Official Implementation of Narrative-Driven Paper-to-Slide Generation via ArcDeck

Tarik Can Ozden^*1   Sachidanand VS^*1   Furkan Horoz^*1,2   Ozgur Kara¹   Junho Kim¹   James M. Rehg¹

¹ University of Illinois Urbana-Champaign    ² Middle East Technical University
^* Equal contribution

Quick Start • Pipeline • CLI Usage • Web UI • Themes • ArcDeck as Claude Skills • ArcBench • Citation

---

ArcDeck is an end-to-end slide generation framework that converts academic PDF papers into polished .pptx presentation slides. ArcDeck frames slide generation around the paper’s narrative structure instead of simple summarization. By combining narrative-driven outline generation with visually strong slide rendering, ArcDeck produces polished and engaging slide decks.

---

Pipeline Overview

Pipeline Details

ArcDeck operates in three main phases:

1. Preprocessing

• Converts the input PDF to Markdown using Docling (falls back to Marker if extraction is too short)
• Extracts an asset dictionary of figures, tables, and citations

2. Narrative-Driven Outline Generation

This is the core intellectual contribution — a multi-agent loop that plans slides using discourse structure:

Step	Agent	Role
1	Discourse Parser	Analyzes each section's rhetorical structure using RST, producing a binary discourse tree that captures how paragraphs relate (elaboration, contrast, evidence, etc.)
2	Commitment Builder	Generates a high-level commitments.md contract capturing the paper's thesis, key takeaways, narrative spine, and slide budget priorities
3	Slide Planner / Reviser	Groups paragraphs into slides using RST relations, then revises based on feedback
4	Narrative Critic	Critiques the slide plan against the commitment and presentation quality rubrics
5	Narrative Judge	Decides pass or revise — if revise, circles back the loop (up to 3 rounds)

3. Slide Generation

• Slide Deck Constructor — An LLM agent plans detailed layouts (template selection, content placement, bullet points) and renders them into .pptx via python-pptx or PptxGenJS
• Aesthetics Refiner — Matches extracted figures/tables to slides and applies visual styling

---

News

• [04/2026]: 🚀 We We have open-sourced the ArcDeck codebase!
• [04/2026]: ✨ The ArcBench benchmark has been released on HuggingFace!
• [04/2026]: 📜 Narrative-Driven Paper-to-Slide Generation via ArcDeck is now available on arXiv!

TODO

• Codex Plugin Release
• Codebase & Claude Plugin Release
• Benchmark Release

---

Quick Start

Install

pip install -r requirements.txt

Set API Keys

Create a .env file in the project root:

OPENAI_API_KEY=sk-...
# Or set other provider keys as needed

---

CLI Usage

Basic:

python -m slide_generation.arcdeck path/to/paper.pdf --model gpt-5

Full options:

python -m slide_generation.arcdeck path/to/paper.pdf \
    --model gpt-5 \
    --template 3 \
    --duration 20 \
    --audience researchers \
    --speaker_script

JS rendering (PptxGenJS):

curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.4/install.sh | bash
source ~/.bashrc

npm install pptxgenjs  # one-time setup

python -m slide_generation.arcdeck path/to/paper.pdf \
    --model gpt-5 \
    --use_js

CLI Flags

Flag	Description	Default
paper_path	Path to input PDF	(required)
--model	LLM model name	4o
--template	PPTX slide template (1–8)	1
--duration	Target talk length (minutes)	20
--audience	Target audience	researchers
--speaker_script	Generate speaker notes	off
--use_js	Generates slides using PptxGenJS Library	False
--js_design	PptxGenJS decoration layout (0–5)	0
--agentic_js	Agentic JS generation (LLM writes the full PptxGenJS script)	False

---

ArcDeck as Claude Skills

ArcDeck also includes a structured Claude Skills configuration for agentic paper-to-slides generation.

