Ph.D. Student, Yonsei University
Statistics and Data Science
Hello! I am Sungjun Lim, a PhD student at Yonsei University's Statistics and Data Science department.
My ultimate goal is to build AI systems that remain trustworthy under uncertainty: models that can recognize what they do not know, reason over alternatives, adapt beyond training conditions, and make their decisions understandable to humans.
Five research axes toward one goal.
Use uncertainty not only to measure confidence, but to guide prediction, selection, and exploration.
Representative paper Uncertainty-driven Embedding ConvolutionBuild models that remain useful when distributions shift, environments change, or evidence is incomplete.
Representative paper Flat Posterior Does Matter For Bayesian Model Averaging (FP-BMA)Move beyond a single deterministic path by exploring diverse plausible solutions before deciding.
Develop explanations that reflect internal model behavior rather than post-hoc stories.
Representative paper Geometry-Adaptive Explainer (GAE)Let models decide when to spend more computation and when the available evidence is enough.
Representative paper Semi-Supervised Preference Optimization with Limited Feedback (SSPO)The workshop will be held at ICML 2026 in Seoul, South Korea.
Our work on uncertainty-aware embedding ensembles was accepted to the ICLR 2026.
The paper was accepted to ICLR 2026 as an oral presentation.
Our paper on flat posterior behavior in Bayesian model averaging was accepted to UAI 2025.
No news in this category yet.
University of Seoul
MLAI Lab, University of Seoul
University of Seoul
Advisor: Kyungwoo Song
Yonsei University
Statistics and Data Science
Advisor: Kyungwoo Song
All publicationsGoogle Scholar
We train Bayesian neural networks to favor flat posteriors, improving generalization in model averaging and transfer learning.
We align language models with fewer preference labels by learning from paired feedback and pseudo-labeled unpaired responses.
We combine embedding models using their uncertainty to improve retrieval and classification across tasks and languages.
No peer-reviewed publications match the selected filters.
We realign explainer dictionaries to shifted activation geometry, restoring faithful explanations without gradient updates.
