Pith. sign in

REVIEW 3 major objections 31 references

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

T0 review · 3 major / 0 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ScaleCUA claims that verifiable GUI task synthesis plus efficient online RL can push open computer-use agents to new open-source highs on desktop and scientific software benchmarks.

desk verdict Solid open systems paper: live-OS verifiable task factory + frontier sampling + sliding visual context actually moves open CUA RL numbers, with real ablations and code claims; residual judge noise and mild OSWorld template reuse are real but secondary to the empirical package. read the letter →

arxiv 2607.11185 v1 pith:FAFZNG2L submitted 2026-07-13 cs.AI

Bowen Lv , Xiao Liu , Yanyu Ren , Hanyu Lai , Bohao Jing , Hanchen Zhang , Yanxiao Zhao , Shuntian Yao
show 2 more authors
Jie Tang Yuxiao Dong
This is my paper
classification cs.AI
keywords computeruseagentsonlinereinforcementlearningverifiablerewardsGUItasksynthesisfrontiersamplingvisualcontextsegmentationOSWorldScienceBoard
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Computer-use agents that act from screenshots are limited by two bottlenecks: almost no large pools of GUI tasks with machine-checkable rewards, and multi-turn online RL that wastes rollouts and training time. ScaleCUA argues both can be broken together. On the data side, VeriGen drives multi-agent loops against live desktop containers to invent tasks with executable judges, then refines them from the model’s own successes and failures, yielding tens of thousands of verifiable tasks and a few thousand RL-ready ones. On the training side, Frontier Sampling steers rollouts to tasks near the model’s current success rate, and Visual Context Segmentation keeps only a sliding window of recent screenshots so long trajectories stay trainable. With that stack, a 9B open model reaches 68.7% on OSWorld and 54.0% on ScienceBoard, above prior open systems including much larger ones. A sympathetic reader cares because the paper offers a concrete recipe—synthetic verifiable tasks plus sampling and context control—for scaling agents that automate real desktop workflows without hand-labeled rewards.

What carries the argument

VeriGen is the data engine: multi-agent propose–judge–check loops against live Docker desktops, plus trajectory-guided splits and compositions, producing 24K+ verifiable tasks and nearly 3K RL tasks. Frontier Sampling tracks per-task exponential-moving-average success and samples near a target mid-difficulty band. Visual Context Segmentation segments long rollouts with a sliding window of recent screenshots (about 2.83× faster than step-wise training) while keeping textual continuity.

What would settle it

Train the same backbone on the same pipeline but with judges that an expert panel rates as systematically wrong on a large held-out set (or replace them with human-verified judges only); if OSWorld and ScienceBoard gains collapse while synthetic task count stays high, the claim that VeriGen-scale verifiable rewards drive the scaling fails.

Watch

Extended reading notes

Core claim

The paper’s central claim is that large-scale online reinforcement learning with verifiable rewards becomes practical for computer-use agents when three pieces work together: an end-to-end synthesis pipeline that produces executable GUI tasks and judges from live OS environments, a sampler that allocates rollouts to tasks at the model’s current capability frontier, and a sliding visual-context scheme that balances rollout and training cost on multi-turn trajectories—together lifting a 9B open agent to 68.7% OSWorld and 54.0% ScienceBoard success, new state-of-the-art among open-source computer-use agents.

Load-bearing premise

The load-bearing premise is that the automatically generated executable judges are accurate enough reward signals for learning—even though human audit agreement is only about four in five cases overall.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. ScaleCUA proposes a unified pipeline for scaling online RL with verifiable rewards for computer-use agents. VeriGen synthesizes executable GUI tasks and judges via multi-agent interaction with live Docker OS environments (24K+ candidates, ~3K RL tasks). Frontier Sampling allocates rollouts using EMA per-task success rates under a Gaussian kernel peaked at intermediate difficulty. Visual Context Segmentation packages multi-turn trajectories with a sliding visual window, claiming a 2.83× training speedup over step-wise decomposition. Applied to GLM-4.6V-Flash, Qwen3-VL-8B-Thinking, and Qwen3.5-9B, the method reports 68.7% on OSWorld and 54.0% on ScienceBoard, presented as open-source SOTA, with component ablations, sampling/speed comparisons, and judge audits in the appendix.

