pith. sign in

arxiv: 2606.18890 · v1 · pith:QMBPQW77new · submitted 2026-06-17 · 💻 cs.AI

Skill-Guided Continuation Distillation for GUI Agents

Zhimin Fan , Hongwei Yu , Yeqing Shen , Haolong Yan , Guozhen Peng , Tianhao Peng , Yudong Zhang , Xiaowen Zhang
show 4 more authors
Kaijun Tan Zheng Ge Xiangyu Zhang Daxin Jiang
This is my paper

Pith reviewed 2026-06-26 21:08 UTC · model grok-4.3

classification 💻 cs.AI
keywords GUI agentsbehavior cloningoff-trajectory statesself-improvementdistillationOSWorldskill guidancecontinuation plans
0
0 comments X

The pith

Skill-guided continuation distillation closes the supervision gap for GUI agents by generating successful trajectories from off-trajectory states.

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

Current GUI agents trained via behavior cloning on expert trajectories fail to act correctly on states outside those trajectories, which arise naturally during closed-loop execution. The paper proposes Skill-Guided Continuation Distillation to fix this by letting the plain policy reach such states, then using a skill-guided policy to finish the task and create usable continuations. These new trajectories are mixed with the original expert data so the policy receives supervision precisely where it was missing. The extracted skills—Continuation Plans, Critical Targets, Failure Traps, and Success Criteria—come from both successful and failed rollouts. On OSWorld-Verified this raises success rates for three base models from the low-30 percent range to over 50 percent.

Core claim

SGCD first runs the plain policy for a few steps to reach realistic policy-induced off-trajectory states. From those states a skill-guided policy then completes the task, producing successful continuations that are combined with expert trajectories to supply supervision over previously unseen states. Skills extracted from rollouts consist of Continuation Plans, Critical Targets, Failure Traps, and Success Criteria.

What carries the argument

Skill-Guided Continuation Distillation, an iterative self-improvement loop that mixes skill-guided completions from off-trajectory states with expert trajectories.

If this is right

  • Success rates on OSWorld-Verified rise from the low-30 percent range to over 50 percent across three base models.
  • The policy receives effective supervision on states that expert trajectories alone cannot cover.
  • Self-improvement occurs iteratively without requiring new expert demonstrations.
  • Skills extracted from rollouts enable the policy to avoid failure traps and meet success criteria.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same continuation-distillation pattern could reduce dependence on expert data in other imitation-learning domains that suffer distribution shift.
  • Extracted skills might transfer across tasks or models if stored separately from the policy weights.
  • The approach suggests a general recipe for turning execution failures into training signals whenever a stronger guided policy is available.

Load-bearing premise

The skill-guided policy can reliably produce successful continuations from the off-trajectory states reached by the plain policy.

What would settle it

Applying SGCD to the three base models on OSWorld-Verified and measuring no increase in success rate above the original low-30 percent range would show the method does not close the supervision gap.

Figures

Figures reproduced from arXiv: 2606.18890 by Daxin Jiang, Guozhen Peng, Haolong Yan, Hongwei Yu, Kaijun Tan, Tianhao Peng, Xiangyu Zhang, Xiaowen Zhang, Yeqing Shen, Yudong Zhang, Zheng Ge, Zhimin Fan.

Figure 1
Figure 1. Figure 1: Failure analysis of GUI agents. Left: failures are concentrated in early execution. Right: representative [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SGCD. (1) Task Trajectories Sampling: collect successful and failed plain-policy trajectories. (2) Skill Construction: extract off-trajectory continuation skills from trajectory evidence. (3) Off-trajectory Continuation Construction: use k-step handoff to collect skill-guided successful continuations. (4) Mixed Trajectories Training: train the plain policy with expert and verified continuation … view at source ↗
Figure 3
Figure 3. Figure 3: Iterative SGCD across training rounds [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Skill-extraction prompt used with Gemini [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Skill extracted for an office-document editing [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Skill extracted for a web browsing task on a [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Skill extracted for a browser-settings task. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: System prompt used at evaluation. Thinking [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
read the original abstract

