The Pulse
Researchers Report CUA-Master-27B Scoring 80.2 on OSWorld Verified
CUA-Master-27B scores 80.2 on OSWorld Verified, with results also reported on three other computer-use benchmarks. The framework combines open-world task discovery, synthetic data construction and iterative model optimization.
AI.info Team · Written with AI · fact-checked ·
CUA-Master-27B scores 80.2 on OSWorld Verified, according to a preprint submitted to arXiv on 2 Nov 2026. The paper also reports results on OSWorld 2.0, WeaveBench and Agents’ Last Exam, and describes CUA-Master as a framework for repeatedly improving computer-use agents through new data and model optimization. Its authors are Xinlei Yu and 17 coauthors.
Two loops address data and training limits
The authors identify limited diversity, controllability and reliability in interaction data as constraints on computer-use agent development. They also point to a gap in conventional optimization pipelines: data is generated and used for training, but the process does not adequately feed the model’s evolving weaknesses back into data construction. As a result, additional data may produce uneven improvements while leaving capability gaps in place.
Existing recursive improvement methods can add iterative training, but the paper says they often depend on experiences generated by the model itself. The authors characterize this as closed-world optimization: a system’s exploration remains bounded by its existing capabilities and can reinforce distributional biases, with diminishing returns. CUA-Master is designed to address both the limited feedback in one-way pipelines and the restricted knowledge expansion of self-generated experience.
The framework has two connected loops. At the system level, an outer loop links open-world exploration, data construction and model optimization. It incorporates open-world corpus and uses weaknesses observed in the evolving model to guide later exploration, data generation and optimization. At the data level, an inner loop combines open-world seeding with autocatalytic propagation.
Videos provide seeds for task construction
For open-world seeding, the system mines authentic workflows from instructional videos and uses them as seed tasks. The paper describes hierarchical retrieval and filtering for finding diverse tasks, followed by visual-semantic dual-path trajectory reconstruction. The aim is to use workflows from outside the agent’s own generated experience as a starting point for data construction.
In the propagation stage, existing seeds are abstracted into reusable primitives and composed into new tasks. The authors say this approach allows control over task complexity and application scope, and supports synthesis of diverse, complex and verifiable tasks. They identify homogeneity in tasks drawn from restricted sources as one obstacle to broader ability coverage. They also note that instruction rewriting and template-based generation offer limited control over operational abilities, task complexity and dependencies, particularly for long-horizon work across applications.
Retry mechanism targets rollout failures
Collecting verified trajectories across real environments can become less reliable as tasks grow more complex, the paper says. CUA-Master uses a feedback-guided snapshot-rollback retry mechanism during trajectory collection. The method preserves valid execution prefixes and recovers from localized failures without restarting the entire rollout, according to the paper’s introduction.
The authors report that the framework improves the quality and efficiency of data synthesis and shows continuous gains across recursive iterations. They also describe improvements in the diversity and coverage of the data distribution, as well as in synthesis reliability and controllability.
Results cover four benchmarks
CUA-Master-27B records 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents’ Last Exam. The authors present these as results across GUI-only and GUI-CLI hybrid benchmarks and describe the performance as competitive and approaching markedly larger or closed-source models. The paper identifies the system as a 27-billion-parameter model and frames the benchmark results alongside its proposed recursive data and optimization process.