Contextual Privacy Leakage Evaluation on PrivaCI-Bench
0Leakage RateCI-parametric steering
Evaluation Results
| Method | Links | |
|---|---|---|
| CI-parametric steeringModel=Llama-3.1, alpha (α)=1.02026.03 | 0 | |
| CI-parametric steeringModel=Mistral, alpha (α)=1.02026.03 | 0.7 | |
| CI-parametric steeringModel=Mistral, alpha (α)=0.52026.03 | 2.4 | |
| Rep TuningModel=Mistral2026.03 | 4 | |
| Additive SteeringModel=Llama-3.1, alpha (α)=0.52026.03 | 10 | |
| No SteeringModel=Mistral2026.03 | 10.7 | |
| CI-parametric steeringModel=Qwen-2.5, alpha (α)=1.02026.03 | 12 | |
| CI-parametric steeringModel=Qwen-2.5, alpha (α)=0.52026.03 | 12.7 | |
| Additive SteeringModel=Qwen-2.5, alpha (α)=1.02026.03 | 13.3 | |
| LoRRAModel=Llama-3.12026.03 | 14 | |
| LoRRAModel=Qwen-2.52026.03 | 14.7 | |
| CI-parametric steeringModel=Llama-3.1, alpha (α)=0.52026.03 | 15.3 | |
| Additive SteeringModel=Qwen-2.5, alpha (α)=0.52026.03 | 15.4 | |
| LoRRAModel=Mistral2026.03 | 16 | |
| No SteeringModel=Qwen-2.52026.03 | 16.5 | |
| No SteeringModel=Llama-3.12026.03 | 18 | |
| LoRRAModel=Llama-22026.03 | 18.7 | |
| Rep TuningModel=Qwen-2.52026.03 | 18.7 | |
| CI-parametric steeringModel=Llama-2, alpha (α)=0.52026.03 | 18.7 | |
| CI-parametric steeringModel=Llama-2, alpha (α)=1.02026.03 | 19.3 | |
| No SteeringModel=Llama-22026.03 | 22.5 | |
| Additive SteeringModel=Mistral, alpha (α)=1.02026.03 | 30.7 | |
| Additive SteeringModel=Mistral, alpha (α)=0.52026.03 | 40 | |
| Additive SteeringModel=Llama-2, alpha (α)=0.52026.03 | 44 | |
| Rep TuningModel=Llama-3.12026.03 | 46.7 | |
| Additive SteeringModel=Llama-3.1, alpha (α)=1.02026.03 | 64.7 | |
| Rep TuningModel=Llama-22026.03 | 65.3 | |
| Additive SteeringModel=Llama-2, alpha (α)=1.02026.03 | 73.3 |