Answers are grounded in this paper and cite their sources.
Key information extraction (KIE) from visually rich documents must identify field values, establish that each value comes from the correct document region, test schema and cross-field constraints, and support interpretable correction after verification failure. Existing document models improve candidate generation, but usually manage candidate values, locatable evidence, verification states, and repair actions separately. We introduce VGA-KIE, a modular post-hoc reliability method that couples three mechanisms as field-level state transitions. An evidence-grounded field state defines the object to be diagnosed. Typed failure diagnosis converts constraint violations into repair conditions. Failure-conditioned repair with re-verification permits a state transition only after renewed evidence linking and verification against the same constraints. Under a shared protocol, LayTokenLLM-8B with VGA-KIE achieved Field-F1 scores of 94.91, 99.17, and 98.86 on FUNSD, SROIE, and CORD. Against direct extraction from the same frozen candidate cache, mean Field-F1 increased from 93.49 to 97.65. VGA-KIE released 4,475 of 4,720 target fields, reached 92.8\% Evidence Accuracy and 87.9\% Verification Failure F1, and produced a 2.7\% Unsupported Output Rate, 70.5\% Repair Success, and 2.1\% False Repair Rate at 1.32× post-hoc latency. The paired comparison and component ablations support the combined reliability layer under the evaluated protocol. They do not establish per-class optimal diagnosis, candidate-generator-independent gains, or priority for any individual grounding, verification, or repair operation.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: