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
cs.CV
2026-09-17
WeVisDoc trains parsers in two data stages: ~40M coverage records, then ~5M residual-aware ones. The 4B model scores 95.38 on OmniDocBench v1.6 and 75.54 mean on PureDocBench.
Document parsing turns a page into structured text: Markdown, LaTeX formulas, HTML tables, and the right reading order. Corpora skew toward common layouts and clean digital pages. Collecting more data does not say which remaining failures are missing support, and extra copies of patterns you already have are a different intervention. High residual error can also mean bad labels, unreadable pixels, or decoding failure.
WeVisDoc, from WeChat Vision at Tencent, budgets supervision in target tokens rather than record counts, because a full page and a short formula are not equal. On a fixed backbone it first builds coverage, then uses residuals to choose new data versus more exposure.
The 2B and 4B runs start from Qwen3-VL-Instruct with a frozen serialization and autoregressive loss. Stage I is about 40 million records: public OCR sets, in-house production pages, targeted crawls, and executable page programs. Pages / text-or-mixed crops / isolated tables / isolated formulas are roughly 40% / 25% / 25% / 10%. Source token shares use an exponent ρ so huge libraries cannot dominate. When labels are missing, three expert parsers (MinerU2.5-Pro, PaddleOCR-VL-1.6, dots.mocr) must agree to enter the clean pool; disagreements go to review or Stage II. Appearance transforms follow acquisition paths (scan, photo, screen recapture) and drop views that are no longer readable.
Stage II is about 5 million records. The frozen Stage I checkpoint is probed on a split disjoint from training, hard-example mining, and final eval. Residuals are aggregated in visual–structural clusters from SigLIP2 embeddings plus layout metadata. High-error items are audited for label and legibility before retrieval, targeted synthesis, or extra exposure. Stage II freezes the vision encoder and updates the language model under its own token budget.
On OmniDocBench v1.6 (1,651 pages), the 4B Overall is 95.38 and the 2B is 95.06, ahead of HunyuanOCR-1.5 at 94.74 and Unlimited-OCR at 93.92 among end-to-end specialists. Pipeline systems still sit higher: PaddleOCR-VL-1.6 at 96.33, MinerU2.5-Pro at 95.75. The paper's "first in all four settings" is among compared end-to-end parsers, not every parser.
On PureDocBench's three aligned tracks, the 4B Avg3 is 75.54 (Clean 79.81, Digital 77.74, Real 69.08). Versus Stage I, the 4B gains +1.16 on OmniDocBench, only +0.49 on Clean, +2.55 Digital, +4.03 Real. The 2B pattern is the same: degraded tracks move more. Scaling 2B to 4B adds 3.48 on Real Degraded and 0.45 on Clean, so extra capacity is spent on acquisition shift.
| Model | OmniDocBench Overall | PureDocBench Avg3 | Real Degraded |
| WeVisDoc-4B Stage II | 95.38 | 75.54 | 69.08 |
| WeVisDoc-4B Stage I | 94.22 | 73.19 | 65.05 |
| WeVisDoc-2B Stage II | 95.06 | 73.86 | 65.60 |
| HunyuanOCR-1.5 | 94.74 | 68.79 | n/a |
| PaddleOCR-VL-1.6 (pipeline) | 96.33 | n/a | n/a |
Qualitative Stage II fixes include dropped tables, column-major reading, split field–value pairs, lost formula serialization, and bleed into another Chinese document. On real degraded pages, structure returns; local character errors remain.
For document VLMs this is an operable data system: coverage and residual repair are booked separately, and supervision is counted in tokens, not pages. 2B already sits on the end-to-end frontier; 4B buys most of its extra on real degradation. A lot of remaining headroom is in the mixture, not another decoder. The pipeline SOTA gap is also honest: modular layout plus regional recognition still wins on clean pages.
Stage II bundles continued training, hard examples, targeted synthesis, and replay, so Table 4 is not a component ablation. Probe set, clustering, and evaluator all steer which data get added. Handwriting, historical scans, severe damage, and more languages remain thinly measured. Synthetic degradation is not a substitute for real capture. High-stakes use still needs human review. OmniDocBench uses annotated gold; v1.5 had many block-level errors, and v1.6 scores are tied to a frozen evaluator.