AI news ·
CUA-Master uses nested recursive loops to improve computer-use agents
CUA-Master, a new training framework for computer-use agents, hit 80.2 on OSWorld Verified with its 27B model. It mines real instructional videos and retries failed runs from saved states.
A research team has released CUA-Master, a training framework that improves computer-use agents by repeatedly feeding them data drawn from real-world instructional videos and live software environments. Its largest model, CUA-Master-27B, scored 80.2 on the OSWorld Verified benchmark and 36.9 on OSWorld 2.0, with further scores of 43.9 on WeaveBench and 25.0 on Agents' Last Exam. The results matter because computer-use agents - software that clicks, types, and navigates applications the way a person does - remain hard to train: interaction data is scarce, and agents trained on self-generated practice tend to plateau.
The framework addresses that plateau with what the authors call a nested double-loop recursive improvement system. An outer loop, described as a "world-data-model flywheel," folds in an open-world corpus and uses the model's emerging weaknesses to decide what data to build next. An inner loop turns that raw material into training tasks.
Where the training data comes from
The inner loop has two stages. The first mines authentic workflows from instructional videos and treats them as seed tasks, using what the paper calls visual-semantic dual-path trajectory reconstruction to rebuild the steps a person took on screen. The second stage, "autocatalytic propagation," converts those seeds into reusable primitives that can be recombined into new tasks - diverse, complex, and checkable.
Collecting reliable trajectories in real environments is the other bottleneck. The team introduced a feedback-guided snapshot-rollback retry mechanism, which lets an agent rewind to an earlier state and try again when a run fails, rather than discarding the whole attempt.
The authors frame the work as a response to limits in current methods: "conventional optimization pipelines lack iterative feedback, while existing recursive improvement approaches often rely on self-generated experiences, limiting ability expansion and yielding diminishing returns."
Gains that compound across iterations
CUA-Master was tested on both GUI-only and hybrid GUI-CLI benchmarks. The paper reports that the framework improves the quality and efficiency of data synthesis and shows continuous gains as the recursive iterations stack up, rather than flattening after the first round. For teams weighing how to build agents that operate business software, that detail matters more than any single benchmark score: it suggests the training loop can keep producing useful data instead of running dry.
The broader push toward agents that act inside applications rather than just answer questions is covered in AI Agent Courses, and the underlying model training methods are covered in Generative AI Courses.
Why this matters for your team
If your work involves software that people operate by hand - scheduling systems in healthcare, booking platforms in hospitality, records tools in education, lab software in research - this is the kind of agent training that determines whether an assistant can actually complete a multi-step task or stalls halfway. The practical takeaway: when evaluating computer-use agents, ask how they were trained on real workflows and whether failures were retried from a saved state. Those two choices separate agents that work in a demo from agents that finish the job on a live system.