Search papers, labs, and topics across Lattice.
1
0
Feature splitting in traditional Sparse Autoencoders can be mitigated by using learned subspace decoders, leading to improved interpretability and efficiency in mechanistic analysis.
