Elamur Research⌕
☰
home / papers / 2026 / Revealing Procedural Reasoning Structures in Chain-of-Thought Training via Span-Level Gradient Organization

Revealing Procedural Reasoning Structures in Chain-of-Thought Training via Span-Level Gradient Organization

Paper · 2026 · Jia Liu, Jiaxin Luo, Weiwen Xu et al.

Summary

The paper asks how step-by-step reasoning patterns in Chain-of-Thought (CoT) training data appear in the training signal itself, rather than in generated outputs or hidden representations. The authors propose Gradient-based Structural Developer (GSD), an unsupervised diagnostic framework. GSD computes per-sample gradients while a student model is fine-tuned on public reasoning benchmarks in a distillation-style setup, then sums token-level gradient contributions over short contiguous spans.

Single-token gradients prove noisy, but span-aggregated gradients point in consistent directions across examples. The authors call these recurring patterns "aligned sequential stresses." This alignment often coincides with semantic similarity. It also links lexically different segments that play the same procedural role, yet segments that share meaning do not necessarily produce gradients that align.

As secondary evidence, fine-tuning on small subsets with high gradient consistency gives better accuracy than comparison subsets of the same size. The work matters because it offers an annotation-free way to inspect how procedural reasoning structure is reinforced in parameter space, complementing behavioral evaluation.

Key points

  • GSD needs no output supervision, predefined reasoning templates or semantic labels. A preliminary study took gradients from the first ten tokens of each example (the reasoning prefix) and visualized them with t-SNE. Examples with similar procedural organization clustered together.
  • A signal–noise analysis argues that the span is the smallest aggregation scale at which a stable alignment signal emerges.
  • In the auxiliary selection experiment, training on N = 16/32/64/128 high-consistency samples reached 59.36%, 59.74%, 62.70% and 64.37%. The comparison subsets reached 49.36–60.35%, so the gains were +10.00, +8.49, +4.02 and +4.02 pp.
  • Training on the full 800-sample set gave 65.05%, barely above the 128-sample subset. This points to diminishing returns beyond N = 128.
  • The authors list several limitations:
    • GSD is diagnostic, not a full causal model of how reasoning is acquired.
    • Alignment reflects procedural consistency, not the correctness of the reasoning.
    • Linear aggregation may miss nonlinear or weakly aligned effects.
    • The selection results should not be read as a practical data-selection algorithm.

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🎓Huazhong University of Science and TechnologyWuhan · China45 works · 6,345 citations🎓South China University of TechnologyGuangzhou · China22 works · 1,554 citations🎓University of Nottingham Ningbo ChinaNingbo · China3 works · 216 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

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Jia Liu, Jiaxin Luo, Weiwen Xu, Jonathan M. Garibaldi, Xiao-Kun Wu, Yixue Hao and Min Chen. 2026. Revealing Procedural Reasoning Structures in Chain-of-Thought Training via Span-Level Gradient Organization. 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.1754

@inproceedings{liu2026revealing,
  title = {{Revealing Procedural Reasoning Structures in Chain-of-Thought Training via Span-Level Gradient Organization}},
  author = {Jia Liu, Jiaxin Luo, Weiwen Xu, Jonathan M. Garibaldi, Xiao-Kun Wu, Yixue Hao and Min Chen},
  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.1754},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1754}
}

