CUA-Master: Nested Dual-Loop Framework Scales Computer-Use Agents via Open-World Recursive Improvements
Xinlei Yu · hf · 2026-11-03
CUA-Master, from CUHK MMLab, THU, HKUST(GZ) and JD Explore Academy, tackles the limited diversity, controllability and reliability of interaction data for computer-use agents (CUAs).
- Addresses diminishing returns of self-generated recursive improvement pipelines by introducing a nested dual-loop framework
- System-level outer loop: a data-model flywheel continuously incorporates open-world corpora and leverages evolving model weaknesses to guide subsequent data construction and optimization iterations
- Data-level inner loop: couples open-world seeding (mining authentic workflows as seed tasks from situational videos with visual-semantic dual-path trajectory reconstruction) and autocatalytic propagation
More from coding & agent
- Live workshop: instrument your AI agent with OpenTelemetry and debug it via Langfuse traces — Al_Grigor · 2026-11-03
- AutoRLBench: A Benchmark Testing AI Agents on Real RL Training Failures — wzenus · 2026-11-03
- Building live TV fact-checking agents on Bedrock AgentCore at re:Invent IND421 — FlolightC · 2026-11-03
- New Tutorial: Driving Gemini for Computer Use with Cua Driver — osanseviero · 2026-11-03
- SecureInput: open-source local secret input for MCP coding agents — rishabhgusain · 2026-11-03
- Sidelight runs browser agents in your real Chrome session to dodge login walls — Equivalent_Type6726 · 2026-11-03