Multimodal retrieval-augmented generation over visually rich documents often needs evidence at a finer granularity than whole pages. According to the latest arXiv announcement, existing visual retrievers usually index or return full pages, which can be too coarse for tasks that depend on a specific figure, table, or paragraph.
The paper introduces GARRET, a method that leverages generative feedback for region retrieval. The abstract suggests the work directly targets the mismatch between page-level indexing and the need for region-level evidence, though the announcement does not yet include experimental details.
Because the abstract is truncated, the exact mechanism and results remain unclear. Still, the problem it addresses is well recognised: making retrieved evidence precise enough for document-grounded reasoning. The paper is available as arXiv:2610.27829v1.