Elamur Research⌕
☰
home / papers / 2026 / ExpSeek: Self-Triggered Experience Seeking for Web Agents

ExpSeek: Self-Triggered Experience Seeking for Web Agents

Paper · 2026 · Wenyuan Zhang, Xinghua Zhang, Haiyang Yu et al.

Summary

Web agents that search the open web often explore poorly or answer too early, particularly when built on small LLMs. Earlier experience-based methods supply lessons from past trajectories only once, as global context before the task begins, so the guidance cannot respond to what the agent sees later. ExpSeek instead lets the agent request help at individual steps.

The system has two parts:

  • When to intervene: the agent's own step entropy (mean token entropy of a response) acts as the trigger. Threshold intervals for each step type are estimated with logistic regression and bootstrap resampling.
  • What to provide: an experience base is built by contrasting successful and failed trajectories. A separate experience model reads the current history, consults the relevant entries and writes guidance tailored to that step.

ReAct-style agents on Qwen3-8B and Qwen3-32B were tested across four difficult web benchmarks. ExpSeek gave clear gains over the base models and over passive-injection baselines. The work matters because an intrinsic signal decides intervention timing without an external judge, and even a small guidance model can help a larger agent.

Key points

  • Average absolute gains over the base models are 9.3% for Qwen3-8B and 7.5% for Qwen3-32B, with a maximum of 14.6%. ExpSeek beats passive experience injection by 6.7% and 6.0%.
  • Each experience entry is a triplet of behavior, mistake and guidance, and entries are grouped by topic. Generated guidance gave clearly higher accuracy than retrieving the single most similar entry with text-embedding-v4.
  • Triggering on entropy matched the accuracy of rule-based and reward-model triggers (the latter using claude-sonnet-4) at lower cost. The rule-based trigger needed 1.7× the steps and 2.6× the time on GAIA.
  • A Qwen3-4B guidance model raised the 32B agent's score by 5.2 points on GAIA and 9.7 points on xbench.
  • Guidance raised entropy in intermediate steps, which encourages exploration, and lowered it at the answer step, which indicates more confident convergence. Code is available at github.com/WYRipple/ExpSeek.

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

🏛️Chinese Academy of SciencesBeijing · China194 works · 23,906 citations🎓University of Chinese Academy of SciencesBeijing · China147 works · 8,480 citations🔬Institute of Information EngineeringBeijing · China30 works · 1,411 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksInformation Systems · 1,437 works

Cite

Cite as Wenyuan Zhang, Xinghua Zhang, Haiyang Yu, Shuaiyi Nie, Bingli Wu, Juwei Yue, Tingwen Liu and Yongbin Li. 2026. ExpSeek: Self-Triggered Experience Seeking for Web Agents. 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.1462

@inproceedings{zhang2026expseek,
  title = {{ExpSeek: Self-Triggered Experience Seeking for Web Agents}},
  author = {Wenyuan Zhang, Xinghua Zhang, Haiyang Yu, Shuaiyi Nie, Bingli Wu, Juwei Yue, Tingwen Liu and Yongbin Li},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1462},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1462}
}

References

5 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Zhengyang Ai, Zikang Shan, Xiaodong Ai, Jingxian Tang, Hangkai Hu, and Pinyan Lu. 2026. Shape: Stage-aware hierarchical advantage via potential estimation for llm reasoning. Preprint, arXiv:2604.06636. find
  2. Sikai Bai, Haoxi Li, Jie Zhang, Yongjiang Liu, and Song Guo. 2026. Ttvs: Boosting self-exploring reinforcement learning via test-time variational synthesis. arXiv preprint arXiv:2604.08468. find
  3. Vance W Berger and YanYan Zhou. 2014. Kolmogorov– smirnov test: Overview. Wiley statsref: Statistics reference online. find
  4. We propose ExpSeek, a framework enabling web agents to actively seek step-level guidance based on step entropy. Even with small open-source agents, ExpSeek demonstrates significant performance improvements and excellent generalization on complex real-world web reasoning tasks, showing great potentia… find
  5. Andrew P Bradley. 1997. The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition, 30(7):1145–1159. \#\# Limitations Zhicheng Cai, Xinyuan Guo, Yu Pei, Jiangtao Feng, Jinsong Su, Jiangjie Chen, Ya-Qin Zhang, Wei-Ying Ma, Mingxuan Wang, and Hao Zhou.… — [arXiv:2511.06449](https://arxiv.org/abs/2511.06449) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-expseek-self-triggered-experience-seeking/