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

CUA-Master is a nested double-loop recursive framework that couples open-world seeding and primitive-based task synthesis with iterative model training to improve computer-use agents.

Developing CUAs that reliably execute diverse long-horizon cross-application workflows is bottlenecked by how data construction and model optimization are coupled. Existing pipelines are often unidirectional, so additional data may not address evolving model weaknesses, while recursive improvement approaches rely mainly on self-generated experiences in a closed world and can reinforce biases with diminishing returns. Synthetic data generation also faces intertwined challenges of diversity, controllability, and reliability, and human-annotated trajectories are costly and limited in coverage.

CUA-Master is a nested double-loop recursive framework: an outer world-data-model flywheel uses evolving model weaknesses to guide open-world data construction and optimization, while an inner data flywheel couples open-world seeding and autocatalytic propagation. Open-world seeding retrieves instructional videos using a hierarchical application taxonomy, filters them, mines seed tasks as temporal episodes, and reconstructs trajectories through visual-semantic dual-path triggers plus annotator rechecking. Autocatalytic propagation primitivizes seed tasks into application-level ability primitives and composes compatible primitives into synthetic tasks with controlled complexity K and application count N, producing task-level verifiers. Trajectory collection uses a feedback-guided snapshot-rollback retry mechanism that restores the latest valid snapshot, preserves the stable prefix, and resumes with diagnostic feedback until verification passes or retry budget is exhausted. Training uses SFT iterations from the same base checkpoint, followed by RL in the final iteration on high-quality synthetic tasks with outcome-based verifier rewards.

• CUA-Master-27B scores 80.2, 36.9, 43.9, and 25.0 on OSWorld Verified, OSWorld 2.0, WeaveBench, and Agents' Last Exam, and improves over UI-Mate-27B by 2.6, 12.9, 25.6, and 11.8 across the four benchmarks. §3.2.1
• On OSWorld Verified it achieves 80.2 in only 8 minutes; on OSWorld 2.0 it slightly surpasses Qwen-3.8-27B while reducing runtime by 24%; on WeaveBench it matches GPT-5.6-Luna with 28% less runtime; on Agents' Last Exam it matches Kimi-K3 and Qwen-3.8-Max with 34-36% less time. §3.2.2
• Hybrid execution on WeaveBench scores 43.9 versus GUI-only 6.6 and CLI-only 11.5, gains of 37.3 and 32.4 points, while reducing steps by 23.8% versus GUI-only; on Agents' Last Exam it improves from 5.9 to 25.0 over GUI-only with 26.7% fewer steps, and on OSWorld 2.0 it reduces steps by 43.8%. §3.3.1
• At a configuration with 40 primitives across 3 applications, rollout success is nearly 50% versus 14.6% for the baseline; at the most challenging configuration, success is 18.7% versus 0.0%. §3.3.2
• Successive training iterations improve performance across all four benchmarks, with OSWorld Verified and WeaveBench increasing by over 25 points; the final RL stage further boosts OSWorld 2.0 and Agents' Last Exam by approximately 15 points. §3.3.3
• CUA-Master-27B achieves 56.2, 43.0, and 59.9 on WindowsAgentArena, MacOSWorld, and MacAgentBench. §3.2.3

• The recursive improvement pipeline incurs substantial computational and infrastructure costs, particularly from large-scale environment interactions, trajectory collection, and iterative model optimization; snapshot-rollback still requires large-scale VM deployment support. (§E)
• The upper limit of synthesized data quality relies on capable annotator models, with stronger annotators consistently yielding greater performance gains; this dependence introduces additional inference costs and may constrain accessibility when strong annotators are unavailable. (§E)

• CUA-Master combines an outer world-data-model flywheel with an inner data flywheel coupling open-world seeding and autocatalytic propagation. §2.2
• Open-world seeding mines seed tasks from instructional videos and reconstructs trajectories via a visual-semantic dual-path mechanism. §2.3
• Autocatalytic propagation extracts application-level ability primitives and composes them into synthetic tasks under controlled complexity and application scope. §2.4.1
• The feedback-guided snapshot-rollback retry mechanism restores the latest valid snapshot, preserves the stable prefix, and retains only the successful and most efficient branch. §2.4.2
• The hybrid action space supports GUI and CLI atomic actions, with pointer coordinates normalized to x,y in [0,1000]. §2.1.2
• The environment infrastructure supports over 100 different CUA environments and over 2,000 concurrently running VMs across Linux, Windows, and MacOS. §2.1.3
• Training uses SFT iterations with 63K, 38K, and 14K tasks and corresponding trajectories, totaling about 115K SFT data, followed by about 12K high-quality tasks for RL. §3.1.3
• The training data are organized around 8 domains, 36 subdomains, and over 650 applications, yielding 127K samples including 115K for SFT and 12K for RL. §B