I study agents that must act without seeing the whole state, in environments that change over time and where no one knows the optimal answer.
People rarely see the whole picture either; we decide anyway and adjust as the results come in. Agents should be able to do the same.
About
I am a Ph.D. student in Industrial Engineering at Seoul National University, advised by Prof. Pilsung Kang in the Data Science and Business Analytics Lab. Studying industrial engineering, I got used to starting from how a process actually runs, constraints and all.
My projects usually start from the gap between what a method assumes and the setting people actually work in. ContAccum began when I needed large-batch contrastive training on limited memory: gradient accumulation was not enough, and gradient caching was hard to build on. AlienLM began from privacy frameworks designed for conventional machine learning, which did not account for the compute, data size, and modality of exchanging text with an LLM API.
Now I build environments that resemble industrial processes: complex, partially observable, and changing over time. Nobody knows the optimal decision in them, so even the evaluator can only score each step greedily. An agent needs reasoning, retrieval, and reinforcement learning to work in such a setting, so I study the three together there.
I like simple sentences for difficult problems, and I usually want to know what we mean by a capability before trying to improve it.
Thought for 10 semestersresearch directions
The three directions I keep coming back to.
Reasoning and reinforcement learning
I am interested in how models reason, and in RL as a way to study long-horizon behavior. The harder question often comes before the reward: defining the agent's role, action space, environment dynamics, and feedback structure in a realistic way. I also work on rationale reduction and decoding strategies that keep reasoning quality while staying efficient.
Retrieval
I view retrieval broadly, as channeling external information into a language model. In a changing environment, it is also how an agent finds out what has changed, for example from a new log entry or sensor reading.
Agents that operate over time
Most agent benchmarks ask whether an agent can finish a task. I am more interested in whether it can keep doing the job: noticing when the environment changes, deciding when to step in, and keeping track of a goal that may not have a clear end.
The output is not the answer. The output is the trajectory. My current project builds an environment modeled on industrial processes to study this.
Selected research
-
AlienLMICML 2026
The first text-only method for API-boundary privacy that works on general tasks: alienized text stays useful to the model but is hard for an outside observer to recover, retaining over 81% of plaintext performance.
-
ContAccumNeurIPS 2024
Improves contrastive learning by accumulating comparison queries, not just comparison targets.
-
CheckEvalEMNLP 2025
Reframes LLM-as-a-judge around stability under model and sample changes, not human correlation alone.
News
- 2026-10step 9Updated my CV with project outcomes and reviewing service. [CV]
- 2026-07step 8AlienLM project page with an in-browser tokenizer demo is up. [project page] [arXiv] [code]
- 2026-05step 7AlienLM was accepted to ICML 2026.
- 2026-05step 6Updated my CV with new under-review manuscripts and archived publications.
- 2026-03step 5New manuscripts on belief-shift rewards and semiconductor equipment log retrieval are under review.
show full trajectory
Beyond research
Outside research, I like books, films, dry humor, and diving. Scuba and freediving have recently become my favorite ways to get out of my head. I also run and play badminton when I want to stop overthinking for a while.
I share my days with two cats, Luce and hoyang, who have firm opinions about my working-from-home schedule.
Get in touch
If you work on LLM reasoning, agents, RL, or retrieval, I would be glad to talk. I am looking for a research internship, especially where the processes are complex, hard to measure, and changing over time.
Click the address to copy it, or open your mail app.
Thanks for reading.
↺ keep generating