About me
I am an Assistant Professor in the School of Electrical Engineering at Korea University. Previously, I was a postdoctoral researcher at KAIST, working with Prof. Juho Lee, and at Mila and Université de Montréal, working with Prof. Yoshua Bengio. I received my Ph.D. from KAIST, where I was advised by Prof. Sung Ju Hwang.
Research
LLMs know many things.
Agents must know what to trust, what to question, and what to do next.
My lab builds AI agents that think before they act.
Today’s language models answer fast, like intuition. That is powerful, but intuition cannot show its work or tell you when it is wrong. We add a second, slower way of thinking: the agent writes down what it believes as an explicit world model, tests it against evidence, and only then decides what to do.
- System 2 reasoning. Check fast guesses with slow, explicit reasoning.
- World models. Write knowledge down so it can be tested and fixed.
- Agentic AI. Plan, act, and learn from what happens.
- Safe AI. Stay within limits that people set.
For example, an agent learning the rules of a box-pushing puzzle:
- GuessThe language model says: pushing a box always moves it.
- Write it downThe agent writes this as a rule it can test, next to a rival rule: a box stays put if a wall is behind it.
- TestIt tries the one move that tells the two rules apart: pushing a box that stands against a wall.
- UpdateThe box does not move. The agent drops the first rule, keeps the second, and plans its next moves with it.
Research themes
Four keywords, one agent: it reasons slowly when it should, keeps a model of the world it can test, acts on what it has verified, and stays within limits.
System 2 reasoning
Slow, careful thinking on top of fast intuition.
A language model answers instantly, like intuition. We study how an agent can notice when that first answer may be wrong, slow down, and check it with explicit hypotheses, probabilities, and experiments.
We askWhen should an agent stop trusting its first answer, and how should it check it?
More detail
Language models propose candidate hypotheses; probabilistic inference weighs them against evidence; and experiments are chosen to tell the remaining alternatives apart. When a prediction fails, competing explanations of the failure are compared and only the implicated component is repaired, while the rest is kept.
One useful signal is available before any experiment: disagreement between a language model’s intuitive predictions and the predictions implied by the theories it states. We study how to turn a language model’s beliefs into calibrated priors over explicit hypotheses without counting the same evidence twice, and we always include the language model’s own direct predictions as a competing candidate.
Related paper: Bayesian Model Discovery with Language Models: Closing the Loop between Hypotheses and Experiments
World models
Knowledge written down in a form that can be tested.
Instead of leaving knowledge hidden in billions of parameters, we state it as equations, causal graphs, rules, and programs. When a prediction fails, we can see which part was wrong and fix only that part.
We askHow can an agent find the part of its knowledge that is wrong, and repair just that?
More detail
An explicit world model serves as the working memory of a language-model agent: a small, active set of variables, hypotheses, and verified models, held with uncertainty over alternatives. We study this across numerical laws, causal structure, relational rules, and the laws that govern how actions change the world.
Working memory is small, so deciding what to load into it matters as much as the inference performed inside it. Errors arise at the boundary between working memory and the knowledge left outside it, and we study how to diagnose them. Our first study asks whether an agent that selectively makes relevant outside assumptions explicit judges better than one with a fixed workspace, and adapts more efficiently than one with a broader explicit model, under the same proposer, evidence, and budget.
Agentic AI
AI that plans, acts, and learns from what happens.
Agents run experiments, use tools, and carry out long tasks. We study how they can plan with what they have verified, and keep improving from their own experience without reinforcing their own mistakes.
We askHow can an agent keep improving from its own experience without reinforcing its own errors?
More detail
A world model is justified by what an agent can do with it: reuse learned laws for new goals, recover quickly when part of the world changes, and check plans before acting. Goals are many and they change, so utility should decide not only what an agent learns but at what resolution it keeps that knowledge and for how long. Verified corrections are first kept as explicit records, and only stable, broadly useful ones are internalized into the model.
We also study self-improving agents in which a language model only proposes candidate rules, the records an environment returns decide which rules are believed, and a planner rather than the model itself supplies the training signal, so that errors are not learned from the agent’s own answers.
Related paper: Let Agents Design How They Search: Automatic Interface Optimization for Agentic Search
Safe AI
Agents we can inspect, correct, and keep within limits.
As agents act with less supervision, people need to see what an agent believes and why it acts. We study agents whose decisions pass through explicit beliefs and stay within limits that goals cannot override, even when the language model inside cannot be fully trusted.
We askIf we cannot fully trust the model inside an agent, what can we still guarantee about what it does?
More detail
Decisions computed from explicit beliefs by a fixed procedure can be traced to those beliefs and tested by intervening on everything else. We study what this requires, where implicit knowledge can still bypass it, and how faster learned policies can be checked against the explicit procedure. A small layer of externally specified norms takes priority over goals; we study how such norms are applied under uncertainty, including why an agent under a probability threshold may prefer not to look for evidence, and how a duty of inquiry can close that gap.
The language model and external sources are treated alike as inputs whose origin may have an interest in the conclusion. Their influence on belief can be bounded, and claims that can be checked are audited against evidence they cannot shape. What cannot yet be bounded is omission, a model that never proposes the decisive alternative, and we study it by evaluating agents against proposers that are adversarial by construction. In the longer term, we ask how a controllable agent might help defend against agents directed to cause harm.
