SCALECUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL

Jul 16, 2026

Computer use agents (CUAs) automate complex digital workflows via visual perception and GUI execution, and online reinforcement learning with verifiable rewards (RLVR) is key to scaling. Existing efforts face two bottlenecks: scarce verifiable data and inefficient online RL. SCALECUA introduces a unified framework that scales online RL through verifiable task synthesis and efficient training, including VERIGEN for generating 24K+ verifiable tasks and Frontier Sampling plus Visual Context Segmentation to improve sample efficiency and training throughput.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages21

Reading Time53 min

Abstract

Introduction

Challenges in Scaling RLVR for CUAs

Data Bottlenecks: Verifiable GUI Task Scarcity

Online RL Inefficiency: Sampling and Multi-turn Training Trade-offs

arXiv :2607 . 1 1 185v 1 [ cs .AI] 13 Jul 2026 SCALECUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL Bowen Lv 1∗†, Xiao Liu 1 ,2∗, Yanyu Ren 1†, Hanyu Lai 1†, Bohao Jing2†, Hanchen Zhang 1†, Yanxiao Zhao 1†, Shuntian Yao 1†, Jie Tang 1 , Yuxiao Dong 1 1Tsinghua University 2Z.AI Abstract Computer use agents (CUAs) are emerging as a powerful interface for automating complex digital workflows through visual perception and GUI execution. Online reinforcement learning with verifiable rewards (RLVR) has emerged as a key direction for scaling their capabilities. However, this paradigm is bottlenecked by verifiable data scarcity and online RL inefficiency. To break these barriers, we introduce SCALECUA, a unified framework that scales online RL for CUAs via verifiable task synthesis and efficient training. At the data level, we design VERIGEN, an end-to-end framework for generating verifiable RL tasks through iterative docker interactions and a multi-agent feedback loop. Scaled to 100+ concurrent agent workers via a shared docker interaction probe, this pipeline produces 24K+ verifiable tasks and nearly 3K high-quality RL tasks. To maximize sample efficiency, we propose Frontier Sampling, which tracks per-task capability and allocates rollouts to the current learning frontier. On the training side, we further design Visual Context Segmentation, a sliding window over recent visual context that balances rollout and training-engine pressure, yielding a 2.83× training speedup over step-wise decomposition. Together, SCALECUA achieves 68.7% on OSWorld and 54.0% on ScienceBoard, establishing new state-of-the-art performance among open-source computer use agents. Code, models, and datasets are available at [https://github.com/THUDM/SCALE-CUA](https://github.com/THUDM/SCALE-CUA). OSWorld Success Rate (%) | 68.7 ScaleCUA | | | --- | --- | | Claude Sonnet 4.5 (62 .9) | | | | EvoCUA-32B (56 .7) | | ComputerRL-9B (48.0) Start Point | | 0K 2K 5K 10K 20K Synthetic OSWorld Task Nums Figure 1: Left: Comparison with open-source models on OSWorld. SCALECUA achieves 68.7%, surpassing most prior open-source models including those 4× larger. Right: Effect of scaling generated verifiable tasks on OSWorld performance under the same training pipeline. 1 Introduction Computer use agents (CUAs) have emerged as a transformative interface for automating everyday digital workflows through visual perception and GUI action execution [Xie et al., 2024, Zhou et al., *Equal contribution. †Work done while these authors interned at Z.AI. Figure 2: Trained SCALECUA agent solving a 47-step multi-app OSWorld task end-to-end. Given the instruction “install the Orchis GTK theme and switch to it for GNOME”, the Qwen3.5- 9B rollout progresses left-to-right, top-to-bottom—browsing [gnome-look.org](gnome-look.org), downloading and extracting the theme, then applying it through the terminal and gsettings—until the OSWorld judge returns score=1 .0. 2023, Hong et al., 2024, Liu et al., 2024] . Unlike traditional automation built on structured APIs, CUAs operate directly on screen pixels and UI elements, enabling profound generalization across diverse applications and operating systems [Gou et al., 2024, Wu et al., 2024, Qin et al., 2025] . Each CUA task is naturally a multi-turn interaction loop: the agent observes a screenshot, actson the interface, and iterates until task completion. Recent commercial breakthroughs such as Claude Computer Use [Anthropic, 2024] and OpenAI Operator [OpenAI, 2025b] demonstrate the immense practical value of this paradigm. To further push the boundaries of CUA capabilities, the research community has begun shifting toward online reinforcement learning with verifiable rewards (RLVR) [Bai et al., 2024, Qi et al., 2024, Lai et al., 2025b, Xu et al., 2025, Li et al., 2025] . However, while RLVR has shown initial promise, scaling it for CUAs remains largely unso

FAQs

What problem does SCALECUA address in scaling computer use agents?

SCALECUA addresses two key issues: verifiable data scarcity and online RL inefficiency when scaling CUAs using RLVR.

How does SCALECUA generate verifiable tasks at scale?

It introduces VERIGEN, an end-to-end pipeline that uses iterative docker interactions and a multi-agent feedback loop to produce 24K+ verifiable tasks.

What training methods does SCALECUA use to improve efficiency?

It proposes Frontier Sampling to allocate rollouts to the current learning frontier, and Visual Context Segmentation to balance rollout and training pressure for a substantial training speedup.

Jul 16, 2026

Upload by Theodora