Prompt-level trigger detection on HarmBench DirectRequest
57.5Trigger Coverage @8AERIC
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| AERICTarget Model=Qwen3-8B, Safe-trigger budget=10%2026.05 | 57.5 | 58.75 | 61.25 | 64.38 | 32.5469 | |
| AERICTarget Model=Gemma-4-E4B-it, Safe-trigger budget=10%2026.05 | 34.38 | 35.94 | 40.62 | 46.56 | 23.5281 | |
| WildGuard-7B PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.05 | 22.19 | 25.31 | 28.75 | 37.5 | 15.3313 | |
| WildGuard-7B PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.05 | 13.13 | 14.38 | 14.38 | 14.38 | 7.3125 | |
| Qwen3Guard-Stream-4BTarget Model=Qwen3-8B2026.05 | 4.69 | 5.31 | 8.13 | 10.63 | 4.3094 | |
| Qwen3Guard-Stream-4BTarget Model=Gemma-4-E4B-it2026.05 | 4.38 | 7.5 | 15.31 | 32.81 | 9.3375 | |
| ShieldGemma-9B PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.05 | 3.12 | 7.81 | 16.56 | 21.88 | 6.7625 | |
| ShieldGemma-9B PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.05 | 2.81 | 6.88 | 14.69 | 20.94 | 6.3937 |