WeVisDoc: From Coverage to Capability for Robust End-to-End Document Parsing
Hao Yu, Kang Liu, Linnan Zhao, Jiabo Zhan, Chong Sun, Chen Li, Jing Lyu
WeChat Vision, Tencent Inc.
Published September 17, 2026 · on the feed September 18, 2026 · ▲ 47 · arXiv 2609.20423
The abstract
Document parsing converts document images into structured content and requires reliable performance across diverse layouts and acquisition conditions. Yet training corpora are biased toward common document types and clean digital pages, while expanding coverage alone does not specify how to address a parser's remaining weaknesses. We present WeVisDoc, a two-stage data-centric framework for robust end-to-end document parsing. Stage I broadens semantic, structural, and appearance coverage through heterogeneous data and structure-preserving degradation synthesis. Stage II uses a held-out probe to measure the Stage I parser's residual errors within fixed visual-structural clusters. These diagnostics guide targeted data construction and reallocation of the target-token budget. WeVisDoc-4B achieves an Overall score of 95.38 on OmniDocBench v1.6 and a mean Overall score of 75.54 across the three PureDocBench tracks, ranking first among the compared end-to-end parsers in all four settings. Compared with Stage I, Stage II improves Overall scores for the 2B and 4B models on both benchmarks, with larger gains on the degraded PureDocBench tracks, including a 4.03-point gain for the 4B model on the Real Degraded track.
The Brief
2:13A short podcast summary of the paper. Written and voiced by AI.
Transcript
WeVisDoc. From Coverage to Capability for Robust End-to-End Document Parsing. By Hao Yu and six others, from WeChat Vision at Tencent. Published on the seventeenth of September. Its code repository is public but slight: eight source files.
Document parsing turns images of documents into structured content, and has to work reliably across many layouts and capture conditions. Training corpora, though, lean toward common document types and clean digital pages. Adding more coverage helps, but it does not say how to fix whatever a parser still gets wrong. A parser that has seen many kinds of document can still fail on particular structures or under particular damage, and more of the same data will not necessarily reach those failures.
WeVisDoc is a two-stage, data-centric framework for robust end-to-end document parsing. The first stage is about coverage. The second is about capability: finding the parser's remaining weaknesses and building data aimed at them.
Stage I broadens semantic, structural and appearance coverage, using heterogeneous data and degradation synthesis that preserves structure. Stage II uses a held-out probe to measure the Stage I parser's residual errors within fixed visual-structural clusters. Those diagnostics then guide targeted data construction, and the reallocation of the target-token budget toward where the errors are. Grouping errors into fixed clusters is what makes them actionable. A weakness becomes a specific kind of layout or appearance where the parser underperforms, and data can be built for exactly that.
WeVisDoc-4B scores an Overall of ninety-five point three eight on OmniDocBench v1.6, and a mean Overall of seventy-five point five four across the three PureDocBench tracks. That ranks it first among the end-to-end parsers compared, in all four settings.
Compared with Stage I, Stage II improves Overall scores for the two billion model and for the four billion model, on both benchmarks. The gains are larger on the degraded PureDocBench tracks, including four point zero three points for the four billion model on the Real Degraded track.
The title names the progression. Coverage gets a parser to see many kinds of document. Measuring where it still fails, cluster by cluster, and spending data there is what turns that coverage into capability, and the degraded tracks are where that second step shows most.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
The model is good. The question is what to do about where it isn't.
You are training a model that reads page images and writes out their text, tables, formulas and reading order. Pages come in every layout, and some are scanned, photographed or damaged.
The road ahead
- 1 · The problem ~7 min A page-reading model trained on broad data still fails on some structures and some damage, and more of the same data gives no aim.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min First broaden coverage, then find the trained model's remaining weaknesses in fixed groups and aim data and training tokens at them.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Build broad data, train a first parser, measure its errors per cluster on a held-out probe, then add targeted data and shift tokens.
- 6 · The experiment ~6 min You’ll guess: After the second, capability-focused stage, how many Overall points did the larger model gain on the Real Degraded track?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.