Read-Parse-Describe
Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation
A unified dataset for document parsing requiring models to output ordered layout elements with bounding boxes, labels, reading order, text, and visual descriptions from 18,163 pages.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info read-parse-describe
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
13 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Dolphin-v2 | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek-OCR | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| ChatGPT-5.4 | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.5-2B + SARA | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| MonkeyOCR | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| MinerU | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemini 3.1 Pro | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemma4-E4B | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.5 2B | 1.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek-OCR | 0.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| PaddleOCR-VL | 0.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Dolphin-v2 | 0.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| ChatGPT-5.6 | 0.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
The benchmark comprises 2,402 scientific documents (18,163 pages) annotated with 125,218 layout elements and 11,338 visual instances. It requires generating a structured page parse including spatial, textual, and semantic information.
Evaluation measures the joint recovery of reading order and visual semantics against a unified contract. Existing systems struggle with this joint task, while the proposed SARA-enhanced model shows improved performance in parsing fidelity and generation quality.