VeriCUA learns to inspect GUI terminal state before judging success
Submitted October 19, VeriCUA trains a Judge to choose what to inspect and when to stop, while a frozen Probe gathers evidence in the live terminal state. The authors report that its 4B Judge beats few-shot GPT-5.4 reward agreement with fewer Probe tokens.
Why it matters
Screenshots and agent self-reports can miss file contents, cross-application state and hidden postconditions. Learning an active evidence strategy gives generative-UI completion checks, reward modeling and failure explanations a stronger chain of evidence than the first screenshot.
A closer look
VeriCUA trains its Judge only on OSWorld, then evaluates zero-shot on MobileWorld and WebArena-Verified with a benchmark-specific frozen Probe. Across four independent runs on terminal states stable under executable re-evaluation, the authors report 74.08% macro agreement for the 4B model versus 68.72% for few-shot GPT-5.4, with 47.6%–59.8% fewer Probe tokens; the 1B model reaches 71.62%. Used as a frozen reward source for the same 4B Actor, it raises overall Pass@4 by 3.21 points and by 5.40 points on 74 held-out MultiApp tasks.
Who this is for
Research and platform teams building GUI-agent evaluation, executable postconditions, reward models, active evidence collection and run audit
Original sources
- VeriCUA authors / arXiv and official repository ↗arxiv.org
- Full paper, evaluation and reproducibility statement ↗arxiv.org
- Official Apache-2.0 repository ↗github.com
Revisions & corrections (1)
— Historical backfill discovered from a new GUI-agent research directory; publication date, method, benchmark scope, exact author-reported metrics, public code status and reproducibility caveats were verified in the arXiv record, full paper and official repository.