Self-Learning Principles of Large-Scale Foundation Models

NEW LINK: Zoom link for remote attendees

Meeting ID: 962 5398 2796​ Passcode: 123123

Abstract: Large-scale foundation models (e.g., GPT) have achieved unprecedented capabilities by leveraging two core learning paradigms: massive self-supervised learning (SSL) (e.g., next word prediction) and flexible test-time learning (TTL) (e.g., reasoning). Both paradigms reflect a model’s ability to learn without explicit human supervision—a capability we term self-learning—which enables models to scale their performance with increasing training-time and test-time compute. However, these self-learning mechanisms remain poorly understood, leading to significant empirical trial and error in practice.

In this talk, I will outline principled approaches to understanding and designing self-learning mechanisms in foundation models. First, I introduce a unified graph-theoretic framework that characterizes the generalization of both discriminative and generative SSL models. This framework explains how seemingly disparate SSL approaches converge on meaningful representations without labels and derives practical strategies to enhance model efficiency, robustness, and interpretability. Second, I investigate how test-time learning operates in language models—particularly in long, self-reflective reasoning processes such as o1 and R1—and present both theoretical insights and scalable designs to enhance model capabilities and safety at test time. By integrating rigorous theoretical understanding with practical algorithmic advancements, this research charts a path toward designing more capable, reliable, and trustworthy foundation models through principled self-learning mechanisms.

Bio: Yifei Wang is a postdoctoral associate at MIT CSAIL, where he works with Professor Stefanie Jegelka. His research focuses on the theoretical and algorithmic foundations of self-learning, foundation models, and AI safety. His research in these areas has earned him three best paper awards, including the sole Best ML Paper Award at ECML-PKDD 2021, the Silver Best Paper Award at the ICML 2021 AML workshop, and the Best Paper Award at the ICML 2024 ICL Workshop. His work was featured by Anthropic and MIT for its contributions to self-learning and AI safety mechanisms, and was recognized by multiple national and president scholarships. Before joining MIT, he earned a PhD in Applied Mathematics, a BS in Data Science, and a BA in Philosophy from Peking University.