한동준 연구실

Focus on algorithmic and theoretical aspects of scalable and heterogeneous federated LLM fine-tuning, addressing privacy concerns when data owners do not want to share privacy-sensitive data.

Developing strategies for personalized, robust, and trustworthy AI services to edge users, including continual learning, LLM safety, and unlearning to handle out-of-distribution scenarios.

Developing algorithms for efficient training, LLM compression, and collaborative reasoning/inference under constraints of limited data, scarce labels, and restricted computation and memory resources.

Studying how representation structures influence model behavior, focusing on interpretable representations and mathematical foundations for principled understanding and control of LLMs, VLMs, and MLLMs.