Scenario Generation on SafeBench CARLA v3 (scenario-level)
50.7Attack Success RateChatScene (CS) + EVODRIVE
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ChatScene (CS) + EVODRIVEEvolution stage=w/ EVODRIVE2026.06 | 50.7 | 92.4 | 93.4 | 45.7 | 143.5 | |
| ChatScene (CS)Evolution stage=Parent2026.06 | 39.4 | 72.9 | 81.2 | 51.6 | — | |
| AdvSim (AS) + EVODRIVEEvolution stage=w/ EVODRIVE2026.06 | 38.2 | 61.5 | 79.1 | 51.7 | 142.5 | |
| AT + EVODRIVEEvolution stage=w/ EVODRIVE2026.06 | 33 | 60.7 | 72.9 | 54.8 | 255.7 | |
| Human + EVODRIVEEvolution stage=w/ EVODRIVE2026.06 | 30.1 | 92.7 | 71.3 | 57.8 | 73.7 | |
| HumanEvolution stage=Parent2026.06 | 26.4 | 85.6 | 66.4 | 60.2 | — | |
| AdvSim (AS)Evolution stage=Parent2026.06 | 26 | 54.4 | 63.1 | 58.9 | — | |
| ATEvolution stage=Parent2026.06 | 19.7 | 51.4 | 53.7 | 63.3 | — |