TokenCarve introduces a novel framework for compressing visual tokens in multimodal LLMs, significantly reducing computational overhead while preserving model performance.

• Achieves 80% reduction in visual tokens with minimal impact on performance
• Implements information-preserving compression without expensive model retraining
• Outperforms existing compression methods in both efficiency and accuracy
• Works across diverse multimodal LLM architectures

By addressing the computational bottleneck of visual processing in MLLMs, TokenCarve enables faster, more efficient multimodal AI systems for practical engineering applications, paving the way for more responsive and cost-effective AI deployment.

TokenCarve: Information-Preserving Visual Token Compression in Multimodal Large Language Models