Elamur Research⌕
☰
home / papers / 2026 / U-Fold: Dynamic Intent-Aware Context Folding for User-Centric Agents

U-Fold: Dynamic Intent-Aware Context Folding for User-Centric Agents

Paper · 2026 · Jin Su, Runnan Fang, Yeqiu Li et al.

Summary

LLM agents that call tools during long conversations with users quickly run out of context. Existing context-folding methods address this by compressing earlier history, but they were designed for tasks with one goal. In multi-turn, user-facing dialogues they fail in two ways: they permanently drop small but decisive details, such as a shop's opening hours, and they lose track of user intent as it shifts. The authors propose U-Fold to fix both problems. U-Fold never deletes the original dialogue or the tool-call record. Instead, at every user turn it builds a fresh, compact working context from two modules:

  • a Conversation Summarization module that follows how the dialogue develops and keeps an explicit to-do list;
  • a Dynamic Data Extraction module that keeps only the fields of structured tool outputs that matter for the current goal.

The method was tested on τ-bench, τ²-bench, VitaBench and a deliberately noisier version of VitaBench. It beats ReAct and earlier folding methods, and the gains are largest when contexts are long and cluttered. This matters for moving context-management ideas from single-query benchmarks into practical assistant applications.

Key points

  • Against ReAct, U-Fold wins 71.4% of comparisons in long-context settings. Over prior folding baselines it improves results by up to 27.0%.

  • On a "hard" VitaBench variant, tool outputs are padded with redundant distractor data, using Qwen3-Thinking-30B-A3B as the backbone. Avg@4 scores improve as follows:

    DomainReActU-Fold
    In-store11.027.0
    Cross-Scenarios4.012.0

  • Ablations on VitaBench show that the two modules complement each other:

    • Without extraction, performance suffers most on the tool-heavy OTA domain.
    • Without summarization, performance suffers most on Delivery, where constraints are spread across many turns.

  • Error analysis identifies three main failure types: misreading user intent, omitting critical user information, and missing user errors. U-Fold reduces all three, with the largest reduction in omissions.

  • Using GPT-4.1 as the "folder" for a weaker Qwen3-4B agent beat both ReAct and standard U-Fold. This suggests a strong model could serve as a cheap context-shaping service for small agents.

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

🏢Alibaba GroupHangzhou · China168 works · 15,386 citations🎓Zhejiang UniversityHangzhou · China172 works · 12,988 citations🎓Westlake UniversityHangzhou · China45 works · 8,028 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 Jin Su, Runnan Fang, Yeqiu Li, Xiaobin Wang, Shihao Cai, Pengjun Xie, Ningyu Zhang and Fajie Yuan. 2026. U-Fold: Dynamic Intent-Aware Context Folding for User-Centric 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.897

@inproceedings{su2026u,
  title = {{U-Fold: Dynamic Intent-Aware Context Folding for User-Centric Agents}},
  author = {Jin Su, Runnan Fang, Yeqiu Li, Xiaobin Wang, Shihao Cai, Pengjun Xie, Ningyu Zhang and Fajie Yuan},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.897},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.897}
}

