Safety Boundary Over-Refusal on ORBench
7.2ORBench Over-Refusal RateMistral-7B-v2
Evaluation Results
| Method | Links | |
|---|---|---|
| Mistral-7B-v2Backbone=Mistral-7B-v22026.05 | 7.2 | |
| QWEN-2.5-7BBackbone=QWEN-2.5-7B2026.05 | 9.6 | |
| LLAMA3-8BBackbone=LLAMA3-8B2026.05 | 14.3 | |
| Ellipsoid ControlBackbone=Mistral-7B-v22026.05 | 14.5 | |
| Ellipsoid ControlBackbone=QWEN-2.5-7B2026.05 | 17.4 | |
| Ellipsoid ControlBackbone=LLAMA3-8B2026.05 | 19.3 | |
| LoRA-SteerBackbone=QWEN-2.5-7B2026.05 | 32.3 | |
| LoRA-SteerBackbone=LLAMA3-8B2026.05 | 35 | |
| PGD-SphereBackbone=QWEN-2.5-7B2026.05 | 37.9 | |
| PGD-SphereBackbone=LLAMA3-8B2026.05 | 39.5 | |
| PGD-SphereBackbone=Mistral-7B-v22026.05 | 42.8 | |
| Jailbreak AntidoteBackbone=LLAMA3-8B2026.05 | 45.4 | |
| HSFBackbone=QWEN-2.5-7B2026.05 | 45.5 | |
| LoRA-SteerBackbone=Mistral-7B-v22026.05 | 45.7 | |
| Jailbreak AntidoteBackbone=Mistral-7B-v22026.05 | 46.1 | |
| Circuit BreakerBackbone=LLAMA3-8B2026.05 | 48.4 | |
| Jailbreak AntidoteBackbone=QWEN-2.5-7B2026.05 | 48.8 | |
| HSFBackbone=Mistral-7B-v22026.05 | 52 | |
| HSFBackbone=LLAMA3-8B2026.05 | 57.2 | |
| Circuit BreakerBackbone=Mistral-7B-v22026.05 | 82.7 |