PaDoc: Layout-Grounded Parallel Decoding for Document Parsing
PaDoc tries to make end-to-end document parsing faster by decoding layout and regional content in parallel instead of one long sequence.
The paper frames the problem as a tradeoff between full-page parsers, which serialize everything, and crop-based parsers, which parallelize regions but lose efficiency and context. PaDoc uses a predicted layout as a branching structure over a shared page representation, letting content branches advance concurrently. On OmniDocBench Full, it reports 91.1 layout F1 and a 94.24 overall end-to-end parser score. In a 384-page test on one A800 GPU, it reports 67.4-118% higher valid-page throughput and 39.2-54.9% lower P95 latency than a same-backbone sequential baseline. HF Daily Papers' note
The paper frames the problem as a tradeoff between full-page parsers, which serialize everything, and crop-based parsers, which parallelize regions but lose efficiency and context. PaDoc uses a predicted layout as a branching structure over a shared page representation, letting content branches advance concurrently. On OmniDocBench Full, it reports 91.1 layout F1 and a 94.24 overall end-to-end parser score. In a 384-page test on one A800 GPU, it reports 67.4-118% higher valid-page throughput and 39.2-54.9% lower P95 latency than a same-backbone sequential baseline. HF Daily Papers' note
score 5