A preprint compares three pruning methods and finds SparseGPT and WANDA keep fixed sparse-autoencoder metrics closer to dense-model results.