This research introduces a novel approach to optimize Large Vision Language Models (LVLMs) by significantly reducing computational costs while maintaining performance for reasoning segmentation tasks.

• Addresses the critical challenge of computational overhead in vision-language systems
• Employs a three-step approach: clustering, scattering, and pruning of image tokens
• Demonstrates that strategic token reduction can maintain model performance while improving efficiency
• Enables more practical deployment of powerful LVLMs in resource-constrained environments

This engineering breakthrough makes advanced vision-language capabilities more accessible for real-world applications by balancing computational requirements with performance needs.

LVLM_CSP: Accelerating Large Vision Language Models via Clustering, Scattering, and Pruning for Reasoning Segmentation