Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

CODA: Coordinating the Cerebrum and Cerebellum for a Dual-Brain Computer Use Agent with Decoupled Reinforcement Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read CODA claims that a trainable planner paired with a frozen executor can learn domain-specific GUI planning from automated judge rewards, and that distilling the resulting specialists into one generalist surpasses every specialist.

desk verdict Clever training recipe, but the numbers are unproven: the judge that produces every training reward is also the likely scorer of the headline results. read the letter →

arxiv 2508.20096 v1 pith:J7JZ3ZCQ submitted 2025-08-27 cs.CV cs.AIcs.LG

classification cs.CVcs.AIcs.LG
keywords computeruseagentGUIautomationreinforcementlearningGRPOplanner-executorarchitecturevision-languagemodelscientificsoftwareScienceBoard
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

CODA sets out to overcome the planner-versus-executor trade-off in GUI agents for scientific software. It pairs Qwen2.5-VL as a trainable planner with UI-TARS-1.5 as a frozen executor, then uses decoupled GRPO to improve only the planner's planning tokens, with rewards supplied by an automated judge instead of human labels. On four ScienceBoard applications the paper reports Average@8 rising from 7.57% for the base planner to 14.39% after specializing per application and 21.04% after the second-stage generalization SFT, with Pass@8 at 39.96%—a new open-source state of the art on that benchmark. The deeper claim is that a fixed execution module can provide a stable grounding interface while the planner adapts through experience, and that a generalist distilled from specialists' successful trajectories can outperform each specialist.

What carries the argument

Decoupled GRPO: a group of G candidate plans from the planner is executed by the frozen executor, each resulting action is scored against a labeled positive action by type-match plus continuous coordinate or bounding-box distance, the scores are normalized into relative advantages, and the GRPO loss is applied only to the planner's planning tokens. The second mechanism is specialist-to-generalist SFT: after four per-application RL specialists are trained, their successful trajectories are aggregated and used as supervised training data for one final planner, transferring cross-application planning knowledge.

What would settle it

Take a random sample of trajectories the judge labeled clean-and-successful and have human annotators score them against the original task goal. If a substantial fraction, near the measured 30% false-positive rate, turn out incomplete or wrong, the RL and SFT training signals are contaminated, and the reported ScienceBoard gains should be re-measured with corrected labels.

Watch

Extended reading notes

Core claim

CODA's load-bearing claim is that you do not need to fine-tune the whole agent to improve a computer-use system: the planner can be reinforced while the executor stays frozen. Stage 1 trains four software-specific planners with decoupled GRPO, scoring each candidate plan by comparing the executor's resulting action with a labeled positive action through an action-type indicator plus a normalized distance term, and backpropagating the advantage only through planning tokens. Stage 2 aggregates 0.77K judge-filtered successful trajectories from these specialists and SFTs a single generalist planner, which the paper reports outperforms the individual specialists (21.04% vs 14.39% Average@8 on Sci

Load-bearing premise

The automated judge's clean-and-successful verdict is a faithful measure of actually completing the task; by the paper's own Table 2, the best judge achieves only 69.5% precision on ScienceBoard, so roughly 30% of the trajectories it certifies as positive training signal are failures by the paper's own measurement.

Editorial extensions

If this is right

  • Adapting a computer-use agent to a new specialized application reduces to RL-tuning the planner; the frozen executor's grounding ability transfers without retraining.
  • The paper's numbers imply that on ScienceBoard the specialist-to-generalist model (21.04% Average@8, 39.96% Pass@8) beats both the base planner (7.57%, 19.49%) and the ensemble of per-application specialists (14.39%, 32.12%).
  • Automatic judging with voting, multi-resolution screenshots, and model ensembling can replace human trajectory labels as the reward source for RL training.
  • Open-source components, a 32B planner plus a 7B executor, can reach the reported benchmark results, so proprietary planner models are not necessary for this level of performance.
  • A generalist trained on aggregated specialist trajectories can outperform its own teachers, supporting the specialist-to-generalist curriculum over direct multi-application RL.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the same judge also selects the Stage-2 SFT trajectories, its 69.5% precision ceiling means part of the reported improvement may be learning to satisfy the judge rather than completing tasks; a more precise judge would likely increase genuine gains, not shrink the method's value.
  • Editorial inference: the framework should transfer to other domains where a mature low-level controller exists, such as robotics, spreadsheet manipulation, or CAD, but the paper only evaluates four ScienceBoard applications and does not test that transfer.
  • A clean experiment the paper does not report: hold one application out of the Stage-2 aggregation and test whether the generalist still improves on it; if it does not, the cross-domain generalization claim is limited to structurally similar workflows.
  • Comparing CODA against a version where the executor is also RL-tuned, with the same compute budget, would isolate how much of the gain comes from decoupling itself.
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

