Events

CDS Colloquium: Virtues and Pitfalls of Weak-to-Strong Generalization: From Intrinsic Dimensions to Spurious Correlations

Speaker: Qi Lei

Location: 60 Fifth Avenue, Room 150

Date: Friday, October 3, 2025

Weak-to-strong (W2S) generalization is an intriguing paradigm where a strong, pre-trained student model adapts to a downstream task using pseudo-labels generated by a weaker teacher. Despite the apparent limitations of the weak teacher, W2S fine-tuning often leads the student to outperform the teacher itself. This talk will present two recent theoretical perspectives on why and when W2S succeeds.

First, I will discuss the phenomenon through the lens of low intrinsic dimension and in a variance-dominant regime where fine-tuning often takes place in sufficiently expressive low-dimensional subspaces. This analysis reveals a surprising virtue of discrepancy between strong and weak models' feature representation: while variance is inherited in overlapping subspaces, it is dramatically reduced in subspaces of discrepancy, with explicitly derived characterizations of sample complexity and scaling behavior.

Second, I will examine W2S under spurious correlations, a common challenge when labeled data shaping the teacher and unlabeled data guiding the student differ in group proportions. High-dimensional asymptotic analysis reveals that alignment between group distributions is critical: under group-balanced teachers, minority enrichment improves W2S, while under imbalanced teachers, it harms performance. To address this, a simple confidence-based retraining scheme with generalized cross-entropy can mitigate the pitfalls and consistently strengthen W2S across synthetic and real-world datasets.
Together, these works explain why W2S emerges-via intrinsic dimension and representation discrepancy-and how it is affected by spurious correlations, providing a sharper theoretical foundation and guidance for its future development.



Bio:

Qi Lei is an assistant professor of Mathematics and Data Science at the Courant Institute of Mathematical Sciences and the Center for Data Science at NYU. Previously she was an associate research scholar or visiting student at the ECE department of Princeton University/IAS. She received her Ph.D. from Oden Institute for Computational Engineering & Sciences at UT Austin. Before that, she was a research fellow at Simons Institute for the Foundations of Deep Learning Program. Her research aims to bridge the theoretical and empirical boundary of modern machine learning algorithms and in particular Al safety, with a focus on data privacy, distributionally robust algorithms, sample- and parameter-efficient learning. Qi has received several awards, including the Outstanding Dissertation Award, Computing Innovative Fellowship, and Rising Stars in EECS, in Machine Learning, and in Statistics and Data Science.