A model's ability to apply learned reasoning patterns to new, unseen problems beyond its training distribution.