Researchers at KAIST have introduced a learning method that trains artificial intelligence to form theories about how the world operates using only observed information. They also created a neural network model that carries out the method. A world model is an internal representation an AI constructs so it can understand and forecast events around it. Earlier world models concentrated mainly on forecasting the next state. Accurate next-state forecasts do not guarantee that the model grasps the governing principle behind the change.
The researchers drew on insights from developmental cognitive science. Young children form internal theories of physical regularities long before they master language. The same idea guided the design of a system that seeks underlying principles instead of mere next-step prediction.
Building theories from observation pairs alone
The model receives no ready-made rules or answers. It is shown only a pair of observations - one before and one after a change - and must itself identify the rule that produced the difference. Conventional systems try to guess what comes next; This one asks why the observed change occurred. It locates simple reusable building blocks, known as primitives, inside the observed transformations and assembles those primitives into executable programs. Once the primitives are learned they can be recombined in new ways. In experiments the neural network independently extracted primitives corresponding to rotation, leftward or downward movement, and coloring. When later shown an unfamiliar sequence such as downward movement followed by coloring and then rotation, the model recombined the already-learned primitives to account for the new case.
Ordinary models often store complex patterns as single entangled units. As a result their accuracy falls sharply when the same elements appear in novel orderings. Systematic tests confirmed that this method achieves stronger compositional generalization - the capacity to reassemble basic learned rules and solve problems never encountered during training. The researchers described the result as a step toward world theory models that go beyond pure prediction. Such models are expected to prove useful for intelligent robots, autonomous agents, and systems that assist scientific discovery. This research is described in a preprint itled "Learning to Theorize the World from Observation."