The costliest ingredient in training computer-use agents is not the model — it is the interaction data showing how a human actually operates software. The mainstream recipe lets the model try, fail, and bank its own experience; the literature calls these self-generated experiences. In plain terms, the model mines itself: the data can only contain what the model already knows, and returns diminish with every loop. CUA-Master, posted to arXiv on Nov 2, flips the source — instructional videos already on the internet are massive libraries of real human demonstrations. A day after release it sat at #2 on Hugging Face's paper-of-the-day board with 90 upvotes.
Two nested flywheels
The framework is a nested double loop. The outer loop is a world-data-model flywheel: open-world corpus keeps flowing in, and the model's current weaknesses decide what data gets built and which abilities get optimized next round. The inner loop owns the data itself. One side mines authentic workflows from instructional videos as seed tasks and reconstructs trajectories through visual-semantic dual paths, intersecting semantic triggers from transcripts with frame-change detection into a boundary structure. The other side runs autocatalytic propagation: seed tasks are abstracted into reusable ability primitives, then recomposed under explicit complexity control — K primitives × N applications — into controllable, verifiable new tasks. Trajectory collection adds a detail worth copying: feedback-guided snapshot-rollback retries, so a derailed step rolls the environment back to the last snapshot and tries again instead of voiding the whole trajectory.
Per the GitHub repo's configuration, seed harvesting spans 8 domains, 36 subdomains, and 650+ apps, and every video passes a 0-5 CUA-suitability score before entering the pipeline.
The 27B scoreboard
CUA-Master-27B reports scores across GUI-only and GUI-CLI hybrid benchmarks: 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam. The abstract's own word is "competitive" — no leaderboard sweep claimed. The gap between 80.2 and 36.9 is the real story: on harder, newer benchmarks the whole field still has a long way to climb. The paper also reports continuous gains across recursive iterations and improved quality and efficiency in data synthesis.
The open-source delivery is solid: code, configs, prompts, and VM image management all live in the GitHub repo; training data converts to ms-swift SFT format; RL uses task-level verifiers as outcome rewards; and the README links the CUAMaster-27B model on Hugging Face. The repo has gathered 86 stars within days.
The real signal: switching the data supply
The industry value here is not any single score — it is switching the CUA data supply from model self-play to open-world corpus. Self-generation caps out at the model's current ability boundary; instructional videos, where real humans operate real software, cap out at all the operational knowledge humanity has ever published. Pair that with verifiable task synthesis and data quality stops being vibes — the verifiers double as RL reward sources.
For teams building agent data, this is a blueprint you can copy module by module: mine seeds from video, compose tasks from primitives, save compute with snapshot rollback. For anyone watching the space, the yardstick is simple: whoever spins the open-world-data x recursive-improvement flywheel first will be the first to crash CUA's data costs.
Reference: arXiv:2611.03126; GitHub: xlyu0106/CUA-Master.