This paper proposes an unsupervised learning-based hallucination detection method that identifies false information generated by large-scale language models (LLMs) without labeled data. Existing unsupervised learning methods rely on surrogate metrics unrelated to factual accuracy, leading to poor generalization performance across datasets and contexts. IRIS addresses this issue by leveraging internal representations related to factual accuracy. The LLM is guided to verify the truth of a given proposition, and the resulting contextualized embeddings are used as features for learning. The uncertainty of each response is used as a soft pseudo-label for truth. Experimental results demonstrate that IRIS outperforms existing unsupervised learning-based methods. IRIS is a fully unsupervised learning method, computationally inexpensive, and effective even with a small amount of training data, making it suitable for real-time detection.
