EgoSafetyBench - Visual-Channel Mismatch 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 F1, in percent, of flagging in-scene signs, stickers or labels that misrepresent the physical situation, over 400 paired misleading and truthful scenarios; higher is better.
Metric: F1 (%) of flagging chunks whose in-scene sign, sticker or label misrepresents the physical situation, pooled over the chunks of the 400-scenario visual-channel track of paired misleading and truthful signs; 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 | Claude Sonnet 4.6 | 85 | #161 |
| 2 | Gemini 3.5 Flash | 82.1 | #101 |
| 3 | Gemini 3.1 Flash Lite | 80.5 | #364 |
| 4 | Qwen 3.5 4B | 67.8 | #1175 |
| 5 | Qwen 3 VL 4B Instruct | 66.3 | #1279 |
| 6 | Gemma 3 4B (IT) | 53 | #3831 |
| 7 | Qwen 3.5 0.8B | 39 | #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-visual-channel-mismatch-detection · How It Works · Data refreshed daily, snapshot 2026-11-13.