Paper review · November 4, 2026
CUA-Master: scaling computer-use agents through open-world recursive improvement
CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements
The authors built a dual-loop data flywheel that mines real tasks from tutorial videos and composes them into reusable Primitives, and achieved a score of 80.2 on OSWorld Verified with a 27B model.
Problem
Interaction data for training computer-use agents (CUAs) lacks diversity, controllability and reliability. Existing pipelines are one-directional, creating data and then training, so the model's weaknesses are not reflected in the next round of data generation. Recursive improvement approaches that rely on self-generated experience stay confined to what the model already knows, so gains gradually shrink.
Method
The structure is a dual loop. The outer loop (the world-data-model flywheel) keeps bringing in external knowledge and uses the model's weaknesses to direct the next round of data exploration and training. The inner data flywheel has two parts. (1) Open-world seeding: tutorial videos from YouTube, TikTok and Bilibili are collected and filtered based on a taxonomy of 8 major categories, 36 subfields and about 650 apps. After tasks are segmented from the videos, Trajectories are reconstructed using both action descriptions from speech transcripts (semantic) and screen changes (visual). (2) Autocatalytic propagation: seed tasks are abstracted into Primitives made up of an app, a capability, preconditions, environment requirements and a verifier. New verifiable tasks are then synthesized, with complexity controlled by the number of Primitives and the number of apps involved. When collecting Rollouts, feedback-based snapshot rollback retry keeps the valid earlier part of the execution and retries from the point of failure. The action space combines GUI and CLI. The infrastructure supports more than 100 environments and more than 2,000 concurrent VMs.
Results
- CUA-Master-27B scored 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents' Last Exam.
- The authors reported that this performance approaches that of much larger or proprietary models, with high execution efficiency.
- The authors reported that the flywheel improved the quality and efficiency of data synthesis and the diversity of the data distribution, and that performance kept rising with each recursive iteration.
Limitations
The authors note that Trajectories reconstructed from videos may contain annotation noise, and that there is no guarantee these Trajectories match actual execution in the target environment. Most task segmentation, annotation and verification depends on a stronger annotation model. The infrastructure needs thousands of VMs, so reproduction is costly. Within the provided text, per-iteration figures and detailed comparisons with baseline models could not be found.
Why it matters
For computer-use agents used in business automation, the bottleneck is often data rather than the model. The pipeline design, which uses public tutorial videos as a source of real workflows and keeps adding data to target model weaknesses, is a useful reference for teams building their own agent data.
Key terms
- CUA(Computer-use Agent)
- An agent that looks at screenshots and performs tasks by operating real computer software through mouse and keyboard input or shell commands.
- Primitive
- A reusable, app-level unit of capability extracted from seed tasks. Because it includes execution conditions and a verification method, units can be combined to create new tasks.
- Snapshot rollback retry
- A method that, when a failure occurs during a Rollout, does not start over but restores a saved earlier environment state and retries only the failed segment.
Original
Authors: Xinlei Yu, Gen Li, Ji Zhao, Zhangquan Chen, Congcong Wang, Shuai Dong, Zelin Peng, Yuqi Xu, et al. · Posted 2026-11-02 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.