Confounding control
lastmod 2026-10-30created 2026-10-303 edits
Confounding control is the process of removing hidden factors that create spurious correlations between a cause (action) and an effect (state change).
Concept
A 'confounder' is a variable that affects both the cause and the effect, creating the illusion that the two have a direct causal relationship. Confounding control is a technique that separates out the influence of such hidden variables so that only the pure causal effect is measured.
Role in world models
It prevents the bias that arises when a world model is trained only on observational data. If confounding is not controlled, the model wrongly learns an unrelated factor as the cause of the outcome and makes completely wrong predictions in new environments or conditions.
Example
Suppose that on a rainy day (confounder) people use umbrellas (action A) and traffic jams occur (outcome B). A model without confounding control may wrongly learn that "using an umbrella causes traffic jams." Confounding control separates out the common cause, 'rain', to identify the true causal relationship.