low-dimensional-structure

This paper studies seed dependence in sparse autoencoders, finding that stable features carry most predictive signal while unstable features reflect reproducible low-dimensional subspaces.