GeoAgent

A benchmark of 1200 Google Street View locations for evaluating AI vision models' geolocation ability, plus outputs from three families of agents evaluated on it.

Project page: https://geoagent-benchmark.github.io Companion code: https://github.com/sohambuilds/geoagent

Configs (subsets)

Load a specific config with load_dataset("ArkaMukherjee/geoagent", "<config_name>").

• locations (1200 rows) — the benchmark itself: one Street View panorama per row (lat/lng, heading, pitch, population tier, continent/country/city, Street View image id).
• autonomous_agent (25,201 rows) — round-level outputs from agents that freely explore a location (rotate/move/look) before submitting a guess. Covers 7 models x 3 runs each: Claude Haiku 4.5, Gemini 3 Flash, Gemma 3 27B, Geo-R1, GPT-5 Mini, Llama 4 Scout, and a random-guess baseline. Includes the full per-action trace (actions) with each action's reasoning, observations, and token usage.
• multiview_baseline (50,400 rows) — round-level outputs from agents shown a single fixed composite image (4 or 8 stitched viewpoints) with no exploration. Same 7 "models" x 3 runs x 2 view modes (4view, 8view).
• random_walk (20,400 rows) — round-level outputs from agents that take randomized navigation actions (rather than deliberate exploration) before guessing. 6 models x up to 3 runs (Gemini 3 Flash has 2 runs; all others have 3).

Each output config carries model, agent_type, run (1-indexed per model, and per view_mode for multiview_baseline), session_id, and source_file (path in the original repo) alongside the round-level guess, ground truth, distance/points scoring, and accuracy flags.

Known data quirks

• One autonomous_agent random-baseline run (run=1) has 1201 rows instead of 1200 (one duplicate round in the original log).
• random_walk for Gemini 3 Flash only has 2 runs; a third file was a mislabeled duplicate of a Gemma 3 27B run and was dropped during dataset construction.

License

Citation