Improving robustness of language models from a geometry-aware perspective

Bin Zhu, Zhaoquan Gu, Le Wang +2

Recent studies have found that removing the norm-bounded projection and increasing search steps in adversarial training can significantly improve robustness. However, we observe th…