DECK: A Consistency x Confidence Taxonomy of LLM Hallucinations
This research proposes a new way to categorize LLM hallucinations by analyzing their detectability signature rather than just what is wrong with the output. It introduces the DECK taxonomy, which sorts errors into four behavioral regimes based on consistency and confidence. This allows researchers to predict which types of scoring methods—like black-box or white-box scorers—are best suited to find specific kinds of model failures.