Lee is leading research on how AI systems interact, focusing on how linguistic resources influence reasoning, how agents can evaluate and discuss problems, and how diverse LLMs can communicate efficiently, leveraging HUFS’s linguistic and cultural strengths.

He stresses cross-agent collaboration and efficient communication among AI systems, aligning with HUFS’s strengths and opening doors to international and Korea-based partnerships.

At ICLR 2026, Lee introduced a continuous evaluation approach where AI agents generate and progressively increase problem difficulty to test models’ reasoning, moving beyond fixed test sets.

This continuous evaluation method uses AI agents to create escalating challenges, enabling flexible assessment of AI reasoning at ICLR 2026.

The work spans the Big Three AI conferences—ICLR, ICML, and NeurIPS—highlighting a broad dissemination footprint.

Related papers appeared at EMNLP 2026 ( Findings ) and in the Information Sciences journal, in collaboration with Korea University.

Two NeurIPS 2026 papers explore (1) direct internal information exchange between AI models without natural language conversion to speed collaboration, and (2) a model-compression approach that preserves critical knowledge in medicine, law, and finance.

The NeurIPS work also covers enabling rapid internal model communication and knowledge-preserving compression during size reduction.

HUFS professor Lee Jun-hyun had four papers accepted across ICML, ICLR, and NeurIPS in 2026, marking a milestone for simultaneous acceptance at all three events.

Lee Junhyun’s four papers (as corresponding author) underscore a prolific year at ICLR, ICML, and NeurIPS for HUFS.

Collaboration involved LG AI Research, University of Illinois Chicago, Capital One AI Foundations, and Chung-Ang University, with participation from Korea-based and international partners.

Industry and academic partners, including LG AI Research, University of Illinois Chicago, Capital One AI Foundations, and Chung-Ang University, contributed to the studies.