OmniDocBench
v1.6 · diverse documents
Δ+0.64
TECHNICAL REPORT COVERAGE → CAPABILITY
WeVisDoc
From Coverage to Capability for Robust
End-to-End Document Parsing
WeChat Vision Team, Tencent Inc. * Equal contribution. † Project leader. ‡ Corresponding author.
Document parsing converts document images into structured content and requires reliable performance across diverse layouts and acquisition conditions. We present WeVisDoc, a two-stage data-centric framework for robust end-to-end document parsing. Stage I broadens semantic, structural, and appearance coverage; Stage II uses residual-error diagnostics to guide targeted data construction and reallocation.
OmniDocBench v1.695.38Overall · 4B
PureDocBench75.54Avg₃ · 4B
Model family2B / 4BEnd-to-end
01 / END-TO-END LEADERBOARD
Four views. One parser.
No pipeline hand-offs.
WeVisDoc-4B against the strongest reported end-to-end specialist on each benchmark view. Every value is the Overall score.
PureDocBench Clean
source-traceable clean pages
Δ+1.43
PureDocBench Digital Degraded
10 algorithmic degradations
Δ+1.41
PureDocBench Real Degraded
4 physical capture chains
Δ+1.44
Overall ↑ · The comparison baseline is named per row. The visual axis begins at 60 to make small score differences legible.
STAGE II EFFECT · 4B
Refinement pays where appearance shifts are hardest.
Capability-aware refinement preserves clean-page accuracy while concentrating the largest gain on real captures.
OmniDocBench+1.16Overall
PDB · Clean+0.49Overall
PDB · Digital+2.55Overall
PDB · Real+4.03Overall
02 / INTERACTIVE RECOVERY LAB
See what Stage II fixes.
Twelve selected pages expose concrete failure modes—not just final scores. Switch benchmark tracks, inspect the marked source, and compare the serialized outputs.
LIVE COMPARISONCase 01 · OmniDocBench (a)01 / 03
SOURCE PAGEScroll to inspect
Stage I omits the 9×9 Sudoku grid; Stage II restores the complete table.
Overall 65.87 → 99.22TableTEDS 0.333 → 1.000
SERIALIZED OUTPUTMarkdown-native
Fig. 10.3 Solved sudoku puzzle. ## 10.4 Hybrid Optimization Hybrid methods may be required to solve particularly difficult real-world optimization problems. Implementation of hybrid methods typically requires non-trivial scripting …
capability-aware refinement
<table><tr><td>5</td><td>3</td><td>4</td><td>6</td><td>7</td><td>8</td><td>9</td><td>1</td><td>2</td></tr> … <tr><td>3</td><td>4</td><td>5</td><td>2</td><td>8</td><td>6</td><td>1</td><td>7</td><td>9</td></tr></table> Fig. 10.3 Solved sudoku puzzle. ### 10.4 Hybrid Optimization …
03 / COVERAGE TO CAPABILITY
Train broad. Diagnose precisely.
Refine deliberately.
The architecture stays fixed. What changes is where useful target-token mass is allocated.
01
STAGE I · SUPPORT
Widen the document manifold
Source-balanced heterogeneous data combines semantic coverage with structure-preserving appearance synthesis.
domainslayoutsdegradations
02
DIAGNOSIS · RESIDUALS
Find capability regions
Held-out probes reveal coherent failure clusters across structure, content, language, and acquisition conditions.
03
STAGE II · ALLOCATION
Spend tokens on residual gaps
Clipped residual-aware weights and curated hard examples strengthen weak regions without displacing natural data.
boundedauditedtargeted
ONE OUTPUT SPACEMarkdown
# Heading$$ E = mc² $$<table>…</table>Text, formulas, tables, and reading order stay in one inspectable sequence.
04 / OPEN RELEASE
Run WeVisDoc.
Start with the compact 2B checkpoint or use the 4B checkpoint for the strongest reported results. Code and tutorials are available on GitHub.
RELEASE LINKS2B / 4B
Model checkpointWeVisDoc-4B ↗Highest reported resultsModel checkpointWeVisDoc-2B ↗Compact end-to-end parser
Technical reportComing soonPaper and citation will be added here.