Adversarial Robustness Evaluation on HarmBench
1Attack Success Rate (ASR)SafeMoE-XL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SafeMoE-XLArchitecture=XL2026.05 | 1 | 1.1 | |
| SafeMoE-QwenBackbone=Qwen2026.05 | 20 | 2.1 | |
| Qwen3-4B-InstructFine-tuning=Instruct, Number of parameters=4B2026.05 | 24 | 2.7 | |
| Mistral-SFTFine-tuning=SFT2026.05 | 42 | 3.15 |