LongEgoRefer - Temporal Grounding: leaderboard
LongEgoRefer (2026) builds long-form egocentric referring expressions on Ego4D: 1,498 detailed descriptions of object interactions in untrimmed videos averaging 45 minutes, where the target object is visible in about 1.3 percent of frames. This board scores the vision-language model finding when the described interaction happens, as mean temporal IoU between the predicted and annotated interval in percent; higher is better.
Metric: Mean temporal IoU (%; predicted versus annotated time interval of the referred object interaction, 1,498 expressions over untrimmed Ego4D videos averaging 45 minutes; zero-shot VLM prompted for start and end times). Source: arxiv.org. Saturation forecast: Around October 2027. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 16.19 | #167 |
| 2 | Gemini 2.5 Pro | 10.73 | #266 |
| 3 | GPT-4o | 8.71 | #741 |
| 4 | Gemini 2.5 Flash | 4.49 | #453 |
| 5 | Gemini 2.0 Flash | 2.29 | #704 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longegorefer-temporal-grounding · How It Works · Data refreshed daily, snapshot 2026-11-13.