The KAIST research team that won the ICML 2026 international AI challenge. From left: Na Byeong-hu, a researcher at the KI Robotics Research Institute; Kwak Ji-seok, a doctoral student at the Graduate School of Data Science; Cho Su-hyeon, a doctoral student in the Department of Industrial and Systems Engineering; and Kim Tae-woo, a master's student. [KAIST]
The KAIST research team that won the ICML 2026 international AI challenge. From left: Na Byeong-hu, a researcher at the KI Robotics Research Institute; Kwak Ji-seok, a doctoral student at the Graduate School of Data Science; Cho Su-hyeon, a doctoral student in the Department of Industrial and Systems Engineering; and Kim Tae-woo, a master's student. [KAIST]

The practical application of AI that can interpret both images and text, apply the laws of physics and reason independently has moved a step closer to reality.

KAIST announced Wednesday that a research team led by professor Moon Il-chul of the AX department won the international challenge at the International Conference on Machine Learning (ICML) 2026.

The competition evaluated AI systems on their ability to understand physics problems presented in both visual and text formats, apply physical laws, and generate not only correct answers but also full solution processes.

A total of 139 teams competed, including those from leading research institutions such as ETH Zurich, Fudan University and the Shanghai Innovation Institute. The KAIST team achieved the highest score in the final evaluation to claim first place.

Competing AI systems were required to integrate multimodal information and apply physical laws — including Newtonian mechanics — to solve problems logically. The challenge assessed not simple computational ability but the capacity to understand and reason about complex physical phenomena.

The result demonstrates that AI can go beyond solving exam-style questions to understand real physical environments and optimize design and operations. The technology is expected to find applications across aviation, robotics, autonomous driving and space systems, where next-generation AI will need to account for real-world physical conditions in design and decision-making.

Five researchers from KAIST's Applied AI Laboratory took part, including Kwak Ji-seok, a doctoral student in the Department of Industrial and Systems Engineering. The team developed a multi-agent AI architecture in which multiple AI foundation models were each configured as independent agentic AI systems that cross-verify and debate each other's answers.

The approach reduced individual AI errors and improved reasoning reliability, enabling the team to achieve top performance worldwide. It also pointed to new possibilities for AI reasoning on complex physical phenomena and for agentic AI model design.

"This research demonstrated that reliable reasoning can be secured simply by orchestrating existing frontier models and explicitly reconstructing visual information, without developing a larger single model," Moon said. "We plan to extend the extraction module to more challenging visual materials such as handwritten diagrams, and to advance adaptive control techniques that dynamically allocate computational resources according to problem difficulty."

KAIST President Bae Choong-sik called the achievement "a remarkable feat showing that AI is advancing beyond merely generating information to understanding the real physical world and solving difficult problems facing industry and society." He added that KAIST would redouble its efforts to develop next-generation AI source technologies to lead future industries and to cultivate global AI talent.


nbgkoo@heraldcorp.com