AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

MaTVLM: Hybrid Mamba-Transformer for Efficient Vision-Language Modeling

Yingyue Li, Bencheng Liao, Wenyu Liu, Xinggang Wang

ICCV 2025 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 8 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 25% Mamba-2 layers instead of 50% (ablation)MaTVLM · 9-benchmark averageaverage score: 62.3 vs 60.1; surpassing the 50% ratio by 2.2 pointsImproved
use 25% Mamba-2 layers instead of 12.5% (ablation)MaTVLM · 9-benchmark averageaverage score: 0.5 points lower for 12.5% (61.8 vs 62.3)Improved
distribute Mamba-2 layers evenly instead of all at the beginning (ablation)MaTVLM · 9-benchmark averageaverage score: 62.3 vs 60.4; 1.9 pointsImproved
distribute Mamba-2 layers evenly instead of all in the middle (ablation)MaTVLM · 9-benchmark averageaverage score: 62.3 vs 58.9; 3.4 pointsImproved
place Mamba-2 layers all at the end instead of evenly (ablation)MaTVLM · distillationresponse qualityHurt
use L_prob instead of L_ce only (ablation)MaTVLM · 9-benchmark averageaverage score: 55.6 -> 61.4; 5.8-point increaseImproved
add L_layer to L_prob (ablation)MaTVLM · 9-benchmark averageaverage score: 61.4 -> 62.3; 0.9 points higherImproved
add L_ce to L_prob + L_layer (ablation)MaTVLM · 9-benchmark averageaverage score: 62.3 -> 61.3; 1.0-point dropHurt
distill into hybrid Mamba-2 model instead of teacher TinyLLaVAMaTVLM · MME-Pperception score: 1466.4 -> 1484.1; 17.6-point improvementImproved
replace transformer layers with Mamba-2 (25%) for inferenceMaTVLM vs TinyLLaVA · generation speedspeedup: 3.6x fasterImproved
replace transformer layers with Mamba-2 for memoryMaTVLM vs TinyLLaVA · 16,384 tokenspeak GPU memory: 27.5% reductionImproved

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