Search papers, labs, and topics across Lattice.
This paper introduces PaDoc, a layout-grounded document parser that optimizes decoding by treating the predicted layout as a branching structure, allowing for concurrent processing of layout and regional content. By leveraging a prefix-conditioned factorization, PaDoc reduces decoding depth and enhances throughput without sacrificing context, achieving state-of-the-art performance on the OmniDocBench Full benchmark. The results show that PaDoc not only outperforms existing end-to-end parsers in terms of layout F1 score and latency but also significantly improves valid-page throughput by up to 118% compared to traditional sequential methods.
PaDoc achieves a remarkable 94.24 F1 score among end-to-end parsers while being the fastest parser at five concurrency levels, revolutionizing document parsing efficiency.
End-to-end document parsers provide a unified interface, but serialize page layouts and regional contents into one autoregressive sequence. This formulation forces independent regions onto a decoding path whose length grows with the total content, whereas crop-based two-stage parsers expose region-level parallelism at the cost of repeated visual prefills and fragmented page context. To retain full-page context while removing dependencies, we propose PaDoc, a layout-grounded parser that treats the predicted layout as a branching structure over a shared page representation. Under a region-sufficiency assumption, we derive a prefix-conditioned factorization in which the layout stream and regional content branches advance concurrently, reducing the decoding depth to the longest layout-content path. We realize this factorization within a single MLLM: packed variable-length ancestor attention preserves the visibility under standard next-token training, while masked parallel decoding creates branches that the evaluated vLLM backend serves as concurrent requests with cache-resident shared-prefix reuse. On OmniDocBench Full, PaDoc attains an Overall layout F1 of 91.1 and, among end-to-end parsers, a top-tier Overall score of 94.24 together with the best Text Edit (0.038) and Formula CDM (95.59). On a 384-page subset and 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. Code is available at https://github.com/Longin-Yu/Padoc
Tsinghua AI