HierarchicalTopK trains a single sparse autoencoder that reconstructs transformer activations well at many sparsity levels, matching or beating separate per-level models.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy
HierarchicalTopK trains a single sparse autoencoder that reconstructs transformer activations well at many sparsity levels, matching or beating separate per-level models.