Safety Control on DeepSeek-R1-Distill-Qwen-1.5B
89.8P_safeguarded (Safety-Quality Score)SafeCtrl-RL
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SafeCtrl-RLModel Type=Our Proposed2026.05 | 89.8 | 0.608 | |
| progressiveModel Type=Hand-crafted2026.05 | 87.9 | 0.589 | |
| ai_enhancedModel Type=Hand-crafted2026.05 | 84.3 | 0.552 | |
| best_worst_recentModel Type=Hand-crafted2026.05 | 83.2 | 0.542 | |
| ai_onlyModel Type=Hand-crafted2026.05 | 82.8 | 0.538 | |
| raw_historyModel Type=Hand-crafted2026.05 | 78.1 | 0.49 | |
| minimalModel Type=Hand-crafted2026.05 | 77.7 | 0.486 | |
| hybridModel Type=Hand-crafted2026.05 | 75.3 | 0.463 | |
| trajectory_learningModel Type=Hand-crafted2026.05 | 73.8 | 0.448 | |
| performance_tieredModel Type=Hand-crafted2026.05 | 64.5 | 0.355 | |
| smart_adaptiveModel Type=Hand-crafted2026.05 | 61.5 | 0.325 | |
| contrast_learningModel Type=Hand-crafted2026.05 | 54.9 | 0.259 | |
| OPROModel Type=Prompt Optimisation2026.05 | 16.7 | -0.123 | |
| DynamicRetrievalModel Type=Prompt Optimisation2026.05 | 15.6 | -0.134 | |
| GRIPSModel Type=Prompt Optimisation2026.05 | 10 | -0.19 | |
| TextGradientModel Type=Prompt Optimisation2026.05 | 6.9 | -0.221 | |
| EvolutionaryModel Type=Prompt Optimisation2026.05 | 6.5 | -0.225 |