Comparison of PDF-to-Markdown parsing tools for scientific research papers (arXiv-style).
| Tool | Tiers Tested | Guide |
|---|---|---|
| Marker | 1, 4 | Guide |
| MarkItDown | 1, 2 | Guide |
| Docling | 1–5 | Guide |
| SciPDF | 3–5 | Guide |
| PaddleOCR | 1 | Guide |
| dots.ocr | — | Guide |
| OCRFlux | — | Guide |
| DeepSeek-OCR | — | Guide |
| Dolphin | — | Guide |
| MinerU2.5 | benchmarked (M2 Pro) | Guide |
| Chandra | benchmarked (M2 Pro) | Guide |
| QARI-OCR | benchmarked (Arabic) | Survey |
parsing tools/ # Setup guides with tier configs and benchmarks
comparison_reports/ # Side-by-side quality comparisons
benchmark/ # Tier configs and benchmark scripts
output/ # Parsed markdown/JSON per tool and tier
papers/ # Source PDFs (gitignored)
Every parse under output/ is of the same source document, so the tools and
tiers are directly comparable:
Zhang, Alex L., Tim Kraska, and Omar Khattab. Recursive Language Models. arXiv:2512.24601v2. Licensed CC BY 4.0.
It was chosen for having the failure modes that separate these tools: a multi-column layout, inline math, footnotes, and a multi-section comparison table with grouped column headers.
- Marker Tier 1 vs Tier 4
- MarkItDown vs Marker
- Docling vs Marker vs MarkItDown
- Marker Tier 1 vs Docling Tier 4
- SciPDF Tier 3 vs Marker
- SciPDF Tier 5 vs Marker
- New Parsers Benchmark — MinerU2.5 vs Marker vs QARI on Apple M2 Pro / 16 GB
Original work in this repository (guides, comparison reports, research reports,
benchmark scripts, notebooks) is MIT licensed: see LICENSE.
Third-party material and the reasoning behind which documents are and are not
republished here are documented in NOTICE.md.