ANLI

ANLI (Adversarial Natural Language Inference) is a benchmark dataset for natural language understanding that uses an adversarial human-in-the-loop process to construct progressively harder examples. Unlike static datasets, ANLI is designed to expose whether models rely on genuine inference or superficial spurious patterns. The benchmark was developed to address the limitations of earlier NLI datasets, which models often mastered through pattern matching rather than true comprehension. ANLI is closely related to discussions of Benchmark overfitting and represents an early attempt at Adversarial evaluation of language models. Its iterative construction protocol also connects to the broader concept of a Dynamic benchmark.