Researchers developed a harness to learn interpretable models from data using LLMs, outperforming state-of-the-art rule-learning methods on tabular and text data.