Sungjun Lim

Trustworthy AI,Reliable AI,Reasoning Models,Uncertainty Quantification,Robustness,Mechanistic Interpretability,Probabilistic Machine Learning

About Me

Sungjun Lim

Sungjun Lim

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 goal is to build trustworthy AI that makes reliable decisions under uncertainty. I pursue this through probabilistic learning, reliable and efficient inference, and the study of model behavior.

Research Philosophy

Three research questions toward trustworthy AI.

Ultimate Goal Trustworthy AI under uncertainty
Axis 01 / 03

Learning under Uncertainty

Learn useful distributions over models, representations, and solutions, using uncertainty to improve prediction and generalization.

Representative paper Flat Posterior Does Matter For Bayesian Model Averaging (FP-BMA) UAI 2025 Representative paper Uncertainty-driven Embedding Convolution ICLR 2026
Explore each axis

Across these questions, I care about robustness to changing conditions and the efficient use of data and computation.

News

Category
Year

2026

Publications
Geometry-Adaptive Explainer accepted to NeurIPS 2026

Our work on faithful dictionary-based interpretability under distribution shift was accepted to NeurIPS 2026.

arXiv Code Project page Publications
Career
Started an internship at RIKEN AIP

Joined the Adaptive Bayesian Intelligence Team at RIKEN AIP as an intern.

Career Team
Etc.
Joined the organizing committee for the Mechanistic Interpretability Workshop at ICML 2026

The workshop will be held at ICML 2026 in Seoul, South Korea.

Workshop
Publications
Uncertainty-driven Embedding Convolution accepted to ICLR 2026

Our work on uncertainty-aware embedding ensembles was accepted to the ICLR 2026.

arXiv Project page Publications
Publications
Semi-Supervised Preference Optimization with Limited Feedback accepted to ICLR 2026

The paper was accepted to ICLR 2026 as an oral presentation.

arXiv Publications

2025

Career
Visited Australian National University as a visiting researcher

Worked with Lexing Xie in Computer Science at ANU.

Career Advisor
Publications
Flat Posterior Does Matter For Bayesian Model Averaging accepted to UAI 2025

Our paper on flat posterior behavior in Bayesian model averaging was accepted to UAI 2025.

arXiv Publications

No news matches these filters.

Career

Earlier Scroll to explore Latest

Mar 2017 – Aug 2022

B.S. in Statistics

University of Seoul

Jun 2021 – Aug 2022

Undergraduate Research Assistant

MLAI Lab, University of Seoul

Sep 2022 – Feb 2024

M.S. in Artificial Intelligence

University of Seoul

Advisor: Kyungwoo Song

Mar 2024 – Present

Ph.D. Student

Current

Yonsei University

Statistics and Data Science

Advisor: Kyungwoo Song

Jul 2025 – Aug 2025

Visiting Researcher

Australian National University

Computer Science

Advisor: Lexing Xie

Publications

All publicationsGoogle Scholar

Year
Type
Category
  1. Language model-guided student performance prediction with multimodal auxiliary information
    Applications
    • Changdae Oh, Minhoi Park, Sungjun Lim, Kyungwoo Song
    • Expert Systems with Applications (ESWA) 2024
  2. GFML: Gravity Function for Metric Learning
    Robustness
    • Hoyoon Byun, Sungjun Lim, Kyungwoo Song
    • Engineering Applications of Artificial Intelligence (EAAI) 2025
  3. Robust Optimization for PPG-based Blood Pressure Estimation
    Robustness Applications
    • Sungjun Lim, Taero Kim, Hyeonjeong Lee, Yewon Kim, Minhoi Park, Kwang-Yong Kim, Minseong Kim, Kyu Hyung Kim, Jiyoung Jung, Kyungwoo Song
    • Biomedical Signal Processing and Control (BSPC) 2025
  4. Brain-inspired Lp-Convolution benefits large kernels and aligns better with visual cortex
    • Jae Kwon, Sungjun Lim, Kyungwoo Song, C. Justin Lee
    • International Conference on Learning Representations (ICLR) 2025
  5. Sufficient Invariant Learning for Distribution Shift
    Robustness
    • Taero Kim, Subeen Park, Sungjun Lim, Yonghan Jung, Krikamol Muandet, Kyungwoo Song
    • Computer Vision and Pattern Recognition (CVPR) 2025
  6. On a synthetic posterior with sharp and flat modes, FP-BMA samples the flat mode while VI, MCMC, and SWAG concentrate on the sharp mode.
    UAI 2025
    Flat Posterior Does Matter For Bayesian Model Averaging

    We train Bayesian neural networks to favor flat posteriors, improving generalization in model averaging and transfer learning.

