Elamur Research⌕
☰
home / papers / 2026 / Distill and Align Decomposition for Enhanced Claim Verification

Distill and Align Decomposition for Enhanced Claim Verification

Paper · 2026 · Jabez Magomere, Elena Kochkina, Samuel Mensah et al.

Summary

The paper tackles a weak spot in decompose-then-verify fact-checking of long LLM outputs. The step that splits sentences into subclaims is usually tuned separately from the verifier. It can therefore produce claims that are too fragmented, too coarse, or missing context for the verifier to judge well. The authors propose Distill-Align-Decompose (DAD), a three-part training recipe for an 8B decomposer:

  • Structured reasoning: decomposition is framed as a four-step reasoning chain inside one model call (claim detection, decontextualization, relationship identification, claim extraction).
  • Warm start: the model is first fine-tuned on decompositions generated by a larger teacher model.
  • RL alignment: it is then trained with GRPO, using a combined reward for output format, verifier correctness, and a decomposition-quality checklist.

Across six evaluation settings, the trained model raises downstream verification to 71.75% macro-F1. It beats prompt-based baselines and an earlier RL method, and it produces a moderate number of subclaims. Human raters scored its subclaims highly on the quality criteria. The work matters because it shows a relatively small open model can reach state-of-the-art claim verification without costly multi-stage pipelines.

Key points

  • Gains in macro-F1: +1.99pp over VeriScore, +5.84pp over the RL method DyDecomp, and +2.58pp over the same model used with prompting only.
  • Average subclaims per sentence: 8.14, compared with 22.92 for FActScore (over-decomposition) and 1.66 for DyDecomp (under-decomposition).
  • Each sentence is decomposed together with the original question and a symmetric local context of 2 preceding and 2 following sentences. VeriScore uses an asymmetric 3/1 window.
  • Quality criteria include atomicity alignment with the verifier, verifiability, entailment, coverage, and decontextualization. Verifier accuracy serves as a proxy for the right atomicity.
  • Stated limitations:
    • only one fixed verifier was tested;
    • training exemplars came from a single teacher model;
    • experiments were English-only and used Wikipedia/Google Search as evidence sources;
    • no multi-hop benchmarks were included.

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

🎓University of OxfordOxford · United Kingdom104 works · 157,183 citations🏢JPMorgan Chase & CoNew York · United States2 works · 3 citations

 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 worksSociology and Political Science · 577 works

Cite

Cite as Jabez Magomere, Elena Kochkina, Samuel Mensah, Simerjot Kaur, Fernando Acero, Arturo Oncevay, Charese Smiley, Xiaomo Liu and Manuela Veloso. 2026. Distill and Align Decomposition for Enhanced Claim Verification. 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.309

@inproceedings{magomere2026distill,
  title = {{Distill and Align Decomposition for Enhanced Claim Verification}},
  author = {Jabez Magomere, Elena Kochkina, Samuel Mensah, Simerjot Kaur, Fernando Acero, Arturo Oncevay, Charese Smiley, Xiaomo Liu and Manuela Veloso},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.309},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.309}
}

