Searching visually rich documents is hard because the relevant evidence often sits in tables, figures, or layouts that don't match plain-text queries. A new arXiv preprint tackles this by proposing a cascaded retrieval system that refines queries across modalities, aiming to make visual document search more efficient.

The paper's title and abstract describe a two-stage approach: first a broad retrieval pass, then a cross-modal refinement step that sharpens the query based on the visual context. This is meant to handle the complex, layout-dependent evidence that current vision-language retrievers struggle with.

However, the abstract as provided is cut off mid-sentence, so the full methodology, experimental results, and comparisons to existing systems are not described in the available text. Readers should consult the full paper for those details.