Portrait of Kyeongyoon Lee

Kyeongyoon Lee

M.S. Student

AIM Lab, SKKU

Seoul, South Korea

lky3685@skku.edu

About Me

I am Kyeongyoon Lee, an M.S. student at AIM Lab, Sungkyunkwan University (SKKU). I received my bachelor’s degrees in Statistics and Computer Science from the University of Seoul. My research explores how multimodal AI systems can listen, see, and reason more efficiently.

Research Direction

I am interested in capable and efficient multimodal models that understand long and dynamic real-world contexts.

My research interests include:

  • Multimodal Models for joint reasoning across audio, vision, and language.
  • Token Compression for reducing computation and memory while preserving useful evidence.
  • Video Understanding through temporal representations and memory mechanisms.

News

2026
Aug Spatio-Temporal Audio Language Modeling for Dynamic Sound Sources is accepted to EMNLP 2026 Main Conference.
Jun Hierarchical Multimodal Memory for Training-Free Video Moment Retrieval will appear at RespMultimodal’26, a KDD 2026 Workshop.

Publications

  1. 🔎 How can language models understand moving sound sources across space and time?
    Figure 1 overview of the ST-Audio Encoder and ST-AudioLM EMNLP

    Spatio-Temporal Audio Language Modeling for Dynamic Sound Sources

    Oh Hyun-Bin, Kazuki Shimada, Yuhta Takida, Kim Sung-Bin, Toshimitsu Uesaka, Takashi Shibuya, Kyeongyoon Lee, Tae-Hyun Oh, Yuki Mitsufuji

    Empirical Methods in Natural Language Processing (EMNLP), 2026. Main Conference.

    OpenReview
  2. 🔎 Can audio-visual dynamics identify the tokens an Omni-LLM should preserve?
    A-PACK deferred audio pruning and local audio-visual dynamics overview arXiv

    Deferred Audio Pruning with Local Audio-Visual Dynamics for Omni-LLMs

    Kyeongyoon Lee, Hongyeob Kim, Youngeun Kim, Sungeun Hong

    arXiv preprint, 2026.

    TL;DR A two-stage, training-free framework that preserves information-dense audio before the LLM, compresses video with local audio-visual dynamics, and prunes low-relevance multimodal tokens after query-conditioned interactions emerge.

    PDF
  3. 🔎 Can reusable multimodal memory retrieve precise moments without training?
    Hierarchical Multimodal Memory architecture overview KDD Workshop

    Hierarchical Multimodal Memory for Training-Free Video Moment Retrieval

    Kyeongyoon Lee, Hongyeob Kim, Sungeun Hong

    RespMultimodal’26: Responsible Multimodal Foundation Models for Knowledge Discovery, KDD 2026 Workshop.

    TL;DR Builds reusable hierarchical memory from visual captions and ASR, then uses query-aware proposals and multimodal re-ranking to retrieve precise video moments without training.

    PDF

Domestic

GPT and Stable Diffusion-based Reading Activity Service
SeYun Bae*, Kyeongyoon Lee*, JinSu Lee, Hogyun Jeon, Hyunggu Jung
KSC 2023 Undergraduate Division · *equal contribution

Background

Education

  • 2026.02 — PresentM.S. in Immersive Media Engineering, AIM Lab, Sungkyunkwan University (SKKU)
  • 2018.03 — 2024.08B.S. in Statistics & B.E. in Computer Science, University of Seoul (UOS)

Experience

  • 2025.09 — 2026.02Lab Intern, Department of Immersive Media Engineering, SKKU
  • 2025.03 — 2025.06Undergraduate Research Program, KAIST School of Computing · Prof. Oh Tae-Hyun
  • 2024.01 — 2025.03Loan Comparison Platform Team, Nonghyup Bank
  • 2023.06 — 2023.12Undergraduate Research Intern, Time-series TFT ML, Department of Statistics, UOS
  • 2023.01 — 2023.04DataHub Intern, Humax Mobility

Honors & Awards

  • AWS Certified Solutions Architect – Associate (C03)
  • Shinhan Big Data Hackathon Award
  • Hanium Contest ICT Company and Media CEO Award
  • KSC 2023 Undergraduate Division Encouragement Award