I’m a Ph.D. candidate at Stanford CS, advised by Chelsea Finn and part of the IRIS lab. I am affiliated with SAIL, CRFM, and the ML Group at Stanford. My research is supported by an OpenAI Superalignment Fellowship and a KFAS PhD Scholarship.
Previously, as mandatory military service for the South Korean army, I was a research scientist at Kakao and AITRICS, working with Juho Lee. Before that, I completed my master’s (CS, advised by Seungjin Choi) and undergraduate (math) degrees at POSTECH.
Here are some key questions that guide my research:
- Teaching strong models: Strong pre-trained models already know much of what we want to teach them. Post-training seems to be more about eliciting the appropriate pre-existing capabilities than instilling entirely new information. Can we develop a more effective paradigm for “teaching” models that leverage the pre-existing capabilities inside pre-trained models?
- Underspecification: No dataset fully specifies its intended task. How can we make models recognize and represent the multitude of possible realities consistent with given data? What is the best way to leverage such diverse hypotheses?
- Understanding information: Within any data, there is an underlying essence (“information”) that exists independently of the specific representation. How can we better conceptualize this notion of information, and understand the mechanisms by which machine learning models extract, store, and communicate it?
- Mitigating risks: What strategies can we employ to handle the reality of machine learning systems generating potentially erroneous or harmful outputs?
Selected Papers
Clarify: Improving Model Robustness with Natural Language Corrections
UIST 2024, NeurIPS 2023 workshops XAIA and ICBINB
AutoFT: Learning an Objective for Robust Fine-Tuning
NeurIPS 2023 workshop DistShift
Project and Probe: Sample-Efficient Domain Adaptation by Interpolating Orthogonal Features
ICLR 2024 (spotlight)
DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature
ICML 2023 (long oral)