PaDoc Decodes Document Layout and Content in Parallel Instead of Sequentially — 67-118% Higher Throughput, 39-55% Lower P95 Latency

arXiv / HuggingFace Daily Papers (14 upvotes)·medium signalXBlueskyLinkedInCopy link

PaDoc (arXiv 2608.06146, submitted August 6) treats predicted page layout as a branching structure over a shared page representation, letting layout and regional content decode simultaneously and cutting decoding depth to the longest layout-content path. On OmniDocBench Full it reports 91.1 overall layout F1, 94.24 overall among end-to-end parsers, 0.038 Text Edit and 95.59 Formula CDM; on a 384-page subset on a single A800 it improves valid-page throughput 67.4-118% and cuts P95 latency 39.2-54.9% versus the sequential baseline across five concurrency levels, with code released. Directly relevant to the same-week finding that PDF-to-markdown conversion is a top enterprise token expense.