VisualWebArena

VisualWebArena is the multimodal sibling of WebArena, built by an overlapping Carnegie Mellon team to close a gap the original left open: most of WebArena’s tasks could, in principle, be solved by reading page text alone. VisualWebArena’s 910 tasks, spread across a new classifieds site plus the shopping and forum sites carried over from WebArena, are constructed so that an agent has to look at an image — a product photo, a listing picture — to know what to do, not just parse the surrounding HTML.

The gap between models and people was, if anything, wider than on the text-only benchmark. At release in January 2024, GPT-4V managed 15% and the best model in the paper’s results, GPT-4o, reached under 20%, against an 88.7% human baseline; stripping vision from GPT-4 entirely dropped it to about 7%, showing that the tasks genuinely needed the image rather than being solvable through clever text guessing.

Progress has followed the same trajectory as most agent benchmarks: better scaffolding and newer frontier multimodal models have pushed scores up substantially, with Gemini 2.5 Flash and GPT-5 both reported above 50% on the same public leaderboard by late 2025, though still well short of the human ceiling. VisualWebArena is tracked on the same open spreadsheet as WebArena and remains a standard reference for whether a “computer use” agent actually grounds its actions in what is on screen.