Elamur Research⌕
☰
home / papers / 2026 / WebAnchor: Anchoring Agent Planning to Stabilize Long-Horizon Web Reasoning

WebAnchor: Anchoring Agent Planning to Stabilize Long-Horizon Web Reasoning

Paper · 2026 · Xinmiao Yu, Liwen Zhang, Xiaocheng Feng et al.

Summary

LLM-based web research agents trained with reinforcement learning often lose strategic coherence over long multi-step searches. The authors find that the agent's opening planning move largely decides whether the whole trajectory succeeds. They call this the "plan anchor" effect. Standard GRPO ignores it because it spreads reward evenly across all steps. Their remedy, Anchor-GRPO, trains one policy model in two phases.

  1. Stage 1 (planning): Only the initial plan is updated. Later tokens are masked out, and the plan receives a dense score from rubrics produced by a Plan Rubrics Learner. This learner extracts lessons from successful, failed and paired self-play trajectories, then refines the rubrics through LLM updates and human checks.
  2. Stage 2 (execution): Tool use with Search and Visit is tuned using sparse, task-level rewards, keeping actions consistent with the plan.

Across models from 3B to 30B, the method beats vanilla GRPO and First-step GRPO on four deep-research benchmarks. Gains also grow with model size and context length. The work shows that targeted credit for early planning can make long-horizon agents more reliable.

Key points

  • Motivation experiments: A wrong first step lowered Avg Pass@1 by 28.7% on BC-ZH, 30.9% on BC-EN and 23.6% on GAIA.
  • Benchmarks: BrowseComp, BrowseComp-Zh, GAIA and XBench-DeepSearch. WebAnchor-30B reaches 46.0% pass@1 on BrowseComp and 76.4% on GAIA.
  • Plan rubrics: The rubrics assess aspects including how the task is broken down, how well the plan matches the goal, and which tools are chosen. A judge LLM scores each dimension, and the scores are normalized into a [0, 1] reward.
  • Agent setup: A ReAct-style agent uses two tools. Search is backed by the Google Search API, and Visit is an LLM-simulated page reader.
  • Ablations: Results support three claims: first-step optimization matters, joint planner–executor training beats planner-only training, and rubric-based dense rewards are essential.

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

🎓Harbin Institute of TechnologyHarbin · China185 works · 29,140 citations🏢Alibaba GroupHangzhou · China201 works · 17,173 citations🏢Alibaba Group (Cayman Islands)George Town · Cayman Islands36 works · 3,645 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksInformation Systems · 1,730 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,685 works

Cite

Cite as Xinmiao Yu, Liwen Zhang, Xiaocheng Feng, Pengjun Xie, Jingren Zhou, Bing Qin and Yong Jiang. 2026. WebAnchor: Anchoring Agent Planning to Stabilize Long-Horizon Web Reasoning. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.2058

@inproceedings{yu2026webanchor,
  title = {{WebAnchor: Anchoring Agent Planning to Stabilize Long-Horizon Web Reasoning}},
  author = {Xinmiao Yu, Liwen Zhang, Xiaocheng Feng, Pengjun Xie, Jingren Zhou, Bing Qin and Yong Jiang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2058},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2058}
}

