Hyungseok Song

Contact me: 7590sok@gmail.com

About Me

I am a research scientist in LG AI Research (2022 ~ present). I received Ph.D. in the School of Electrical Engineering at Korea Advanced Institute of Science and Technology (KAIST), advised by Prof. Yung Yi, in 2022. I received my B.S. (2014) and M.S. (2016) in Mathematics from KAIST, Daejeon, Republic of Korea, where my M.S. was supervised by Prof. Wanmo Kang.

My research interests lie in applying reinforcement learning for industrial optimization. Followings are my recent/past research topics with brief descriptions.

AI for Industrial Optimization – Solving real-world industrial optimization problems, including neural combinatorial optimization, design AI for PCB, and retrosynthesis for molecular synthesis.
Generative Models for Specialized Domains – Adapting generative models such as diffusion models and large language models (LLMs) to specialized industrial data.
Constrained Reinforcement Learning – Designing RL algorithms that satisfy real-world industrial constraints, including budget-constrained multi-armed bandits and objective alignment for generative and from-scratch models.
Theoretical Analysis in AI – Theoretical study of the expressive power of graph neural networks, the sample complexity of graphical models, and multi-armed bandits.

Publication

Journal

[J3] Hyungseok Song*, Deunsol Yoon*, Kanghoon Lee, Han-Seul Jeong, Soonyoung Lee, and Woohyung Lim, "CADO: Cost-Aware Diffusion Solvers for Combinatorial Optimization through RL Fine-tuning," Transactions on Machine Learning Research (TMLR), accepted, 2026. (* equal contribution; earlier version presented at SPIGM @ ICML 2024)
[J2] Hyeryung Jang, Hyungseok Song, and Yung Yi, "On Cost-efficient Learning of Data Dependency," IEEE/ACM Transactions on Networking (TON), vol. 30, no. 3, pp. 1382–1394, June 2022. (Part of this work was published in MOBIHOC 2018)
[J1] Daewoo Kim, Hyojung Lee, Hyungseok Song, Nakjung Choi, and Yung Yi, "Economics of Fog Computing: Interplay among Infrastructure and Service Providers, Users, and Edge Resource Owners" IEEE Transactions on Mobile Computing (TMC), PP(99):1-1, July 2019. (Part of this work has been published ICC 2018)

Conference

[C4] Se-eun Yoon, Hyungseok Song, Kijung Shin and Yung Yi, "How Much and When Do We Need Higher-order Information in Hypergraphs? A Case Study on Hyperedge Prediction", to appear at The Web Conference (WWW), 2020.
[C3] Hyungseok Song, Hyeryung Jang, Hai Hong Tran, Seeun Yoon, Kyunghwan Son, Donggyu Yun, Hyoju Chung, and Yung Yi, "Solving Continual Combinatorial Selection via Deep Reinforcement Learning,", to appear at International Joint Conference on Artificial Intelligence (IJCAI), 2019. (Acceptance rate: 17.8%)
[C2] Hyeryung Jang, HyungSeok Song, and Yung Yi, "Learning Data Dependency with Communication Cost," in Proceedings of ACM Mobile Ad Hoc Networking and Computing (MOBIHOC), 2018. (Acceptance rate: 16.8%) (arXiv preprint version is available at arXiv:1804.10942) [slide]
[C1] Daewoo Kim, Hyojung Lee, Hyungseok Song, Nakjung Choi, and Yung Yi, “On the Economics of Fog Computing: Inter-play among Infrastructure and Service Providers, Users, and Edge Resource Owners”, Proceedings of IEEE ICC 2018 [slide]

Workshop

[W2] Han-Seul Jeong, Youngjoon Park, Hyungseok Song, and Woohyung Lim, "ARC: Leveraging Compositional Representations for Cross-Problem Learning on VRPs," Workshop on Differentiable Learning of Combinatorial Algorithms (DiffCoALG @ NeurIPS), 2025. (Oral)
[W1] Deunsol Yoon*, Hyungseok Song*, Kanghoon Lee, and Woohyung Lim, "CADO: Cost-Aware Diffusion Solvers for Combinatorial Optimization through RL Fine-tuning," 2nd Workshop on Structured Probabilistic Inference & Generative Modeling (SPIGM @ ICML), 2024. (* equal contribution)

