KAIST Researchers Create Safe AI Training Method
A team from KAIST, led by Professor Changick Kim, has developed an innovative approach to make personalized AI safer while improving its performance. This new training method, called “Buffer-and-Reinforce,” focuses on keeping AI safe as it’s customized using individual or corporate data.
Personalized AI technology is becoming increasingly popular, but it often presents challenges, notably in maintaining safety standards when fine-tuning models like ChatGPT. Fine-tuning can enhance the model’s performance for specific tasks but might weaken its existing safety features.
The team’s breakthrough comes from understanding that fine-tuning AI when it’s in a temporarily “jailbroken” state—where it might be susceptible to harmful requests—doesn’t significantly lower its safety standards. Their approach involves a buffering module named “BufferLoRA,” which temporarily shields the model during the fine-tuning process to prevent harmful data from interfering with the core functions.
Once fine-tuning is complete, this protective layer is removed, allowing the model to operate normally while maintaining strong safety protocols. To reinforce safety further, the researchers then apply a module called “ReinforceLoRA,” which uses a mathematical technique called QR decomposition. This helps to selectively keep useful information from user data while enhancing safety features.
In simpler terms, the researchers placed a temporary shield over the AI model during adjustments to stop harmful data from affecting it. After the adjustments, the shield was removed, and safety protocols were strengthened. As a result, the AI not only performed well but also became safer than before.
Tests showed that even when trained with harmful questions, the AI maintained a safety rate of only about 8% in generating harmful responses—much lower than the 18% rate of the untrained model. This new framework allows for effective customization without needing extensive safety data or significantly increased computational costs.
Professor Kim highlighted the significance of this research, stating it lays the groundwork for anyone to create safer, customized AI solutions. The research was spearheaded by Ph.D. student Seokil Ham and has garnered international attention, being selected for a prestigious presentation at the International Conference on Machine Learning (ICML) 2026.
This work was supported by the Institute of Information & Communication Technology Planning & Evaluation, under a grant from the Ministry of Science and ICT.
