Pro2Bench - Intervention Timing: leaderboard
Pro2Bench (Plan, Watch, Recover, 2026) tests proactive procedural assistants on egocentric and instructional video: at each decision point the model either stays silent or interrupts with guidance, evaluated zero-shot over six sets including the new EgoProactive recordings. This board is the geometric mean of interrupt and silent F1 in percent, which scores only when to speak and drops to zero for always-interrupt or always-silent policies; higher is better.
Metric: Geometric mean of the interrupt-class and silent-class F1 (G-Mean F1, 0-1 scaled to %) for deciding when to speak; zero-shot monolithic assistant (8 uniformly sampled frames per clip, greedy decoding for the interrupt or silent decision), mean over the six evaluation sets (EgoExo4D, EPIC-KITCHENS, HowTo100M, Ego4D, HoloAssist, EgoProactive); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 51 | #102 |
| 2 | Gemini 3.1 Pro (Preview) | 48 | #94 |
| 3 | Llama 4 Maverick | 45 | #1099 |
| 4 | Qwen 3 VL 235B A22B | 43 | #479 |
| 5 | Qwen 3.6 27B | 31 | #341 |
| 6 | GPT-5.2 | 7 | #176 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pro2bench-intervention-timing · How It Works · Data refreshed daily, snapshot 2026-11-13.