LLM Red-teaming on GFN-defended Target Model
0.33Unsuccessful Attack Rate (UA)PPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PPO2026.05 | 0.33 | 0.03 | |
| SFT2026.05 | 0.67 | 0.07 | |
| PPO + Curiosity2026.05 | 1 | 0.1 | |
| DPO2026.05 | 2 | 0.2 | |
| Rainbow Teaming2026.05 | 3.33 | 0.33 | |
| GFN2026.05 | 5 | 4.69 | |
| ICL2026.05 | 5.33 | 0.52 | |
| Jailbreak R1CoT=enabled2026.05 | 30.33 | 2.96 | |
| S-GFN2026.05 | 43.33 | 22.53 |