July23: Korean researchers have developed a core source technology that allows multimodal artificial intelligence to reliably retain existing knowledge without losing it even when repeatedly learning new knowledge, attracting global attention
Electronics and Telecommunications Research Institute announced that the “Continual and Compositional Knowledge Editing Technology,” jointly developed by the research team led by Director Lim Soo Jong of the Language Intelligence Research Section with Pohang University of Science and Technology and Sungkyunkwan University, was presented at NeurIPS 2025, the world's most prestigious AI conference, and held in San Diego, U.S., late last year.
Recently, multimodal AI that can understand images and text simultaneously, such as ChatGPT, Gemini, and Claude, has been spreading rapidly. For example, it is AI that can look at a photo and describe it or answer when asked in text about what is in an image.
However, this type of AI had one major problem. When AI learns new information or revises existing information, a phenomenon called “catastrophic forgetting” occurs, in which it forgets previously learned knowledge as well. In other words, when AI learns something new, it develops a kind of “AI forgetfulness” in which it forgets what it previously knew.
In particular, when visual information and language information had to be revised at the same time, the two types of knowledge often became mixed together, causing the AI to misunderstand and frequently give wrong answers to compositional questions.
For example, if an AI is sequentially taught the visual information, “The dessert shown in the photo is Dubai chewy cookie ,” and the language information, “Dujjonku is popular in Korea,” and is then asked, “In which country is this dessert popular?”, existing AI models showed limitations in properly connecting the photo with the relevant knowledge.
In such cases, existing models frequently produced hallucinations, such as misidentifying the dessert in the photo and generating inaccurate answers such as, “The image shown in the photo is a chocolate truffle and is popular in Europe.”
To solve this problem, ETRI researchers developed a knowledge-editing AI technology capable of accurately answering even compositional questions.
Conventional methods mainly used an approach that changed knowledge by directly modifying the AI's internal parameters. This was a kind of “brain-surgery-like approach” that fundamentally altered the structure of the existing model, and it was limited in that even previously stored information could be affected in the process of modifying knowledge.
Instead, the researchers proposed a method of storing new information in external memory rather than inside the AI. This approach of adding an auxiliary memory has a structure that retrieves and uses information only when needed, making it possible to flexibly add new information while maintaining the stability of the existing model, thereby ensuring scalability as well.
MemEIC was designed with inspiration from the structure of the human brain. Just as the human brain is divided into the left and right hemispheres, which play different roles, the AI was designed to store knowledge separately as well.
Image-related visual information is stored independently in a “visual adapter,” while text-related language information is stored independently in a “language adapter.” Then, when the AI receives a compositional question that requires understanding both images and text, a “knowledge connector” links the two pieces of information according to context to produce an answer.
It was confirmed that AI applying the MemEIC technology developed by ETRI accurately combined visual and language information and correctly answered, “The dessert shown in the photo is Dubai chewy cookie, and it is popular in Korea.”
In this way, through a separated-storage and selective-combination structure that stores knowledge separately and connects it only when needed, an AI architecture capable of compositional reasoning and answering compositional questions was implemented by minimizing the problems of internal interference, in which different information becomes mixed, and the degradation of existing knowledge.
To verify the technology's performance, the researchers built a compositional knowledge editing benchmark consisting of 1,278 items and conducted experiments that sequentially edited hundreds of pieces of knowledge. As a result, MemEIC achieved an accuracy level of approximately 70% in answering compositional questions. Compared with the 36% to 52% range of existing technologies, this represents more than double the performance. In addition, the preservation characteristic of “locality,” in which response stability is maintained because answers to existing questions do not change even after new knowledge is added, was also confirmed.
This study is highly meaningful in that it goes a step beyond merely alleviating AI's forgetting phenomenon and simultaneously solves the two challenges of continuous knowledge editing and compositional reasoning. In particular, it is expected to have high practical applicability in intelligent service fields that require continual updates as new information is constantly added and changed, such as policy and legal information, product information, and industrial data.
Director Lim Soo Jong of ETRI's Language Intelligence Research Section said,
“This study is an achievement that has laid the technological foundation for multimodal AI to simultaneously reflect up-to-date information required in real service environments and secure reliability. We will further advance the technology so that it can reliably reflect diverse information from industrial sites in the future.”
Seong Jin, the first author of the paper and a researcher at ETRI's Language Intelligence Research Section, explained,
“Conventional methods had the problem of interference occurring while revising visual knowledge and language knowledge at the same time. MemEIC overcame these limitations through a structure that stores the two types of knowledge independently and connects them only when needed.”
