global Commitment, and iterative Narrative Refinement Loop.The discourse tree preserves local rhetorical relationships, persistent commitments maintain paper-level intent and consistency, and critique–judge–revise cycles improve the outline before slide rendering.Source: §1 · p. 2
• ArcDeck’s closed-loop multi-agent refinement supports coherent structural planning, consistent narrative flow across the slide sequence, and presentation customization such as length.The framework balances content organization with visual formatting rather than relying on one-shot outlining.Source: §1 · p. 2
• ArcBench is an open-source paper–slide benchmark from diverse academic venues that standardizes evaluation of text coverage, narrative coherence, visual quality, and content retention.It is curated from author-prepared presentations and provides a human reference standard for evaluating slide-generation quality.Source: §1 · p. 2

2.1 Agentic Slide Generation.

• ArcDeck introduces the first multi-agent narrative refinement loop and global commitment document, supporting coherent slide generation and customization for different presentation lengths.Prior systems explored multi-agent code refinement, specialized agents, intermediate representations, and template editing, but did not combine these mechanisms with explicit global narrative commitments.Source: §2.1 · p. 3

2.2 Rhetorical Structure Theory (RST).

• ArcDeck is the first to apply RST parsing to slide generation, using unfine-tuned state-of-the-art LLMs and discourse trees to guide coherent narrative content.Prior RST research developed top-down, bottom-up, neural, and text-generation parsers, and used predicted trees to guide other text-generation tasks.Source: §2.2 · p. 3

2.3 Paper-Slide Datasets.

• ArcBench curates 100 high-quality oral paper-slide pairs from 994 CV/ML conference pairs, addressing prior datasets’ domain, presentation-quality, and public-release limitations.Human references are validated against generated baselines, revealing a clear quality gap that motivates the benchmark; all ArcBench assets will be released.Source: §2.3 · p. 3

3.1 Preprocessing

• Preprocessing converts the raw PDF into Markdown, extracts figures and tables with captions and sizes, and maps short-form in-text citations to full numbered references.Docling and Marker perform parsing while excluding the bibliography and appendix to focus downstream generation on narrative content.Source: §3.1 · p. 3, 3) Slide Generation · p. 4

3.2 Narrative-Driven Outline Generation

• ArcDeck’s outline-generation stage combines an RST discourse tree, a global commitment, and a critique–judge–revise loop to preserve rhetorical dependencies and narrative flow.The commitment encodes the snapshot, core content, talk contract, narrative spine, and section plan; refinement evaluates alignment, flow, balance, coherence, redundancy, and missing content, with at most three cycles.Source: §3.2 · pp. 4–6

3.3 Slide Generation

• The slide-generation pipeline selects visuals and templates for each slide, generates narrative-conditioned text, then refines visual grounding, content density, color, and emphasis before rendering the final deck.The Constructor chooses among 14 layouts using text volume and visual-element properties, while the Aesthetics Refiner performs four targeted polishing operations and outputs a python-pptx presentation.Source: §3.3 · pp. 6–7

3.4 ArcBench Dataset

• ArcBench curates 100 author-prepared paper–slide pairs from six 2022–2025 AI venues, filtered for oral presentations and papers containing at least 3 figures and 3 tables.The released resource includes both the 100-pair evaluation subset and the full pool of 994 candidate pairs.Source: §3.4 · p. 7

4.1 Experimental Setup

• The experiments compare ArcDeck with four systems and author-prepared slides across GPT-4o, GPT-5, and Qwen3-VL-32B on 100 ArcBench paper–slide pairs under fixed themes.Generated slides use a 13.33×7.5-inch format, and ArcDeck’s reference template contains 14 layout configurations.Source: §4.1 · pp. 7–8

