Hi, I’m Kihyun! I’m a PhD Candidate in EECS at MIT.
Click or tap the objects below to explore my research and learn more about me.

A modern white pixel-art research lab with a bookshelf, globe, CV notebook, news magazine, robot, and panoramic windows overlooking the Charles River and Boston skyline.

News

Research Projects

Reward Learning

Visualization for Inverse Reinforcement Learning without an Optimal Demonstrator: A Feasible Reward Set Approach
Under review · 2026

Inverse Reinforcement Learning without an Optimal Demonstrator: A Feasible Reward Set Approach

Kihyun Kim, Shripad Deshmukh, Nikos Vlassis, Jiawei Zhang

We study how to infer rewards from multiple imperfect demonstrators with different levels of suboptimality. Our framework translates their suboptimality into linear constraints and combines them into a feasible reward set, with guarantees that can hold without a near-optimal demonstrator under sufficient coverage. An offline algorithm extends the approach to function approximation, with experiments in grid worlds and LLM fine-tuning.

Paper
Visualization for A Unified Linear Programming Framework for Offline Reward Learning from Human Demonstrations and Feedback
ICML 2024

A Unified Linear Programming Framework for Offline Reward Learning from Human Demonstrations and Feedback

Kihyun Kim, Jiawei Zhang, Asuman Ozdaglar, Pablo A. Parrilo

We formulate offline reward learning from demonstrations and pairwise feedback within a unified linear programming framework. Using primal-dual optimality conditions, we estimate a feasible reward set from precollected trajectories and establish optimality and sample-efficiency guarantees. The approach incorporates preference feedback without conventional probabilistic preference models, with examples demonstrating advantages over maximum-likelihood estimation.

Paper

AI Alignment

Visualization for Beyond RLHF and NLHF: Population-Proportional Alignment under an Axiomatic Framework
ICLR 2026

Beyond RLHF and NLHF: Population-Proportional Alignment under an Axiomatic Framework

Kihyun Kim, Jiawei Zhang, Asuman E. Ozdaglar, Pablo A. Parrilo

We address how AI should aggregate diverse human preferences without disproportionately favoring particular groups. Our framework infers feasible population distributions from pairwise comparisons and constructs policies satisfying axioms for proportional representation, Pareto efficiency, and bounded manipulation. A softmax relaxation offers a tunable trade-off between population-proportional alignment and Condorcet-winner selection, with evaluations in recommendation and LLM alignment.

Paper

RL / Control

Visualization for Distributional robustness in minimax linear quadratic control with Wasserstein distance
SIAM Journal on Control and Optimization · 2023

Distributional robustness in minimax linear quadratic control with Wasserstein distance

Kihyun Kim, Insoon Yang

We design linear-quadratic controllers that remain robust when the disturbance distribution estimated from data is inaccurate. A Wasserstein penalty controls how far an adversary can shift this distribution, yielding policies computed through Riccati equations for finite and infinite horizons. We establish closed-loop stability and out-of-sample performance guarantees, and connect the method to classical H-infinity control.

Paper
Visualization for Optimizing large-scale fleet management on a road network using multi-agent deep reinforcement learning with graph neural network
IEEE ITSC 2021

Optimizing large-scale fleet management on a road network using multi-agent deep reinforcement learning with graph neural network

Juhyeon Kim, Kihyun Kim

We model ride-hailing fleet management directly on a road-network graph and learn vehicle allocation policies using multi-agent reinforcement learning. Graph neural networks approximate action values, while stochastic policy updates improve on greedy updates. Experiments in a simulator based on empirical taxi demand demonstrate the approach across different operating conditions.

Paper
Visualization for Minimax control of ambiguous linear stochastic systems using the Wasserstein metric
IEEE CDC 2020

Minimax control of ambiguous linear stochastic systems using the Wasserstein metric

Kihyun Kim, Insoon Yang

We propose a minimax controller for linear stochastic systems whose disturbance distributions are only approximately known. An adversary selects a worst-case distribution subject to a Wasserstein penalty, allowing the controller to adjust its conservativeness. We derive finite-horizon policies through a Riccati equation and give conditions for an infinite-horizon solution that stabilizes the expected system state under the worst-case distribution.

Paper
Visualization for Generative Autoregressive Networks for 3D Dancing Move Synthesis from Music
IEEE Robotics and Automation Letters · 2020

Generative Autoregressive Networks for 3D Dancing Move Synthesis from Music

Hyemin Ahn, Jaehun Kim, Kihyun Kim, Songhwai Oh

We generate sequences of 3D dance poses directly from music by combining a music encoder, an autoregressive pose generator, and a genre classifier. The model synthesizes dance sequences exceeding 5,000 frames, with experiments across different songs demonstrating human-like motion. We also transfer a generated sequence to a humanoid robot, showing how music can drive robotic dance.

Paper