HPD-Parsing reaches 4,752 tokens/s by replacing page-wide autoregression

PaddlePaddle · hf · 2026-07-22

HPD-Parsing replaces page-wide autoregression with hierarchical parallel decoding

This paper tackles the sequential bottleneck in unified VLM-based document parsers.

• Existing parsers process the whole page jointly, but still generate outputs through one token-by-token autoregressive path, which becomes a bottleneck as documents get longer.
• The authors argue that document parsing has two different needs: global layout analysis and parallel block-level content parsing.
• Their method, HPD-Parsing, uses a main layout branch to organize the page structure and dynamically assigns block content to concurrent decoding branches.
• They further add progressive multi-token prediction (P-MTP) to cut the number of decoding steps inside each branch.
• On public benchmarks, the system reaches 4,752 tokens/sec, which is 2.62× faster than the fastest prior document parsing model and 3.06× faster than the vanilla autoregressive baseline.
• Accuracy remains competitive, suggesting hierarchical parallel decoding is a viable alternative to full-page autoregression for document parsing.

Related event: HPD-Parsing Achieves High-Throughput Document Parsing via Hierarchical Parallel Decoding(2 posts)→

Original post →

More from Research

• EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
• EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
• TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
• Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
• UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
• CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27

Research channel →