MultiGlobeQA (KG Retrieval Agent): leaderboard

Metric: Exact-match accuracy (%; Tier 3a: a CodeAgent writing and executing Python with spatial knowledge-graph retrieval tools, at most 10 steps; exact match with answer-type tolerances (geodesic or relative tolerance, set equality, polygon IoU of at least 0.5) over the 4,979 true-premise questions of the English text small split, refusals and non-answers counted incorrect; greedy decoding, output constrained to the answer-type JSON schema; mean of three seeds). Source: arxiv.org. Saturation forecast: Around 2029. 4 models tracked.

Top models

#ModelScore
1Qwen 3.5 27B FP844.3
2Gemini 3 Flash (Preview)29.5
3Gemma 3 27B (IT)23.9

Interactive version: theaggregate.ai/benchmark?slug=multiglobeqa-kg-retrieval-agent · How It Works · Data refreshed daily, snapshot 2026-09-26.