Linxuan Wang

Department of Statistics · Purdue University

I am a third-year PhD student in Statistics at Purdue University, advised by Prof. Guang Lin and Prof. Qifan Song.

My research focuses on efficient large language models, ranging from diffusion language models to speculative decoding. I develop pretraining methods that teach diffusion language models to self-correct, and tree-based speculative decoding methods that accelerate LLM inference.

In summer 2026, I was a research intern with the Inclusion AI LLaDA team at Ant Group, mentored by Dr. Jianguo Li.

Previously, I received my M.S. in Statistics from Duke University, where I worked with Prof. Surya Tokdar, and my B.S. in Mathematics and B.A. in Economics from Wuhan University.

Email: wang6559 [at] purdue [dot] edu

GitHubLinkedInCV

Portrait of Linxuan Wang

News

Sep 2026UBTree, our work on parallel tree drafting for speculative decoding, is available. Paper ↗
Sep 2026Our work on adaptive Hamiltonian Monte Carlo for federated sampling was accepted to NeurIPS 2026.
Apr 2026Generalized Discrete Diffusion with Self-Correction was accepted to ICML 2026. Project ↗

Publications & Preprints

* Equal contribution.

UBTree architecture: parallel unigram candidate proposals and bigram transition scoring construct a verification tree.

UBTree: Parallel Tree Drafting via Unigram and Bigram Models for Speculative Decoding

C. Liang*, Linxuan Wang*, X. Peng*, H. Liu, Y. Chen, G. Liu, G. Lin, Q. Song, J. Li

Preprint, 2026

Parallel tree drafting with a bigram selector and tree-native training for efficient speculative decoding.

SCDD training and inference: clean tokens are corrupted, a denoiser learns token predictions with an ELBO objective, and generation uses ancestral sampling.

Generalized Discrete Diffusion with Self-Correction

Linxuan Wang*, Z. Wang*, Y. Bai*, W. Deng, G. Lin, Q. Song

ICML 2026

Learning remask-free self-correction during discrete diffusion pretraining to improve parallel language generation.

Paper Code Project NVIDIA NeMo

SE(3)-MeanFlow few-step inference: an average-velocity network transforms a noisy protein backbone into a generated backbone using SE(3) updates.

SE(3)-MeanFlow: Few-Step Protein Backbone Generation on Lie Groups

Y. Bai, B. Lu, Y. Liu, E. Akbari, S. Kolouri, Linxuan Wang, P. He, S. Wang, R. Zhang, G. Lin

Preprint, 2026

Centralized and federated Hamiltonian Monte Carlo: local leapfrog trajectories alternate with communication and synchronization across clients.

Breaking the √d Communication Barrier in Federated Sampling with Adaptive Hamiltonian Monte Carlo

J. Liang, Linxuan Wang, G. Lin, Q. Song

NeurIPS 2026

Bayesian Sparse Learning of Joint Quantile Planes Estimation

Linxuan Wang

Master’s thesis · Duke University, 2024

Experience

Research Intern

Ant Group · Inclusion AI LLaDA Team

May – Sep 2026

Education

Purdue University

PhD in Statistics

2024 – 2028 (expected)

Duke University

MS in Statistics

2022 – 2024

Wuhan University

BS in Mathematics & BA in Economics

2018 – 2022

Academic Service

Conference reviewer: ICLR 2027.