Significance. If the results hold under stricter contamination controls, this is a substantial systems contribution: it attacks the two real bottlenecks of CUA RLVR—scarce executable rewards and multi-turn multimodal training inefficiency—and ships a large verifiable task factory, sampling and trajectory packaging methods, multi-model gains, and public code/models/datasets. The ScienceBoard result is especially useful as a check beyond OSWorld-style office workflows. The work is empirical rather than theoretical, but the combination of live-OS judge synthesis at 100+ worker scale, frontier-aware sampling, and a concrete multi-turn training speedup is of clear interest to the agent RL community.

major comments (3)
  1. Appendix B.6 / Table 8: the SOTA claim on OSWorld rests on a training pool with 2.10% exact judge reuse (22/1049 traj_verify) and rare objective near-duplicates (0.19%). The paper treats source IDs as non-leakage and only excludes full-JSON / exact-instruction duplicates. For a benchmark SOTA claim this is too weak. Please report OSWorld (and, if applicable, subset) scores after removing exact judge-reuse and objective near-duplicate tasks from the RL pool, and state whether the 68.7% result is materially unchanged. ScienceBoard’s 0% judge reuse already helps; OSWorld needs the same decontamination table.
  2. Appendix B.3 / Table 6: expert–judge agreement is 82.5% overall (78% OSWorld; 11 FP / 17 FN on 160 audits). Mechanical executability (94.5%) does not imply unbiased reward. Because residual false positives/negatives can systematically shape GRPO advantages, the manuscript should either (i) quantify policy sensitivity to label flips on a held-out audit set, or (ii) retrain/evaluate with human-corrected judges on a non-trivial subset and report the delta. Without this, the claim that VeriGen supplies reliable RLVR signal remains only partially supported.
  3. §4.2 / Tables 2–3: several strong proprietary and open baselines are evaluated at 100 steps while ScaleCUA is reported at 50 steps (and some rows lack per-subset scores or are taken from reports). The paper already lists step budgets, but the open-source SOTA comparison should include a matched-budget column or a 100-step ScaleCUA run where feasible, and should flag any baseline whose action interface or observation modality differs (screen-only vs s+a11y). Otherwise the headline ranking overstates cross-system comparability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical systems paper whose SOTA claims rest on external benchmarks, not on quantities forced by construction from fitted inputs or self-citation uniqueness.

full rationale

ScaleCUA’s load-bearing chain is (1) VeriGen synthesizes executable-judge GUI tasks via live Docker multi-agent loops, (2) Frontier Sampling allocates rollouts by EMA success rates with a fixed Gaussian kernel (μ=0.5, σ=0.25), (3) Visual Context Segmentation packages multi-turn trajectories with a sliding visual window, and (4) GRPO/DAPO online RL is evaluated on OSWorld and ScienceBoard. None of these steps reduces a claimed prediction to its own definition or fit. Benchmark scores are external success rates under fixed harnesses, not quantities derived from training-task pass rates. Hyperparameters (μ, σ, K, α, γ) are stated design choices validated by ablations and compute sweeps, not parameters fitted to the reported SOTA numbers. Trajectory-guided synthesis reuses the model’s own rollouts to mint easier/harder tasks—standard curriculum self-bootstrapping, not a tautology that forces the held-out benchmark scores. Self-citations (ComputerRL, AgentRL, related author work) supply infrastructure and baselines; they are not uniqueness theorems that forbid alternatives or smuggle an ansatz that alone produces the result. Appendix B.6’s overlap audit and B.3’s judge–human agreement concern contamination and reward noise (correctness risk), not circular derivation. The paper is self-contained against external benchmarks; score 0 is the honest finding.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

