XL-DocBench

"Benchmarking Evidence-Grounded Extra-Long Document Understanding." A benchmark for the document work professionals actually do: answering questions from annual reports, regulations, clinical guidelines, and technical manuals that run to hundreds or thousands of pages, sometimes by comparing related reports, with every answer tied to evidence in the source. The public "conservative" release on HuggingFace (September 2, 2026) holds 1,345 QA rows over 292 documents after removing questions whose sources were not approved for redistribution. Authors span Wuhan University and Microsoft.