Document parsing improved by fixing specific failure cases not complexity
Collapse, Not Complexity: Failure-Conditioned Decomposition Repair for End-to-End Document Parsing
Computer Vision and Pattern RecognitionComputation and Language
Summary
Parsing complex documents can sometimes fail badly, but these failures aren’t always due to the documents being complicated. The authors found that the usual strategy of using more reasoning or increasing effort often doesn’t help. Instead, they show that detecting when parsing fails and breaking the document into smaller parts can fix many errors more efficiently. This new approach improves accuracy while using fewer computing resources compared to traditional methods.
What this means in practice
- •For software engineers: Implement failure detection and region-wise re-parsing to improve document parsing software accuracy and efficiency.
- •For enterprise automation teams: Enhance automated document processing pipelines to recover from parsing collapses without costly increases in computation.
Authors
Xingyu Lin, Dehui Du
Abstract
End-to-end document parsers increasingly offer an optional reasoning mode for complex pages. On a 180-page entropy-stratified discovery sample with one frozen 4B checkpoint, complexity is the wrong decision variable. Reasoning lowers mean quality by 2.21 Overall at 1.54x tokens; a preregistered input-only model cannot predict its signed benefit (held-out AUROC 0.47, indistinguishable from chance). The benefit concentrates on pages whose ordinary pass has already collapsed, and they do not look complex: shared collapses have lower layout entropy than healthy ones yet consume 19x the tokens as degenerate repetition that doubling the budget does not cure. Switching modes rarely repairs them: 83% recur under reasoning. We instead detect collapse from the ordinary-pass trace, decompose the page by projection, and re-parse each region. Repair gains 1.40 Overall (95% CI [0.68, 2.16]) at 1.13x tokens, replicates across three checkpoints, and, with all parameters frozen, gains 2.41 (CI [1.64, 3.46]) on the remaining 1,175 benchmark pages.