Researchers have developed a new benchmark called ComponentBench to evaluate the performance of computer-use agents in modern web interfaces. The benchmark focuses on component-level interactions, such as toggling buttons or selecting options from menus, which are often challenging for current AI models. ComponentBench consists of a library-agnostic ontology of 97 UI components and 2,910 programmatically verified tasks, along with human reference trajectories to evaluate task success and interaction efficiency. The authors demonstrate that changing the observation and action space can significantly impact performance, with some configurations taking up to 3.7 times longer than humans to complete tasks. --- Why it matters: This matters because current AI models struggle with realistic component-centered interactions in modern web interfaces, which are crucial for many applications. ComponentBench provides a standardized way to evaluate and compare the performance of different models on these challenging tasks. Source: https://arxiv.org/abs/2608.18307