EgoSafetyBench - Sequential Detection: 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. This board is the percent of unsafe videos whose first alarm falls within half a second after the annotated hazard onset, neither late nor premature; higher is better.
Metric: Caught rate (%; share of unsafe situational videos whose first alarm falls within one half-second chunk after the annotated hazard onset, chunks streamed in order; 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 May 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 57.4 | #161 |
| 2 | Gemini 3.5 Flash | 54.2 | #101 |
| 3 | Gemini 3.1 Flash Lite | 52.3 | #364 |
| 4 | Qwen 3 VL 4B Instruct | 47.7 | #1279 |
| 5 | Qwen 3.5 4B | 45.6 | #1175 |
| 6 | Gemma 3 4B (IT) | 43.5 | #3831 |
| 7 | Qwen 3.5 0.8B | 15.8 | #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-sequential-detection · How It Works · Data refreshed daily, snapshot 2026-11-13.