HPD-Parsing: Hierarchical Parallel Document Parsing
The paper claims a document parser can remove much of its decoding bottleneck by splitting layout planning from parallel block parsing.
HPD-Parsing uses a main layout branch to organize the page, then assigns block-level content decoding to concurrent branches. It also adds progressive multi-token prediction to cut decoding steps inside each branch. The authors report 4,752 tokens per second, or 2.62x the throughput of the fastest existing document parsing model they compare against, while keeping competitive accuracy. HF Daily Papers' note
HPD-Parsing uses a main layout branch to organize the page, then assigns block-level content decoding to concurrent branches. It also adds progressive multi-token prediction to cut decoding steps inside each branch. The authors report 4,752 tokens per second, or 2.62x the throughput of the fastest existing document parsing model they compare against, while keeping competitive accuracy. HF Daily Papers' note
score 4