Elamur Research⌕
☰
home / papers / 2026 / WebRollback: Enhancing Web Agents with Explicit Rollback Mechanisms

WebRollback: Enhancing Web Agents with Explicit Rollback Mechanisms

Paper · 2026 · Zhisong Zhang, Tianqing Fang, Kaixin Ma et al.

Summary

Most LLM-based web agents navigate greedily in a single direction. On live, changing websites they can enter a wrong state and waste many steps before recovering, if they recover at all. Parallel search methods are costly here because each branch needs its own browser environment. Best-first search, as in Koh et al., keeps the search serial but tends to jump between states too often. WebRollback adds two model-driven modules to a standard agent. A critique module checks each new state and decides whether to keep going or revert. A rollback module then picks which earlier step to return to, so the agent can undo several steps in one move. The authors test the approach on Mind2Web-Live and WebVoyager, both zero-shot with large open and closed LLMs and with smaller models distilled from a larger teacher. WebRollback gives the best overall success scores and needs fewer state switches than BestFirst. This suggests that letting the agent decide when and where to backtrack makes better use of extra search budget, which matters for deploying agents on real websites.

Key points

  • Setup: The agent works in a text-based environment where pages are represented as accessibility trees and the browser is driven by Playwright.
    • Rollback is done by reloading the URL recorded for the chosen step.
    • The default budget is 16 steps.
    • Results are averaged over three runs.
    • GPT-4o acts as the judge and assigns scores between 0.0 and 1.0, reported as Full% and Partial%.
  • Zero-shot comparison: Llama-3.3-70B-Instruct and Qwen2.5-72B-Instruct were tested, plus GPT-4o-mini and GPT-4o on a 100-task subset.
    • For Qwen2.5-72B, the Full% score rose from 24.53 (OneWay) to 27.36 on one benchmark and from 49.56 to 51.90 on the other.
    • Rollback switched states less often than BestFirst, for example 5.0 vs 8.1 switches with Llama.
  • Fine-tuning: Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct were distilled from Qwen2.5-72B-Instruct.
    • Training data came from the Mind2Web-Live train split and OpenWebVoyager imitation data.
    • The relative ranking of the strategies stayed the same as in the zero-shot setting.
  • Analysis:
    • With larger step budgets, the strategies that can roll back improve faster than OneWay.
    • The share of tasks needing recovery (the "struggle-ratio") was 19% for OneWay vs 7% for Rollback.
  • Limitations:
    • URL-based rollback cannot reverse irreversible actions such as payments or sending emails.
    • Only models of around 7B parameters were fine-tuned.

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

🏢TencentShenzhen · China134 works · 19,743 citations🎓City University of Hong KongHong Kong · China45 works · 3,014 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 works

Cite

Cite as Zhisong Zhang, Tianqing Fang, Kaixin Ma, Wenhao Yu, Hongming Zhang, Haitao Mi and Dong Yu. 2026. WebRollback: Enhancing Web Agents with Explicit Rollback Mechanisms. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-short.12

@inproceedings{zhang2026webrollback,
  title = {{WebRollback: Enhancing Web Agents with Explicit Rollback Mechanisms}},
  author = {Zhisong Zhang, Tianqing Fang, Kaixin Ma, Wenhao Yu, Hongming Zhang, Haitao Mi and Dong Yu},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-short.12},
  url = {https://doi.org/10.18653/v1/2026.eacl-short.12}
}

References

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

  1. Ziru Chen, Michael White, Ray Mooney, Ali Payani, Yu Su, and Huan Sun. 2024. When is tree search useful for LLM planning? it depends on the discriminator. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13659–13678, Bangkok, T…
  2. Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. 2023. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36:28091–28114.
  3. Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. 2024a. WebVoyager: Building an end-to-end web agent with large multimodal models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),…
  4. Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Hongming Zhang, Tianqing Fang, Zhenzhong Lan, and Dong Yu. 2024b. Openwebvoyager: Building multimodal web agents via iterative real-world exploration, feedback and optimization. arXiv preprint arXiv:2410.19609.
  5. Jian Hu, Xibin Wu, Zilin Zhu, Weixun Wang, Dehao Zhang, Yu Cao, and 1 others. 2024. Openrlhf: An easy-to-use, scalable and high-performance rlhf framework. arXiv preprint arXiv:2405.11143.
  6. Jing Yu Koh, Stephen McAleer, Daniel Fried, and Ruslan Salakhutdinov. 2024. Tree search for language model agents. arXiv preprint arXiv:2407.01476. find
  7. Xingzuo Li, Kehai Chen, Yunfei Long, Xuefeng Bai, Yong Xu, and Min Zhang. 2025. Generator-assistant stepwise rollback framework for large language model agent. arXiv preprint arXiv:2503.02519. find
  8. Michael Lutz, Arth Bohra, Manvel Saroyan, Artem Harutyunyan, and Giovanni Campagna. 2024. Wilbur: Adaptive in-context learning for robust and accurate web agents. arXiv preprint arXiv:2404.05902. — [arXiv:2404.05902](https://arxiv.org/abs/2404.05902) find
  9. Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314.
  10. Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In The Thirteenth International Conference on Learning Representations. find
  11. Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, and 1 others. 2024. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345.
  12. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations.
  13. Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2023. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems, 36:41618–41650.
  14. Xiao-Wen Yang, Xuan-Yi Zhu, Wen-Da Wei, Ding-Chu Zhang, Jie-Jing Shao, Zhi Zhou, Lan-Zhe Guo, and Yu-Feng Li. 2025. Step back to leap forward: Self-backtracking for boosting reasoning of language models. arXiv preprint arXiv:2502.04404. — [arXiv:2502.04404](https://arxiv.org/abs/2502.04404) find
  15. Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744–20757.
  16. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. find
  17. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, and 1 others. 2023. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36:46534–46594. find
  18. Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. 2024a. Language agent tree search unifies reasoning, acting, and planning in language models. In Forty-first International Conference on Machine Learning. find
  19. Yichen Pan, Dehan Kong, Sida Zhou, Cheng Cui, Yifei Leng, Bing Jiang, Hangyu Liu, Yanyi Shang, Shuyan Zhou, Tongshuang Wu, and 1 others. 2024. Webcanvas: Benchmarking web agents in online environments. arXiv preprint arXiv:2406.12373. — [arXiv:2406.12373](https://arxiv.org/abs/2406.12373) find
  20. Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. 2024b. Webarena: A realistic web environment for building autonomous agents. In The Twelfth International Conference on Learning Representatio…
  21. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-webrollback-enhancing-web-agents-explicit/