4.3 Analysis and Discussion

• ArcDeck achieves the highest overall narrative-flow win rate against all baselines across open- and closed-source judges, with the strongest advantages over PPTAgent and HTML.Its advantage is attributed to discourse-aware outline generation and the Narrative Refinement Loop; VLM-as-Judge scores also place ArcDeck first or second across backbones and judges.Source: §4.3 · p. 9, Table 4 · p. 9
• ArcDeck achieves the highest or second-highest quiz scores across content categories, exceeding the next-best baseline by 6.65 on Hard and 3.43 on Depth under Qwen-3-VL.It leads Story across all three models, performs strongly on Visuals, and uses a hierarchical discourse tree to preserve conceptual explanations and methodological nuances.Source: §4.3 · pp. 9–10, Table 3 · p. 9
• ArcDeck achieves the highest ROUGE-L under GPT-4o and Qwen-3-VL while remaining competitive under GPT-5, with fluent slide text across models.Perplexity results indicate fluency, while metric differences reflect the concise bullet style favored by slides.Source: Table 5 · p. 9, §4.3 · p. 10
• ArcDeck attains the highest average selection rate against author-prepared slides, indicating the strongest holistic performance among automated methods.The comparison jointly evaluates content organization, visual design, information delivery, and slide structure.Source: §4.3 · pp. 9–10, Table 4 · p. 9
• Removing the Discourse Parser omits key method details, while removing the Commitment Builder disorganizes slide ordering; the full pipeline restores coherent progression.The full sequence runs from existing methods to the method overview and detailed methodology, while token usage is analyzed across agents.Source: §4.3 · pp. 10–11, Figure 8 · p. 11, Figure 9 · p. 11

5 Conclusion

• ArcDeck formulates automatic paper-to-slide generation as a structured narrative reconstruction problem.The framework is explicitly narrative-driven rather than limited to direct text summarization.Source: §5 · p. 11
• RST-based discourse parsing models the paper’s structure and preserves its high-level intent through a global commitment.These structural priors guide the generation process toward structural alignment with the source paper.Source: §5 · p. 11
• A multi-agent critique–judge–revise loop iteratively refines presentation outlines before slide generation.Role-specific refinement is a central mechanism for improving narrative coherence.Source: §5 · p. 11
• ArcBench is a curated paper–slide paired dataset for evaluating content, narrative coherence, and visual quality.The benchmark supports systematic, multidimensional assessment of generated slide decks.Source: §5 · p. 11
• Explicit discourse modeling substantially improves generated presentations’ narrative structure and information retention.The conclusion attributes these improvements to making discourse structure an explicit generation prior.Source: §5 · p. 11
• ArcDeck remains competitive in textual and visual quality while improving narrative structure and information retention.The reported results indicate that the structural gains do not come at the expense of textual or visual quality.Source: §5 · p. 11

Human Evaluation Setup

• Twenty-five undergrad or MS/PhD students ranked ArcDeck, SlideGen, and PPTAgent on five oral papers within their most familiar research area.The study covered Generative Models, Vision-Language/Multimodal Learning, and Computer Vision Core, using a custom web interface for inspection and ranking.Source: §A.1 · p. 16

Human Evaluation Results

• ArcDeck achieves the best average ranking across evaluated papers, indicating the strongest narrative flow and overall content coherence among compared baseline frameworks.Participants ranked slide decks by content quality and narrative flow, with Rank 1 receiving 3 points, Rank 2 receiving 2, and Rank 3 receiving 1.Source: §A.1 · p. 16, Table 6 · p. 16

Presentation Duration

• Global Commitment conditioning condenses information for shorter presentations while preserving key points and retains more detailed content for longer presentations.In the 5-minute version, multiple slides may be combined, text reduced, and only the most significant quantitative result retained; the 20-minute version includes more detail and alternative configurations.Source: §A.2 · p. 16, Figure 10 · p. 16

Target Audience

• For General Public audiences, ArcDeck uses accessible language and high-level explanations, whereas Research Scientist decks provide technical terminology and detailed methods, experiments, and results.The two audience settings also differ in abstraction: General Public presentations remain conceptual, while Research Scientist presentations are more technically grounded.Source: §A.2 · pp. 16–17, Figure 11 · p. 17

A.3 Analysis of Discourse Parser

• Generated discourse trees exhibit section-specific relation patterns, with Elaboration dominant overall and Explanation, Context, and Evaluation concentrated in structurally appropriate paper sections.Elaboration is especially frequent in Methodology and Evaluation; Explanation in Introduction and Methodology; Context in Introduction and Related Work; and Evaluation in Evaluation and Conclusion.Source: §A.3 · p. 17, Figure 12 · p. 18
• The generated trees show the expected near-linear relationship between EDU and group counts, while tree height and average leaf depth also correlate with paragraph count.These relationships support the soundness of the discourse parser and alignment with academic-paper discourse structure.Source: §A.3 · p. 17, Figure 13 · p. 18

A.4 Ablation Study of Narrative-Driven Outline Generation

• Removing the Discourse Parser, Global Commitment, or Narrative Refinement Loop significantly decreases Narrative Flow, while full ArcDeck wins pairwise comparisons against each ablated variant.The evaluation uses VLM-as-Judge scores and A/B testing against the full system.Source: §A.4 · p. 18, Table 7 · p. 18

