PolyOCR-Venus: Unified OCR Foundation Models for Text-Centric Visual Intelligence

PolyOCR-Venus. Unified OCR Foundation Models for Text-Centric Visual Intelligence. By GuangJian Team and twenty-four others, from the GuangJian Team at Ant Group. Published on the twenty-ninth of September. Its code repository is worked out: about seven hundred and fifty source files, with a readme and a licence.

Optical character recognition is no longer only about transcribing plain text. It is moving toward general visual intelligence: models are expected to recognise text, localise it, and reason over it inside complex visual scenes. Existing OCR systems tend to excel at only some of these tasks, and they struggle to balance recognition, parsing and reasoning across different scenarios.

This report presents PolyOCR, a family of unified OCR foundation models at several scales. Rather than one specialist per task, a single shared instruction-following framework covers them, fed by a large-scale data engine that turns heterogeneous visual resources into quality-verified OCR supervision.

Training adds a technique the authors call Competence-Guided Policy Optimization. It combines two things: verifier-based Group Relative Policy Optimization, and on-policy distillation from a teacher. Which one a given sample goes through is decided by sample-wise routing, based on how reliable the teacher is and on the competence gap between teacher and student.

Alongside the models comes OCRBench v2.1, a revision of OCRBench v2. It carries manually verified annotation corrections and scoring metrics aligned to each task.

Experiments span OCRBench v2.1, CC-OCR, an in-house KIE Benchmark, OmniDocBench v1.6 and MDPBench. Across them, PolyOCR achieves state-of-the-art or highly competitive performance. The abstract reports that result in those terms, without giving the individual scores.

The organising idea is breadth without fragmentation. One instruction framework, one data engine and one training recipe are asked to cover reading, parsing and reasoning at once, and the routing between reinforcement and distillation decides, sample by sample, which kind of signal the student needs.