FoodMonitor: leaderboard

FoodMonitor (2026) asks multimodal models to find food-safety violations in 477 one-minute commercial kitchen videos against a 27-item rule document, naming the rule, describing the violation and, for workers, tracking the person with boxes over time. The C score, 0 to 100 and higher is better, averages the detection F1 for environment and for person violations.

Metric: C score (0-100; the paper's 0-1 score times 100): the mean of the environment-violation F1 and the person-violation F1, on the 477 one-minute FoodMonitor kitchen surveillance clips (3,307 annotated violations of a 27-item food safety rule document), frames at 1 fps plus the full rule document, structured JSON output; a prediction counts only if its rule category matches and a text comparison model accepts its description as the same violation, one-to-one; person violations must first localize the worker with at least ten box-and-time anchors (mean IoU 0.3, coverage 0.6); all models in thinking mode; higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Seed 2.0 ProThinking36#226
2Gemini 3 Flash31.6#180
3Doubao Seed 2.0 LiteThinking29.8#204
4GLM-4.6V27.9#721
5Gemini 3 Pro25#146
6Qwen 3.5 397B A17B23.1#270
7Kimi K2.520.8#272
8Qwen 3 VL 235B A22B (Thinking)19.3#462
9GLM-4.6V-Flash13.5#1049
10MiMo-VL-7B-RL13.2#1160
11Qwen 3 VL 8B (Thinking)13#924

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=foodmonitor · How It Works · Data refreshed daily, snapshot 2026-11-22.