Elamur Research⌕
☰
home / papers / 2026 / Anchor: Branch-Point Data Generation for GUI Agents

Anchor: Branch-Point Data Generation for GUI Agents

Paper · 2026 · Jinbiao Wei, Yilun Zhao, Kangqi Ni et al.

Summary

Training end-to-end GUI agents for desktop operating systems requires a lot of interaction data. Human demonstrations are costly, and automatic pipelines tend to produce either narrow task sets or noisy trajectories that drift from the goal. Anchor addresses this by growing a large training corpus from a small number of human-verified seed trajectories collected in OSWorld (Ubuntu) and WindowsAgentArena environments.

Along each seed, the pipeline uses GPT-5.1 to locate branch points, which are steps where the interface changes substantially or reveals new content. From these states it proposes new tasks grounded in what is on screen. A GUI agent then executes each task while the task description is revised if execution drifts. A summarizer rewrites the final task, a verifier checks completion, and step-level filtering plus intent-consistency denoising clean the resulting trajectories.

Fine-tuning several open VLMs on this data improves results on OSWorld and WindowsAgentArena over zero-shot agents and other synthesis baselines, including across applications and operating systems. This makes affordable, high-signal desktop supervision feasible.

Key points

  • Seeds come from 117 OSWorld and 51 WindowsAgentArena task environments. Each has one human-checked, shortest successful trajectory, and each seed is expanded at 3–5 branch points.
  • With 1K Ubuntu trajectories, OSWorld success rates rise as follows:
    • Qwen2.5-VL-7B: from 0.93 to 7.94.
    • GLM-4.1V-9B: from 0.47 to 7.01.
    • Qwen3-VL-8B: from 16.82 to 20.56, after a temporary dip at 0.3K.
  • Adding 600 WindowsAgentArena trajectories to the Ubuntu set lifts Qwen2.5-VL-7B from 7.94 to 9.95, so there is no negative cross-platform transfer.
  • Removing step-level filtering and post-branch denoising lowers scores for all backbones (e.g., 20.56 → 19.15 and 7.94 → 7.01).
  • Data, models and code are released as yale-nlp/Anchor.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of North Carolina at Chapel HillChapel Hill · United States67 works · 16,474 citations🎓Yale UniversityNew Haven · United States95 works · 12,765 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSoftware · 92 works

Cite

Cite as Jinbiao Wei, Yilun Zhao, Kangqi Ni and Arman Cohan. 2026. Anchor: Branch-Point Data Generation for GUI Agents. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.774

@inproceedings{wei2026anchor,
  title = {{Anchor: Branch-Point Data Generation for GUI Agents}},
  author = {Jinbiao Wei, Yilun Zhao, Kangqi Ni and Arman Cohan},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.774},
  url = {https://doi.org/10.18653/v1/2026.acl-long.774}
}

