They should prompt the model to ignore irrelevant information and test if the model performs better and is good at ignoring those statements?