Elamur Research⌕
☰
home / papers / 2026 / Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors

Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors

Paper · 2026 · Zhiwei Zhang, Fei Zhao, Rui Wang et al.

Summary

Small tool-calling LLMs handle multi-turn tasks poorly once a tool call fails. Instead of reading the error message and adjusting, they often repeat the same invalid call until the dialogue breaks down. On BFCL v4 Multi-Turn, Qwen3-8B recovers in only about 20% of such cases, while Claude Sonnet 4 recovers in over 50%. The authors trace this to training. Standard GRPO reduces a failure to a sparse negative reward and gives no signal when every rollout fails. Offline error-correction datasets drift away from the mistakes the policy currently makes. FISSION-GRPO addresses both problems inside the RL loop:

  • Ordinary GRPO exploration runs as usual.
  • Failed trajectories are intercepted.
  • A fine-tuned Error Simulator adds realistic diagnostic feedback that does not reveal the answer.
  • Each augmented failure "fissions" into several fresh on-policy recovery rollouts.

The result is dense corrective signal built from the model's own current errors. The method sets new top results on BFCL v4 Multi-Turn for Qwen3 models and transfers to TAU-Bench and TAU2-Bench. This matters for compact, low-latency or on-device agents that must cope with unreliable APIs.

Key points

  • Main result: On BFCL v4 Multi-Turn, Qwen3-8B gains 5.7 points in error recovery rate and 4.0 points in overall accuracy (42.75% → 46.75%). It beats RL baselines and specialized tool-use agents, and improvements hold across Qwen3 1.7B, 4B and 8B.
  • Transfer: The method generalizes to TAU-Bench and TAU2-Bench, which use different tool ecosystems. It leads in most settings, with the largest gain of +17.4% on TAU1 Retail.
  • Error Simulator quality: Human raters judged 96% of 100 sampled outputs non-leaking (Cohen's κ = 0.71). Automated quality scores dropped by at most 0.24 points on unseen tool schemas.
  • Compute-matched check: With an equal budget of 432 updates on Qwen3-8B, FISSION-GRPO scores 46.75 versus 41.75 for GRPO. Most of the gain comes from the Miss Param and Miss Func subsets.
  • Code: Released at github.com/zxzadm/Fission-GRPO. In a case study, the trained model checks the environment state (e.g., with find) before retrying a failed action, whereas the GRPO-trained model invents arguments that do not exist.

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 University of Hong KongHong Kong · China126 works · 14,073 citations🏢Xiaohongshu (Little Red Book)Shanghai · China11 works · 107 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

Cite

Cite as Zhiwei Zhang, Fei Zhao, Rui Wang, Zezhong Wang, Bin Liang, Jiakang Wang, Yao Hu, Shaosheng Cao and Kam-Fai Wong. 2026. Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors. 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.1880

@inproceedings{zhang2026robust,
  title = {{Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors}},
  author = {Zhiwei Zhang, Fei Zhao, Rui Wang, Zezhong Wang, Bin Liang, Jiakang Wang, Yao Hu, Shaosheng Cao and Kam-Fai Wong},
  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.1880},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1880}
}