    Sungjun Lim, Jeyoon Yeom, Sooyon Kim, Hoyoon Byun, Jinho Kang, Yohan Jung, Jiyoung Jung, Kyungwoo Song

    Uncertainty Robustness
    Paper Code
  7. Uncertainty Aware Contrastive Decoding
    Uncertainty
    • Hakyung Lee, Subeen Park, Joowang Kim, Sungjun Lim, Kyungwoo Song
    • Association for Computational Linguistics (ACL) 2025 Findings
  8. COVID-19 Prediction with Doubly Multi-task Gaussian Process
    Uncertainty Applications
    • Sooyon Kim, Yongtaek Lim, Sungjun Lim, Gyeongdeok Seo, Jihee Kim, Hojun Park, Jeahun Jung, Kyungwoo Song
    • Journal of Biomedical Informatics 2025
  9. Causal Effect Variational Transformer for Public Health Measures and COVID-19 Infection Cluster Analysis
    Uncertainty Applications
    • Jinho Kang, Sungjun Lim, Kyungwoo Song
    • Conference on Information and Knowledge Management (CIKM) 2025
  10. Data Adaptive Stochastic Ensemble Net: Optimizing Infection Predictions for COVID-19 Cluster Analysis
    Uncertainty Applications
    • Sungjun Lim, Yongtaek Lim, Hojun Park, Junggu Lee, Jaehun Jung, Kyungwoo Song
    • IEEE Journal of Biomedical and Health Informatics 2025
  11. RAILL : Retrieval-Augment and Instruction Tuning for Low-resource Language Model Training
    Efficiency
    • Youngjun Choi, Sungjun Lim, Minhoi Park, Jaekyeong Jung, TaeKyung Kim, Hosik Choi, Kyungwoo Song
    • IEEE Big data 2025 (Short)
  12. SSPO uses a reward threshold to pseudo-label unpaired responses, then learns from both paired preferences and pseudo-labeled responses.
    ICLR 2026 Oral
    Semi-Supervised Preference Optimization with Limited Feedback

    We align language models with fewer preference labels by learning from paired feedback and pseudo-labeled unpaired responses.

    Seonggyun Lee, Sungjun Lim, Seojin Park, Soeun Cheon, Kyungwoo Song

    Alignment & Safety
    Data efficiency
    Paper Code
  13. UEC converts deterministic embeddings into probabilistic embeddings, combines them using uncertainty, and computes uncertainty-aware similarity.
    ICLR 2026
    Uncertainty-driven Embedding Convolution

    We combine embedding models using their uncertainty to improve retrieval and classification across tasks and languages.

    Sungjun Lim, Kangjun Noh, Youngjun Choi, Heeyoung Lee, Kyungwoo Song

    Uncertainty
    Paper Code Project page
  14. Eigen-Value : Efficient Domain-Robust Data Valuation via eigenvalue-Based Approach
    Robustness
    Efficiency
    • Youngjun Choi, Junseong Kang, Sungjun Lim, Kyungwoo Song
    • Computer Vision and Pattern Recognition (CVPR) 2026 Findings
  15. Distribution shift misaligns an explainer dictionary with the model's active subspace; GAE realigns the dictionary to restore faithfulness.
    NeurIPS 2026
    Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift

    We realign explainer dictionaries to shifted activation geometry, restoring faithful explanations without gradient updates.

    Sungjun Lim, Heedong Kim, Andrew Lee, Kyungwoo Song

    Interpretability Robustness
    Paper Code Project page

No peer-reviewed publications match the selected filters.

  1. DDRL: A Diffusion-Driven Reinforcement Learning Approach for Enhanced TSP Solutions
    Applications
    Reasoning
    • Joowang Kim, Jeyoon Yeom, Gyeongdeok Seo, Sungjun Lim, Jae Ha Kwak, Heejun Ahn, Gyeong-moon Park, Kyungwoo Song
  2. When to Stop and Which to Return: Trajectory-Level Minimum Bayes Risk Decoding for Diffusion Language Models
    Efficiency
    • Subeen Park, Sungjun Lim, Heejin Jung, Somin Kim, Gyeong-Moon Park, Kyungwoo Song
  3. RRD: Routing-and-Residual Distillation for Recovery after Dense-to-MoE Conversion
    Efficiency
    • Hoyoon Byun, Kangjun Noh, Somin Kim, Heedong Kim, Jaeyoon Shim, Sungjun Lim, Youngjun Choi, Kyungwoo Song
  4. Train for Many, Update with One: Improving Multi-Sample Reasoning via Fenchel-Bregman Posterior Learning
    Uncertainty
    Reasoning Efficiency
    • Sungjun Lim, Adrian Robert Minut, Nico Daheim, Mohammad Emtiyaz Khan, Kyungwoo Song, Thomas Möllenhoff
  5. SPHERE: Signal-Aware Prototype Harmonization for Continual Specific Emitter Identification
    Applications
    • Sungjun Lim, Sumin Park, Gyungmin Kim, Kyuha Song, Kyungwoo Song

  1. A Double-Edged Sword: Benchmarking the Trade-off Between Bias Mitigation and Helpfulness of LLM Guardrails in Finance
    Alignment & Safety Applications
    Evaluation
    • Sungjun Lim, Hoyoon Byun, Jihee Kim, Kyungwoo Song
    • Advances in Financial AI: Innovations, Risk, and Responsibility in the Era of LLMs at CIKM 2025
  2. Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift
    Interpretability Robustness
    • Sungjun Lim, Heedong Kim, Andrew Lee, Kyungwoo Song
    • Mechanistic Interpretability Workshop at ICML 2026