MR-Pruner: Multi-Resolution Token Pruning Method

📌 Introduction

We propose MR-Pruner, a training-free, graph-based token pruning framework explicitly designed for multi-resolution MLLMs. Unlike prior single-resolution pruning methods, MR-Pruner accounts for the informativeness distribution across resolutions and their mutual complementarity.

Key highlights:

• 🔍 Resolution-aware pruning: Separately handles high- and low-resolution tokens with adaptive ratios.
• 🔗 Graph-based scoring: Builds intra- and cross-resolution token graphs to propagate informativeness.
• 🚀 Training-free: Plug-and-play, no model retraining needed.
• 🎯 Robust: Preserves complementary tokens across resolutions, enabling resilience in extreme pruning.

📰 News

• [2025/11/11] Our paper MR-Pruner has been accepted to WACV 2026!🎉

---

⚙️ Installation

MR-Pruner is built on Lmms-eval, which is evaluation toolkit for MLLMs.

git clone https://github.com/gooriiie/MR-Pruner.git
cd MR-Pruner

# create environment
conda create -n mrpruner python=3.10
conda activate mrpruner
pip install -r requirements.txt

# create environment with yaml
conda env create -f environment.yml

---

🚀 Run

Run evaluation with default settings.

# TextVQA
bash ./scripts/eval_lmms_eval_textvqa.sh  

# ScienceQA-IMG
bash ./scripts/eval_lmms_eval_sqa.sh

# POPE
bash ./scripts/eval_lmms_eval_pope.sh