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.
Figures & tables
Figure 1: Framework Comparison. (a) Prior methods rely on raw-text summarization , isolated section-level organization , or direct outline extraction , yielding condensed text without narrative guidance. (b) In contrast, ArcDeck explicitly models discourse structure and establishes global commitments to produce narrative-driven outlines through an iterative multi-agent refinement loop.
Figure 2: ArcDeck Overview. 1) Preprocessing extracts textual and visual assets from the source paper; 2) Narrative-Driven Outline Generation combines discourse parsing and global commitment to guide the Narrative Refinement Loop , producing a structured Slide Outline; and 3) Slide Generation constructs the initial deck and refines its aesthetic layout to yield the final presentation. The numbered indices of each agent, e.g . , 1 , denote their temporal execution order, while the icons at the bottom of each agent box indicate the specific structural inputs they receive.
Relation
Explanation
Type
Elaboration
Adds detail, examples, or implementation specifics.
NS
Explanation
Clarifies why or how a nucleus claim holds.
NS
Context
Background or definitions needed to interpret the nucleus.
NS
Purpose
Goal or motivation associated with the nucleus.
NS
Evaluation
Assessment, strength, or limitation of the nucleus.
NS
Organization
Roadmap or meta-structural framing.
NS
Table 1: RST Relation Taxonomy. Definitions of the rhetorical relations used to construct our discourse trees (NS: Nucleus-Satellite; MN: Multinuclear).
Figure 3: Example Discourse Trees from Secs. 1 & 3 of this paper. Leaf nodes correspond to elementary discourse units (EDUs, corresponding to paragraphs), while internal nodes and arrows denote rhetorical relations that recursively group these spans.
Figure 4: Global Commitment Example. The global commitment establishes the high-level narrative progression of the presentation through five key components: a snapshot, core content, a talk contract, a narrative spine, and a light section plan.
Figure 5: Slide Generation Stage. Slide Deck Constructor 6 handles asset mapping, layout selection, and text generation, while the Aesthetics Refiner 7 polishes the deck via figure matching and content formatting.
Figure 6: ArcBench Statistics. Overview of our curated 100-pair dataset, showing the distribution of research topics across six major CV & ML venues, publication years, visual density, and author-prepared presentation lengths.
Category
Specific Metric
Input
Output
High-Level Idea
Story
Pres. Text & 25 MCQs
Storyline capture and high-level information retention.
Visuals
Slide Images & 25 MCQs
Information conveyed through visual elements and slide layout.
Hard
Pres. Text & 25 MCQs
Conceptual understanding of methodology and limitations.
VLM-based Q/A Quiz
Depth
Pres. Text & 25 MCQs
Accuracy (%)
Coverage of important technical/implementation details.
Text Quality (TQ)
Technical substance and specificity of the text.
Narrative Flow (NF)
Pres. Text
Logical progression and comprehensive section coverage.
Table 2: Summary of Evaluation Metrics
VLM-based Q/A Quiz (Open) ↑
VLM-based Q/A Quiz (Closed) ↑
VLM-as-Judge (Open) ↑
VLM-as-Judge (Closed) ↑
Method
Story
Visuals
Hard
Depth
Story
Visuals
Hard
Depth
TQ
NF
VL
VT
TQ
NF
VL
VT
Generation Model: GPT-5
HTML
95.16
64.13
81.83
69.73
95.60
38.40
80.00
82.20
86.11
89.06
95.22
76.77
58.50
61.50
63.67
10.00
P2Poster
95.22
65.51
76.70
67.67
87.00
27.00
66.40
64.60
62.20
70.55
97.08
91.35
35.17
41.67
72.00
78.50
PPTAgent
92.96
65.08
69.49
65.67
75.00
12.80
32.00
36.20
35.95
55.94
84.86
81.95
16.50
37.67
71.33
57.50
SlideGen
95.26
68.37
78.39
69.43
97.60
35.80
86.00
82.40
69.02
88.18
97.36
95.28
36.17
56.67
88.33
84.83
Table 3: VLM-Based Evaluation. The best value in each generation model is highlighted in green , and runner-up is underlined.
Table 4: VLM Pairwise Preference (A/B Test). Holistic evaluation of generated presentations using two comparison protocols: assessing narrative flow against baselines (left), and overall quality against Author-Prepared (AP) slides (right). To mitigate single-model bias, both closed- (GPT-5) and open-source (Qwen3-VL) judges are used.
Generation Model
GPT-5
GPT-4o
Qwen3
Automated Metrics
R-L ↑
PPL ↓
R-L ↑
PPL ↓
R-L ↑
PPL ↓
HTML
64.7
39.90
33.6
31.51
108.5
31.08
Paper2Poster
71.3
21.04
53.8
22.07
76.8
16.96
PPTAgent
33.6
91.43
34.6
57.21
64.6
23.33
SlideGen
55.1
80.46
45.2
43.84
83.9
23.09
ArcDeck
57.1
46.01
84.8
33.47
156.0
16.95
Table 5: Automated Metrics Evaluation. The best value in each column is highlighted in green , and runner-up is underlined. R-L values are scaled by 10−3 .
