Lifelong Safety Adaptation on Safety Benchmarks (Final-day)
96.2Final-day Macro F1 ScoreLiSA
Evaluation Results
| Method | Links | |
|---|---|---|
| LiSAlabel-flip ratio (ρ)=0%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 96.2 | |
| ReasoningBanklabel-flip ratio (ρ)=0%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 93.6 | |
| LiSAlabel-flip ratio (ρ)=10%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 93.3 | |
| LiSAlabel-flip ratio (ρ)=20%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 91.7 | |
| Synapselabel-flip ratio (ρ)=0%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 89.1 | |
| ReasoningBanklabel-flip ratio (ρ)=10%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 88.2 | |
| Synapselabel-flip ratio (ρ)=10%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 87.8 | |
| Synapselabel-flip ratio (ρ)=20%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 86.3 | |
| AGraillabel-flip ratio (ρ)=0%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 85.7 | |
| ReasoningBanklabel-flip ratio (ρ)=20%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 85.7 | |
| AGraillabel-flip ratio (ρ)=10%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 84.6 | |
| AGraillabel-flip ratio (ρ)=20%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 83.2 | |
| Pure Predictionlabel-flip ratio (ρ)=0%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 81.8 | |
| Pure Predictionlabel-flip ratio (ρ)=10%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 81.8 | |
| Pure Predictionlabel-flip ratio (ρ)=20%, Base Model=Gemini-3.1-flash-lite, Number of seeds=52026.05 | 81.8 |