Safety Control on Macro Metrics Aggregate across LLMs
81.8Macro-P Safeguarded Safety-Quality ScoreSafeCtrl-RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SafeCtrl-RLModel Type=Our Proposed2026.05 | 81.8 | 0.513 | |
| raw_historyModel Type=Hand-crafted2026.05 | 80.7 | 0.502 | |
| progressiveModel Type=Hand-crafted2026.05 | 79.3 | 0.488 | |
| minimalModel Type=Hand-crafted2026.05 | 79.1 | 0.486 | |
| ai_onlyModel Type=Hand-crafted2026.05 | 78.6 | 0.481 | |
| hybridModel Type=Hand-crafted2026.05 | 78.1 | 0.476 | |
| ai_enhancedModel Type=Hand-crafted2026.05 | 77.6 | 0.471 | |
| best_worst_recentModel Type=Hand-crafted2026.05 | 76.5 | 0.46 | |
| performance_tieredModel Type=Hand-crafted2026.05 | 74.2 | 0.437 | |
| trajectory_learningModel Type=Hand-crafted2026.05 | 70 | 0.395 | |
| contrast_learningModel Type=Hand-crafted2026.05 | 67.8 | 0.373 | |
| smart_adaptiveModel Type=Hand-crafted2026.05 | 64.5 | 0.34 | |
| OPROModel Type=Prompt Optimisation2026.05 | 34.2 | 0.037 | |
| DynamicRetrievalModel Type=Prompt Optimisation2026.05 | 29.3 | -0.012 | |
| GRIPSModel Type=Prompt Optimisation2026.05 | 26.4 | -0.041 | |
| TextGradientModel Type=Prompt Optimisation2026.05 | 21.2 | -0.093 | |
| EvolutionaryModel Type=Prompt Optimisation2026.05 | 16.7 | -0.138 |