Shucheng Liu

Welcome to my homepage! I am a second-year Ph.D. student in Statistics at the University of North Carolina at Chapel Hill, advised by Prof. Hongtu Zhu and Prof. Kai Zhang. I also work closely with Prof. Chengchun Shi. Before UNC, I received my B.S. in Statistics from the University of Science and Technology of China in 2025. During my undergraduate study, I was fortunate to be advised by Prof. Jing Zeng and Prof. Hong Zhang.

Previously, I was a large language model research intern at Tencent YouTu Lab. I was also a summer research intern at the University of Wisconsin–Madison, advised by Prof. Guanhua Chen.

My research interests include:

Generative modeling. Diffusion and flow-matching models and low-dimensional subspace learning from complex, high-dimensional data.
Foundation models. Post-training of large language and multimodal models (RL, reward design), LLM agents, representation learning and knowledge integration with applications to biomedical data.
Causal inference. Robust treatment effect estimation from high-dimensional observational data to support reliable decision-making.

I’m open to exploring any area that matters. I am currently seeking internship opportunities in Summer 2027.

shucheng@unc.edu LinkedIn Google Scholar GitHub

Internship Experience

Tencent YouTu Lab2025.02 - 2025.05

Large Language Model Research Intern

Designed a difficulty-aware RL reward for post-training large reasoning models, framing when-to-stop as a sequential decision problem (keep thinking, or stop and answer): a learned question-difficulty coefficient modulates the length penalty so answers stay brief on easy questions while deliberation is preserved on hard ones; characterized when the length-penalized reward drifts or destabilizes training.
Analyzed how large reasoning models (QwQ-32B, DeepSeek-R1-Distill-32B) allocate inference-time compute: information-theoretic metrics (InfoBias, InfoGain) show that longer chains of thought accumulate bias with diminishing per-step gain; contributed to an entropy-based adaptive stopping strategy that cut token usage by about 50% across six reasoning benchmarks while slightly improving accuracy.

Publications

First-author papers are highlighted. Click a figure to enlarge.

click to enlarge	FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model Shucheng Liu, Chengchun Shi, Kai Zhang, Hongtu Zhu NeurIPS 2026 [arXiv] A flow-matching model that generates the clean signal instead of noise or velocity, with a theoretical proof that its approximation error scales with the intrinsic rather than the ambient dimension. A history-conditioned Transformer is paired with a multi-graph GCN fusing three anatomical priors through a learnable Laplacian polynomial.
click to enlarge	Weighted Spline-Expanded Networks with Distributional Balancing for Continuous Treatment Effects Shucheng Liu, Chan Park, Guanhua Chen Under review [arXiv] WSENet deconfounds through distance-covariance balancing weights instead of estimating or inverting a generalized propensity score, and models the dose-response with a spline-expanded embedding. Weighted targeted regularization from the efficient influence function yields a doubly robust estimator; best integrated RMSE on IHDP, News and TCGA.
click to enlarge	MoSAIC: Motion and Supervision-Aware Inference for Myocardial Scar and Edema Segmentation Shucheng Liu, Mingcheng Hu, Liangqiao Gui, Yineng Chen, Peng Huang, Hui Guo, Shu Hu, Xin Wang, Hongtu Zhu MICCAI CARE 2026 (Oral, top 10%) [PDF] A coarse-to-fine framework across centers with missing modalities and partial labels, combining anatomical priors, modality gating, and masked supervision. Presence gates and supervision masking pool ragged multi-center cohorts; a cine branch reads scar from motion alone.
click to enlarge	Progressive Optimization of UU-Mamba for Left Atrial Multi-Structure Segmentation in CT Yineng Chen, Liangqiao Gui, Shucheng Liu, Mingcheng Hu, Peng Huang, Hui Guo, Shu Hu, Hongtu Zhu, Xin Wang MICCAI CARE 2026 [PDF] A Mamba-based semi-supervised segmentation pipeline; logit adjustment for class imbalance and automated pseudo-label quality control lift mean Dice.

Service and Teaching

Reviewer: NeurIPS 2026, AISTATS 2026, 2027

Lab Instructor & Teaching Assistant: STOR 120 Foundations of Statistics and Data Science, UNC-Chapel Hill

Teaching Assistant: STOR 320 Methods and Models of Data Science; STOR 415 Introduction to Optimization; STOR 435 Introduction to Probability, UNC-Chapel Hill

✅ Copied

Keep Exploring