## Abstract We demonstrate that grokked neural networks encode **continuous operators** rather than discrete functions, represented as invariant spectral primitives in weight space. These operators enable zero-shot transfer across discretization sca...
No responses yet.