CUA-Master-27B: Self-Evolving 27B Agent Rivals GPT-5.6 with 28% Faster Execution

Benchmark Results: 27B Model Approaches Flagship Performance

CUA-Master-27B is a 27B-parameter computer-use agent (CUA) developed by researchers from CUHK MMLab and JD Exploration Research Institute. It employs a dual recursive iteration framework that learns from real-world tutorial videos. Across four core benchmarks, the model delivers competitive scores while significantly reducing execution time and steps.

WeaveBench : 43.9 score vs. GPT-5.6-Luna's 44.0; average execution time dropped from 47 minutes to 34 minutes ( ~28% reduction ).

Agents' Last Exam : 25.0% full-task pass rate, on par with Kimi-K3 and Qwen-3.8-Max; average time reduced by ~34% and ~36% respectively.

Cross-platform : WindowsAgentArena 56.2, MacOSWorld 43.0, MacAgentBench 59.9.

The model achieves the Pareto frontier on the score–execution trade-off across all four benchmarks.

Dual Recursive Self-Improvement Framework

The system builds a scalable data flywheel that connects open-world knowledge acquisition, controllable task construction, real-environment execution, and model training.

Inner Loop: From Real Workflows to Capability Primitives

High-quality data originates from public software tutorial videos covering 8 domains, 36 subdomains, and 650+ software applications . Videos are filtered and segmented; a dual-path reconstruction using semantic cues and visual changes converts unstructured demonstrations into reusable seed tasks . These tasks are further decomposed into capability primitives containing preconditions, environment requirements, and verification conditions. Primitives are then recombined by complexity and software scope to generate new executable, verifiable tasks.

The inner-loop flywheel:

Real workflows → Seed tasks → Reusable capability primitives → Cross-application complex tasks → Collected execution trajectories → Next iteration

Outer Loop: Model Feedback Guides Data Collection

The outer loop closes the cycle between model training and data acquisition: weak model performance triggers targeted exploration of relevant software, workflows, and task difficulties.

Open-world corpus collection → CUA data construction → Model training improvement → Model feedback → Next iteration

This dual-layer recursive improvement framework uses three SFT stages totaling ~115K samples and 12K high-quality RL samples.

Hybrid GUI-CLI Action Space for Efficiency

Human computer use fluidly switches between GUI (visual inspection, window interaction) and CLI (batch file processing, structured computation). CUA-Master unifies both in a single action space, allowing the model to choose the appropriate modality per step. This hybrid mode reduces execution steps compared to GUI-only, benefiting long-horizon tasks. Examples include batch grading answer sheets and entering results into spreadsheets, processing documents and completing web forms, and combining professional visualization software for data verification.

Training Data Mechanisms: Diversity, Controllability, Reliability

The framework addresses three core challenges in CUA training data.

Diversity via Capability Primitive Composition

Extracting primitives from real tutorial videos and composing them continuously expands task types. t-SNE visualization shows generated tasks cover existing CUA dataset distributions and extend into sparser regions.

Controllability via K and N Parameters

Task difficulty is precisely controlled along two dimensions: K (number of composed capability primitives) and N (number of involved applications). Preconditions, dependencies, and verification conditions enable construction of tasks ranging from single-app simple operations to cross-app complex workflows.

Reliability via Snapshot Rollback Retry

Long tasks often fail due to local errors. A feedback-guided snapshot rollback mechanism locates the failure node, rolls back to the latest valid snapshot, and re-executes only the failed segment instead of restarting entirely.

Training Details and Resources

The paper reports three SFT stages accumulating ~115K training samples and 12K high-quality RL samples. The base model is Qwen-2.5-27B (implied by comparisons). The model, code, and paper are publicly available:

CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements
https://arxiv.org/pdf/2611.03126
https://github.com/xlyu0106/CUA-Master.git
https://huggingface.co/yxl66666/CUAMaster-27B