Many computer-use agents can perform isolated actions but struggle to carry a task through changing software states and verify the final result. The report’s central idea is to treat complete, checked attempts—not just clicks or screenshots—as the experience that drives improvement. That links what an agent decides to what it actually produces.
AI Agent Papers2026
NorthCUA Technical Report: Turbocharging Computer-Use Agents through Experience at Scale
Trains computer-use agents on verified trajectories from long interactions, connecting task generation, execution, evaluation, and policy improvement in a feedback loop. Reports Nano, Mini, and Pro models across desktop and web benchmarks.
Introduction
Information
- Websitearxiv.org
- AuthorsChaoyun Zhang, Bingxuan Li, Pinxue Guo, Yangfan Gao, Zikun Cai, Tianci Xie, Hui Li, Yuqi Huang, Yifei Cao, Bo Mao …
- Published date2026/11/01
More Items
AI Agent Papers2026
Mines real computer-use workflows from instructional videos, turns them into reusable skills, and composes them into verifiable tasks. An open-data and model-feedback loop guides each round of dataset expansion.
AI Agent Papers2026
Provides a unified tool and execution harness for open-weight language models to operate robots and support further training. The paper reports a RoboDojo score of 31.39 for Qwen3.8-Flash-Next and competitive results across 19 real-world tasks.
AI Video Papers2026
Explores how agent harnesses coordinate video models for multi-shot production, interactive world modeling, and recursive self-improvement. It distinguishes improvements to outputs, tools, model parameters, and the development process itself.