Text Moderation on WildJailbreak Adv. Benign n = 210
3Flagged CountAdaSteer
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AdaSteerBackbone=Llama 3.1-8B2026.04 | 3 | 0.022 | 0.002 | 0.359 | |
| Gemma 3-1BBackbone=Gemma 3-1B2026.04 | 3 | 0.023 | 0.006 | 0.354 | |
| Qwen3-14BBackbone=Qwen3-14B2026.04 | 2 | 0.014 | 0.001 | 0.253 | |
| SmoothLLMBackbone=Llama 3.1-8B2026.04 | 2 | 0.018 | 0.004 | 0.358 | |
| RESTABackbone=Llama 3.1-8B2026.04 | 2 | 0.013 | 0.004 | 0.351 | |
| Llama 3.1-8BBackbone=Llama 3.1-8B2026.04 | 1 | 0.017 | 0.003 | 0.356 | |
| GPT-OSS-20BBackbone=GPT-OSS-20B2026.04 | 0 | 0.011 | 0.003 | 0.27 | |
| TSABackbone=GPT-OSS-20B, # Iterations=1.0, Time (s)=8.0, Net (s)=6.32026.04 | 0 | 0.009 | 0.003 | 0.097 | |
| TSABackbone=Qwen3-14B, # Iterations=2.1, Time (s)=29.2, Net (s)=26.32026.04 | 0 | 0.01 | 0.001 | 0.098 | |
| TSABackbone=Llama 3.1-8B, # Iterations=1.0, Time (s)=12.6, Net (s)=10.92026.04 | 0 | 0.009 | 0.002 | 0.098 | |
| Phi-3.5-4BBackbone=Phi-3.5-4B2026.04 | 0 | 0.015 | 0.001 | 0.205 | |
| TSABackbone=Phi-3.5-4B, # Iterations=1.3, Time (s)=10.2, Net (s)=8.22026.04 | 0 | 0.009 | 0.001 | 0.098 | |
| TSABackbone=Gemma 3-1B, # Iterations=3.1, Time (s)=12.4, Net (s)=8.42026.04 | 0 | 0.019 | 0.004 | 0.199 |