Today for AI
HOT RADAR
llmHEAT 6.7°

AI CURATED HOT STORY · 10/28/2026

Ant Group Releases PolyOCR: Unified 2B Model for Full-Scenario Document Parsing

1 reports archived1 independent sourcesupdated 10/28/2026, 17:55:00
Synthesis & Latest Updates
1 Sources Cross-Validated

Ant AI Security Lab introduced PolyOCR, a unified model that integrates text recognition, layout parsing, and field extraction using Capability-Guided Policy Optimization (CGPO) with online policy distillation and reinforcement learning. Despite its compact 2B parameter size, the model achieves SOTA performance on benchmarks like OCRBenchV2.1, addressing error accumulation issues in traditional segmented OCR pipelines for complex business documents.

LATEST/Ant AI Security Lab introduced PolyOCR, a unified model that integrates text recognition, layout parsing, and field extraction using Capability-Guided Policy Optimization (CGPO) with online policy distillation and reinforcement learning. Despite its compact 2B parameter size, the model achieves SOTA performance on benchmarks like OCRBenchV2.1, addressing error accumulation issues in traditional segmented OCR pipelines for complex business documents.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first
  1. 量子位 (微信公众号)T2·78 pts

    Ant Group Releases PolyOCR: Unified 2B Model for Full-Scenario Document Parsing

    • Unified architecture outputs characters, positions, and field relationships simultaneously to prevent error accumulation in segmented OCR
    • Introduces CGPO algorithm combining online policy distillation and reinforcement learning to enhance layout and semantic understanding
    • Constructed 60 million high-quality supervised data samples focusing on long-tail layouts and complex tables
Ant Group Releases PolyOCR: Unified 2B Model for Full-Scenario Document Parsing | Today for AI