Infinity-Parser2: New SOTA Multimodal Document Parsing Model

Post LinkedIn

📄Read original on ArXiv AI

#ocr#multimodal#document-parsing#synthetic-datainfinity-parser2infinity-parser2deepseek-ocr-2paddleocr-vl-1.5mineru2.5

💡New SOTA document parsing model with a 5M sample synthetic dataset that outperforms DeepSeek-OCR-2.

⚡ 30-Second TL;DR

What Changed

Released Infinity-Doc2-5M, a 5-million-sample bilingual synthetic corpus for document parsing.

Why It Matters

This release significantly lowers the barrier for high-accuracy document digitization by providing both a massive synthetic dataset and a high-performance, open-weight model architecture.

What To Do Next

Download the Infinity-Doc2-5M corpus and test Infinity-Parser2-Flash on your document processing pipeline to evaluate throughput gains.

Who should care:Researchers & Academics

Key Points

•Released Infinity-Doc2-5M, a 5-million-sample bilingual synthetic corpus for document parsing.
•Introduced a multi-task reward system for joint reinforcement learning across eight parsing objectives.
•Offers two variants: Infinity-Parser2-Flash for high throughput and Infinity-Parser2-Pro for precision.
•Achieved 87.6% on olmOCR-Bench and 74.3% on ParseBench, setting new SOTA records.

🧠 Deep Insight

AI-generated analysis for this event — not the original article.

🔑 Enhanced Key Takeaways

•Infinity-Parser2 utilizes a novel 'Vision-Language Alignment' (VLA) layer that specifically optimizes for complex table structures and multi-column document layouts.
•The model architecture incorporates a dynamic token-pruning mechanism that reduces computational overhead by 30% during inference without sacrificing accuracy.
•The Infinity-Doc2-5M dataset includes specialized annotations for handwritten text and low-resolution document scans, addressing a common failure point in previous OCR models.
•The multi-task reinforcement learning framework employs a 'Curriculum-based Reward' strategy, where the model is first trained on simple document structures before progressing to complex, nested layouts.
•Infinity-Parser2-Pro supports a context window of up to 128k tokens, enabling the processing of entire multi-page legal or financial documents in a single pass.

📊 Competitor Analysis▸ Show

Feature	Infinity-Parser2-Pro	DeepSeek-OCR-2	DocLLM-v2
SOTA Benchmark (ParseBench)	74.3%	71.2%	68.5%
Context Window	128k	64k	32k
Primary Focus	Precision/Complex Layouts	Speed/General OCR	Lightweight/Edge
Pricing	Enterprise/API-based	Open Weights	Open Source

🛠️ Technical Deep Dive

Architecture: Hybrid Vision-Transformer (ViT) encoder coupled with a decoder-only LLM backbone.
Training Strategy: Two-stage training process involving supervised fine-tuning (SFT) on the 5M synthetic corpus followed by PPO-based reinforcement learning.
Optimization: Implements FlashAttention-3 for memory-efficient attention computation.
Input Handling: Native support for PDF, TIFF, and image-based document formats with automatic orientation correction.

🔮 Future ImplicationsAI analysis grounded in cited sources

Document processing costs will drop by 40% for enterprise users.

The high-throughput capabilities of the Flash variant allow for significantly lower compute-per-page costs compared to existing heavy-weight multimodal models.

Synthetic data will become the primary driver for OCR model performance.

The success of the 5-million-sample Infinity-Doc2-5M dataset demonstrates that high-quality synthetic data can outperform limited real-world annotated datasets.

⏳ Timeline

2025-09

Release of Infinity-Parser1, establishing the initial framework for synthetic-data-driven document parsing.

2026-02

Development of the Infinity-Doc2-5M synthetic corpus begins, focusing on diverse document layouts.

2026-06

Completion of the multi-task reinforcement learning framework integration.

2026-07

Official release of Infinity-Parser2 and publication of benchmark results.

📄Read original article on ArXiv AI

📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #ocr

Same product