Simulated neuron activations derived from natural language explanations yield loss increases no better than zero-ablating the transcoder or SAE component.
These features of text are normally identified by looking for specific words or patterns in the text
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
-
Partially Rewriting a Transformer in Natural Language
Simulated neuron activations derived from natural language explanations yield loss increases no better than zero-ablating the transcoder or SAE component.