References

26 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Isabelle Augenstein, Timothy Baldwin, Meeyoung Cha, Tanmoy Chakraborty, Giovanni Luca Ciampaglia, David Corney, Renee DiResta, Emilio Ferrara, Scott Hale, Alon Halevy, and 1 others. 2024. Factuality challenges in the era of large language models and opportunities for fact-checking. Nature Machine In…
  2. William Brown. 2025. Verifiers: Environments for llm reinforcement learning. https://github.com/ willccbb/verifiers. ↗
  3. Shiqi Chen, Yiran Zhao, Jinghan Zhang, I-Chun Chern, Siyang Gao, Pengfei Liu, and Junxian He. 2023. Felm: Benchmarking factuality evaluation of large language models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track. find
  4. Mehul Damani, Isha Puri, Stewart Slocum, Idan Shenfeld, Leshem Choshen, Yoon Kim, and Jacob Andreas. 2025. Beyond binary rewards: Training lms to reason about their uncertainty. Preprint, arXiv:2507.16806. find
  5. Michael Han Daniel Han and Unsloth team. 2023. Unsloth. find
  6. 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
  7. Anisha Gunjal and Greg Durrett. 2024. Molecular facts: Desiderata for decontextualization in LLM fact verification. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3751–3768, Miami, Florida, USA. Association for Computational Linguistics. find
  8. Dasha Metropolitansky and Jonathan Larson. 2025. Towards effective extraction and evaluation of factual claims. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6996–7045, Vienna, Austria. Association for Computational Linguist…
  9. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  10. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural L…
  11. Qisheng Hu, Quanyu Long, and Wenya Wang. 2025. Decomposition dilemmas: Does claim decomposition boost or burden fact-checking performance? In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volum… find
  12. Ryo Kamoi, Tanya Goyal, Juan Diego Rodriguez, and Greg Durrett. 2023. WiCE: Real-World Entailment for Claims in Wikipedia. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7561–7583, Singapore. Association for Computational Linguistics.
  13. Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2024a. The dawn after the dark: An empirical study on factuality hallucination in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long…
  14. Miaoran Li, Baolin Peng, Michel Galley, Jianfeng Gao, and Zhu Zhang. 2024b. Self-checker: Plug-and-play modules for fact-checking with large language models. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 163–181, Mexico City, Mexico. Association for Computational Li…
  15. Xin Liu, Lechen Zhang, Sheza Munir, Yiyang Gu, and Lu Wang. 2025a. Verifact: Enhancing long-form factuality evaluation with refined fact extraction and reference facts. Preprint, arXiv:2505.09701.
  16. Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025b. Understanding r1-zero-like training: A critical perspective. In Conference on Language Modeling (COLM).
  17. Yining Lu, Noah Ziems, Hy Dang, and Meng Jiang. 2025. Optimizing decomposition for optimal claim verification. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5095–5114, Vienna, Austria. Association for Computational Linguisti… find
  18. Rishanth Rajendhran, Amir Zadeh, Matthew Sarte, Chuan Li, and Mohit Iyyer. 2025a. Verifastscore: Speeding up long-form factuality evaluation. Preprint, arXiv:2505.16973. find
  19. Rishanth Rajendhran, Amir Zadeh, Matthew Sarte, Chuan Li, and Mohit Iyyer. 2025b. VeriFastScore: Speeding up long-form factuality evaluation. arXiv preprint. ArXiv:2505.16973 [cs]. find
  20. Wooseok Seo, Seungju Han, Jaehun Jung, Benjamin Newman, Seungwon Lim, Seungbeen Lee, Ximing Lu, Yejin Choi, and Youngjae Yu. 2025. Verifying the verifiers: Unveiling pitfalls and potentials in fact verifiers. In Second Conference on Language Modeling. find
  21. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
  22. Yixiao Song, Yekyung Kim, and Mohit Iyyer. 2024. VERISCORE: Evaluating the factuality of verifiable claims in long-form text generation. arXiv preprint. ArXiv:2406.19276 [cs].
  23. Liyan Tang, Philippe Laban, and Greg Durrett. 2024. MiniCheck: Efficient fact-checking of LLMs on grounding documents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8818–8847, Miami, Florida, USA. Association for Computational Linguistics.
  24. Miriam Wanner, Benjamin Van Durme, and Mark Dredze. 2024a. Dndscore: Decontextualization and decomposition for factuality verification in long-form text generation. Preprint, arXiv:2412.13175. find
  25. Miriam Wanner, Seth Ebner, Zhengping Jiang, Mark Dredze, and Benjamin Van Durme. 2024b. A closer look at claim decomposition. In Proceedings of the 13th Joint Conference on Lexical and Computational Semantics (*SEM 2024), pages 153–175, Mexico City, Mexico. Association for Computational Linguistics. find
  26. Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V. Le. 2024. Long-form factuality in large language models. In Advances in Neural Information Processing Systems, volume 37, pages 80756–80827. Curran Associates,…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/magomere-2026-distill-align-decomposition-enhanced-claim/