NorthCUA Technical Report: Turbocharging Computer-Use Agents through Experience at Scale

NorthCUA Technical Report. Turbocharging Computer-Use Agents through Experience at Scale. By Chaoyun Zhang and twenty-nine others, from Shanghai Qiji Zhifeng. Published on the first of November.

Moving computer-use agents from individual actions to complex software work needs a particular kind of experience: experience that connects decisions to complete outcomes. Knowing that one action was reasonable is not the same as knowing the whole task got done. NorthCUA is presented as an integrated system for supplying that experience at scale, and for turning it into better agents.

The system is organised as a loop. An Experience Flywheel produces tasks and curated data. A Harness runs the agent, records what it does, and checks the result. A staged training recipe turns the data into a policy, and that policy then goes back into the Harness to generate more experience.

The Experience Flywheel supplies diverse executable tasks to the NorthCUA Harness. The Harness sustains long interactions, records decisions, and checks task outcomes. The Flywheel then uses those records to curate verified trajectories. For tasks that remain unsolved, recursive self-improvement evolves reusable skills and preserves verified milestones, so that later attempts build on earlier progress instead of starting over.

Training happens in stages. Staged Learning combines that experience with grounding data from the Flywheel, developing GUI competence through mid-training and then long-horizon supervised fine-tuning. After that, the policy learns from fresh attempts in the same Harness, through joint desktop-web reinforcement learning. That last step closes the loop between producing experience and improving the policy.

The result is a family of three models: Nano, Mini, and Pro. Under the reported evaluation settings, they achieve state-of-the-art results among open-source models across GUI grounding, desktop execution, and Web interaction. For Web workflows that need to be more responsive, a variant called NorthCUA-Ultrafast pairs direct action selection with open-ended generation, built on the Nano model before reinforcement learning. It offers a measured tradeoff between quality and latency.

The authors frame NorthCUA as both a high-performing model family and a reusable foundation for agents that carry complex software work from intent to completion. Its central design choice is that one Harness serves twice: first to gather and verify experience, and then as the place where the policy learns from its own fresh attempts.

NorthCUA Technical Report. Turbocharging Computer-Use Agents through Experience at Scale. By Chaoyun Zhang and twenty-nine others, from Shanghai Qiji Zhifeng. Published on the first of November. This is the long version: the whole paper rather than the abstract.

A computer-use agent that clicks the right button is not yet an agent that finishes the job. In a long workflow, earlier actions shape the software state that later decisions depend on. A value read from a spreadsheet has to survive until it is typed into a report, and an export that was issued is not the same as a file that exists. Training for that needs experience tying whole sequences of decisions to whole outcomes. Getting it at scale means building many tasks, each with a request, a working environment and a reliable way to judge success. Demonstrations alone are not enough either, because a policy acting on its own reaches states its demonstrations never showed. The NorthCUA team argues that these needs are linked, and that producing experience and learning from it have to be designed together.

Three components share one set of interfaces. The Experience Flywheel builds tasks, each packaged as a CaseSpec: the request, the input files, the environment and its setup, and the requirements for judging the result. The NorthCUA Harness runs each task in its own isolated sandbox, records every observation and action, and keeps the workspace alive until the outcome has been checked. Staged Learning turns the records into a policy in three steps: broad mid-training, then supervised fine-tuning on verified long workflows, then reinforcement learning on desktop and web tasks together. The models come in three sizes, Nano, Mini and Pro, the largest at three hundred and ninety-seven billion parameters, all built on Qwen3.5. Because the updated policy returns to the same Harness, the loop can turn again.

Long tasks overflow any context, so the Harness separates what the model sees from what is recorded. The model sees a sliding window of recent screenshots. Older frames carry overlays marking what was done: a click marker, a drag path, a scroll cue, a badge for typed text, so the agent can connect an action with its effect. When the text history grows too long, a summarizer compacts it, keeping goals, constraints, facts and pending work while dropping old coordinates that may have gone stale. It is told to keep an issued action distinct from an observed effect. The agent can also write itself notes, such as an overdue total carried from a workbook into a report.

Outcomes are judged by task verifiers and by an evaluator agent that reads the record and inspects the live workspace with shell, screen and browser tools. It opens the exported file rather than trusting that export was clicked, and its report separates whether requirements were met from how well the work was done. For tasks ordinary attempts cannot solve, there are two recursive routes. One writes skills from failed explorations, and a separate agent tests each skill before it is reused. The other breaks a task into milestones, snapshots the sandbox when a milestone passes, and branches from there, so a late failure does not mean rebuilding everything. Either way, the original verifier must pass the whole task.

The subtlest piece is in reinforcement learning. Because images leave the window and summaries rewrite history, the final transcript no longer shows what an earlier decision actually saw. So each response is trained on its recorded input, and a new training segment starts whenever the history stops being a clean extension. That splits attempts unevenly, so the advantage is computed once per attempt and shared by its segments, and the loss is averaged over each attempt's own tokens before averaging across attempts. Being split more often earns an attempt no extra weight.

The headline numbers are strong. Pro reaches eighty-eight point six percent on OSWorld-Verified and seventy-three point nine percent on WebArena-Verified, the highest in the paper's comparison tables. The test closer to the thesis is the stage comparison. The large jump comes from verified long-horizon fine-tuning. On the longer OSWorld 2.0 tasks, Pro's partial credit roughly doubles, from thirty point six to sixty-three point six percent, and its full completion rate nearly triples. Joint reinforcement learning then adds smaller but consistent gains, between two and a little over four points, at every size and on both platforms.

One case shows what changed. Before fine-tuning, Mini matched an exam question to the wrong source paper on similar wording, assembled all five files, and earned ten percent credit after two hundred and forty-three decisions. After fine-tuning, it checked the specific test, rejected a near miss, recovered the right paper from the Trash, and earned full credit in one hundred and twenty-four decisions. The weak spot is cross-site web work: Pro completes seventeen of forty-eight such tasks, against about three quarters of single-site ones.

The paper has a section called Limitations and Responsible Use. It says that unfamiliar software and changing interfaces remain hard. It says feedback is only as good as verifier coverage and the calibration of model judgments on open-ended content, and that long workflows funnel many decisions into one terminal outcome, so better intermediate feedback could help credit assignment. For deployment it asks for least-privilege credentials, confirmation for consequential actions, and adversarial testing against prompt injection hidden in pages and documents. Beyond that, many comparison scores are taken from other reports under their own protocols, the grounding benchmark uses the best of three complete runs, and Pro's web score comes from one complete evaluation.

The idea to carry away is that an agent's experience is only as useful as two things around it: the judgment of what each attempt actually achieved, and a faithful record of what each decision actually saw. Evaluators open the files. Milestones are checked before they are kept. Training segments preserve the decision-time input, and advantages stay with whole attempts. Get those right, and the same loop that collects demonstrations can keep teaching the policy that produced them.