Benchmark RadarLongDocBench
LongDocBench is a benchmark for Table-of-Contents Hierarchy Recovery and Contextual Relationship Recovery in long documents. It includes 85 real-world documents (financial reports, textbooks, academic papers) spanning 2,582 pages, with human-verified annotations for 3,937 heading nodes and 3,258 contextual relationships.
- Released
- 2026-08-15
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Existing document parsing benchmarks focus on page-level tasks, leaving document-level structure recovery unevaluated. LongDocBench provides a standardized evaluation for tasks that are critical for understanding long documents, enabling comparison of parsers on hierarchy and relationship recovery.
Motivation
Parsing visual documents into machine-readable representations is fundamental to document intelligence.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.