References

29 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Reyna Abhyankar, Qi Qi, and Yiying Zhang. 2025. Osworld-human: Benchmarking the efficiency of computer-use agents. arXiv preprint arXiv:2506.16042. — [arXiv:2506.16042](https://arxiv.org/abs/2506.16042) find
  2. 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. In The Thirteenth International Conference on Learning Representations. find
  3. Anthropic. 2024. Developing a computer use model. https://www.anthropic.com/news/ developing-computer-use. Accessed: 2025-12-19. ↗
  4. Anthropic. 2025. Introducing claude sonnet 4.5. https://www.anthropic.com/news/ claude-sonnet-4-5. Accessed: 2025-12-19. ↗
  5. Shuai Bai and 1 others. 2025a. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923.
  6. Shuai Bai and 1 others. 2025b. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631.
  7. Rogerio Bonatti, Dan Zhao, Francesco Bonacci, Dillon Dupont, Sara Abdali, Yinheng Li, Yadong Lu, Justin Wagle, Kazuhito Koishida, Arthur Bucker, and 1 others. 2024. Windows agent arena: Evaluating multi-modal os agents at scale. arXiv preprint arXiv:2409.08264. — [arXiv:2409.08264](https://arxiv.org/abs/2409.08264) find
  8. Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track.
  9. Guo Gan, Yuxuan Ding, Cong Chen, Yuwei Ren, Yin Huang, and Hong Zhou. 2026. Android coach: Improve online agentic training efficiency with single state multiple actions. Preprint, arXiv:2604.07277.
  10. Wenbo Hong, Wenhao Yu, and 1 others. 2025. GLM-4.1V-Thinking and GLM-4.5V: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning. arXiv preprint arXiv:2507.01006.
  11. Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, and Jie Tang. 2024. Autowebglm: A large language model-based web navigating agent. Preprint, arXiv:2404.03648.
  12. Wei Li, William E Bishop, Alice Li, Christopher Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. 2024. On the effects of data scale on ui control agents. Advances in Neural Information Processing Systems, 37:92130–92154. find
  13. Musen Lin, Minghao Liu, Taoran Lu, Lichen Yuan, Yiwei Liu, Haonan Xu, Yu Miao, Yuhao Chao, and Zhaojian Li. 2025. Gui-rewalk: Massive data generation for gui agent via stochastic exploration and intent-aware reasoning. Preprint, arXiv:2509.15738. — [arXiv:2509.15738](https://arxiv.org/abs/2509.15738) find
  14. Xing Han Lù, Zdeněk Kasner, and Siva Reddy. 2024. Weblinx: Real-world website navigation with multi-turn dialogue. arXiv preprint arXiv:2402.05930.
  15. OpenAI. 2025. Gpt-5.1: A smarter, more conversational chatgpt. https://openai.com/index/gpt-5-1/. Accessed: 2025-12-19. ↗
  16. Vardaan Pahuja, Yadong Lu, Corby Rosset, Boyu Gou, Arindam Mitra, Spencer Whitehead, Yu Su, and Ahmed Hassan. 2025. Explorer: Scaling exploration-driven web trajectory synthesis for multimodal web agents. In Findings of the Association for Computational Linguistics: ACL 2025, pages 6300–6323. find
  17. Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, and 1 others. 2025. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326.
  18. Ram Ramrakhya, Andrew Szot, Omar Attia, Yuhao Yang, Anh Nguyen, Bogdan Mazoure, Zhe Gan, Harsh Agrawal, and Alexander Toshev. 2025. Scaling synthetic task generation for agents via exploration. arXiv preprint arXiv:2509.25047. — [arXiv:2509.25047](https://arxiv.org/abs/2509.25047) find
  19. Christopher Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, Folawiyo Campbell-Ajala, and 1 others. 2024. Androidworld: A dynamic benchmarking environment for autonomous agents. arXiv preprint arXiv:2405.14573. find
  20. Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy P Lillicrap. Androidinthewild: A large-scale dataset for android device control. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track. find
  21. Yu Su, Diyi Yang, Shunyu Yao, and Tao Yu. 2024. Language agents: Foundations, prospects, and risks. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Tutorial Abstracts, pages 17–24, Miami, Florida, USA. Association for Computational Linguistics. find
  22. Qiushi Sun, Kanzhi Cheng, Zichen Ding, Chuanyang Jin, Yian Wang, Fangzhi Xu, Zhenyu Wu, Chengyou Jia, Liheng Chen, Zhoumianze Liu, and 1 others. 2025. Os-genesis: Automating gui agent trajectory construction via reverse task synthesis. In Proceedings of the 63rd Annual Meeting of the Association for…
  23. Xinyuan Wang, Bowen Wang, Dunjie Lu, Junlin Yang, Tianbao Xie, Junli Wang, Jiaqi Deng, Xiaole Guo, Yiheng Xu, Chen Henry Wu, and 1 others. Opencua: Open foundations for computer-use agents. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. Zhiyong Wu, Chengcheng Han, Zi… find
  24. Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh J Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, and 1 others. 2024. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing Systems, 37…
  25. Yiheng Xu, Dunjie Lu, Zhennan Shen, Junli Wang, Zekun Wang, Yuchen Mao, Caiming Xiong, and Tao Yu. 2024. Agenttrek: Agent trajectory synthesis via guiding replay with web tutorials. arXiv preprint arXiv:2412.09605. — [arXiv:2412.09605](https://arxiv.org/abs/2412.09605) find
  26. An Yan, Zhengyuan Yang, Wanrong Zhu, Kevin Lin, Linjie Li, Jianfeng Wang, Jianwei Yang, Yiwu Zhong, Julian J McAuley, Jianfeng Gao, and 1 others. 2023. Gpt-4v in wonderland: Large multimodal models for zero-shot smartphone gui navigation. CoRR. find
  27. Bofei Zhang, Zirui Shang, Zhi Gao, Wang Zhang, Rui Xie, Xiaojian Ma, Tao Yuan, Xinxiao Wu, Song-Chun Zhu, and Qing Li. 2025. Tongui: Building generalized gui agents by learning from multimodal web tutorials. arXiv preprint arXiv:2504.12679. — [arXiv:2504.12679](https://arxiv.org/abs/2504.12679) find
  28. Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. 2024. Gpt-4v (ision) is a generalist web agent, if grounded. In International Conference on Machine Learning, pages 61349–61385. PMLR.
  29. Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, and 1 others. 2023. Webarena: A realistic web environment for building autonomous agents. In NeurIPS 2023 Foundation Models for Decision Making Workshop.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wei-2026-anchor-branch-point-data-generation/