Reproduce “Weights Read and Write Features: Scalable Parameter Decomposition Grounded in Activation Space” 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
Tue M. Cao, Lisiane Pruinelli, My T. Thai
ASPD links weight components to model activation features, using shared sparse representations and reconstruction objectives to ground components and reproduce matrix outputs. In GPT-2 small’s indirect object identification task, removing an identified query–key component pair substantially suppresses attention to the repeated-name position. This demonstrates an effect on that attention pattern, not that the model gives the wrong final answer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.