List your company

Jailbreak Robustness on Jailbreak Attacks

88.8Prefill Success RateUnaligned
6.01627.5084970.492May 25, 2026

Evaluation results

MethodLinks
UnalignedBackbone=Qwen3-8B2026.0588.868.8—
UnalignedBackbone=LLaMA-3-8B2026.0588.267.6—
UnalignedBackbone=Yi-1.5-9B2026.058267.1—
Sqrt-CompetenceBackbone=Qwen3-8B2026.055927.64
SequentialBackbone=Qwen3-8B2026.0552.826.10.2
Standard DPOBackbone=Qwen3-8B2026.0551.726.9—
Standard DPOBackbone=LLaMA-3-8B2026.0546.523.6—
Curri-DPOBackbone=LLaMA-3-8B2026.0536.817.1-8.1
SequentialBackbone=LLaMA-3-8B2026.053117.8-10.7
Curri-DPOBackbone=Qwen3-8B2026.0530.524.1-12
Sqrt-CompetenceBackbone=LLaMA-3-8B2026.0529.215.6-12.7
Standard DPOBackbone=Yi-1.5-9B2026.0526.511.8—
Staged-CompetenceBackbone=LLaMA-3-8B2026.0524.28.3-18.8
Curri-DPOBackbone=Yi-1.5-9B2026.0522.55-5.4
SequentialBackbone=Yi-1.5-9B2026.0519.26.8-6.2
Staged-CompetenceBackbone=Qwen3-8B2026.0516.28.3-27.1
Sqrt-CompetenceBackbone=Yi-1.5-9B2026.0515.25-9.1
Staged-CompetenceBackbone=Yi-1.5-9B2026.059.21.5-13.8