AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

AIM: Adaptive Inference of Multi-Modal LLMs via Token Merging and Pruning

Yiwu Zhong, Zhuoming Liu, Yin Li, Liwei Wang

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 32 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
merge tokens within frames (spatial) rather than across frames (ablation)AIM token merging · VideoMME (25% retention)accuracy: 55.8 -> 58.0Improved
merge tokens within frames (spatial) rather than across frames (ablation)AIM token merging · VideoMME (3.1% retention)accuracy: 47.4 -> 52.3Improved
do not prune text tokens during LLM pruning (ablation)AIM · VideoMMEaccuracy: 45.7 -> 58.2Improved
prune all visual tokens after layer 22 instead of layer 8 (ablation)AIM scheduler · VideoMMEaccuracy: 41.9 -> 58.1Improved
remove visual tokens after layer 15 instead of layer 22 (ablation)AIM scheduler · VideoMMEaccuracy: 58.1 -> 54.3Hurt
reduce token-merging retention from 100% to 25% (ablation)AIM token merging · VideoMMEaccuracy: 58.2 -> 58.0No clear effect
reduce token-merging retention from 25% to 1.6% (ablation)AIM token merging · VideoMMEaccuracy: 58.0 -> 50.9Hurt
start pruning at layer 14 instead of 21 (l2=22) (ablation)AIM pruning scheduler · VideoMMEaccuracy: 58.1 -> 58.2No clear effect
sample 192 frames instead of 32 with AIM (ablation)AIM · MLVUaccuracy: 63.7 -> 69.3Improved
add AIM to LLaVA-OV with 128 frames (ablation)LLaVA-OV · MLVUaccuracy: 67.7 -> 69.0Improved
apply AIM instead of FastV on Qwen2-VL base (ablation)Qwen2-VL-7B · VideoMMEaccuracy: 51.2 -> 52.8Improved
use AIM instead of LLaVA-Prumerge at 1.00 FLOPs on TextVQALLaVA-1.5-7B · TextVQAaccuracy: 53.8 -> 48.4Hurt

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