Tencent's WeVisDoc tops OmniDocBench with 95.38 via two-stage data-centric training

tencent · hf · 2026-09-18

Tencent introduces WeVisDoc, a two-stage data-centric framework for robust end-to-end document parsing. Stage I broadens semantic, structural, and appearance coverage via heterogeneous data and structure-preserving degradation synthesis; Stage II uses a held-out probe to diagnose residual errors within visual-structural clusters and reallocates the target-token budget accordingly. WeVisDoc-4B scores 95.38 overall on OmniDocBench v1.6 and 75.54 mean across three PureDocBench tracks, ranking first in all four settings, with the largest gains on degraded inputs (e.g., +4.03 on Real Degraded for the 4B model).

Original post →

More from Research

Research channel →