4 major / 6 minor

Summary. The paper proposes CODA, a compositional computer-use agent that pairs a fixed executor (UI-TARS-1.5-7B) with a tunable planner (Qwen2.5-VL-32B). Training is two-stage: (Stage 1) per-application specialization via a decoupled GRPO update that backpropagates through planner tokens only, using rewards derived from an automatic judge system; (Stage 2) aggregation of successful specialist trajectories into a 0.77K-sample SFT corpus to train a generalist planner. The authors report large gains on four ScienceBoard applications, e.g., Average@8 rising from 7.57% (Qwen2.5-VL-32B) and 5.14% (UI-TARS-1.5-7B) to 14.39% (Stage-1 ensemble) and 21.04% (Stage-2). They also present a judge evaluation in Table 2, with the best ensemble reaching 69.5% precision on ScienceBoard trajectories. The paper includes the training prompts, distributed VM infrastructure details, and states that code and models will be released.

Significance. If the empirical claims are correct, CODA is a meaningful contribution: it shows that a frozen specialist executor can be combined with an RL-tuned planner in a data-efficient way, and that a specialist-to-generalist SFT stage can consolidate per-domain gains. The decoupled GRPO design, the use of an automatic judge to supply rewards, and the distributed trajectory-collection system are all useful engineering and methodological contributions. The work is also clearly positioned against non-trainable compositional baselines and releases artifacts. However, the central empirical claim depends heavily on the reliability of the automatic judge and on a clean separation between auto-generated training tasks and the evaluation benchmark, both of which are currently under-specified. The paper therefore presents a promising framework and a credible intended result, but the evidence as written is not yet sufficient to support the headline 'new state of the art' assertion.

