KAIST researchers who conducted the study. From left: doctoral candidate Lee Won-jun, professor Kim Chang-ik, and doctoral candidates Ham Seok-il and Jang Jae-hyeok. [KAIST]
KAIST researchers who conducted the study. From left: doctoral candidate Lee Won-jun, professor Kim Chang-ik, and doctoral candidates Ham Seok-il and Jang Jae-hyeok. [KAIST]

South Korean researchers have developed a key technology that prevents large language models from losing their safety guardrails during custom training — and can make them safer than they were before fine-tuning.

KAIST announced Wednesday that a research team led by professor Kim Chang-ik of the School of Electrical Engineering has developed a training framework called "Buffer-and-Reinforce" that addresses safety degradation in fine-tuning — the process of retraining large language models (LLMs) such as ChatGPT on data specific to an individual or company.

The erosion of safety rules during custom training has long been considered one of the biggest challenges of the AI personalization era: while fine-tuning improves a model's task-specific capabilities, it tends to weaken the safety constraints built into the original model. The team drew on earlier findings showing that fine-tuning a "jailbroken" AI — one modified to respond to dangerous requests it would normally refuse — paradoxically causes far less safety degradation.

Rather than deploying a jailbroken model in a live service, the team devised an approach that uses a buffer module called "BufferLoRA" to apply the jailbroken state temporarily during custom training only, then removes it entirely before deployment.

The team was also the first in the world to explain why this phenomenon occurs. Their findings confirmed that a jailbroken AI is less susceptible to harmful information while still effectively learning the new task capabilities a user wants.

A schematic diagram of the Buffer-and-Reinforce framework. [KAIST]
A schematic diagram of the Buffer-and-Reinforce framework. [KAIST]

In experiments, the AI maintained high safety even under extreme conditions where all user training data consisted of harmful questions and answers. The rate at which the model generated dangerous responses after retraining was about 8 percent — lower than the roughly 18 percent recorded by the baseline model that had not been retrained. The framework achieved both strong personalized performance and top-level safety without requiring additional safety retraining or extra computational cost, suggesting it can be applied efficiently to real-world AI personalization services.

The technology can be applied across any field where LLMs are retrained for individual or corporate purposes — including AI assistants built on internal corporate documents, specialized AI handling sensitive information in medicine, law and finance, customized education AI, customer service chatbots and research assistance AI. In real-world service environments where user data is diverse and difficult to control, technology that preserves personalization performance while guaranteeing safety is essential.

"This research provides the core foundational technology that allows anyone to freely build a personalized AI using their own data while using it more safely," Kim said. "We plan to continue follow-up research to enable safe personalization not only in text-based large language models but also in multimodal AI and agentic AI systems."

The findings were accepted as a spotlight paper at the International Conference on Machine Learning 2026 — the world's top academic conference in AI — placing the work among the top approximately 2.2 percent of all submissions in July, drawing global attention.


nbgkoo@heraldcorp.com