Reproduce “Comparing Sparsified MLPs and Transcoders: Many Transcoder Features Are Neurons in Disguise” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Stepan Shabalin, Mateusz Piotrowski
The authors compare transcoder features with sign-split, sparsified neurons to test what learned sparse features add to MLP analysis. Across 13 open-source transcoders, neuron-aligned features account for more than half of feature activity in late layers despite forming a smaller share of dictionaries. In Gemma 3 4B, they account for nearly all measured Jacobian alignment, a local faithfulness result, not every feature’s behavioral role.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.