PolyMATH: leaderboard

5,000 visual multiple-choice puzzles in 10 categories (pattern recognition, spatial reasoning, odd one out) from Arizona State and Georgia Tech (2024); exact match, top model 41%, humans 66.6%.

Source: arxiv.org.

Interactive version: theaggregate.ai/benchmark?slug=polymath · How It Works · Data refreshed daily, snapshot 2026-09-05.