Preprint finds a cheaper way to gauge whether AI agents will click correctly

Researchers report that a new method estimates whether an AI agent will click the right place on a screen in a single processing pass. Their preprint says it matched or beat methods that require multiple samples in its primary comparisons on ScreenSpot-Pro and ScreenSpot-v2. The method gives more weight to uncertainty in the digits that most affect a click's position. Its reported gains apply to the fixed-scale agents and benchmarks tested; the abstract does not establish how it performs in deployed agents.