References

37 references, 20 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. Peter Belcak, Greg Heinrich, Shizhe Diao, Yonggan Fu, Xin Dong, Saurav Muralidharan, Yingyan Celine Lin, and Pavlo Molchanov. 2025. Small language models are the future of agentic ai. arXiv preprint arXiv:2506.02153. find
  3. Wenlong Deng, Yi Ren, Muchen Li, Danica J Sutherland, Xiaoxiao Li, and Christos Thrampoulidis. 2025. On the effect of negative gradient in group relative deep reinforcement optimization. arXiv preprint arXiv:2505.18830. — [arXiv:2505.18830](https://arxiv.org/abs/2505.18830) find
  4. Tianqing Fang, Hongming Zhang, Zhisong Zhang, Kaixin Ma, Wenhao Yu, Haitao Mi, and Dong Yu. 2025a. Webevolver: Enhancing web agent self-improvement with co-evolving world model. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 8970– 8986.
  5. Tianqing Fang, Zhisong Zhang, Xiaoyang Wang, Rui Wang, Can Qin, Yuxuan Wan, Jun-Yu Ma, Ce Zhang, Jiaqi Chen, Xiyun Li, Hongming Zhang, Haitao Mi, and Dong Yu. 2025b. Cognitive kernel-pro: A framework for deep research agents and agent foundation models training. Preprint, arXiv:2508.00414. — [arXiv:2508.00414](https://arxiv.org/abs/2508.00414) 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. 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. In ACL (Findings). find
  8. Shiting Huang, Zhen Fang, Zehui Chen, Siyu Yuan, Junjie Ye, Yu Zeng, Lin Chen, Qi Mao, and Feng Zhao. 2025. Critictool: Evaluating self-critique capabilities of large language models in tool-calling error scenarios. arXiv preprint arXiv:2506.13977.
  9. Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, and 1 others. 2024. Training language models to self-correct via reinforcement learning. arXiv preprint arXiv:2409.12917.
  10. Mukai Li, Qingcheng Zeng, Tianqing Fang, Zhenwen Liang, Linfeng Song, Qi Liu, Haitao Mi, and Dong Yu. 2026. Verified critical step optimization for LLM agents. CoRR, abs/2602.03412. find
  11. Zihan Lin, Xiaohan Wang, Hexiong Yang, Jiajun Chai, Jie Cao, Guojun Yin, Wei Lin, and Ran He. 2025. Awpo: Enhancing tool-use of large language models through adaptive integration of reasoning rewards. arXiv preprint arXiv:2512.19126. — [arXiv:2512.19126](https://arxiv.org/abs/2512.19126) find
  12. Weiwen Liu, Xu Huang, Xingshan Zeng, Xinlong Hao, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Zhengying Liu, Yuanqing Yu, and 1 others. 2024. Toolace: Winning the points of llm function calling. arXiv preprint arXiv:2409.00920. find
  13. Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025. Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783.
  14. Gongrui Nan, Siye Chen, Jing Huang, Mengyu Lu, Dexun Wang, Chunmei Xie, Weiqi Xiong, Xianzhou Zeng, Qixuan Zhou, Yadong Li, and 1 others. 2025. Ngrpo: Negative-enhanced group relative policy optimization. arXiv preprint arXiv:2509.18851. — [arXiv:2509.18851](https://arxiv.org/abs/2509.18851) find
  15. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744.
  16. Zhuoshi Pan, Yu Li, Honglin Lin, Qizhi Pei, Zinan Tang, Wei Wu, Chenlin Ming, H Vicky Zhao, Conghui He, and Lijun Wu. 2025. Lemma: Learning from errors for mathematical advancement in llms. arXiv preprint arXiv:2503.17439. find
  17. 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
  18. Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. 2024. Gorilla: Large language model connected with massive apis. Advances in Neural Information Processing Systems, 37:126544–126565.
  19. 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.
  20. 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.
  21. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. find
  22. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  23. Raghav Sharma and Manan Mehta. 2025. Small language models for agentic systems: A survey of architectures, capabilities, and deployment trade offs. arXiv preprint arXiv:2510.03847. find
  24. Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256.
  25. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652.
  26. David Silver and Richard S Sutton. 2025. Welcome to the era of experience. Google AI, 1:11. find
  27. Junhao Su, Yuanliang Wan, Junwei Yang, Hengyu Shi, Tianyang Han, Junfeng Luo, and Yurui Qiu. 2025. Failure makes the agent stronger: Enhancing accuracy through structured reflection for reliable tool interactions. arXiv preprint arXiv:2509.18847.
  28. 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.
  29. Qwen Team. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  30. Sri Vatsa Vuddanti, Aarav Shah, Satwik Kumar Chit-tiprolu, Tony Song, Sunishchal Dev, Kevin Zhu, and Maheep Chaudhary. 2025. Paladin: Self-correcting language model agents to cure tool-failure cases. arXiv preprint arXiv:2509.25238. — [arXiv:2509.25238](https://arxiv.org/abs/2509.25238) find
  31. Rui Wang, Ce Zhang, Jun-Yu Ma, Jianshu Zhang, Hongru Wang, Yi Chen, Boyang Xue, Tianqing Fang, Zhisong Zhang, Hongming Zhang, Haitao Mi, Dong Yu, and Kam-Fai Wong. 2025. Explore to evolve: Scaling evolved aggregation logic via proactive online exploration for deep research agents. Preprint, arXiv:25… — [arXiv:2510.14438](https://arxiv.org/abs/2510.14438) find
  32. Kaishuai Xu, Tiezheng Yu, Wenjun Hou, Yi Cheng, Chak Tou Leong, Liangyou Li, Xin Jiang, Lifeng Shang, Qun Liu, and Wenjie Li. 2025. Subtle errors in reasoning: Preference learning via error-injected self-editing. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguist… find
  33. 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.
  34. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. In The eleventh international conference on learning representations.
  35. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, and 1 others. 2025. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476.
  36. Jianguo Zhang, Tian Lan, Ming Zhu, Zuxin Liu, Thai Quoc Hoang, Shirley Kokane, Weiran Yao, Juntao Tan, Akshara Prabhakar, Haolin Chen, and 1 others. 2025a. 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…
  37. Kangning Zhang, Wenxiang Jiao, Kounianhua Du, Yuan Lu, Weiwen Liu, Weinan Zhang, and Yong Yu. 2025b. Looptool: Closing the data-training loop for robust llm tool calls. arXiv preprint arXiv:2511.09148.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-robust-tool-use-fission-grpo/