過剰拒否緩和

Latent-space adversarial training with post-aware calibration for defending large language models against jailbreak attacks