What it measures
Self-hosted websites support realistic multi-step web tasks with checks on whether the intended state was reached.
How to read the result
Measure the actual outcome of browser actions rather than relying on the assistant’s completion message.
Limits and comparison traps
Website snapshot, browser interface, observation mode, and task revision matter. A visual agent and a DOM-based agent see different evidence.
Published evaluation
Task success under executable final-state evaluation; chain-of-thought and unachievable-task hints are source configuration variables.
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
Task success under executable final-state evaluation; chain-of-thought and unachievable-task hints are source configuration variables.
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
- WebArena: A Realistic Web Environment for Building Autonomous Agents — published result table ↗
All 12 authors
Zhou, Shuyan, Xu, Frank F., Zhu, Hao, Zhou, Xuhui, Lo, Robert, Sridhar, Abishek, Cheng, Xianyi, Ou, Tianyue, Bisk, Yonatan, Fried, Daniel, Alon, Uri, Neubig, Graham
- Primary benchmark documentation or model evaluation report ↗
Reviewed 2026-10-27. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.