Seungdong Yoa
AI Research Scientist · Builder

Turning frontier research into systems that ship.

I am an AI Research Scientist at LG AI Research, where I work on large language models — multi-agent reasoning, reliability, and efficient inference.

My work spans language and vision — from multi-agent reasoning and LLM reliability to efficient models for image recognition — and I'm increasingly drawn to multimodal, interactive intelligence: systems that perceive across modalities and collaborate with people and one another. Alongside research, I design and ship the products and ventures that put it to work.

LinkedIn · Email · Google Scholar · CV (PDF)
Seoul, Korea
6
Publications
3
ICML · ICLR · AAAI
'24
LG Award
3
Ventures Built
01

News

2026
Two papers accepted — ICML 2026 (efficient multi-agent reasoning) and ICLR 2026 (agent-centric evaluation of LLM reasoning).
2025
ImagePiece accepted at AAAI 2025 — content-aware re-tokenization for efficient image recognition.
2024
Received the LG Award, LG Group's highest honor, for the Hero LensAA deployment.
02

Selected Research

My long-term goal is multimodal, interactive intelligence — agentic systems that perceive and reason across language and vision, stay reliable in real-world use, and collaborate with people. I pursue it through multi-agent reasoning, model reliability, and efficient inference that makes such systems deployable.

ICML
2026
Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate
Seungdong Yoa, Ye Seul Sim, Suhee Yoon, Sanghyu Yoon, Dongmin Kim, Soonyoung Lee, Bumsoo Kim, Junhyun Lee
A leader-centric debate framework reaching multi-agent reasoning accuracy at near single-agent token cost.
Link
ICLR
2026
From Static Benchmarks to Dynamic Protocol: Agent-Centric Text Anomaly Detection for Evaluating LLM Reasoning
Seungdong Yoa, Sanghyu Yoon, Suhee Yoon, Dongmin Kim, Ye Seul Sim, Junhyun Lee, Woohyung Lim
Link
AAAI
2025
ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition
Seungdong Yoa, Seungjun Lee, Hyeseung Cho, Bumsoo Kim, Woohyung Lim
SOTA-efficiency re-tokenization for ViTs — 1.5× faster with a +0.39% accuracy gain.
Link
Three additional papers (IEEE Access, 2021) on self-supervised anomaly detection, surrogate losses, and meta-learning
Self-Supervised Learning for Anomaly Detection With Dynamic Local Augmentation
Seungdong Yoa, S. Lee, C. Kim, H. J. Kim
IEEE ACCESS · 2021
Learning Non-Parametric Surrogate Losses With Correlated Gradients
Seungdong Yoa, J. Park, H. J. Kim
IEEE ACCESS · 2021
Learning to Balance Local Losses via Meta-Learning
Seungdong Yoa, M. Jeon, Y. Oh, H. J. Kim
IEEE ACCESS · 2021

Full list on Google Scholar →

03

Education

M.S., Computer Science & Engineering, Korea University
2019 — 2022
GPA 4.02 / 4.5. Advised by Prof. Hyunwoo J. Kim. Research on self-supervised anomaly detection, surrogate-loss learning, and meta-learning.
B.S., Computer Science & Engineering, Korea University
2010 — 2017
Major in Computer Science with a minor in Business Administration. National Scholarship for Science & Engineering (full merit-based).
04

Experience

AI Research Scientist, LG AI Research
2022 — Present
Research on LLM reasoning, reliability, and efficient inference — from multi-agent debate protocols and hallucination detection to semantic-aware INT4 quantization. Led the Hero LensAA industrial-AI project, reducing test-set overkill from 99% to near zero — recognized with the 2024 LG Award.
LLM Reasoning Multi-Agent Systems Quantization Hallucination Detection MCP
AI Researcher Intern, Hyundai Motor Group AIRS
Jul — Sep 2021
Self-supervised anomaly detection algorithms for smart-factory quality control.
Graduate Research Assistant, Korea University
2019 — 2022
Led a joint AI project with Samsung Electronics on real-time human–object interaction and anomaly detection for industrial drone monitoring.
Software Engineer, SSG.COM (Shinsegae)
2017 — 2019
Built an OAuth 2.0 server and a scalable integration platform for AI speakers (Google Home, Naver Clova), plus a customer-service chatbot.
05

Ventures & Building

Alongside research, I design and ship products — taking ideas from prototype to live service.

Effectgen Founder · Generative AI
An AI fitting-shot studio for fashion sellers. Upload a garment, pick a model template, and Effectgen renders the piece worn on the model — production-ready fitting shots without a photoshoot.
Effectgen Homeliving Founder · Generative AI
The home-goods vertical of Effectgen, for furniture sellers. Upload a product image, choose an interior template, and the furniture is composited naturally into the scene — generating listing-ready detail and focus shots.
AgentMon Founder · AI Agents
A marketing-automation agent. Set a campaign from a Discord command and it generates and posts content to X (Twitter) on a recurring schedule — an experiment in always-on, agent-driven content pipelines.
06

Honors

2024
LG Award
The most prestigious honor across all LG Group affiliates — for the Hero LensAA project's commercial deployment.
2020
AI Grand Challenge — 2nd Place
Awarded a USD 0.6M research grant for AI problem-solving excellence.
—
National Scholarship for Science & Engineering
Full merit-based scholarship, Korea — top-tier academic performance.
07

Also

Patents
Two registered patents (KR) in consumer-product design — wet-type vacuum cleaner and wet-type air cleaner.
Stack
Python · PyTorch · SQL · Next.js · Java · Spring

Let's talk.

Open to conversations on LLM research, applied AI, and early-stage building. Reach out anytime.

Email LinkedIn Google Scholar CV (PDF)