CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements
Xinlei Yu, Gen Li, Ji Zhao, Zhangquan Chen, Congcong Wang, Shuai Dong, Zelin Peng, Yuqi Xu and 10 others
MMLab, CUHK · THU · HKUST(GZ) · Explore Academy, JD
Published November 2, 2026 · on the feed November 3, 2026 · ▲ 47 · arXiv 2611.03126
The 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.
The Brief
2:34A short podcast summary of the paper. Written and voiced by AI.
Transcript
CUA-Master. Scaling Computer-use Agents through Open-World Recursive Improvements. By Xinlei Yu and seventeen others, from MMLab at the Chinese University of Hong Kong, Tsinghua University, and two other institutions. Published on the second of November.
Computer-use agents have advanced, but their development is held back by the interaction data they learn from, which is limited in diversity, controllability and reliability. Two further problems compound this. Conventional optimization pipelines lack iterative feedback. And existing recursive improvement approaches often rely on experiences the model generated itself, which limits how far its abilities can expand and yields diminishing returns. A model that learns only from its own experience keeps meeting the same kinds of tasks it can already handle. The data problem and the feedback problem are linked, and the framework is built to address them together.
CUA-Master answers with a nested, double-loop recursive improvement framework. The outer loop works at the system level: a world, data and model flywheel that keeps bringing in open-world corpus and uses the model's evolving weaknesses to steer the next round of data construction and optimization. Each pass through the outer loop, then, is shaped by what the model got wrong in the previous one. The inner loop works at the data level, and it is where new tasks come from.
That inner flywheel couples two parts. Open-world seeding mines authentic workflows from instructional videos as seed tasks, and reconstructs trajectories along two paths, one visual and one semantic. Autocatalytic propagation then turns those seeds into reusable primitives, from which diverse, complex and verifiable tasks can be synthesized in a controllable way. The seeds bring authenticity, real workflows that people actually perform. The primitives bring scale and control, since one seed can be recombined into many tasks, each of which can be checked.
Tasks still have to be carried out to produce training trajectories, and that collection must be efficient and reliable. For this the framework adds a feedback-guided snapshot-rollback retry mechanism, used in various real environments.
The resulting model, CUA-Master-27B, is reported as competitive on both GUI-only and GUI-CLI hybrid benchmarks. It scores eighty point two on OSWorld Verified, thirty-six point nine on OSWorld two point zero, forty-three point nine on WeaveBench, and twenty-five on Agents' Last Exam. The framework also substantially improves the quality and efficiency of data synthesis, and keeps gaining across recursive iterations.
The central bet is on where new ability comes from. Instead of a model feeding on its own experience, CUA-Master keeps pulling in the open world, from real workflows in instructional videos, and lets its current weaknesses decide what to learn next.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
An agent that only learns from itself soon stops learning. Where does fresh material come from?
You are training an agent that operates computers: clicking through apps and typing commands. Human-annotated trajectories are costly and narrow, and synthetic tasks drawn from similar sources look alike.
The road ahead
- 1 · The problem ~6 min Agents that operate computers need varied, reliable training data, and pipelines that build data once, or reuse the model's own experience, stop improving.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Two nested loops: an outer one brings in outside material and targets model weaknesses each round; an inner one mines, multiplies and reliably collects tasks.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Mine videos into seeds, rebuild their trajectories, turn seeds into primitives, compose verifiable tasks, collect with rollback retries, then train and let weaknesses steer.
- 6 · The experiment ~6 min You’ll guess: What score did the 27B CUA-Master model reach on OSWorld Verified?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.