Jang-Hyun Kim

Hello! I am a machine learning researcher at Apple Foundation Models team, focusing on post-training of LLMs. I earned PhD in computer science at Seoul National University, advised by Hyun Oh Song. I was a visiting scholar at New York University in 2024, hosted by Kyunghyun Cho. Before that, I interned at NAVER AI in 2018 for a speech enhancement project. I completed BSc with Mathematics at Seoul National University in 2019. My studies were supported by the Korea Foundation for Advanced Studies (PhD) and Presidential Science Scholarship (BSc).

Email | CV | Scholar | Github | LinkedIn

Research

My latest research focused on improving contextual memory in AI models, addressing challenges in attention mechanisms (Context Memory) and KV caching (KVzip). These enables effective context management of AI models, facilitating continual personalization, large-database inference, and streamed input processing.

Previously, I led research projects in synthetic data generation, dataset cleaning, causal discovery for human genes, and speech enhancement.

Publications

	Fast KVzip: Efficient and Accurate LLM Inference with Gated KV Eviction Jang-Hyun Kim, Dongyoon Han, Sangdoo Yun NeurIPS, 2026 Paper \| Code \| Blog \| Bibtex
	LensVLM: Selective Context Expansion for Compressed Visual Representation of Text Roy Xie, Dan Friedman, Donghan Yu, Bowen Pan, Christopher Fifty, Jang-Hyun Kim, Xianzhi Du, Zhe Gan, Vivek Rathod, Bhuwan Dhingra NeurIPS, 2026 Paper \| Code \| Bibtex
	KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction Jang-Hyun Kim, Jinuk Kim, Sangwoo Kwon, Jae W. Lee, Sangdoo Yun, Hyun Oh Song NeurIPS, 2025 - Oral Presentation (77/21575=0.35%) Paper \| Code \| Blog \| Bibtex
	Large-Scale Targeted Cause Discovery via Learning from Simulated Data Jang-Hyun Kim, Claudia Skok Gibbs, Sangdoo Yun, Hyun Oh Song, Kyunghyun Cho TMLR, 2025 Paper \| Code \| LM podcast \| Bibtex
	Compressed Context Memory For Online Language Model Interaction Jang-Hyun Kim, Junyoung Yeom, Sangdoo Yun, Hyun Oh Song ICLR, 2024 Paper \| Code \| Project Page \| Bibtex
	Neural Relation Graph: A Unified Framework for Identifying Label Noise and Outlier Data Jang-Hyun Kim, Sangdoo Yun, Hyun Oh Song NeurIPS, 2023 Paper \| Code \| Bibtex
	Dataset Condensation via Efficient Synthetic-Data Parameterization Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, Hyun Oh Song ICML, 2022 Paper \| Code \| Bibtex
	Uncertainty-Based Offline Reinforcement Learning with Diversified Q-Ensemble Gaon An, Seungyong Moon, Jang-Hyun Kim, Hyun Oh Song NeurIPS, 2021 Paper \| Code \| Bibtex
	Co-Mixup: Saliency Guided Joint Mixup with Supermodular Diversity Jang-Hyun Kim, Wonho Choo, Hosan Jeong, Hyun Oh Song ICLR, 2021 - Oral Presentation (53/2997=1.7%) Paper \| Code \| Bibtex
	Spherical Principal Curves Jongmin Lee, Jang-Hyun Kim, Hee-Seok Oh TPAMI, 2021 \| R Journal, 2022 Paper \| R Journal \| Code \| Bibtex
	Puzzle Mix: Exploiting Saliency and Local statistics for Optimal Mixup Jang-Hyun Kim, Wonho Choo, Hyun Oh Song ICML, 2020 Paper \| Code \| Bibtex
	Phase-Aware Speech Enhancement with Deep Complex U-Net Hyeong-Seok Choi, Jang-Hyun Kim, Jaesung Huh, Adrian Kim, Jung-Woo Ha, Kyogu Lee arxiv, 2019 Paper \| Bibtex
	Multi-Domain Processing via Hybrid Denoising Networks for Speech Enhancement Jang-Hyun Kim, Jaejun Yoo, Sanghyuk Chun, Adrian Kim, Jung-Woo Ha arxiv, 2018 Paper \| Code \| Bibtex \| Demo

Projects

	Google's Speaker Verification Code \| Kaggle
	Caricature Generation Code
	Image Mosaic via Mixed Integer Programming Code

Dissertation

Data Optimization for Efficient Deep Learning, PhD Dissertation, 2025 | Paper

Mathematical Backgrounds for Machine Learning, Undergraduate Dissertation (in Korean), 2019 | Paper

Academic Services

Workshop Program Committee / Reviewer

Curated Data for Efficient Learning (ICCV 2025) | Website
Interpolation Regularizers and Beyond (NeurIPS 2022) | Website
ImageNet: Past, Present, and Future (NeurIPS 2021) | Website

Reviewing Activities

NeurIPS (2021-), ICLR (2022-), ICML (2022-), TMLR (2022-)

Template based on Jon Barron's website.