VGA-KIE adds a verification gate to document extraction: mean Field-F1 rises from 93.49 to 97.65 on a shared candidate cache
Synopsis
Key information extraction from visually rich documents can produce schema-valid field values without showing which page region supports them or whether schema and cross-field constraints hold. VGA-KIE binds each field value, locatable evidence, verification state, failure reason and repair history into one field-level state, and releases a repaired result only after evidence is relinked and the same constraints are re-checked. In a paired comparison sharing one frozen candidate cache with LayTokenLLM-8B direct extraction, mean Field-F1 rose from 93.49 to 97.65, with 4,475 of 4,720 target fields released.
Interpretation
The output of key information extraction becomes a field-level state carrying evidence, verification status and repair history rather than a bare value, and a field is released only when its evidence is non-empty and locatable, its constraints are testable and its state is pass. Earlier document models mainly improved candidate generation and kept candidate values, locatable evidence, verification states and repair actions in separate components, so they could not form a unified field state with auditable transition rules. In a paired comparison sharing the same frozen LayTokenLLM-8B checkpoint, input processing, decoding configuration and candidate cache with direct extraction, mean Field-F1 rose from 93.49 to 97.65, with gains of 4.05, 4.45 and 3.98 percentage points on the three datasets.
The system processes each target field through extraction, grounding, verification, repair and verification again, and repair acts only on fields judged fail or uncertain rather than re-running the whole document. The grounding module links each candidate to a token span, text line, text block, table cell or image region, and the verifier then tests format, spatial, evidence, semantic, cross-field and table-structure constraints. On the fixed test inventory of FUNSD, SROIE and CORD, covering 497 documents and 4,720 target fields, VGA-KIE released 4,475 fields, an acceptance coverage of 94.8%.
A repaired candidate must be re-grounded and pass the same constraint checks before entering the final output, and this second verification gate mainly limits harmful rewrites. Removing verification again slightly raised repair success to 71.8% but increased the false repair rate from 2.1% to 5.8%. The full system corrected 282 of 400 initially incorrect fields that entered repair, a repair success of 70.5%, and changed 86 of 4,100 initially correct fields to incorrect values, a false repair rate of 2.1%.
Perspective
The results are limited to FUNSD, SROIE and CORD under fixed splits and a shared evaluation script, with LayTokenLLM-8B fixed as the candidate generator; the 1.32x latency measures only post-hoc overhead above the frozen candidate cache and excludes OCR and candidate generation.
Failure detection is reported as a binary metric, and per-class precision, recall and F1 for the six dominant failure reasons are not given; the results do not establish stable gains across candidate generators and do not cover long multi-page documents; all values are point estimates on fixed test sets without document-level bootstrap confidence intervals or across-seed variance.