References

37 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Anthropic. 2025. Introducing claude sonnet 4.5. find
  2. Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. 2025. 𝜏 2 -bench: Evaluating conversational agents in a dual-control environment. arXiv preprint arXiv:2506.07982. find
  3. Fouad Bousetouane. 2026. Ai agents need memory control over more context. arXiv preprint arXiv:2601.11653. — [arXiv:2601.11653](https://arxiv.org/abs/2601.11653) find
  4. Guoxin Chen, Zile Qiao, Xuanzhong Chen, Donglei Yu, Haotian Xu, Wayne Xin Zhao, Ruihua Song, Wenbiao Yin, Huifeng Yin, Liwen Zhang, and 1 others. 2025. Iterresearch: Rethinking long-horizon agents via markovian state reconstruction. arXiv preprint arXiv:2511.07327. — [arXiv:2511.07327](https://arxiv.org/abs/2511.07327) find
  5. Maximillian Chen, Ruoxi Sun, Tomas Pfister, and Sercan Ö Arık. 2024. Learning to clarify: Multi-turn conversations with action-based contrastive self-training. arXiv preprint arXiv:2406.00222. — [arXiv:2406.00222](https://arxiv.org/abs/2406.00222) find
  6. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  7. Wei He, Yueqing Sun, Hongyan Hao, Xueyuan Hao, Zhikang Xia, Qi Gu, Chengcheng Han, Dengchang Zhao, Hui Su, Kefeng Zhang, Man Gao, Xi Su, Xiaodong Cai, Xunliang Cai, Yu Yang, and Yunke Zhao. 2025. Vitabench: Benchmarking llm agents with versatile interactive tasks in real-world applications. arXiv pr… — [arXiv:2509.26490](https://arxiv.org/abs/2509.26490) find
  8. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.
  9. Zhiyu Li, Shichao Song, Chenyang Xi, Hanyu Wang, Chen Tang, Simin Niu, Ding Chen, Jiawei Yang, Chunyu Li, Qingchen Yu, and 1 others. 2025. Memos: A memory os for ai system. arXiv preprint arXiv:2507.03724. — [arXiv:2507.03724](https://arxiv.org/abs/2507.03724) find
  10. Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, and 1 others. 2025. Deepseek-v3. 2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556.
  11. Jiarui Lu, Thomas Holleis, Yizhe Zhang, Bernhard Aumayer, Feng Nan, Haoping Bai, Shuang Ma, Shen Ma, Mengyu Li, Guoli Yin, and 1 others. 2025. Toolsandbox: A stateful, conversational, interactive evaluation benchmark for llm tool use capabilities. In Findings of the Association for Computational Lin…
  12. 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.
  13. Kushan Mitra, Dan Zhang, Hannah Kim, and Estevam Hruschka. 2026. Recap: Rewriting conversations for intent understanding in agentic planning. In Findings of the Association for Computational Linguistics: EACL 2026, pages 2015–2033.
  14. OpenAI. 2025. Introducing gpt-4.1 in the api. find
  15. Cheng Qian, Zuxin Liu, Akshara Prabhakar, Zhiwei Liu, Jianguo Zhang, Haolin Chen, Heng Ji, Weiran Yao, Shelby Heinecke, Silvio Savarese, and 1 others. 2025a. Userbench: An interactive gym environment for user-centric agents. arXiv preprint arXiv:2507.22034. — [arXiv:2507.22034](https://arxiv.org/abs/2507.22034) find
  16. Cheng Qian, Zuxin Liu, Akshara Prabhakar, Jielin Qiu, Zhiwei Liu, Haolin Chen, Shirley Kokane, Heng Ji, Weiran Yao, Shelby Heinecke, and 1 others. 2025b. Userrl: Training interactive user-centric agent via reinforcement learning. arXiv preprint arXiv:2509.19736. — [arXiv:2509.19736](https://arxiv.org/abs/2509.19736) find
  17. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36:68539–68551.
  18. Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36:38154–38180. 18051 9
  19. Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2020. Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768.
  20. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022b. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations.
  21. Weiwei Sun, Miao Lu, Zhan Ling, Kang Liu, Xuesong Yao, Yiming Yang, and Jiecao Chen. 2025. Scaling long-horizon llm agent via context-folding. arXiv preprint arXiv:2510.11967. — [arXiv:2510.11967](https://arxiv.org/abs/2510.11967) find
  22. Rui Ye, Zhongwang Zhang, Kuan Li, Huifeng Yin, Zhengwei Tao, Yida Zhao, Liangcai Su, Liwen Zhang, Zile Qiao, Xinyu Wang, and 1 others. 2025. Agentfold: Long-horizon web agents with proactive context management. arXiv preprint arXiv:2510.24699. — [arXiv:2510.24699](https://arxiv.org/abs/2510.24699) find
  23. Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023a. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv: Arxiv-2305.16291. find
  24. Xingyao Wang, Zihan Wang, Jiateng Liu, Yangyi Chen, Lifan Yuan, Hao Peng, and Heng Ji. 2023b. Mint: Evaluating llms in multi-turn interaction with tools and language feedback. arXiv preprint arXiv:2309.10691.
  25. 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. find
  26. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  27. 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.
  28. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2025. A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110.
  29. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  30. Jianguo Zhang, Tian Lan, Ming Zhu, Zuxin Liu, Thai Quoc Hoang, Shirley Kokane, Weiran Yao, Juntao Tan, Akshara Prabhakar, Haolin Chen, and 1 others. 2025. xlam: A family of large action models to empower ai agent systems. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter o…
  31. Xuanming Zhang, Sitong Wang, Jenny Ma, Alyssa Hwang, Zhou Yu, and Lydia B Chilton. 2024. Jump-starter: Human-ai planning with task-structured context curation. arXiv preprint arXiv:2410.03882. — [arXiv:2410.03882](https://arxiv.org/abs/2410.03882) find
  32. 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. arXiv preprint arXiv:2504.03160.
  33. Peilin Zhou, Bruce Leon, Xiang Ying, Can Zhang, Yifan Shao, Qichen Ye, Dading Chong, Zhiling Jin, Chenxuan Xie, Meng Cao, and 1 others. 2025a. 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
  34. Yuyang Zhou, Jin Su, Jiawei Zhang, Wangyang Hu, Tianli Tao, Guanqi Li, Xibin Zhou, Li Fan, and Fajie Yuan. 2025b. Prime: A multi-agent environment for orchestrating dynamic computational workflows in protein engineerings. bioRxiv, pages 2025–09. find
  35. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. 2025c. Mem1: Learning to synergize memory and reasoning for efficient long-horizon agents. arXiv preprint arXiv:2506.15841. — [arXiv:2506.15841](https://arxiv.org/abs/2506.15841) find
  36. Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022a. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744–20757.
  37. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2024. 𝜏-bench: A benchmark for tool-agent-user interaction in real-world domains. arXiv preprint arXiv:2406.12045. 18052 10

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/su-2026-u-fold-dynamic-intent-aware/