Direct SAE decompositions of steering vectors are dominated by encoder bias and cannot capture negative feature projections, making them unreliable for interpretation.
Progress update #1 from the GDM mech interp team
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Can sparse autoencoders be used to decompose and interpret steering vectors?
Direct SAE decompositions of steering vectors are dominated by encoder bias and cannot capture negative feature projections, making them unreliable for interpretation.