compositional generalization

Compositional generalization is the ability to combine familiar pieces into something genuinely new. If a model knows how to translate to French and knows how to write a limerick, can it write a limerick in French it has never seen? Humans do this constantly: we recombine known words, rules, and skills to handle situations we have never met. The hope is that a model which has truly learned the parts can assemble them on demand, rather than only repeating whole combinations it saw during training.

In practice models are partly good at this and partly not, and the gap is one of the deepest questions in the field. They blend styles, formats, and domains impressively, yet they can also fail on a clean composition of two skills they each handle alone, especially as the number of pieces grows or the combination drifts far from the training distribution. Strong compositional generalization is a sign of real understanding rather than memorised templates, so probing where it holds and where it breaks is a careful way to map what a model has actually grasped.