PolyOCR-Venus: A Unified OCR Route to Text-Centric Visual Intelligence

Introduction

OCR is no longer limited to transcribing pixels into plain text. Practical systems must also identify where text appears, recover document structure, extract fields, and reason about text in relation to the surrounding image. PolyOCR-Venus addresses this broader requirement with a family of OCR foundation models designed to share one instruction-following interface across text-centric visual tasks.

What the project contributes

• A common task interface. The 2B and 9B models cover text recognition, grounding, document parsing, information extraction, and OCR reasoning within the same framework. This is a move away from treating each OCR capability as an isolated model or pipeline.
• A quality-oriented data engine. PolyOCR converts heterogeneous visual resources into OCR supervision and applies quality verification along the way. The supplied material does not disclose the complete data composition or scale, so the contribution should be understood as a more systematic supervision pipeline rather than simply a claim about having a larger dataset.
• Selective teacher guidance. The proposed Competence-Guided Policy Optimization, or CGPO, combines verifier-based Group Relative Policy Optimization with on-policy distillation. It routes supervision on a sample-by-sample basis using two signals: how reliable the teacher is on that sample and how large the current teacher–student competence gap is. The underlying idea is practical: teacher output is not equally useful everywhere, so distillation should not be applied uniformly.
• A revised benchmark. OCRBench v2.1 manually verifies and corrects 2,449 samples and introduces task-aligned scoring metrics. This matters because OCR comparisons can be distorted by annotation errors or by metrics that reward formatting behavior unrelated to the task being tested.

Reported results and limitations

The project page reports that PolyOCR-9B scores 80.42 on the English track and 78.37 on the Chinese track of OCRBench v2.1. It also reports 82.43 on the three evaluated CC-OCR tracks, excluding multilingual OCR. A dedicated 2.7B multilingual variant reaches 87.0 on MDPBench. The paper further evaluates PolyOCR on an in-house KIE Benchmark and OmniDocBench v1.6, but the supplied summary does not provide the full comparison tables. These figures therefore indicate strong reported performance, not proof of universal superiority across every document or language setting.

Why it matters

PolyOCR-Venus is notable because it treats the model, the training signal, and the benchmark as parts of one OCR system. The shared interface encourages transfer between recognition, parsing, and reasoning. CGPO attempts to reduce the downside of blindly copying a teacher by deciding when that supervision is useful. The benchmark revision, meanwhile, aims to make reported gains easier to interpret.

This direction is relevant to receipts, forms, tables, complex documents, and image-question answering where text cannot be separated from layout and context. Still, the available material does not establish how stable the models are on severely degraded images, unusual layouts, or broad multilingual deployment. Teams considering production use should validate the models on their own documents and weigh accuracy against inference cost and error tolerance.

Source: Hugging Face Daily Papers