GeoScope Study - Zero-shot - RC: leaderboard
GeoScope v1,RC,80testquestions,native accuracy percent,Zero-shot. Seeded20%development/80%test,configuration frozen. Direct wholeimage4Mpx,firsttoken optionlogit; numeric-only response options form separate condition. GeoScope uses fixedQwen3-4Bsynthesizer,greedy,horizon5,depth5 withnamedperceptionmodel. Exact rerun protocol retained,not a framework-named checkpoint. No constant-answer test-tie upperbound or cited external scores.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3-VL-8B-Instruct (GeoScope Study Run: method=Zero-shot) | 47.5 | |
| 2 | Qwen2.5-VL-7B-Instruct (GeoScope Study Run: method=Zero-shot) | 41.25 | |
| 3 | Qwen3-VL-4B-Instruct (GeoScope Study Run: method=Zero-shot) | 40 | |
| 4 | llava-onevision-qwen2-7b-ov-hf (GeoScope Study Run: method=Zero-shot) | 35 |
Interactive version: theaggregate.ai/benchmark?slug=geoscope-study-zero-shot-rc · How It Works · Data refreshed daily, snapshot 2026-11-12.