Perception, Layout, and Validation: Calibrated Confidence for Reliable Straight-Through Processing of Financial Documents

A decomposed confidence layer combining perception, layout, and validation channels with conformal risk control raised AUROC for financial-document key-value extraction from 0.54-0.74 with native Vision Language Model verbalized confidence to 0.90-0.99, according to a paper published on arXiv (2609.20110v1). Validated on three public datasets of real invoices, synthetic invoices, and ad-buy forms using Qwen3.6-27B and Gemini-3.1-Flash-Lite, the method auto-approves 49-72% of fields while holding the accepted tier's empirical error at or below a 10% target, versus 0.1%-7.0% for native VLM confidence signals. The result enables usable straight-through processing of financial documents without human review.

arXiv:2609.20110v1 Announce Type: new Abstract: Straight-through processing STP on extracted key-value fields from financial documents without human review requires a calibrated probability together with a bounded guarantee on the residual error of the auto-approved tier. The emergence of modern Vision Language Models VLMs provides an out-of-the-box capability for extracting the key-values, but their verbalized confidence signals are unreliable and weakly track field correctness. This paper introduces a decomposed confidence layer along three interpretable channels, including perception, layout, and validation. Together with a final conformal risk control, the score can be used for reliable STP of financial documents. The method is validated on three public datasets covering real invoices, synthetic invoices, and ad-buy forms, using two different VLM families Qwen3.6-27B and Gemini-3.1-Flash-Lite . Our decomposed score consistently improves the separation of correct from incorrect extractions, substantially raising the AUROC from 0.54-0.74 for VLM verbalized signals to 0.90-0.99 with contributions from all three designed channels. Crucially for industrial deployment, this enables usable STP. The native VLM confidence signals could clear only 0.1%-7.0% of fields under risk control at a target error of <10%. In contrast, the proposed method auto-approves 49-72% of fields while holding the empirical error of the accepted tier at or below the target.