Drive-P2D - High-Risk Scenarios: leaderboard
Drive-P2D (Tang et al., Zhejiang University, 2026) poses human-annotated selection questions on front-camera driving images from nuScenes, KITTI and BDD100K; models reason inside think tags and only the final answer is scored. This board is the mean in percent of the six task accuracies on the high-risk split (about 1.6K questions from scenes whose mean rating by two annotators is at least 4 on a five-point risk scale), zero-shot and averaged over two runs; higher is better.
Metric: Average score (%; unweighted mean of the six task accuracies on the high-risk split, scenes whose mean rating by two annotators is at least 4 on a five-point risk scale, about 1.6K questions; zero-shot, mean of two runs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 71.82 | #492 |
| 2 | Gemini 2.5 Pro | 64.66 | #262 |
| 3 | Llama 3.2 90B Vision Instruct | 59.24 | #1757 |
| 4 | Qwen 2.5 VL 72B Instruct | 58.25 | #845 |
| 5 | Llama 3.2 11B Instruct | 49.52 | #5039 |
| 6 | Qwen 2.5 VL 7B Instruct | 49.4 | #1842 |
| 7 | Gemma 3 27B (IT) | 48.01 | #1306 |
| 8 | Phi-4 Multimodal Instruct | 46.77 | #3914 |
| 9 | Gemma 3 4B (IT) | 36.84 | #3848 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=drive-p2d-high-risk-scenarios · How It Works · Data refreshed daily, snapshot 2026-11-12.