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

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

Computer-use agents are moving beyond isolated clicks. The harder goal is completing software work that spans many decisions, applications, files, and changing interface states. NorthCUA, from the NorthCUA Team at Shanghai Qiji Zhifeng Co., Ltd., presents a system designed around that problem: generate executable tasks, run agents in persistent environments, verify the results, and feed the resulting experience back into training.

The core idea: build an experience flywheel

The report’s central contribution is not a single model trick but an integrated training loop. The NorthCUA Harness prepares isolated desktop or web environments, executes tasks, records actions and observations, retains the workspace when necessary, and evaluates the final result. The Experience Flywheel then turns these attempts into learning material.

Task generation follows three routes. Scaling Intents creates requests grounded in available applications and content. Scaling Environments builds the files, applications, or local websites needed by sourced requests. Scaling Verifiers adapts existing task-and-checker pairs into related tasks while updating the acceptance logic. Each route aims to produce a complete CaseSpec: a request, initial environment, setup instructions, and outcome checks.

For difficult tasks, the authors describe recursive self-improvement. Skill-guided exploration converts successful discoveries into reusable procedures, while milestone-guided search saves verified software states and resumes from them. A late failure therefore need not force the agent to repeat all earlier work. The full task verifier still decides whether a completed trajectory is accepted.

Training is staged rather than one-shot

NorthCUA uses three learning stages. Broad GUI mid-training combines grounding examples with diverse interaction trajectories. The reported corpus contains more than 100,000 trajectories and more than four million GUI steps, plus approximately 40,000 single-step grounding examples.

Verified long-horizon supervised fine-tuning then concentrates on complete workflows. The authors report approximately 3,000 trajectories and 310,000 GUI steps in this cohort. Finally, joint desktop-and-web reinforcement learning lets the policy generate fresh attempts. Groups of attempts at the same task receive verifier-based outcomes, and GRPO updates the policy using relative performance within each group.

A notable infrastructure detail is the separation between the complete execution record and the context shown to the model at each decision. NorthCUA uses recent visual history, task-preserving compaction, and agent-written notes to manage long interactions. During reinforcement learning, the Experience Adapter preserves each decision’s original input and assigns the complete attempt’s reward across any resulting training segments.

Reported results across grounding, desktop, and web work

The family includes Nano at 9B parameters, Mini at 35B total and 3B active parameters, and Pro at 397B total and 17B active parameters. The authors evaluate six benchmarks spanning GUI grounding, desktop execution, web interaction, and web testing.

  • Grounding: Pro reaches 93.3% on ScreenSpot-Pro and 91.0% on OSWorld-G. Mini reaches 92.4% and 87.9%, while Nano reaches 72.9% and 83.0%.
  • Desktop execution: On OSWorld-Verified, the reported scores are 78.9% for Nano, 85.3% for Mini, and 88.6% for Pro.
  • Long-horizon desktop work: On OSWorld 2.0, Pro reaches 30.6% full completion and 66.8% partial credit. The gap between these metrics shows that completing some requirements remains much easier than finishing an entire dependent workflow.
  • Web interaction: On WebArena-Verified, Nano scores 67.2%, Mini 70.4%, and Pro 73.9%.
  • Web testing: On WebTestBench with a supplied checklist, Pro reaches 60.3% F1, compared with 54.4% for Mini and 48.0% for Nano.

The stage comparisons support the report’s training narrative, although they should not be read as isolating one component perfectly. Long-horizon SFT improves desktop and OSWorld 2.0 performance across all three sizes. Subsequent joint RL improves OSWorld-Verified and WebArena-Verified for every tier; Mini’s OSWorld 2.0 binary score stays at 18.5% while its partial score rises from 46.3% to 50.1%.

Speed is a separate engineering problem

NorthCUA-Ultrafast explores a hybrid approach for web tasks. A choice head directly selects finite actions such as clicking or toggling, while open-ended text generation remains available when the agent must formulate a query or response. In the authors’ measured setup, the Head plus LoRA variant reaches 60.6% task success on WebArena-Verified, versus 64.2% for full generation. On a fixed success-selected timing subset, mean task time falls from 616.8 seconds to 266.8 seconds, a reported 2.3-fold reduction. This is a quality-latency tradeoff, not a free improvement.

Limitations and practical implications

The authors acknowledge that unfamiliar software, changing interfaces, weak verifiers, and open-ended content remain difficult. Terminal rewards can provide sparse feedback across very long trajectories, and agent-as-a-judge evaluation depends on evidence quality and calibration. Cross-site web tasks are especially challenging: Pro completes 17 of 48 reported cross-site tasks, versus 583 of 764 single-site tasks.

Benchmark comparisons also require care. The report combines its own evaluations with published baseline scores, and protocols, task versions, step limits, tools, and aggregation rules can differ. Its state-of-the-art claims therefore apply to the reported settings, not automatically to every deployment scenario.

For practitioners, the most transferable lesson is architectural. A useful computer-use system needs more than a capable multimodal model. It needs reproducible task initialization, isolated workspaces, outcome checks, retained artifacts, failure diagnosis, and explicit permission boundaries. Teams building internal agents should begin with narrow workflows whose final state can be verified, log every action and observation, and treat credentials, prompt injection, privacy, and consequential actions as first-class safety concerns.

NorthCUA’s demonstrated results suggest that experience production and policy training can be designed as one loop. Whether that loop generalizes reliably to real business software remains an open question, but the report offers a concrete blueprint for investigating it.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close