A research team led by Professor Namhoon Lee at the POSTECH Graduate School of Artificial Intelligence has developed a methodology that separates and measures the uncertainty contained in large language model (LLM) predictions by cause. This study has been accepted as an oral presentation at the international conference ‘ACL 2026,’ which will be held in San Diego, USA, from the 2nd to the 7th. The core of the team’s approach is the concept of a "Self-Function Vector," which effectively separated and measured uncertainty arising from data ambiguity and uncertainty caused by the model’s lack of knowledge. The researchers combined research on "Mechanistic Interpretability" with Bayesian inference theory to propose a new analytical framework, and also developed an evaluation framework to objectively validate their technology.