Figure 7: Qualitative Comparison. Unlike baselines, ArcDeck shows clear narrative flow instead of mirroring paper section structure and avoids cross-slide content overlap.
Figure 8: Qualitative Ablation Study. Comparison between ArcDeck and variants without the Discourse Parser (DP) or Commitment Builder (CB).
Figure 9: Token Usage Analysis.
Appendix figures & tables60 assets
Supplementary material from the paper’s appendix.
Appendix
Human Evaluation Topic
Method
V-L / MM
Gen. Models
CV Core
Overall
PPTAgent
1.03
1.13
1.20
1.10
SlideGen
2.42
2.23
2.14
2.30
ArcDeck
2.55
2.63
2.66
2.60
Appendix
Table 6: Human Evaluation Results. Participants ranked slide decks based on content quality and narrative flow, where Rank 1 = 3 points, Rank 2 = 2 points, and Rank 3 = 1 point. The best value in each topic is highlighted in green , and runner-up is underlined.
Figure 10: Qualitative Comparison of ArcDeck-generated Slides under Different Presentation Durations. 20-minute presentation includes more detailed content, while the 5-minute presentation preserves the key points of the paper.
Figure 11: Qualitative Comparison of ArcDeck-generated Slides under Different Target Audiences. The figure illustrates how ArcDeck adapts content selection, terminology, and level of technical detail according to the specified target audience.
Figure 12: Distribution of Discourse Tree Relations across Section Groups. Stacked bars represent the relative proportion of each discourse relation within each section group.
Figure 13: Relationship of Generated Discourse-Tree Height, Average Leaf Depth, and Group Count per Paragraph to the Number of Paragraphs. Pearson correlation coefficients are shown in the top-left corner of each figure.
VLM-as-Judge ↑
Pairwise Comparison
Method
NF
ArcDeck Win Rate %
ArcDeck w/o Discourse Parser
7.50
89.1
ArcDeck w/o Global Comm.
7.52
94.5
ArcDeck w/o Narr. Ref. Loop
8.68
61.8
ArcDeck
9.70
–
Appendix
Table 7: Narrative-Driven Outline Generation Ablation Study. Narrative Flow (NF) scores for different ablated versions, along with the full ArcDeck model’s narrative flow pairwise comparison win rate against these ablated versions, are shown.
VLM-as-Judge ↑
Pairwise Comparison
NRL Iterations #
NF
NF Win Rate vs. 1st It.
1 iteration
4.18
–
2 iterations
7.00
56.4 %
3 iterations (Final)
7.34
100 %
Appendix
Table 8: Narrative Refinement Loop (NRL) Improvement Study. Narrative Flow (NF) scores for different numbers of iterations, together with the win rate relative to the first iteration, are shown.
Figure 14: Human Evaluation Interface. Participants were shown slide decks generated by different methods and asked to rank them based on content quality and narrative flow.
Venue
2022
2023
2024
2025
Total
CVPR
—
—
52
—
1
176
—
1
228
—
2
ECCV
—
—
222
—
27
—
222
—
27
ICCV
—
—
—
296
—
4
296
—
4
ICLR
—
—
12
—
—
51
—
27
63
—
27
ICML
52
—
—
38
—
3
40
—
17
22
—
8
152
—
28
NeurIPS
—
14
—
4
10
—
4
9
—
4
33
—
12
Appendix
Table 9: ArcBench Venue and Year Distribution. Each cell: pool — bench. Dashes indicate no coverage.
Figure 15: Key Statistics of Papers and Slide Decks in ArcBench . We report distributions for the full paper pool and the curated benchmark, including the number of figures, tables, character counts, and Oral papers.
Figure 16: ArcBench Topic Distribution. We report the distribution of papers across research topics.
Dataset
Year
Pairs
Venues
Domain
Type
Pub.
DOC2PPT
2022
5,873
General academic
Multi-domain
Oral & Poster
✓
SciDuet
2021
1,088
ICML, NeurIPS, ACL
NLP / ML
Oral & Poster
✓
Zenodo10K
2025
10K
—
Multi-domain
—
✓
SLIDESBENCH †
2025
7,585
—
General
—
✓
SlideGen
2025
200
ICLR, ICML, NeurIPS
AI / ML
Oral-only
✗
ArcBench
2025
100
6 venues ‡
CV / ML
Oral-only
✓
Appendix
Table 10: Comparison of ArcBench with Prior Paper-to-Slide Datasets. † Generates individual slides from NL instructions rather than full papers. ‡ CVPR, ECCV, ICCV, ICML, ICLR, NeurIPS.
Figure 17: Pearson Correlation of Normalized Scores Across Methods. (a,b) Correlation between presentations generated by Qwen3-VL-32B and GPT-5, evaluated by Qwen3 and GPT-5, respectively. (c) Correlation between Qwen3 and GPT-5 evaluator scores across all generated presentations.