Load-bearing content is empirical and systems-level: GRPO-style online RL with executable environment rewards, Docker desktops as training/eval worlds, and hyperparameterized sampling/windowing. No formal derivation; free parameters are training knobs chosen for efficiency and pass rate. Invented pieces are algorithmic frameworks validated by ablations rather than postulated physical entities.

free parameters (4)
  • Frontier Sampling target μ and bandwidth σ
    Gaussian kernel peaks sampling at μ=0.5 with σ=0.25 (plus γ=0.2 uniform mix and EMA α=0.2); chosen to concentrate rollouts near the learning frontier and affect convergence curves in Fig. 6.
  • Visual window size K (and Δ)
    K=5 (OSWorld) / K=8 (ScienceBoard), Δ=K, selected from a compute sweep (Fig. 5) and pass@k study (Fig. 7); directly controls the reported 2.83× speedup and success trade-off.
  • GRPO/DAPO clip and KL settings
    ε_l=0.2, ε_h=0.28, clip cap c=3.0, λ_KL=1e-4, incomplete-episode penalty −0.2 (Table 10); standard RL knobs that shape the optimized policy.
  • Initial pass-rate filter [p_min, p_max] and task-pool filtering
    Tasks outside a capability band are dropped before RL; determines which of the ~3K ‘high-quality’ tasks enter training and thus the reported gains.
assumptions (4)
  • domain assumption Executable environment judges (filesystem/browser/Python checks) are a sufficient and mostly correct reward for multi-turn GUI RL.
    Core of RLVR framing in §1 and VeriGen design §3.2; partially stress-tested by 94.5% executable rate and 82.5% human agreement (Appendix B).
  • domain assumption Screenshot-only observation plus the exposed mouse/keyboard (and ScienceBoard extras) action interface is adequate for the evaluated tasks.
    Stated in §4.1 and Appendix A; no a11y trees used for perception.
  • domain assumption OSWorld and ScienceBoard success rates under ≤50 turns are valid proxies for open computer-use capability.
    Main evaluation protocol §4; Limitations note ultra-long workflows and non-Ubuntu OSes are out of scope.
  • domain assumption GRPO with group-relative advantages and DAPO-style asymmetric clipping is a sound online RL objective for these agents.
    Training setup §4.1 and Appendix D; standard in recent LLM RL literature the paper builds on.
invented entities (3)
  • VeriGen multi-agent verifiable task factory independent evidence
    purpose: Scale synthesis of GUI tasks with executable judges via proposer/judger/checker loops, Docker probe parallelism, and trajectory-guided split/compose.
    Primary data contribution (§3.2); independent handle is the released task pool and judge audit metrics, not an external physical prediction.
  • Frontier Sampling independent evidence
    purpose: Allocate rollouts using EMA success rates and a Gaussian kernel around a target pass rate.
    §3.3; validated against uniform/DAPO/curriculum in Fig. 6(a).
  • Visual Context Segmentation independent evidence
    purpose: Sliding-window packaging of multi-turn multimodal trajectories to balance rollout vs training engine load.
    §3.4 / Algorithm 1–2; validated by 2.83× speedup and pass@k vs K.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCALECUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL." pith.science (2026). https://pith.science/paper/FAFZNG2L

@misc{pith2026260711185,
  author       = {Pith},
  title        = {Pith review of: SCALECUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FAFZNG2L}},
  note         = {Machine review of arXiv:2607.11185}
}
read the original 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 2.83x 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.

Figures

Figures reproduced from arXiv: 2607.11185 by the authors.

