S-EMBER - Free-Form Overall Accuracy: leaderboard
Streaming egocentric memory retrieval (S-EMBER, Wang et al., Meta, 2026): 9,448 questions about 388 hours of first-person video from Ray-Ban Meta glasses, each asked at a trigger time and answerable only from what the wearer saw earlier, with open models given 128 uniformly sampled frames and proprietary models their platform limits. Free-form answers are graded by a Gemini 3.1 Flash judge against three reference answers and count unless they contradict the references, extra details allowed; accuracy in percent, higher is better.
Metric: Overall accuracy (%; free-form answer judged correct by Gemini 3.1 Flash, extra details allowed unless they contradict the references) (S-EMBER: 9,448 questions on 3,141 egocentric Ray-Ban Meta videos, answered from the stream up to the question's trigger time; open models at 128 uniformly sampled frames, proprietary models at their platform limits). Source: arxiv.org. Saturation forecast: Around April 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 47.83 | #91 |
| 2 | GPT-5.4 | 37.92 | #121 |
| 3 | Qwen 3 VL 32B Instruct | 30.32 | #582 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=s-ember-free-form-overall-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.