Research map / Vision-language and multimodal / Multimodal LLMs
P-MTP: Efficient Document Parsing via Multi-Token Prediction with Progressive Depth Scaling
Le Xiang, Chenxi Zhai, Shu Wei, Jingjing Wu, Qunyi Xie, Xiao Tan, KunbinChen KunbinChen, Wei He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Serial MLP MTP head instead of Parallel separate LM head (ablation) | P-MTP · PubTabNet | average acceptance length: 6.28 -> 7.29 | Improved |
| use Serial DecoderLayer instead of Serial MLP (ablation) | P-MTP · PubTabNet | latency: 5.53 -> 7.85 | Hurt |
| add dynamic loss weight instead of equal weight (ablation) | P-MTP (K=9) · PubTabNet | speedup: 3.42 -> 4.83 | Improved |
| add dynamic loss weight instead of static decay (ablation) | P-MTP (K=9) · PubTabNet | speedup: 4.49 -> 4.83 | Improved |
| add confidence-gated dynamic drafting (ablation) | P-MTP (K=3) · PubTabNet | speedup: 2.86x to 3.43x | Improved |
| add confidence-gated dynamic drafting (ablation) | P-MTP (K=6) · PubTabNet | speedup: 4.10x to 4.57x | Improved |
| add confidence-gated dynamic drafting (ablation) | P-MTP (K=9) · PubTabNet | speedup: 4.83x to 5.24x | Improved |
| combine sequential and retrospective constraints in loss weight (ablation) | P-MTP · PubTabNet | speedup: 4.22 -> 4.83 | Improved |
| increase training look-ahead K from 9 to 18 (ablation) | P-MTP · PubTabNet | speedup: 5.24x to 5.13x | Hurt |
| increase base model size from 2B to 8B (ablation) | P-MTP · PubTabNet | acceptance length: 8.60 to 11.67 | Improved |
| apply P-MTP in vLLM | Qwen3VL-2B · PubTabNet throughput | TPS: 647 -> 1028 | Improved |
| apply P-MTP to document parsing | Qwen3VL-2B · Document parsing | Overall: 86.31 -> 81.28 | Hurt |
Closest papers by meaning
- Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training (CVPR 2026 · similarity 0.88)
- MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios (NeurIPS 2026 · similarity 0.87)
- Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing (CVPR 2026 · similarity 0.87)
- L-MTP: Leap Multi-Token Prediction Beyond Adjacent Context for Large Language Models (NeurIPS 2025 · similarity 0.86)
- MPDocBench-Parse: Benchmarking Practical Multi-page Document Parsing (NeurIPS 2026 · similarity 0.86)
- DocPTBench: Benchmarking End-to-End Photographed Document Parsing and Translation (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
