LG AI Research Institute Papers Gain Recognition at Prestigious Conference

LG AI Research Institute's Papers Accepted at ICML 2025
One Paper Selected as 'Spotlight,' Representing Top 3%

Three papers published by LG AI Research Institute, focusing on artificial intelligence (AI) technologies applicable to industrial sites, have been accepted at 'ICML 2025,' the most prestigious conference in the field of machine learning.

According to industry sources on June 16, LG AI Research Institute achieved notable results at ICML 2025, with one of its papers being selected as a 'Spotlight' paper, which represents the top 3% of all submissions.

A paper co-authored by a doctoral student from the research team of Professor Sung Youngchul at Korea Advanced Institute of Science and Technology (KAIST) proposed a method in which AI first learns process operation data accumulated by humans and then incorporates additional experiences from real-world environments.

By utilizing this approach, it is possible to fully leverage existing data while maintaining the exploration capabilities and high adaptability that are advantages of reinforcement learning. The algorithm proposed by LG AI Research Institute has been evaluated as 'world-class' in terms of performance.

Additionally, another accepted paper proposed a structure in which each task at industrial sites is handled by independent AI agents, who collaborate with each other as needed.

This research was highly regarded for designing an effective learning process using actual operational data, as it applied the technology to the naphtha cracking center (NCC) process at LG Chem's Daesan facility.

The industrial AI agents developed by LG AI Research Institute are recognized as highly scalable technologies that can be applied not only to petrochemical processes but also to various fields such as robotics and physics-based AI systems.

Furthermore, the conference acknowledged another study that created a transfer stage to seamlessly connect pre-training using accumulated industrial data with reinforcement learning for real-world applications.