Question Answering on NuScenes-QA v1.0 (test)
88.55Accuracy (Exist)KLDrive-GT
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| KLDrive-GTTraining Protocol=use frozen LLM with few-shot ICL only, variant=KG from ground-truth annotations2026.03 | 88.55 | 78.92 | 80.11 | 85.72 | 82.01 | 84.49 | |
| MAPLMTraining Protocol=trained since they have task-specific parameters2026.03 | 79.29 | 18.45 | 60.67 | 58.38 | 70.55 | 60.17 | |
| DriveLMTraining Protocol=trained since they have task-specific parameters2026.03 | 77.29 | 16.92 | 53.67 | 50.64 | 67.41 | 56.42 | |
| LiDAR-LLMTraining Protocol=trained since they have task-specific parameters2026.03 | 76.41 | 15.84 | 50.63 | 48 | 69.11 | 55.08 | |
| KLDrive-KGTraining Protocol=use frozen LLM with few-shot ICL only, variant=full system with predicted perception outputs2026.03 | 74.67 | 64.46 | 54.3 | 57.84 | 53.23 | 65.04 | |
| CREMATraining Protocol=trained since they have task-specific parameters2026.03 | 73.17 | 15.69 | 53.69 | 52.63 | 72.03 | 55.26 | |
| Qwen3Vision-8BTraining Protocol=use frozen LLM with few-shot ICL only2026.03 | 64.16 | 10.3 | 27.66 | 26.11 | 39.63 | 39.53 | |
| IS-FusionTraining Protocol=trained since they have task-specific parameters2026.03 | 52.7 | 0.04 | 0.42 | 0.67 | 44.87 | 24.86 | |
| FocalFormer3DTraining Protocol=trained since they have task-specific parameters2026.03 | 50.28 | 0.01 | 0.55 | 0.22 | 47.36 | 24.02 | |
| Internvl-3.5-8bTraining Protocol=use frozen LLM with few-shot ICL only2026.03 | 49.86 | 11.16 | 35.74 | 22.29 | 38.72 | 34.28 | |
| LLaVA-v1.6-mistral-7bTraining Protocol=use frozen LLM with few-shot ICL only2026.03 | 48.64 | 6.83 | 1.48 | 5.09 | 52.8 | 26.17 | |
| RayDNTraining Protocol=trained since they have task-specific parameters2026.03 | 44.13 | 0.04 | 1.48 | 0.28 | 44.38 | 21.46 |