Hugging Face has published a research report introducing PolyOCR-Venus, a family of unified OCR foundation models designed to move optical character recognition beyond plain-text transcription toward what the authors call text-centric visual intelligence. The models are intended to recognize, localize, and reason over text in complex visual environments.
According to the report, existing OCR systems tend to excel at only some tasks and struggle to balance recognition, parsing, and reasoning across scenarios. PolyOCR addresses this by combining a shared instruction-following framework with a large-scale data engine that converts heterogeneous visual resources into quality-verified OCR supervision. The 2B and 9B models bring text recognition, grounding, document parsing, information extraction, and OCR reasoning into one interface.
The report introduces Competence-Guided Policy Optimization (CGPO), which combines verifier-based Group Relative Policy Optimization with on-policy distillation. Teacher supervision is adjusted per sample based on teacher reliability and the evolving teacher–student competence gap.
On evaluation, PolyOCR-9B scores 80.42 / 78.37 on OCRBench v2.1 (English / Chinese) and 82.43 on CC-OCR across the three evaluated tracks, excluding multilingual OCR. A dedicated 2.7B multilingual variant reaches 87.0 on MDPBench.
The authors also present OCRBench v2.1, a revision of OCRBench v2 that corrects 2,449 samples through manual verification and introduces task-aligned scoring metrics for more reliable evaluation. Experiments across OCRBench v2.1, CC-OCR, an in-house KIE Benchmark, OmniDocBench v1.6, and MDPBench show state-of-the-art or highly competitive performance, the report states.