LegibilityBench: A Human-Centered Benchmark for Behavior Legibility in Computer-Use Agents

Existing computer-use agent (CUA) benchmarks primarily evaluate task success and efficiency, but task completion alone does not make agents suitable collaborators if their behavior is difficult for humans to understand, verify, or recover from. We propose LegibilityBench, a human-centered CUA benchmark for behavior legibility: how understandable and actionable CUA behavior is to human observers. LegibilityBench evaluates everyday computer tasks that vary in horizon length, interface complexity, goal ambiguity, and action visibility. Human evaluators observe recorded or paused agent trajectories and complete probes spanning recall, goal inference, next-action prediction, error diagnosis, and task handoff from the agent's paused state. Beyond agent task success, the benchmark measures human accuracy, confidence, cognitive load in understanding agent behavior as well as handoff success. By comparing these measures across models, tasks, and behaviors, LegibilityBench tests whether more capable CUAs are also more overseeable, and provides a foundation for improving CUA legibility and developing automated proxies for human legibility evaluation.