BLINK: leaderboard

3,807 multiple-choice questions over 7.3K images recasting 14 classic vision tasks (relative depth, correspondence, jigsaw, forensics), from UPenn, UW and AI2 (2024); humans 95.7%, GPT-4V 51.26%.

Source: arxiv.org.

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