Integrated M.S./Ph.D. Student in Electronic Engineering
Sogang University, Seoul, South Korea
I am advised by Prof. Suk-Ju Kang in the VIDS Lab, and I received my B.S. in Computer Engineering from Hansung University.
My research focuses on efficient AI: making large models lighter and faster at inference. I am particularly interested in token pruning, sparse attention, and KV cache compression for multimodal and generative models.
* Equal contribution · † Corresponding author
arXiv 2026
Sparse attention that folds the unread tail of the KV cache into a few count-aware cluster proxies, with an ErrorGate that spends the exact-read budget on the clusters with the largest estimated error. Keeps 99% of dense accuracy at a 1% read on RULER-32K HARD with Llama-3.1-8B, and its kernel runs decode attention 10× faster than dense at a 1% read.
arXiv 2026
A training-free block-sparse attention method for video visual autoregressive generation. At each expensive scale, it builds a fresh per-head mask instead of reusing one from an earlier scale or clip, with low selector overhead.
ICML 2026 Workshop on Resource-Adaptive Foundation Model Inference (AdaptFM)
Shows that the Top-K oracle used to evaluate sparse attention in video diffusion transformers cuts off the tail of the attention distribution, and proposes a low-delta oracle policy with Unified Tail Aggregation, which adds one aggregated logit to restore it. Up to 97.4% lower MSE than the Top-K oracle at 50% sparsity.
Integrated M.S./Ph.D. in Electronic Engineering, VIDS Lab
Advisor: Prof. Suk-Ju Kang
B.S. in Computer Engineering