Yuanhe Zhang


I am a PhD student in Statistics at University of Warwick, supervised by Dr. Fanghui Liu, Dr. Thomas Berrett, and Prof. Chenlei Leng.

I am interested in formal mathematical reasoning and theory-grounded algorithm design for post-training of LLMs. [CV]

Email: firstname.lastname@warwick.ac.uk

Yuanhe Zhang Photo

Publications

New Statistical Learning Theory in Lean 4: Empirical Processes from Scratch

Lean4-SLT

Yuanhe Zhang, Jason D. Lee, Fanghui Liu
Preprint
[PDF]  [Code]  [HF Dataset]  [Bibtex]  
@article{zhang2026statistical,
  title={Statistical Learning Theory in Lean 4: Empirical Processes from Scratch},
  author={Zhang, Yuanhe and Lee, Jason D and Liu, Fanghui},
  journal={arXiv preprint arXiv:2602.02285},
  year={2026}
}

TLDR: We present the first large-scale Lean 4 formalization of statistical learning theory (SLT) grounded in empirical process theory from scratch. It covers scross High-Dimensional Gaussian Analysis Toolbox, Dudley’s Entropy Integral Bound, and sharp regression rates. Building upon our library, we present a high-quality Lean 4 training dataset for LLM's formal reasoning.

New DAG-Math: Graph-Guided Mathematical Reasoning in LLMs

DAG-MATH

Yuanhe Zhang, Ilja Kuzborskij, Jason D. Lee, Chenlei Leng, Fanghui Liu
International Conference on Learning Representations (ICLR), 2026
[PDF]  [Code]  [HF Dataset]  [Bibtex]  
@article{zhang2025dag,
  title={DAG-Math: Graph-Guided Mathematical Reasoning in LLMs},
  author={Zhang, Yuanhe and Kuzborskij, Ilja and Lee, Jason D and Leng, Chenlei and Liu, Fanghui},
  journal={arXiv preprint arXiv:2510.19842},
  year={2025}
}

TLDR: We propose a new framework by modeling CoT as a rule-based stochastic process on directed acyclic graphs (DAGs), introduce the concept of logic closeness, and then precisely evaluates the mathematical reasoning ability of LLMs via the proposed DAG-MATH format.

LoRA-One: One-Step Full Gradient Could Suffice for Fine-Tuning Large Language Models, Provably and Efficiently

LoRA-One Demo Animation

Yuanhe Zhang, Fanghui Liu, Yudong Chen
International Conference on Machine Learning (ICML), 2025 (Oral, top 120/12107=1.0% of papers).
[PDF]  [Code]  [Website]  [Bibtex]  
@InProceedings{pmlr-v267-zhang25ax,
  title = 	 {{L}o{RA}-One: One-Step Full Gradient Could Suffice for Fine-Tuning 
		Large Language Models, Provably and Efficiently},
  author =       {Zhang, Yuanhe and Liu, Fanghui and Chen, Yudong},
  booktitle = 	 {Proceedings of the 42nd International Conference on Machine Learning},
  pages = 	 {75513--75574},
  year = 	 {2025},
  volume = 	 {267},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {13--19 Jul},
  publisher =    {PMLR},
  url = 	 {https://proceedings.mlr.press/v267/zhang25ax.html},
}

TLDR: We show how theory (from subspace alignment to gradient dynamics as well as preconditioners) contributes to fine-tuning algorithm design in practice.

Quasi-Bayes meets Vines
David Huk, Yuanhe Zhang, Ritabrata Dutta, Mark Steel
Advances in Neural Information Processing Systems (NeurIPS), 2024.
[PDF]  [Bibtex]  

@article{huk2024quasi,
  title={Quasi-Bayes meets Vines},
  author={Huk, David and Zhang, Yuanhe and Dutta, Ritabrata and Steel, Mark},
  journal={Advances in Neural Information Processing Systems},
  volume={37},
  pages={40359--40392},
  year={2024}
}

Education