Soowon Oh

About

Research Profile

I am currently a Ph.D. student focusing on artificial intelligence research, with particular interests in efficient large language models (LLMs), reinforcement learning (RL), and machine learning systems.

My background spans system software development, including firmware, kernel drivers, simulators, and compilers. This experience gives me a deep understanding of the vertical software stack from low-level systems to high-level applications, and now shapes how I study the efficiency and scalability of AI models.

My current research interests include efficient AI/ML deployment under constrained resources and the foundations of genuine intelligence beyond data-driven probabilistic prediction. I am especially interested in RL as a promising path toward understanding what is needed to build more adaptive, robust, and intelligent systems.

I am motivated to bridge systems and AI, leveraging my prior expertise to design efficient learning frameworks for the next generation of intelligent systems.

Updates

News

  • 🎉 BASTION was accepted for an oral presentation at the ICML 2026 AdaptFM Workshop.

  • 🎉 Started my Ph.D. program at KAIST Graduate School of AI.

Academic Path

Education

KAIST, Graduate School of AI

Ph.D. Student in Artificial Intelligence

Research interests include efficient LLMs, reinforcement learning, and machine learning systems.

Sep 2025 — Present

Seoul National University

Master of Science in Computer Science and Engineering

Design and Implementation of the A2 Operating System on the Intel Single-Chip Cloud Computer (SCC).

2014 — 2016

Yonsei University

Bachelor of Science in Electrical & Electronic Engineering, Minor in Computer Science

2008 — 2014

Research Output

Publications

  • 2026

    Bastion: Budget-Aware Speculative Decoding with Tree-structured Block Diffusion Drafting

    Soowon Oh*, Nam Cao*, Yujin Kim, Hojung Jung, Huzama Ahmad, Sangmin Bae, Se-Young Yun

    ICML 2026 Workshop on Resource-Adaptive Foundation Model Inference (AdaptFM). Oral Presentation.

    arXiv Code

  • 2026

    PerMix-RLVR: Preserving Persona Expressivity under Verifiable-Reward Alignment

    Jihwan Oh, Soowon Oh, Murad Aghazada, Minchan Jeong, Sungnyun Kim, Se-Young Yun

    Preprint, 2026.

    arXiv

  • 2026

    mSFT: Addressing Dataset Mixtures Overfitting Heterogeneously in Multi-task SFT

    Woosung Koh, Jeyoung Jeon, Youngjin Song, Yujin Cheon, Soowon Oh, Jaehyeong Choi, Se-Young Yun

    Preprint, 2026.

    arXiv Code

Selected Work

Experience

Samsung Advanced Institute of Technology (SAIT)

Staff Software Engineer

Worked on AI accelerator system software and neural graphics processor projects, with emphasis on workload analysis, performance profiling, software stack design, and functional simulation.

  • LLM workload analysis and profiling for AI accelerator systems.
  • Software stack design and functional simulator implementation for neural graphics processors.

Sep 2022 — Present

Samsung Electronics Co., Ltd. System LSI

Software Engineer

Developed and validated GPU system software across compiler, kernel driver, and firmware layers for mobile GPU projects.

  • GPU GLSL compiler performance tuning, benchmark profiling, and compiler debugging.
  • GPU kernel driver validation and ARM Cortex-A53 firmware development.

Feb 2016 — Aug 2022

Computer System & Platform Laboratory, SNU

Graduate School Researcher

Researched operating-system design for many-core platforms through the implementation of the A2 operating system on the Intel SCC.

  • Designed an SMP kernel model for a many-core platform.
  • Implemented bootloader and network-driver components and validated the OS with multicore benchmarks.

Jan 2014 — Feb 2016

Honors

Recognitions

Samsung SW Certificate - Expert Programmer

Samsung Electronics Co., Ltd.

Certified at the Expert Programmer level in Samsung's internal software engineering certification program.

2020