< (From left) KAIST Professor Hyunwoo J. Kim, Postdoctoral Researcher Sanghyeok Lee, M.S candidate Taehoon Song, Korea University Ph.D candidate Jihwan Park >
How inconvenient would it be if you had to manually transfer every contact and photo from scratch every time you switched to a new smartphone? Current Artificial Intelligence (AI) models face a similar predicament. Whenever a superior new AI model—such as a new version of ChatGPT—emerges, it has to be retrained with massive amounts of data and at a high cost to acquire specialized knowledge in specific fields. A Korean research team has developed a "knowledge transplantation" technology between AI models that can resolve this inefficiency.
KAIST announced on January 27th that a research team led by Professor Hyunwoo J. Kim from the School of Computing, in collaboration with a research team from Korea University, has developed a new technology capable of effectively "transplanting" learned knowledge between different AI models.
Recently, Vision-Language Models (VLM), which understand both images and text simultaneously, have been evolving rapidly. These are easily understood as multimodal AIs, like ChatGPT, which can provide explanations when a user shows them a photo and asks a question. These models have the advantage of adapting relatively quickly to new fields using small amounts of data by pre-learning large-scale image and language data.
However, the need to repeat this "adaptation process" from scratch every time a new AI model is released has been pointed out as a major inefficiency. Existing adaptation techniques also faced limitations: they were difficult to use if the model structure changed even slightly, or they significantly increased memory and computational costs because multiple models had to be used simultaneously.
To solve these problems, the research team proposed "TransMiter," a transferable adaptation technique that allows learned knowledge to be reused regardless of the model's structure or size. The core of this technology is directly transferring the "adaptation experience" accumulated by one AI as it learns to another AI model.
< TransMiter: A transferable adaptation technique reusable regardless of model structure, size, etc. >
The researchers' technology does not overhaul the complex internal structure of the AI; instead, it adopts a method of passing on "know-how" learned by observing only the prediction results (output) to another AI. Even if the AI models have different architectures, if the know-how learned by one AI is organized based on the answers given to the same questions, another AI can utilize that knowledge immediately. Consequently, there is no need to undergo the complex and time-consuming retraining process, and there is almost no slowdown in speed.
This study is highly significant as it is the first to prove that AI adaptation knowledge—previously considered almost impossible to reuse if model structures or sizes differed—can be precisely transplanted regardless of the model type. This is expected to not only reduce repetitive learning costs but also be utilized as a so-called "knowledge patch" technology that updates Large Language Models (LLMs) in real-time according to specific needs.
Professor Hyunwoo J. Kim explained, "By extending this research, we can significantly reduce the cost of post-training that had to be performed repeatedly whenever a rapidly evolving hyper-scale language model appears. It will enable 'model patches' that easily add expertise in specific fields."
The study involved Taehoon Song (Master's student, KAIST School of Computing), Sanghyeok Lee (Postdoctoral researcher), and Jihwan Park (Doctoral student, Korea University) as co-authors, with Professor Hyunwoo J. Kim serving as the corresponding author. The research results were accepted for oral presentation (4.6% acceptance rate as of 2025) at AAAI 2026 (Association for the Advancement of Artificial Intelligence), the most prestigious international conference in the field of AI, and were presented on January 25th.
Paper Title: Transferable Model-agnostic Vision-Language Model Adaptation for Efficient Weak-to-Strong Generalization
Meanwhile, Professor Hyunwoo J. Kim's laboratory presented a total of three papers at the conference, including this paper and "TabFlash," a technology developed in collaboration with Google Cloud AI to enhance the understanding of tables within documents.
