TAR-Bench: leaderboard
TAR-Bench (2026) evaluates vision-language models on 960 human-curated questions about 80 held-out traffic-anomaly clips, zero-shot and without reasoning mode. Score is the unweighted mean of ten task scores covering question answering, temporal reasoning and scene understanding, in percent; higher is better.
Metric: Mean score (%; unweighted mean of the ten task scores: accuracy for binary and multiple-choice QA, mean IoU for temporal localization, scaled BERTScore F1 for the seven open-ended tasks; zero-shot in non-reasoning (direct-answer) mode through VLMEvalKit, mean of 3 runs, on TAR-Bench: 960 human-curated annotations for 80 held-out traffic-anomaly clips from 17 public YouTube videos). Source: arxiv.org. Saturation forecast: Around 2031. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 27B (Non-reasoning) | 39.5 | #359 (Qwen 3.5 27B) |
| 2 | Qwen 3.5 9B (Non-reasoning) | 39.1 | #803 (Qwen 3.5 9B) |
| 3 | Gemini 3.1 Pro (Preview) | 36.3 | #94 |
| 4 | Gemma 4 31B (IT) | 33.3 | #342 |
| 5 | Qwen 3 VL 8B Instruct | 30.9 | #981 |
| 6 | Qwen 3 VL 32B Instruct | 28 | #585 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tar-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.