About me

I look for what actually limits a model, then resolve that limit through analysis and experiment. My work runs along two directions: efficient LLMs through architecture compression, and grounding vision-language-action models in speech.

I am an integrated M.S.–Ph.D. student in Artificial Intelligence at Ajou University, and a member of the MMAI Lab (Multi-Modal Artificial Intelligence).
Advised by Prof. Jongbin Ryu.

Research interests

  • Efficient large language models
  • Efficient finetuning of quantized models
  • Application: sound

News

  • 2026

    Paper accepted to EMNLP 2026: Colla-Q, collaborative experts in MoE quantization via minimax precision balancing.

Publications

2026

  1. Colla-Q: Toward Collaborative Experts in MoE Quantization via Minimax Precision Balancing

    Colla-Q takes an ensemble view of mixture-of-experts quantization, using activation entropy to balance expert performance so that zero-shot accuracy and calibration-domain stability both improve at the same average bit budget.

    EunJu Shin, Jongbin Ryu

    Empirical Methods in Natural Language Processing, 2026papercode

    • LLM
    • MoE
    • Quantization
    • EfficientAI

  2. Don’t Overlook the Learning Trajectory of the LoRA Adapter on Quantized LLMs

    A-LEMA is a trajectory-aware framework that adaptively controls decay during quantized LLM fine-tuning, preserving consistently learned adapter components and improving general-task performance over existing methods.

    EunJu Shin*, Hankyul Kang*, Jongbin Ryu

    Under review2026* equal contribution

    • LLM
    • Quantization
    • LoRA
    • EfficientAI
2025

  1. Pluggable Decoder for Slimmable Neural Networks

    A decoder that plugs into a slimmable network lets one trained model be served at several width budgets, instead of retraining a separate model for each one.

    Hankyul Kang, EunJu Shin, Jongbin Ryu

    Under review2025

    • Vision
    • EfficientAI
    • Slimmable Networks

  2. Music Reactive Video Generation: Highlighted Chromagram and Its Evaluation Metric

    A highlighted chromagram conditions video generation on music, paired with an evaluation metric for how well the generated video actually reacts to it.

    EunJu Shin*, Hankyul Kang*, Jongbin Ryu

    Under review* equal contribution

    • Sound
    • Multimodal
    • Video Generation

In progress

  • Speech-conditioned VLA policies

    Vision-language-action models cannot resolve ambiguous deixis — spatial referring expressions such as “that one over there.” Speech-spatial conditioning grounds the referred object from the speaker’s location, so the policy acts on what was actually meant.

Experience

  • 2024–

    Student Researcher

    MMAI Lab, Ajou UniversityJan 2024 – Present

    Efficient finetuning: quantized-LoRA finetuning. Model compression: MoE quantization and slimmable networks. AI applications: speech-conditioned VLA.

Education

  • Ph.D.

    Integrated M.S.–Ph.D. in Artificial Intelligence

    Ajou University, Suwon, KoreaSep 2024 – Present

    Advisor: Prof. Jongbin Ryu

  • B.S.

    Business Intelligence

    Ajou University, Suwon, KoreaMar 2020 – Aug 2024

    GPA 4.34/4.5, ranked 2nd in the Department of Business Intelligence

Reference

Prof. Jongbin Ryu, Associate Professor at Ajou University and my Ph.D. advisor, jongbinryu@gmail.com.