KAIST Develops Training Method to Improve Multimodal AI Performance
- Yul So
- 2 days ago
- 2 min read
Oct 17, 2025
Yul So
Unrelated Data Mixing Encourages Models to Interpret Conflicting Signals More Effectively
Korea Advanced Institute of Science and Technology (KAIST) announced on October 14 that a research team led by Professor Hwang Eui-jong has developed a new training technique designed to enhance the performance of multimodal artificial intelligence (AI).
Multimodal AI refers to technology that can process and understand various types of data at the same time, including text, images, speech, and video. However, just as people tend to focus on images before reading accompanying text, multimodal AI often relies too heavily on one specific type of data. This imbalance can lead to lower prediction accuracy.
To solve this problem, Professor Hwang’s team in the Department of Electrical Engineering developed a data augmentation method that helps AI systems interpret all types of data more evenly. The technique involves combining unrelated or mismatched data to create samples with inconsistent meanings. By learning from these conflicting signals, the model becomes more balanced and less dependent on a single input type.
The researchers also applied a weighting strategy that gives more importance to confusing mismatched samples that have similar meanings, which further improves the system’s performance. “Improving AI performance is not only about changing the model’s structure or algorithm. It is more important to decide how and what data are used for training,” said Professor Hwang. “This research introduces a new approach to designing and refining data so that multimodal AI can use information in a more balanced and effective way.”
The study will be presented this December at the Conference on Neural Information Processing Systems (NeurIPS), one of the most prestigious AI conferences, held in San Diego, United States, and Mexico City, Mexico.
