A New Method to Detect “Confabulations” Hallucinated by Large Language Models

By calculating semantic entropy with a second LLM, we can better flag answers as unreliable due to lack of knowledge