Figure 18: ArcDeck Failure Case.
Figure 19: Template Slide Layout. used by ArcDeck
Figure 20: Aesthetic Refiner Ablation. Comparison of slides generated before and after applying the refiner agent.
VLM-as-Judge ↑
Pairwise Comparison
Method
VL
VT
Win Rate %
Generation Model: GPT-5
ArcDeck wo Refiner
9.62
9.37
25
ArcDeck
9.75
9.5
75
Appendix
Table 11: VLM-as-Judge Evaluation (Open Setting) and Pairwise Visual Aesthetic Comparison (Open Setting). The best value is highlighted in green , and runner-up is underlined.
Figure 21: Qualitative Comparison. Sample generated by ArcDeck -(4o)
Figure 22: Qualitative Comparison. Sample generated by SlideGen-4o
Figure 23: Qualitative Comparison. Sample generated by PPTAgent-4o
Figure 24: Qualitative Comparison. Sample generated by Paper2Poster-4o
Figure 25: Qualitative Comparison. Sample generated by HTML-4o
Figure 26: Additional Qualitative Results. generated by ArcDeck GPT-5 backbone.
Figure 27: Additional Qualitative Results. generated by ArcDeck GPT-5 backbone.
Figure 28: Additional Qualitative Results. generated by ArcDeck GPT4o backbone.
Figure 29: ArcDeck Extension. to JavaScript output format.
Figure 30: ArcDeck Extension. to Latex Beamer output format.
Figure 31: Additional Qualitative Results for diverse domain. Slides generated by ArcDeck (GPT-5 backbone) for a document from the Physics domain.
Figure 32: Additional Qualitative Results for diverse domain. Slides generated by ArcDeck (GPT-5 backbone) for a document from the Biology domain.
Figure 49: VLM as Judge - Narrative Flow Evaluation Prompt.
Figure 50: VLM as Judge - Text Quality Evaluation Prompt.
Figure 51: VLM as Judge - Visual Layout Evaluation Prompt.
Figure 52: VLM as Judge - Visual Thematic Evaluation Prompt.
Figure 53: Example Global Commitment generated for ArcDeck.
Figure 54: Example Global Commitment generated for Attention Is All You Need.
Figure 55: Example full Discourse Parser outputs in JSON format for introduction sections. In (a), the discourse tree for Section 1 of this paper is shown. In (b), the discourse tree for Section 1 of Attention Is All You Need is shown, together with snippets from the corresponding paragraphs.
Figure 56: Example Discourse Parser output in JSON format. Discourse tree for Section 3 (Model Architecture) of Attention Is All You Need, shown together with snippets from the corresponding paragraphs.
Figure 57: Example Slide Outline output in JSON format.
Figure 58: Example Narrative Critic output in JSON format.
Figure 59: Example Failing Narrative Judge output.
Figure 60: Example Passing Narrative Judge output.
Automatic academic paper-to-slide generation is inherently iterative, because creating an effective presentation requires repeated cycles of generation, critique, and revision. Recent multi-agent systems partially acknowledge this through internal critique-and-revise loops, while conversational approaches allow users to refine generated slide decks through dialog. However, these refinement processes either remain largely closed to the user or introduce feedback only after a complete deck has been produced, limiting the user's ability to participate in the iterative refinement of narrative flow, content allocation, and presentation emphasis. To address this gap, we introduce ConvDeck, a multi-agent pipeline for conversational paper-to-slide generation that distributes interaction across the pipeline through stage-specific loops, allowing users to iteratively refine both the presentation outline and the final slide deck at the stages where each kind of decision is made. These loops are driven by a refinement mechanism in which agents can think, speak, and act, enabling them to either directly apply edits or respond conversationally to clarify user feedback and discuss revision options. Our evaluation shows that stage-specific conversational feedback improves user-goal satisfaction while preserving narrative coherence, content quality, and visual presentation.
Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide--script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.
Ming Yang, Zhiwei Zhang, Jiahang Li +3
School of Data Science, Fudan University, Shanghai, China
Scientific presentations are more than summaries of research papers. They need to present the work in a coherent sequence, explain the main ideas clearly, and help the audience follow the presentation. We present SlideLab, a training-free multi-agent framework for generating scientific presentations from research papers. SlideLab first plans the presentation narrative, then builds and iteratively refines a shared slide deck using agents for content planning, visual generation, layout refinement, and grounding verification. In a blind human preference study, SlideLab was preferred over both open-source and commercial systems on 77% of papers while using roughly 4 times fewer inference tokens than the strongest open-source baseline. We also introduce ConfArena, an audience-oriented evaluation framework that simulates a conference room and assesses presentations slide by slide. ConfArena matches human system rankings and detects injected presentation problems, including falsified numbers, degraded figures, dropped slides, and shuffled slide order.