VeriCUA: Learning Proactive Verification for Reliable GUI Agent Rewards

VeriCUA trains a Judge to select verification checks and stopping decisions while a frozen Probe inspects the post-task GUI. Single-attempt reinforcement learning balances verdict correctness and Probe use.

AI-assisted editorial review against a primary source. This is a catalog annotation, not an independent reproduction or endorsement.

What is this resource?

Resource type

Research papers

Environment

desktop, linux, web, android

Topics

reward-models, reinforcement-learning, evaluation, efficiency

Record publication date

2026-10-19

Evidence retrieved

2026-10-20T08:27:26.236797+00:00

Resource ID

arxiv:2610.26469

Recorded paper version

arXiv v1

Version updated: 2026-10-19T14:17:33Z.

This identifies the retained arXiv metadata snapshot; newly observed changes require separate review.

Authors

Zhizhou Wang, Yaxi Lu, Zhong Zhang, Zikun Hu, Xin Cong, Yankai Lin, Fangming Liu

Cite this paper

Download BibTeX · Download CSL JSON

Citations preserve the selected source edition, its authors and recorded date precision. Unrecorded venues and identifiers are omitted.

Explore related topics

Scope and limitations

Reviewed the complete abstract and selected v1 method/evaluation sections. The Judge is trained on OSWorld and frozen for MobileWorld/WebArena transfer; downstream Actor RL is separate. Evaluation uses screened terminal outcomes, and reported cost counts Probe tokens. No independent reproduction.

Primary evidence and verification

Source retrieval is separate from experiment reproduction. Use the original publication or publisher page when citing claims about the resource.

1. arxiv api

   Retrieved 2026-10-20T05:52:00Z. Supports: name, authors, published_at, updated_at, identifiers.arxiv, details.arxiv_snapshot.

   Content fingerprint

   af310824eea97aac8209023ff53bad7bea77778816586b10bb94110148347af9
2. editorial primary review

   Retrieved 2026-10-20T08:27:26.236797+00:00. Supports: name, kind, summary, topics, scope.

   Content fingerprint

   80bed80786f939973d0d13b4f892300fee02a6110e326247131fc83d257796de
3. editorial primary review

   Retrieved 2026-10-20T08:28:59.073204+00:00. Supports: summary, environments, topics, scope, notes.

   Content fingerprint

   f13f0225c721d5aa430bb8aea1d1e7a7da5973760c4c3b82f26ba648aade82ba

Suggest a correction

Open a source-backed correction. Include this record ID and the original source. Read our editorial policy and coverage boundaries.