Inner alignment vs outer alignment

A research distinction between specifying an adequate training objective and whether a learned optimizer, if one exists, pursues that objective.

Outer alignment asks whether the objective supplied to training represents the intended goal. Inner alignment asks whether a model that learns an internal search or optimization process would retain that goal outside the training conditions. The second concept comes from mesa-optimization theory and should not be collapsed into every case of shortcut learning or ordinary distribution shift.

Builder example

For most product teams, the practical checks are still concrete: does the metric represent the desired outcome, and does behavior remain acceptable on shifted and adversarial cases? Those tests do not reveal a model's internal objective, but they can expose proxy failures and unsafe behavior.

Common confusion: A strong evaluation score establishes performance on the tested cases. It neither proves a particular internal goal nor by itself demonstrates an inner-alignment failure.