References

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

  1. Stéphane Boucheron, Gábor Lugosi, and Pascal Massart. 2013. Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press. find
  2. Tadeusz Caliński and Jerzy Harabasz. 1974. A dendrite method for cluster analysis. Communications in Statistics-theory and Methods, 3(1):1–27. find
  3. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
  4. David L Davies and Donald W Bouldin. 2009. A cluster separation measure. IEEE transactions on pattern analysis and machine intelligence, (2):224–227. find
  5. Yuntian Deng, Kiran Prasad, Roland Fernandez, Paul Smolensky, Vishrav Chaudhary, and Stuart Shieber. 2023. Implicit chain of thought reasoning via knowledge distillation. arXiv preprint arXiv:2311.01460. find
  6. Subhabrata Dutta, Joykirat Singh, Soumen Chakrabarti, and Tanmoy Chakraborty. 2024. How to think step-by-step: A mechanistic understanding of chain-of-thought reasoning. arXiv preprint arXiv:2402.18312. find
  7. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  8. Shangmin Guo, Yi Ren, Stefano V Albrecht, and Kenny Smith. 2024. lpntk: Better generalisation with less data via sample interaction during learning. arXiv preprint arXiv:2401.08808. — [arXiv:2401.08808](https://arxiv.org/abs/2401.08808) find
  9. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769. find
  10. 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.
  11. Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. 2020. Estimating training data influence by tracing gradient descent. Advances in Neural Information Processing Systems, 33:19920– 19930. find
  12. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728– 53741.
  13. Leonardo Ranaldi, Marco Valentino, and Andre Freitas. 2025. Improving chain-of-thought reasoning via quasi-symbolic abstractions. pages 17222–17240.
  14. Jianhao Huang, Zixuan Wang, and Jason D Lee. 2025. Transformers learn to implement multi-step gradient descent with chain of thought. arXiv preprint arXiv:2502.21212. — [arXiv:2502.21212](https://arxiv.org/abs/2502.21212) find
  15. Yi Ren, Shangmin Guo, Linlu Qiu, Bailin Wang, and Danica J Sutherland. 2024. Bias amplification in language model evolution: An iterated learning perspective. Advances in Neural Information Processing Systems, 37:38629–38664. find
  16. Arthur Jacot, Franck Gabriel, and Clément Hongler. 2018. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31. find
  17. Yi Ren, Shangmin Guo, and Danica J Sutherland. 2022. Better supervisory signals by observing learning paths. arXiv preprint arXiv:2203.02485. — [arXiv:2203.02485](https://arxiv.org/abs/2203.02485) find
  18. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199– 22213.
  19. Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, and 1 others. 2023. Measuring faithfulness in chain-of-thought reasoning. arXiv preprint arXiv:2307.13702.
  20. Ayeong Lee, Ethan Che, and Tianyi Peng. 2025. How well do llms compress their own chain-of-thought? a token complexity approach. arXiv preprint arXiv:2503.01141. — [arXiv:2503.01141](https://arxiv.org/abs/2503.01141) find
  21. Jindong Li, Yali Fu, Li Fan, Jiahong Liu, Yao Shu, Chengwei Qin, Menglin Yang, Irwin King, and Rex Ying. 2025. Implicit reasoning in large language models: A comprehensive survey. arXiv preprint arXiv:2509.02350.
  22. Peter J Rousseeuw. 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53–65. find
  23. Ozan Sener and Silvio Savarese. 2017. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489. find
  24. Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  25. Roman Vershynin. 2018. High-Dimensional Probability: An Introduction with Applications in Data Science. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press. find
  26. Jia Liu, Yue Wang, Zhiqi Lin, Min Chen, Yixue Hao, and Long Hu. 2024. Natural language fine-tuning. arXiv preprint arXiv:2412.20382. — [arXiv:2412.20382](https://arxiv.org/abs/2412.20382) find
  27. Yiming Wang, Pei Zhang, Baosong Yang, Derek F Wong, and Rui Wang. 2024a. Latent space chain-of-embedding enables output-free LLM self-evaluation. arXiv preprint arXiv:2410.13640. — [arXiv:2410.13640](https://arxiv.org/abs/2410.13640) find
  28. José I Orlicki. 2025. Beyond words: A latent memory approach to internal reasoning in llms. arXiv preprint arXiv:2502.21030. — [arXiv:2502.21030](https://arxiv.org/abs/2502.21030) find
  29. Yu Wang, Shiwan Zhao, Zhihu Wang, Heyuan Huang, Ming Fan, Yubo Zhang, Zhixing Wang, Haijun Wang, and Ting Liu. 2024b. Strategic chain-of-thought: Guiding accurate reasoning in llms through strategy elicitation. arXiv preprint arXiv:2409.03271. — [arXiv:2409.03271](https://arxiv.org/abs/2409.03271) find
  30. Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652.
  31. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  32. Skyler Wu, Eric Meng Shen, Charumathi Badrinath, Jiaqi Ma, and Himabindu Lakkaraju. 2023. Analyzing chain-of-thought prompting in large language models via gradient-based feature attributions. arXiv preprint arXiv:2307.13339.
  33. Weihao Yu, Zihang Jiang, Yanfei Dong, and Jiashi Feng. 2020. Reclor: A reading comprehension dataset requiring logical reasoning. Preprint, arXiv:2002.04326.
  34. Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, and Noah D Goodman. 2024. Quiet-star: Language models can teach themselves to think before speaking.
  35. Hugh Zhang and David C Parkes. 2023. Chain-of-thought reasoning is a policy improvement operator. arXiv preprint arXiv:2309.08589. find
  36. Xuan Zhang, Chao Du, Tianyu Pang, Qian Liu, Wei Gao, and Min Lin. 2024. Chain of preference optimization: Improving chain-of-thought reasoning in llms. Advances in Neural Information Processing Systems, 37:333–356. find
  37. Contents \#\# Introduction Related Work 2.1 Chain-of-Thought Reasoning . . . 2.2 Training-Time Analysis of LLMs . B.2.1 2 2 3 Problem Definition and Preliminary Observations 3 Methodology 4.1 Training-time gradient decomposition and interpretation. . . . . . . 4.2 Span-level aggregation. . . . . . . 4… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2026-revealing-procedural-reasoning-structures-chain/