A new paper introduces VeriCUA, a verifier that inspects app data instead of trusting an AI agent's claim that a task is done
AI agents have a habit familiar to anyone who has managed a summer intern. They say the job is done, and sometimes it isn’t.
A new research paper from Google and the University of Hong Kong takes direct aim at that problem. It introduces VeriCUA, a framework that checks the actual state of apps and files before accepting an agent’s claim that a task is complete.
The core idea is simple. An agent saying “finished” counts as a claim, not as proof.
What VeriCUA actually does
The paper focuses on computer-using agents, or CUAs. These are AI systems that operate software the way a person would, by clicking through interfaces, typing, and moving between apps.
VeriCUA takes what the authors describe as an agentic approach to verification. It first builds a rubric tailored to the specific task, drawing on the task’s context.
Then it goes looking for evidence. The verifier runs GUI-based or code-based checks, such as reading files and inspecting system state, before it issues a verdict.
The paper, listed on arXiv as 2610.40578, was submitted on October 29, 2026. Its authors include Qiushi Sun and Pritish Kamath.
The numbers behind the pitch
The researchers tested VeriCUA across desktop, mobile, and web environments. According to the paper, it beat baseline systems across the board, with the biggest gains showing up in harder scenarios.
The headline metric is the false-positive rate. In this context, a false positive means the verifier signs off on a task the agent did not actually complete.
On the OSWorld benchmark, VeriCUA cut that rate from 27.9% to 13.9%. That is roughly half the bogus approvals of the comparison method.
The ScienceBoard results were more dramatic. The false-positive rate fell from 24.8% to 5.4%.
On AndroidWorld, a mobile benchmark, VeriCUA posted 90.5% accuracy with a false-positive rate of 7.7%. The baseline, an auto-generated script verifier, recorded 82.4% accuracy and a 16.9% false-positive rate.
The paper also reports that when VeriCUA is confident in its judgment, decision accuracy exceeds 95%. The system also showed higher agreement with reference labels than the baselines did.
In online tests, VeriCUA rejected approximately 80% of false completion claims.
Why self-reports are a problem
The broader issue the paper raises is inflated success rates. When verification relies on weak checks, agents can appear more capable than they really are.
The researchers position their dynamic verification approach as a counter to that inflation. If the checker only looks at surface signals, a confident-sounding agent can pass tests it should fail.
The paper drew discussion in the tech community shortly after it appeared on arXiv.
What this means for AI agents
For teams building computer-using agents, the takeaway is uncomfortable but useful. If older verification methods let through false positives at rates like 27.9% or 24.8%, some reported success numbers may be flattering the underlying systems.
The approximately 80% rejection rate for false completion claims in online testing suggests that layer can catch a meaningful share of those quiet failures.
The confidence signal is also notable. Decision accuracy above 95% when the system is confident means builders could potentially route uncertain cases to humans and let the verifier handle the clearer ones.
For now, the research makes a straightforward argument. When an AI agent says it finished the job, somebody should check. Google and the University of Hong Kong are proposing that the somebody be another agent, one that actually opens the files.
