Minkyoung Song

prof_pic.jpg

Hi! I’m a second-year M.S. student in Artificial Intelligence at POSTECH, advised by Prof. Namhoon Lee.

I’m currently an AI Research Engineer Intern at Nota AI.

My recent work quantifies uncertainty for explainable AI (XAI), and I’m now working on model compression — pruning, quantization, and joint compression-optimization methods.

I hope to make ML and DL effortlessly accessible in everyone’s daily life.

Feel free to reach out to me via minkyoung@postech.ac.kr anytime!😊

news

  1. 💼 Joined Nota AI as an AI Research Engineer Intern (NetsPresso Tech · XPU Enabler Team)! 🚀 💼
  2. 🏆 Won the Grand Prize (1st Place) at the 5th Human Rights & Gender Equality Contest hosted by the POSTECH Human Rights Center! 🥳 🏆
  3. 🏆 Our paper has been selected for an Oral Presentation and received the Outstanding Paper Award at the KCC 2026 XAI Workshop! 🥳 🏆
  4. 🎤 Our ACL 2026 paper has been selected as an Oral Presentation!🤩 🎤

publications

  1. ACL
    acl2026.png
    Quantifying Aleatoric Uncertainty of In-Context Learning for Robust Measure of LLM Prediction Confidence
    Jinseok Chung , Minkyoung Song* , Hyunji Jung* , and Namhoon Lee
    ACL (Oral), 2026
    KCC 2026 XAI Workshop Outstanding Paper Award (Oral)
    * Equal contribution
    arXiv Code
    경향신문 전자신문 KPI뉴스 매일신문 경북매일신문 경북도민일보 전기신문

latest posts

Efficient Inference Inference speed isn't just a model-size problem PagedAttention, speculative decoding, and early exit all make inference faster without removing a single parameter. Three papers that show "smaller" and "faster" are different goals. Read → Efficient Inference Why joint compression beats doing it in sequence Pruning and quantization interact. APQ (CVPR 2020) stops pretending they don't — and the hard part turns out to be making the joint search cheap, not smart. Read → Efficient Inference Quantizing LLMs: three ways to stop outliers from setting the scale Naive low-bit quantization breaks on large language models because of a few huge activations. LLM.int8(), SmoothQuant, and GPTQ each take a different route around them. Read →