Workflow Steering Defense on ASB-Bench
68TASRG-Safeguard
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| G-SafeguardModel=Gemini-2.5-Flash2026.05 | 68 | 71 | |
| ARGUSModel=Qwen-3.5-Flash2026.05 | 53 | 59 | |
| w/o DefenseModel=GPT-4o-mini2026.05 | 52 | 59 | |
| w/o DefenseModel=Gemini-2.5-Flash2026.05 | 51 | 63 | |
| G-SafeguardModel=Qwen-3.5-Flash2026.05 | 50 | 64 | |
| w/o DefenseModel=Qwen-3.5-Flash2026.05 | 49 | 62 | |
| G-SafeguardModel=GPT-4o-mini2026.05 | 44 | 59 | |
| ARGUSModel=DeepSeek-V32026.05 | 41 | 64 | |
| G-SafeguardModel=DeepSeek-V32026.05 | 41 | 53 | |
| w/o DefenseModel=DeepSeek-V32026.05 | 40 | 54 | |
| ARGUSModel=GPT-4o-mini2026.05 | 36 | 47 | |
| FLOWGUARDModel=Gemini-2.5-Flash2026.05 | 34 | 34 | |
| FLOWGUARDModel=Qwen-3.5-Flash2026.05 | 33 | 46 | |
| FLOWGUARDModel=GPT-4o-mini2026.05 | 22 | 27 | |
| FLOWGUARDModel=DeepSeek-V32026.05 | 22 | 30 | |
| ARGUSModel=Gemini-2.5-Flash2026.05 | 17 | 22 |