Jailbreaking on AdvBench (LG4 ASR & Judge Score)
43.6LG4 ASRRefusal-Direction
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Refusal-DirectionTarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 43.6 | 0.67 | |
| CRaFTTarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 42.1 | 2.3 | |
| Refusal-SAETarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 41.9 | 0.99 | |
| PAPTarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 7.9 | 1.11 | |
| GCGTarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 7.1 | 0.41 | |
| Steering-SAETarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 2.1 | 0.25 | |
| No attackTarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 0.4 | 0.09 | |
| AutoDANTarget Model=google/gemma-3-1b-it, Decoding=greedy, max_new_tokens=5122026.04 | 0.2 | 0.07 |