ERGeoBench - Embodied GLS: leaderboard

Embodied geo-localization on ERGeoBench (2026): a model locates a real-world street scene as an embodied agent that turns, tilts and zooms to gather views, scored by the Geo-localization Score, the mean of hierarchical street, city and country accuracy, distance-threshold hit rates and a log-scaled median error. Scale 0 to 100; higher is better.

Metric: Geo-localization Score, GLS (0-100): mean of the street, city and country accuracy, the hit rate within 1, 25, 200, 750 and 2,500 km, and a log-scaled median-error score, over the 2,207 street-level panoramas, located as an embodied agent that chooses yaw, pitch and zoom actions to gather sequential egocentric views; unified structured prompt with egocentric observations, action history and in-context examples, temperature 0.1, first valid JSON object parsed (invalid outputs count as failures); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro69.02#276
2Gemini 3 Flash68.81#175
3GPT-4o67.55#751
4Gemini 2.0 Flash67.13#723
5InternVL3-8B43.91#1782
6Qwen 3 VL 8B43.87#917
7Qwen-VL-Max42.54#1400
8InternVL2.5-8B33.46#2370
9Qwen 2.5 VL 7B33.38#1862

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=ergeobench-embodied-gls · How It Works · Data refreshed daily, snapshot 2026-11-21.