AnyGroundBench - Public Security TVG: leaderboard

Metric: tIoU@0.3 (%; share of queries whose temporal IoU is at least 0.3) on the 250 public security-domain test queries (UCA surveillance and DoTA traffic-accident videos), temporal video grounding (the queried event's start and end), zero-shot, default inference settings (Qwen3.5 with thinking disabled); higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 15 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)69.4
2Gemini 3 Flash66.8
3Gemini 2.5 Pro65.8
4GPT-5.161.2
5GPT-4o56
6Gemini 2.5 Flash51.2
7Qwen 3.5 9B (Non-reasoning)50.4
8Qwen 3.5 4B (Non-reasoning)49.2
9InternVL3-14B13.2
10InternVL3-8B4.4
11InternVL3.5-8B3.6

Interactive version: theaggregate.ai/benchmark?slug=anygroundbench-public-security-tvg · How It Works · Data refreshed daily, snapshot 2026-09-29.