Jailbreak Defense on DAN
42.9Drop in ASRLatent Similarity- Best Layer
Evaluation Results
| Method | Links | |
|---|---|---|
| Latent Similarity- Best LayerAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | 42.9 | |
| Latent Similarity- Layer 12Alignment Strategy=Deliberative Alignment (GRPO)2026.04 | 21.9 | |
| RandomAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | -2.3 | |
| KLAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | -14.2 | |
| PerplexityAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | -38 | |
| Self CertaintyAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | -54.2 |