Skill Installation

1. Create the skill directory and symlinks:

   mkdir -p .claude/skills/arcdeck
   ln -s "$(pwd)/arcdeck-skill/SKILL.md" .claude/skills/arcdeck/SKILL.md
   ln -s "$(pwd)/arcdeck-skill/agents" .claude/skills/arcdeck/agents
   ln -s "$(pwd)/arcdeck-skill/prompts" .claude/skills/arcdeck/prompts
   ln -s "$(pwd)/arcdeck-skill/reference" .claude/skills/arcdeck/reference

2. Open Claude Code in the project directory:

   claude

3. Run the skill:

   /arcdeck paper.pdf
   

Invocation

/arcdeck <pdf_path> [--audience <type>] [--duration <minutes>] [--template <theme>] [--speaker-notes]

Defaults:

• audience=researchers
• duration=20
• template=ocean
• speaker-notes=off

Skill Templates

Template	Style	Title Font	Colors
ocean	Ocean Gradient — navy/teal/cyan, light content slides	Georgia	065A82 1C7293 0891B2
minimal	Clean Minimal — white background, coral accent	Calibri Light	E63946 457B9D 2A9D8F
dark	Dark Mode — charcoal backgrounds, violet accent	Trebuchet MS	7F5AF0 2CB67D FF6B6B
warm	Warm Academic — cream background, terracotta accent	Georgia	C45B3E 5B8A72 D4A84B

You can pass the theme via --template <name> or set ARCDECK_TEMPLATE=<name>.

Skill Pipeline Details

For the full multi-phase skill workflow, see arcdeck-skill/README.md.

---

Web UI

python app.py

Opens a Gradio interface at http://localhost:7860 where you can upload a PDF, configure options, and download the generated .pptx.

---

Slide Themes & Designs

PPTX Templates (--template 1–8)

Template 1	Template 2	Template 3	Template 4
Template 5	Template 6	Template 7	Template 8

JS Designs (--js_design 0–5)

Designs control geometric/ambient decoration style (background shapes, footer, title bar framing).

Design 0 — Circles & Bokeh	Design 1 — Geometric Angular	Design 2 — Sunset Gradient
Design 3 — Aurora Neon	Design 4 — Teal Minimalist	Design 5 — Glassmorphism

---

Supported Models

Configured via utils/llm/config.py. Supported platforms:

• OpenAI — GPT-4o, GPT-5
• Anthropic — Claude
• vLLM — Local models (e.g., Qwen 32B)
• DeepInfra / OpenRouter / Qwen — OpenAI-compatible endpoints

---

ArcBench

ArcBench is a curated benchmark dataset of 100 oral presentation papers from top-tier machine learning conferences (CVPR, ICCV, ICLR, ICML, NeurIPS), spanning 2022–2025. Each entry includes the full paper PDF, presentation slides PDF, and rich metadata. Papers were selected from oral presentations only, and filtered to ensure rich content (≥3 figures, ≥3 tables) and availability of both the original paper PDF and presentation slides.

You can access the benchmark from ArcBench HuggingFace.

---

Repository Structure

ArcDeck/
├── slide_generation/
│   ├── arcdeck.py                 # Entry point — orchestrates all stages
│   ├── outline_generation/         # Discourse parsing & multi-agent refinement
│   ├── content_generation/         # PDF parsing, figure matching, layout planning
│   └── renderer/                   # PPTX output (python-pptx & PptxGenJS)
├── prompts/                        # All prompt templates
│   ├── outline/                    #   Discourse parser, planner, critic, judge
│   └── pipeline/                   #   Layout agent, refinement, figure match
├── utils/                          # Shared utilities (LLM, PPTX, styling)
├── asset/                          # Design preview images
├── app.py                          # Gradio web UI
└── requirements.txt

Outputs

Generated files are saved under contents/<paper_name>/:

File	Description
*_raw_content.json	Parsed paper content
*_raw_content_rst.json	RST-organized content
*_slide_plan.json	Final slide plan
*_figures.json	Figure/table metadata
*_output_slides.pptx	Generated PowerPoint

---

Citation

If you find ArcDeck useful in your research, please cite our paper:

@article{ozden2026arcdeck,
  title     = {Narrative-Driven Paper-to-Slide Generation via ArcDeck},
  author    = {Ozden, Tarik Can and VS, Sachidanand and Horoz, Furkan
               and Kara, Ozgur and Kim, Junho and Rehg, James M.},
  journal   = {arXiv preprint arXiv:2604.11969},
  year      = {2026}
}

Acknowledgements

Selected prompts and code components in this repository’s visual generation components were adapted from SlideGen. We thank the SlideGen authors for making their work publicly available.

Contact, Support & Maintenance

For feedback, improvements, code support, or general inquiries, please contact tozden2@illinois.edu.