Geonwoo Cho (趙乾佑)

Hi! I am an undergraduate student at GIST, where I work with Prof. Sundong Kim at GIST, Prof. Lexin Li at UC Berkeley, and Prof. Yuhua Zhu at UCLA.

I am interested in developing foundational models for decision-making.

My current research focuses on three key directions:
(1) Scalable reinforcement learning, to build agents that scale with data and compute across large task families.
(2) Unsupervised Reinforcement learning, to pretrain generalist agents via intrinsic-motivation-driven exploration that discovers diverse, reusable skills.
(3) Open-ended curricula, to continuously expose agents to novel yet learnable tasks.

Previously, I worked as a machine learning software engineer at Match Group/Hyperconnect LLC and several other companies, where I gained hands-on experience deploying large-scale machine learning systems in production.

Feel free to reach out if you'd like to have a chat!
Contact: gwcho.public AT gmail.com

Google Scholar /  Twitter /  Github /  CV (Oct. 2025)

profile photo

Annealing Bridges Offline and Online RL
Geonwoo Cho, Jaegyun Im, Doyoon Kim, Lexin Li
Preprint
Paper

SOAR is an offline-to-online RL framework that jointly stabilizes early adaptation and preserves long-run performance via dual annealing of offline reliance and conservatism, explicitly countering spurious Q-optimism that misranks inferior actions during early finetuning.

Causal-Paced Deep Reinforcement Learning
Geonwoo Cho, Jaegyun Im, Doyoon Kim, Sundong Kim
RLC Workshop on The Causal Reinforcement Learning, 2025
Paper / Code

CP-DRL is a causally-aware curriculum that infers SCM differences from interaction data and sequences tasks according to the resulting causal structural variations across environments.

TRACED: Transition-aware Regret Approximation with Co-learnability for Environment Design
Geonwoo Cho, Jaegyun Im, Jihwan Lee, Hojun Yi, Sejin Kim, Sundong Kim
CoRL Workshop on Resource-Rational Robot Learning, 2025
Paper / Website / Code

TRACED is a regret-based curriculum that augments value loss with transition-prediction loss and a lightweight co-learnability metric.

AMPED: Adaptive Multi-objective Projection for balancing Exploration and skill Diversification
Geonwoo Cho*, Jaemoon Lee*, Jaegyun Im, Subi Lee, Jihwan Lee, Sundong Kim
CoRL Workshop on Resource-Rational Robot Learning, 2025
Paper / Website / Code

AMPED is a framework for skill-based reinforcement learning that simultaneously maximizes state coverage and skill diversity through several carefully designed components.


UCLA, Statistics and Data Science (June 2025 – Present)
Continuous-time Reinforcement Learning
Advised by Prof. Yuhua Zhu
Berkeley, Biostatistics (Jan 2025 – Present)
Offline-to-Online Reinforcement Learning
Advised by Prof. Lexin Li
GIST, Data Science Lab (Apr 2024 – Present)
Unsupervised Environment Design, Skill-based Reinforcement Learning
Advised by Prof. Sundong Kim
GIST, AITER Lab (Jun 2020 – Dec 2020)
Time-series Prediction
Advised by Prof. Hongkook Kim

Team Learners (Aug 2023 – Jan 2024)
Machine Learning Software Engineer
Match Group/Hyperconnect LLC (Jun 2022 – Jul 2023)
Machine Learning Software Engineer
Worked as part of the mandatory military service in the Republic of Korea
Business Canvas (Dec 2021 – Jun 2022)
Software Engineer
Worked as part of the mandatory military service in the Republic of Korea
Algorima (Dec 2020 – Jun 2021)
Software Engineer
Worked as part of the mandatory military service in the Republic of Korea

Gwangju Institute of Science and Technology (Feb 2019 – Present)
B.S. in Electrical Engineering and Computer Science, Minor in Mathematics
Leave of absence for military service: Sep 2021 - Sep 2023 (2 years)
University of California, Berkeley (Jan 2025 – Aug 2025)
Exchange Student
Korea Science Academy of KAIST (Mar 2016 – Feb 2019)
University of Wisconsin–Madison (Jun 2018 – Jul 2018)
Exchange Student


Website templates from here and here.