Improving GUI agents typically relies on behavior cloning on expert trajectories. However, as the current policy deviates from the expert policy, it inevitably encounters policy-induced off-trajectory states during closed-loop execution, i.e., states that fall outside the expert trajectories. Since expert trajectories provide no demonstrations for these unseen states, such states receive no effective supervision, leaving the policy unable to select the correct action. To close this supervision gap, we propose Skill-Guided Continuation Distillation (SGCD), an iterative self-improvement framework. SGCD first runs the plain policy without skill guidance for a few steps to reach realistic off-trajectory states. From these states, a skill-guided policy then completes the task and produces successful continuations, which are mixed with expert trajectories to supply supervision over policy-induced off-trajectory states. The skills are extracted from both successful and failed rollouts, consisting of Continuation Plans, Critical Targets, Failure Traps, and Success Criteria. On OSWorld-Verified, SGCD improves the success rate of three base models from the low-30\% range to over 50\%, demonstrating its effectiveness and generality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes Skill-Guided Continuation Distillation (SGCD), an iterative self-improvement framework for GUI agents. It rolls out the current policy for a few steps to induce realistic off-trajectory states, extracts skills (Continuation Plans, Critical Targets, Failure Traps, Success Criteria) from prior successful and failed rollouts, uses a skill-guided policy to generate successful continuations from those states, and mixes the resulting trajectories with expert data to provide supervision for policy-induced states. The central empirical claim is that SGCD raises success rates on OSWorld-Verified from the low-30% range to over 50% across three base models.

Significance. If the results and ablations hold, the work addresses a practically important supervision gap in closed-loop GUI agent execution and offers a concrete mechanism for leveraging both successes and failures to generate corrective data. The explicit extraction of structured skills from rollouts is a clear methodological contribution that could generalize beyond the reported setting.

major comments (3)
  1. [Abstract / Experiments] Abstract and Experiments section: The reported improvement (low-30% to >50% on OSWorld-Verified) is stated without any description of the base models, number of evaluation runs, statistical tests, or comparison against standard behavior-cloning or self-improvement baselines, preventing assessment of whether the gain is attributable to SGCD.
  2. [Method / Experiments] Method (iterative loop description) and Experiments: No quantitative results are supplied showing that the skill-guided policy achieves higher task-completion rates than the plain policy when started from the specific off-trajectory states reached by rolling out the plain policy. Without this measurement, it is impossible to confirm that the mixed trajectories supply corrective supervision rather than redundant or noisy data.
  3. [Experiments] Ablation studies: The manuscript provides no ablation that isolates the contribution of the skill-guided continuations versus other factors (e.g., simply increasing the amount of expert data or using random continuations), which is required to substantiate the claim that the skill-extraction mechanism is load-bearing for the observed gains.
minor comments (2)
  1. [Method] The four skill categories (Continuation Plans, Critical Targets, Failure Traps, Success Criteria) are introduced narratively; formalizing their extraction and representation with pseudocode or a short algorithm box would improve reproducibility.
  2. [Figures / Tables] Figure captions and table headers should explicitly state the evaluation metric (success rate) and the exact OSWorld-Verified split used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the constructive review and the recommendation for major revision. We appreciate the emphasis on providing clearer experimental details, direct measurements of the skill-guided policy's effectiveness on off-trajectory states, and targeted ablations. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation of results and evidence.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and Experiments section: The reported improvement (low-30% to >50% on OSWorld-Verified) is stated without any description of the base models, number of evaluation runs, statistical tests, or comparison against standard behavior-cloning or self-improvement baselines, preventing assessment of whether the gain is attributable to SGCD.

    Authors: We agree that additional details are required for full assessment. In the revised manuscript we will name the three base models, specify the number of evaluation runs (with error bars), report any statistical tests performed, and include direct comparisons against behavior cloning on expert trajectories as well as other self-improvement baselines such as iterative fine-tuning without skill guidance. These additions will clarify that the observed gains are attributable to SGCD. revision: yes

  2. Referee: [Method / Experiments] Method (iterative loop description) and Experiments: No quantitative results are supplied showing that the skill-guided policy achieves higher task-completion rates than the plain policy when started from the specific off-trajectory states reached by rolling out the plain policy. Without this measurement, it is impossible to confirm that the mixed trajectories supply corrective supervision rather than redundant or noisy data.

