Multi-task Language Understanding on MMLU (Adversarial Robustness)
100Accuracy under AttackReporting-and-penalty mechanism
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Reporting-and-penalty mechanismAttack Type=All Attack Types, LLM Model=All Above Models2026.04 | 100 | 100 | |
| SentinelNetAttack Type=Collaboration Attack, LLM Model=Qwen2.5-72B-128K2026.04 | 76 | 93 | |
| GroupGuardAttack Type=False Consensus, LLM Model=Deepseek-V3.12026.04 | 52 | 92 | |
| GroupGuardAttack Type=False Consensus, LLM Model=Qwen3-32B2026.04 | 44 | 97 | |
| GroupGuardAttack Type=False Consensus, LLM Model=Qwen3-235B2026.04 | 43 | 84 |