The Electronics and Telecommunications Research Institute (ETRI) has developed a technology that enables multimodal artificial intelligence (AI) to avoid losing existing information even when repeatedly learning new knowledge. The institute explains that this reduces the "fatal forgetting" problem, which frequently occurs in AI that handles both images and text, while also improving the accuracy of answers to complex questions.

ETRI announced that the 'Continuous and Complex Knowledge Editing Technology (MemEIC),' jointly developed by the research team led by Director Lim Su-jong of the Language Intelligence Research Lab with Pohang University of Science and Technology and Sungkyunkwan University, was accepted and presented at the artificial intelligence conference 'NeurIPS 2025' held in San Diego, USA, at the end of last year.

Recently, generative AI has been rapidly spreading through multimodal methods that understand both photos and sentences, but limitations have been pointed out regarding the loss of previously learned knowledge during the process of incorporating new information. In particular, when visual and linguistic information were modified simultaneously, different knowledge became mixed up, resulting in answers that did not fit the context of the question.

Instead of directly modifying core internal AI parameters as in conventional methods, the research team adopted a approach that stores new information in external memory. Image-related information is stored separately in a 'visual adapter' and text-related information in a 'language adapter,' and when a complex question is input, a 'knowledge connector' combines the two sets of information to generate an answer. The design allows for the addition of necessary knowledge while maintaining stability without significantly altering the existing model itself.

For example, when a related question was asked after sequentially training a model with specific dessert images and information on the food's popular regions, existing models sometimes answered with a different food or region because they failed to properly combine image and text information. In contrast, ETRI explained that the model applying MemEIC connected visual and linguistic information to provide an answer that matched the intent of the question.

The research team conducted an experiment to sequentially edit hundreds of pieces of knowledge based on a complex knowledge editing benchmark consisting of 1,278 items. As a result, MemEIC achieved a complex question accuracy of approximately 70%. This represented a significant improvement compared to the 36–52% of existing technologies, and stability was also confirmed, with responses to existing questions remaining stable even after new knowledge was added.

This research is significant in that it addresses the challenge of AI maintaining existing knowledge while reflecting constantly changing information. ETRI believes there is potential for application in services that require continuous updates, such as policies, laws, product information, and industrial data. The research was conducted with support from the Next-Generation Generative AI Technology Development Project of the Ministry of Science and ICT and the Institute of Information and Communication Technology Planning and Evaluation (IITP).