Stable SAE features dominate functional signal while unstable features concentrate in reproducible subspaces, allowing more stable SAEs via cross-seed pooling.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders
Stable SAE features dominate functional signal while unstable features concentrate in reproducible subspaces, allowing more stable SAEs via cross-seed pooling.