Jeiyoon Park

I'm an AI Research Engineer at Smilegate.

My research interests include Role-Playing Agents (RPAs), Game AI, AI-driven Non-Player Characters (NPCs), Retrieval-Augmented Generation (RAG), On-Device AI, and Machine Translation.

Contact: k4ke@korea.ac.kr

News

Sep 2026Our paper was accepted to the AACL-IJCNLP 2026 Main Track.
May 2026I will join Smilegate as an AI Research Engineer, where I will work with the Behavior AI Team to research and develop AI-driven NPC technologies for Lost Ark Mobile.
Apr 2025I will serve on the program committee of BEA-ACL 2025.
Feb 2025Our paper was accepted to the NAACL 2025 Industry Track (Oral).
Nov 2024We showcased SOOPi, an AI assistant based on streamer personas, at G-STAR 2024.
Mar 2024I will serve on the program committee of BEA-NAACL 2024.
Jun 2022Our team won first place at the 2022 Microsoft Azure Virtual Hackathon.
Mar 2021I interned as an NLP researcher at Naver Webtoon (Webtoon AI), where I had a wonderful time working with Chaeheon Gwak and Seho Kim under the mentorship of Daesik Kim.

Research

An Empirical Study of On-Device Translation for Real-Time Live-Stream Chat on Mobile Devices

Jeiyoon Park, Daehwan Lee, Changmin Yeo, Yongshin Han, Minseop Kim

AACL-IJCNLP, 2026

[arXiv] [dataset] [github]

Dynamic Context Adaptation for Consistent Role-Playing Agents with Retrieval-Augmented Generations

Jeiyoon Park, Yongshin Han, Minseop Kim, Kisu Yang

Preprint, 2025

[arXiv] [dataset] [github]

CharacterGPT: A Persona Reconstruction Framework for Role-Playing Agents

Jeiyoon Park, Chanjun Park, Heuiseok Lim

NAACL, 2025 (Oral Presentation)

[arXiv] [bibtex] [github] [poster]

ChatLang-8: An LLM-Based Synthetic Data Generation Framework for Grammatical Error Correction

Jeiyoon Park, Chanjun Park, Heuiseok Lim

Preprint, 2024

[arXiv]

Multimodal Frame-Scoring Transformer for Video Summarization

Jeiyoon Park, Kiho Kwoun, Chanhee Lee, Heuiseok Lim

Preprint, 2023

[arXiv]

Analysis of Utterance Embeddings and Clustering Methods Related to Intent Induction for Task-Oriented Dialogue

Jeiyoon Park, Yoonna Jang, Chanhee Lee, Heuiseok Lim

The Eleventh Dialog System Technology Challenge (DSTC11), SIGdial-INLG 2023

[arXiv] [bibtex] [github] [poster]

Variational Reward Estimator Bottleneck: Learning Robust Reward Estimator for Multi-Domain Task-Oriented Dialog

Jeiyoon Park, Chanhee Lee, Chanjun Park, Kuekyeng Kim, Heuiseok Lim

Applied Sciences (SCI), 2021

[journal] [bibtex]

Projects

SARSA: SOOP AI Realtime Streaming Assistant

SOOP

2026.01.01 - 2026.05.10 (released 2025.04.17)

[video]

SARSA (쌀사) is an AI manager that supports streamers with their broadcast content and hosting, and helps them communicate with their fans. In the SARSA project, I worked on improving the overall framework of the existing service, advancing individual modules such as Automatic Speech Recognition (ASR) and speaker diarization, and developing Live Recap, which provides post-stream feedback to help streamers grow. Through these enhancements, SARSA is evolving into an AI manager that grows alongside streamers.

SOOPi: SOOP Personalized Intelligence

SOOP

2024.09.27 - 2025.12.31 (released 2025.03.26)

[video] [video2] [video3] [video4] [video5]

SOOPi (수피) is an AI assistant based on streamer personas. Through SOOPi, users can chat freely with their favorite streamers and get personalized recommendations for broadcast content they might enjoy. To prevent hateful and harmful behavior, such as sexual harassment, we introduced SHIELD (SOOP's Hate speech Identification Engine for Lasting Defense), a safety classifier that detects various types of adversarial safety risks. Anyone who subscribes to a streamer available on SOOPi can use it. Try it now!

