Reinforcement learning for computer-use agents is increasingly turning to proactive reward agents. Instead of relying only on recorded screenshots, these agents interact with the GUI environment after an action has been executed.
That interaction is meant to verify what screenshots cannot reveal about the post-execution state. The reliability of such reward agents is a central concern, according to the arXiv abstract.
The paper introduces VeriCUA, a method for learning proactive verification. The abstract does not detail experimental outcomes, but the stated aim is to make GUI agent rewards more reliable.