Seohui Bae

I'm a research scientist at LG AI Research in South Korea.

At LG AI Research, I've worked on reasoning, out-of-distribution extrapolation, and neural functionals. I completed my bachelor's and master's studies at KAIST, where I was fortunate to be advised by Prof. Eunho Yang.

Email  /  Google Scholar  /  LinkedIn

Research

I work at the intersection of reasoning, adaptation, and learning under distribution shifts. My research interest include the following topics:

• Post-Training and Inference-Time Adaptation of Language Agents for Out-of-Distribution Decision Making
• Align While Search: Belief-Guided Exploratory Inference for Test-Time World Alignment
• Scalable and Multimodal Learning for Out-of-Distribution Extrapolation
• Adaptive Information Routing for Multimodal Time Series Forecasting
• Science of Learning (e.g. parameter space geometry, loss landscape, model bias)
• ShERPA: Shifting basin for Enhanced Robustness via Permuted Activations
• Gradient Surgery for One-shot Unlearning on Generative Model
• GTA: Graph Truncated Attention for Retrosynthesis

I regularly contribute to academic publications and collaborative research projects. I’m especially interested in bridging industrial challenges with generalizable solutions in: inference-time scaling, long-tail generalization, and extrapolation

Interests

• decision making, reasoning/
  acting, reinforcement learning
• out-of-distribution, test-time adaptation
• science of learning

Education

• M.S. in Artificial Intelligence, Mar 2020-Feb 2022
  Korea Advanced Institute of Science and Technology (KAIST)
• B.S. in Biological Science, Computer Science (minor), Mar 2015-Feb 2020
  Korea Advanced Institute of Science and Technology (KAIST)
• Korea Science of Academy of KAIST, Mar 2012-Feb 2015

Selected Publications

A training-free LLM agent that aligns with hidden world states via language-driven Bayesian inference.
keyword: epistemic exploration, language model agent, test-time adaptation

Projects

LG AI Research

• EXAONE-Futurecast
• Demand Forecasting

Ongoing Research

We are currently finalizing two projects expected to be submitted in the coming months:

• Extrapolation via Symbolic Time-Series Logic: combining rule-based temporal logic with dynamic forecasting.
• Decision Tree-Based Model Adaptation: using symbolic structure to guide low-cost adaptation of pretrained agents.

Education

M.S. in Graduate School of Artificial Intelligence, Mar 2020–Feb 2022

• Korea Advanced Institute of Science and Technology (KAIST)
• Machine Learning and Intelligence Lab
• Advisor: Eunho Yang

B.S. in Biological Science, Computer Science (minor), Mar 2015–Feb 2020

• Korea Advanced Institute of Science and Technology (KAIST)

Korea Science Academy of KAIST, Mar 2012–Feb 2015

Academic Service

Conference / Journal Reviewer

• Conferences: ICLR 2025, NeurIPS 2024, ICLR 2024
• Workshops / Shorts: ICLR 2024, ICML 2023, AAAI 2023
• Journals: ACM Computing Surveys 2024

Last date of update: 2025-06-12 / template