Instruction Hierarchy Robustness on Human Red-teaming IH-Challenge
271Number of TasksGPT-5-Mini
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| GPT-5-Minisystem mitigation=none2026.03 | 271 | 32.84 | 36.2 | 1.5 | |
| GPT-5-Mini-R + Monitoroutput monitor=true2026.03 | 268 | 43.08 | 7.1 | 0.2 | |
| GPT-5-Mini-Rsystem mitigation=none2026.03 | 265 | 52.39 | 11.7 | 0.4 |