Tuesday, 3 November 2026 · 6 min read
CUA-Master improves computer-use agents via nested recursive data and model refinement
CUA-Master introduces a double-loop recursive framework combining open-world data mining from instructional videos with primitive-based task synthesis and feedback-driven rollout retries, achieving strong performance on GUI and CLI benchmarks with a 27B parameter model.
Based on the paper CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements by Xinlei Yu, Gen Li, Ji Zhao, Zhangquan Chen and 14 others
Challenges in developing capable computer-use agents
Computer-use agents (CUAs) automate workflows by interacting with software through GUIs and CLIs, but building agents that reliably handle diverse, long-horizon, cross-application tasks is difficult.
A key bottleneck is the lack of iterative feedback between data construction and model training; existing pipelines generate data then train models without leveraging model weaknesses to guide data collection.
Recursive improvement methods often rely on self-generated data, limiting exploration to the model's existing knowledge and causing diminishing returns.
Additionally, large-scale, high-quality interaction data is scarce due to annotation costs, and synthetic data methods struggle with limited diversity, controllability, and rollout reliability, constraining agent capabilities.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Two figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short