A mathematical theory of balancing relational generalization and memorization

In three linesTheoretical study on balancing relational generalization and memorization in learning systems. Authors introduce transitive inference with exceptions task and analytically characterize kernel ridge regression models across representations. Validation on pretrained language models shows successful generalization depends on representational geometry, with systematic errors predicted by theory.

Summary generated by Claude — human-verified