Researchers used GPT-3.5 and GPT-4, trying three ways to guide the model. The most effective version, GPT-4, gave more accurate results but took slightly more time and cost.