Text Moderation on WildJailbreak Adv. Harmful n = 2,000
278Flagged CountGemma 3-1B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Gemma 3-1BBackbone=Gemma 3-1B2026.04 | 278 | 13.7 | 6.3 | 94.6 | |
| Llama 3.1-8BBackbone=Llama 3.1-8B2026.04 | 218 | 10.6 | 2 | 95.1 | |
| RESTABackbone=Llama 3.1-8B2026.04 | 199 | 9.8 | 1 | 94.4 | |
| Qwen3-14BBackbone=Qwen3-14B2026.04 | 165 | 9.7 | 2.8 | 94.4 | |
| Phi-3.5-4BBackbone=Phi-3.5-4B2026.04 | 126 | 8.1 | 2 | 92.1 | |
| SmoothLLMBackbone=Llama 3.1-8B2026.04 | 122 | 6.8 | 0.6 | 95.4 | |
| AdaSteerBackbone=Llama 3.1-8B2026.04 | 85 | 5.7 | 1.6 | 94.7 | |
| TSABackbone=Gemma 3-1B, # Iterations=5.7, Time (s)=21.1, Net (s)=13.72026.04 | 59 | 6.5 | 3.1 | 84.4 | |
| GPT-OSS-20BBackbone=GPT-OSS-20B2026.04 | 17 | 2.3 | 0.4 | 84.9 | |
| TSABackbone=GPT-OSS-20B, # Iterations=1.2, Time (s)=12.1, Net (s)=10.22026.04 | 0 | 0.9 | 0.3 | 9.9 | |
| TSABackbone=Qwen3-14B, # Iterations=2.1, Time (s)=28.3, Net (s)=25.42026.04 | 0 | 2.3 | 1.6 | 10 | |
| TSABackbone=Llama 3.1-8B, # Iterations=1.0, Time (s)=12.2, Net (s)=10.52026.04 | 0 | 1.2 | 0.2 | 9.8 | |
| TSABackbone=Phi-3.5-4B, # Iterations=1.8, Time (s)=12.4, Net (s)=9.82026.04 | 0 | 1.9 | 0.8 | 9.9 |