NaturalBench: leaderboard

Vision-language benchmark with 10,000 human-verified natural adversarial VQA samples, pairing each question with two images to force visual grounding rather than language priors.

Source: linzhiqiu.github.io.

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