ProAct-75 - Trigger Detection: leaderboard
Metric: Trigger detection macro-F1 (%, 0-100): whether the current window needs a robot intervention, averaged over the two classes; ProAct-75 best-view test split (516 videos across 75 assistance, maintenance and safety-monitoring tasks; 5-keyframe sliding windows, stride 3); zero-shot two-stage protocol: the model decides trigger and task from the window and the task list, then, if triggered, the current step and future steps from the task graph; nucleus sampling at temperature 0.7; higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 30B A3B Instruct | 66.25 | #365 |
| 2 | Gemini 2.5 Flash | 64.24 | #237 |
| 3 | Qwen 2.5 VL 32B Instruct | 62.36 | #443 |
| 4 | Gemini 2.5 Pro | 55.94 | #145 |
| 5 | GPT-4o | 46.7 | #333 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=proact-75-trigger-detection · How It Works · Data refreshed daily, snapshot 2026-10-11.