Figure 1
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 act… view at source ↗
Figure 2
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, 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., 202… view at source ↗
Figure 3
Figure 3. Overview of SCALECUA. VERIGEN autonomously synthesizes verifiable GUI tasks with executable rewards from live OS environments at 100+ parallel worker scale; combined with upstream Frontier Sampling and long-horizon Visual Context Segmentation, SCALECUA enables efficient large-scale online RL for computer use agents. executes actions and observes the resulting state updates (screenshots and accessibility trees) to te… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Visual Context Segmentation. Left: sliding-window trajectory processing with a token-ID stream. Right: when the visual window overflows, the oldest screenshots are dropped and replaced by a text summary; trainable masks (green, loss=1) cover only assistant responses, w…
Figure 5
Figure 5. Figure 5: Sliding-Window training step time across K at 2/4/8-node scales; moderate K is optimal. To balance this trade-off, we propose Visual Context Segmen￾tation, which keeps the recent visual context within a sliding window while preserving textual continuity. This simultane…
Figure 6
Figure 6. Figure 6: (a) Frontier Sampling vs. uniform, DAPO, and curriculum learning: higher reward sustained [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Pass@k on OSWorld un￾der different sliding-window sizes K (Qwen3.5-9B). Bounded visual context does not reduce task success, and mod￾erate windows (K=3–5) perform best. Visual Context Segmentation. We compare Visual Con￾text Segmentation (K=5) against the standard step…
Figure 8
Figure 8. Figure 8: VERIGEN analysis. Left: VERIGEN produces an order-of-magnitude larger verifiable task pool than prior desktop CUA methods. Right: augmenting the RL pool with trajectory-guided tasks consistently raises task-level reward throughout training. 4.5 Case Study: Impact of Vi…
Figure 9
Figure 9. Figure 9: Anonymized Python judge corresponding to the task schema in Table 5. The concrete [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Per-domain task distribution of VERIGEN-generated tasks. (a) OSWorld: 10 application domains, total 22,322 tasks. (b) ScienceBoard: 6 scientific domains, total 5,529 tasks. Counts aggregate across all generation stages. and 90.0% on ScienceBoard. This audit complement…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 linked inside Pith

  1. [1]

    Agent s2: A compositional generalist-specialist framework for computer use agents.arXiv preprint arXiv:2504.00906,

    Saaket Agashe, Kyle Wong, Vincent Tu, Jiachen Yang, Ang Li, and Xin Eric Wang. Agent s2: A compositional generalist-specialist framework for computer use agents.arXiv preprint arXiv:2504.00906,

    arXiv

  2. [2]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

    arXiv

  3. [3]

    Windows agent arena: Evaluating multi-modal os agents at scale.arXiv preprint arXiv:2409.08264,

    Rogerio Bonatti, Dan Zhao, Francesco Bonacci, Dillon Dupont, Sara Abdali, Yinheng Li, Yadong Lu, Justin Wagle, Kazuhito Koishida, Arthur Bucker, et al. Windows agent arena: Evaluating multi-modal os agents at scale.arXiv preprint arXiv:2409.08264,

    arXiv

  4. [4]

    Gui-genesis: Automated synthesis of efficient environments with verifiable rewards for gui agent post-training.arXiv preprint arXiv:2602.14093,

    Yuan Cao, Dezhi Ran, Mengzhou Wu, Yuzhe Guo, Xin Chen, Ang Li, Gang Cao, Gong Zhi, Hao Yu, Linyi Li, et al. Gui-genesis: Automated synthesis of efficient environments with verifiable rewards for gui agent post-training.arXiv preprint arXiv:2602.14093,

    arXiv

  5. [5]

    Agentic reinforced policy optimization.arXiv preprint arXiv:2507.19849,

    Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, et al. Agentic reinforced policy optimization.arXiv preprint arXiv:2507.19849,

    arXiv

  6. [6]

    Navigating the digital world as humans do: Universal visual grounding for gui agents

    Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243,

    arXiv

  7. [7]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  8. [8]

    Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006,

    Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006,

    arXiv

Show all 31 references
  1. [9]

    Coact: A global-local hierarchy for autonomous agent collaboration.arXiv preprint arXiv:2406.13381,

