Elamur Research⌕
☰
home / papers / 2026 / FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based Agents

FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based Agents

Paper · 2026 · Chiwei Zhu, Benfeng Xu, Mingxuan Du et al.

Summary

Deep research agents must browse many webpages and produce long reports, but their trajectories quickly exceed the model's context window. This squeezes the token budget for both gathering evidence and writing, and blocks test-time scaling. FS-Researcher addresses this by moving the agent's working state into a persistent file-system workspace stored as Markdown. Two ReAct-style agents split the work:

  • Context Builder searches the web and records cited notes in a hierarchical knowledge base (index.md, knowledge_base/, sources/) that can grow far past context length.
  • Report Writer has no web access. It drafts the report section by section, using the knowledge base as its only source of facts.

Control files (todos, a checklist, logs) let each agent review its own output and refine it over multiple sessions. On DeepResearch Bench and DeepConsult, the framework reaches state-of-the-art report quality across several backbone models. Report quality also rises as more compute goes to context building, which shows that test-time scaling works without being capped by the context window.

Key points

  • Tools: file-system operations (ls, grep, read_file, insert/delete/replace) plus search_web (Google SERP API) and read_webpage (Jina AI API). File I/O takes under 0.03% of total wall-clock time.
  • Ablations with GPT-5: removing the persistent workspace, the dual-agent split, or section-wise writing all lower scores. Dropping the dual-agent design hurts most, by up to about 17 points on one metric.
  • More context-building rounds: going from 3 to 5 to 10 rounds makes reports denser and more evidence-grounded. Readability falls accordingly (59.32 → 55.49 → 54.79).
  • Smaller models: GPT-5-mini, given extra context-building rounds, performs comparably to OpenAI-DeepResearch at much lower cost. However, weaker backbones tend to stop prematurely and make more file-operation errors.
  • Availability: code and data are released at github.com/Ignoramus0817/FS-Researcher.

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 Science and Technology of ChinaHefei · China108 works · 112,698 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksDecision Sciences · 209 worksInformation Systems and Management · 41 works

Cite

Cite as Chiwei Zhu, Benfeng Xu, Mingxuan Du, Shaohan Wang, Xiaorui Wang, Zhendong Mao and Yongdong Zhang. 2026. FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based 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.288

@inproceedings{zhu2026fs,
  title = {{FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based Agents}},
  author = {Chiwei Zhu, Benfeng Xu, Mingxuan Du, Shaohan Wang, Xiaorui Wang, Zhendong Mao and Yongdong Zhang},
  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.288},
  url = {https://doi.org/10.18653/v1/2026.acl-long.288}
}

