I build AI agents that play games — currently fine-tuning vision-language-action models and running large-scale distributed RL on PUBG at Krafton's Gameplay AI Team.
Before Krafton, I built LLM agent frameworks at Samsung Research. PhD in EE from KAIST, advised by Prof. Youngchul Sung & Prof. Sae-Young Chung.
My research interests span AI agents (agent harnessing, automated research, planning), reinforcement learning (meta-RL, generalization, sample efficiency, offline RL), and vision-language-action models for gameplay and screen-control agents.
Large-scale distributed reinforcement learning and vision-language-action model training on PUBG, at the Gameplay AI Team.
Released April 2026 — an LLM-powered system that turns natural-language behavior descriptions into trained RL policies.
Agentic engineering for reinforcement learning. An LLM-powered pipeline that writes reward functions, runs parallel PPO training across seeds & configs, evaluates with a dual code/VLM judge, and auto-revises on failure — across MuJoCo and IsaacLab environments.
QCS adaptively combines return-conditioned supervised learning with a Q-function — leaning on RCSL for optimal trajectories and on Q-aid for sub-optimal ones — to recover trajectory-stitching ability without sacrificing stability.
A new MetaFormer-based decision-maker for offline RL using three convolution filters; excels at understanding local associations and generalizes better than prior decision transformers.
Enhances meta-RL generalization on non-parametrically varying tasks by decomposing them into elementary subtasks and conducting virtual training.
An intrinsic-reward coefficient adaptation scheme equipped with intrinsic-motivation awareness, adjusting the coefficient online to maximize extrinsic return.
Trains an RL agent on imaginary tasks generated from mixtures of learned latent dynamics, improving generalization to unseen test tasks.
A computationally efficient recursive deep RL algorithm that propagates sparse and delayed rewards directly through all transitions of a sampled episode.
Meta-Reinforcement Learning with Imaginary Tasks · KAIST EE
Paper competition for graduate students · Qualcomm
Excellence in 2017 entrance examination · KAIST EE
I view life as a meta-reinforcement learning task — reminiscent of MuJoCo's Ant-direction, where an agent learns to walk toward a hidden goal direction it can never observe directly.
Everyone has their own optimal life direction T: unique, often obscured. The objective is to maximize the cumulative reward
r = M · T,the dot product of M — how we choose to live — and the unseen true direction T. Two things matter: the angle θ between them, and the magnitude |M|.
I was fortunate to have guidance from two advisors who instilled in me the importance of both — minimizing |θ| by choosing the right direction, and maximizing |M| by moving with conviction.
The same agent harnessing I do at work, I do for fun on weekends — pair-programming with LLMs to ship small games end-to-end.
This one has a back-story. I picked up Go as a kid and stuck with it long enough to get reasonably serious. During my PhD I worked on Bayes-adaptive policy optimization — agents that infer hidden task structure from observation. Stonecode is what falls out when those two things meet.
A 1v1 game played on a Go board, but the rule space is enormous: each match has a different hidden scoring rule, and players must adapt and infer it Bayes-adaptively from the score feedback after every move. Built end-to-end with an agentic dev workflow — feature branches, parallel sub-agents for client / server / shared layers, and an automated PR-review loop.
