Jailbreak Robustness on JBB
2ASRBCT
Evaluation Results
| Method | Links | |
|---|---|---|
| BCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.06 | 2 | |
| MLPCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.06 | 2 | |
| AttCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.06 | 2 | |
| BCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.06 | 3 | |
| MLPCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.06 | 3 | |
| AttCTModel=Gemma-3-4B-IT, Stage=Post2026.06 | 11 | |
| AttCTModel=Gemma-3-27B-IT, Stage=Post2026.06 | 14 | |
| AttCTTarget (f_theta)=attention weights2026.06 | 16 | |
| AttCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.06 | 17 | |
| AttCTModel=Qwen3-8B, Stage=Post2026.06 | 18 | |
| BCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.06 | 19 | |
| MLPCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.06 | 19 | |
| AttCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.06 | 19 | |
| BCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.06 | 20 | |
| BCTModel=Qwen3-8B, Stage=Pre2026.06 | 20 | |
| MLPCTModel=Qwen3-8B, Stage=Pre2026.06 | 20 | |
| AttCTModel=Qwen3-8B, Stage=Pre2026.06 | 20 | |
| MLPCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.06 | 21 | |
| AttCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.06 | 22 | |
| BCTModel=Qwen3-8B, Stage=Post2026.06 | 23 | |
| Base model (no training)Target (f_theta)=—2026.06 | 23 | |
| BCTTarget (f_theta)=output logits2026.06 | 23 | |
| ACTTarget (f_theta)=residual stream2026.06 | 23 | |
| MLPCTModel=Qwen3-8B, Stage=Post2026.06 | 25 | |
| MLPCTTarget (f_theta)=MLP hidden state2026.06 | 25 | |
| MLPCTModel=Gemma-3-27B-IT, Stage=Post2026.06 | 32 | |
| BCTModel=Gemma-3-27B-IT, Stage=Post2026.06 | 33 | |
| BCTModel=Gemma-3-4B-IT, Stage=Post2026.06 | 34 | |
| BCTModel=Gemma-3-27B-IT, Stage=Pre2026.06 | 34 | |
| MLPCTModel=Gemma-3-27B-IT, Stage=Pre2026.06 | 34 | |
| AttCTModel=Gemma-3-27B-IT, Stage=Pre2026.06 | 34 | |
| MLPCTModel=Gemma-3-4B-IT, Stage=Post2026.06 | 37 | |
| BCTModel=Gemma-3-4B-IT, Stage=Pre2026.06 | 42 | |
| MLPCTModel=Gemma-3-4B-IT, Stage=Pre2026.06 | 42 | |
| AttCTModel=Gemma-3-4B-IT, Stage=Pre2026.06 | 42 |