CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements
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.
Figures & tables
| Method | OSWorld Verified † | OSWorld 2.0 † | WeaveBench ‡ | Agents’ Last Exam ‡ | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Step | Token | Time | Score | Step | Token | Time | Score | Step | Token | Time | Score | Step | Token | Time | Score | |
| Generalist | ||||||||||||||||
| Qwen-3.5-27B [ 35 ] | 32.2 | 396K | 7m | 54.6 | 224.0 | 1.9M | 1h48m | 0.8 | 75.6 | 1.0M | 41m | 13.5 | 165.3 | 4.1M | 1h25m | 1.3 |
| Qwen-3.5-397B-A17B [ 35 ] | 32.8 | 412K | 11m | 61.0 | 208.4 | 4.3M | 2h2m | 10.1 | 52.4 | 2.4M | 1h4m | 19.6 | 133.7 | 9.3M | 1h53m | 5.3 |
| Qwen-3.6-27B [ 36 ] | 34.4 | 408K | 7m | 61.4 | 190.6 | 1.8M | 1h31m | 8.9 | 86.9 | 1.3M | 47m | 16.5 | 128.8 | 3.7M | 1h11m | 5.3 |
| Qwen-3.8-27B [ 37 ] | 28.1 | 330K | 9m | 81.2 | 179.7 | 3.7M | 1h28m | 36.5 | 52.6 | 3.4M | 57m | 14.1 | 187.1 | 9.4M | 1h48m | 19.1 |
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Action Type | Definition |
|---|---|
| computer.click | Performs a single mouse click at a specified position. |
| computer.double_click | Performs a double click for activation or selection. |
| computer.triple_click | Performs a triple click for text or row selection. |
| computer.move | Moves the pointer to a specified screen position. |
| computer.drag | Drags the pointer between two screen positions. |
| computer.scroll | Performs horizontal or vertical scrolling. |
| Type | Function | Prompt |
|---|---|---|
| Open-world Seeding | ||
| Score Evaluator | Video suitability scoring and filtering | Fig. 11 |
| Task Generator | Seed task mining and temporal segmentation | Fig. 12 |
| Action Rechecker | Visual-semantic action boundary refinement | Fig. 13 |
| Action Annotator | Atomic GUI action annotation | Fig. 14 |
| Trajectory Verifier | Trajectory-level verification and filtering | Fig. 15 |
| Domain | Sub Domain | Applications |
|---|---|---|
| Business &Administration | Career | ADP Workforce Now, Indeed, Glassdoor, Greenhouse, Workable, Simplify Jobs, We Work Remotely, Idealist, CharityVillage, LinkedIn, ZipRecruiter, BambooHR, Workday, Upwork, Fiverr |
| Education | Moodle, OpenBoard, GeoGebra, Google Classroom, Canvas LMS, Blackboard, Kahoot, Quizlet, Anki, Coursera, Khan Academy, Duolingo, edX, LeetCode, Outschool, RateMyProfessors, Udemy, Skillshare, Codecademy, Grammarly, Wolfram Alpha, Photomath, Chegg, Schoology, Pluralsight, DataCamp, HackerRank, freeCodeCamp | |
| Finance | QuickBooks, Xero, GnuCash, FreshBooks, Mint, QuantLib, Pro Expense, Stocks, PayPal, Revolut, YNAB, Zoho Books, Quicken, TurboTax, Robinhood, Coinbase | |
| Sales & Marketing | Salesforce, SuiteCRM, G2, Capterra, Trustpilot, Mailchimp, Freshdesk, Twenty, HubSpot, Zoho CRM, Pipedrive, Zendesk, Constant Contact, Klaviyo, Hootsuite | |
| Creative Production | Audio & Music | MuseScore, Dorico 6, Cubase, Audacity, FL Studio, Ableton Live, Logic Pro, Pro Tools, Ardour, LMMS, GarageBand, Adobe Audition, Avid Sibelius, Guitar Pro, REAPER, Retro Music, Audio Recorder, Music, Voice Memos |
| Graphic Design | GIMP, Inkscape, Microsoft Paint, darktable, Krita, Scribus, Adobe Photoshop, Adobe Illustrator, Adobe InDesign, Affinity Photo, Affinity Designer, CorelDRAW, Figma, Canva, RawTherapee, Adobe Lightroom Classic, Font Book, PhotoPrism |
Computer-Use Agents
Computer-use agents perceive and operate graphical interfaces and software tools to complete tasks across websites, apps, and desktops.
See all 171 papers