General Knowledge Question Answering on MMLU (Adversarial Evaluation)
86Task Success Rate (TSR)STAR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| STARBackbone=Qwen-Plus, Attack Setting=Independent2026.05 | 86 | 8 | |
| STARBackbone=Qwen-Plus, Attack Setting=Cooperative2026.05 | 85.75 | 7.25 | |
| G-SafeguardBackbone=Qwen-Plus, Attack Setting=Cooperative2026.05 | 84.75 | 11.25 | |
| G-SafeguardBackbone=Qwen-Plus, Attack Setting=Independent2026.05 | 82.25 | 12.75 | |
| ARGUSBackbone=Qwen-Plus, Attack Setting=Independent2026.05 | 81.5 | 15 | |
| STARBackbone=GPT-5-Nano, Attack Setting=Cooperative2026.05 | 80 | 12.25 | |
| ARGUSBackbone=Qwen-Plus, Attack Setting=Cooperative2026.05 | 79.25 | 16.5 | |
| Blind-GuardBackbone=Qwen-Plus, Attack Setting=Independent2026.05 | 78.25 | 17 | |
| STARBackbone=GPT-5-Nano, Attack Setting=Independent2026.05 | 77.75 | 15 | |
| STARBackbone=GPT-3.5-Turbo, Attack Setting=Cooperative2026.05 | 77 | 11.5 | |
| C&IBackbone=Qwen-Plus, Attack Setting=Independent2026.05 | 76.75 | 20.5 | |
| STARBackbone=GPT-3.5-Turbo, Attack Setting=Independent2026.05 | 76.75 | 11.25 | |
| Blind-GuardBackbone=Qwen-Plus, Attack Setting=Cooperative2026.05 | 75.25 | 20.5 | |
| C&IBackbone=Qwen-Plus, Attack Setting=Cooperative2026.05 | 74 | 22.5 | |
| G-SafeguardBackbone=GPT-5-Nano, Attack Setting=Independent2026.05 | 73.5 | 21.5 | |
| G-SafeguardBackbone=GPT-5-Nano, Attack Setting=Cooperative2026.05 | 70.25 | 25.25 | |
| Blind-GuardBackbone=GPT-5-Nano, Attack Setting=Independent2026.05 | 67.25 | 26.75 | |
| ARGUSBackbone=GPT-5-Nano, Attack Setting=Independent2026.05 | 62.25 | 34.5 | |
| Blind-GuardBackbone=GPT-3.5-Turbo, Attack Setting=Cooperative2026.05 | 61.5 | 32 | |
| C&IBackbone=GPT-3.5-Turbo, Attack Setting=Cooperative2026.05 | 59.75 | 33.5 | |
| Blind-GuardBackbone=GPT-3.5-Turbo, Attack Setting=Independent2026.05 | 59.25 | 34 | |
| G-SafeguardBackbone=GPT-3.5-Turbo, Attack Setting=Cooperative2026.05 | 59 | 35.75 | |
| C&IBackbone=GPT-5-Nano, Attack Setting=Independent2026.05 | 58 | 38.25 | |
| C&IBackbone=GPT-3.5-Turbo, Attack Setting=Independent2026.05 | 57.25 | 36.25 | |
| ARGUSBackbone=GPT-3.5-Turbo, Attack Setting=Cooperative2026.05 | 56.75 | 40.25 | |
| ARGUSBackbone=GPT-3.5-Turbo, Attack Setting=Independent2026.05 | 55.5 | 41.5 | |
| G-SafeguardBackbone=GPT-3.5-Turbo, Attack Setting=Independent2026.05 | 54.75 | 41.25 | |
| ARGUSBackbone=GPT-5-Nano, Attack Setting=Cooperative2026.05 | 51.5 | 46.25 | |
| Blind-GuardBackbone=GPT-5-Nano, Attack Setting=Cooperative2026.05 | 45 | 50 | |
| C&IBackbone=GPT-5-Nano, Attack Setting=Cooperative2026.05 | 41 | 57 |