Highlighted Research

Project 1: Incorporating LLM Embeddings for Variation Across the Human Genome

Genomics · LLM embeddings
Hongqian Niu, Jordan G. Bryan, Jacob Williams, Hufeng Zhou, Haoyu Zhang, Xihao Li, Didong Li

Overview of the LLM embedding framework for human genome variation

A systematic framework for generating variant-level embeddings across the entire human genome using large language models, built from curated annotations (FAVOR, ClinVar, GWAS Catalog) at scales up to ~9 billion possible variants. Applied to embedding-augmented polygenic risk score prediction on UK Biobank data.

arXiv · Hugging Face

Project 2: Deep Generative Models: Complexity, Dimensionality, and Approximation

Statistical/ML theory
Kevin Wang, Hongqian Niu, Yixin Wang, Didong Li

1D-to-2D generator trajectory over training iterations

A theoretical study of how generative networks approximate distributions on manifolds, showing that arbitrarily low-dimensional latent inputs (below the manifold's intrinsic dimension) still suffice, at the cost of a super-exponential complexity trade-off between approximation error, dimensionality, and model size.

JMLR

Project 3: Can Generative Artificial Intelligence Survive Data Contamination? Theoretical Guarantees under Contaminated Recursive Training

Statistical/ML theory
Kevin Wang, Hongqian Niu, Didong Li

$Diffusion model trained under contaminated recursive training (CRT) on MNIST at varying real-data fractions$

A theoretical analysis of generative models trained recursively on data contaminated by earlier model-generated outputs, establishing guarantees on when and how such training remains stable versus degrades (model collapse).

arXiv