GeoDrive-Bench (Reasoning): leaderboard

GeoDrive-Bench (2026) overall accuracy when the model first writes a four-step chain of thought (infer the region, recall its traffic rules, reason over the scene, then answer) on driving questions from six countries. Percent; higher is better.

Metric: Overall accuracy (%) on 5,053 human-validated multiple-choice questions on driving scenes from China, the US, the UK, Japan, Singapore and India (perception, prediction, planning and region-rule reasoning), without country labels; Reasoning prompting (a four-step free-form chain of thought: geographic inference, cultural rule recall, visual reasoning, answer selection); vLLM inference, 512 output tokens; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked. Updated 30 Sep 2026.

Top models

#ModelScoreOverall rank
1Gemma 3 12B58.5#2059
2InternVL3-8B57.4#1727
3InternVL3.5-8B54.6#1558
4Qwen 2.5 VL 7B Instruct51.8#1867
5Qwen 3 VL 8B50.5#899
6Llama 3.2 11B Vision47#4055
7LLaVA-1.6-7B43.2#3932

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=geodrive-bench-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-16.