Sangjin Choi

Postdoctoral Researcher

Information & Electronics Research Institute, KAIST

sjchoi [at] casys.kaist.ac.kr

CV Google Scholar GitHub
Sangjin Choi

I design systems software for cost-efficient AI infrastructure by exploiting hardware and workload characteristics. I am a postdoctoral researcher in the CASYS Lab at KAIST.

During my Ph.D., I was advised by Prof. Youngjin Kwon and received additional mentorship from Prof. Jeongseob Ahn and Prof. Myeongjae Jeon. I have also worked closely with Yifan Xiong and Ziyue Yang at Microsoft Research Asia.

ML Systems LLM Serving Memory Systems Energy-Efficient ML

Education

Work Experience

Publications

SIGOPS ATC '26
First author Acceptance rate: 13.9% (135 / 973)
Artifacts Available Artifacts Evaluated: Functional Results Reproduced
ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving
Sangjin Choi, Sukmin Cho, Yifan Xiong, Ziyue Yang, Youngjin Kwon, Peng Cheng

(Work done during an internship at Microsoft Research.)

Paper Code
MLSys '26
Co-author Acceptance rate: 26.8% (135 / 504)
Artifacts Available Artifacts Evaluated: Functional Results Reproduced
BEAM: Joint Resource–Power Optimization for Energy-Efficient LLM Inference under SLO Constraints
Hyunjae Lee, Sangjin Choi, Seungjae Lim, Youngjin Kwon
Paper Code
EuroSys '26
Co-author Acceptance rate: 19.1% (138 / 723)
Artifacts Available Artifacts Evaluated: Functional Results Reproduced
MTTM: Dynamic Fast Memory Partitioning with Bandwidth Optimization for Multi-tenant Cloud
Changjun Lee, Sangjin Choi, Youngjin Kwon
Paper Code
Findings of NAACL '25
Co-author Overall acceptance: 37% (est.)
(Main + Findings)
Lossless Acceleration of Large Language Models with Hierarchical Drafting based on Temporal Locality in Speculative Decoding
Sukmin Cho, Sangjin Choi, Taeho Hwang, Jeongyeon Seo, Soyeong Jeong, Huije Lee, Hoyun Song, Jong C. Park, Youngjin Kwon
Paper Code
ICCAD '23
Co-author Acceptance rate: 22.9% (172 / 750)
PRIMO: A Full-Stack Processing-in-DRAM Emulation Framework for Machine Learning Workloads
Jaehoon Heo, Yongwon Shin, Sangjin Choi, Sungwoong Yune, Jung-Hoon Kim, Hyojin Sung, Youngjin Kwon, Joo-Young Kim
Paper
USENIX ATC '23
First author Acceptance rate: 18% (65 / 353)
Artifacts Available Artifacts Evaluated: Functional Results Reproduced
EnvPipe: Performance-preserving DNN Training Framework for Saving Energy
Sangjin Choi, Inhoe Koo, Jeongseob Ahn, Myeongjae Jeon, Youngjin Kwon
Paper Code Talk
USENIX ATC '22
Co-first author Acceptance rate: 16.5% (64 / 389)
Artifacts Available Artifacts Evaluated: Functional Results Reproduced
Memory Harvesting in Multi-GPU Systems with Hierarchical Unified Virtual Memory
Sangjin Choi*, Taeksoo Kim*, Jinwoo Jeong, Rachata Ausavarungnirun, Myeongjae Jeon, Youngjin Kwon, Jeongseob Ahn (*equal contribution)
Paper Code Talk