VisAnomBench: leaderboard
Time-series anomaly detection from plots: a vision-language model marks the anomalous intervals of univariate series drawn from the KPI, GutenTAG, UCR-EGI and UCR-TSAD collections; VisAnomBench (Zhou et al., 2026). Interval-level F1 in percent over 740 held-out series, where a predicted interval that overlaps a true anomaly counts as a hit; higher is better.
Metric: Interval-level F1 (%; 740 test series, a predicted interval overlapping a ground-truth anomaly interval is a true positive, one per ground-truth interval). Source: arxiv.org. Saturation forecast: Around July 2028. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 4 Maverick | 37.96 | #1105 |
| 2 | Qwen 2.5 VL 7B | 33.79 | #1796 |
| 3 | Grok 4 Fast | 26.33 | #438 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=visanombench · How It Works · Data refreshed daily, snapshot 2026-11-12.