A.5 Effect of the Narrative Refinement Loop

• ArcDeck Final is the overall winner after iterative narrative refinement, and two iterations outperform one iteration in win rate across five papers.The study compares slide decks generated from outlines at each of three refinement iterations.Source: §A.5 · p. 19, Table 8 · p. 19

B ArcBench: Additional Information and Statistics

• ArcBench is built in two stages: collecting paper-slide pairs from major AI/CV venues and applying principled filters to produce a high-quality benchmark, summarized across a 994-pair pool and 100-pair subset.Table 9 and Figure 15 summarize venue, year, paper, and slide-deck statistics for both datasets.Source: B ArcBench: Additional Information and… · p. 21, Table 9 · p. 21, Figure 15 · p. 21

B.1 Full Dataset Collection.

• The full ArcBench pool contains 994 paper-slide pairs from CVPR, ECCV, ICCV, ICML, ICLR, and NeurIPS, spanning 2022–2025 with 295 oral and 699 poster presentations.The collection was drawn from official proceedings and presentation archives.Source: §B.1 · p. 21

B.2 Filtering to the 100-Pair Benchmark.

• The curated 100-pair benchmark retains only oral presentations with at least 3 figures and 3 tables per paper, ensuring visual material and quantitative content for evaluation.These sequential criteria are intended to support figure fidelity, information coverage, and reliable human reference standards.Source: §B.2 · p. 21

B.3 Benchmark Statistics.

• ArcBench spans diverse CV/ML topic clusters and contains substantial visual and quantitative content density across its venue and year splits.This diversity challenges generation methods with varied content structures, notation conventions, and domain-specific visual elements.Source: §B.3 · p. 22

B.4 Comparison with Prior Datasets.

• ArcBench focuses exclusively on oral, author-prepared presentations from six top-tier AI/CV venues and applies figure-and-table content-density thresholds.These design choices distinguish it from datasets such as DOC2PPT, SciDuet, Zenodo10K, SLIDESBENCH, and SlideGen [11].Source: §B.4 · p. 22, Table 10 · p. 22
• ArcBench provides both a curated 100-pair benchmark and a full 994-pair pool for community use.The benchmark is designed for the academic paper-to-slide setting, unlike datasets targeting broader presentations or individual slide generation.Source: §B.4 · p. 22

Narrative-Driven Outline Generation.

• The pipeline stores discourse relations, global commitments, agent outputs, and slide outlines in structured JSON or Markdown formats to preserve narrative structure across agents.Slide outlines retain section and slide identifiers, titles, and paragraph assignments for each slide.Source: Narrative-Driven Outline Generation · p. 23
• The Slide Deck Constructor matches figures and tables to slide content, selects layouts and citations, and produces structured plans for final .pptx generation.Reference assets are matched during rendering, with short-form citations added to slide corners.Source: Narrative-Driven Outline Generation · p. 23
• The Aesthetic Refiner adds or improves visual elements, formatting, and sparse content before producing the refined slide plan.Text annotations such as textbf and textcolor are converted into python-pptx formatting instructions.Source: Narrative-Driven Outline Generation · p. 23

C.2.1 Judge Models & Iteration Protocol.

• Evaluation uses Qwen3-VL-30B-A3B-Instruct for 11 iterations per baseline and GPT-5 for 3 iterations to assess generated presentations.The evaluation framework combines VLM judging, pairwise comparison, and quiz-based coverage.Source: §C.2.1 · p. 23

C.2.2 VLM-as-Judge.

• VLM-as-Judge scores presentations from 0–10 on text quality, narrative flow, visual layout, and visual thematic criteria using strict binary checklists.These dimensions respectively assess technical substance, logical ordering, rendering quality, and whether visuals communicate scientific content.Source: §C.2.2 · p. 24

C.2.3 Pairwise Comparison.

• Pairwise comparison evaluates ArcDeck against baselines for narrative flow and compares Author-Prepared presentations with all methods for overall quality using majority votes across iterations.The criteria include technical substance, visual quality, presentation effectiveness, and logical story structure.Source: §C.2.3 · p. 24

C.2.4 Quiz-Based Coverage.

• Quiz-based coverage generates 100 multiple-choice questions per paper across StoryFlow, Visual, Hard, and Depth dimensions, with answers scored as accuracy percentages from slide text or images.References, acknowledgements, and appendices are removed during question generation, while Visual questions use rendered images and the other types use extracted slide text.Source: §C.2.4 · pp. 24–25

D.1 Cross-Judge Model Agreement.

