A new model called COLAGUARD enhances the safety of large language models by employing latent reasoning, significantly improving efficiency and reducing resource usage compared to existing methods. This advancement allows for safer and faster deployment of LLMs in real-world applications.