Infinity-Parser2: 5M Infinity-Doc2 dataset and SOTA parsing

Infinity-Parser2, described in an arXiv technical report submitted 8 Jul 2026, pairs a controllable data-synthesis pipeline with multi-task reinforcement learning to tackle document parsing. The project includes Infinity-Doc2-5M, a 5-million-sample bilingual (Chinese/English) corpus, and two model variants: Infinity-Parser2-Flash and Infinity-Parser2-Pro.

What was released and what does it contain?

They released Infinity-Doc2-5M, a 5-million-sample bilingual corpus (Chinese/English) annotated with element bounding boxes, canonical content forms, and full-page reading order, plus two model variants under a shared architecture. The dataset includes canonical content forms such as Markdown, HTML, LaTeX, SMILES, and structured charts and is the foundation for training the models described in the technical report.

The release bundles a controllable rendering framework and an iterative refinement loop used to synthesize diverse document types with faithful annotations. The two model variants are Infinity-Parser2-Flash, optimized for low-latency inference, and Infinity-Parser2-Pro, engineered for precision-critical settings.

How does Infinity-Parser2 train and what tasks does it cover?

Infinity-Parser2 uses a verifiable, multi-task reward system to enable 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. The report frames these objectives as a single optimization signal that unifies perception, structure, and reasoning.

Training relies on the scalable synthesis engine that pairs controllable rendering with iterative refinement to produce the Infinity-Doc2-5M corpus. The corpus supplies element bounding boxes, canonical text representations (Markdown, HTML, LaTeX, SMILES, structured charts), and full-page reading order annotations, which together support end-to-end parsing tasks spanning visual layout and structured content extraction.

What performance and efficiency claims does the report make?

Infinity-Parser2-Flash delivers a 3.68x throughput gain over Infinity-Parser-7B, positioning it for low-latency inference use cases. Infinity-Parser2-Pro reaches 87.6% on olmOCR-Bench and 74.3% on ParseBench, surpassing DeepSeek-OCR-2, PaddleOCR-VL-1.5, and MinerU2.5 according to the report.

Those two specific performance figures serve as the clearest, source-attributed benchmarks: 87.6% on olmOCR-Bench and 74.3% on ParseBench for Infinity-Parser2-Pro, and a 3.68x throughput improvement for Infinity-Parser2-Flash compared with Infinity-Parser-7B.

Why it matters

The combination of a large, openly released synthetic corpus and a joint-reward reinforcement learning approach tackles a persistent bottleneck: the scarcity of faithfully annotated parsing corpora. By publishing Infinity-Doc2-5M and claiming both higher accuracy on established benchmarks and substantially higher inference throughput for a latency-optimized variant, the project addresses both data and deployment constraints that have limited document parsing systems.

Consolidating layout, content extraction, formula parsing, chart parsing, chemical parsing, and VQA into a single optimization signal also signals a shift from many single-task pipelines toward unified, multi-capability models for document understanding.

What to watch

Look for external evaluations and code or data releases tied to the arXiv entry (arXiv:2607.07836) to verify the claimed 87.6% on olmOCR-Bench, 74.3% on ParseBench, and the 3.68x throughput gain. The availability and adoption of Infinity-Doc2-5M will determine how broadly the community can reproduce or build on the reported results.

References: Infinity-Parser2 Technical Report, arXiv:2607.07836, submitted 8 Jul 2026.