References

22 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, Guorui Zhou, Yutao Zhu, Ji-Rong Wen, and Zhicheng Dou. 2025a. Agentic reinforced policy optimization. Preprint, arXiv:2507.19849. — [arXiv:2507.19849](https://arxiv.org/abs/2507.19849) find
  2. Zhichen Dong, Zhanhui Zhou, Zhixuan Liu, Chao Yang, and Chaochao Lu. 2025b. Emergent response planning in llms. arXiv preprint arXiv:2502.06258. — [arXiv:2502.06258](https://arxiv.org/abs/2502.06258) find
  3. Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. 2025. Deepresearch bench: A comprehensive benchmark for deep research agents. arXiv preprint arXiv:2506.11763.
  4. Junteng Liu, Yunji Li, Chi Zhang, Jingyang Li, Aili Chen, Ke Ji, Weiyu Cheng, Zijia Wu, Chengyu Du, Qidi Xu, and 1 others. 2025. Webexplorer: Explore and evolve for training long-horizon web agents. arXiv preprint arXiv:2509.06501. — [arXiv:2509.06501](https://arxiv.org/abs/2509.06501) find
  5. Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. 2023. Gaia: a benchmark for general ai assistants. In The Twelfth International Conference on Learning Representations.
  6. OpenAI. 2025. Deep research system card. find
  7. Zile Qiao, Guoxin Chen, Xuanzhong Chen, Donglei Yu, Wenbiao Yin, Xinyu Wang, Zhen Zhang, Baixuan Li, Huifeng Yin, Kuan Li, and 1 others. 2025. Webresearcher: Unleashing unbounded reasoning capability in long-horizon agents. arXiv preprint arXiv:2509.13309. — [arXiv:2509.13309](https://arxiv.org/abs/2509.13309) find
  8. Akshit Sinha, Arvindh Arun, Shashwat Goel, Steffen Staab, and Jonas Geiping. 2025. The illusion of diminishing returns: Measuring long horizon execution in llms. arXiv preprint arXiv:2509.09677. — [arXiv:2509.09677](https://arxiv.org/abs/2509.09677) find
  9. Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. 2025. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. arXiv preprint arXiv:2503.05592.
  10. Liangcai Su, Zhen Zhang, Guangyu Li, Zhuo Chen, Chenxi Wang, Maojia Song, Xinyu Wang, Kuan Li, Jialong Wu, Xuanzhong Chen, Zile Qiao, Zhongwang Zhang, Huifeng Yin, Shihao Cai, Runnan Fang, Zhengwei Tao, Wenbiao Yin, and 1 others. 2025. Scaling agents via continual pre-training. find
  11. Yuan Sui, Yufei He, Tri Cao, Simeng Han, Yulin Chen, and Bryan Hooi. 2025. Meta-reasoner: Dynamic guidance for optimized inference-time reasoning in large language models. arXiv preprint arXiv:2502.19918.
  12. Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, and 1 others. 2025. Tongyi deepresearch technical report. arXiv preprint arXiv:2510.24701. for long-horizon decision making through multi-turn reinforcement learn… — [arXiv:2510.24701](https://arxiv.org/abs/2510.24701) find
  13. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR).
  14. Guibin Zhang, Hejia Geng, Xiaohang Yu, Zhenfei Yin, Zaibin Zhang, Zelin Tan, Heng Zhou, Zhongzhi Li, Xiangyuan Xue, Yijiang Li, Yifan Zhou, Yang Chen, Chen Zhang, Yutao Fan, Zihu Wang, Songtao Huang, Francisco Piedrahita-Velez, Yue Liao, Hongru Wang, and 6 others. 2025a. The landscape of agentic rei… — [arXiv:2509.02547](https://arxiv.org/abs/2509.02547) find
  15. Zijing Zhang, Ziyang Chen, Mingxiao Li, Zhaopeng Tu, and Xiaolong Li. 2025b. Rlvmr: Reinforcement learning with verifiable meta-reasoning rewards for robust long-horizon agents. Preprint, arXiv:2507.22844. — [arXiv:2507.22844](https://arxiv.org/abs/2507.22844) find
  16. Yida Zhao, Kuan Li, Xixi Wu, Liwen Zhang, Dingchu Zhang, Baixuan Li, Maojia Song, Zhuo Chen, Chenxi Wang, Xinyu Wang, and 1 others. 2025. Repurposing synthetic data for fine-grained search agent supervision. arXiv preprint arXiv:2510.24694. — [arXiv:2510.24694](https://arxiv.org/abs/2510.24694) find
  17. Peilin Zhou, Bruce Leon, Xiang Ying, Can Zhang, Yifan Shao, Qichen Ye, Dading Chong, Zhiling Jin, Chenxuan Xie, Meng Cao, and 1 others. 2025. Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese. arXiv preprint arXiv:2504.19314. — [arXiv:2504.19314](https://arxiv.org/abs/2504.19314) find
  18. Ziyu Wan, Yunxiang Li, Xiaoyu Wen, Yan Song, Hanjing Wang, Linyi Yang, Mark Schmidt, Jun Wang, Weinan Zhang, Shuyue Hu, and Ying Wen. 2025. Rema: Learning to meta-think for llms with multi-agent reinforcement learning. Preprint, arXiv:2503.09501. — [arXiv:2503.09501](https://arxiv.org/abs/2503.09501) find
  19. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. 2025. Browsecomp: A simple yet challenging benchmark for browsing agents. arXiv preprint arXiv:2504.12516.
  20. Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Xinmiao Yu, Dingchu Zhang, Yong Jiang, and 1 others. 2025. Resum: Unlocking long-horizon search intelligence via context summarization. arXiv preprint arXiv:2509.13313.
  21. Xbench Team. 2025. Xbench-deepsearch. find
  22. Zhiheng Xi, Jixuan Huang, Chenyang Liao, Baodai Huang, Honglin Guo, Jiaqi Liu, Rui Zheng, Junjie Ye, Jiazheng Zhang, Wenxiang Chen, Wei He, Yiwen Ding, Guanyu Li, Zehui Chen, Zhengyin Du, Xuesong Yao, Yufei Xu, Jiecao Chen, Tao Gui, and 4 others. 2025. Agentgym-rl: Training llm agents find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yu-2026-webanchor-anchoring-agent-planning-stabilize/