New task tests how learning systems balance rules and exceptions

Researchers introduced a novel task called transitive inference with exceptions to study how learning systems balance relational generalization and memorization. The task requires learning a relational rule while also memorizing an exception to that rule. Using kernel ridge regression, they analytically characterized model behavior across various representations and parameters, finding that these models can balance generalization and memorization. The work addresses a gap in task paradigms for studying this essential ability.