lolismek/bsf-analysis
A stress-test of block-sparse featurizers (SAEs whose atomic unit is a subspace, not a direction) on a synthetic manifold zoo and real vision models. BSFs beat classic SAEs at recovering manifolds but still split features across blocks; we characterize the split geometry, fix much of it with Tournament Top-K, and extend blocks to crosscoders.
正在读取项目信息…