Awards

2022 Azure Virtual Hackathon

Yeram Lim, Moongi Lee, Juhui Heo, Jeiyoon Park

Microsoft

First Place, 2022.06.10

[slide] [github] [video1] [video2]

The Azure Virtual Hackathon aims to tackle social problems such as climate change and accessibility using Azure, Microsoft's cloud platform. Our entry, "Hi Azure!: Helping video creators who are visually-impaired edit videos using Azure Cloud AI!", is a framework that makes video editing and content creation more accessible to visually impaired creators by extracting visual features from video, converting them to speech, and providing them as frame-level descriptions. We won first place in the hackathon, and I also placed first in the Microsoft Cloud Skills Challenge (1/63 = 1.58%).

The 3rd AI SPARK Challenge

Kiho Kwoun, Moongi Lee, Juhui Heo, Jeiyoon Park

Innopolis Foundation

Second Place (2/61 = 3.27%), 2022.04.15

[slide] [github] [press]

We postulate that frames in which emotional information appears frequently are correlated with video highlights. Our model first extracts visual and audio features from a video simultaneously, and then applies an attention mechanism to these features. Finally, it predicts importance scores from the emotion-attended features and generates highlights based on those scores.

2020 ETRI Autonomous Growth Artificial Intelligence Contest

Jeiyoon Park, Yewon Park, Jaehoon Cho, Heuiseok Lim

Electronics and Telecommunications Research Institute (ETRI)

Runner-Up Award (14/43 = 32.55%), 2020.10.20

[video] [slide]

In this project, we propose a novel meta-reinforced multi-domain task-oriented dialog agent that exploits an autonomously growing knowledge base and extends an existing single-domain agent algorithm to multiple domains. Our method is designed to infer information unseen during training, as well as the relations among domain, slot, and value triples, and to respond to domains beyond those it was trained on.

2019 Korean Information Processing System Competition

Jeiyoon Park, Jihwan Kim, Boemjoo Park, Heuiseok Lim

Korean Institute of Information Scientists and Engineers

Encouragement Award, 2019.10

[url]

The 31st Annual Conference on Human and Cognitive Language Technology (HCLT 2019) was held at the Korea Advanced Institute of Science and Technology (KAIST) in Daejeon. We received the Encouragement Award at the 2019 Korean Information Processing System Competition.

ITRC Global Makerthon 2019

Jeiyoon Park, Jihwan Kim, Boemjoo Park, Woowon Jhang, Heuiseok Lim

Ministry of Science and ICT, Hyundai (Sponsor)

Second Place (2/34 = 5.88%), 2019.08

[video] [vlog] [news] [news2] [press]

ITRC Global Makerthon 2019 was held at Hanyang University. Our team proposed simple noise-canceling software for target noises (e.g., cafe or factory noise) and won second place.

The 3rd DongBu Global Electronics Competition 2016

Jeiyoon Park, Jeongse Park

DB Foundation

Runner-Up Award (90/1200 = 7.5%), 2016.11

[slide]

We proposed a simple closet that helps people keep their clothes clean and organized, and we invented a novel hanger sensor and actuator for this project. I collaborated with the designer Jeongse Park, and we were awarded the runner-up prize.

The 6th K-ICT Creative Idea Contest

Jeiyoon Park, Minji Kim, Yeonkyung Kim, Kihwan Shin

Ministry of Science, ICT and Future Planning

Encouragement Award (19/473 = 4.01%), 2016.10

[pdf] [news]

We proposed a smart hospital scenario based on future robot technology. I collaborated with Minji Kim (a nurse), Yeonkyung Kim (a designer), and Kihwan Shin (a computer scientist) on this project, and we won the Encouragement Award.

International Robot Contest & R-BIZ Challenge 2016: Self-Driving Car Mission

Jeiyoon Park, Seunghee Woo

Ministry of Trade, Industry and Energy

Bronze Medal, 2016.10

[map]

This challenge consisted of four missions: (1) traffic sign recognition, (2) autonomous parking, (3) stopping in front of a pedestrian crossing, and (4) a free mission of our choice. We won the bronze medal.

Texas Instruments Innovation Challenge Korea Final: MCU Design Contest 2015

Joonwoo Ahn, Hoosang Lee, YoonJoo Choi, Jeiyoon Park

Texas Instruments Korea

Second Place (2/70 = 2.85%), 2015.11

[slide]

We proposed a novel racing simulator that is fun and safe to enjoy at a much lower price than existing racing simulators. To this end, we invented a new elastic actuator. We won second place.