Jiseung Jeong

I am a B.S. student in Computer Science and Engineering at UNIST, and an undergraduate research intern at the Ubiquitous AI Lab, advised by Prof. Taesik Gong. My research interests are in efficient and interpretable on-device foundation models, with a current focus on Vision-Language-Action (VLA) systems.

wjdwltmd1151 [at] unist [dot] ac [dot] kr

CV  /  GitHub  /  LinkedIn

Jiseung Jeong

Research
Premover Premover: Fast Vision-Language-Action Control by Acting Before Instructions Are Complete
Joonha Park, Jiseung Jeong, Taesik Gong
Preprint, 2026
arXiv

Projects
Fused Dequantization Fused Dequantization for Quantized LLM Attention
CUDA kernels for 4-bit dequantization (NF4, MXFP4, NVFP4) fused into QKV projection, cutting attention latency 20–25% on RTX 3090. Course project for CSE412 (UNIST), 1st / 13 teams.
code
QuantShift QuantShift
Studied self-speculative decoding under 4/8-bit quantization; showed the speedup is bounded by the AR baseline, so NF4 gains nothing from tuning exit depth. Course project for CSE402 (UNIST), 1st / 21 students.
code
Shadow Boxing Tracking with 3 Devices Shadow Boxing Tracking with 3 Devices
Recognized and logged boxing motions by fusing IMU data from an earbud, phone, and smartwatch. Course project for CSE465 (UNIST), 1st in final presentation / 8 teams.
code
YOLOkemon App YOLOkemon App
A YOLO-based Pokedex object-collection app using a post-training-quantized (PTQ) model. Course project for CSE465 (UNIST).
video

Design and source code from Jon Barron's website.