PaDoc: Layout-Grounded Parallel Decoding for Document Parsing
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - PaDoc is an end-to-end document parser that treats predicted layout as a branching structure over a shared page representation, letting layout and per-region content decode in parallel instead of as one long autoregressive sequence. It matters because it removes the speed penalty of end-to-end parsers while keeping full-page context, hitting top-tier accuracy and large throughput gains.
- Under a region-sufficiency assumption, the authors derive a prefix-conditioned factorization so the layout stream and regional content branches advance concurrently, cutting decoding depth to the longest layout-content path rather than total content length.
- Implemented inside a single MLLM: packed variable-length ancestor attention preserves visibility under standard next-token training, and masked parallel decoding spawns branches served by vLLM as concurrent requests with cache-resident shared-prefix reuse.
- On OmniDocBench Full: 91.1 Overall layout F1, 94.24 Overall score among end-to-end parsers, 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 across five concurrency levels — 67.4–118% higher valid-page throughput and 39.2–54.9% lower P95 latency vs. a same-backbone Sequential SFT baseline; code released on GitHub.
Sources (1)
PaDoc: Layout-Grounded Parallel Decoding for Document Parsing
TL;DR - PaDoc is an end-to-end document parser that treats predicted layout as a branching structure over a shared page representation, letting layout and per-region content decode in parallel instead of as one long autoregressive sequence. It matters because it removes the speed penalty of end-to-end parsers while keeping full-page context, hitting top-tier accuracy and large throughput gains.
- Under a region-sufficiency assumption, the authors derive a prefix-conditioned factorization so the layout stream and regional content branches advance concurrently, cutting decoding depth to the longest layout-content path rather than total content length.
- Implemented inside a single MLLM: packed variable-length ancestor attention preserves visibility under standard next-token training, and masked parallel decoding spawns branches served by vLLM as concurrent requests with cache-resident shared-prefix reuse.
- On OmniDocBench Full: 91.1 Overall layout F1, 94.24 Overall score among end-to-end parsers, 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 across five concurrency levels — 67.4–118% higher valid-page throughput and 39.2–54.9% lower P95 latency vs. a same-backbone Sequential SFT baseline; code released on GitHub.