Original title: Closing Cost-Quality Gap in Document VLMs: Difficulty-Aware Data Curation and Quality-Adjusted Deployment Economics
Maksim Evdokimov, Matvey Ivanov, Dmitrii Tsiupin, Olga Tsymboi +2 · September 2026
Pulling structured text from huge piles of documents is tricky. Custom software misses many layouts, privacy laws ban external models, and good open-source models cost more to run than human workers. Researchers solved this by building a practical document-reading model. They used a smart setup where only a small part of the network runs at one time. To teach it, they blended internal business files with public documents filtered to test layout variety, extractable facts, and consistent answers across different models. It runs on a single computer chip and guides different tasks using prompts. It outperforms rival models up to ten times its size, cuts costs by over 80% compared to humans, and beats competing open-source models by more than 50%.
Imagine sorting through a huge stack of tax forms to find key numbers. Normally, you'd read every page, which takes forever and might miss details. In this paper, the system acts like a focused assistant who only glances at the most relevant parts of each form, saving time and reducing errors. It's trained to know which documents are tricky, so it pays extra attention where needed.
Regulated industries like finance or healthcare process millions of documents annually, but current methods are slow, expensive, or limited by privacy rules. This system makes automation affordable, so companies can handle paperwork faster and cheaper without compromising security.
The abstract doesn't explicitly state limitations, but it implies the system is optimized for regulated industries and uses in-house production data, which may limit generalizability. It focuses on non-reasoning tasks, so complex reasoning might not be covered.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Closing Cost-Quality Gap in Document VLMs: Difficulty-Aware Data Curation and Quality-Adjusted Deployment Economics — Maksim Evdokimov, Matvey Ivanov, Dmitrii Tsiupin et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.