Detecting and defining hallucinations in NLG tasks using a quantitative framework.
The author introduces MIND, an unsupervised training framework utilizing internal states of Large Language Models for real-time hallucination detection. The approach outperforms existing methods in detecting hallucinations.