A transformer LM trained with TopK activations in its hidden layers produces sparse, SAE-like internal representations without needing post-hoc sparse autoencoder training.
Rigging the lottery: Making all tickets winners
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
TopK Language Models
A transformer LM trained with TopK activations in its hidden layers produces sparse, SAE-like internal representations without needing post-hoc sparse autoencoder training.