Contrast sets can effectively probe the depth of language understanding in large language models, revealing their reliance on superficial patterns rather than true comprehension.
