KAIST unveils AI that learns to theorize like a child

The KAIST research team has developed a next-generation global model. This is an internal model that AI builds to understand and make predictions about the world, learning actionable theories solely from observation.

KAIST (President Bae Choong-sik) announced on July 15 that a team led by Professor Sung-Jin Ahn of the School of Computing has proposed a new learning paradigm called Learning-to-Theorize (L2T), which trains AI to theorize about how the world works using only observed information. The team also built Neural Theorizer (NEO), a neural network-based model that implements this paradigm.

This research was selected for oral presentation at the 43rd International Conference on Machine Learning (ICML 2026), held in Seoul from July 6th to 11th, and was presented on July 9th. This resulted in this study being in the top 0.7% (168 posts) out of a total of 23,918 posts. This paper was also selected for the Compositional Learning Workshop’s Best Paper Award.

World models are foundational technologies that span robotic control, autonomous driving, generative AI, and autonomous agents (AI systems that can make decisions and act on their own). To date, world models have primarily focused on predicting what will happen next. Even if a model accurately predicts the next scene, it doesn’t necessarily understand why the change occurred – the underlying principles that govern the world.

A team of researchers has found a solution to how humans learn. Long before children acquire language, they develop their own internal theories about how the world works. Applying this view of developmental cognitive science to AI, researchers have built systems that understand the principles behind the world, not just systems that predict the future.

The team’s proposed L2T framework has no predetermined answers or rules. Given only the “before” and “after” observations, the AI ​​independently discovers which rule caused the change. While traditional AI models focus on “guessing what will happen next,” this approach is built to understand “why change happened.”

To implement this, the team also developed Neural Theorizier NEO. The model discovers reusable primitives hidden within the observed transformations and composes them into executable programs. These learned primitives can be systematically recombined to explain new situations.

For example, NEO independently learns primitives that correspond to basic operations such as rotation, moving to the bottom left, and coloring. Even if it is presented with a combination it has never encountered during training, such as “move down, then color, then rotate,” it can recombine the primitives it has already learned to explain and solve new situations.

In contrast, conventional AI tends to memorize intertwined patterns in isolation, and its performance drops rapidly when faced with unknown combinations. Through a variety of experiments, the team demonstrated that NEO outperforms existing approaches at constructive generalization, or the ability to combine learned basic rules to solve never-before-seen problems.

“It points to a new direction beyond the prediction-centric world model, which we call the ‘world theory model.’ We hope this will develop into a core technology across fields such as intelligent robots, autonomous agents, and AI to support scientific discovery,” said Professor Ahn Seong-jin.

Doojin Baek and Gyubin Lee, master’s students in the School of Computing, served as co-first authors of the study.

*Paper title: “Learning to theorize the world from observation” Paper link: https://arxiv.org/abs/2605.03413

*Authors: Baek Doo-jin*, Lee Gyu-bin*, Baek Jun-yeop, Lee Ho-sung, Ahn Sung-jin (*co-first author)

This research was supported by the National Research Foundation of Korea (NRF).