PerceptionTest: leaderboard

11,600 videos averaging 23 seconds with multiple-choice and grounded video QA, object and point tracks, and action and sound segments, from DeepMind (2023); humans 91.4% vs models 46.2% on video QA.

Source: arxiv.org.

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