GST-Bench - Global Position (Semantic): 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. The model marks where an object it saw earlier, named by its category, lies on a top-down image of the scene; mean accuracy within five pixel thresholds, in percent; higher is better.
Metric: Mean accuracy (%; predicted top-down map position of a previously seen object within 100 to 300 pixels, averaged over five thresholds; target named by its category; 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 February 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 43.67 | #137 |
| 2 | GPT-5 | 43.23 | #167 |
| 3 | Seed 1.8 | 36.07 | #242 |
| 4 | Gemini 2.5 Pro | 32.4 | #266 |
| 5 | GPT-4o | 25.24 | #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-global-position-semantic · How It Works · Data refreshed daily, snapshot 2026-11-13.