AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

DTP: Delta-Guided Two Stage Pruning for Mamba-based Multimodal Large Language Models

Seong Yeol Park, Min Jung Kwon, XIANGHUA PIAO, Yeong Hyeon Gu

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DTP, a delta-guided two-stage visual token pruning method that progressively removes redundant visual tokens in Mamba-based multimodal LLMs to cut prefill cost.

Problem

Mamba-based multimodal LLMs are slowed by redundant visual tokens inflating inference cost, with prefill accounting for most inference time.

Approach

Progressively prunes visual tokens in two stages, selective pruning at an early layer followed by complete pruning at a late layer, with pruning guided by the Mamba delta (step-size) parameter.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use delta_t as token importance instead of C_t (ablation)DTP on Cobra · TextVQA (~50% FLOPs)accuracy: 44.9 -> 56.1Improved
use delta_t instead of y_t (ablation)DTP on Cobra · TextVQAaccuracy: 48.0 -> 56.1Improved
restrict top-k pruning to visual tokens only (do not prune text tokens) (ablation)DTP on Cobra · TextVQAaccuracy: 10.3 -> 56.1Improved
top-k delta selection instead of random visual token selection (ablation)DTP on Cobra · TextVQAaccuracy: 53.7 -> 56.1Improved
add complete pruning of visual tokens at late layer 45 (ablation)DTP on Cobra · GQA/TextVQA/VizWiz (~50% FLOPs)accuracy and FLOPs: FLOPs 1.26 -> 0.97; GQA 61.5 -> 61.4Improved
add complete late pruning (ablation)DTP on RoboMamba · GQA/POPE/OKVQAaccuracy and FLOPs: FLOPs 0.44 -> 0.34; GQA 55.1 -> 54.9Improved
DTP two-stage delta-guided pruning vs FastV/DART at ~70% FLOPsCobra · six-benchmark averagescore: 65.3 (FastV), 65.2 (DART) -> 65.8Improved
DTP at ~50% FLOPs vs other pruningCobra · six-benchmark averagescore: 64.9 vs FastV 64.7, DART 64.3, VTW 54.0Mixed
fix VTW withdrawal layer at 32 instead of 45VTW on Cobra · six-benchmark averagescore: 65.7 -> 54.0Hurt
DTP pruning on RoboMamba (256 tokens) at ~50% FLOPsRoboMamba · five-benchmark averagescore: 67.0 -> 65.9Mixed
apply DTP to reduce prefillCobra · POPE latencyprefill ms: 98.04 -> 61.54Improved

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