INCHEON — Researchers have identified how artificial intelligence models that forecast time-series data can learn temporal patterns even without a core component. Inha University said on the 12th that a paper by undergraduate student Kim Min-kyu and a team led by Professor Kim Do-kook of the Department of Artificial Intelligence Engineering has been accepted by the world's most prestigious conference in the field.
The research focused on models that forecast time-series data — information accumulated over time, such as stock prices, electricity demand and weather observations. The team concentrated on analyzing how the transformer model learns temporal patterns. To do so, the researchers designed 15 diagnostic tasks of their own to verify what information each stage inside the model learns.
The analysis found that the role of learning temporal patterns is not fixed to any single component. The function of each component changed depending on how tokenization — dividing data into smaller units — was combined with the model architecture. The team demonstrated that even when attention, the core component, is removed entirely, the ability to learn temporal patterns is preserved under certain conditions.
By analyzing data segmentation methods together with model architecture, the team unified previously conflicting findings in the field into a single explanatory framework. The research is expected to provide an important benchmark for designing highly efficient AI models in industries where time-series forecasting is essential, including finance, energy and manufacturing. It is also expected to help improve the reliability of AI models and sharply reduce computing costs.
Kim Min-kyu is the first author and Professor Kim Do-kook the corresponding author of the paper, which will be presented at NeurIPS 2026 in December. NeurIPS ranks second in computer science and first in artificial intelligence on Google Scholar's h5-index, with a score of 371, making it the most prestigious conference in the field. Having an undergraduate lead an entire research project through to a paper at the world's top conference is described as a highly unusual achievement.
"It is meaningful that the fruit of an undergraduate's active participation in every stage of the research has been recognized on the world's biggest stage," said Kim Do-kook, a professor in Inha University's Department of Artificial Intelligence Engineering. The professor added, "We will continue to build an environment where students can demonstrate world-class research capabilities from their undergraduate years." The professor also said the team would continue research on next-generation AI models to broaden their applicability in industrial settings.