    Authors: This observation is correct; the current manuscript does not include a direct head-to-head evaluation of the skill-guided versus plain policy on the exact off-trajectory states induced by the plain policy. In the revision we will add a targeted experiment that collects such states, evaluates both policies from those states, and reports the resulting task-completion rates. This will provide direct evidence that the skill-guided continuations supply corrective rather than redundant supervision. revision: yes

  3. Referee: [Experiments] Ablation studies: The manuscript provides no ablation that isolates the contribution of the skill-guided continuations versus other factors (e.g., simply increasing the amount of expert data or using random continuations), which is required to substantiate the claim that the skill-extraction mechanism is load-bearing for the observed gains.

    Authors: We acknowledge that the manuscript lacks ablations isolating the skill-guided mechanism from volume or randomness effects. The revised version will include new ablation experiments that (i) match the volume of added trajectories using only expert data, (ii) replace skill-guided continuations with random or unguided ones, and (iii) ablate individual extracted skills. These results will demonstrate the load-bearing role of the skill-extraction and guidance components. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method with external benchmark validation

full rationale

The paper describes an iterative procedure for generating corrective trajectories via skill-guided continuation from off-trajectory states reached by the base policy, then mixing them with expert data for distillation. The central result is an empirical success-rate lift on the external OSWorld-Verified benchmark. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the provided text. The derivation chain is self-contained against the benchmark and does not reduce the reported gains to a tautology by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Ledger constructed from abstract alone; full paper may contain additional fitted parameters or unstated assumptions.

axioms (1)
  • domain assumption Expert trajectories provide no demonstrations for policy-induced off-trajectory states
    Explicitly stated as the core supervision gap the method targets.
invented entities (1)
  • Continuation Plans, Critical Targets, Failure Traps, Success Criteria no independent evidence
    purpose: Skills extracted from rollouts to guide continuation from off-trajectory states
    Introduced as the components of the skill-guided policy

pith-pipeline@v0.9.1-grok · 5760 in / 1157 out tokens · 19977 ms · 2026-06-26T21:08:52.572999+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

68 extracted references · 1 canonical work pages

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    2024 , month = oct, howpublished =

    Introducing Computer Use, a New Claude 3.5 Sonnet, and Claude 3.5 Haiku , author =. 2024 , month = oct, howpublished =

    2024

  9. [9]

    2024 , howpublished =

    Computer Use Tool , author =. 2024 , howpublished =

    2024

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    Mind2web: Towards a generalist agent for the web , author=. Advances in Neural Information Processing Systems , volume=

  11. [11]

    International Conference on Learning Representations , volume=

    Webarena: A realistic web environment for building autonomous agents , author=. International Conference on Learning Representations , volume=

  12. [12]

    arXiv preprint arXiv:2401.01614 , year=

    Gpt-4v (ision) is a generalist web agent, if grounded , author=. arXiv preprint arXiv:2401.01614 , year=

    Pith/arXiv arXiv

  13. [13]

    International Conference on Learning Representations , volume=

    Androidworld: A dynamic benchmarking environment for autonomous agents , author=. International Conference on Learning Representations , volume=

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    arXiv preprint arXiv:2508.15144 , year=

    Mobile-agent-v3: Fundamental agents for gui automation , author=. arXiv preprint arXiv:2508.15144 , year=

    Pith/arXiv arXiv

  16. [16]

    5: Multi-platform fundamental gui agents , author=

    Mobile-agent-v3. 5: Multi-platform fundamental gui agents , author=. arXiv preprint arXiv:2602.16855 , year=

    arXiv

  17. [17]

    2025 , month = nov, howpublished =

    2025

  18. [18]

    Proceedings of the fourteenth international conference on artificial intelligence and statistics , pages=

    A reduction of imitation learning and structured prediction to no-regret online learning , author=. Proceedings of the fourteenth international conference on artificial intelligence and statistics , pages=. 2011 , organization=

    2011

  19. [19]

    International Conference on Learning Representations , volume=

    On-policy distillation of language models: Learning from self-generated mistakes , author=. International Conference on Learning Representations , volume=

  20. [20]

