DriveSpatial - Construction: leaderboard

DriveSpatial (2026) tests whether vision-language models can build and reason over a coherent driving scene from synchronized multi-view camera frames, with human-verified questions generated from scene graphs of nuScenes, Waymo, TruckScenes, AV2 and ONCE. This board covers the scene-construction tasks on the full 15,670-question set, scored as accuracy in percent; higher is better.

Metric: Accuracy (%) on the scene-construction tasks (building a coherent scene from multi-view, multi-frame camera input) on the full benchmark (15,670 human-verified questions over 20 tasks from nuScenes, Waymo, TruckScenes, AV2 and ONCE); open-source generalist and driving or spatial specialist VLMs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B42.76#917
2InternVL3.5-8B39#1604
3DriveSpatial SenseNova-SI (checkpoint unspecified)37.91
4Gemma 3 12B (IT)35.55#1887
5LLaVA-OneVision-7B33.73#2981
6InternVL3-8B30.63#1782
7DriveSpatial Ego3D (checkpoint unspecified)30.21
8SpaceThinker27.16#1448
9LLaVA-Video-7B26.94#2276
10LongVA-7B26.84#3456
11RoboTron-Drive24.31
12DeepSeek-VL2-Small23.92#2464

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=drivespatial-construction · How It Works · Data refreshed daily, snapshot 2026-11-21.