For researchers: the technical picture
Four parties shape what an agent does. Only one of them decides what it believes.
| What is considered | What is believed | What is done | |
|---|---|---|---|
| Language model | proposes hypotheses and plans | a bounded prior, nothing more | no role |
| Evidence | no role | decides | no role |
| Goals | direct attention | no role | choose among permitted actions |
| Norms | require inquiry where harm is possible | no role | set the limits |
- Explicit world model
- Knowledge stated as executable hypotheses (equations, programs, causal graphs, logical rules) that make testable predictions and expose the mechanisms and assumptions behind them.
- Working memory
- The small, active set of variables, hypotheses, and verified models that an agent needs for its current inference, test, or decision, held with uncertainty over alternatives.
- Untrusted model
- A pretrained language model proposes and summarizes, but is not assumed to be honest. What it contributes is recorded, bounded in specified ways, and checked against evidence it cannot shape.
Language models know a great deal about the world, but their knowledge is stored in parameters. It is not attached to the evidence that supports it, it carries no calibrated estimate of its reliability, and when part of it is wrong for a particular system, there is no explicit place where the error can be found and repaired. Our research asks how an agent that knows many things can also know what to trust, what to question, and what to do next.
The question is becoming urgent for two reasons. In domains where experiments are costly, generating candidate hypotheses is now far cheaper than testing them, so the bottleneck shifts from generation to verification and selection. And language models are increasingly deployed as agents that act with decreasing supervision. For an agent, the deficit extends from what it knows to what it does: the beliefs behind an action cannot be inspected before it is taken, and the reasons a model gives need not be the reasons that produced its behavior. As models grow more capable, the value of explicit knowledge shifts from an instrument of accuracy to an instrument of oversight, and in that role it grows with capability rather than against it.
We therefore treat an explicit world model as the working memory of a language-model agent. A pretrained language model proposes hypotheses and supplies background knowledge; probabilistic inference, under stated evidence models, decides what is believed; and goals decide what the agent attends to, never what it believes. Knowledge is brought into working memory when it must be tested, repaired there, and archived with its evidence. Records that prove stable are internalized back into the model; the rest are kept for exact computation or forgotten. Because working memory reasons explicitly about only a few variables, its central difficulty lies at its boundary, and understanding that boundary tells an agent what to make explicit.
The same workspace can make action accountable. If decisions are computed from its contents by a fixed procedure, the language model can influence behavior only through items it places there, where they carry a status and can be examined. A small layer of externally specified norms constrains what may be pursued, with priority over the agent’s goals, and neither goals nor norms are allowed to reweight belief.
One assumption runs through this work: the language model cannot be assumed honest, and training a capable model from scratch is beyond the resources of an academic laboratory like ours. We take this constraint as a question rather than an excuse. When the model that proposes hypotheses and summarizes the unmodeled world cannot be trusted, how much can the structure around it guarantee, and where exactly does it fail? The answer is informative either way, and because the structure does not depend on what fills the model’s place, a more trustworthy predictor could later be studied in exactly the roles where this design is weakest.
The first claim to be tested is the simplest: that selectively making the relevant assumptions explicit improves judgment and adaptation enough to justify its cost. The rest of the agenda builds on that answer.
Prospective students
I welcome motivated students from electrical engineering, computer science, mathematics, and related fields who are excited by questions about intelligence, reasoning, and learning. If these themes interest you, please feel free to contact me.
- Undergraduate interns
- No prior AI or ML experience required. Curiosity, initiative, and enthusiasm for research matter most.
- Master’s students
- Some background in AI or ML, shown through coursework, research internships, or relevant experience.
- Ph.D. students
- At least one first-authored paper at a leading AI/ML venue, such as NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR, or ICCV.
For graduate applicants, I strongly recommend a research internship in my group for at least three months before applying.
Contact haebeomlee [at] korea [dot] ac [dot] kr
Awards
- 2021Google Ph.D. Fellowship
- 2019–2021Global Ph.D. Fellowship
- 2020, 2022Outstanding Reviewer, ICML (Top 33% and Top 10%)
Publications
Recent preprints
- arXiv 2026
- arXiv 2026
Let Agents Design How They Search: Automatic Interface Optimization for Agentic Search
*: equal contribution
- arXiv 2026
Bayesian Model Discovery with Language Models: Closing the Loop between Hypotheses and Experiments
*: equal contribution as second authors, †: co-corresponding authors
Conference publications
- NeurIPS 2026
ProPolar: Progressive Polar Decomposition for Implicit Neural Representations
†: co-corresponding authors
- NeurIPS 2025
Cost-Sensitive Freeze-thaw Bayesian Optimization for Efficient Hyperparameter Tuning
- ICML 2025
- ICLR 2024
Delta-AI: Local Objectives for Amortized Inference in Sparse Graphical Models
*: equal contribution
- ICLR 2022Spotlight
Online Hyperparameter Meta-Learning with Hypergradient Distillation
- ICLR 2022
Sequential Reptile: Inter-Task Gradient Alignment for Multilingual Learning
*: equal contribution
- ICLR 2022
Meta-Learning Low Rank Covariance Factors for Energy-Based Deterministic Uncertainty
- ICML 2021
- NeurIPS 2020Spotlight
- Interspeech 2020
- ICML 2020
Meta Variance Transfer: Learning to Augment from the Others
- ICLR 2020Oral
- ICLR 2020
- NeurIPS 2018
- NeurIPS 2018
- ICML 2018
Earlier preprints
- arXiv 2022
Dataset Condensation with Latent Space Knowledge Factorization and Sharing
*: equal contribution
- arXiv 2020
- arXiv 2018
