Changwoo Baek

I'm a master's student working on Efficient Multimodal LLMs in the Department of Electronics Engineering at Pusan National University.

My goal is to advance algorithmic efficiency, making AI models lightweight, fast, and reliable enough to run on edge devices (e.g., Mobile, NPU, Robotics platforms). I’m particularly interested in uncovering how (M)LLMs work internally, using that understanding to design generalizable lightweighting algorithms rather than relying on task-specific heuristics.

Research Interests:

KV Cache Compression/Sharing
2D/3D Visual Token Pruning
Parameter-Efficient Fine-Tuning (LoRA)
Multi-Agent Systems

News

[2026/08] Our paper RestoreKV ranked 1st on the NVIDIA KVPress leaderboard 🏆

[2026/06] Our paper 3DZip is accepted at ECCV 2026 🇸🇪

[2026/06] Our paper HYBRIDKV is accepted at ICML 2026 Workshop AdaptFM 🇰🇷

[2026/05] Our paper TA-OCD is accepted at IEEE Robotics and Automation Letters (RA-L)

[2026/01] Our paper AgilePruner is accepted at ICLR 2026 🇧🇷

[2025/06] Our paper KDLR is accepted at IEEE Signal Processing Letters (SPL)

Education

Pusan National University, Busan, Republic of Korea (Mar. 2025 - Feb. 2027 (Expected))
M.S. Department of Electronics Engineering
Advisor: Prof. Kyeongbo Kong

Pusan National University, Busan, Republic of Korea (Mar. 2019 - Feb. 2025)
B.S. Department of Electronics Engineering

Selected Publications

(* equally-contributed, † corresponding author)

Preprint

[P1] RestoreKV: Recovering Full-Cache Behavior Under Aggressive Query-Agnostic KV Cache Eviction

Changwoo Baek, Seungjun Shin†, Kyeongbo Kong†

[Project page]

Rank #1 NVIDIA KVPress leaderboard (26.8)

International Conference

[C2] 3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering

Changwoo Baek and Kyeongbo Kong†

ECCV 2026, Sep. 2026 [Project page]

3D Token Pruning [3D VLM]

[C1] AgilePruner: An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models

Changwoo Baek*, Jouwon Song*, Sohyeon Kim*, Kyeongbo Kong†

ICLR 2026, Apr. 2026 [Project page]

Visual Token Pruning [2D VLM]

International Journals

[J2] Target-Aware Change Detection for Efficient VLM Gating in Mobile Robots

Changwoo Baek*, Taehoon Kim*, Woojin Lee*, Jungkwan Son, Seoyoung Lee, Jiyoung Kim, Kyeongbo Kong†

IEEE Robotics and Automation Letters (RA-L), Jun. 2026

ICRA 2027 presentation

Efficient Architecture for Change Detection [Robot Vision]

[J1] Unsupervised Knowledge Distillation via Local Representations for Vision-Language Models

Changwoo Baek, Jouwon Song, Kyeongbo Kong†

IEEE Signal Processing Letters (SPL), Aug. 2025

Knowledge Distillation for Image Classification [VLM]

International Workshop papers

[W1] HYBRIDKV: Exploiting Head-Dominant Reconstruction for Efficient Query-Agnostic KV Cache Compression

Changwoo Baek and Kyeongbo Kong†

ICML 2026 Workshop AdaptFM, Jul. 2026

KV Cache Compression [LLM]

Projects

Developed vision-language model-based technologies for visual perception and situation awareness in home robot systems

Collaborating with LG Electronics, Seoul, Korea (June. 2025 - )

Development of Anomaly and Environment Detection Technology for Home Robot Applications

Collaborating with LG Electronics, Seoul, Korea (June. 2025 - Nov. 2025)

Development of Nighttime Human Detection and Daytime Face Recognition Enhancement for Home Robot Environments

Collaborated with LG Electronics, Seoul, Korea (Dec. 2023 - Dec. 2024)

Awards & Scholarship

Graduate Research Encouragement Scholarship

12,000,000₩, National Research Foundation of Korea (NRF), 2025

Academic Service

Reviewer

NeurIPS'26, ICLR'27, IEEE Transactions on Image Processing

Extra

Teaching Experiences

[EE1600575] C Programming, Teaching Assistant

Dept. of EE, Pusan National University, Spring 2025

[EE3100695] Digital Circuits Laboratory, Teaching Assistant

Dept. of EE, Pusan National University, Fall 2025

Programming Skills

Python, Pytorch, C, C++, ROS2

Page updated

Google Sites

Report abuse