major comments (4)
  1. [Sec. 3.4, Sec. 4.2, Table 2, Eq. (1)] The paper's own judge evaluation reports a best precision of only 69.5% on ScienceBoard trajectories. Because Stage-1 GRPO rewards in Eq. (1) are computed against judge-labeled positive actions a_T, and Stage-2 SFT uses 0.77K trajectories filtered by the same judge (Sec. 4.1), roughly 30% of the positive training signal is, by the paper's own metric, a false positive. More importantly, the manuscript does not state how the success rates in Table 1 are obtained. If the same judge is used to decide evaluation success, then both training and the headline metric share the same potentially biased signal, and the reported 21.04% improvement could reflect improved judge-pleasing behavior rather than genuine task completion. Please specify the evaluation protocol (e.g., final file state, human verification, or a separate judge), and provide an analysis or ablation using human-verified or ground-
  2. [Sec. 4.1, Table 1] Training tasks are auto-generated from ScienceBoard (Sec. 4.1), and Table 1 evaluates on ScienceBoard, but the paper never states whether the generated training tasks are disjoint from the evaluation tasks. If any of the auto-generated tasks coincide with evaluation tasks, the model is evaluated on training data, making the comparison against baseline models (which presumably did not train on those tasks) invalid. Please report the overlap between generated and evaluated tasks, and ideally re-evaluate on a held-out subset of ScienceBoard tasks that were not used during auto-generation or training.
  3. [Table 1, note (*)] The Stage-1 row is labeled as 'specialist agents trained separately for each software with ensembled results,' but the ensemble protocol is never defined. It is unclear whether Average@8 and Pass@8 are computed by averaging the four specialists' per-task scores, by taking a majority vote over their actions, or by some other selection method. This makes the Stage-1 result irreproducible and obscures the comparison with Stage-2. In addition, all reported numbers in Table 1 appear to be single runs with no variance estimates; given the stochasticity of GRPO training and the judge's voting procedure, please report at least 3 seeds or multiple evaluation rollouts with standard errors.
  4. [Sec. 3.3.1, Eq. (1)] The reward function depends on an unspecified 'labeled positive action' a_T and on distance metrics whose normalization is only loosely described ('L1 distance for coordinates or IoU for bounding boxes'). The paper does not define how action types are matched, how coordinates are normalized relative to screen size, or how the two reward terms (type match and distance) are combined. This is a reproducibility gap for the central learning signal, and it is also directly tied to the judge-precision concern: since a_T comes from judge-labeled trajectories, any false-positive trajectory provides an incorrect target for RL.
minor comments (6)
  1. [Abstract, Sec. 1] Grammar issues such as 'challenging domains like scientific computing, require' should be corrected; also 'UI-Tars' vs 'UI-TARS' and 'Qwen2.5VL-32B' vs 'Qwen2.5-VL-32B' are used inconsistently.
  2. [Fig. 2 caption] The caption contains garbled notation: '?(1) ?(1) ?(2) ?(2) ?? ??' should be replaced with proper symbols for planner outputs, executor actions, and rewards.
  3. [Eq. (3)] The GRPO loss expression has formatting issues, and the definition of D_KL(πθ∥πref) appears reversed compared to standard usage. Please check the sign and notation to avoid ambiguity.
  4. [Sec. 4.2] The phrase 'building upon difference description fine-tuning' is not defined at that point; the reader is referred only to a citation. Please explain the fine-tuning objective or direct the reader to Appendix A where this is described.
  5. [Sec. 4.2, last paragraph] The sentence 'With methods proposed in . This judge system provide high quality reward signal...' is truncated and ungrammatical; it should be rewritten.
  6. [Sec. 4.1] The definitions of Average@8 and Pass@8 are not provided; please state explicitly how these metrics are computed in the ScienceBoard benchmark.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the headline result is measured on an external benchmark, and the judge-based training signal is a data-quality concern, not a construction that equates inputs with outputs.

full rationale

The paper's central claim is an empirical success-rate improvement on ScienceBoard (Table 1), and nothing in the described pipeline makes that outcome equal to a fitted input by construction. Training rewards (Eq. 1) and Stage-2 SFT trajectories are labeled by the judge system of Sec. 3.4, but the reported evaluations are on ScienceBoard tasks with success defined by completing the instructed workflow, not by the judge's verdict. The judge's own limitations are disclosed in Table 2 (best ScienceBoard precision 69.5%), which is a real validity threat for the training signal, but it is not circularity: the training signal and the benchmark metric are not the same object. The specialist-to-generalist stage is attributed to the authors' prior work (Sun et al., 2025b), and the judge fine-tuning is 'inspired by' the same paper, but these are method attributions rather than load-bearing proofs; the paper's Stage-1 versus Stage-2 comparison provides independent empirical evidence for the claimed benefit. The auto-generation of training tasks from ScienceBoard without an explicit train/test separation is a possible contamination risk, but it is not an equation-level reduction of the prediction to the input, and the paper does not define evaluation success in terms of its own judge or task generator. Accordingly, no circular step meeting the required evidence standard can be exhibited.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The counts here are honest: the framework contributes a training recipe (decoupled GRPO on the planner) and a distillation step, but every learning signal flows through hand-chosen reward normalization, hyperparameters, judge configurations, and corpus sizes, none of which are accompanied by sensitivity analysis. The axioms are mostly domain assumptions about the executor, the judge, the task generator, and the benchmark, the first three of which are the paper's own artifacts. Cerebrum and Cerebellum are metaphors for existing models; the judge and distributed VM system are engineering artifacts, not new theoretical entities.

