Published Document Processing Benchmarks | Extend
Benchmarks for the hardest parts of document processing
Published document processing benchmarks for parsing, structured extraction, and splitting. Each report states its corpus, metric, comparison scope, methodology, and public data or code.
Benchmarks
Public benchmarks for document parsing, extraction, and splitting, with published methodology and source data. Each card separates key results, benchmark details, and sources.
Parsing
Measures whether parsers deliver accurate layouts, preserve reading order, and enable agents to correctly answer objective questions against real-world documents.
Q&A accuracy 95.7% Extend Parse 2.0
Layout F1 0.847
Benchmark details for RealDoc-Bench
| Field | Detail |
|---|---|
| Test corpus | 1,500 layout samples; 1,359 document Q&A prompts across 581 documents in four regulated industries |
| Metric | Adjusted F1 for layout; field-level document Q&A accuracy |
| Comparison | Same-run comparison with LlamaParse, Reducto, Azure Document Intelligence, and AWS Textract |
| Published | May 26, 2026 |
Sources
- Research paper arXiv
- Source code GitHub repository
- Layout dataset Hugging Face
- Document Q&A dataset Hugging Face
Extraction
Tests whether extraction systems preserve cardinality and return complete, schema-faithful arrays when the output grows from a dozen of rows to thousands.
Mean accuracy 99.2% Extend across 45 PDFs
Run completion 100% 45 of 45 PDFs completed
Benchmark details for LongArray-Extract
| Field | Detail |
|---|---|
| Test corpus | 45 financial, clinical, and legal PDFs with repeated arrays of 27 to about 2,200 records |
| Metric | Mean per-document extraction accuracy, with failed and timed-out runs scored as zero |
| Comparison | - Document-AI platforms: Reducto, Pulse, LlamaParse - Raw-model providers: Anthropic, Google, OpenAI |
| Published | June 2, 2026 |
Sources
Splitting
Evaluates document splitting on long, compound tax filings where frontier models miss subtle boundaries across hundreds of pages.
Best harness F1 72.48% Claude Opus 4.6
Recall lift 17-44 Points across models
Benchmark details for PoliTax Split
| Field | Detail |
|---|---|
| Test corpus | The 30 largest compound documents in the public PoliTax corpus |
| Metric | Boundary-detection F1 |
| Comparison | Same-model comparison between the Extend splitting harness and direct frontier-model use |
| Published | March 30, 2026 |
Sources
( fig.11 )