• Cross-judge evaluation found strong agreement for text quality and narrative flow, but Qwen3’s saturated visual-layout scores caused low agreement with GPT-5 for Ours.TQ correlated at r = 0.874 and NF at r = 0.831 for HTML; Ours had VL correlation r = 0.095 because Qwen3 assigned over 75% maximum scores with std=0.027 versus GPT-5’s std=0.103.Source: Figure 17 · p. 26, §D.1 · p. 27

D.3 Limitations

• ArcDeck may produce different results with GPT-4o and Qwen because their capabilities differ, so optimal performance may require model-specific prompt tuning.The limitation concerns adapting prompts to account for differences between open-source and closed-source models.Source: §D.3 · p. 27

E.1 Template Layouts for Slide Generation

• ArcDeck uses 14 reference layouts spanning text, image, and table combinations, with the Slide Constructor selecting layouts according to content, matched figures, and sizes.These layouts are intended to support coherent organization and visually balanced slides.Source: §E.1 · p. 28, Figure 19 · p. 28

E.2 Impact of the Aesthetic Refiner

• 75% of pairwise comparisons preferred slides generated with the Aesthetic Refiner, which consistently improves VLM-as-Judge visual layout and thematic quality.The refiner also adds textual content and formatting such as boldface and colored text, improving readability and visual aesthetics.Source: §E.2 · p. 28, Table 11 · p. 29, Figure 20 · p. 29

E.3 Additional Qualitative Comparison

• Paper2Poster gives Methods and Experiments coverage similar to introductory sections, limiting emphasis on technical contributions.Its slides follow section-wise summarization of the source document.Source: §E.3 · p. 28
• ArcDeck generates more informative slides with richer explanations and clearer key contributions by prioritizing Methods and Experiments through its global content commitment mechanism.Paper2Poster, PPTAgent, HTML-4o, and SlideGen either distribute emphasis poorly, omit visuals, misalign layouts, or provide overly concise content.Source: §E.3 · pp. 28–29
• PPTAgent provides more detailed slides than Paper2Poster but often fails to prioritize Methods, combines Methods and Experiments, and sometimes selects layouts misaligned with content.These issues are illustrated by the combined Methods-and-Experiments slide and examples such as Slides 8 and 12.Source: §E.3 · p. 29
• HTML-4o generates subsection-wise content but lacks visual elements, whereas SlideGen produces visually appealing yet overly concise slides with limited Methods and Experiments coverage.ArcDeck instead provides richer textual explanations and clearer presentations of key contributions.Source: §E.3 · p. 29

E.4 ArcDeck Results Across Different Slide Themes

• ArcDeck supports multiple user-defined slide themes while preserving the underlying layout and content structure, enabling varied visual styles without disrupting slide organization.Additional qualitative results demonstrate this theme flexibility across several aesthetically pleasing designs.Source: §E.4 · pp. 29–30, Figure 26 · p. 34, Figure 27 · p. 35
• The theme-flexible rendering further enhances the visual appearance and overall aesthetic quality of generated slides while maintaining consistent content placement.The supplied examples illustrate aesthetic enhancement under fixed layout structure rather than changes to the slide organization.Source: §E.4 · pp. 29–30

E.5 ArcDeck Extensions to Alternative Rendering Formats

• ArcDeck generates presentations in JavaScript and LaTeX, including PPTxGenJS and LaTeX Beamer pipelines, without modifying its core generation process.Figures 29 and 30 show examples rendered in these alternative output formats.Source: §E.5 · p. 30, Figure 29 · p. 37, Figure 30 · p. 38
• Support for multiple output formats makes ArcDeck adaptable across different slide-generation and rendering pipelines while retaining visually appealing presentation outputs.The demonstrated formats are JavaScript and LaTeX Beamer.Source: §E.5 · p. 30, Figure 29 · p. 37, Figure 30 · p. 38

E.6 ArcDeck Results on Papers from Diverse Scientific Domains

• Although experiments primarily use AI papers, ArcDeck is designed to generalize to diverse structured-document domains without domain-specific modifications.The framework is explicitly described as not being limited to a specific document domain.Source: §E.6 · p. 30
• ArcDeck generates slides for papers from both Physics and Biology, demonstrating that its document-to-slide capability is not restricted to AI-related content.Additional qualitative results show examples for one Physics document and one Biology document.Source: §E.6 · p. 30, Figure 31 · p. 39, Figure 32 · p. 39