4 papers co-authored by Hankuk University of Foreign Studies professor accepted in 3 AI conferences

Four papers co-authored by Lee Jun-hyun, a professor in the Division of Computer Engineering at Hankuk University of Foreign Studies (HUFS), have been accepted at three of the world’s leading artificial intelligence (AI) conferences this year, the university said Wednesday.

The three conferences are the International Conference on Learning Representations (ICLR), the International Conference on Machine Learning (ICML) and the Conference on Neural Information Processing Systems (NeurIPS). Together, they are known as the “Big Three” AI conferences.

Researchers from academia and industry worldwide present their latest findings at the conferences, where papers undergo rigorous peer review, the university said.

Lee was the corresponding author of all four papers: one accepted at ICLR 2026, one at ICML 2026 and two at NeurIPS 2026.

This marks the first time Lee has had papers accepted at all three conferences in a single year, highlighting the breadth of his research in AI system evaluation and collaboration, efficient information exchange and model compression.

The studies were conducted in collaboration with researchers from academic institutions and industry in Korea and abroad, including Chung-Ang University, LG AI Research, the University of Illinois Chicago and Capital One AI Foundations.

The study accepted at ICLR 2026 proposed a new method for evaluating AI models’ reasoning capabilities. Instead of relying on predetermined test questions, the approach uses AI agents to generate new problems and progressively increase their difficulty.

The approach enables more flexible, continuous evaluation than conventional methods based on fixed test sets, making it better suited to rapidly evolving AI systems.

The ICML 2026 study addressed the inefficiency of conventional methods that require multiple AI agents to discuss every problem.

The proposed method activates multiagent discussion only when the system has low confidence in its answer, reducing computational costs to nearly the level of a single AI model while maintaining performance.

Two studies were accepted by NeurIPS 2026, focusing on efficient information exchange among different large language models (LLMs) and AI model compression.

The first study introduced a method that allows different AI models to exchange internal information directly, without converting it into natural language.

The method improved performance while cutting communication time to about one-eleventh of that required by conventional approaches.

The second study proposed a model-compression method designed to reduce model size while preserving knowledge in critical fields such as medicine, law and finance.

The method aims to minimize the loss of essential knowledge and performance as AI models are compressed.

“As different agents increasingly need to make use of their respective strengths and collaborate effectively, our research has focused on interactions among AI systems,” Lee said.

“These include how linguistic resources and characteristics affect AI models’ reasoning performance (EMNLP 2026 Findings), how different agents can evaluate and discuss problems (ICLR and ICML 2026), and how heterogeneous LLMs can communicate efficiently (NeurIPS 2026),” he added.

“This research direction aligns closely with HUFS’ strengths in linguistic and cultural diversity. Building on this foundation, we will continue to expand our research through collaboration with researchers in Korea and abroad,” he said.