PolyOCR-Venus: Unified OCR Foundation Models for Text-Centric Visual Intelligence
PolyOCR is pitched as one OCR model family for recognition, parsing, localization, and reasoning.
The report says PolyOCR uses a shared instruction-following setup and a data engine that turns varied visual material into verified OCR supervision. It introduces Competence-Guided Policy Optimization, mixing verifier-based GRPO with on-policy distillation based on teacher reliability and teacher-student gaps. The authors also revise OCRBench into v2.1 with corrected annotations and task-aligned scoring. They report state-of-the-art or highly competitive results across OCRBench v2.1, CC-OCR, KIE, OmniDocBench, and MDPBench. HF Daily Papers' note
The report says PolyOCR uses a shared instruction-following setup and a data engine that turns varied visual material into verified OCR supervision. It introduces Competence-Guided Policy Optimization, mixing verifier-based GRPO with on-policy distillation based on teacher reliability and teacher-student gaps. The authors also revise OCRBench into v2.1 with corrected annotations and task-aligned scoring. They report state-of-the-art or highly competitive results across OCRBench v2.1, CC-OCR, KIE, OmniDocBench, and MDPBench. HF Daily Papers' note
score 5