# CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements

**Authors:** Xinlei Yu, Gen Li, Ji Zhao, Zhangquan Chen, Congcong Wang, Shuai Dong, Zelin Peng, Yuqi Xu, Hongrui Wu, Kaiwen Tuo, Cheng Yang, Kaituo Feng, Yifan Wu, Jiayi Zhang, Dongming Wu, Wenqi Shao, Jiaqi Wang, Xiangyu Yue  
**Lab:** CUHK · **Published:** 2 Nov 2026 · **Upvotes:** 92 · **Level:** Advanced  
**Links:** [arXiv](https://arxiv.org/abs/2611.03126) · [PDF](https://arxiv.org/pdf/2611.03126) · [Code](https://github.com/xlyu0106/CUA-Master) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.03126.html)

## Curator's notes

**TL;DR:** CUA-Master introduces a double-loop recursive framework that uses open-world data to continuously improve computer-use agents.

- **What's new:** It combines an outer world-data-model flywheel with an inner seed‑generation and autocatalytic propagation flywheel, adding feedback‑guided rollback for reliable trajectory collection.
- **Key result:** CUA-Master-27B scores 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam.
- **Why it matters:** Builders gain a scalable way to generate diverse, controllable interaction data and improve agent performance without manual data curation.

## Key numbers

- **25.0**: and 25.0 on Agents' Last Exam

## Abstract

Despite recent advances in CUAs, their development remains constrained by the limited diversity, controllability, and reliability of interaction data. Moreover, conventional optimization pipelines lack iterative feedback, while existing recursive improvement approaches often rely on self-generated experiences, limiting ability expansion and yielding diminishing returns. To address these challenges, we introduce CUA-Master, a nested double-loop recursive improvement framework. At the system level, an outer world-data-model flywheel continuously incorporates open-world corpus and leverages evolving model weaknesses to guide subsequent data construction and optimization iterations. At the data level, an inner flywheel couples open-world seeding and autocatalytic propagation. The former mines authentic workflows as seed tasks from instructional videos, and employs visual-semantic dual-path trajectory reconstruction; The latter transforms these seeds into reusable primitives for controllable synthesis of diverse, complex, and verifiable tasks. To enable efficient and reliable trajectory collection, we further introduce a feedback-guided snapshot-rollback retry mechanism in various real environments. Extensive experiments demonstrate that CUA-Master-27B achieves competitive performance across GUI-only and GUI-CLI hybrid benchmarks, scoring 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam. Our framework also substantially improves the quality and efficiency of data synthesis, and exhibits continuous gains across recursive iterations.

## What you could build

1. Create a task generator that mines instructional videos into reusable GUI primitives.
2. Implement a feedback‑guided rollback system for robust agent execution in real environments.
3. Develop a continuous training pipeline that iterates on model weaknesses using the outer flywheel.

## Quiz yourself

What two flywheel loops does CUA-Master employ and what purpose does each serve?

**Answer:** An outer world‑data‑model flywheel that incorporates open‑world corpus to guide data construction, and an inner flywheel that seeds tasks from videos and autocatalytically propagates them into reusable primitives.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.