Gihun Cho; 조기훈

I work on post-training and evaluation for large language models.

Projects

2026

Motif 3, Motif Technologies

A 314B-A13B mixture-of-experts LLM, open-weight under MIT. Scored 47 on the Artificial Analysis Intelligence Index v4.1.1, level with MiniMax M3. At release that placed the lab 10th in the world, 5th among those releasing open weights, and 1st in Korea.

  • Generated the supervised fine-tuning data, and ran SFT on a customised torchtitan.
  • Ran reinforcement learning: asynchronous GRPO over a customised NeMo-Gym with vLLM.
  • Built the evaluation pipeline for the agentic and long-context benchmarks, GDPval and AA-LCR.

Weights Tech report Artificial Analysis

Experience

2026 –

AI Research Engineer, Motif Technologies

Post-training: data preparation, supervised fine-tuning, reinforcement learning, and the evaluation pipelines around them.

2023 – 2026

Graduate Researcher, Innovative Radiology AI Lab (iRAIL)

Clinical evaluation of medical LLMs and VLMs; metrics and benchmarks for radiology report generation.

Publications

CREPE: Rapid Chest X-ray Report Evaluation by Predicting Multi-category Error Counts

Gihun Cho, Seunghyun Jang, Hanbin Ko, Inhyeok Baek, Chang Min Park

Proceedings of EMNLP 2025, Main Conference (poster). Suzhou, China.

Six radiology-specific error categories, each regressed by its own head on a BiomedBERT encoder and summed into one score. Trained on 32k synthetic report pairs; agrees with radiologists on ReXVal at τ = 0.786 while running in 9.5 ms per pair, roughly 280× faster than an LLM judge.

ACL Anthology OpenReview Project page Code

3 more publications

Evaluating Open and Closed-Source Language and Vision-Language Models for Multicenter Image-Based Diagnosis in Radiology: A Comparative Study with Reader Performance

Dabin Min*, Gihun Cho*, Kyungmin Jeon*, Jiyoung Lee, Donguk Kim, Kwang Nam Jin, Chang Min Park

RSNA 2024, Cutting-Edge Research (abstract). Chicago, USA. *Equal contribution.

Seven open-source LLMs, two open-source VLMs and two commercial models over 9,409 Radiopaedia cases and 175 external chest cases. Given history and findings, Claude scored 60.0% against readers' 41.8%; given image and history alone, every model fell below them, Claude's own accuracy dropping 38.2% and 50.3% on the two sets. Fine-tuning lifted the open-source models by 9.2 ± 4.3 points.

Exploring the Capabilities of Large Language Model Encoders for Image-Text Retrieval in Chest X-rays

Hanbin Ko, Rong Yang, Gihun Cho, Inhyeok Baek, Donguk Kim, Joonbeom Koo, Changi Kim, Dongheon Lee, Chang Min Park

IEEE Journal of Biomedical and Health Informatics (2026).

A bidirectional LLM text encoder adapted to chest radiograph reports, trained with masked token prediction and supervised contrastive learning over stylistically different but clinically equivalent variants of the same report, then dropped into a dual-tower vision-language framework.

DOI arXiv

SeamXSim: Seamless-textured virtual colonoscopy simulator via unpaired long-term video translation

Seunghyun Jang, Dongheon Lee, Yisak Kim, Gihun Cho, Kwang Woo Kim, Sihyun Kim, Jong Pil Im, Byeong Gwan Kim, Chang Min Park

Computers in Biology and Medicine 198, 111217 (2025).

Seamless colon textures synthesized from a single exemplar by inpainting and outpainting, then carried through an unpaired video translation stage to produce long-term colonoscopic sequences from simulation.

DOI PubMed

Education

2024 – 2026

M.S. Bioengineering, Seoul National University

Advised by Chang Min Park, M.D., Ph.D., at the Innovative Radiology AI Lab.

2018 – 2024

B.S. Biomedical Engineering, Hanyang University

Summa cum laude.

2015 – 2018

Software Development, Sunrin Internet High School

17th president of IWOP, the web development club.