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

NorthCUA is an integrated harness, experience-production flywheel, and staged-learning pipeline that trains Nano, Mini, and Pro computer-use agents and reports strong GUI grounding, desktop, web, and latency results.

Scaling computer-use agents from single actions to complex software work requires experience that connects decisions to complete outcomes. Constructing varied executable tasks with reliable verifiers, sustaining long GUI interactions, and turning records into learning samples are linked bottlenecks. Independent execution can enter states absent from demonstrations, so fresh interaction and outcome feedback are needed for continued improvement.

The Experience Flywheel builds single-step grounding examples and executable CaseSpec tasks through scaling intents, environments, and verifiers, then curates assessed trajectories; for unsolved tasks, recursive self-improvement uses skill guidance or milestone checkpoints. The NorthCUA Harness runs desktop and web tasks through a shared ReAct-style runtime, action-aware visual context, task-preserving compaction, structured notes, an agent-as-judge evaluator, and a batch rollout service. Staged Learning uses broad GUI mid-training with grounding data, verified long-horizon SFT, and joint desktop-web GRPO reinforcement learning from fresh verifier-scored attempts. An Experience Adapter preserves decision-time inputs and attempt-level reward attribution across context changes. NorthCUA-Ultrafast adds a choice head for finite web actions while falling back to native generation for open-ended outputs.

• NorthCUA-Pro reaches the highest grounding scores in the compared table, with Mini and Nano retaining much of the capability. §6.2 GUI Grounding Results
• On OSWorld-Verified desktop workflows, final RL checkpoints score highest for Pro, then Mini, then Nano. §6.3 Desktop Agent Results
• On OSWorld 2.0 Pro reaches 30.6% binary completion and 66.8% partial credit; within the 108-task reference block it exceeds GPT-5.6 Sol by 3.2 and 4.1 percentage points. §6.3 Desktop Agent Results
• On WebArena-Verified Pro reaches 73.9%, Mini 70.4%, and Nano 67.2%; Pro is 2.3 points above the joint-highest reference score of 71.6%. §6.4 Web Agent Results
• Long-horizon SFT improves desktop success by 7.4 points for Nano, 15.0 for Mini, and 10.0 for Pro; OSWorld 2.0 partial scores rise by 16.7, 20.7, and 33.0 points. §6.5 Progress Across Training Stages
• Ultrafast Head + LoRA reaches 60.6% WebArena success at 266.8 s mean task time versus Full generation 64.2% at 616.8 s, a factor-of-2.3 reduction; direct-action bypass is 31.1% versus 40.2% for Head-only. §6.8 NorthCUA-Ultrafast

• The provided text's §8 Limitations and Responsible Use section contains no stated limitations, so broader safety and ethics limits are left unaddressed there.
• Ultrafast adaptation results come from one training seed and states can share a task; the LoRA result is a validation comparison, as stated in E.7.3 and 6.8.
• The online Ultrafast timing comparison uses a fixed 320-task success-selected subset rather than all 812 tasks, as stated in 6.8.
• Stage comparisons measure complete training transitions including data selection and additional optimization, not isolated component effects, as stated in 6.5.

• The Experience Flywheel constructs single-step grounding pairs and executable tasks through scaling intents, environments, and verifiers, then curates assessed trajectories. §4 Experience Flywheel: Scaling Grounding and Agentic Experience
• Long-horizon execution uses a sliding visual window with action overlays, task-preserving compaction, and agent-written structured notes. §3.2 Sustaining Long-Horizon Interaction
• The evaluator acts as an agent-as-a-judge that combines recorded evidence and live workspace inspection under task-grounded multidimensional rubrics. §3.3 Agentic Outcome Evaluation
• Batch rollout separates centralized agent control from isolated device workers, pipelines environment preparation, and applies stage-specific concurrency and recovery. §3.4 Reliable Batch Rollout at Scale
• Recursive self-improvement addresses hard tasks through skill-guided knowledge accumulation and milestone-guided progress accumulation while keeping original acceptance conditions fixed. §4.3 Recursive Self-Improvement for Hard-Task Collection
• Broad GUI mid-training reports over 100,000 trajectories and over 4 million GUI interaction steps plus approximately 40,000 single-step grounding examples. §5.1 Broad GUI Mid-Training
• Verified long-horizon SFT uses approximately 3,000 trajectories and 310,000 GUI interaction steps; the RL pool has approximately 1,500 desktop tasks and 500 Web tasks. §5.2 Verified Long-Horizon SFT
• Joint desktop-web RL screens tasks by collecting at least five SFT-policy rollouts and selecting cases with variation among valid verifier scores. §5.3.1 Selecting Tasks with Outcome Variation