What if, after purchasing the latest mobile phone, you couldn't transfer your personal data from your previous device? Users would undoubtedly find this inconvenient and troublesome. The same goes for artificial intelligence (AI) models. Even if a new AI model with improved performance is released, if it must go through the process of retraining to acquire expertise in a specific field, the burden of massive data and costs makes the situation far from welcome. A technology enabling 'knowledge transplantation' between AI models, designed to solve such inefficiencies, has now been developed in Korea.


KAIST announced on January 27 that Professor Hyunwoo Kim's research team from the School of Computing, in collaboration with a research team from Korea University, has developed a technology that enables effective mutual 'transplantation' of knowledge learned by different AI models.


Image generated by artificial intelligence (AI). Provided by KAIST

Image generated by artificial intelligence (AI). Provided by KAIST

View original image

Recently, the AI field has seen rapid advances in 'vision-language models (VLMs)' that can understand both images and text. VLMs operate on the same principle as multimodal AI such as 'ChatGPT,' which provides explanations when a user submits a photo and asks a question. These models, having been pre-trained on large-scale image and language datasets, can adapt relatively quickly to new domains even with small amounts of data.


However, it has been pointed out that having to repeat this 'adaptation process' from scratch each time a new AI model is introduced is inefficient. Existing adaptation techniques are also difficult to apply as-is if the model architectures differ even slightly, or they require using multiple models simultaneously, leading to increased memory and computational costs.


To address these issues, the joint research team proposed 'TransMiter,' a transferable adaptation technique that allows the reuse (a form of transplantation) of learned knowledge regardless of model architecture or size. The core of this technology is to directly transfer the 'adaptation experience' accumulated by one AI model to another.


This technique allows the know-how learned from observing prediction results (outputs) to be shared with another AI without modifying the complex internal structure of the model. If the answers provided by different AI models to the same question are organized based on a set reference point, the know-how acquired by one AI can be immediately utilized by another. This eliminates the need for a complex and time-consuming retraining process, while also maintaining almost the same speed.


(From left) Professor Hyunwoo Kim, Postdoctoral Researcher Sanghyuk Lee, Master's Course Student Taehoon Song from KAIST, and Ph.D. Course Student Jihwan Park from Korea University. Courtesy of KAIST

(From left) Professor Hyunwoo Kim, Postdoctoral Researcher Sanghyuk Lee, Master's Course Student Taehoon Song from KAIST, and Ph.D. Course Student Jihwan Park from Korea University. Courtesy of KAIST

View original image

Most notably, this research is significant in that it is the first to demonstrate that adaptation knowledge can be precisely transplanted regardless of model type. It shows the possibility of knowledge transfer between different AI models, which was previously considered impossible.


The joint research team believes this will not only reduce repetitive training costs but also enable the so-called 'knowledge patch' technology, which allows for real-time updates of large language models tailored to specific domains.


Professor Kim stated, "If this research is expanded, it could drastically reduce the post-training costs that had to be repeatedly incurred whenever rapidly advancing large language models appear. In effect, it will make it possible to easily add domain expertise through a 'model patch.'"



This research included KAIST master's student Taehoon Song, postdoctoral researcher Sanghyuk Lee, and Korea University Ph.D. student Jihwan Park as co-authors, with Professor Kim as the corresponding author. The results (paper) were recently accepted and presented as an oral presentation at 'AAAI 2026 (Association for the Advancement of Artificial Intelligence),' an international conference in the AI field.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing