Elamur Research⌕
☰
home / papers / 2026 / Unlocking Implicit Experience: Synthesizing Tool-Use Trajectories from Text

Unlocking Implicit Experience: Synthesizing Tool-Use Trajectories from Text

Paper · 2026 · Zhihao Xu, Rumei Li, Jiahuan Li et al.

Summary

Training LLM agents for multi-turn tool use is limited by a shortage of realistic trajectories. Most existing synthesis methods start from a fixed API collection, so the data can only be as broad as that collection. The authors invert this approach. They argue that ordinary web text, such as how-to guides and procedural descriptions, already records human multi-step problem solving that can be turned into agent data. Their pipeline, GEM, works in four stages:

  1. It keeps text segments that describe multi-step procedures.
  2. It derives abstract workflows and OpenAI-format tool definitions from them.
  3. It has a strong model ground these into user–assistant–tool dialogues.
  4. It increases complexity and diversity, then checks the results with rule-based and LLM-based verification.

To reduce cost, they also fine-tune a Trajectory Synthesizer that performs the whole text-to-trajectory conversion in one step and reportedly matches the pipeline's quality. Models trained on GEM data improve substantially on BFCL V3 Multi-turn and τ²-bench. Their out-of-domain data competes with τ-bench in-domain training, which suggests the text-derived data generalizes well.

Key points

  • Source corpus analysis: Of roughly 250,000 sampled Ultra-fineweb segments, about 14% contain explicit multi-step workflows. These cover many categories, including customer support, research and data, and education. Qwen3-8B was used for annotation and filtering.
  • Headline gains: GEM-32B gains 14.9% on BFCL V3 Multi-turn, and GEM training yields a 13.8% gain on τ²-bench.
  • Ablation on BFCL V3 (overall accuracy):
    • Qwen3-32B rises from 28.35 to 44.88; without the refinement stage it reaches only 32.50.
    • Qwen3-8B rises from 18.00 to 30.25.
  • Data scale and complexity: Scaling GEM-8B training data from 10K to 70K trajectories keeps improving τ-bench Airline/Retail scores. The synthesized trajectories are long, averaging about 16 messages each.
  • Limitations: The framework currently handles English, text-only data. The authors name multilingual and multimodal extensions, and use of the data for mid-training, as future work.

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

🎓Renmin University of ChinaBeijing · China114 works · 16,118 citations🏢MeituanBeijing · China17 works · 3,769 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 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Zhihao Xu, Rumei Li, Jiahuan Li, Rongxiang Weng, Jingang Wang, Xunliang Cai and Xiting Wang. 2026. Unlocking Implicit Experience: Synthesizing Tool-Use Trajectories from Text. 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.452

@inproceedings{xu2026unlocking,
  title = {{Unlocking Implicit Experience: Synthesizing Tool-Use Trajectories from Text}},
  author = {Zhihao Xu, Rumei Li, Jiahuan Li, Rongxiang Weng, Jingang Wang, Xunliang Cai and Xiting Wang},
  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.452},
  url = {https://doi.org/10.18653/v1/2026.acl-long.452}
}

References

24 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 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
  2. Chen Chen, Xinlong Hao, Weiwen Liu, Xu Huang, Xingshan Zeng, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Yuefeng Huang, and 1 others. 2025. Acebench: Who wins the match point in tool usage? arXiv preprint arXiv:2501.12851. — [arXiv:2501.12851](https://arxiv.org/abs/2501.12851) find
  3. Zhicheng Guo, Sijie Cheng, Hao Wang, Shihao Liang, Yujia Qin, Peng Li, Zhiyuan Liu, Maosong Sun, and Yang Liu. 2024. Stabletoolbench: Towards stable large-scale benchmarking on tool learning of large language models. arXiv preprint arXiv:2403.07714. find
  4. Ai Jian, Xiaoyun Zhang, Wanrou Du, Jingqing Ruan, Jiangbo Pei, Weipeng Zhang, Ke Zeng, and Xunliang Cai. 2026. Trust-sql: Tool-integrated multi-turn reinforcement learning for text-to-sql over unknown schemas. arXiv preprint arXiv:2603.16448. — [arXiv:2603.16448](https://arxiv.org/abs/2603.16448) find
  5. Mahnaz Koupaee and William Yang Wang. 2018. Wikihow: A large scale text summarization dataset. arXiv preprint arXiv:1810.09305. find
  6. Jian Li, Yizhang Jin, Dongqi Liu, Hang Ding, Jiafu Wu, Dongsheng Chen, Yunhang Shen, Yulei Qin, Ying Tai, Chengjie Wang, and 1 others. 2026. Se-search: Self-evolving search agent via memory and dense reward. arXiv preprint arXiv:2603.03293. — [arXiv:2603.03293](https://arxiv.org/abs/2603.03293) find
  7. Yuetai Li, Huseyin A Inan, Xiang Yue, Wei-Ning Chen, Lukas Wutschitz, Janardhan Kulkarni, Radha Poovendran, Robert Sim, and Saravan Rajmohan. 2025. Simulating environments with reasoning models for agent training. arXiv preprint arXiv:2511.01824. — [arXiv:2511.01824](https://arxiv.org/abs/2511.01824) find
  8. 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.
  9. Weiwen Liu, Xu Huang, Xingshan Zeng, Xinlong Hao, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Zhengying Liu, Yuanqing Yu, and 1 others. 2024a. Toolace: Winning the points of llm function calling. arXiv preprint arXiv:2409.00920. find
  10. Zuxin Liu, Thai Hoang, Jianguo Zhang, Ming Zhu, Tian Lan, Juntao Tan, Weiran Yao, Zhiwei Liu, Yihao Feng, Rithesh RN, and 1 others. 2024b. Apigen: Automated pipeline for generating verifiable and diverse function-calling datasets. Advances in Neural Information Processing Systems, 37:54463–54482. find
  11. Shishir G Patil, Huanzhi Mao, Fanjia Yan, Charlie Cheng-Jie Ji, 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
  12. Akshara Prabhakar, Zuxin Liu, Ming Zhu, Jianguo Zhang, Tulika Awalgaonkar, Shiyu Wang, Zhiwei Liu, Haolin Chen, Thai Hoang, Juan Carlos Niebles, and 1 others. 2025. Apigen-mt: Agentic pipeline for multi-turn data generation via simulated agent-human interplay. arXiv preprint arXiv:2504.03601.
  13. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, and 1 others. 2023. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789.
  14. Jinbo Su, Lingzhe Gao, Wei Li, Shihao Liu, Haojie Lei, Xinyi Wang, Yuanzhao Guo, Ke Wang, Daiting Shi, and Dawei Yin. 2025. Racqc: Advanced retrieval-augmented generation for chinese query correction. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 675–689. find
  15. 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.
  16. Zhiyuan Yao, Zishan Xu, Yifu Guo, Zhiguang Han, Cheng Yang, Shuo Zhang, Weinan Zhang, Xingshan Zeng, and Weiwen Liu. 2026. Toolace-mcp: Generalizing history-aware routing from mcp tools to the agent web. arXiv preprint arXiv:2601.08276. — [arXiv:2601.08276](https://arxiv.org/abs/2601.08276) find
  17. Fan Yin, Zifeng Wang, I-Hung Hsu, Jun Yan, Ke Jiang, Yanfei Chen, Jindong Gu, Long Le, Kai-Wei Chang, Chen-Yu Lee, and 1 others. 2025. Magnet: Multi-turn tool-use data synthesis and distillation via graph translation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Lin… find
  18. Xingshan Zeng, Weiwen Liu, Lingzhi Wang, Liangyou Li, Fei Mi, Yasheng Wang, Lifeng Shang, Xin Jiang, and Qun Liu. 2025. Toolace-mt: Non-autoregressive generation for agentic multi-turn interaction. arXiv preprint arXiv:2508.12685. — [arXiv:2508.12685](https://arxiv.org/abs/2508.12685) find
  19. Xiaoyun Zhang, Jingqing Ruan, Xing Ma, Yawen Zhu, Jiansong Chen, Ke Zeng, and Xunliang Cai. 2025. A reasoner for real-world event detection: Scaling reinforcement learning via adaptive perplexity-aware sampling strategy. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language… find
  20. Weikang Zhao, Xili Wang, Chengdi Ma, Lingbin Kong, Zhaohua Yang, Mingxiang Tuo, Xiaowei Shi, Yitao Zhai, and Xunliang Cai. 2025. Mua-rl: Multi-turn user-interacting agent reinforcement learning for agentic tool use. arXiv preprint arXiv:2508.18669. — [arXiv:2508.18669](https://arxiv.org/abs/2508.18669) find
  21. Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of 100+ language models. arXiv preprint arXiv:2403.13372.
  22. Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534.
  23. Yudong Wang, Zixuan Fu, Jie Cai, Peijun Tang, Hongya Lyu, Yewei Fang, Zhi Zheng, Jie Zhou, Guoyang Zeng, Chaojun Xiao, and 1 others. 2025. Ultrafineweb: Efficient data filtering and verification for high-quality llm training data. arXiv preprint arXiv:2505.05427. — [arXiv:2505.05427](https://arxiv.org/abs/2505.05427) find
  24. Zhangchen Xu, Adriana Meza Soria, Shawn Tan, Anurag Roy, Ashish Sunil Agrawal, Radha Poovendran, and Rameswar Panda. 2025. Toucan: Synthesizing 1.5 m tool-agentic data from real-world mcp environments. arXiv preprint arXiv:2510.01179. — [arXiv:2510.01179](https://arxiv.org/abs/2510.01179) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/xu-2026-unlocking-implicit-experience-synthesizing-tool/