Thesis

[T2] Hyungseok Song, "Sequential Item Selection via Reinforcement Learning," Ph.D. Thesis in Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), 2022. (Advisor: Prof. Yung Yi)
[T1] Hyungseok Song, "Importance Sampling for multifactor portfolio credit risk in t-copula model" M.S. Thesis in Mathematics from Korea Advanced Institute of Science and Technology (KAIST), 2016. (Advisor: Prof. Wanmo Kang)

History

Career

Research scientist, Data Intelligence Lab, LG AI Research, 2022.03 ~ Present

Education

Ph.D: Electrical Engineering, KAIST, 2016.03~2022.02 (under supervision of Prof. Yung Yi)
M.S: Department of Mathematical Sciences, KAIST, 2014.03~2016.02 (under supervision of Prof. Wanmo Kang)
B.S: Department of Mathematical Sciences, KAIST, 2010.03~2014.02
Highschool: Korea Science Academy (KSA), 2007.03~2010.02

Honors and Awards

Outstanding Reviewer, ACM SIGKDD (KDD) Research Track, (February Cycle), 2025
Samsung HumanTech Thesis Prize Awards: Encouragement Prize, 2018
BK21 Plus Scholarship, Republic of Korea, 2017~2022
Government Scholarship, Republic of Korea, 2014~2022

Project Experiences

2026.01 ~ Present: Research on Agentic AI for PCB Design, LG AI Research (in collaboration with LG Electronics), Republic of Korea
- Reformulated PCB design as a routing problem in which an AI agent directly leverages EDA tool functions
2025.01 ~ 2025.12: Research on Generative AI for Molecular Synthesis in Cosmetics, LG AI Research (in collaboration with LG Household & Health Care), Republic of Korea
- Developed a retrosynthesis model based on Monte Carlo Tree Search (MCTS)
- Developed a diffusion-based reaction condition prediction model to synthesize novel molecules for cosmetic applications
2022.09 ~ 2024.12: Research on PCB Routing Algorithms, LG AI Research (in collaboration with LG Electronics), Republic of Korea
- Developed a PCB routing algorithm, built on the Pathfinder algorithm, that satisfies diverse real-world constraints on actual PCB boards.
- Improved runtime through a range of algorithmic optimizations
2022.03 ~ 2022.08: Research on Sequential Item Recommendation, LG AI Research, Republic of Korea
- Developed a transformer model that predict a user's next item selection based on their selection history
2018.09~2019.12: Research on Predicting Economic Indicator from Electronic Power Index, Korea Electric Power Corporation (KEPCO), Republic of Korea
- Analyze the pattern between the economic index and the electronic power index via machine learning method
- Design sample efficient deep learning algorithm to handle the lack of the economic indexes
2017.07~2018.08: Research on Learning-based Service Improvement Framework in Large-scale Online Request System, Naver, Republic of Korea
- Design and analyze Reinforcement Learning based service improvement framework in large-scale online request system
- Implement deep reinforcement learning algorithm for service improvement in large-scale online request system

Teaching Experiences

Teaching Assistant, Special Topics in Electronic Engineering I <My Life and Career in EE> (EE485), KAIST, Fall 2017, Spring 2018
Teaching Assistant, Calculus I (MAS101), KAIST, Spring, 2016
Teaching Assistant, Probability and Statistics (CC511), KAIST, Spring, 2016
Teaching Assistant, Probability and Statistics (MAS250), KAIST, Fall 2014, Spring 2015, Fall 2015
Teaching Assistant, Analysis for Engineers (MAS501), KAIST, Spring 2015

Reference Available on Request

Professor Yung Yi: Professor at the Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), South Korea, yiyung at kaist dot edu
Professor Wanmo Kang: Associate Professor at the Department of Mathematical Sciences, Korea Advanced Institute of Science and Technology (KAIST), South Korea, wanmo dot kang at kaist dot edu

Google Sites

Report abuse