PaDoc: Layout-Grounded Parallel Decoding for Document Parsing
Hao Yu, Jiabo Zhan, Kang Liu, Linnan Zhao, Dongxu Yue, Rui Chen, Jinglin Wang, Chong Sun, Chen Li, Jing Lyu, Chun Yuan
cs.AI
2026-08-06
PaDoc turns document layout into parallel content branches over a shared page prefix: same-backbone throughput up to 2x, P95 latency nearly halved on one A800, with top-tier OmniDocBench quality.
Document parsing turns a page image into structured output: layout regions, reading order, text, formulas, tables. The dominant end-to-end approach uses a single multimodal LLM (MLLM) to emit everything for a page as one autoregressive sequence, with layout boxes first and then the content of each region one after another. The problem is that many regions on a page are independent (a title in the top-left has no real dependency on a formula in the bottom-right), yet they must wait in line. Decoding length grows with the total content of the page, and pages with many regions get slow.
The alternative is crop-based two-stage parsing: detect the layout, crop each region, and recognize them separately. That does expose parallelism across regions, but every crop re-runs the visual encoder and loses page-level context.
PaDoc wants both at once: keep the full-page context of end-to-end parsing and the region-level parallelism of two-stage parsing.
The key observation is that a document's dependency structure is a tree, not a chain. A layout box Bk splits the page into regions, and the content Yk of region k depends only on the page image X and its own layout Bk, not on the contents of other regions. The paper calls this the region-sufficiency assumption: given the page and the layout, recognizing one region does not require having seen the others.
Under that assumption the joint probability factorizes so that a layout stream predicts the next box B(k+1) and a content branch predicts Yk, both conditioned on a shared prefix (the page image X plus layouts up to k). Because the two do not depend on each other, they can be decoded concurrently. The payoff is in decoding depth: sequential decoding scales with total content length, while PaDoc reduces it to the longest single layout-prefix-plus-content path.
This factorization is learned inside one MLLM, with no extra detector head and no draft model.
The upshot is no architecture change and no auxiliary heads, just turning serial decoding into layout-grounded parallel branches that ride on prefix caching vLLM already provides.
Layout analysis (OmniDocBench, Table 1): PaDoc reaches Overall F1 of 91.1, matching the best two-stage system PaddleOCR-VL 1.5 (91.2), and posts the highest Overall precision in the table at 93.3. It leads all four table-layout metrics decisively (IoU 92.7, F1 97.0).
End-to-end quality (Table 2, OmniDocBench v1.6 Full):
| System | Size | Overall | Text Edit down | Formula CDM | Table TEDS |
| MinerU2.5-Pro (two-stage) | 1.2B | 95.69 | 0.036 | 97.29 | 93.42 |
| HunyuanOCR 1.5 (end-to-end) | 1.0B | 94.74 | 0.039 | 94.50 | 93.67 |
| Qianfan-OCR (end-to-end) | 4.7B | 93.90 | 0.040 | 95.08 | 90.53 |
| PaDoc (end-to-end) | 2.1B | 94.24 | 0.038 | 95.59 | 90.94 |
PaDoc's Overall of 94.24 ranks second among end-to-end parsers, half a point behind HunyuanOCR 1.5 (94.74), but it is the best end-to-end system on Text Edit (0.038, lower is better) and Formula CDM (95.59). Table content (Table TEDS 90.94) is its weak spot, about 3 points behind HunyuanOCR. As a side note, general VLMs lag badly here: GPT-5.2 scores only 86.52 Overall, so document parsing is still the province of specialized models.
Serving efficiency (384-page subset, one A800, Table 3): against a same-backbone Sequential SFT baseline, throughput rises 67.4% to 118% and P95 latency falls 39.2% to 54.9%, peaking at 1.722 pages per second at 64-way concurrency. The counterintuitive part is that at 2.1B parameters, PaDoc beats the 1.0B HunyuanOCR 1.5 on both throughput and latency at every concurrency level, and outpaces the 0.7B MonkeyOCRv2 by a wide margin.
Document parsing is the front end of RAG pipelines, OCR services, and document digitization, so latency and throughput map directly to serving cost. PaDoc's contribution is not a new top score. It is a demonstration that layout-grounded parallel decoding can cut end-to-end parser serving cost substantially without touching the model architecture, adding heads, or pulling in a draft model, and it runs on stock vLLM.
For teams doing parsing at scale, the practical takeaway is that a single GPU handles nearly twice the pages while quality stays in the end-to-end top tier. It has not overtaken the best two-stage system (MinerU2.5-Pro), but it removes the extra encoding round-trip that two-stage pays.
The region-sufficiency assumption is an approximation, not a theorem. The paper itself notes in the appendix that the prefix-conditioned form it uses is strictly weaker than the full region-specific factorization and cannot recover it. Real cross-region dependencies, such as reading-order cues, cross-references, and spanning table cells, could in principle violate it, and the paper gives no error analysis for when the assumption breaks.
Quality is not best-in-class either. Overall 94.24 trails two-stage MinerU2.5-Pro (95.69) and GLM-OCR (95.15), and within end-to-end it trails HunyuanOCR 1.5. Table content recognition (Table TEDS) is an explicit weak point.
The efficiency baseline is the same-backbone Sequential SFT model, a clean ablation for isolating the parallel-decoding contribution. That means the throughput gain is attributable to parallel decoding itself, not to a stronger model. All serving numbers come from single-GPU, single-replica runs with no tensor parallelism, so real multi-node cluster behavior is untested.