Agentic Driving on Agentic Driving environment
85.2Success RatePrivileged
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| PrivilegedAccess Level=Privileged2025.10 | 85.2 | 13.2 | 100 | 100 | 100 | 0 | |
| Myopic-PrivilegedAccess Level=Privileged, Evaluation Protocol=Fine-tuned via PPO2025.10 | 81.9 | 16.2 | — | — | — | — | |
| ReGuardEvaluation Protocol=Proposed method2025.10 | 77.3 | 18.8 | 99 | 99 | 56 | 1 | |
| Myopic-RealisticEvaluation Protocol=Fine-tuned via PPO2025.10 | 44.4 | 55.6 | — | — | — | — | |
| GPT-5 (Think: minimal)Evaluation Protocol=Zero-shot2025.10 | 39.5 | 59.4 | 29 | 2 | 92 | 94 | |
| GPT-4oEvaluation Protocol=Zero-shot2025.10 | 31.7 | 68.2 | 29 | 0 | 1 | 100 | |
| GPT-4o-miniEvaluation Protocol=Zero-shot2025.10 | 23.4 | 70.2 | 29 | 25 | 74 | 77 | |
| Llama-3.1-70B-InstructEvaluation Protocol=Zero-shot2025.10 | 22.8 | 76.8 | 0 | 95 | 5 | 3 | |
| Qwen-2.5-72B-InstructEvaluation Protocol=Zero-shot2025.10 | 22.3 | 72.5 | 78 | 52 | 58 | 58 |