Generative AI Output Safety on HarmBench and AdvBench (test)
82.88Safe RateReliable Consensus Sampling
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Reliable Consensus Samplingadversary_condition=f < ⌈n/2⌉2025.12 | 82.88 | 0 | 24.26 | |
| Reliable Consensus Samplingadversary_condition=f = ⌈n/2⌉2025.12 | 81.49 | 0 | 22.21 | |
| Reliable Consensus Samplingadversary_condition=AVG2025.12 | 77.42 | 0 | 21.63 | |
| Reliable Consensus Samplingadversary_condition=Collusion2025.12 | 67.9 | 0 | 18.42 | |
| Consensus Samplingadversary_condition=f < ⌈n/2⌉2025.12 | 20.39 | 71.79 | 22.6 | |
| Consensus Samplingadversary_condition=f = ⌈n/2⌉2025.12 | 15.83 | 72.52 | 26.73 | |
| Consensus Samplingadversary_condition=AVG2025.12 | 12.56 | 75.75 | 23.59 | |
| Consensus Samplingadversary_condition=Collusion2025.12 | 1.45 | 82.95 | 21.43 |