ITML

Foundations of modern machine learning

We build mathematical frameworks that explain the behavior of foundation models — large language models, vision-language models, and the agents built on them — and use that understanding to design better algorithms.

Our research follows the lifecycle of a foundation model — pre-training → post-training → inference — and at every stage we ask the same two questions: what is really happening inside the model? (theory) and how can we do it better? (algorithms).

Pillar 1 · Pre-training

How foundation models learn representations

A foundation model begins by learning representations from massive multi-modal data. We build the theory of contrastive and multi-modal pre-training — the geometry embeddings converge to, why failures such as class collapse arise, and how concepts are encoded across modalities — and turn that understanding into pre-training methods that learn better representations.

Ongoing · in preparation
Theory of multi-modal embeddings beyond two modalitiesRepresentation learning under long-tailed distributions
Pillar 2 · Post-training

How foundation models adapt

A pre-trained model is then adapted: fine-tuned to downstream tasks, continually updated, aligned via reinforcement — and sometimes made to unlearn. We characterize exactly what changes at this stage (in-context learning ability, prior knowledge, memorization capacity, representational structure) and design post-training methods with provable guarantees rather than purely empirical recipes.

Ongoing · in preparation
LLM post-training with verifiable rewards (RLVR)Token-level LLM unlearning
Pillar 3 · Inference

How foundation models reason & reach deployment

At inference time the model meets the world: it reasons, learns in context, and must be evaluated and served under real budgets. We study the computations transformers can carry out at inference (reasoning, in-context learning), put LLM evaluation on rigorous statistical footing, and compress models — asking what is essential and what is disposable — so capability survives deployment.

Ongoing · in preparation
Theory of LLM reasoningLong-context QA for agentic tasksDiffusion vs. autoregressive decoding for agentic tasksCompression of image generative models
Applications

From theory to practice

Every stage of this pipeline is tested against real-world domains, alongside domain experts. Current applications include healthcare and psychiatry (cognitive behavioral therapy alignment, psychiatric interviews, cognitive-distortion classification), finance (question answering and retrieval-augmented generation), and law (Korean legal QA). These deployments create real impact while surfacing new theoretical questions that feed back into the three pillars.

Ongoing · in preparation
LLMs for mental healthcareAI agents for educationRAG & evaluation for financeVision-language-action models for roboticsAI agents for research acceleration

Earlier research directions

Distributed ML

Distributed and federated learning over networks.

  • Buffer-based Gradient Projection for Continual Federated Learning TMLR · 2025
  • Election Coding for Distributed Learning: Protecting SignSGD against Byzantine Attacks NeurIPS · 2020
  • Attack of the Tails: Yes, You Really Can Backdoor Federated Learning NeurIPS · 2020
  • Hierarchical Coding for Distributed Computing IEEE ISIT · 2018

Information / Communication Theory

Coding theory for storage, computing, and wireless systems.

  • Capacity of Clustered Distributed Storage IEEE Trans. Information Theory · 2019
  • Secure Clustered Distributed Storage Against Eavesdropping IEEE Trans. Information Theory · 2019
  • On Reusing Pilots Across Interfering Cells in Massive MIMO IEEE Trans. Wireless Comm. · 2017
  • Pilot Reuse Strategy Maximizing the Weighted-Sum-Rate in Massive MIMO Systems IEEE JSAC · 2017

Funding

  • · IITP — Next-Generation AI development (2025–2029)
  • · NRF — Basic Research Laboratory (2024–2027)
  • · NRF — Outstanding Young Scientist Award (2024–2027)