Evidence Attribution in Visual Document Understanding without Coordinates or Region Labels

Researchers proposed a language-based interface for visual document QA that predicts verbatim quotes as evidence instead of coordinate boxes.

A parser and retriever then map those quotes to page regions, enabling training without region-level evidence labels.

On CiteVQA, evidence recall jumped from at most 8% to about 47%, while attribution hallucination was cut roughly in half.

With GRPO training, strict attributed accuracy improved from 22.4% to 33.8%.