free parameters (5)
  • reward distance normalization = normalized to [0,1]; L1 for coordinates, IoU for boxes
    Defines the dense reward in Eq. 1 that drives all Stage-1 RL; the normalization constants are hand-chosen and not reported.
  • GRPO hyperparameters (group size G, clip epsilon, KL beta) = not reported
    Eq. 3 depends on G, epsilon, beta and the reference-policy terms; none of the values are given, yet they determine every planner update.
  • judge voting configuration = T=1.0, top_p=0.6, 4 votes, multi-resolution mixing, 2-model ensemble
    The reward source's precision (Table 2) is tuned through these hand-chosen settings; the best configuration determines the labels used for training.
  • Stage-2 SFT corpus size = 0.77K trajectories
    The generalist planner is distilled from this hand-chosen corpus of judge-filtered trajectories (Sec. 4.1).
  • judge fine-tuning corpus = 4.7K samples, LoRA rank 8, 370 steps
    The judge model itself is built with these hand-chosen quantities (Sup. A); judge outputs are the reward source for the planner.
assumptions (6)
  • standard math The task is a POMDP with visual observations and pyautogui action programs (Sec. 3.1).
    Standard formulation; not controversial.
  • domain assumption The fixed executor (UI-TARS-1.5-7B) provides stable, software-agnostic grounding, so only the planner needs training (Sec. 3.3.1).
    The entire decoupled design and the data-efficiency claim rest on this; the paper supports it with an anecdotal observation, not a controlled comparison.
  • domain assumption Judge verdicts (Correctness, Redundant, First Error Step) are a valid proxy for task success and can label positive actions a_T (Sec. 3.4).
    Load-bearing for both RL rewards and SFT filtering; Table 2 shows 69.5% precision, so the premise is only partially satisfied.
  • domain assumption Auto-generated tasks, bootstrapped from a small seed set of human tasks per software, produce executable, skill-relevant training distributions (Sec. 3.4, Fig. 5).
    Task generation by Qwen2.5-72B is unverified beyond the prompt template; training coverage of the evaluated skills is assumed.
  • domain assumption Specialist-to-generalist SFT distillation improves over the specialists (Sec. 3.3.2).
    Adopted from the authors' own SEAgent preprint (Sun et al., 2025b); reported as observed, not derived.
  • standard math GRPO's group-relative advantage is a valid update rule for the planner (Eq. 2-3).
    Standard algorithm from Shao et al. 2024 and Guo et al. 2025; not a new claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CODA: Coordinating the Cerebrum and Cerebellum for a Dual-Brain Computer Use Agent with Decoupled Reinforcement Learning." pith.science (2026). https://pith.science/paper/J7JZ3ZCQ

@misc{pith2026250820096,
  author       = {Pith},
  title        = {Pith review of: CODA: Coordinating the Cerebrum and Cerebellum for a Dual-Brain Computer Use Agent with Decoupled Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7JZ3ZCQ}},
  note         = {Machine review of arXiv:2508.20096}
}
read the original abstract

Autonomous agents for Graphical User Interfaces (GUIs) face significant challenges in specialized domains such as scientific computing, where both long-horizon planning and precise execution are required. Existing approaches suffer from a trade-off: generalist agents excel at planning but perform poorly in execution, while specialized agents demonstrate the opposite weakness. Recent compositional frameworks attempt to bridge this gap by combining a planner and an actor, but they are typically static and non-trainable, which prevents adaptation from experience. This is a critical limitation given the scarcity of high-quality data in scientific domains. To address these limitations, we introduce CODA, a novel and trainable compositional framework that integrates a generalist planner (Cerebrum) with a specialist executor (Cerebellum), trained via a dedicated two-stage pipeline. In the first stage, Specialization, we apply a decoupled GRPO approach to train an expert planner for each scientific application individually, bootstrapping from a small set of task trajectories. In the second stage, Generalization, we aggregate all successful trajectories from the specialized experts to build a consolidated dataset, which is then used for supervised fine-tuning of the final planner. This equips CODA with both robust execution and cross-domain generalization. Evaluated on four challenging applications from the ScienceBoard benchmark, CODA significantly outperforms baselines and establishes a new state of the art among open-source models.

Figures

Figures reproduced from arXiv: 2508.20096 by the authors.

