The headline numbers are disputed, self-reported, or under revision, so treat the ranking as a claim, not a settled fact.
OmniDocBench
OmniDocBench tests how well a system turns a real PDF page into clean structured output, getting the text, tables, formulas, layout, and reading order all right at once, versus just running plain OCR on tidy text.
It spans 1,651 real pages across many document types, so it measures the messy real-world version of document parsing.
A system is handed a real PDF page, a scanned form, a paper with equations, or a multi-column report with tables, and must reproduce its full content as structured data; the score combines how well it recovered the text, the tables, and the formulas.
Turning PDFs and scans into usable data is one of the most common practical AI tasks, for research, finance, and paperwork, and OmniDocBench is the most comprehensive public test of it. The catch, below, is that its leaderboard cannot be taken at face value.
A composite Overall score (0 to 100, higher is better) that averages a text edit-distance metric, a table-structure metric (TEDS), and a formula metric (CDM). Maintained by OpenDataLab; published at CVPR 2025. Text scoring currently covers only Chinese and English.
The Overall score blends three different sub-metrics, so it is a rough composite rather than a single accuracy number; the leaders cluster in the mid-90s, which sounds near-perfect but reflects the metric's generosity more than solved document parsing.
Treat the ranking with care for two reasons. The top scores are self-reported in the project's own repository rather than independently re-run, and the number-one system is maintained by OpenDataLab, the same organization that maintains the benchmark, a direct conflict of interest. Text evaluation also covers only Chinese and English, even though the dataset spans more languages. There is no independent third-party board to check these numbers against.
The benchmark creates the need to know. The catalog explains the ideas behind it: