A mechanistic approach for understanding and detecting factual errors of large language models. - View it on GitHub