    Thinking Machines Lab: Connectionism , year =

    Kevin Lu and Thinking Machines Lab , title =. Thinking Machines Lab: Connectionism , year =

  21. [21]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  22. [22]

    Advances in neural information processing systems , volume=

    Self-refine: Iterative refinement with self-feedback , author=. Advances in neural information processing systems , volume=

  23. [23]

    International Conference on Learning Representations , volume=

    Agent s: An open agentic framework that uses computers like a human , author=. International Conference on Learning Representations , volume=

  24. [24]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  25. [25]

    Advances in Neural Information Processing Systems , volume=

    Gui-reflection: Empowering multimodal gui models with self-reflection behavior , author=. Advances in Neural Information Processing Systems , volume=

  26. [26]

    2025 , month = nov, howpublished =

    A New Era of Intelligence with. 2025 , month = nov, howpublished =

    2025

  27. [27]

    2026 , month = mar, howpublished =

    Introducing. 2026 , month = mar, howpublished =

    2026

  28. [28]

    arXiv preprint arXiv:2501.12326 , year=

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

    Pith/arXiv arXiv

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    Opencua: Open foundations for computer-use agents , author=. Advances in Neural Information Processing Systems , volume=

  30. [30]

    arXiv preprint arXiv:2512.15431 , year=

    Step-gui technical report , author=. arXiv preprint arXiv:2512.15431 , year=

    arXiv

  31. [31]

    arXiv preprint arXiv:2601.15876 , year=

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

    arXiv

  32. [32]

    arXiv preprint arXiv:2605.07505 , year=

    LiteGUI: Distilling Compact GUI Agents with Reinforcement Learning , author=. arXiv preprint arXiv:2605.07505 , year=

    Pith/arXiv arXiv

  33. [33]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Screenspot-pro: Gui grounding for professional high-resolution computer use , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  34. [34]

    Advances in Neural Information Processing Systems , volume=

    Scaling computer-use grounding via user interface decomposition and synthesis , author=. Advances in Neural Information Processing Systems , volume=

  35. [35]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Visual test-time scaling for gui agent grounding , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  36. [36]

    Advances in Neural Information Processing Systems , volume=

    Gui-g1: Understanding r1-zero-like training for visual grounding in gui agents , author=. Advances in Neural Information Processing Systems , volume=

  37. [37]

    arXiv preprint arXiv:2507.22025 , year=

    Ui-agile: Advancing gui agents with effective reinforcement learning and precise inference-time grounding , author=. arXiv preprint arXiv:2507.22025 , year=

    Pith/arXiv arXiv

  38. [38]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    GUI-G ^2 : Gaussian Reward Modeling for GUI Grounding , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  39. [39]

    2026 , month = oct, howpublished =

    Introducing Claude Sonnet 4.6 , author =. 2026 , month = oct, howpublished =

    2026

  40. [40]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  41. [41]

    International Conference on Learning Representations , volume=

    Learn-by-interact: A data-centric framework for self-adaptive agents in realistic environments , author=. International Conference on Learning Representations , volume=

  42. [42]

    International Conference on Learning Representations , volume=

    Agenttrek: Agent trajectory synthesis via guiding replay with web tutorials , author=. International Conference on Learning Representations , volume=

  43. [43]

    arXiv preprint arXiv:2508.14040 , year=

    Computerrl: Scaling end-to-end online reinforcement learning for computer use agents , author=. arXiv preprint arXiv:2508.14040 , year=

    arXiv

  44. [44]

    arXiv preprint arXiv:2512.14895 , year=

    Imitation Learning for Multi-turn LM Agents via On-policy Expert Corrections , author=. arXiv preprint arXiv:2512.14895 , year=

    arXiv

  45. [45]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Seeclick: Harnessing gui grounding for advanced visual gui agents , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  46. [46]

    International Conference on Learning Representations , volume=

    Navigating the digital world as humans do: Universal visual grounding for gui agents , author=. International Conference on Learning Representations , volume=

  47. [47]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Showui: One vision-language-action model for gui visual agent , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  48. [48]

    Advances in Neural Information Processing Systems , volume=

    Look before you leap: A gui-critic-r1 model for pre-operative error diagnosis in gui automation , author=. Advances in Neural Information Processing Systems , volume=

