NorthCUA :Turbocharging Computer-Use Agents through Experience at Scale
retained text
Scaling [computer-use agents](https://huggingface.co/papers?q=computer-use%20agents) from individual actions to complex software work requires experience that connects decisions to complete outcomes. We present NorthCUA, an integrated system for turbocharging [computer-use agents](https://huggingface.co/papers?q=computer-use%20agents) through experience at scale. Its [Experience Flywheel](https://huggingface.co/papers?q=Experience%20Flywheel) supplies diverse executable tasks to the NorthCUA Harness, which sustains long interactions, records decisions, and checks task outcomes. The Flywheel uses these records to curate verified trajectories. For unsolved tasks, [recursive self-improvement](https://huggingface.co/papers?q=recursive%20self-improvement) evolves [reusable skills](https://huggingface.co/papers?q=reusable%20skills) and preserves [verified milestones](https://huggingface.co/papers?q=verified%20milestones) so later attempts build on earlier progress. [Staged Learning](https://huggingface.co/papers?q=Staged%20Learning) combines this experience with grounding data from the Flywheel to develop GUI competence through mid-training and long-horizon [supervised fine-tuning](https://huggingface.co/papers?q=supervised%20fine-tuning). The policy then learns from fresh attempts in the same Harness through joint desktop-web [reinforcement learning](https://huggingface.co/papers?q=reinforcement%20learning), closing the loop between experience production and policy improvement. The resulting Nano, Mini, and Pro family achieves state-of-the-art results among open-source models across [GUI grounding](https://huggingface.co/papers?q=GUI%20grounding), [desktop execution](https://huggingface.co/papers?q=desktop%20execution), and [Web interaction](https://huggingface.co/papers?q=Web%20interaction) under the reported evaluation settings. For more responsive Web workflows, NorthCUA-Ultrafast pairs direct action selection with open-ended generation on pre-RL Nano, offering a measured quality-latency tradeoff. Together, these advances establish NorthCUA as a high-performing model family and a reusable foundation for agents that carry complex software work from intent to completion. [https://github.com/beipo-ai/NorthCUA](https://github.com/beipo-ai/NorthCUA) [https://beipo.top/blog/northcua/](https://beipo.top/blog/northcua/) [https://arxiv.org/abs/2611.02102](https://arxiv.org/abs/2611.02102) weights [https://huggingface.co/china-beipo/NorthCUA-Nano](https://huggingface.co/china-beipo/NorthCUA-Nano) [https://huggingface.co/china-beipo/NorthCUA-Mini](https://huggingface.co/china-beipo/NorthCUA-Mini) [https://huggingface.co/china-beipo/NorthCUA-Pro](https://huggingface.co/china-beipo/NorthCUA-Pro)
document fetched 2026-11-04 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xmlwjv/northcua_turbocharging_computeruse_agents_through