Benchmark
VisualWebArena
Test visually guided web tasks, including product selection and shopping actions.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
910 tasks spanning shopping, classifieds, and a forum; tasks require visual understanding, and some use image-based requests.
Methodology
Agents act in reproducible websites. Execution-based evaluators check answers and final states; model, screenshots/captions, interaction limits, and site subsets affect success rates.