TAR-Bench - Temporal Description: leaderboard

Metric: Scaled BERTScore F1 (%; a description of what happens in a given interval, 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 2035. 11 models tracked.

Top models

#ModelScore
1Qwen 3.5 27B (Non-reasoning)14.9
2Qwen 3.5 9B (Non-reasoning)13.1
3Gemma 4 31B (IT)12.4
4Qwen 3 VL 8B Instruct11.2
5Gemini 3.1 Pro (Preview)10.6
6Qwen 3 VL 32B Instruct6.6

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