HuggingFace Papers
8/10 signal
HPD-Parsing: Hierarchical Parallel Document Parsing
infrastructureresearchmodels
Summary
This paper introduces HPD-Parsing, a document parsing model that replaces full-page autoregressive generation with a Hierarchical Parallel Decoding (HPD) paradigm. A main layout branch coordinates the document structure and dynamically assigns block-level content decoding to concurrent branches, while Progressive Multi-Token Prediction (P-MTP) accelerates generation within each branch. On public benchmarks, HPD-Parsing achieves a throughput of 4,752 tokens per second, a 3.06x improvement over its autoregressive baseline and 2.62x over the fastest existing model, while maintaining competitive parsing accuracy. The work establishes HPD as an effective alternative for efficient, unified document parsing.
Problem
Recent Vision-Language Model (VLM)-based document parsers have improved accuracy by modeling diverse elements in a unified framework. However, they generate output via a single, token-by-token autoregressive trajectory. This creates a sequential bottleneck, where inference latency grows with document length, making it difficult to meet the efficiency demands of large-scale document parsing.
Method
The proposed method, HPD-Parsing, uses a Hierarchical Parallel Decoding paradigm built on an InternVL3.5-1B backbone. A main layout branch generates the global document structure and dynamically forks concurrent content branches to parse localized regions in parallel, reusing the shared-prefix KV cache. Within each branch, Progressive Multi-Token Prediction (P-MTP) is used to predict and verify multiple tokens per step. The model is trained using a three-stage adaptation strategy with data from an automated difficulty-aware curation pipeline and evaluated on the OmniDocBench v1.6 benchmark.
Details
Accuracy on OmniDocBench v1.6:
HPD-Parsing achieves an overall score of 94.91, establishing a new state-of-the-art among unified parsers and remaining competitive with leading pipeline-based models.
| Model | Size | Overall ↑ | TextEdit ↓ | FormulaCDM ↑ | TableTEDS ↑ | ReadOrderEdit ↓ |
|---|---|---|---|---|---|---|
| HPD-Parsing (Ours) | 1B | 94.91 | 0.039 | 97.28 | 91.35 | 0.124 |
| PaddleOCR-VL-1.6 (Pipeline SOTA) | 0.9B | 96.3 | 0.032 | 97.5 | 94.8 | 0.127 |
| MinerU2.5-Pro (Pipeline) | 1.2B | 95.75 | 0.036 | 97.45 | 93.42 | 0.120 |
| GLM-OCR (Pipeline) | 0.9B | 95.22 | 0.044 | 97.18 | 92.83 | 0.133 |
| HunyuanOCR-1.5 (Unified) | 1B | 94.74 | 0.039 | 94.50 | 93.67 | 0.129 |
| Qianfan-OCR (Unified) | 4B | 93.90 | 0.040 | 95.08 | 90.53 | 0.130 |
Inference Efficiency on OmniDocBench v1.6 (Batch Size=512):
HPD-Parsing achieves the highest throughput among all compared models, significantly outperforming its autoregressive baseline and other specialized VLMs.
| Model | Size | Avg Input Tokens | TPS ↑ | PPS ↑ |
|---|---|---|---|---|
| HPD-Parsing (Ours) | 1B | 4809.3 | 4752.1 | 2.68 |
| Baseline (Autoregressive) | 1B | 4809.3 | 1554.8 | 1.02 |
| DeepSeek-OCR-2 (Unified) | 3B-A0.5B | 1100.2 | 2932.1 | 2.05 |
| Unlimited OCR (Unified) | 3B-A0.5B | 1485.6 | 2901.5 | 2.03 |
| GLM-OCR (Pipeline) | 0.9B | - | 2133.8 | 1.86 |
| MinerU2.5-Pro (Pipeline) | 1.2B | - | 1890.3 | 1.58 |
- Compared to its autoregressive baseline, HPD-Parsing delivers a 3.06x improvement in Tokens Per Second (TPS) and a 2.62x improvement in Pages Per Second (PPS).
- It achieves 1.62x the TPS of DeepSeek-OCR-2, the fastest prior model.
- The efficiency advantage grows with output length. For the longest documents, HPD-Parsing achieves up to 18.04x fewer decoding steps, 3.67x higher request throughput, and 5.80x lower single-request latency than the baseline.
- P-MTP achieves an average accepted length of 6.6 tokens per decoding step.
Model and Training:
- The model is built on InternVL3.5-1B, with a 0.3B InternViT visual encoder and a 0.8B LLM decoder (Qwen3-0.6B architecture).
- Training follows a three-stage strategy: 1) Full-page pre-training on 2.8M samples for general capability. 2) Adaptation to HPD on 100K curated hard-case samples. 3) Reward-guided optimization on 600 representative hard cases.
- The data curation pipeline involves feature-based clustering, multi-model annotation (PaddleOCR-VL-1.5, MinerU-2.5 Pro), difficulty estimation, VLM-based refinement, and distribution balancing.
What's new
The paper introduces Hierarchical Parallel Decoding (HPD), a novel decoding paradigm for document parsing. It reframes the task from a single, full-page sequential generation problem into one of global layout coordination and concurrent, localized content decoding, fundamentally changing the inference process to improve throughput.
Conclusion
HPD-Parsing demonstrates that hierarchical parallel decoding is an effective alternative to conventional full-page autoregressive generation for document parsing. By combining global layout coordination with localized parallel decoding, the model achieves state-of-the-art throughput—3.06x faster than its autoregressive baseline—while maintaining competitive parsing accuracy. This approach opens a new direction for developing efficient, unified document parsing systems.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.