Closed-loop End-to-End Autonomous Driving on Bench2Drive (base set)
86.23Driving Score (DS)DeepSight
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DeepSightParadigm=VLM, Expert=Think2Drive2026.05 | 86.23 | 71.36 | 201.71 | 16.11 | |
| SimLingoParadigm=VLM, Expert=PDM-Lite2026.05 | 85.94 | 66.82 | 244.18 | 25.49 | |
| DeepSight w/o adaptive CoTParadigm=VLM, Expert=Think2Drive2026.05 | 84.52 | 65.91 | 198.8 | 14.25 | |
| AutoVLAParadigm=VLM, Expert=Think2Drive2026.05 | 78.84 | 57.73 | 146.93 | 39.33 | |
| ORIONParadigm=VLM, Expert=Think2Drive2026.05 | 77.74 | 54.62 | 151.48 | 17.38 | |
| DiffusionDriveParadigm=E2E, Expert=PDM-Lite2026.05 | 77.68 | 52.72 | — | — | |
| DriveAdapter*Paradigm=E2E, Expert=Think2Drive, Expert feature distillation=true2026.05 | 64.22 | 33.08 | 70.22 | 16.01 | |
| ReasonPlanParadigm=VLM, Expert=Think2Drive2026.05 | 64.01 | 34.55 | 180.64 | 25.63 | |
| DriveTransParadigm=E2E, Expert=Think2Drive2026.05 | 63.46 | 35.01 | 100.64 | 20.78 | |
| ThinkTwice*Paradigm=E2E, Expert=Think2Drive, Expert feature distillation=true2026.05 | 62.44 | 31.23 | 69.33 | 16.22 | |
| TCP-traj*Paradigm=E2E, Expert=Think2Drive, Expert feature distillation=true2026.05 | 59.9 | 30 | 76.54 | 18.08 | |
| TCP-traj w/o distillationParadigm=E2E, Expert=Think2Drive2026.05 | 49.3 | 20.45 | 78.78 | 22.96 | |
| GenADParadigm=E2E, Expert=Think2Drive2026.05 | 44.81 | 15.9 | — | — | |
| MomADParadigm=E2E, Expert=Think2Drive2026.05 | 44.54 | 16.71 | 170.21 | 48.63 | |
| VADParadigm=E2E, Expert=Think2Drive2026.05 | 42.35 | 15 | 157.94 | 46.01 | |
| TCP*Paradigm=E2E, Expert=Think2Drive, Expert feature distillation=true2026.05 | 40.7 | 15 | 54.26 | 47.8 | |
| TCP-ctrl*Paradigm=E2E, Expert=Think2Drive, Expert feature distillation=true2026.05 | 30.47 | 7.27 | 55.97 | 51.51 |