Safety Evaluation on HEx-PHI (out-of-domain)
0ASR (OM)Trajectory Alignment
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Trajectory AlignmentModel=Llama-3.1-8B-Instruct, Iteration=12026.06 | 0 | 0 | |
| Trajectory AlignmentModel=Qwen2.5-7B-Instruct, Iteration=22026.06 | 0 | 0 | |
| Trajectory AlignmentModel=Mistral-7B-Instruct2026.06 | 0.37 | 1.86 | |
| Trajectory AlignmentModel=Llama-3.1-8B-Instruct, Iteration=22026.06 | 0.37 | 0 | |
| Circuit BreakersModel=Mistral-7B-Instruct2026.06 | 0.74 | 1.86 | |
| Trajectory AlignmentModel=Llama-3.1-8B-Instruct2026.06 | 1.12 | 0 | |
| No AttackModel=Llama-3.1-8B-Instruct2026.06 | 1.49 | 0 | |
| Trajectory AlignmentModel=Mistral-7B-Instruct, Iteration=22026.06 | 1.49 | 0 | |
| BaseModel=Llama-3.1-8B-Instruct2026.06 | 1.49 | 0 | |
| No AttackModel=Qwen2.5-7B-Instruct2026.06 | 2.6 | 2.97 | |
| BaseModel=Qwen2.5-7B-Instruct2026.06 | 2.6 | 2.97 | |
| Trajectory AlignmentModel=Mistral-7B-Instruct, Iteration=12026.06 | 3.35 | 1.86 | |
| LATModel=Llama-3.1-8B-Instruct2026.06 | 3.72 | 0.37 | |
| SafeProbingModel=Qwen2.5-7B-Instruct2026.06 | 4.46 | 8.18 | |
| Trajectory AlignmentModel=Qwen2.5-7B-Instruct, Iteration=12026.06 | 4.83 | 1.49 | |
| Trajectory AlignmentModel=Qwen2.5-7B-Instruct2026.06 | 9.67 | 4.83 | |
| SafeProbingModel=Mistral-7B-Instruct2026.06 | 10.41 | 18.96 | |
| No AttackModel=Mistral-7B-Instruct2026.06 | 23.42 | 19.7 | |
| BaseModel=Mistral-7B-Instruct2026.06 | 23.42 | 19.7 | |
| Circuit BreakersModel=Llama-3.1-8B-Instruct2026.06 | 28.25 | 37.55 | |
| SafeProbingModel=Llama-3.1-8B-Instruct2026.06 | 35.69 | 47.21 | |
| Egida-DPOModel=Llama-3.1-8B-Instruct2026.06 | 40.15 | 42.38 | |
| Injection (no defense)Model=Qwen2.5-7B-Instruct2026.06 | 56.13 | 58.74 | |
| Egida-DPOModel=Qwen2.5-7B-Instruct2026.06 | 59.85 | 55.02 | |
| Injection (no defense)Model=Llama-3.1-8B-Instruct2026.06 | 61.71 | 51.67 | |
| Injection (no defense)Model=Mistral-7B-Instruct2026.06 | 65.06 | 49.07 |