This research addresses the critical challenge of safety alignment in Multimodal Large Language Models through a novel preference dataset and optimization technique.

• Created MMSafe-PO, a specialized safety preference dataset for multimodal models
• Developed Blind Preference Optimization (BPO) to enhance safety without compromising capabilities
• Achieved significant reduction in unsafe responses across multiple benchmarks
• Demonstrated effective transfer of safety capabilities to new harmful scenarios

This work provides a scalable approach to mitigating safety risks in multimodal AI systems while maintaining their performance on core tasks, addressing a key security concern for enterprise AI deployment.

Original Paper: Towards Harmless Multimodal Assistants with Blind Preference Optimization