Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words

Open in new window