Lin Gui
I am a final-year PhD student in the Department of Statistics at the University of Chicago. I am fortunately advised by Professor Victor Veitch and Professor Jingshu Wang. Previously, I received my master's degree in statistics at the University of Chicago. Prior to UChicago, I was an undergraduate student at University of Science and Technology of China (USTC).
I am broadly interested in utilizing statistical methodologies and AI tools to solve real-world problems. In particular, I currently focus on: (1) developing post-training techniques for large generative models to enhance their task-specific capabilities and alignment with human preferences, and (2) designing reliable statistical inference methods for scientific discovery, with emphasis on multiple testing, replicability, and causal inference.
Research Interests
Machine Learning
Developing post-training techniques for large generative models
Statistics
Designing reliable statistical inference methods, including selective inference and causal inference
Publications
Aggregating Dependent Signals with Heavy-Tailed Combination Tests
Lin Gui, Yuchao Jiang, Jingshu Wang
Biometrika, 2025
We introduce the heavy-tailed combination test encompassing the widely used Cauchy combination test and harmonic mean p-value as special cases. Under asymptotic independence, we establish its validity but show that it is equivalent to the Bonferroni test in this theoretical setting. In contrast, under stronger dependence, empirical results suggest that the heavy-tailed combination test can substantially outperform Bonferroni.
BoNBoN Alignment for Large Language Models and the Sweetness of Best-of-n Sampling
Lin Gui, Cristina Gârbacea, and Victor Veitch
Advances in Neural Information Processing Systems, 2024
We prove that the underlying sampling distribution of the best-of-n sampling is essentially the optimal model (distribution) for post-training alignment of large language models. It strikes a balance between enhancing output quality, such as helpfulness and harmlessness, and preserving closeness to the base model. Based on this theoretical finding, we propose a fine-tuning method, BoNBoN, to explicitly approximate the best-of-𝑛 distribution.
Concept Algebra for (Score-Based) Text-Controlled Generative Models
Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch
Advances in Neural Information Processing Systems, 2023
We link the score representations of the text-controlled diffusion models to real-world concepts, enabling controlled concept shifts in generated images through direct manipulation of these score representations.
Causal Estimation for Text Data with (Apparent) Overlap Violations
Lin Gui, and Victor Veitch
The Eleventh International Conference on Learning Representations, 2023
We formulated a formal causal estimand tailored to the causal inference of the text-attribute question, and verified its identifiability under minimal conditions. We provided a computationally efficient estimation of the uncertainty quantification of this causal estimand, supported by theoretical assurances.
Detecting Multiple Replicating Signals Using Adaptive Filtering Procedures
Jingshu Wang, Lin Gui, Weijie J. Su, and Chiara Sabatti, Art B. Owen
The Annals of Statistics, 2022
We introduced an innovative multiple testing procedure that enhances detection power by adaptively filtering out unlikely candidates of PC nulls, and theoretically established the control of both Family-Wise Error Rate (FWER) and False Discovery Rate (FDR) for this method.
Select Preprints
Validity and Power of Heavy-Tailed Combination Tests under Asymptotic Dependence
Lin Gui, Tiantian Mao, Jingshu Wang, Ruodu Wang
We establish the asymptotic validity of the heavy-tailed combination test under a broad class of dependence structures, including both asymptotic independence and dependence. We further show that its power advantage over the Bonferroni's test is pronounced under asymptotic dependence, with the gain increasing as the level of dependence strengthens.
On the Significance of Softmax Geometry: Interpretability and Token Decoding
Yibo Jiang, Lin Gui, Sean M. Richardson, Mark Muchane, Yo Joong Choe, Victor Veitch
Recent findings show that learned embeddings of large language models do not lie in a Euclidean space, questioning the default use of cosine similarity. We adopt a more geometry-aware similarity metric and investigate it with two tasks: (1) learning interpretable features via sparse autoencoders, and (2) efficiently retrieving top-𝑘 likely next tokens.
Statistical Inference for Cell Type Deconvolution
Dongyue Xie, Lin Gui, Jingshu Wang
We study the cell type deconvolution problem, which seeks to estimate the proportions of different cell types in a mixed sample from gene expression data. We identify key conditions required for the identifiability of these proportions. To address this problem, we propose MEAD, a comprehensive statistical framework for both estimating and conducting inference on cell type proportions. MEAD further enables the comparison of deconvolved proportions across individuals, accounting for both gene–gene correlations and biological variability.