What it measures

910 multimodal web tasks requiring visual inspection of images, memes, product photos, and layout state.

Reading the measurement

The useful outcome is a verified change or deliverable. Inspect the tool interface, initial state, user simulator and policy constraints. An agent can claim completion without reaching the required state, or reach it through an action the protocol forbids.

Published evaluation

End-to-end task success across the three source websites; base model, visual backbone and input representation remain visible.

Where this fits in an evaluation suite

A transcript that says “done” is weaker evidence than the verified final state. Track policy violations, recoveries, and invalid attempts alongside success.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

End-to-end task success across the three source websites; base model, visual backbone and input representation remain visible.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

  1. VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks — published result table ↗

    Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang et al. · 2024

    All 10 authors

    Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, Daniel Fried

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.