Red-teaming on Llama-3-70B
45Attack Success Rate (ASR)Adaptive Instruction Composition
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Adaptive Instruction CompositionBandit Type=Aggressive, lambda=0.012026.04 | 45 | 4,500 | |
| Adaptive Instruction CompositionBandit Type=Subtle, lambda=0.012026.04 | 15.5 | 1,552 | |
| WildTeamingAttack Strategy=WildTeaming2026.04 | 8.8 | 876 |