While larger language models exhibit impressive capabilities, they remain vulnerable to adversarial prompts. Empirical findings show that robustness against such attacks significantly improves with adversarial training, but not with model scaling alone.
