Xuhan Huang

MPhil in Computer Science

CUHK-Shenzhen

About

I am an MPhil student in Computer Science at CUHK-Shenzhen. I completed my undergraduate studies in Applied Mathematics there, where I was advised by Professors Benyou Wang and Zhongxiang Dai. I am currently mentored by Professor Jie Fu.

My research interests lie broadly in AI safety, with a focus on mechanistic understanding and verifiable reasoning. I am interested in understanding the internal mechanisms that shape model behavior, and in using verifiable feedback to support more reliable and scalable training and evaluation.

Email: xuhanhuang@link.cuhk.edu.cn

Links: Google Scholar GitHub

Recent Work

All research ↗

Sparse Weight Decomposition for Efficient Circuit Extraction

Chuanhao Yan^*, Xuhan Huang^*, Yawen Duan, Zhenfei Yin, Hang Zhao, Bryan Dai, Jie Fu

Extracting faithful model circuits through a sparse decomposition of dense neural-network weights.

Preprint · 2026 Paper Code Blog

Re:Form reinforcement-learning pipeline connecting policy sampling, specification-subset reward, verification reward, and policy updates.

Re:Form—Reducing Human Annotations in Scalable Formal Software Verification with RL in LLMs: A Preliminary Study on Dafny

Chuanhao Yan^*, Fengdi Che^*, Xuhan Huang^*, Xu Xu^*, Xin Li^*, et al.

Training code agents with objective feedback from a formal verifier.

TMLR · 2026 Paper Code

Comparison of FLDB-GD and FLDB-OGD communication patterns derived from the original Draw.io figure.

Federated Linear Dueling Bandits

Xuhan Huang, Yan Hu, Zhiyan Li, Zhiyong Wang, Benyou Wang, Zhongxiang Dai

Learning from distributed pairwise preferences while reducing communication rounds.

AAAI · 2026 Paper

MAMO pipeline in which a language model produces a mathematical model whose solved answer is checked against the ground answer.

LLMs for Mathematical Modeling: Towards Bridging the Gap between Natural and Mathematical Languages

Xuhan Huang, Qingning Shen, Yan Hu, Anningzhe Gao, Benyou Wang

Evaluating whether language models can translate real-world problems into executable mathematical models.

NAACL Findings · 2025 Paper Code

* Equal contribution.

More Research

VeriEquivBench: An Equivalence Score for Ground-Truth-Free Evaluation of Formally Verifiable Code ICLR · 2026

Lingfei Zeng^*, Fengdi Che^*, Xuhan Huang, Fei Ye, Xu Xu, Binhang Yuan, Jie Fu

Paper Code

CALM Before the STORM: Unlocking Native Reasoning for Optimization Modeling ICML · 2026

Zhengyang Tang^*, Zihan Ye^*, Chenyu Huang^*, Xuhan Huang, Chengpeng Li, Sihang Li, Guanhua Chen, Ming Yan, Zizhuo Wang, Hongyuan Zha, Dayiheng Liu, Benyou Wang

Paper

Differentiable Evolutionary Reinforcement Learning ICLR RSI · 2026 Poster

Sitao Cheng^*, Tianle Li^*, Xuhan Huang^*, Xunjian Yin, Difan Zou

Paper Code

* Equal contribution.