Hojoon Leo Kim
Undergraduate Student, Electrical and Computer Engineering, Seoul National University
hojoon.kim@snu.ac.kr
37, Samseong-ro 51-gil
Gangnam-gu, Seoul 06280
Republic of Korea
Hello, my name is Hojoon Kim, and I also go by Leo.
As an undergraduate researcher passionate about advancing ML systems through hardware-software co-design, I explore how traditional computer architecture principles such as caching, branch prediction, and multiprocessing can address inefficiencies in modern ML workloads. My work spans low-bit quantization, storage-assisted inference, and cache-driven planning for embodied AI agents, with publications at OSDI'25, ICML'25 (Spotlight), and submissions to MLSys'26. I aim to develop practical system architectures that make next-generation ML applications both efficient and deployable at scale by rethinking system abstractions across the entire computing stack.
When I’m not deep in code, you can probably find me on the tennis court🎾!
News
| Jan, 2026 | Our work on QUESO, a storage-assisted quantization error compensation method for on-device LLM inference, has been submitted to ICML'26! |
|---|---|
| Jan, 2026 | Working on HELIOS, a scheduler for multi-robot serving on embedded devices. It hooks CUDA kernels using a light-weighted shared object (.so). Submitted to ICML'26! |
| Dec, 2025 | Working on a scheduler for multi-robot serving on embedded devices for ICML'26. It hooks CUDA kernels using a light-weighted shared object (.so). |
| Nov, 2025 | Our team won the Grand Prize (1st Place) at the 2025 AI Chip Contest - NPU Optimization Track hosted by the Ministry of Science and ICT (MSIT), Republic of Korea, receiving a prize of KRW 10,000,000. |
| Nov, 2025 | Our work on AgenticCache, a cache-driven asynchronous planning system for embodied AI agents, has been accepted to MLSys'26! |
| May, 2025 | Our DecDEC has been accepted to OSDI'25! |
| May, 2025 | Excited to share that our ICML'25 submission, FlashTP, is currently under consideration for a Spotlight or Oral presentation. Also, it has already been successfully deployed in real-world industrial applications. |
Selected Publications
* indicates equal contribution
- ICML’26 (In Submission)
HELIOS: Heterogeneous Lightweight VLA Model Serving System2026In submission - MLSys’26
AgenticCache: Cache-Driven Asynchronous Planning for Embodied AI Agents2026 - ICML’26 (In Submission)
QUESO: Storage-Assisted Quantization Error Compensation for On-Device LLM Inference2026In submission
