Reproduce “Sparse Transformer: a Controlled Testbed for Mechanistic Interpretability” 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
Anton Korznikov, Alexey Dontsov, Elena Tutubalina
The Sparse Transformer provides a language-model testbed with known neuron value vectors by retaining only the top eight of 8,192 post-ReLU activations per token. It scores 0.866 on automated AutoInterp, compared with 0.605 for a loss-matched dense model. A Transcoder also recovers value vectors better than an SAE, but results concern five-layer TinyStories models, not production-scale interpretability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.