GuangJian Team researchers introduced PolyOCR-Venus, a family of optical character recognition (OCR) models, in a paper published on September 29 and submitted to Hugging Face on October 12. The models are designed to recognize, locate, parse and reason about text in complex visual settings. The authors say existing OCR systems often perform well on individual tasks but struggle to balance recognition, parsing and reasoning across different scenarios.

PolyOCR uses a shared instruction-following framework and a large-scale data engine that turns varied visual sources into quality-verified OCR training data. Its 2B and 9B models combine text recognition, grounding, document parsing, information extraction and OCR reasoning in one interface. The team also developed Competence-Guided Policy Optimization (CGPO), which combines verifier-based Group Relative Policy Optimization with on-policy distillation. It routes training samples according to teacher reliability and the difference between teacher and student competence.

The authors report that PolyOCR-9B scored 80.42 for English and 78.37 for Chinese on OCRBench v2.1, and 82.43 across the three evaluated CC-OCR tracks, excluding multilingual OCR. A separate 2.7B multilingual version scored 87.0 on MDPBench. The team also revised OCRBench v2.1 with manual corrections to 2,449 samples and task-aligned scoring metrics. Experiments on OCRBench v2.1, CC-OCR, an in-house KIE benchmark, OmniDocBench v1.6 and MDPBench showed state-of-the-art or highly competitive performance, according to the paper.

GuangJian Team researchers introduced PolyOCR-Venus, a family of optical character recognition (OCR) models, in a paper published on September 29 and submitted to Hugging Face on October 12. The models are designed to recognize, locate, parse and reason about text in complex visual settings. The authors say existing OCR systems often perform well on individual tasks but struggle to balance recognition, parsing and reasoning across different scenarios.

PolyOCR uses a shared instruction-following framework and a large-scale data engine that turns varied visual sources into quality-verified OCR training data. Its 2B and 9B models combine text recognition, grounding, document parsing, information extraction and OCR reasoning in one interface. The team also developed Competence-Guided Policy Optimization (CGPO), which combines verifier-based Group Relative Policy Optimization with on-policy distillation. It routes training samples according to teacher reliability and the difference between teacher and student competence.

The authors report that PolyOCR-9B scored 80.42 for English and 78.37 for Chinese on OCRBench v2.1, and 82.43 across the three evaluated CC-OCR tracks, excluding multilingual OCR. A separate 2.7B multilingual version scored 87.0 on MDPBench. The team also revised OCRBench v2.1 with manual corrections to 2,449 samples and task-aligned scoring metrics. Experiments on OCRBench v2.1, CC-OCR, an in-house KIE benchmark, OmniDocBench v1.6 and MDPBench showed state-of-the-art or highly competitive performance, according to the paper.