Enterprises increasingly rely on frontier multimodal models to extract structured fields from documents, following predefined schemas. Accuracy has steadily improved as new models and systems arrive, but the paper argues that two aspects have not kept pace: the evidence supporting each extracted field and the confidence in that extraction.
To address this, the authors propose VERDICT, a visual extraction-result decoder that integrates confidence and traceable evidence directly into the extraction process. Rather than focusing solely on improving accuracy, VERDICT aims to make the results more transparent and verifiable.
The paper's contribution is a decoder that pairs each extracted field with both a confidence measure and evidence that can be traced back to the source document. This is intended to give users a clearer basis for trusting or auditing automated extractions in real-world enterprise workflows.