Infinity-Parser2 tops benchmarks with 87.6% on olmOCR-Bench
Outperforms DeepSeek-OCR-2 and uses multi-task RL for end-to-end parsing.
Researchers behind Infinity-Parser2 have introduced a new generation of document parsing AI that tackles the chronic shortage of high-quality annotated training data. They built a scalable synthesis engine pairing a controllable rendering framework with an iterative refinement loop, producing Infinity-Doc2-5M—a 5-million-sample bilingual (Chinese/English) corpus. Each sample includes element bounding boxes, canonical content forms (Markdown, HTML, LaTeX, SMILES, structured charts), and full-page reading order. This corpus is open-sourced, enabling further research.
Beyond data, the team implemented a verifiable multi-task reward system enabling Joint Reinforcement Learning across eight co-trained objectives: document parsing, layout analysis, table parsing, math formula parsing, chart parsing, chemical formula parsing, document VQA, and general multimodal understanding. Two variants are released: Infinity-Parser2-Flash, optimized for low-latency inference with a 3.68x throughput gain over the previous Infinity-Parser-7B, and Infinity-Parser2-Pro, engineered for precision-critical settings. Pro achieves state-of-the-art results: 87.6% on olmOCR-Bench and 74.3% on ParseBench, surpassing DeepSeek-OCR-2, PaddleOCR-VL-1.5, and MinerU2.5, with strong generalization to charts, chemical formulas, and document VQA.
- Open-sourced 5M bilingual document corpus (Infinity-Doc2-5M) with bounding boxes, multiple output formats, and reading order.
- Multi-task reinforcement learning across 8 objectives unifies perception, structure, and reasoning in a single optimization signal.
- Infinity-Parser2-Pro achieves 87.6% on olmOCR-Bench and 74.3% on ParseBench, outperforming DeepSeek-OCR-2 and PaddleOCR-VL-1.5.
Why It Matters
Enables accurate, low-latency document parsing for enterprise workflows, from OCR to chemical formula extraction.