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

My long-term research vision is to develop highly adaptive, general-purpose AI systems that continuously improve their capabilities and expand their skill sets through "active engagement" within a dynamic, ever-changing multimodal world. These systems aim to become more effective over time, responding to an unpredictable environment and developing human-like problem-solving abilities. My research includes neuro-symbolic learning, language agent reasoning/acting, reinforcement learning, and cost-efficient test-time adaptation in out-of-distribution scenarios. While my past work has drawn inspiration from neuroscience and cognitive science, these interests continue to inform my perspective on building adaptable AI systems.

My research interest include the following topics:

• Post-Training of Language Agent for Decision Making in Out-of-Distribution Scenarios
• Assimilate Your Neuron: Decision Tree-Based Optimization for Training LLM Agents in Dynamic Action Space (ongoing; as 1st-equal author)
• Retrieval-Guided Environment Reflection Enables Exploration-Exploitation Trade-offs for LLM Agents
• Scalable and Multimodal Learning for Out-of-Distribution Extrapolation
• Learning to Extrapolate Implicit Neural Representation in Parameter Manifold
• 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

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 retrieval-guided environment reflection enables exploration-exploitation trade-offs for LLM agents
keyword: LLM agent, partial observation, knowledge retrieval

A latent identifying neural functional along parameter manifold for modality-agnostic generalized INR
keyword: generalization, extrapolation, neural functional, parameter space

An adaptive routing technique conditioned with text modality for out-of-distribution forecasting
keyword: out-of-distribution forecasting

A gradient projection technique that removes the data influence in the generative model
keyword: generalization, unlearning, generative model

A graph truncated attention regularizer outperforms vanilla transformer 7%p in top-1 single-turn translation task
keyword: graph transformer

Building MIDAS system
keyword: human-controllable mice locomotive system, maze, neuroscience

Full Publications List

A retrieval-guided environment reflection enables exploration-exploitation trade-offs for LLM agents
keyword: LLM agent, partial observation, knowledge retrieval

A latent identifying neural functional along parameter manifold for modality-agnostic generalized INR
keyword: generalization, extrapolation, neural functional, parameter space

An adaptive routing technique conditioned with text modality for out-of-distribution forecasting
keyword: out-of-distribution forecasting

A fine-tuning technique with neural alignment under parameter symmetry for domain generalization
keyword: domain generalization, parameter space

A gradient projection technique that removes the data influence in the generative model
keyword: generalization, unlearning, generative model

A graph truncated attention regularizer outperforms vanilla transformer 7%p in top-1 single-turn translation task
keyword: graph transformer

Configuring brain functionals in bilinguals
keyword: linguistic, statistical parametric mapping, neuroscience

Building MIDAS system
keyword: human-controllable mice locomotive system, maze, neuroscience

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

• Korea Advanced Institute of Science and Technology(KAIST)
• visiting: Tokyo Institute of Technology(TIT), 2018
• visiting: École Polytechnique Fédérale de Lausanne(EPFL), 2017

• visiting: National University of Singapore(NUS), 2014
• visiting: University of Michigan, 2013

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

Last date of update: 2025-02-15 / template