B.S. Candidate, Chungbuk National University · Research Intern, Artificial Intelligence Laboratory (AI Lab)
World Models · World-Action Models · Model-Based Planning · Tool-Using Language Agents

Homepage · CV (PDF) · Email · LinkedIn

I am an undergraduate researcher in Information and Communication Engineering at Chungbuk National University. My research focuses on world models and world-action models: learning compact models of the environment, planning inside them, and reading actions out of them efficiently. I also explore world models as decision modules for tool-using language agents. Previously, I worked on multimodal reinforcement learning for autonomous driving.

Seeking research internship positions in graduate labs working on world models and robot learning (Winter–Spring 2027).

• How Many Candidates Does a One-Step World Model Need? A Random-Codebook Account of Nearest-of-m Training on Frozen DINO Tokens — TMLR in prep.
  A one-step nearest-of-m world model on frozen DINO tokens covers 0.93 of the enumerated future modes vs. 0.79 for 50-step flow matching, at about 30× lower latency. [code]
• Planners Exploit Deterministic Error: Dithered Low-Bit World Models — KAIA 2026 poster · RSS 2027 in prep.
  Planners exploit deterministic rounding error in low-bit TD-MPC2 world models; training-free dithering removes a median 95% of the success drop on Meta-World. [code]
• Near-Miss Blindness: Latent World-Model Imagination Loses Resolution at the Outcome Boundary — KAIA 2026 poster · TMLR in prep.
  Latent world-model imagination blurs near misses: success-vs-near-miss AUROC falls from 0.953 to 0.853 on DINO-WM PushT, while far misses stay separable. [code]
• Imagine Once, Act Fast: Lightweight Action Readout from a Frozen Driving World-Action Model — KAIA 2026 poster · TMLR in prep.
  Reads the driving trajectory from a frozen 6B world-action model's first imagination pass, cutting decision latency up to 11.9× on PhysicalAI-AV. [code]
• A World Model for LLM Agents: Comparing a Decision Model with General-Purpose LLMs — KIIS 2026 oral
  A frozen decision model (JEV) works as an LLM agent's world model without task training: 93.4% success in TextWorld vs. 79.5% for a failure-memory baseline. [code]

• Oct. 2026 · Selected as a top-10 finalist at the 2026 Southeast Region Multi-Domain Mobility Business Hackathon (drift world model for sea rescue).
• Oct. 2026 · Paper accepted for an oral presentation at the KIIS 2026 Fall Conference.
• Oct. 2026 · Three papers submitted to the KAIA 2026 Conference as posters.
• Oct. 2026 · Advanced to the final round of the KRAFTON AI R&D Hackathon (world-action models).
• Oct. 2026 · Folding Next passed round 1 of the Startup for All Project (모두의 창업).
• Jun. 2026 · Joined the Artificial Intelligence Laboratory as a research intern (Prof. Keon Myung Lee).

Education, projects, competitions, awards, and skills are in the full CV: Homepage · PDF
_{Updated November 9, 2026.}