Drive-P2D - Action Safety (Decision-2): leaderboard
The Drive-P2D task of rating the safety of a specified, possibly suboptimal action (single-choice). 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. Accuracy in percent (exact match), zero-shot over all scenarios and averaged over two runs; higher is better.
Metric: Accuracy (%; rate the safety of a specified, possibly suboptimal action (single choice); zero-shot, all scenarios, mean of two runs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 55.11 | #262 |
| 2 | GPT-4.1 | 51.91 | #492 |
| 3 | Llama 3.2 90B Vision Instruct | 49.55 | #1757 |
| 4 | Qwen 2.5 VL 72B Instruct | 38.2 | #845 |
| 5 | Llama 3.2 11B Instruct | 37.99 | #5039 |
| 6 | Qwen 2.5 VL 7B Instruct | 35.48 | #1842 |
| 7 | Gemma 3 27B (IT) | 31.65 | #1306 |
| 8 | Phi-4 Multimodal Instruct | 30.72 | #3914 |
| 9 | Gemma 3 4B (IT) | 30.66 | #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-action-safety-decision-2 · How It Works · Data refreshed daily, snapshot 2026-11-12.