Driving View Evidence QA: leaderboard
A driving VQA benchmark (2026) built from nuScenes: 122 conflict-centric questions about causality, counterfactuals and road-user intent, each grounded in one of six synchronized camera views or in none. This board is view-selection accuracy in percent, the share of questions for which the model names the camera that holds the evidence (or none), mean of five runs; higher is better.
Metric: View-selection exact-match accuracy (%), mean of 5 runs: pick the camera that holds the evidence (or None when no view supports the question) on 122 conflict-centric questions over six synchronized nuScenes camera views from 73 scenes (causality, counterfactual and intent), zero-shot, deterministic decoding; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 77.54 | #123 |
| 2 | Qwen 2.5 VL 7B Instruct | 12.62 | #1833 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=driving-view-evidence-qa · How It Works · Data refreshed daily, snapshot 2026-11-13.