Elamur Research⌕
☰
home / papers / 2026 / Multi-Agent Procedural Graph Extraction with Structural and Logical Refinement

Multi-Agent Procedural Graph Extraction with Structural and Logical Refinement

Paper · 2026 · Wangyang Ying, Yanchi Liu, Xujiang Zhao et al.

Summary

The paper addresses the conversion of natural-language procedure descriptions into procedural graphs. These are directed graphs whose nodes are actors, actions, gateways (XOR/OR/AND), constraints and start/end markers, and whose edges encode sequence, condition and constraint flows. One-pass LLM prompting often yields graphs that cannot be executed or that misread conditional logic.

The authors propose text2flow, a multi-agent loop that needs no training. A graph builder agent first drafts the graph via few-shot prompting, using textual triples. An evaluation sandbox then inspects it with two agents:

  • a simulation agent runs the graph to expose topological defects such as dead nodes or paths that never reach End;
  • a semantic agent checks whether gateway types match the conditions stated in the text.

Their findings are ranked, phrased in natural language, and fed into the next prompt. Because this feedback comes from observable execution results and explicit text–graph conflicts, it avoids relying on the model's own judgment.

Across backbones, the method outperforms zero-shot, few-shot, Self-Refine and Actor-Critic baselines on both structural and logical metrics. It offers a supervision-free, interpretable route to executable workflows, though at a higher token cost.

Key points

  • The framework has three iterative stages: a graph builder agent, a simulation agent for structural feedback, and a semantic agent for logical feedback. All feedback is top-k prioritized and requires no parameter updates.
  • Tested backbones include GPT-4o, Gemma-27B and Llama-3.1-8B. On GPT-4o, gains mostly appear in the first two refinement rounds and then plateau, so the default is two iterations.
  • On Llama-3.1-8B, text2flow raises Actor F1 from 0.403 to 0.577 and Action F1 from 0.649 to 0.742 compared with few-shot prompting.
  • LoRA supervised fine-tuning on Llama-3.1-8B beats text2flow on AND/OR gateway typing. However, it lags on Flow-Seq and Flow-Constrain and on semantic roles.
  • Average tokens per sample are 305 for zero-shot, 1,256 for few-shot and 2,489 for text2flow, the highest of all methods. The main limitations are this cost and the reliance on rule-based heuristics and LLM judgments.

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

🎓Arizona State UniversityTempe · United States54 works · 2,356 citations🏢NEC (United States)Irving · United States10 works · 658 citations

 Fields, topics and concepts

📚 Shelves 5The 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 worksComputer Vision and Pattern Recognition · 3,486 works

Cite

Cite as Wangyang Ying, Yanchi Liu, Xujiang Zhao, Wei Cheng, Zhengzhang Chen, Wenchao Yu, Yanjie Fu and Haifeng Chen. 2026. Multi-Agent Procedural Graph Extraction with Structural and Logical Refinement. Findings of the Association for Computational Linguistics: EACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-eacl.158

@inproceedings{ying2026multi,
  title = {{Multi-Agent Procedural Graph Extraction with Structural and Logical Refinement}},
  author = {Wangyang Ying, Yanchi Liu, Xujiang Zhao, Wei Cheng, Zhengzhang Chen, Wenchao Yu, Yanjie Fu and Haifeng Chen},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.158},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.158}
}

