ShortV introduces an efficient approach to reduce computational costs in Multimodal Large Language Models by identifying and freezing visual tokens in layers where they contribute minimally.

• Introduces Layer Contribution (LC) metric to quantify each layer's impact on visual and text tokens
• Achieves up to 60% computation reduction with negligible performance impact
• Demonstrates effectiveness across multiple MLLM architectures including LLaVA and MiniGPT-4
• Provides theoretical framework for understanding layer-wise redundancy in MLLMs

This research matters for Engineering by offering practical techniques to deploy more efficient multimodal AI systems with lower computational requirements, making advanced vision-language models more accessible for real-world applications.

ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers