New Delhi: Is it possible to fix the forgetfulness of artificial intelligence AI? Researchers from South Korea believe they have done just that by creating a system that keeps LLM models from losing what they know.
A significant shortcoming in AI learning is that acquiring new knowledge often means old information disappears. South Korean scientists now claim that they have found a solution.
A team comprising researchers from the Electronics and Telecommunications Research Institute (ETRI), Pohang University of Science and Technology (POSTECH), and Sungkyunkwan University has invented a new technique called MemEIC. This knowledge-editing method allows multimodal AI to acquire new knowledge while still retaining old knowledge. The research paper was presented at NeurIPS 2025, one of the leading AI conferences.
The problem known as ‘catastrophic forgetting‘ is more concerning as applications such as ChatGPT, Gemini, and Claude are expected to be updated and understand both text and images.
LLM models also struggle to parse visual and language data together. For example, if an AI model is taught that a picture contains the Dujjonku, a chewy cookie sold in Dubai, then separately taught that the Dujjonku is popular in South Korea, it may be unable to link the two. It may offer incorrect responses for queries related to Dujjonku.
Currently, most methods update an LLM model by changing the AI’s internal parameters. Although useful, the process could inadvertently remove earlier information, resulting in incorrect answers or ‘hallucinations’ from the models.
MemEIC takes a different path. Instead of altering a model’s memory, it stores the new knowledge in an external memory, which can be referenced when needed.
The image and text data are also kept separate. Visual knowledge is processed through a ‘visual adapter’, language through a ‘language adapter’, with a ‘knowledge connector’ linking the two when a query requires both.
This design is said to have been inspired by the human brain, where different regions process various forms of information before merging it.
Also read: IIT Madras professor who shaped Startup India says country now needs 20 million entrepreneurs
How effective is MemEIC?
To demonstrate the effectiveness of MemEIC, researchers created the Compositional Continual Knowledge Editing Benchmark (CCKEB), which involved 1,278 evaluation instances and hundreds of sequential knowledge updates.
The method was found to perform 70 per cent better on compositional reasoning tasks than current systems, which could only achieve 36 per cent or 52 per cent. It also preserved what researchers call ‘locality’, meaning that the new information did not interfere with responses on previously acquired knowledge.
“This study laid the technological foundation for multimodal AI to reflect up-to-date information required in real service environments while maintaining reliability,” Lim Soo Jong of ETRI told Newswise.
First author Seong Jin suggested that previous systems struggled due to interference between visual and language data, and that MemEIC effectively stores them separately and uses them when necessary. If MemEIC remains effective outside research labs, it could allow AI to keep learning without forgetting.
(Edited by Prasanna Bachchhav)