DashBench - Mean Agreement: leaderboard

DashBench evaluates video understanding on 1,709 driving clips. Five explanation prompts and binary event detection are scored separately; mean agreement excludes the binary task. The 823-clip corner-case-window subset and training-free decoder variants retain their own conditions. Q5 asks for definition-guided corner-case identification and description. This board reports lLM-judged answer accuracy (%). Higher is better.

Metric: LLM-judged answer accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 22 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-534.75#167
2Gemini 3 Flash34.41#163
3Qwen 3 VL 235B A22B20.67#484
4Qwen 3 VL 8B18.48#899
5Qwen 3 VL 32B17.21#609
6Ovis2.5-9B16.43#1236
7InternVL3.5-241B16.07#625
8InternVL3.5-38B12.78#986
9Tarsier2-7B12.61#844
10MiniCPM-V-4.512.34#1390
11LlaVA-OV-RS (7B)10.01
12InternVL3.5-8BEvaluation run: DashBench TRiCD9.8#1558
13InternVideo2.5-Chat-8B7.76#2017
14InternLM-XComposer2.57.74#3981
15GLM-4.1V-9B (Thinking)6.33

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

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