Bench2Drive-VL (Camera, CoT): leaderboard

Closed-loop driving on Bench2Drive-VL (2026), which plugs a vision-language model into the CARLA simulator: the model answers a chain of driving questions every few frames and its chosen direction and speed keys steer the ego vehicle. This board uses front camera images as visual input and graph chain-of-thought prompting (answers to earlier questions passed as context). Driving Score from 0 to 100 (route completion discounted by collisions, red-light violations and timeouts, the Bench2Drive protocol); higher is better.

Metric: Driving Score (0-100; Bench2Drive closed-loop protocol: route completion times violation penalties such as 0.6 per collision, 0.8 per red-light or stop-sign violation, 0.7 per timeout) of the VLM driving the CARLA ego vehicle through the Bench2Drive-VL text-to-control action module, with front camera images as visual input and graph chain-of-thought prompting (answers to earlier questions passed as context); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 VL 3B Instruct42.24#2720
2InternVL3-2B33#3318
3Gemma 3 4B (IT)15.04#4227

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=bench2drive-vl-camera-cot · How It Works · Data refreshed daily, snapshot 2026-11-22.