← News

ResearchSources: arXiv · 18 Nov 2026

Quote-and-retrieve approach cuts evidence hallucination in document AI study

A study on arXiv reports that having vision-language models quote evidence in text, rather than output box coordinates, sharply improves attribution on a bilingual benchmark subset.

In plain words

When document-reading AI models answer questions, they often point to the wrong part of the page even when the answer itself is right. The researchers tested a different method: the model quotes its supporting text, and a separate retrieval tool locates that quote on the page. This made evidence much more accurate without hurting answer quality.

What was announced

The authors argue that a known weakness, which they call Attribution Hallucination, may partly stem from the coordinate interface itself: models frequently mark incorrect regions even when their answers are correct. They compared the usual box-output method against a text-based one across six open vision-language models. With the text method, evidence recall climbed from a maximum of 8 points to a range of 26 to 47, and hallucinations fell by roughly half, while answer quality stayed nearly unchanged.

The team then reused the quote-and-retrieve pipeline for training. Because labelling regions in long documents is costly, they applied a GRPO recipe in which a reward comes from a judge reading the gold answer alongside crops of the retrieved regions. This raised strict attributed accuracy for an 8B model from 22.4 to 33.8, suggesting attribution can improve without coordinates or region-level labels.

What it does not show

  • The source is an abstract, so full experimental details, datasets and caveats are not available for verification.
  • Results come from a verified subset of one bilingual benchmark (CiteVQA) and six open models; performance on other documents or closed models is not established.
  • The reported gains are the authors' own measurements and have not been independently confirmed.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.