Elamur Research⌕
☰
home / papers / 2026 / LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining

LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining

Paper · 2026 · Phuong Huu Vu Tran, Long Minh Vo, Son Nguyen Minh Le et al.

Summary

LLM-INSTRUCT won the UZH Shared Task at ArgMining 2026. The task asks systems to analyze paragraphs of UN and UNESCO resolutions in three ways: label each paragraph as preambular or operative, assign a subset of 141 official thematic tags, and predict labeled directed links between paragraphs. Output must follow a strict JSON schema, and only open-weight models of up to 8B parameters are allowed.

The authors treat this as constrained structured prediction rather than free generation. Paragraph type comes from a rule-based heuristic. For tags, intfloat/e5-base-v2 retrieves candidates using prototype strings built from organizer metadata. Qwen3-8B then chooses only from that retrieved set, under caps on the number of tags. Ambiguous cases go to a three-model debate. A final validation layer checks schema compliance.

The system ranked first overall. During development, its Task 1b Micro-F1 rose from 35.83% to 40.08%. The main takeaway is that shrinking the label space before generation improves both accuracy and the reliability of valid submissions. Recognition of rare tags remains weak.

Key points

  • Data:
    • The training release is unlabelled and comes from 2,695 UN resolutions in the UN-RES corpus (linked to SpiritRAG).
    • Evaluation uses a held-out UNESCO split.
    • Texts are in French, with English translations from the organizers. The system reads English when available.
  • Official ranking: 1st overall, 1st in F1, and 5th in LLM-as-a-Judge. The internal Task 2 score was 4.421.
  • Tag controls:
    • Each tag is represented by a Dimension | Category | CODE prototype.
    • Each paragraph gets at most 5 tags, with no more than 2 per dimension.
    • The prompt includes up to 3 retrieved examples with similarity above 0.70.
  • Ablations:
    • Using CODE-only prototypes cut subset F1 by 7.74 points.
    • Dropping retrieved examples cut subset F1 by 5.17 points.
    • Tags with at most five occurrences reached only 2.51% F1.
  • Debate and relations:
    • The debate agents are Qwen3-8B, Llama-3.1-8B-Instruct and Ministral-8B-Instruct.
    • Relation candidates combine a one-paragraph locality window with the top-6 most similar paragraphs. Edges need confidence ≥0.40, with at most 5 per source paragraph.
    • The final output has 13,323 edges, a graph density of 10.03%.
    • The full run took about 1h42m on 2×RTX 3090 GPUs.

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

Also on the byline: RMIT Vietnam.

 Fields, topics and concepts

📚 Shelves 6The 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 worksSocial Sciences · 1,772 worksPolitical Science and International Relations · 182 works

🧭 Topics 1Research subjects and languages the work is about.Information retrieval & search · 473 works

Cite

Cite as Phuong Huu Vu Tran, Long Minh Vo, Son Nguyen Minh Le and Hoang Van. 2026. LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining. Proceedings of the 13th Workshop on Argument Mining and Reasoning.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.argmining-1.11

@inproceedings{tran2026llm,
  title = {{LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining}},
  author = {Phuong Huu Vu Tran, Long Minh Vo, Son Nguyen Minh Le and Hoang Van},
  booktitle = {Proceedings of the 13th Workshop on Argument Mining and Reasoning},
  year = {2026},
  doi = {10.18653/v1/2026.argmining-1.11},
  url = {https://doi.org/10.18653/v1/2026.argmining-1.11}
}

References

15 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jérémie Cabessa, Hugo Hernault, and Umer Mushtaq. 2025. Argument mining with fine-tuned large language models. In Proceedings of the 31st International Conference on Computational Linguistics, pages 6624–6635, Abu Dhabi, UAE. Association for Computational Linguistics. find
  2. Lang Cao. 2023. Autoam: An end-to-end neural model for automatic and universal argument mining. In Advanced Data Mining and Applications, pages 517– 531, Cham. Springer Nature Switzerland. find
  3. Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. 2024. Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org.
  4. Yingqiang Gao, Fabian Winiger, Patrick Montjourides, Anastassia Shaitarova, Nianlong Gu, Simon Peng-Keller, and Gerold Schneider. 2025. SpiritRAG: A Q&A System for Religion and Spirituality in the United Nations Archive. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language… find
  5. Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. 2023. Grammar-constrained decoding for structured NLP tasks without finetuning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10932– 10952, Singapore. Association for Computational Lingu… find
  6. Deniz Gorur, Antonio Rago, and Francesca Toni. 2025. Can large language models perform relation-based argument mining? In Proceedings of the 31st International Conference on Computational Linguistics, pages 8518–8534, Abu Dhabi, UAE. Association for Computational Linguistics. find
  7. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 o… — The Llama 3 Herd of Models
  8. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769–6781, Onli…
  9. John Lawrence and Chris Reed. 2020. Argument mining: A survey. Computational Linguistics, 45(4):765– 818.
  10. Tianyu Liu, Yuchen Eleanor Jiang, Nicholas Monath, Ryan Cotterell, and Mrinmaya Sachan. 2022. Autoregressive structured prediction with language models. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 993–1005, Abu Dhabi, United Arab Emirates. Association for Computat… find
  11. Mistral AI. 2024. Model card for ministral-8b-instruct-2410. Accessed 2026-04-16. find
  12. Giovanni Paolini, Ben Athiwaratkun, Jason Krone, Jie Ma, Alessandro Achille, Rishita Anubhai, Cicero Nogueira dos Santos, Bing Xiang, and Stefano Soatto. 2021. Structured prediction as translation between augmented natural languages. Preprint, arXiv:2101.05779. find
  13. Christian Stab and Iryna Gurevych. 2017. Parsing argumentation structures in persuasive essays. Computational Linguistics, 43(3):619–659.
  14. Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533.
  15. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tran-2026-llm-instruct-uzh-shared-task/