Youtu-Parsing: Perception, Structuring and Recognition via High-Parallelism Decoding

TL;DR

Youtu-Parsing introduces a decoupled, prompt-guided architecture that combines a NaViT-based visual encoder with Youtu-LLM-2B for layout-aware, region-prompted decoding. By implementing dual-track high-parallelism decoding—Token Parallelism ($5\text{--}11\times$ speedup) and Query Parallelism ($2\times$ throughput)—the model achieves state-of-the-art accuracy on OmniDocBench and olmOCR-bench while maintaining competitive latency for large-scale deployment. The framework supports diverse elements (text, formulas, tables, charts, seals, and hierarchical structures) and demonstrates robustness to multilingual and handwritten content. Extensive training (pre-training, SFT, RL) and synthetic/open-source data pipelines underpin strong generalization, improving end-to-end document understanding for real-world applications.

Abstract

This paper presents Youtu-Parsing, an efficient and versatile document parsing model designed for high-performance content extraction. The architecture employs a native Vision Transformer (ViT) featuring a dynamic-resolution visual encoder to extract shared document features, coupled with a prompt-guided Youtu-LLM-2B language model for layout analysis and region-prompted decoding. Leveraging this decoupled and feature-reusable framework, we introduce a high-parallelism decoding strategy comprising two core components: token parallelism and query parallelism. The token parallelism strategy concurrently generates up to 64 candidate tokens per inference step, which are subsequently validated through a verification mechanism. This approach yields a 5--11x speedup over traditional autoregressive decoding and is particularly well-suited for highly structured scenarios, such as table recognition. To further exploit the advantages of region-prompted decoding, the query parallelism strategy enables simultaneous content prediction for multiple bounding boxes (up to five), providing an additional 2x acceleration while maintaining output quality equivalent to standard decoding. Youtu-Parsing encompasses a diverse range of document elements, including text, formulas, tables, charts, seals, and hierarchical structures. Furthermore, the model exhibits strong robustness when handling rare characters, multilingual text, and handwritten content. Extensive evaluations demonstrate that Youtu-Parsing achieves state-of-the-art (SOTA) performance on both the OmniDocBench and olmOCR-bench benchmarks. Overall, Youtu-Parsing demonstrates significant experimental value and practical utility for large-scale document intelligence applications.

Figures (26)

• Figure 1: Performance of Youtu-Parsing on OmniDocBench v1.5. Youtu-Parsing surpasses both general-purpose vision-language models and specialized domain models, setting new benchmarks in text recognition, formula recognition, table recognition, and reading order prediction across multiple evaluation tasks.
• Figure 2: Parsing Samples
• Figure 3: Youtu-Parsing Framework Overview. The system implements a three-stage cascaded pipeline: (1) Shared Visual Feature Extraction, utilizing a NaViT encoder to extract multi-scale, high-fidelity visual representations; (2) Layout Analysis, which performs cross-modal fusion to precisely identify and categorize structural document elements; (3) Region-prompted Decoding, where localized regions are processed via customized block queries within an LLM module for fine-grained recognition. By decoupling spatial perception from semantic parsing, Youtu-Parsing supports an expansive range of document elements while integrating a high-parallelism decoding strategy to achieve superior inference efficiency.
• Figure 4: The framework of the parallel decoding strategy.
• Figure 5: The framework of hierarchical structure analysis.

Theorems & Definitions (1)

• proof