References

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

  1. Anthropic. 2025a. Meet your thinking parterner. https://claude.com/product/overview. ↗
  2. Anthropic. 2025b. Work with claude directly in your codebase. build, debug, and ship from your terminal, ide, slack, or the web. describe what you need, and claude handles the rest. https://claude.com/ product/claude-code. ↗
  3. Cursor. 2025. Built to make you extraordinarily productive, cursor is the best way to code with ai. https://cursor.com/. ↗
  4. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, and 181 others. 2025. Deepseek-r1: Incentivizing reasoning capability in ll…
  5. Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. 2025. Deepresearch bench: A comprehensive benchmark for deep research agents. Preprint, arXiv:2506.11763.
  6. Google. 2025. Gemini deep research. https://gemini.google/overview/deep-research/. ↗
  7. Zikang Guo, Benfeng Xu, Chiwei Zhu, Wentao Hong, Xiaorui Wang, and Zhendong Mao. 2025. Mcp-agentbench: Evaluating real-world language agent performance with mcp-mediated tools. Preprint, arXiv:2509.09734. — [arXiv:2509.09734](https://arxiv.org/abs/2509.09734) find
  8. LangChain. 2025. Open deep research. https://blog.langchain.com/open-deep-research/. ↗
  9. Yu Lei, Shuzheng Si, Wei Wang, Yifei Wu, Gang Chen, Fanchao Qi, and Maosong Sun. 2025. Rhinoinsight: Improving deep research through control mechanisms for model behavior and context. Preprint, arXiv:2511.18743. — [arXiv:2511.18743](https://arxiv.org/abs/2511.18743) find
  10. Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yongkang Wu, Ji-Rong Wen, Yutao Zhu, and Zhicheng Dou. 2025a. Webthinker: Empowering large reasoning models with deep research capability. Preprint, arXiv:2504.21776. find
  11. Zijian Li, Xin Guan, Bo Zhang, Shen Huang, Houquan Zhou, Shaopeng Lai, Ming Yan, Yong Jiang, Pengjun Xie, Fei Huang, Jun Zhang, and Jingren Zhou. 2025b. Webweaver: Structuring web-scale evidence with dynamic outlines for open-ended deep research. Preprint, arXiv:2509.13312. — [arXiv:2509.13312](https://arxiv.org/abs/2509.13312) find
  12. Abel Lim, Saahil Jain, Vincent Seng, and adamp. 2025. Deepconsult: A deep research benchmark for consulting / business queries. find
  13. Heng Lin and Zhongwen Xu. 2025. Understanding tool-integrated reasoning. Preprint, arXiv:2508.19201. — [arXiv:2508.19201](https://arxiv.org/abs/2508.19201) find
  14. OpenAI, :, Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, Alex Iftimie, Alex Karpenko, Alex Tachard Passos, Alexander Neitz, Alexander Prokofiev, Alexander Wei, Allison Tam, and 244 others. 2024. Openai o1 sy…
  15. OpenAI. 2025. Deep research system card. https://openai.com/index/ deep-research-system-card/. ↗
  16. Shishir G. Patil, Huanzhi Mao, Charlie Cheng-Jie Ji, Fanjia Yan, Vishnu Suresh, Ion Stoica, and Joseph E. Gonzalez. 2025. The berkeley function calling leaderboard (bfcl): From tool use to agentic evaluation of large language models. In Forty-second International Conference on Machine Learning. find
  17. Akshara Prabhakar, Roshan Ram, Zixiang Chen, Silvio Savarese, Frank Wang, Caiming Xiong, Huan Wang, and Weiran Yao. 2025. Enterprise deep research: Steerable multi-agent deep research for enterprise analytics. Preprint, arXiv:2510.17797. — [arXiv:2510.17797](https://arxiv.org/abs/2510.17797) find
  18. Aymeric Roucher, Albert Villanova del Moral, merve, Thomas Wolf, and Clémentine Fourrier. 2025. Open-source deepresearch – freeing our search agents. https://huggingface.co/blog/ open-deep-research. ↗
  19. Yijia Shao, Yucheng Jiang, Theodore Kanell, Peter Xu, Omar Khattab, and Monica Lam. 2024. Assisting in writing Wikipedia-like articles from scratch with large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human…
  20. Zhengwei Tao, Jialong Wu, Wenbiao Yin, Junkai Zhang, Baixuan Li, Haiyang Shen, Kuan Li, Liwen Zhang, Xinyu Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. 2025. Webshaper: Agentically data synthesizing via information-seeking formalization. Preprint, arXiv:2507.15061. — [arXiv:2507.15061](https://arxiv.org/abs/2507.15061) find
  21. Tavily. 2025. Building deep research: How we achieved state of the art. https://blog.tavily.com/research-en/. ↗
  22. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alexandre Passos, William Fedus, and Amelia Glaese. 2025. Browsecomp: A simple yet challenging benchmark for browsing agents. Preprint, arXiv:2504.12516. find
  23. John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. Swe-agent: Agent-computer interfaces enable automated software engineering. Preprint, arXiv:2405.15793.
  24. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2024. τ -bench: A benchmark for tool-agent-user interaction in real-world domains. Preprint, arXiv:2406.12045.
  25. Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. 2025. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. Preprint, arXiv:2504.03160.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhu-2026-fs-researcher-test-time-scaling/