AI Paper Atlas

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

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 3 citations (OpenAlex)

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 changedWhereMeasuredEffect
use Serial MLP MTP head instead of Parallel separate LM head (ablation)P-MTP · PubTabNetaverage acceptance length: 6.28 -> 7.29Improved
use Serial DecoderLayer instead of Serial MLP (ablation)P-MTP · PubTabNetlatency: 5.53 -> 7.85Hurt
add dynamic loss weight instead of equal weight (ablation)P-MTP (K=9) · PubTabNetspeedup: 3.42 -> 4.83Improved
add dynamic loss weight instead of static decay (ablation)P-MTP (K=9) · PubTabNetspeedup: 4.49 -> 4.83Improved
add confidence-gated dynamic drafting (ablation)P-MTP (K=3) · PubTabNetspeedup: 2.86x to 3.43xImproved
add confidence-gated dynamic drafting (ablation)P-MTP (K=6) · PubTabNetspeedup: 4.10x to 4.57xImproved
add confidence-gated dynamic drafting (ablation)P-MTP (K=9) · PubTabNetspeedup: 4.83x to 5.24xImproved
combine sequential and retrospective constraints in loss weight (ablation)P-MTP · PubTabNetspeedup: 4.22 -> 4.83Improved
increase training look-ahead K from 9 to 18 (ablation)P-MTP · PubTabNetspeedup: 5.24x to 5.13xHurt
increase base model size from 2B to 8B (ablation)P-MTP · PubTabNetacceptance length: 8.60 to 11.67Improved
apply P-MTP in vLLMQwen3VL-2B · PubTabNet throughputTPS: 647 -> 1028Improved
apply P-MTP to document parsingQwen3VL-2B · Document parsingOverall: 86.31 -> 81.28Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy