Google and HKU propose VeriCUA to verify computer-use agent task completion

Google and the University of Hong Kong propose VeriCUA, a framework that checks whether a computer-use agent actually finished a task.

According to the paper, the agent's own claim of success does not count as proof. VeriCUA explores the app to build a requirement checklist, then runs screen or code checks, such as reading saved files, until each item is settled.

On the ScienceBoard benchmark, the paper reports a 5.4% false-positive rate for VeriCUA, compared with 24.8% for a vision-language model judge and 21.7% for a generated script verifier. The figures come from the paper as reported by Rohan Paul on X.