Found 1 document(s)
Page 1 of 1
-
MLP neurons in transformer language models are widely understood to be polysemantic, motivating transcoders - sparse, learned approximations of MLP layers - as interpretable replacement models for mechanistic analysis. H...