Figure 1
Figure 1. Overall architecture of the proposed learnable Planner–Executor framework. Anal￾ogous to the relationship between the cerebrum and the cerebellum in the human brain, the Planner (cerebrum) generates high-level thoughts based on the history and screenshots, while the Executor (cerebellum) executes concrete GUI actions accordingly. “cerebellum” (Agashe et al., 2024; 2025). While promising, these pioneering approaches … view at source ↗
Figure 2
Figure 2. Overall training process of the proposed Planner–Executor framework. The Planner generates high-level thoughts based on the history and screenshots, while the Executor executes concrete GUI actions accordingly. During training, the rewards are calculated from a (i) and applied to p (i) to calculate loss. This sequential process induces a state trajectory τ = (s0, s1, . . . , sT ) with the maximum time step T. A task… view at source ↗
Figure 3
Figure 3. Exploration pipeline for training support. 4 EXPERIMENTS 4.1 AGENT PERFORMANCE EVALUATION. Our planner-executor approach is based on Qwen2.5VL-32B (Bai et al., 2025) serve as planner and UI-TARS-1.5-7B (Qin et al., 2025) serve as executor. We use method proposed in Sec.3.4 to gen￾erate high level tasks for each software from ScienceBoard (Sun et al., 2025a). through decoupled reinforcement learning proposed in Sec.3… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Case studies. Qwen2.5-VL-72B Bai et al. (2025) struggles with precise grounding, whereas UI-TARS-1.5 Qin et al. (2025), though specialized, fails to generalize to out-of-distribution software. 4.2 TOWARDS PRECISE JUDGING SYSTEM Our reinforcement learning framework heav…
Figure 5
Figure 5. Figure 5: Detailed prompt for task generation. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Detailed prompt for the judge model. Text in gray all task type based on whether finish given task or answer the question from user. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Detailed prompt for the planner agent. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models

    cs.AI 2026-07 accept novelty 6.0 of 10

    VLM judges of CUA trajectories are systematically lenient; OSReward measures this with human gold, and OS-Shepherd open models close most of the reliability gap cheaply.

Reference graph

Works this paper leans on

70 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Agent s: An open agentic framework that uses computers like a human

    Saaket Agashe, Jiuzhou Han, Shuyu Gan, Jiachen Yang, Ang Li, and Xin Eric Wang. Agent s: An open agentic framework that uses computers like a human. arXiv preprint arXiv:2410.08164, 2024

  2. [2]

    Agent s2: A compositional generalist-specialist framework for computer use agents

    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, 2025

  3. [3]

    Claude computer use

    Anthropic. Claude computer use. 2024. URL https://www.anthropic.com/news/3-5-models-and-computer-use

  4. [4]

    Claude’s extended thinking

    Anthropic. Claude’s extended thinking. 2025. URL https://www.anthropic.com/research/visible-extended-thinking

  5. [5]

    Digirl: Training in-the-wild device-control agents with autonomous reinforcement learning

    Hao Bai, Yifei Zhou, Jiayi Pan, Mert Cemri, Alane Suhr, Sergey Levine, and Aviral Kumar. Digirl: Training in-the-wild device-control agents with autonomous reinforcement learning. Advances in Neural Information Processing Systems, 37: 0 12461--12495, 2024

  6. [6]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

  7. [7]

    Grounding large language models in interactive environments with online reinforcement learning

    Thomas Carta, Cl \'e ment Romac, Thomas Wolf, Sylvain Lamprier, Olivier Sigaud, and Pierre-Yves Oudeyer. Grounding large language models in interactive environments with online reinforcement learning. In International Conference on Machine Learning, pp.\ 3676--3713. PMLR, 2023

  8. [8]

    Bail: Best-action imitation learning for batch deep reinforcement learning

    Xinyue Chen, Zijian Zhou, Zheng Wang, Che Wang, Yanqiu Wu, and Keith Ross. Bail: Best-action imitation learning for batch deep reinforcement learning. Advances in Neural Information Processing Systems, 33: 0 18353--18363, 2020

Show all 70 references
  1. [9]

    Seeclick: Harnessing gui grounding for advanced visual gui agents

    Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024

  2. [10]

    Neuroplasticity

    Vida Demarin and Sandra Morovi \'c . Neuroplasticity. Periodicum biologorum, 116 0 (2): 0 209--211, 2014

