Integrated M.S./Ph.D. Student · Korea University

Jeongho Yoon

I study representation-level methods for language-model systems, with work in raw-text-free inference, multilingual retrieval, post-hoc uncertainty estimation, and input-side safety evaluation. I am now extending this foundation toward efficient reasoning and multimodal models.

Open to research internships and collaborations.

Email me View publications Google Scholar

GitHub Hugging Face LinkedIn Paper notes (Korean)

Selected Publications

Research through concrete interventions

I work at several model interfaces - embeddings, latent inputs, and logits - and evaluate how those interventions affect utility, efficiency, and model behavior.

Jeongho Yoon is shown in bold; an asterisk (*) denotes equal contribution.

2026

ACL 2026 Long Paper First Author Privacy-Aware Inference

Towards Privacy-Preserving Large Language Model: Text-free Inference Through Alignment and Adaptation

Jeongho Yoon, Chanhee Park, Yongchan Chun, Hyeonseok Moon, Heuiseok Lim.

PPFT trains a client-side encoder, server-side projection, and LLM so the server can condition on pooled, noise-injected prompt embeddings rather than raw text, and evaluates the resulting privacy-utility trade-off.

My role: I originated the core idea and led the experiments, analysis, and manuscript preparation.

ACL 2026, Volume 1: Long Papers, pp. 25961-25981. DOI: 10.18653/v1/2026.acl-long.1191.

Paper arXiv

ACL 2026 Long Paper Equal Contribution Multilingual Retrieval

LangSAE Editing: Improving Multilingual Information Retrieval via Post-hoc Language Identity Removal

Dongjun Kim*, Jeongho Yoon*, Chanjun Park, Heuiseok Lim.

LangSAE uses a sparse autoencoder to identify and suppress language-associated features in pooled embeddings. On the reported multilingual retrieval benchmarks, it improves ranking quality and cross-language coverage without retraining the base encoder or re-encoding raw text.

My role: I led the experiments and analysis; I also co-wrote the manuscript.

ACL 2026, Volume 1: Long Papers, pp. 36374-36389. DOI: 10.18653/v1/2026.acl-long.1685.

Paper arXiv Code

CVPR 2026 Highlight Uncertainty Estimation

Evidential Transformation Network: Turning Pretrained Models into Evidential Models for Post-hoc Uncertainty Estimation

Yongchan Chun, Chanhee Park, Jeongho Yoon, Jaehyung Seo, Heuiseok Lim.

ETN learns a lightweight transformation of pretrained logits into Dirichlet evidence for in- and out-of-distribution uncertainty estimation across vision classification and LLM question answering.

My role: I contributed foundational experimental work and helped prepare the manuscript.

CVPR 2026, pp. 6157-6166. Selected as a Highlight.

Paper arXiv Code

Mathematics 2026 Equal Contribution Code Safety Evaluation

Code Redteaming: Probing Ethical Sensitivity of LLMs Through Natural Language Embedded in Code

Chanjun Park*, Jeongho Yoon*, Heuiseok Lim.

Code Redteaming inserts ethically problematic natural language into comments, identifiers, and other natural-language-bearing regions of Python and C code, then evaluates how reliably 18 LLMs detect it.

My role: I led the experiments and analysis; I also co-wrote the manuscript.

Mathematics 14(1), Article 189. DOI: 10.3390/math14010189.

Article

2025

First Author Efficient Embeddings

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

Jeongho Yoon, Aram So, Heuiseok Lim.

Repurposes a decoder-only LLM as a text embedder by truncating upper transformer layers and measuring the retained retrieval quality.

My role: I designed and conducted the study end to end.

37th Annual Conference on Human & Cognitive Language Technology (HCLT 2025), pp. 14-18.

HCLT 2025 Outstanding Paper RAG

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

Jungseob Lee, Minhyuk Kim, Jeongho Yoon, Seongtae Hong, Youngjun Jang, Seungyoon Lee, Jaehyung Seo, Chanjun Park, Jeongbae Park, Heuiseok Lim.

Fine-tunes a Korean-English RAG-specialized model for retrieval-need prediction, distractor rejection, grounded response generation, and self-verification.

My role: I contributed to the experiments and analysis.

37th Annual Conference on Human & Cognitive Language Technology (HCLT 2025), pp. 135-140.

Research Agenda

Making advanced models reason more efficiently

My next goal is to reduce redundant computation and intermediate generation in reasoning models while maintaining answer quality, and then extend these methods from language to multimodal systems.

Efficient Reasoning

Develop methods that improve the trade-off between answer quality and computational cost by reducing redundant computation, context processing, and intermediate generation.

Primary direction for the next 2-3 years

Representations & Interfaces

Study interventions at the level of embeddings, latent inputs, and logits, including methods that leave the base model unchanged and methods that adapt selected components.

Current publication through-line

Multimodal Generalization

Extend methods for efficient reasoning and representation-level control from language models to multimodal models.

Emerging research direction

Background

Research training with an engineering foundation

Mar 2025 - Feb 2030 (expected) Korea University

Integrated M.S./Ph.D. Program in Computer Science and Engineering

Graduate researcher at the NLP & AI Lab, advised by Prof. Heuiseok Lim.

Mar 2019 - Feb 2025 Sejong University

B.Eng. in Mechanical Engineering

Double major in Software Engineering, with project work spanning simulation, autonomous driving, reinforcement learning, and machine learning.

Earlier Research & Engineering

Jul 2023 - Feb 2025

Undergraduate Researcher, Sejong Optimal Structure Lab

Implemented and debugged finite-element simulation components for beam and shell structures using C++ and MATLAB, including 6-DOF representations and coordinate transformations.

2019 - 2024

Autonomous-Driving Systems

Worked across perception, route planning, and control using LiDAR, cameras, ROS, RRT*, Pure Pursuit, PID control, MATLAB/Simulink, and Gazebo-style simulation.

Selected Recognition

Outstanding PaperHCLT 2025 · KULLM-RAG
Second PrizeMathWorks MATLAB Student AI Challenge 2024
First PrizeSejong University College of Engineering Academic Conference 2024
Silver PrizeSejong SW-AI Hackathon, Python Track · 2023
Academic Excellence ScholarshipSejong University · 2023

Methods & Skills

Methods tied to published and implemented work

Research Methods

Sparse autoencoders, representation alignment, dense retrieval, RAG, layer truncation, evidential deep learning, inversion attacks, and ID/OOD evaluation.

Implementation

Python, PyTorch, Hugging Face Transformers, TensorFlow, LoRA fine-tuning, model evaluation pipelines, data processing, and experiment automation.

Systems Foundation

C/C++, MATLAB/Simulink, ROS, Gazebo, finite-element simulation, image-processing pipelines, autonomous-driving control, and reinforcement learning.

Contact

Research internships and collaborations

For internship and collaboration inquiries, please contact me by email.

aa007878@korea.ac.kr

PublicationsGoogle Scholar CodeGitHub Models & CollectionsHugging Face Professional ProfileLinkedIn