Safety Evaluation on Safety Evaluation Suite WJ, FH, SR
4.6ASR (WJ)WaltzRL
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| WaltzRLMethod Identity=Method 72025.10 | 4.6 | 6.2 | 0.3 | 3.7 | |
| Single-model RL + SafeguardSafeguard=Llama Guard 42025.10 | 7.3 | 8.4 | 0.3 | 5.3 | |
| Oracle label-converted feedbackFeedback Source=Ground truth labels2025.10 | 10.6 | 10.4 | 0 | 7 | |
| Single-model RLDescription=Traditional single-model RL without feedback agent2025.10 | 13.2 | 22.8 | 0.6 | 12.2 | |
| Baseline response + SafeguardBackbone=Llama-3.1-8B-Instruct, Safeguard=Llama Guard 42025.10 | 16 | 11 | 0 | 9 | |
| Inference-time collaborationBackbone=Llama-3.1-8B-Instruct, Training=None2025.10 | 19.4 | 17 | 3.8 | 13.4 | |
| Baseline responseBackbone=Llama-3.1-8B-Instruct, Training=None2025.10 | 39 | 40.4 | 0 | 26.5 |