  3. [11]

    Mm-ifengine: Towards multimodal instruction following

    Shengyuan Ding, Shenxi Wu, Xiangyu Zhao, Yuhang Zang, Haodong Duan, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Dahua Lin, and Jiaqi Wang. Mm-ifengine: Towards multimodal instruction following. arXiv preprint arXiv:2504.07957, 2025

  4. [12]

    Gemini 2.5 Pro Preview (03-25)

    Google DeepMind . Gemini 2.5 Pro Preview (03-25) . https://deepmind.google/technologies/gemini, 2025

  5. [13]

    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, 2024

  6. [14]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

  7. [15]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

  8. [16]

    Neuroplasticity and rehabilitation

    Mark Hallett. Neuroplasticity and rehabilitation. Journal of rehabilitation research and development, 42 0 (4): 0 R17, 2005

  9. [17]

    Webvoyager: Building an end-to-end web agent with large multimodal models

    Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. Webvoyager: Building an end-to-end web agent with large multimodal models. arXiv preprint arXiv:2401.13919, 2024

  10. [18]

    Cogagent: A visual language model for GUI agents

    Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxuan Zhang, Juanzi Li, Bin Xu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogagent: A visual language model for GUI agents. CoRR, abs/2312.08914, 2023. doi:10.48550/ARXIV.2312.08914. U...

  11. [19]

    Cogagent: A visual language model for gui agents

    Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 14281--14...

  12. [20]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

  13. [21]

    Openrlhf: An easy-to-use, scalable and high-performance rlhf framework

    Jian Hu, Xibin Wu, Zilin Zhu, Xianyu, Weixun Wang, Dehao Zhang, and Yu Cao. Openrlhf: An easy-to-use, scalable and high-performance rlhf framework. arXiv preprint arXiv:2405.11143, 2024 a

  14. [22]

    Os agents: A survey on mllm-based agents for general computing devices use, 2024 b

    Xueyu Hu, Tao Xiong, Biao Yi, Zishu Wei, Ruixuan Xiao, Yurun Chen, Jiasheng Ye, Meiling Tao, Xiangxin Zhou, Ziyu Zhao, et al. Os agents: A survey on mllm-based agents for general computing devices use, 2024 b

  15. [23]

    Mechanisms of motor learning in the cerebellum

    Masao Ito. Mechanisms of motor learning in the cerebellum. Brain research, 886 0 (1-2): 0 237--245, 2000

  16. [24]

    Autowebglm: A large language model-based web navigating agent

    Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, et al. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data...

  17. [25]

    Showui: One vision-language-action model for gui visual agent

    Kevin Qinghong Lin, Linjie Li, Difei Gao, Zhengyuan Yang, Shiwei Wu, Zechen Bai, Weixian Lei, Lijuan Wang, and Mike Zheng Shou. Showui: One vision-language-action model for gui visual agent. arXiv preprint arXiv:2411.17465, 2024

  18. [26]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36: 0 34892--34916, 2023 a

  19. [27]

    Bolaa: Benchmarking and orchestrating llm-augmented autonomous agents

    Zhiwei Liu, Weiran Yao, Jianguo Zhang, Le Xue, Shelby Heinecke, Rithesh Murthy, Yihao Feng, Zeyuan Chen, Juan Carlos Niebles, Devansh Arpit, et al. Bolaa: Benchmarking and orchestrating llm-augmented autonomous agents. arXiv preprint arXiv:2308.05960, 2023 b

  20. [28]

    Agentrewardbench: Evaluating automatic evaluations of web agent trajectories

    Xing Han L \`u , Amirhossein Kazemnejad, Nicholas Meade, Arkil Patel, Dongchan Shin, Alejandra Zambrano, Karolina Sta \'n czak, Peter Shaw, Christopher J Pal, and Siva Reddy. Agentrewardbench: Evaluating automatic evaluations of web agent trajectories. arXiv preprint arXiv:250...

