LLM Jailbreaking on HarmBench text (test N = 320)
93.75ASR-MPEO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PEOModel=Llama-3.2-3B-Instruct2026.04 | 93.75 | 50 | |
| SPTModel=Llama-3.2-3B-Instruct2026.04 | 92.19 | 16.88 | |
| PEOModel=Vicuna-7B v1.32026.04 | 90.94 | 49.69 | |
| nanoGCGModel=Vicuna-7B v1.32026.04 | 86.25 | 21.88 | |
| PEOModel=Llama-2-7B-Chat2026.04 | 82.81 | 45.62 | |
| SPTModel=Vicuna-7B v1.32026.04 | 80 | 33.44 | |
| PEOModel=Qwen3-4B-Instruct2026.04 | 70.31 | 37.5 | |
| BEASTModel=Vicuna-7B v1.32026.04 | 69.38 | 43.75 | |
| BEASTModel=Llama-3.2-3B-Instruct2026.04 | 68.44 | 16.88 | |
| SPTModel=Llama-2-7B-Chat2026.04 | 66.56 | 39.38 | |
| nanoGCGModel=Llama-3.2-3B-Instruct2026.04 | 64.69 | 19.38 | |
| nanoGCGModel=Llama-2-7B-Chat2026.04 | 58.44 | 26.25 | |
| BEASTModel=Qwen3-4B-Instruct2026.04 | 53.44 | 12.5 | |
| SPTModel=Qwen3-4B-Instruct2026.04 | 47.19 | 24.06 | |
| nanoGCGModel=Qwen3-4B-Instruct2026.04 | 40 | 6.88 | |
| BEASTModel=Llama-2-7B-Chat2026.04 | 39.06 | 8.12 |