Google and University of Hong Kong researchers want AI agents to show their work
Researchers from Google and the University of Hong Kong have introduced VeriCUA, a verification framework that checks actual app data instead of trusting AI agents' claims about completed tasks.
The research paper, submitted on October 29, 2026 and listed on arXiv as 2610.40578, introduces VeriCUA as a solution to the common problem of AI agents claiming tasks are complete when they aren't. This framework takes an agentic approach to verification by first building task-specific rubrics before conducting evidence-based checks.
How VeriCUA works
The system focuses on computer-using agents (CUAs) - AI systems that operate software through interfaces similar to human interaction. VeriCUA performs GUI-based or code-based checks, such as reading files and inspecting system states, before issuing its verdict on task completion.
Impressive test results
Testing across desktop, mobile, and web environments showed VeriCUA outperforming baseline systems significantly. On the OSWorld benchmark, it reduced false-positive rates from 27.9% to 13.9%. The improvement was even more dramatic on ScienceBoard, where error rates fell from 24.8% to 5.4%.
In mobile testing on AndroidWorld, VeriCUA achieved 90.5% accuracy with a 7.7% false-positive rate, compared to the baseline's 82.4% accuracy and 16.9% error rate. When confident in its judgment, the system's decision accuracy exceeded 95%.
Online tests revealed that VeriCUA rejected approximately 80% of false completion claims, demonstrating its effectiveness at catching agents that overstate their capabilities.
Addressing inflated success rates
The research highlights how verification methods relying on surface signals can create inflated success rates for AI agents. By implementing dynamic verification, the researchers counter this inflation by requiring actual evidence of task completion rather than accepting self-reports at face value.
If the checker only looks at surface signals, a confident-sounding agent can pass tests it should fail.
Implications for AI development
For teams building computer-using agents, the findings suggest that previously reported success metrics may have been overly optimistic. The high rejection rate for false claims indicates that verification layers can catch a substantial number of failures that might otherwise go undetected.
The system's high confidence accuracy also suggests potential for human-AI collaboration, where uncertain cases could be routed to human review while the verifier handles clearer ones. The research ultimately makes a straightforward argument: when an AI agent claims completion, someone should verify it - and VeriCUA proposes that verification should be another agent that actually examines the files.
Related
