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 want to build AI systems that can discover how the world works, accumulate what they learn, and reason with that knowledge.
- Propose hypotheses
- Make predictions
- Seek informative evidence
- Update the world model
- Reason and act, then repeat
- System 2 intelligence
- Deliberate, evidence-grounded reasoning that forms hypotheses, weighs alternatives, recognizes uncertainty, and revises beliefs. Reasoning that is reliable and adaptive, not just longer.
- World model
- A structured, uncertainty-aware account of how a system works: its causal relationships, relational rules, mathematical laws, and mechanisms of change.
Read the full research statement
My research is motivated by a fundamental question: How can an AI system go beyond recognizing patterns to understand the mechanisms behind its observations, and use that understanding to solve new problems? I view this as an important step toward System 2 intelligence: deliberate, evidence-grounded reasoning that can formulate hypotheses, evaluate alternatives, recognize uncertainty, and revise beliefs when new evidence arrives. The goal is not simply to make language models generate longer reasoning traces, but to make their reasoning more reliable, grounded, and adaptive.
I believe world models are central to this goal. By a world model, I mean a structured and uncertainty-aware representation of how a system works: its causal relationships, relational rules, mathematical laws, and mechanisms of change. Rather than storing experience only as unstructured text or implicit model parameters, I am interested in explicit models that an agent can inspect, test, revise, compose, and reuse across different problems.
Ultimately, I envision agents that operate in a continuous loop: propose hypotheses, make predictions, seek informative evidence, update their world models, and use the resulting knowledge to reason and act. In this view, scientific discovery, learning from experience, and intelligent planning are closely connected aspects of the same problem.
These capabilities matter well beyond science. Any LLM agent that acts in the world, whether it is optimizing an engineering process, diagnosing a failure, or carrying out a long task, needs a world model that tells it what is grounded in evidence, what remains uncertain, and what to try next. I see explicit, uncertainty-aware world models as part of the epistemic infrastructure that reliable agents require.
Research directions
Language models bring broad prior knowledge to propose hypotheses; probabilistic inference holds those hypotheses accountable to evidence.
Loop steps 1–3
1Discovering explicit world models
How can AI infer the laws and mechanisms of an unfamiliar system from limited observations, interaction, and existing knowledge?
More
I study the discovery of causal, relational, and mathematical models, as well as the laws that govern how actions change the world: when an action applies and what it does. Language models propose candidate explanations, probabilistic inference evaluates them against the data, and active experimentation selects the observations that best distinguish between competing explanations.
World models can also be built from what is already written down. I am interested in compiling papers, logs, and notes into structured knowledge: the hypotheses, mechanisms, and assumptions they contain, and the evidence that supports each of them.
Loop step 4
2Revising and building on knowledge
How can agents turn experience into lasting knowledge that they can correct and build on?
More
An intelligent system should learn more than isolated solutions to individual tasks. I am interested in how agents can consolidate experiences into persistent, structured knowledge; reconcile new evidence with prior beliefs; and identify and correct mistaken assumptions while preserving what they already explain well. Crucially, agents should be able to build on existing scientific knowledge, determine which known mechanisms apply to a new system, and discover what remains unknown. The aim is to develop world models that can be revised, composed, and reused instead of being reconstructed from scratch for every task.
Reusing knowledge safely requires knowing its scope: under which conditions a piece of knowledge holds and to which tasks it transfers. As a world model grows, it must also be maintained by resolving conflicts and removing redundant or stale knowledge. I am also interested in the boundary between explicit and implicit knowledge: what should be kept in an external world model, what can be left to the prior knowledge of a language model, and how stable knowledge can be internalized into compact skills or model weights.
Loop step 5
3Reasoning and planning with world models
How can learned laws and their uncertainty guide what an agent does next?
More
A world model should do more than explain past observations: it should guide future decisions. I study how agents can plan with the laws they have discovered, for example in open-ended game environments, and how amortized Bayesian inference can deliver fast, calibrated predictions so that uncertainty informs every decision.
The same world models can support engineering applications: planning safe interventions to optimize and control a process, and diagnosing why an experiment, process, or system failed in terms of the assumptions, mechanisms, and constraints the model makes explicit.
I also study self-improving agents that learn from their own experience. A successful trajectory shows what worked, but not why it worked or when it applies. I am interested in extracting reliable, reusable knowledge from interaction trajectories, checking which claims the evidence actually supports, and consolidating that knowledge into the agent itself so that it improves even without the knowledge in its prompt. More broadly, I want agents to improve not only their world models but also how they discover, reason, and learn.
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