Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models?

Re-assign community