End-to-end document parsers typically serialize page layouts and regional contents into a single autoregressive sequence, forcing independent regions onto a decoding path that grows with total content. Crop-based two-stage parsers offer region-level parallelism but suffer from repeated visual prefills and fragmented page context.
To address this, researchers from Hugging Face propose PaDoc, a layout-grounded parser that treats the predicted layout as a branching structure over a shared page representation. By assuming region sufficiency, they derive a prefix-conditioned factorization that allows the layout stream and regional content branches to advance concurrently, reducing decoding depth to the longest layout-content path.
PaDoc is implemented within a single multimodal large language model (MLLM). It uses packed variable-length ancestor attention to maintain visibility during standard next-token training, and masked parallel decoding to create branches that the vLLM backend serves as concurrent requests with cache-resident shared-prefix reuse.
On the OmniDocBench Full benchmark, PaDoc achieves an Overall layout F1 of 91.1 and a top-tier Overall score of 94.24 among end-to-end parsers, along with the best Text Edit (0.038) and Formula CDM (95.59). On a 384-page subset with one A800 GPU, it is the fastest end-to-end parser at five concurrency levels, improving valid-page throughput by 67.4–118% and reducing P95 latency by 39.2–54.9% relative to a same-backbone Sequential SFT baseline.
The code is available on GitHub at https://github.com/Longin-Yu/Padoc.