Reproduce “VeriCUA: Learning Proactive Verification for Reliable GUI Agent Rewards” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhizhou Wang, Yaxi Lu, Zhong Zhang, Zikun Hu, Xin Cong, Yankai Lin, Fangming Liu
VeriCUA trains a Judge to select GUI checks and decide when to stop, using a frozen Probe to inspect unresolved evidence and reinforcement learning to reward verdicts and stopping decisions. On three benchmarks, VeriCUA-4B achieves 74.08% macro-average reward agreement, 5.36 percentage points above few-shot GPT-5.4, with 47.6%–59.8% fewer Probe tokens; Actor gains are limited to reported training setups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.