References

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

  1. Patrizio Bellan, Mauro Dragoni, and Chiara Ghidini. 2022. Leveraging pre-trained language models for conversational information seeking from text. Preprint, arXiv:2204.03542. — [arXiv:2204.03542](https://arxiv.org/abs/2204.03542) find
  2. Patrizio Bellan, Han van der Aa, Mauro Dragoni, Chiara Ghidini, and Simone Paolo Ponzetto. 2023. Pet: An annotated dataset for process extraction from natural language text tasks. In Business Process Management Workshops - BPM 2022 International Workshops, Revised Selected Papers, Lecture Notes in B… find
  3. Weihong Du, Wenrui Liao, Hongru Liang, and Wenqiang Lei. 2024. PAGED: A benchmark for procedural graphs extraction from documents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10829–10846, Bangkok, Thailand. Association for… find
  4. Marlon Dumas, L Marcello Rosa, Jan Mendling, and A Hajo Reijers. 2018. Fundamentals of business process management. Springer. find
  5. Elena Viorica Epure, Patricia Martín-Rodilla, Charlotte Hug, Rebecca Deneckère, and Camille Salinesi. 2015. Automatic process model discovery from textual methodologies. In 2015 IEEE 9th international conference on research challenges in information science (RCIS), pages 19–30. IEEE. find
  6. Fabian Friedrich, Jan Mendling, and Frank Puhlmann. 2011. Process model generation from natural language text. In International conference on advanced information systems engineering, pages 482–496. Springer. find
  7. Joachim Herbst and DIMITRIS Karagiannis. 1999. An inductive approach to the acquisition and adaptation of workflow models. In Proceedings of the IJCAI, volume 99, pages 52–57. Citeseer. find
  8. Krzysztof Honkisz, Krzysztof Kluza, and Piotr Wiśniewski. 2018. A concept for generating business process models from natural language description. In International Conference on Knowledge Science, Engineering and Management, pages 91– 103. Springer. find
  9. Shengran Hu, Cong Lu, and Jeff Clune. 2025. Automated design of agentic systems. In The Thirteenth International Conference on Learning Representations.
  10. Knowing-how & knowing-that: A new task for machine comprehension of user manuals. arXiv preprint arXiv:2306.04187. find
  11. Xuan Liu, John Ahmet Erkoyuncu, Jerry Ying Hsi Fuh, Wen Feng Lu, and Bingbing Li. 2025. Knowledge extraction for additive manufacturing process via named entity recognition with llms. Robotics and Computer-Integrated Manufacturing, 93:102900. find
  12. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
  13. Hugo A López, Rasmus Strømsted, Jean-Marie Niy-odusenga, and Morten Marquard. 2021. Declarative process discovery: Linking process and textual views. In International Conference on Advanced Information Systems Engineering, pages 109–117. Springer. find
  14. Xuezhe Ma and Eduard Hovy. 2016. End-to-end sequence labeling via bi-directional LSTM-CNNs-CRF. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1064–1074, Berlin, Germany. Association for Computational Linguistics.
  15. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023a. Self-refine: iterative refinement wit… find
  16. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, and 1 others. 2023b. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36:46534–46594. find
  17. Bilal Maqbool, Farooque Azam, Muhammad Waseem Anwar, Wasi Haider Butt, Jahan Zeb, Iqra Zafar, Aiman Khan Nazir, and Zuneera Umair. 2018. A comprehensive investigation of bpmn models generation from textual requirements—techniques, tools and trends. In International Conference on Information Science… find
  18. Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and Philip S Yu. 2021. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE transactions on neural networks and learning systems, 33(2):494–514. find
  19. Julian Neuberger, Lars Ackermann, and Stefan Jablonski. 2023. Beyond rule-based named entity recognition and relation extraction for process model generation from natural language text. In International Conference on Cooperative Information Systems, pages 179–197. Springer. find
  20. Hongru Liang, Jia Liu, Weihong Du, Dingnan Jin, Wenqiang Lei, Zujie Wen, and Jiancheng Lv. 2023. find
  21. Kuntal Kumar Pal, Kazuaki Kashihara, Pratyay Banerjee, Swaroop Mishra, Ruoyu Wang, and Chitta Baral. 2021. Constructing flow graphs from procedural cybersecurity texts. arXiv preprint arXiv:2105.14357. find
  22. Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. 2024. Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering, 36(7):3580–3599.
  23. Haopeng Ren, Yushi Zeng, Yi Cai, Bihan Zhou, and Zetao Lian. 2023a. Constructing procedural graphs with multiple dependency relations: A new dataset and baseline. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8474– 8486, Toronto, Canada. Association for Computational… find
  24. Haopeng Ren, Yushi Zeng, Yi Cai, Bihan Zhou, and Zetao Lian. 2023b. Constructing procedural graphs with multiple dependency relations: A new dataset and baseline. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8474– 8486. find
  25. Florian Rupp and Kai Eckert. 2024. G-pcgrl: Procedural graph data generation via reinforcement learning. In 2024 IEEE Conference on Games (CoG), pages 1–8. IEEE. find
  26. Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning, 2023. URL https://arxiv.org/abs/2303.11366, 1. ↗
  27. Sholiq Sholiq, Riyanarto Sarno, and Endang Siti Astuti. 2022. Generating bpmn diagram from textual requirements. Journal of King Saud University-Computer and Information Sciences, 34(10):10079–10093. find
  28. Riad Sonbol, Ghaida Rebdawi, and Nada Ghneim. 2023. A machine translation like approach to generate business process model from textual description. SN Computer Science, 4(3):291. find
  29. Xingwei Tan, Yuxiang Zhou, Gabriele Pergola, and Yulan He. 2024a. Cascading large language models for salient event graph generation. arXiv preprint arXiv:2406.18449. find
  30. Xingwei Tan, Yuxiang Zhou, Gabriele Pergola, and Yulan He. 2024b. Set-aligning framework for autoregressive event temporal graph generation. arXiv preprint arXiv:2404.01532. — [arXiv:2404.01532](https://arxiv.org/abs/2404.01532) find
  31. Niket Tandon, Keisuke Sakaguchi, Bhavana Dalvi Mishra, Dheeraj Rajagopal, Peter Clark, Michal Guerquin, Kyle Richardson, and Eduard Hovy. 2020. A dataset for tracking entities in open domain procedural text. arXiv preprint arXiv:2011.08092. find
  32. Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, and 3 others. 2025. A survey of large language models. Prepri…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ying-2026-multi-agent-procedural-graph-extraction/