Hyunjae Ra
Hyunjae Ra wearing a navy hanbok and holding flowers in front of traditional Korean architecture

Hyunjae Ra

M.S. Student · Sungkyunkwan University
AI & Media Lab (AIM Lab) · Seoul, Korea

About Me

I'm an M.S. student in Artificial Intelligence at Sungkyunkwan University, advised by Prof. Sungeun Hong. My research spans efficient model learning, model merging, multimodal learning, hallucination mitigation, and trustworthy multimodal AI.

I work on making large multimodal models efficient to adapt and reliable in their reasoning, with the goal of making capable AI practical under limited compute. I'm always happy to discuss research and collaborations.

Research interestsEfficient Model LearningModel MergingMultimodal LearningHallucination MitigationTrustworthy AI

News & Updates

A few recent updates

Publications

Preprints & international

2026

Module-level analysis of adaptation strategies for multimodal model merging
KDDW 2026August 2026

Where to Adapt for MLLM Merging? A Module-Level Analysis

Hyunjae Ra, Aecheon Jung, Sungeun Hong

KDD Workshop on Responsible Multimodal Foundation Models for Knowledge Discovery

TL;DR We examine where to adapt multimodal large language models before merging them. A module-level comparison explores how the choice of adaptation target relates to the behavior of the merged model.

Paper ↗Workshop ↗
Domestic publications5Show publicationsHide publications

2026

Training scope comparison for multimodal model integration
KIBME 2026June 2026

Analysis of Model Integration Performance according to Training Scope in Multimodal Large Language Models

Hyunjae Ra, Aecheon Jung, Sungeun Hong

The Korean Institute of Broadcast and Media Engineers · Summer Conference

TL;DR We study how the scope of training affects model integration in multimodal large language models. The analysis compares different training scopes to examine their relationship with integration performance.

Clustering-based task-vector quantization for model merging
IPIU 2026February 2026

Clustering-Based Task Vector Quantization for Ultra-Low-Bit Model Merging

Hyunjae Ra, Aecheon Jung, Seunghwan Lee, Sungeun Hong

Workshop on Image Processing and Image Understanding

TL;DR We explore clustering-based quantization of task vectors for ultra-low-bit model merging. The approach connects compact representations of model updates with the task of combining independently trained models.

↗ Outstanding Poster Presentation Award

2025

Comparison of VGGT and SfM initialization for 3D Gaussian Splatting
JBE 2025August 2025

Comparison of VGGT and SfM in Generating Initial Points for 3D Gaussian Splatting

Reagan Koo, Jun-Hyeong Park, Hyunjae Ra, Eun-Seok Ryu

Journal of Broadcasting Engineering · Letters

TL;DR We compare VGGT and structure-from-motion as sources of initial 3D points for Gaussian Splatting. The study focuses on the choice of initialization method within the 3D reconstruction pipeline.

Point cloud augmentation and mesh reconstruction pipeline
KIBME 2025June 2025

Point Cloud Augmentation via 3D Gaussian Splatting and Mesh Reconstruction using NKSR

Hyunjae Ra, Jun-Hyeong Park, Minsu Park, Eun-Seok Ryu

The Korean Institute of Broadcast and Media Engineers · Summer Conference

TL;DR We combine point cloud augmentation through 3D Gaussian Splatting with mesh reconstruction using NKSR. The pipeline connects the generation of additional point data with the construction of a mesh surface.

Level-of-detail mesh generation based on object density
KIBME 2025June 2025

LoD Mesh Generation Based on Object Density Using 3D Gaussian Splatting

Minsu Park, Jun-Hyeong Park, Hyunjae Ra, Yeong-Gyu Kim, Eun-Seok Ryu

The Korean Institute of Broadcast and Media Engineers · Summer Conference

TL;DR We explore level-of-detail mesh generation using object density and 3D Gaussian Splatting. The approach uses density as a guide for deciding how much geometric detail to represent in a reconstructed mesh.

Research Projects

Multimodal AI in practice
AI+X immersive learning innovation program: education, research, industry collaboration, and international exchange
NRF · BK21 FOURParticipation – Present

AI+X-Based Research and Education Unit for Immersive Learning Innovation

AI+X 기반 몰입형 학습혁신 교육연구단

Connecting AI, education, XR, and games to support learner analysis, personalized interventions, and immersive learning experiences.

AI in Education · Immersive Learning · Learning Analytics

Multimodal framework for depression and anxiety screening
MOTIE · KEITParticipation – Present

Multimodal Depression & Anxiety Screening

우울·불안장애 진단 및 중재적 치료를 위한 멀티모달 AI 디지털 의료기기 개발

Screening depression (PHQ-9) and anxiety (GAD-7) from counseling text, audio, and video with frozen foundation-model encoders and a trained decision module.

Multimodal Learning · Mental Health · Digital Diagnostics

Education & Experience

Education

B.S. in Computer Education

Sungkyunkwan University

Overall GPA: 4.14/4.5 · Major GPA: 4.26/4.5

Early graduation in 7 semesters

Research experience

Honors & Awards

Teaching

Teaching Assistant

Introduction to Artificial Intelligence · AAI2009_03

–
Teaching Practicum Student

Computer Education · Hagye Middle School

Cite RED

@misc{ra2026relevantevidence,
  title={Relevant Evidence Decoding for Audio-Visual Hallucination Mitigation},
  author={Hyunjae Ra and Aecheon Jung and Jungin Park and Sungeun Hong},
  year={2026},
  eprint={2610.02976},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2610.02976}
}