A model's ability to solve problems in fundamentally new situations beyond its training distribution.