Google and the University of Hong Kong are introducing VeriCUA, a verifier that checks the actual state of applications and files before accepting an AI agent’s claim that a task is complete. The system builds a task-specific rubric and then searches for evidence through graphical or code-based checks. On the OSWorld benchmark, the false-positive rate, meaning tasks approved without actually being completed, fell from 27.9% to 13.9%. On ScienceBoard, it dropped from 24.8% to 5.4%, while AndroidWorld achieved 90.5% accuracy with a 7.7% false-positive rate, compared with 82.4% and 16.9% for the baseline. Online tests rejected approximately 80% of false completion claims, and decision accuracy exceeded 95% when the verifier was confident in its judgment.
Source: Read the original article
