JY Jeongho Yoon

NLP / LLM Research

Jeongho Yoon

I work on representation-level methods for efficient, privacy-aware, and reliable language-model systems.

View publications Google Scholar Research notes GitHub Email

My work focuses on practical constraints in pretrained language models: lower compute, reliable retrieval, privacy-preserving inference, calibrated uncertainty, and safety evaluation. I study post-hoc methods for transforming, editing, compressing, and evaluating model representations.

Research Directions

Representation-centered research interests

01

Representation Control

Post-hoc methods for editing, transforming, compressing, and interpreting pretrained model representations.

  • Sparse autoencoders
  • Language identity removal
  • Evidential transformation
02

Efficient Retrieval Systems

Embedding and RAG systems that improve retrieval quality, self-verification, and model efficiency without relying only on scale.

  • Layer truncation
  • Dense retrieval
  • Self-verifying RAG
03

Privacy, Uncertainty, and Safety

Practical methods for privacy-preserving inference, post-hoc uncertainty estimation, and adversarial safety evaluation.

  • Text-free inference
  • Code red teaming
  • Safety evaluation

Publications

Selected publications

2026

2025

First Author Efficiency

Less Is Enough: Turning LLMs into Efficient Embedders via Layer Truncation

Jeongho Yoon, A. So, Heuiseok Lim.

Annual Conference on Human and Language Technology, pp. 14-18.

RAG Self-Verification

KULLM-RAG: A RAG-Specialized Large Language Model with Intelligent Self-Verification

J. Lee, M. Kim, Jeongho Yoon, S. Hong, Y. Jang, S. Lee, J. Seo, C. Park, J. Park, et al.

Annual Conference on Human and Language Technology, pp. 135-140.

Background

Engineering background for language-model research

Education

Studied Software Engineering and Mechanical Engineering at Sejong University.

Systems Foundation

Built projects across FEM simulation, autonomous-driving perception/control, and reinforcement-learning workflows before focusing on NLP systems.

Recognition

Received an Academic Excellence Scholarship and Silver Prize in the Sejong SW-AI Hackathon Python track.

Themes and Skills

Methods and systems experience

Technical Themes

  • Representation-level control: post-hoc transformation, sparse-autoencoder editing, language identity removal, evidential transformation.
  • Efficient LLM use: layer truncation, pruning-to-encoder, context compression, long-context chunking.
  • Trustworthy deployment: privacy-preserving inference, uncertainty estimation, red teaming, safety evaluation.
  • Retrieval systems: RAG-specialized LLMs, dense retrieval, embedding training, self-verification.

Selected Skills

LLMs Transformers Embedding Models Dense Retrieval RAG Long-Context Processing Hugging Face Sparse Autoencoders Uncertainty Estimation Privacy-Preserving NLP Safety Evaluation Python PyTorch / TensorFlow LoRA Fine-tuning Simulation and Control