List your company

Jailbreak Robustness on JBB

2ASRBCT
0.411.22232.8Jun 4, 2026

Evaluation results

MethodLinks
BCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.062
MLPCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.062
AttCTModel=Qwen3-4B-Instruct-2507, Stage=Pre2026.062
BCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.063
MLPCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.063
AttCTModel=Gemma-3-4B-IT, Stage=Post2026.0611
AttCTModel=Gemma-3-27B-IT, Stage=Post2026.0614
AttCTTarget (f_theta)=attention weights2026.0616
AttCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.0617
AttCTModel=Qwen3-8B, Stage=Post2026.0618
BCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.0619
MLPCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.0619
AttCTModel=Llama-3.1-8B-Instruct, Stage=Pre2026.0619
BCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.0620
BCTModel=Qwen3-8B, Stage=Pre2026.0620
MLPCTModel=Qwen3-8B, Stage=Pre2026.0620
AttCTModel=Qwen3-8B, Stage=Pre2026.0620
MLPCTModel=Llama-3.1-8B-Instruct, Stage=Post2026.0621
AttCTModel=Qwen3-4B-Instruct-2507, Stage=Post2026.0622
BCTModel=Qwen3-8B, Stage=Post2026.0623
Base model (no training)Target (f_theta)=—2026.0623
BCTTarget (f_theta)=output logits2026.0623
ACTTarget (f_theta)=residual stream2026.0623
MLPCTModel=Qwen3-8B, Stage=Post2026.0625
MLPCTTarget (f_theta)=MLP hidden state2026.0625
MLPCTModel=Gemma-3-27B-IT, Stage=Post2026.0632
BCTModel=Gemma-3-27B-IT, Stage=Post2026.0633
BCTModel=Gemma-3-4B-IT, Stage=Post2026.0634
BCTModel=Gemma-3-27B-IT, Stage=Pre2026.0634
MLPCTModel=Gemma-3-27B-IT, Stage=Pre2026.0634
AttCTModel=Gemma-3-27B-IT, Stage=Pre2026.0634
MLPCTModel=Gemma-3-4B-IT, Stage=Post2026.0637
BCTModel=Gemma-3-4B-IT, Stage=Pre2026.0642
MLPCTModel=Gemma-3-4B-IT, Stage=Pre2026.0642
AttCTModel=Gemma-3-4B-IT, Stage=Pre2026.0642