Adversarial NLI (ANLI) is built with a human-and-model-in-the-loop process where annotators craft premise–hypothesis pairs that specifically fool strong NLI systems. Compared to earlier NLI datasets, ANLI contains harder, more diverse phenomena and fewer superficial patterns, making it a robust stress test of entailment, contradiction, and neutrality judgments.

Because items were collected against competitive models, success on ANLI often requires nuanced world knowledge, careful reading, and multi-step reasoning rather than pattern matching. Performance is typically reported across progressively harder rounds, reflecting increasing adversarial pressure and out-of-distribution generalization.