VisualWebBench: leaderboard

1.5K human-curated items from 139 websites across seven tasks (captioning, WebQA, heading and element OCR, element grounding, action prediction and grounding), from a CMU-led team (2024).

Source: arxiv.org.

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