A second, independent VeriCUA turns GUI-task acceptance into active evidence gathering with abstention
A Google and University of Hong Kong team submitted a separate project also named VeriCUA on October 29. It inspects an environment or recording before constructing a task rubric, then gathers missing evidence through GUI, accessibility-tree, file and database reads, marking requirements met, unmet or unknown. In online experiments it reduces false positives against baselines on AndroidWorld, OSWorld and ScienceBoard.
Why it matters
An agent's completion claim or final screenshot cannot prove that a file was saved, a setting persisted or every constraint held. Active verification with typed source attributes and abstention when evidence is missing can reduce failed trajectories entering training data or triggering downstream business actions.
A closer look
Online verification maps packages, directories and database schemas before synthesizing cross-surface checkers for hard requirements. The harness exposes allowlisted reads, rejects arguments that could smuggle shell control, and compares digests of persistent targets before and after execution. Fixed rules require attested evidence for success and corroborating refutations for failure; unresolved tasks abstain. With gemini-3-flash-preview as the shared backbone, accuracy and false-positive rate are 90.5% and 7.7% on AndroidWorld, 79.5% and 13.9% on OSWorld, and 86.4% and 5.4% on ScienceBoard, versus VLM-judge false-positive rates of 15.4%, 27.9% and 24.8%. Offline accuracy also leads compared judges across OSReward-Hard, OSReward, WebArena, AndroidWorld and OSWorld. Hybrid GUI plus code reads raise AndroidWorld decision coverage from 45% to 82% without increasing the reported 3/41 false-positive count.
Who this is for
Research and platform teams designing GUI-agent acceptance, training-data filtering, reward models, generative-UI postconditions, audit evidence and high-risk action gates
Original sources
- Google and University of Hong Kong VeriCUA arXiv paper ↗arxiv.org
- Full paper, experiments, cost and limitations ↗arxiv.org
Revisions & corrections (1)
— Publish a new primary-source study that unifies online and offline GUI-agent verification through source-aware active evidence, while explicitly distinguishing it from an unrelated same-name VeriCUA project.