LLM Jailbreaking on AdaSteer Evaluation Set (test)
1SRFSCAV
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SCAVTarget LLM=Llama-3.1-8B-AdaSteer, Train:Val=1:9, Strength=1e-42026.05 | 1 | 0 | 0 | |
| RepETarget LLM=Gemma-2-9b-AdaSteer, Train:Val=1:0, Strength=1.002026.05 | 1 | 0 | 1 | |
| AngularTarget LLM=Gemma-2-9b-AdaSteer, Train:Val=1:0, Strength=180°2026.05 | 2 | 1 | 0 | |
| RepETarget LLM=Qwen2.5-7B-AdaSteer, Train:Val=1:0, Strength=1.002026.05 | 4 | 5 | 6 | |
| RepETarget LLM=Llama-3.1-8B-AdaSteer, Train:Val=1:0, Strength=1.002026.05 | 11 | 21 | 21 | |
| RD-CTarget LLM=Gemma-2-9b-AdaSteer, Train:Val=4:1, Strength=1.02026.05 | 18 | 26 | 29 | |
| SCAVTarget LLM=Qwen2.5-7B-AdaSteer, Train:Val=1:9, Strength=1e-42026.05 | 23 | 51 | 37 | |
| SCAVTarget LLM=Gemma-2-9b-AdaSteer, Train:Val=1:9, Strength=1e-42026.05 | 30 | 47 | 53 | |
| RD-ATarget LLM=Gemma-2-9b-AdaSteer, Train:Val=4:12026.05 | 33 | 35 | 42 | |
| Adaptive Probe-based SteeringTarget LLM=Gemma-2-9b-AdaSteer2026.05 | 50 | 51 | 73 | |
| AngularTarget LLM=Llama-3.1-8B-AdaSteer, Train:Val=1:0, Strength=180°2026.05 | 62 | 73 | 82 | |
| Adaptive Probe-based SteeringTarget LLM=Qwen2.5-7B-AdaSteer2026.05 | 62 | 73 | 82 | |
| Adaptive Probe-based SteeringTarget LLM=Llama-3.1-8B-AdaSteer2026.05 | 65 | 79 | 88 | |
| AngularTarget LLM=Qwen2.5-7B-AdaSteer, Train:Val=1:0, Strength=180°2026.05 | 65 | 75 | 84 |