CUA-Recovery is a benchmark that evaluates error detection and recovery for long‑horizon computer‑use agents (CUAs). As CUAs operate across multiple applications, undetected errors can silently corrupt persistent state and trigger cascading failures. Reliable error awareness and recovery have become critical bottlenecks for building robust CUAs. Existing benchmarks mainly report end‑to‑end task success, offering little insight into an agent’s ability to spot and fix errors during execution. The few recovery‑focused benchmarks rely on manually injected faults, which may not reflect the error types or horizons produced by real policies. To address this, we build CUA‑Recovery on top of the ReRail framework. ReRail synthesizes long‑horizon tasks and trajectories in a persistent desktop environment, reuses MyPCBench tasks to compose new workflows with executable ground truth, and gathers verifier‑confirmed failures from multiple policies, turning them into verified recovery trajectories. CUA‑Recovery includes a human‑verified test suite of 2,250 erroneous states covering 100 tasks and a training suite of 8.6K trajectories covering 377 tasks. Evaluation shows that six agents across six takeover depths achieve at most 44.0% Error Awareness Rate (EAR) and 25.5% Pass@3. After fine‑tuning Qwen3.5‑35B‑A3B on the training suite, average EAR rises from 19.9% to 35.2% and Rubric Score from 5.0% to 16.1%. The model also gains 15.3 points on MyPCBench and 10.8 points on Odysseys for Pass@3, indicating that the training trajectories improve both recovery and clean‑start task completion. Code is released at https://github.com/xu-hu-2002/CUA-Recovery
Review