PaSBench-Video - Healthcare: leaderboard
PaSBench-Video (2026) streams videos to multimodal models, which see only past frames in 3-second windows and must decide when to issue a safety warning before an accident. This board covers 33 synthesized healthcare risk videos. The score is First-Act at a 1-second lead time in percent: the first warning must fall between the annotated risk onset and 1 second before the accident and name the correct risk source; higher is better, and false alarms on safe videos are not part of the score.
Metric: First-Act@1.0 (%) on the healthcare (33 risk videos, synthesized clips): share of risk videos whose first warning, issued while the model streams 3-second windows of past frames only, lands between the annotated risk onset and 1 s before the accident and names the correct risk source; temperature 0, identical prompts; false positives on the no-risk split are not part of this score; higher is better. Source: arxiv.org. Saturation forecast: Around September 2028. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 27.3 | #161 |
| 2 | Qwen 3.5 27B | 18.2 | #359 |
| 3 | GPT-5.4 | 12.1 | #123 |
| 4 | Claude Opus 4.7 | 12.1 | #79 |
| 5 | Qwen 3.5 4B | 12.1 | #1175 |
| 6 | Qwen 3.5 35B A3B | 12.1 | #482 |
| 7 | Qwen 3.5 9B | 9.1 | #803 |
| 8 | GPT-5.4 Mini | 9.1 | #389 |
| 9 | Gemma 4 26B | 6.1 | #612 |
| 10 | Gemma 4 31B | 3 | #394 |
| 11 | Claude Haiku 4.5 | 0 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pasbench-video-healthcare · How It Works · Data refreshed daily, snapshot 2026-11-13.