송용휘
B.S. Candidate, Chungbuk National University
School of Information and Communication Engineering · Cheongju, Korea
I am an undergraduate researcher at Chungbuk National University and a research intern at the Artificial Intelligence Laboratory. My research centers on world models and world-action models: how an agent can learn a compact model of its environment, plan inside it, and read actions out of it cheaply. I study where these models fail under efficiency constraints, such as planners exploiting the errors of low-bit world models and latent imagination losing near-miss outcomes, and how to read driving actions from a frozen world-action model. I also explore world models as decision modules for tool-using language agents. Previously I worked on multimodal reinforcement learning for autonomous driving at the MIP Lab.
Chungbuk National University, School of Information and Communication Engineering · Cheongju, Korea
GPA 3.83 / 4.5
Artificial Intelligence Laboratory (AI Lab), Chungbuk National University · Advisor: Prof. Keon Myung Lee
Researching efficient and reliable world models and their use in LLM agents.
Multimedia Information Processing (MIP) Laboratory, Chungbuk National University · Advisor: Prof. Hyun Soo Kang
Conducted research on multimodal reinforcement learning for autonomous driving.
Pseudo Lab (가짜연구소)
Reviewing world-model research papers, participating in technical discussions, and contributing structured review documents to the project website.
HyperCore AI & Data Analytics Club
Formed and led a club team for hackathons and paper reviews.
Under Review
Planners Exploit Deterministic Error: Dithered Low-Bit World Models
Submitted to the Korean Artificial Intelligence Association (KAIA) Conference (poster), 2026. Extended version to be submitted to Robotics: Science and Systems (RSS) 2027. · Code
Near-Miss Blindness: Latent World-Model Imagination Loses Resolution at the Outcome Boundary
Submitted to the Korean Artificial Intelligence Association (KAIA) Conference (poster), 2026. Extended version to be submitted to Transactions on Machine Learning Research (TMLR). · Code
Imagine Once, Act Fast: Lightweight Action Readout from a Frozen Driving World-Action Model
Submitted to the Korean Artificial Intelligence Association (KAIA) Conference (poster), 2026. · Code
In Preparation
How Many Candidates Does a One-Step World Model Need? A Random-Codebook Account of Nearest-of-m Training on Frozen DINO Tokens
To be submitted to Transactions on Machine Learning Research (TMLR). · Code
K-PragShift: Diagnosing Korean Pragmatic Revisions in Tool-Using Language Models
To be submitted to the EACL 2027 Student Research Workshop.
Imagine Once, Act Fast: Lightweight Action Readout from a Frozen Driving World-Action Model
To be submitted to Transactions on Machine Learning Research (TMLR). · Code
Conference Presentations
A World Model for LLM Agents: Comparing a Decision Model with General-Purpose LLMs
Oral presentation, 2026 Fall Conference of the Korean Institute of Intelligent Systems (KIIS). · Code
Additional Publications
A Comparative Study on the Prevalence of Forensic Flies Using Chicken Corpse
Journal of Science Education for the Gifted, 2020.
* First author.
2026 NASA Space Apps Challenge Team Project
Building a React web app that visualizes CLPS lunar landing sites, Sun/Earth visibility windows, and mission timelines from NASA data.
Team Project
Built the initial LLM-wiki platform that turns HWP, Office, and PDF files into a linked knowledge base.
3rd MixUp AI Hackathon (Upstage × BDAI)
Built AI-agent workflows and backend infrastructure for an educational AI platform.
Personal Project
Compared linear, LSTM, Transformer, and diffusion trajectory forecasters; PCA latent diffusion reached a minFDE of 0.86 at K=16.
Independent Research Project
Analyzing 93 RL-based autonomous driving studies and summarizing key research trends.
Faculty-Supervised Research Project
Implemented Shapley-guided multimodal fusion using RGB, LiDAR, route, and ego-state inputs.
Accepted as a participant in the Seoul Local Event with Team Moonkeeper.
Finalist · SeaSight · Selected as one of the top 10 teams for the final round.
Built a drift world model and drone search planner to find fishers who fall overboard; on 88 real drifter cases, the model cut the search area for 90% containment 2.7×.
Passed the document screening round.
Advanced to the final round.
Built a world-action model that predicts cart–double-pendulum dynamics and selects the next action from the same representation.
Folding Next · Passed the first round.
A desktop agent that helps public servants fill forms and draft documents from in-house files, with user-checked sources.
Advanced to the Phase 2 online hackathon (top 20%).
Worked on predicting pitch-control success probabilities from game context, player history, and tracking data.
Advanced to the final round.
Developed an AI system for automated evaluation of Korean argumentative writing.
Finalist · Selected with Folding for the offline final and demo day (about 20 teams).
Finalist · Solar TutorBoard
Built the agent workflows and backend of a Solar Pro3-based tutoring assistant for lesson reports, payment reminders, and scheduling.
Chungbuk National University
Chungbuk National University
Chungbuk National University
Chungbuk National University