  49. [49]

    arXiv preprint arXiv:2504.08942 , year=

    Agentrewardbench: Evaluating automatic evaluations of web agent trajectories , author=. arXiv preprint arXiv:2504.08942 , year=

    arXiv

  50. [50]

    Advances in neural information processing systems , volume=

    Judging llm-as-a-judge with mt-bench and chatbot arena , author=. Advances in neural information processing systems , volume=

  51. [51]

    arXiv preprint arXiv:2505.20023 , year=

    Training LLM-Based Agents with Synthetic Self-Reflected Trajectories and Partial Masking , author=. arXiv preprint arXiv:2505.20023 , year=

    arXiv

  52. [53]

    arXiv preprint arXiv:2509.17336 , year=

    Mano Technical Report , author=. arXiv preprint arXiv:2509.17336 , year=

    arXiv

  53. [54]

    arXiv preprint arXiv:2504.00906 , year=

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

    Pith/arXiv arXiv

  54. [55]

    arXiv preprint arXiv:2510.02250 , year=

    The unreasonable effectiveness of scaling agents for computer use , author=. arXiv preprint arXiv:2510.02250 , year=

    arXiv

  55. [56]

    arXiv preprint arXiv:2510.03853 , year=

    Uground: Towards unified visual grounding with unrolled transformers , author=. arXiv preprint arXiv:2510.03853 , year=

    Pith/arXiv arXiv

  56. [57]

    arXiv preprint arXiv:2505.11821 , year=

    Reinforcing Multi-Turn Reasoning in LLM Agents via Turn-Level Credit Assignment , author=. arXiv preprint arXiv:2505.11821 , year=

    arXiv

  57. [58]

    The Fourteenth International Conference on Learning Representations , year=

    Information Gain-based Policy Optimization: A Simple and Effective Approach for Multi-Turn Search Agents , author=. The Fourteenth International Conference on Learning Representations , year=

  58. [59]

    arXiv preprint arXiv:2603.24533 , year=

    Ui-voyager: A self-evolving gui agent learning via failed experience , author=. arXiv preprint arXiv:2603.24533 , year=

    arXiv

  59. [60]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Swift: a scalable lightweight infrastructure for fine-tuning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  60. [61]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Os-genesis: Automating gui agent trajectory construction via reverse task synthesis , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  61. [62]

    International Conference on Learning Representations , volume=

    OS-ATLAS: Foundation action model for generalist GUI agents , author=. International Conference on Learning Representations , volume=

  62. [63]

    arXiv preprint arXiv:2503.23434 , year=

    Towards trustworthy gui agents: A survey , author=. arXiv preprint arXiv:2503.23434 , year=

    arXiv

  63. [64]

    arXiv preprint arXiv:2509.23866 , year=

    Efficient Multi-turn RL for GUI Agents via Decoupled Training and Adaptive Data Curation , author=. arXiv preprint arXiv:2509.23866 , year=

    arXiv

  64. [65]

    5: Visual Agentic Intelligence , author=

    Kimi K2. 5: Visual Agentic Intelligence , author=. arXiv preprint arXiv:2602.02276 , year=

    Pith/arXiv arXiv

  65. [66]

    8 model card: Towards generalized real-world agency , author=

    Seed1. 8 model card: Towards generalized real-world agency , author=. arXiv preprint arXiv:2603.20633 , year=

    Pith/arXiv arXiv

  66. [67]

    arXiv preprint arXiv:2601.09668 , year=

    Step3-vl-10b technical report , author=. arXiv preprint arXiv:2601.09668 , year=

    arXiv

  67. [68]

    arXiv preprint arXiv:2505.21964 , year=

    UI-Evol: Automatic Knowledge Evolving for Computer Use Agents , author=. arXiv preprint arXiv:2505.21964 , year=

    arXiv

  68. [69]

    arXiv preprint arXiv:2601.05787 , year=

    From Off-Policy to On-Policy: Enhancing GUI Agents via Bi-level Expert-to-Policy Assimilation , author=. arXiv preprint arXiv:2601.05787 , year=

    arXiv