|
Seohui Bae
배서희
I'm a research scientist at LG AI Research in South Korea.
At LG AI Research, I'm working on AI systems that reason over structure, make decisions, and improve through interaction. 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
|
|
News
- (26.05) I will attend ICML 2026 in Seoul 🇰🇷
- (26.02) 1 paper accepted to CVPR 2026. See you all in Denver 🇺🇸
|
Research
I study how language models and AI agents can acquire stronger reasoning and decision-making abilities through post-training and interaction, with reinforcement learning as a central methodology. My research focuses on training models to reason over complex decision spaces, building scalable environments and data for learning, and incorporating structural constraints into generation and decision-making.
- Agent Reasoning & Post-Training: Developing reinforcement learning, verifier-guided training, and structured supervision methods for improving reasoning and sequential decision-making. I am particularly interested in understanding which reasoning abilities are trainable and transferable across tasks.[P2, P4, C1]
- Data, Environment & Curriculum Scaling: Building executable environments, benchmarks, and data pipelines for scalable training, evaluation, and curriculum learning. I study how environments and data distributions can be designed to measure and improve model capabilities.[P3, P4, P5]
- Structured & Constrained Generation: Developing generative and learning methods that incorporate discrete structure, hard constraints, and domain knowledge, with applications to reliable reasoning, planning, and constrained generation.[P6]
I regularly contribute to academic publications and collaborative research projects.
I’m especially interested in bridging industrial challenges with generalizable solutions in: RL post-training, constrained diffusion policies, and data and environment scaling.
|
Selected Publications
(* equal contribution; † corresponding author). For the full list, see Google Scholar.
[C#] conferences; [P#] preprints/under review
-
[P6] Conditional Feasibility Steering in Diffusion Models
Seohui Bae, Han-Seul Jeong, Won-Seok Choi, Hyungseok Song, Junseok Park, Youngjoon Park, Soonyoung Lee
under review
-
[P5] RePlanBench: Benchmarking Reuse, Repair, or Replan When the World Changes
Seohui Bae, Hyungseok Song, Han-Seul Jeong, Won-Seok Choi, Junseok Park, Youngjoon Park†, Soonyoung Lee†
under review
-
[P4] CADET: Construction of PCB Routing Instances with Executable Demonstrations
Hyungseok Song, Won-Seok Choi, Seohui Bae, Han-Seul Jeong, Junseok Park, Youngjoon Park, Soonyoung Lee
under review
-
[P3] PCBWorld: A Benchmark Environment for Engine-Grounded PCB Design Automation
[pdf] [code]
Hyungseok Song*, Junseok Park*, Won-Seok Choi*, Seohui Bae, Han-Seul Jeong, Youngjoon Park†, Soonyoung Lee†
KDD Workshop on Evaluation and Trustworthiness of Agentic AI, 2026
-
[P2] Align as Act: Innovations-Based Reward Decomposition for LLM Agents
[pdf]
Sojeong Rhee*, Seohui Bae*, Jongeui Park, Whiyoung Jung, Soonyoung Lee, Woohyung Lim, Youngchul Sung
COLM Workshop on Agent Behavior, 2026
-
[C1] Align While Search: Belief-Guided Exploratory Inference for World-Grounded Embodied Agents
[pdf]
Seohui Bae, Jeonghye Kim, Youngchul Sung, Woohyung Lim
Conference on Computer Vision and Pattern Recognition (CVPR), 2026
ICML Workshop on Exploration in AI Today, 2025
|
Projects
EXAONE Lab - Current
-
Foundation Models: Large-scale foundation model training and evaluation, with a focus on agent capabilities.
Data Intelligence Lab
-
CAD Agents: LLM post-training, agent evaluation, and executable environment development for electronic design, including PCB routing.
-
Industrial RL: Reinforcement learning for sequential decision-making and optimization in manufacturing systems.
-
Forecasting: Demand forecasting and foundation-model-based forecasting, including EXAONE-Futurecast.
|
Academic Service
Conference Reviewer
- Main: ICLR, ICML, NeurIPS, AAAI, AISTATS
- Workshops: AAAI, ICLR, ICML
Journal Reviewer
|
Last date of update: 2026-05-16 / template
|
|