EgoDyn-Bench: leaderboard

EgoDyn-Bench (2026) tests whether vision-language models understand the motion of the ego vehicle in driving video, through multiple-choice questions on speed, acceleration and turning answered from video alone. This board is balanced accuracy in percent on the semantic ego-motion questions; higher is better.

Metric: Balanced accuracy (%) of semantic ego-motion questions (14 templates of discrete choices on speed, acceleration and turning) from driving video alone, no explicit dynamic state inputs; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro (Preview)47#119
2GPT-5.145.2#240
3Cosmos-Reason2-8B39.9#755
4Qwen 3 VL 4B39.8#1151
5Qwen 3 VL 8B38.9#917
6InternVL3.5-8B38.8#1604
7RoboTron-Drive38.6
8Claude Sonnet 4.538#261
9ImpromptuVLA (EgoDyn-Bench checkpoint unspecified)37.8
10InternVL3.5-38B37.6#1021
11Qwen 3 VL 2B37.4#2134
12CamReasoner-8B36.9
13Gemini 2.0 Flash36.5#723
14InternVL3.5-4B36#1930
15Cosmos-Reason2-2B35.1#1425

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

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