author
Postdoctoral Scholar at Princeton University
,
Princeton University
Postdoctoral Scholar
I am a final-year PhD student in Computer Science at University of Wisconsin-Madison, where I am fortunate to be advised by Frederic Sala. Before that, I was a master’s student at Seoul National University, where I was lucky to learn deep learning, exploratory data analysis, and information theory from Wonjong Rhee. Prior to that, I received B.A in Psychology and B.S. in Computer Science and Engineering from Seoul National University.
My research focuses on data-centric AI, particularly programmatic weak supervision and weak-to-strong generalization in foundation models. These approaches use weaker models as supervision sources to train stronger models, providing labels, reward signals, and verification signals that guide more capable systems. I have also explored inference-time steering, which involves intervening on internal representations to improve robustness, alignment, and personalization of foundation models at inference time without fine-tuning.
Looking ahead, my vision is to develop strategies for supervising superhuman-level intelligence, where traditional human oversight is no longer sufficient. My research currently focuses on two directions. The first is weak-to-strong generalization, where weaker models are used as supervision sources to train stronger ones, providing labels, reward signals, and verification for more capable systems. The second is out-of-distribution (OOD) generalization, including challenges such as easy-to-hard generalization, length generalization, and compositional generalization.
The latest from Ph.D. Student
Weak-to-Strong Generalization Through the Data-Centric Lens
The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. We propose a simple data-centric mechanism that characterizes weak-to-strong generalization: the overlap density. Intuitively, generalization tracks the number of points that contain overlaps, i.e., both easy patterns (learnable by a weak model) and challenging patterns (only learnable by a stronger model), as with such points, weak predictions can be used to learn challenging patterns...
Research Paper
Weak-to-Strong Generalization Through the Data-Centric Lens
The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. We propose a simple data-centric mechanism that characterizes weak-to-strong generalization: the overlap density. Intuitively,…
Blog
Weak supervision for non-categorical applications + superalignment
We need more labeled data than ever, so we have explored weak supervision for non-categorical applications—with notable results.
Zero-Shot Robustification of Zero-Shot Models with Foundation Models
Zero-shot inference is a powerful paradigm that enables the use of large pretrained models for downstream classification tasks without further training. However, these models are vulnerable to inherited biases that can impact their performance. The traditional solution is fine-tuning, but this undermines the key advantage of pretrained models, which is their ability to be used out-of-the-box. We propose ROBOSHOT, a method that improves the robustness of pretrained model embeddings in a fully zero-shot fashion. First, we use zero-shot language models (LMs) to obtain useful insights from task descriptions. These insights are embedded and used to remove harmful and boost useful...
Research Paper
Zero-Shot Robustification of Zero-Shot Models with Foundation Models
Zero-shot inference is a powerful paradigm that enables the use of large pretrained models for downstream classification tasks without further training. However, these models are vulnerable to inherited biases that can impact their performance. The traditional solution is fine-tuning, but this undermines the key advantage of pretrained models, which is their ability to be used out-of-the-box. We propose ROBOSHOT, a…
