Relationship-Aware Safety Unlearning for Multimodal LLMs

1. 1Identify relational concepts that can lead to safety failures
2. 2Develop a framework for relationship-aware safety unlearning
3. 3Implement the framework in multimodal LLMs to mitigate safety failures
4. 4Evaluate the effectiveness of the framework in reducing safety failures

AI engineers and researchers working on multimodal LLMs can benefit from this framework to improve model safety, while data scientists and ML researchers can apply these concepts to other areas of AI development

💡 Relational concepts can cause safety failures in multimodal LLMs, and a targeted unlearning approach can help mitigate these issues

💡 Mitigating safety failures in multimodal LLMs with relationship-aware safety unlearning