TAR-Bench - Scene Description: leaderboard

Metric: Scaled BERTScore F1 (%; a physical description of the scene, compared with the reference; 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 2034. 11 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)24
2Qwen 3.5 9B (Non-reasoning)22.1
3Qwen 3.5 27B (Non-reasoning)22
4Qwen 3 VL 8B Instruct21.9
5Gemma 4 31B (IT)15.6
6Qwen 3 VL 32B Instruct14.1

Interactive version: theaggregate.ai/benchmark?slug=tar-bench-scene-description · How It Works · Data refreshed daily, snapshot 2026-09-29.