Failure Modes in Multi-Page Visually Rich Document Understanding

Empirical analysis of multi-page document understanding reveals that errors stem from representation, selection, or reasoning failures. Findings show that vision cannot replace text extraction and that reasoners frequently fail to integrate evidence even when all pages are provided in context.