Large language models (LLMs) often produce false or unsubstantiated outputs, hindering their adoption in various fields. This paper introduces entropy-based uncertainty estimators to detect a subset of these hallucinations—confabulations—which are arbitrary and incorrect generations. The method focuses on semantic meaning rather than word sequences, working across datasets and tasks without prior knowledge and generalizing well to new tasks. By identifying prompts likely to produce confabulations, it helps users assess LLM reliability and expands the potential uses of these otherwise unreliable models.

Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, Yarin Gal