    Xinming Hou, Mingming Yang, Wenxiang Jiao, Xing Wang, Zhaopeng Tu, and Wayne Xin Zhao. Coact: A global-local hierarchy for autonomous agent collaboration.arXiv preprint arXiv:2406.13381,

    arXiv

  2. [10]

    Androidgen: Building an android language agent under data scarcity

    11 Hanyu Lai, Junjie Gao, Xiao Liu, Yifan Xu, Shudan Zhang, Yuxiao Dong, and Jie Tang. Androidgen: Building an android language agent under data scarcity. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages ...

  3. [11]

    Visualagentbench: Towards large multimodal models as visual foundation agents.arXiv preprint arXiv:2408.06327,

    Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Yifan Xu, Xixuan Song, Shudan Zhang, Hanyu Lai, Xinyi Liu, Hanlin Zhao, et al. Visualagentbench: Towards large multimodal models as visual foundation agents.arXiv preprint arXiv:2408.06327,

    arXiv

  4. [12]

    GPT-5.4 system card.https://openai.com/index/introducing-gpt-5-4/, 2025a

    OpenAI. GPT-5.4 system card.https://openai.com/index/introducing-gpt-5-4/, 2025a. OpenAI. Introducing Operator. https://openai.com/index/introducing-operator/, 2025b. Zehan Qi, Xiao Liu, Iat Long Iong, Hanyu Lai, Xueqiao Sun, Wenyi Zhao, Yu Yang, Xinyue Yang, Jiadai Sun, Shunt...

    arXiv

  5. [13]

    Ui-tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326,

    Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326,

    arXiv

  6. [14]

    Bytedance Seed. Seed1. 8 model card: Towards generalized real-world agency.arXiv preprint arXiv:2603.20633,

    arXiv

  7. [15]

    Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    arXiv

  8. [16]

    Mobilegui-rl: Advancing mobile gui agent through reinforcement learning in online environment.arXiv preprint arXiv:2507.05720,

    Yucheng Shi, Wenhao Yu, Zaitang Li, Yonglin Wang, Hongming Zhang, Ninghao Liu, Haitao Mi, and Dong Yu. Mobilegui-rl: Advancing mobile gui agent through reinforcement learning in online environment.arXiv preprint arXiv:2507.05720,

    arXiv

  9. [17]

    Megatron-lm: Training multi-billion parameter language models using model parallelism

    Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catan- zaro. Megatron-lm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053,

    1909 arXiv

  10. [18]

    Os-genesis: Automating gui agent trajec- tory construction via reverse task synthesis

    Qiushi Sun, Kanzhi Cheng, Zichen Ding, Chuanyang Jin, Yian Wang, Fangzhi Xu, Zhenyu Wu, Chengyou Jia, Liheng Chen, Zhoumianze Liu, et al. Os-genesis: Automating gui agent trajec- tory construction via reverse task synthesis. InProceedings of the 63rd Annual Meeting of the Asso...

    arXiv

  11. [19]

    Ui-tars-2 technical report: Advancing gui agent with multi-turn reinforcement learning.arXiv preprint arXiv:2509.02544, 2025a

    Haoming Wang, Haoyang Zou, Huatong Song, Jiazhan Feng, Junjie Fang, Junting Lu, Longxiang Liu, Qinyu Luo, Shihao Liang, Shijue Huang, et al. Ui-tars-2 technical report: Advancing gui agent with multi-turn reinforcement learning.arXiv preprint arXiv:2509.02544, 2025a. Xinyuan W...

    arXiv

  12. [20]

    Agentsynth: Scalable task generation for generalist computer-use agents.arXiv preprint arXiv:2506.14205, 2025a

    Jingxu Xie, Dylan Xu, Xuandong Zhao, and Dawn Song. Agentsynth: Scalable task generation for generalist computer-use agents.arXiv preprint arXiv:2506.14205, 2025a. Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh J Hua, Zhoujun Cheng, Dongc...

