ArcDeck/ArcBench

ArcBench: ML Conference Oral Paper-Presentation Benchmark

This benchmark is from the paper [Narrative-Driven Paper-to-Slide Generation via ArcDeck](https://huggingface.co/papers/2604.11969).

A curated benchmark dataset of 100 oral presentation paper-slide deck link pairs from top-tier machine learning conferences (CVPR, ICCV, ICLR, ICML, NeurIPS), spanning 2022–2025. Each entry provides rich metadata together with links to the original paper PDF and presentation slides, plus a script that downloads them all in one step.

   

Dataset Summary

This benchmark is designed to support research on multimodal document understanding, slide generation, paper-to-slide alignment, and LLM evaluation tasks. Papers were selected from oral presentations only — the highest-quality subset of each conference — and filtered to ensure rich content (≥3 figures, ≥3 tables) and availability of both the original paper PDF and presentation slides.

Dataset Structure

Files

benchmark.csv            # Metadata + source links for all 100 papers
download_pdfs.py         # Downloads every paper/slide PDF from its source link
download_openreview.py   # Companion downloader for the OpenReview-hosted papers
papers/                  # Created by the download script: 100 paper PDFs
└── paper{i}_{Title}_{Conference}_{Year}.pdf
slides/                  # Created by the download script: 100 slide PDFs
└── slide{i}_{Title}_{Conference}_{Year}.pdf

The papers/ and slides/ folders are populated by running the download script (see Downloading the PDFs).

Metadata Fields (benchmark.csv)

Naming Convention

Files are named as {type}{index}_{CleanTitle}_{Conference}_{Year}.pdf where:

• —index is 0-based, consistent across papers/ and slides/ for matched pairs
• —CleanTitle has special characters removed and spaces replaced by underscores (max 100 chars)

The download script reconstructs these exact filenames from benchmark.csv, so file i always corresponds to row i of the metadata.

Downloading the PDFs

You can download the PDFs from their original sources with the included script.

# Download all 100 papers + 100 slides into ./papers and ./slides
python download_pdfs.py

Useful options:

python download_pdfs.py --type slides          # slides only (or: papers, both)
python download_pdfs.py --indices 0,5,84        # just a few entries
python download_pdfs.py --limit 10              # first 10 entries
python download_pdfs.py --out /data/arcbench    # choose the output directory
python download_pdfs.py --workers 8             # more parallelism
python download_pdfs.py --force                 # re-download existing files

The script verifies every download is a real PDF, writes atomically, and skips files that are already present, so it is safe to re-run to resume an interrupted download. Any files it could not fetch are listed in download_failures.csv.

A note on versions

Links point to the canonical, live source for each work. For papers hosted on arXiv this is the latest revision, which may differ slightly from the exact PDF originally archived for the benchmark (updated figures, camera-ready edits, etc.). The content is the same paper. Slide decks served by conference media servers are typically byte-for-byte identical.

Dataset Statistics

Distribution by Conference

Distribution by Year

Content Statistics

• —92% of papers include an appendix
• —100% are oral presentations

Top Research Topics

Extracted via GPT-4o-mini from paper abstracts:

Contrastive Learning · Graph Neural Networks · Causal Inference · Multimodal Large Language Models · Federated Learning · Sampling Efficiency · Reinforcement Learning · Diffusion Models · Self-Supervised Learning · Vision-Language Models

Selection Criteria

Papers were selected using the following filters applied to a broader 994-paper dataset:

• —Presentation type: Oral only
• —Minimum figures: ≥ 3
• —Minimum tables: ≥ 3
• —Original paper available: Must have the full (non-anonymized) version
• —Balanced sampling: Proportional stratified sampling across year × conference to reach exactly 100 papers

Intended Uses

This dataset is suited for:

• —Slide generation / paper-to-slide summarization: Given papers/, generate slides comparable to slides/
• —Slide-grounded QA: Answer questions about a paper using its slides as context
• —Cross-modal retrieval: Match papers to their corresponding slides
• —LLM evaluation: Benchmark LLM understanding of dense scientific documents
• —Multimodal document analysis: Study relationships between figures, tables, equations, and slide content

Source

Papers were collected from official proceedings of:

• —ICML (2022–2025)
• —ICLR (2024–2025)
• —NeurIPS (2022–2025)
• —CVPR (2024–2025)
• —ICCV (2025)

Citation

If you use this dataset in your research, please cite:

@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}
}

License

The benchmark metadata (benchmark.csv), the source links, and the download scripts in this repository are released under the MIT license.