PolyOCR-Venus Presents Unified Optical Character Recognition Foundation Models for Text-Centric Visual Intelligence.

Meta description: PolyOCR-Venus 2B and 9B models unify OCR tasks, use CGPO training, score 80.42 and 78.37 on OCRBench v2.1 English and Chinese, and 82.43 on CC-OCR.

Tags: PolyOCR-Venus, OCR models, CGPO, OCRBench v2.1, CC-OCR

Featured image: Abstract technology style with document grids, bounding boxes, and neural network nodes, no real people, no company logos, no text.

PolyOCR-Venus has been introduced as a family of optical character recognition (OCR) models for reading, locating, parsing, and reasoning over text in images. The release matters because it places several common document-processing tasks inside one instruction-following framework, reducing the need to treat each task as a separate pipeline.

Background: One Interface for Five OCR Tasks

OCR workloads often split into separate stages: one system reads text, another locates it, another parses document structure, another extracts specific fields, and another reasons across the content. PolyOCR-Venus frames these functions as a shared interface.

The 2B and 9B models cover text recognition, grounding, document parsing, information extraction, and OCR reasoning within the same instruction-following setup. That gives the project a common surface for tasks that are usually handled by different model types.

Development Process: Training Uses Teacher Guidance Sample by Sample

The training method is described as Competence-Guided Policy Optimization, or CGPO. It combines GRPO with on-policy distillation.

The approach adjusts teacher supervision per sample rather than applying the same guidance across every example. It weighs teacher reliability and the evolving teacher-student competence gap. In other words, the student model receives more or less teacher influence depending on whether the guidance is useful and how close the student is to the teacher in a given situation.

Latest Progress: Results at Modest Model Sizes

The reported results focus on compact model sizes. PolyOCR-9B scores 80.42 on the English side of OCRBench v2.1 and 78.37 on the Chinese side. On CC-OCR, it scores 82.43 across the three evaluated tracks, excluding multilingual OCR.

A dedicated 2.7B multilingual variant reaches 87.0 on MDPBench.

The evaluation layer also receives attention. OCRBench v2.1 revises 2,449 samples with manual verification and introduces task-aligned scoring for more reliable measurement.

Current Stage: Shared OCR Framework and Revised Benchmarks

PolyOCR-Venus currently sits at the intersection of two developments: a model family designed to handle multiple OCR tasks through one interface, and a benchmark that has been manually revised for more consistent evaluation. The next point to watch is how the 2B, 9B, and 2.7B multilingual variants perform when their shared task interface is tested across more OCR combinations.