VT-Bench DVM-Car QA - Conditional Mean: leaderboard
Averaging a numeric attribute over rows filtered relative to the matching car. DVM-Car QA, the vision-tabular reasoning set of VT-Bench (Jia et al., Nanjing University, 2026), pairs a front-view car photo with a table of 10, 20 or 50 candidate vehicles, exactly one of which matches the photo through a visual key such as colour or maker (5,400 items). Models answer zero-shot at temperature 0; exact-match accuracy in percent, higher is better.
Metric: Exact-match accuracy (%; average a numeric attribute over rows filtered relative to the target car; zero-shot at temperature 0). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 49.67 | #141 |
| 2 | GPT-4.1 | 21.11 | #492 |
| 3 | GLM-4.1V-9B (Thinking) | 20.56 | #1061 (GLM-4.1V-9B) |
| 4 | Pixtral-12B | 8.56 | #2493 |
| 5 | Qwen 3 VL 8B Instruct | 5.78 | #982 |
| 6 | InternVL3-8B | 4.33 | #1695 |
| 7 | Llama 3.2 11B Instruct | 4.33 | #5039 |
| 8 | Qwen 3 VL 8B (Thinking) | 2.3 | #887 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vt-bench-dvm-car-qa-conditional-mean · How It Works · Data refreshed daily, snapshot 2026-11-12.