Tag
The study finds that neural language models degrade similarly under word-level noise but differently under character-level noise, with tokenization identified as the key hidden variable. It provides a method to predict model robustness without noisy evaluation and suggests noise-augmented training for install robustness.