Agents meet the world's paperwork

Editing is its own benchmark now

In a preprint, the VDE Bench authors argue that editing a document image is a fundamentally different task from editing a photo: dense text and tables have to change while layout integrity survives. Their benchmark supplies 3,506 instruction-guided editing samples for document images, split into 1,578 text-region edits and 1,928 table-region edits, with an evaluation protocol that combines region-level and full-page measurements of textual correctness, structural fidelity, and visual content preservation. m:…0089

The audience is explicit: teams building document automation, financial reporting, or enterprise content management, where a model that garbles a table is not a demo failure but a liability. General image-editing leaderboards never measured this; now there is a ruler, and this is reasoning, the first step toward procurement checklists that mention document editing at all. m:…0089

The non-English blind spot

A second preprint finds that existing multimodal models struggle with Korean visual document retrieval, particularly where structured content and diverse query types are involved. The KoViDoRe benchmark covers Korean visual documents with tables, figures, and multi-column structures, and the authors release Ko-VDR Train Public, a large-scale training dataset aimed at closing the gap. m:…5589

The pattern generalizes, and this is reasoning: every market whose documents use complex layouts in a non-English language is a place where frontier models underperform and local training data is the moat. Document agents that work in English and fail in Korean are not global products. m:…5589

Retrieval is not enough

In a preprint, the VELD-QA authors evaluate nine vision-language models and seven retrieval-augmented systems on long-document question answering with explicit visual evidence annotations. Retrieval-based methods significantly underperform oracle settings with ground-truth supporting pages, which means the failure is not just finding the right page. Incorporating visual-element annotations boosts multimodal document QA performance by 17 to 24%. m:…7052

The sobering number is the ceiling: oracle accuracy for visual evidence grounding in long documents peaks at 81.7%, suggesting inherent limitations in current visual interpretation and reasoning. Even with perfect retrieval, roughly one question in five defeats the models. Long-document agents need better visual reasoning, not just better search. m:…7052

Zoom before you describe

The fourth preprint offers a mechanism. ZoomCap implements dynamic, evidence-routed zooming so vision-language models stop missing small objects and spatial relations, paired with FactCycle, a verification system that scores claims against both source images and reconstructions. The reported results are startling: an 81.9% human tie-adjusted preference score against a model over ten times larger, using a teacher more than 10x smaller, with FactCycle transferring 52.4 claims per image against 31.5 for the ScaleCap baseline. m:…9296

ZoomCap also works as a data engine: continual pre-training with 450K ZoomCap-captioned images improves an 11-benchmark average by 1.31 to 2.38 points. The second-order read, and this is reasoning, is that document understanding is becoming the defensible layer in vertical AI. The models are commoditizing; the ability to read, edit, and verify the world's paperwork, in every language and layout, is not. m:…9296

Read together, the four preprints sketch a full document stack: a benchmark for editing, training data for overlooked languages, evidence annotations that retrieval alone cannot replace, and a zoom-and-verify loop that turns captions into pre-training fuel. The teams that assemble this stack first will own the document workflows everyone else demos on top of. m:…0089 m:…7052