  13. [21]

    Scaling computer-use grounding via user interface decomposition and synthesis.arXiv preprint arXiv:2505.13227, 2025b

    Tianbao Xie, Jiaqi Deng, Xiaochuan Li, Junlin Yang, Haoyuan Wu, Jixuan Chen, Wenjing Hu, Xinyuan Wang, Yuhui Xu, Zekun Wang, et al. Scaling computer-use grounding via user interface decomposition and synthesis.arXiv preprint arXiv:2505.13227, 2025b. Haiyang Xu, Xi Zhang, Haowe...

  14. [22]

    Mobilerl: Online agentic reinforcement learning for mobile gui agents.arXiv preprint arXiv:2509.18119,

    Yifan Xu, Xiao Liu, Xinghan Liu, Jiaqi Fu, Hanchen Zhang, Bohao Jing, Shudan Zhang, Yuting Wang, Wenyi Zhao, and Yuxiao Dong. Mobilerl: Online agentic reinforcement learning for mobile gui agents.arXiv preprint arXiv:2509.18119,

  15. [23]

    Agenttrek: Agent trajectory synthesis via guiding replay with web tutorials.arXiv preprint arXiv:2412.09605,

    Yiheng Xu, Dunjie Lu, Zhennan Shen, Junli Wang, Zekun Wang, Yuchen Mao, Caiming Xiong, and Tao Yu. Agenttrek: Agent trajectory synthesis via guiding replay with web tutorials.arXiv preprint arXiv:2412.09605,

    arXiv

  16. [24]

    Evocua: Evolving computer use agents via learning from scalable synthetic experience.arXiv preprint arXiv:2601.15876,

    Taofeng Xue, Chong Peng, Mianqiu Huang, Linsen Guo, Tiancheng Han, Haozhe Wang, Jianing Wang, Xiaocheng Zhang, Xin Yang, Dengchang Zhao, et al. Evocua: Evolving computer use agents via learning from scalable synthetic experience.arXiv preprint arXiv:2601.15876,

  17. [25]

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jian- hong Tu, Jingren Zhou, Junyang Lin, et al. Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement.arXiv preprint arXiv:2409.12122,

    arXiv

  18. [26]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a. Yan Yang, Dongxu Li, Yutong Dai, Yuhao Yang, Ziyang Luo, Zirui Zhao, Zhiyuan Hu, Junzhe ...

    arXiv

  19. [27]

    Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476,

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476,

    arXiv

  20. [28]

    Agentrl: Scaling agentic reinforcement learning with a multi-turn, multi-task framework.arXiv preprint arXiv:2510.04206,

    Hanchen Zhang, Xiao Liu, Bowen Lv, Xueqiao Sun, Bohao Jing, Iat Long Iong, Zhenyu Hou, Zehan Qi, Hanyu Lai, Yifan Xu, et al. Agentrl: Scaling agentic reinforcement learning with a multi-turn, multi-task framework.arXiv preprint arXiv:2510.04206,

  21. [29]

    Webarena: A realistic web environment for building autonomous agents.arXiv preprint arXiv:2307.13854,

    Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents.arXiv preprint arXiv:2307.13854,

    arXiv

  22. [30]

    The concrete artifact type, function suffix, target property, and file paths are masked, while the generated evaluator structure is preserved. 16 Calc Chrome Impress OS Writer GIMP T-bird VLC MultiVS Code 0 1000 2000 3000 4000 5000Number of Tasks 4515 3558 3418 2863 2239 1614 ...

    2000

  23. [31]

    Larger K values produce fewer but longer segments, requiring more rollout workers to keep the training engine fed

    Smaller K values produce more segments per episode with shorter sequences, so fewer rollout workers and more training workers are allocated. Larger K values produce fewer but longer segments, requiring more rollout workers to keep the training engine fed. In practice, 4-node a...

    2025

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.