List your company

Prompt-level trigger detection on SocialHarmBench

67.64Trigger Coverage @8AERIC
-1.457616.481234.4252.3588May 13, 2026

Evaluation results

MethodLinks
AERICTarget Model=Gemma-4-E4B-it, Safe-trigger budget=10%2026.0567.6467.8169.1873.6341.8579
AERICTarget Model=Qwen3-8B, Safe-trigger budget=10%2026.0563.5364.3865.9268.4932.7688
WildGuard PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.0526.7127.2327.2333.2215.2911
WildGuard PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.0511.311.8211.9911.996.5565
Qwen3Guard-Stream-4BTarget Model=Gemma-4-E4B-it2026.055.9910.7917.6429.6210.0856
Qwen3Guard-Stream-4BTarget Model=Qwen3-8B2026.052.743.946.517.193.411
ShieldGemma-9B PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.051.23.2511.4720.385.3904
ShieldGemma-9B PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.051.21.716.6811.992.7432