VeriCUA verifies computer-using agents by gathering GUI and code evidence
Original title: VeriCUA: Verification as Agentic Exploration for Computer-Using Agents
The framework builds a task rubric, then runs GUI or code checks and rejects unsupported completion claims.
VeriCUA treats verification of computer-using agents as agentic exploration rather than a single vision-language judgment or a generated program. After building a task rubric from available context, it selects GUI or code checks and combines findings under evidence rules. The authors report better agreement with reference labels than baselines, especially offline on hard cases, and stronger online rejection of false completions.
- One procedure covers online and offline checks across desktop, mobile, and web.
- Largest offline gains are reported on hard cases.
- Analyses attribute gains to complementary evidence and test-time exploration.
From the original article
1 min read at arxiv.orgBuilding and scaling computer-using agents (CUAs) across desktop, mobile, and web environments requires reliable task-completion verification for data curation, evaluation, and reinforcement learning. Typical automated approaches prompt vision-language models (VLMs) to judge trajectories or generate programs to inspect application state. However, both rely on partial views of complex digital environments.
By Qiushi Sun · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.