Prompt-level trigger detection on SocialHarmBench
67.64Trigger Coverage @8AERIC
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| AERICTarget Model=Gemma-4-E4B-it, Safe-trigger budget=10%2026.05 | 67.64 | 67.81 | 69.18 | 73.63 | 41.8579 | |
| AERICTarget Model=Qwen3-8B, Safe-trigger budget=10%2026.05 | 63.53 | 64.38 | 65.92 | 68.49 | 32.7688 | |
| WildGuard PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.05 | 26.71 | 27.23 | 27.23 | 33.22 | 15.2911 | |
| WildGuard PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.05 | 11.3 | 11.82 | 11.99 | 11.99 | 6.5565 | |
| Qwen3Guard-Stream-4BTarget Model=Gemma-4-E4B-it2026.05 | 5.99 | 10.79 | 17.64 | 29.62 | 10.0856 | |
| Qwen3Guard-Stream-4BTarget Model=Qwen3-8B2026.05 | 2.74 | 3.94 | 6.51 | 7.19 | 3.411 | |
| ShieldGemma-9B PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.05 | 1.2 | 3.25 | 11.47 | 20.38 | 5.3904 | |
| ShieldGemma-9B PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.05 | 1.2 | 1.71 | 6.68 | 11.99 | 2.7432 |