Newswise — Korean researchers have achieved a breakthrough in the safety of generative AI. They developed a vision language model optimized for safety and released it for the first time. In this model, AI can preemptively analyze both images and text, even detecting risks. The research team is paving the way for a safe AI era.
ETRI announced that it has unveiled a new type of vision language model called “Safe LLaVA,” which structurally enhances safety in generative AI models.
This technology departs from the existing data-centric fine-tuning method and directly incorporates a visual guard module that detects 20 safety categories into the model itself, providing safe responses along with the reasoning grounds behind them when harmful input…