Elamur Research⌕
☰
home / papers / 2026 / InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous Agents

InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous Agents

Paper · 2026 · Chenglin Yu, Yuchen Wang, Songmiao Wang et al.

Summary

InfiAgent addresses the problem that LLM agents become unreliable over long tasks. Their prompts keep growing, and they must then truncate, summarize or retrieve, which loses information and lets early mistakes compound. The authors propose a different model of agent state. The file system holds the full, authoritative record of progress, including plans, artifacts, logs and code. The agent's working context is rebuilt at every step from a snapshot of that workspace plus a small, fixed window of recent actions, so context size stays constant no matter how long the run lasts. A separate thinking model periodically rewrites a structured record of todos, file descriptions, carryover state and upcoming steps, and the action window is then cleared. Agents are arranged in a hierarchy, and an external attention component processes heavy documents outside the main context. The framework needs no task-specific fine-tuning. On DeepResearch Bench and an 80-paper literature review, a 20B open model keeps pace with larger proprietary agents and sustains much better long-horizon coverage. This suggests that treating persistent state as a first-class object is a practical basis for durable agents.

Key points

  • Context is built as c_t = g(F_t, a_{t−k:t−1}) from workspace files and the last k actions, giving O(1) context size with respect to task horizon; checkpointed traces allow resuming after interruption.
  • The agent hierarchy has three levels:
    • an Alpha Agent at the root that splits user requests;
    • domain agents such as data_collection_agent and coder_agent;
    • atomic agents such as get_data_from_arxiv and answer_from_papers.
  • Evaluation covered DeepResearch Bench and a literature review task of up to 80 papers. With GPT-OSS-20B as the backbone, InfiAgent was competitive with larger proprietary research agents, though the authors do not claim absolute superiority.
  • In the ablation, replacing file-centric state with compressed long-context prompts sharply reduced average coverage: 3.2 for GPT-OSS-20B, 21.1 for Gemini-3-Flash and 27.7 for Claude-4.5-Sonnet.
  • Stated limitations include:
    • latency from the agent hierarchy;
    • strictly serial execution with no parallelism;
    • hallucinations that can persist in files, especially with smaller models;
    • evaluation limited to research-style tasks.

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 Hong KongPok Fu Lam · China118 works · 13,496 citations🎓Hong Kong Polytechnic UniversityHong Kong · China106 works · 5,190 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

Cite

Cite as Chenglin Yu, Yuchen Wang, Songmiao Wang, Hongxia Yang and Li Ming. 2026. InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous 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.1787

@inproceedings{yu2026infiagent,
  title = {{InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous Agents}},
  author = {Chenglin Yu, Yuchen Wang, Songmiao Wang, Hongxia Yang and Li Ming},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1787},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1787}
}

References

13 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Xinye Cai, Zhiwei Mei, Zhun Fan, and Qingfu Zhang. 2018. A constrained decomposition approach with grids for evolutionary multiobjective optimization. IEEE Transactions on Evolutionary Computation, 22(4):564–577. find
  2. 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.
  3. Zhenyu Guan, Xiangyu Kong, Fangwei Zhong, and Yizhou Wang. 2024. Richelieu: Self-evolving LLM-based agents for AI diplomacy. In Advances in Neural Information Processing Systems, volume 37, pages 123471–123497. Curran Associates, Inc. find
  4. Wenhao Li, Dan Qiao, Baoxiang Wang, Xiangfeng Wang, Bo Jin, and Hongyuan Zha. 2023. Semantically aligned task decomposition in multi-agent reinforcement learning. arXiv preprint arXiv:2305.10865. — [arXiv:2305.10865](https://arxiv.org/abs/2305.10865) find
  5. Haowei Liu, Xi Zhang, Haiyang Xu, Yuyang Wanyan, Junyang Wang, Ming Yan, Ji Zhang, Chunfeng Yuan, Changsheng Xu, Weiming Hu, and Fei Huang. 2025. PC-Agent: A hierarchical multi-agent collaboration framework for complex task automation on PC. arXiv preprint arXiv:2502.14282. — [arXiv:2502.14282](https://arxiv.org/abs/2502.14282) find
  6. Elliot Meyerson, Giuseppe Paolo, Roberto Dailey, Hormoz Shahrzad, Olivier Francon, Conor F. Hayes, Xin Qiu, Babak Hodjat, and Risto Miikkulainen. 2025. Solving a million-step LLM task with zero errors. arXiv preprint arXiv:2511.09030. — [arXiv:2511.09030](https://arxiv.org/abs/2511.09030) find
  7. David J. Moore. 2025. A taxonomy of hierarchical multi-agent systems: Design patterns, coordination mechanisms, and industrial applications. arXiv preprint arXiv:2508.12683. — [arXiv:2508.12683](https://arxiv.org/abs/2508.12683) find
  8. Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. 2025. A comprehensive overview of large language models. ACM Transactions on Intelligent Systems and Technology, 16(5):1–72.
  9. Xiaodong Qu, Andrews Damoah, Joshua Sherwood, Peiyan Liu, Christian Shun Jin, Lulu Chen, Minjie Shen, Nawwaf Aleisa, Zeyuan Hou, Chenyu Zhang, Lifu Gao, Yanshu Li, Qikai Yang, Qun Wang, and Cristabelle De Souza. 2025. A comprehensive review of AI agents: Transforming possibilities in technology and… — [arXiv:2508.11957](https://arxiv.org/abs/2508.11957) find
  10. Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O’Sullivan, and Hoang D. Nguyen. 2025. Multi-agent collaboration mechanisms: A survey of LLMs. arXiv preprint arXiv:2501.06322.
  11. Honguk Woo, Gwangpyo Yoo, and Minjong Yoo. 2022. Structure learning-based task decomposition for reinforcement learning in non-stationary environments. Proceedings of the AAAI Conference on Artificial Intelligence, 36(8):8657–8665. find
  12. Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Xin Guo, Dingwen Yang, Chenyang Liao, Wei He, Songyang Gao, Lu Chen, Rui Zheng, Yicheng Zou, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang, Zuxuan Wu, and Yu-Gang Jiang. 2025. AgentGym: Evaluating and training large langua… find
  13. Xinjie Zhao, Moritz Blum, Fan Gao, Yingjian Chen, Boming Yang, Luis Marquez-Carpintero, Mónica Pina-Navarro, Yanran Fu, So Morikawa, Yusuke Iwasawa, Yutaka Matsuo, Chanjun Park, and Irene Li. 2025. AGENTiGraph: A multi-agent knowledge graph framework for interactive, domain-specific LLM chatbots. In… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yu-2026-infiagent-infinite-horizon-framework-general/