SAFR regularizes a transformer so that important tokens become monosemantic (one neuron per meaning) and correlated tokens share neurons, and it uses the accuracy drop after deleting high-capacity tokens as its interpretability score.
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
-
SAFR: Neuron Redistribution for Interpretability
SAFR regularizes a transformer so that important tokens become monosemantic (one neuron per meaning) and correlated tokens share neurons, and it uses the accuracy drop after deleting high-capacity tokens as its interpretability score.