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 vision
LLMs know many things.
Agents must know what to trust, what to question, and what to do next.
I build explicit world models that serve as the working memory of language-model agents: small, testable, uncertainty-aware workspaces where knowledge is checked against evidence, repaired where it fails, and consolidated for future use.
- Externalize what is uncertain, fallible, or decision-relevant
- Verify it against evidence and revise what fails
- Archive verified results with their evidence
- Internalize what proves stable; forget the rest
- 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.
Read the full research statement
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. My 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. In domains where experiments are costly, generating candidate hypotheses is now far cheaper than testing them, so the bottleneck of intelligent behavior shifts from generation to verification and selection: deciding which explanations to believe, which experiment would best tell them apart, and which part of existing knowledge to revise when evidence disagrees.
Both language models and explicit models can make predictions that are checked against observations. The difference lies in what is falsified when a prediction fails. An explicit model commits to mechanisms and assumptions, so a failure can be traced to specific components, tested by targeted experiments, and repaired locally. This structure makes calibrated uncertainty, local revision, and extrapolation possible, though it does not guarantee them; they have to be earned empirically.
I therefore treat an explicit world model as the working memory of a language-model agent. This builds on the distinction in model-based machine learning between a compact world model and a large amortized inference machine, but starts from a pretrained language model as a proposal mechanism and a source of background knowledge, one that arrives entangled with an implicit world model of its own and is not assumed to perform calibrated inference for the explicit one. Knowledge is brought into working memory when it must be tested, verified and revised there, and archived with its evidence as an explicit record. Records that prove stable and recurrent are internalized back into the model; the rest are retained for exact computation or forgotten.
Because working memory reasons explicitly about only a few variables while the language model stands in for the rest of the world, its central difficulty lies at that boundary. Errors arise when the language model's summary of the outside world is wrong, or when evidence depends on variables left outside. Understanding this boundary tells an agent what to make explicit, and the utilities it pursues decide at what resolution, and for how long, that knowledge is kept.
These capabilities matter well beyond science. Any agent that acts in the world, whether it is optimizing an engineering process, diagnosing a failure, or carrying out a long task, needs to know what is grounded in evidence, what remains uncertain, and what to check next.
Research directions
Language models bring broad prior knowledge to propose hypotheses; probabilistic inference holds those hypotheses accountable to evidence.
Inside working memory
1Verifying and revising explicit knowledge
How accurately, and with how few experiments, can explicit knowledge be verified and repaired?
More
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. I study this verification engine across numerical laws, causal structure, relational rules, and the laws that govern how actions change the world.
Because the value of explicit verification lies where implicit knowledge fails, I evaluate it on mechanisms that language models do not already know, and I always include the language model's own direct predictions as a competing candidate.
At the boundary
2Deciding what to make explicit
When a language model knows some things and not others, what should an agent bring into working memory?
More
Working memory is small, so deciding what to load into it matters as much as the inference performed inside it. Knowledge deserves to be made explicit when it is uncertain, when the language model is likely to be wrong about it, and when a decision depends on it. 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.
I study how to turn a language model's beliefs into calibrated priors over explicit hypotheses without counting the same evidence twice; how errors arise at the boundary between working memory and the knowledge outside it, and how to diagnose them; and how to use domain-specific records that no language model has seen, kept with their provenance and collection conditions.
Across tasks and time
3Knowledge for action, over time
Under multiple and changing goals, how should knowledge be used, kept, and internalized?
More
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. I study planning with discovered laws, for example in open-ended game environments, and engineering uses such as optimizing a process or diagnosing why a system failed in terms of the assumptions a model makes explicit.
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 supplied in context, and only stable, broadly useful ones are internalized into the model, with the explicit record serving as a regression test. I also study self-improving agents that extract reliable knowledge from their own trajectories and consolidate it into themselves.
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 directions 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
- 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