Layout Grounded Parallel Decoding for Documents
Listen to the summary
Uses a voice available on your device
Audio options
On this page
Key Takeaways
- End-to-end document parsers are typically sequential, creating artificial dependencies, while crop-based two-stage parsers fragment page context.
- PaDoc uses a prefix-conditioned factorization where a layout stream predicts the next region and a content branch predicts regional content.
- The model implements a purpose-built ancestor-attention mask during training, realized via packed variable-length attention.
- On OmniDocBench Full, PaDoc attains an Overall layout F1 of 91.1 and an Overall score of 94.24.
Summary & Methodology Analysis
End-to-end document parsers are typically autoregressive and sequential, which creates artificial dependencies and inefficiency, whereas crop-based two-stage parsers fragment page context and require repeated visual prefills. This paper addresses the need for a document-parsing formulation that exposes region-level parallelism while maintaining a shared page-level context. The authors formulate a prefix-conditioned factorization where a layout stream predicts the next region and a content branch predicts the regional content from the shared prefix of the image and the layout generated so far. To support this, they implement a purpose-built ancestor-attention mask during training that allows nodes to attend to the document image and layout ancestors while excluding sibling contents, realized via packed variable-length attention. The model is trained using a standard next-token objective on a packed sequence containing the shared multimodal prefix and independent branched content.
For inference, PaDoc performs parallel serving by instantiating branch-isolated content streams as concurrent requests, using shared-prefix KV reuse to maintain full-page visual context without repeated visual prefills. This system design allows the backend to handle requests efficiently despite the architectural shift away from strictly sequential processing. The models and tools involved in the broader ecosystem include OmniDocBench, Qwen3-VL-2B, vLLM, HunyuanOCR-1.5, MonkeyOCRv2, Donut, Pix2Struct, Nougat, SmolDocling, olmOCR, Qianfan-OCR, DocLayout-YOLO, PP-DocLayout, DocBed, Dolphin, HSD, SPTS v2, SoT, APAR, PASTA, ASPD, SPRINT, Hogwild! Inference, and DeFT.
Regarding performance and limitations, PaDoc attains an Overall layout F1 of 91.1 and an Overall score of 94.24 on OmniDocBench Full. However, the serving throughput can be impacted by scheduling overhead and imbalances among branches of unequal length. The paper does not specify hardware requirements, exact dollar costs, or training durations beyond what is listed.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What is the main problem addressed by the paper?
End-to-end document parsers are typically autoregressive and sequential, which creates artificial dependencies and inefficiency, whereas crop-based two-stage parsers fragment page context and require repeated visual prefills.
Q2. What is the core solution proposed by the authors?
A document-parsing formulation that exposes region-level parallelism while maintaining a shared page-level context using a prefix-conditioned factorization.
Q3. What are the primary performance results reported?
On OmniDocBench Full, PaDoc attains an Overall layout F1 of 91.1 and an Overall score of 94.24.
Q4. How does the layout stream and content branch interact in PaDoc?
A layout stream predicts the next region and a content branch predicts the regional content from the shared prefix of the image and the layout generated so far.
Q5. What specific mask is used during training?
A purpose-built ancestor-attention mask that allows nodes to attend to the document image and layout ancestors while excluding sibling contents, realized via packed variable-length attention.
Q6. How is parallel serving performed at inference time?
By instantiating branch-isolated content streams as concurrent requests, using shared-prefix KV reuse to maintain full-page visual context without repeated visual prefills.
Q7. What objective function is used for training the model?
A standard next-token objective on a packed sequence containing the shared multimodal prefix and independent branched content.
Q8. What are the known limitations of the serving approach?
The serving throughput can be impacted by scheduling overhead and imbalances among branches of unequal length.
Q9. Which models and datasets are referenced in relation to the study?
The paper references OmniDocBench, Qwen3-VL-2B, vLLM, HunyuanOCR-1.5, MonkeyOCRv2, Donut, Pix2Struct, Nougat, SmolDocling, olmOCR, Qianfan-OCR, DocLayout-YOLO, PP-DocLayout, DocBed, Dolphin, HSD, SPTS v2, SoT, APAR, PASTA, ASPD, SPRINT, Hogwild! Inference, and DeFT.