Less is More: A Simple yet Effective Token Reduction Method for Efficient Multi-modal LLMs