PolyOCR-Venus: Unified OCR Foundation Models for Text-Centric Visual Intelligence
Moving beyond surface-level text transcription, PolyOCR introduces a unified multimodal foundation model family designed for text-centric visual intelligence. The framework integrates a curated data engine with Competence-Guided Policy Optimization, a training paradigm that routes samples between verifier-based group policy optimization and on-policy distillation according to teacher-student competence gaps. Alongside corrected evaluation benchmarks in OCRBench v2.1, the work establishes a coordinated recipe for dense recognition, document parsing, and visual reasoning.
Why it's worth reading
It offers a disciplined training recipe combining policy optimization with on-policy distillation, tackling the longstanding tension between transcription and reasoning in multimodal OCR.