GST-Bench: leaderboard
GST-Bench (2026) asks vision-language models about long egocentric videos rendered in a simulator, where answers require a global map of the scene rather than the current frame. Score is the unweighted mean of 12 subtask scores covering object localization, self localization and scene structure, in percent; higher is better.
Metric: Average score (%; unweighted mean of the 12 subtask scores (mean accuracy over tolerance thresholds for direction and position, mean relative accuracy for distance, accuracy for multiple choice); zero-shot, greedy decoding, official prompt templates, over human-verified questions on 6,790 minutes of simulator-rendered egocentric exploration video). Source: arxiv.org. Saturation forecast: Around June 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 42.68 | #137 |
| 2 | Gemini 2.5 Pro | 40.95 | #266 |
| 3 | GPT-5 | 40.85 | #167 |
| 4 | Seed 1.8 | 34.04 | #242 |
| 5 | GPT-4o | 30.33 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gst-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.