The paper presents NorthCUA, a family of computer-use agents for GUI and web tasks, built through an "experience flywheel" and reinforcement learning.
Training: The Flywheel runs executable tasks in a harness that records decisions and checks outcomes, then curates verified trajectories. Unsolved tasks go through recursive self-improvement, and staged learning combines mid-training, supervised fine-tuning and joint desktop-web reinforcement learning.
Models: The Nano, Mini and Pro family, which the paper says achieves state-of-the-art results among open-source models on GUI grounding, desktop execution and web interaction under its reported evaluation settings.
Fast variant: NorthCUA-Ultrafast pairs direct action selection with open-ended generation on pre-RL Nano, which the paper calls a measured quality-latency tradeoff.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper presents NorthCUA, a family of computer-use agents for GUI and web tasks, built through an "experience flywheel" and reinforcement learning.