  21. [29]

    Webgpt: Browser-assisted question-answering with human feedback

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021

  22. [30]

    Gui agents: A survey

    Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, et al. Gui agents: A survey. arXiv preprint arXiv:2412.13501, 2024

  23. [31]

    GPT-4 technical report

    OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023. doi:10.48550/ARXIV.2303.08774. URL https://doi.org/10.48550/arXiv.2303.08774

  24. [32]

    Operator

    OpenAI. Operator. 2025. URL https://openai.com/research/operator

  25. [33]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 2...

  26. [34]

    Autonomous evaluation and refinement of digital agents

    Jiayi Pan, Yichi Zhang, Nicholas Tomlin, Yifei Zhou, Sergey Levine, and Alane Suhr. Autonomous evaluation and refinement of digital agents. arXiv preprint arXiv:2404.06474, 2024

  27. [35]

    Agent q: Advanced reasoning and learning for autonomous ai agents

    Pranav Putta, Edmund Mills, Naman Garg, Sumeet Motwani, Chelsea Finn, Divyansh Garg, and Rafael Rafailov. Agent q: Advanced reasoning and learning for autonomous ai agents. arXiv preprint arXiv:2408.07199, 2024

  28. [36]

    Webrl: Training llm web agents via self-evolving online curriculum reinforcement learning

    Zehan Qi, Xiao Liu, Iat Long Iong, Hanyu Lai, Xueqiao Sun, Wenyi Zhao, Yu Yang, Xinyue Yang, Jiadai Sun, Shuntian Yao, et al. Webrl: Training llm web agents via self-evolving online curriculum reinforcement learning. arXiv preprint arXiv:2411.02337, 2024

  29. [37]

    UI-TARS: pioneering automated GUI interaction with native agents

    Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, Wanjun Zhong, Kuanye Li, Jiale Yang, Yu Miao, Woyu Lin, Longxiang Liu, Xu Jiang, Qianli Ma, Jingyu Li, Xiaojun Xiao, Kai Cai, Chuang Li, Yaowei Zheng, C...

  30. [38]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36: 0 53728--53741, 2023

  31. [39]

    High-dimensional continuous control using generalized advantage estimation

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015

  32. [40]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

  33. [41]

    Coact-1: Computer-using agents with coding as actions

    Linxin Song, Yutong Dai, Viraj Prabhu, Jieyu Zhang, Taiwei Shi, Li Li, Junnan Li, Silvio Savarese, Zeyuan Chen, Jieyu Zhao, et al. Coact-1: Computer-using agents with coding as actions. arXiv preprint arXiv:2508.03923, 2025

  34. [42]

    A survey of neural code intelligence: Paradigms, advances and beyond

    Qiushi Sun, Zhirui Chen, Fangzhi Xu, Kanzhi Cheng, Chang Ma, Zhangyue Yin, Jianing Wang, Chengcheng Han, Renyu Zhu, Shuai Yuan, et al. A survey of neural code intelligence: Paradigms, advances and beyond. arXiv preprint arXiv:2403.14734, 2024 a

  35. [43]

    Os-genesis: Automating gui agent trajectory 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 trajectory construction via reverse task synthesis. arXiv preprint arXiv:2412.19723, 2024 b

  36. [44]

    Scienceboard: Evaluating multimodal autonomous agents in realistic scientific workflows

    Qiushi Sun, Zhoumianze Liu, Chang Ma, Zichen Ding, Fangzhi Xu, Zhangyue Yin, Haiteng Zhao, Zhenyu Wu, Kanzhi Cheng, Zhaoyang Liu, et al. Scienceboard: Evaluating multimodal autonomous agents in realistic scientific workflows. arXiv preprint arXiv:2505.19897, 2025 a

  37. [45]

    X-prompt: Towards universal in-context image generation in auto-regressive vision language foundation models, 2024 c

    Zeyi Sun, Ziyang Chu, Pan Zhang, Tong Wu, Xiaoyi Dong, Yuhang Zang, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. X-prompt: Towards universal in-context image generation in auto-regressive vision language foundation models, 2024 c . URL https://arxiv.org/abs/2412.01824

  38. [46]

    Bootstrap3d: Improving 3d content creation with synthetic data

    Zeyi Sun, Tong Wu, Pan Zhang, Yuhang Zang, Xiaoyi Dong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Bootstrap3d: Improving 3d content creation with synthetic data. arXiv e-prints, pp.\ arXiv--2406, 2024 d

