EgoSafetyBench: leaderboard
EgoSafetyBench (2026) evaluates vision-language models as runtime safety guards on generated egocentric robot videos from homes and factories, judged one half-second chunk at a time. The headline is video-level balanced accuracy on 800 situational scenarios, the mean of the recall on unsafe and on safe videos, in percent; higher is better.
Metric: Video-level balanced accuracy (%; mean of the recall on unsafe videos and on safe videos over the 800 situational-track scenarios, a video called unsafe when any of its chunks is flagged, macro-averaged over videos; ten VLM safety guards judging egocentric robot videos split into ten half-second chunks (ten frames each at 1280x720), each chunk judged alone under the current-chunk protocol with a single-word verdict and greedy decoding; higher is better). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Flash Lite | 87.3 | #364 |
| 2 | Qwen 3.5 4B | 85.5 | #1175 |
| 3 | Gemini 3.5 Flash | 85.4 | #101 |
| 4 | Qwen 3 VL 4B Instruct | 84.9 | #1279 |
| 5 | Claude Sonnet 4.6 | 82.5 | #161 |
| 6 | Gemma 3 4B (IT) | 65.3 | #3831 |
| 7 | Qwen 3.5 0.8B | 65.3 | #8081 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=egosafetybench · How It Works · Data refreshed daily, snapshot 2026-11-13.