Work
Public research and policy write-ups: methods, results, and the approaches that did not work.
-
Which representations do small models learn for finite-group multiplication?
Small models trained to multiply in a finite group learn only a compressed, near-minimal slice of the group's representation theory, far short of the whole regular representation. To work out which slice, and to tell the competing accounts of the algorithm apart, the project builds matched pairs of groups those accounts cannot distinguish, so any difference the models show counts against a theory.