Interaction traces—records of how people and software agents navigate interfaces—offer a window into usability problems, expertise levels, and how closely AI behavior matches human behavior. But comparing such traces is not straightforward, because they are captured as low-level, interface-specific events that are hard to align across different users or systems.
A new arXiv preprint, titled "Comparing Computer Use Behavior Across Multiple Levels of Abstraction," takes on this problem. The abstract argues that comparing traces can surface usability barriers, differences in expertise, and human–agent similarity, but notes that the raw format of traces makes such comparisons difficult.
The work appears to propose a multi-level abstraction approach, as suggested by the title, though the abstract itself is truncated and does not detail the method or results. The key contribution, based on the available text, is framing trace comparison as a problem of choosing the right level of abstraction—rather than relying solely on raw, interface-specific events.