Steering Multimodal Large Language Models Decoding for Context-Aware Safety