  39. [47]

    Seagent: Self-evolving computer use agent with autonomous learning from experience

    Zeyi Sun, Ziyu Liu, Yuhang Zang, Yuhang Cao, Xiaoyi Dong, Tong Wu, Dahua Lin, and Jiaqi Wang. Seagent: Self-evolving computer use agent with autonomous learning from experience. arXiv preprint arXiv:2508.04700, 2025 b

  40. [48]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023

  41. [49]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

  42. [50]

    Voyager: An open-ended embodied agent with large language models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023

  43. [51]

    Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024

  44. [52]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

  45. [53]

    Gui-actor: Coordinate-free visual grounding for gui agents

    Qianhui Wu, Kanzhi Cheng, Rui Yang, Chaoyun Zhang, Jianwei Yang, Huiqiang Jiang, Jian Mu, Baolin Peng, Bo Qiao, Reuben Tan, et al. Gui-actor: Coordinate-free visual grounding for gui agents. arXiv preprint arXiv:2506.03143, 2025

  46. [54]

    Os-copilot: Towards generalist computer agents with self-improvement

    Zhiyong Wu, Chengcheng Han, Zichen Ding, Zhenmin Weng, Zhoumianze Liu, Shunyu Yao, Tao Yu, and Lingpeng Kong. Os-copilot: Towards generalist computer agents with self-improvement. arXiv preprint arXiv:2402.07456, 2024 a

  47. [55]

    Os-atlas: A foundation action model for generalist gui agents

    Zhiyong Wu, Zhenyu Wu, Fangzhi Xu, Yian Wang, Qiushi Sun, Chengyou Jia, Kanzhi Cheng, Zichen Ding, Liheng Chen, Paul Pu Liang, et al. Os-atlas: A foundation action model for generalist gui agents. arXiv preprint arXiv:2410.23218, 2024 b

  48. [56]

    Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments

    Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh J Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing S...

  49. [57]

    Scaling computer-use grounding via user interface decomposition and synthesis

    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, 2025

  50. [58]

    Scalecap: Inference-time scalable image captioning via dual-modality debiasing

    Long Xing, Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jinsong Li, Shuangrui Ding, Weiming Zhang, Nenghai Yu, et al. Scalecap: Inference-time scalable image captioning via dual-modality debiasing. arXiv preprint arXiv:2506.19848, 2025

  51. [59]

    Gpt-4v in wonderland: Large multimodal models for zero-shot smartphone gui navigation

    An Yan, Zhengyuan Yang, Wanrong Zhu, Kevin Lin, Linjie Li, Jianfeng Wang, Jianwei Yang, Yiwu Zhong, Julian McAuley, Jianfeng Gao, et al. Gpt-4v in wonderland: Large multimodal models for zero-shot smartphone gui navigation. arXiv preprint arXiv:2311.07562, 2023

  52. [60]

    Fine-tuning large vision-language models as decision-making agents via reinforcement learning

    Simon Zhai, Hao Bai, Zipeng Lin, Jiayi Pan, Peter Tong, Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma, et al. Fine-tuning large vision-language models as decision-making agents via reinforcement learning. Advances in neural information processing systems, 37: 0 110935...

  53. [61]

    Appagent: Multimodal agents as smartphone users

    Chi Zhang, Zhao Yang, Jiaxuan Liu, Yanda Li, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pp.\ 1--20, 2025

  54. [62]

    Android in the zoo: Chain-of-action-thought for gui agents

    Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. Android in the zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713, 2024

  55. [63]

    Webarena: A realistic web environment for building autonomous agents

    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, 2023

  56. [64]

    Archer: Training language model agents via hierarchical multi-turn rl

    Yifei Zhou, Andrea Zanette, Jiayi Pan, Sergey Levine, and Aviral Kumar. Archer: Training language model agents via hierarchical multi-turn rl. arXiv preprint arXiv:2402.19446, 2024

  57. [65]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025

  58. [66]

    Fine-tuning language models from human preferences

    Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019

  59. [67]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  60. [68]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  61. [69]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  62. [70]

    planner" and an

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.