Pairwise Trajectory Preference Evaluation on DriveCritic (test)
82.83Accuracy (ACCTPS)DriveJudge-8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DriveJudge-8BFinetuning=true2026.06 | 82.83 | — | |
| DriveJudge-2BFinetuning=true2026.06 | 80.63 | — | |
| DriveCriticFinetuning=true2026.06 | 76.33 | — | |
| ADE/FDEFinetuning=false2026.06 | 70.53 | — | |
| Supervised ClassifierFinetuning=true2026.06 | 64.8 | — | |
| EPDMSFinetuning=false2026.06 | 56.5 | — | |
| Gemini3.1Finetuning=false2026.06 | 55.45 | — | |
| GPT-5Finetuning=false2026.06 | 55.2 | — | |
| RandomFinetuning=false2026.06 | 50 | — | |
| DriveCriticFine-tuning=✓2025.10 | — | 76 | |
| EPDMSFine-tuning=✗2025.10 | — | 41.4 | |
| GPT-5Fine-tuning=✗, zero-shot=true2025.10 | — | 55.2 | |
| OpenAI-o3Fine-tuning=✗, zero-shot=true2025.10 | — | 53.3 | |
| Qwen2.5-VL-7BFine-tuning=✗, zero-shot=true2025.10 | — | 48 | |
| Supervised Pairwise ClassifierFine-tuning=✓, Backbone=ResNet-1012025.10 | — | 64.8 |